text
stringlengths 2.85k
2.55M
| label
class label 11
classes |
|---|---|
Finding Approximate Local Minima
Faster than Gradient Descent
arXiv:1611.01146v4 [math.OC] 24 Apr 2017
Naman Agarwal
Zeyuan Allen-Zhu
Brian Bullins
[email protected]
[email protected]
[email protected]
Princeton University
Institute for Advanced Study
Princeton University
Elad Hazan
Tengyu Ma
[email protected]
[email protected]
Princeton University
Princeton University
November 3, 2016
Abstract
We design a non-convex second-order optimization algorithm that is guaranteed to return
an approximate local minimum in time which scales linearly in the underlying dimension and
the number of training examples. The time complexity of our algorithm to find an approximate
local minimum is even faster than that of gradient descent to find a critical point. Our algorithm
applies to a general class of optimization problems including training a neural network and other
non-convex objectives arising in machine learning.
1
Introduction
Finding a global minimizer of a non-convex optimization problem is NP-hard. Thus, the standard
goal of efficient non-convex optimization algorithms is instead to find a local minimum. This
problem has become increasingly important as the state-of-the-art in machine learning is attained
by non-convex models, many of which are variants of deep neural networks. Experiments in [10,
11, 21] suggest that fast convergence to a local minimum is sufficient for training neural nets, while
convergence to critical points (points with vanishing gradients) is not. Theoretical works have also
affirmed the same phenomenon for other machine learning problems (see [5, 6, 18, 19] and the
references therein).
In this paper we give a provable linear-time algorithm for finding an approximate local minimum
in smooth non-convex optimization. It applies to a general setting of machine learning optimization,
and in particular to the optimization problem of training deep neural networks. Furthermore, the
running time bound of our algorithm is the fastest known even for the more lenient task of computing
a point with vanishing gradient (called a critical point), for a wide range of parameters.
Formally, the problem of unconstrained mathematical optimization is stated in general terms
as that of finding the minimum value that a function attains over Euclidean space, i.e.
min f (x) .
x∈Rd
(1.1)
If f is convex, the above formulation is convex optimization and is solvable in (randomized)
polynomial time even if only a valuation oracle to f is provided. A crucial property of convex
functions is that “local optimality implies global optimality”, allowing for greedy algorithms to
reach the global optimum efficiently. Unfortunately, this is no longer the case if f is nonconvex;
indeed, even a degree four polynomial can be NP-hard to optimize [23], or even just to check
whether a point is not a local minimum [25]. Thus, for non-convex optimization one has to settle
for the more modest goal of reaching approximate local optimality efficiently.
Note that a particular interest to machine learning is the optimization of functions f : Rd 7→ R
of the finite-sum form
n
1X
f (x) =
fi (x) .
(1.2)
n
i=1
Such functions arise when minimizing loss over a training set, where each example i in the set
corresponds to one loss function fi in the summation.
We say that the function f is second-order smooth if it has Lipschitz continuous gradient and
Lipschitz continuous Hessian. We say that a point x is an ε-approximate local minimum if it
satisfies (following the tradition of [28]):
√
k∇f (x)k ≤ ε and ∇2 f (x) − εI ,
where k · k denotes the Euclidean norm of a vector. We say that a point x is an ε-critical point
if it satisfies the gradient condition above, but not necessarily the second-order condition. Critical
points include saddle points in addition to local minima. We remark that ε-approximate local
minima (even with ε = 0) are not necessarily close to any local minimum, neither in domain nor in
function value. However, if we assume in addition the function satisfies the (robust) strict-saddle
property [15, 24] (see Section 2 for the precise definition), then an ε-approximate local minimum
is guaranteed to be close to a local minimum for sufficiently small ε.
Our main theorem below states the time required for the proposed algorithm FastCubic to find
an ε-approximate local minimum for second-order smooth functions.
1
Theorem (informal). Ignoring smoothness parameters, the running time of FastCubic to return
an ε-approximate local minimum is
!
n
n3/4
1
+ 7/4 · Th,1 for (1.2) or Õ
· Th for the general (1.1).
Õ
ε3/2
ε
ε7/4
Above, Th is the time to compute Hessian-vector product for ∇2 f (x) and Th,1 is that for an
arbitrary ∇2 fi (x).
The full statement of Theorem 1 can be found in Section 2.
Hessian-vector products can be computed in linear time —meaning Th,1 = O(d) and Th =
O(nd)— for many machine learning problems such as generalized linear models and training neural
networks [1, 29]. We explain this more generally in Appendix A. Therefore,
Corollary 1.1. Algorithm FastCubic returns an ε-approximate local minimum for the optimization problem of training a neural network in time
!
nd
n3/4 d
Õ
+ 7/4
.
ε3/2
ε
Another important aspect of our algorithm is that even in terms of just reaching an ε-critical
point, i.e. a point that satisfies k∇f (x)k ≤ ε without any second-order guarantee, FastCubic is
faster than all previous results (see Table 1 for a comparison).
The fastest methods to find critical points for a smooth non-convex function are gradient descent
and its extensions, jointly known as first-order methods. These methods are extremely efficient in
terms of per-iteration complexity; however, they necessarily suffer from a 1/ε2 convergence rate [27],
to the best of our knowledge, in previous results only higher-order methods seem capable of breaking
this 1/ε2 bottleneck [28]. For certain ranges of parameters, our FastCubic finds local minima even
faster than first-order methods, even though they only find critical points. This is depicted in
Table 1.
Paper
Gradient Descent (GD)
SVRG [2]
SGD [20]
noisy SGD [16]
a
cubic regularization [28]
Total Time Achieving k∇f (x)k ≤ ε
O nd
ε2
2/3
O nd + n ε2 d
O εd4
O
Õ
this paper
this paper
ndω−1 +dω
ε3/2
Õ
Õ
dC1
ε4
nd
ε3/2
nd
ε7/4
+
n3/4 d
ε7/4
Second-Order Guarantee
n/a
n/a
n/a
∇2 f (x) −ε1/C2 I
∇2 f (x) −ε1/2 I
Table 1: Comparison of known methods.
a
Here C1 , C2 are two constants that are not explicitly written. We believe C1 ≥ 4.
2
∇2 f (x) −ε1/2 I
∇2 f (x) −ε1/2 I
1.1
Related work
Methods that Provably Reach Critical Points. Recall that only a gradient oracle is needed
to reach a critical point. The most commonly used algorithm in practice for training non-convex
learning machines such as deep neural networks is stochastic gradient descent (SGD), also known
as stochastic approximation [30] and its derivatives. Some practical enhancements widely used in
practice are based on Nesterov’s acceleration [26] and adaptive regularization [12]. The variance
reduction technique, introduced in [32], was extremely successful in convex optimization, but only
recently there was a non-convex counterpart with theoretical benefits introduced [2].
Methods that Provably Reach Local Minima. The recent work of Ge et al. [17] showed
that a noise-injected version of SGD in fact converges to local minima instead of critical points,
as long as the underlying non-convex function is strict-saddle. Their theoretical running time is a
large polynomial in the dimension and not competitive with our method (see Table 1).
The work of Lee et al. [24] shows that gradient descent, starting from a random point, almost
surely converges to a local minimum of a strict-saddle function. The rates of convergence and
precise step-sizes that are required are, however, yet unknown.
If second-order information (i.e., the Hessian oracle) is provided, the cubic-regularization method
1
of Nesterov and Polyak [28] converges in O( ε3/2
) iterations. However, each iteration of NesterovPolyak requires solving a cubic function which, in general, takes time super-linear in the input
representation.
One natural direction is to apply an approximate trust region solver, such as the linear-time
solver of [22], to approximately solve the cubic regularization subroutine of Nesterov-Polyak. However, the approximation needed by a naive calculation makes this approach even slower than vanilla
gradient descent. Our main challenge is to obtain approximate second-order local-minima and simultaneously improve upon gradient descent.
Independently of this paper and concurrently1 , Carmon et al. [7] develop an accelerated gradient
descent method that achieves the same running time for finding an approximate local minimum as
in our paper. Remarkably, the same running time is obtained via a very different technique.
1.2
Our Techniques
Our algorithm is based on the cubic regularization method of Nesterov and Polyak [8, 9, 28]. At
a high level, cubic regularization states that if we can minimize a cubic function m(h) , g > h +
1 >
L
3
2
2 h Hh + 6 khk exactly, where g = ∇f (x), H = ∇ f (x), and L is the second-order smoothness of
the function f , then we can iteratively perform updates x0 ← x + h, and this algorithm converges
to an ε-approximate local minimum in O(1/ε3/2 ) iterations. Unfortunately, solving this cubic
minimization problem exactly, to the best of our knowledge, requires a running time of O(dω )
where ω is the matrix multiplication constant. Getting around this requires five observations.
The first observation is that, minimizing m(h) up to a constant multiplicative approximation
(plus a few other constraints) is sufficient for showing an iteration complexity of O(1/ε3/2 ).2 The
proof techniques to show this observation are based on extending Nesterov and Polyak.
The second observation is that the minimizer h∗ of m(h) must be of the form h∗ = (H+λ∗ I)+ g+
v, where λ∗ ≥ 0 is some constant satisfying H + λ∗ I 0, and v is the smallest eigenvector of H
and + denotes the pseudo-inverse of a matrix. This can be viewed as moving in a mixture direction
1
To be precise, their manuscript appeared online approximately 24 hours before ours.
More specifically, we need mt (h) ≤ C1 minh {mt (h)} for some constant C. In addition, we need to have good
bounds on khk and k∇m(h)k.
2
3
between choosing h ← v, and choosing h to follow a shifted Newton’s direction h ← (H + λ∗ I)+ g.
Intuitively, we wish to reduce both the computation of (H+λ∗ I)+ g and v to Hessian-vector products.
The first task of computing (H + λ∗ I)+ g can be slow, and even if H + λ∗ I is strictly positivedefinite, computing it has a complexity depending on the (possibly huge) condition number of
H + λ∗ I [34]. The third observation is that it suffices to pick some λ0 > λ∗ so both (1) the condition
number of H + λ0 I is small and (2) the vectors (H + λ∗ I)−1 g and (H + λ0 I)−1 g are close. This relies
on the structure of m(h).
√
The second task of computing v has a complexity depending on 1/ δ where δ is the target
√
additive error [13, 14]. The fourth observation is that the choice δ = ε suffices for the outer loop
of cubic regularization to make sufficient progress. This reduces the complexity to compute v.
Finally, finding the correct value λ∗ itself is as hard as minimizing mt (h). The fifth step is to
design an iterative scheme that makes only logarithmic number of guesses on λ∗ . This procedure
either finds the correct one (via binary search), or finds an approximate one, λ0 , but satisfying
(H + λ∗ I)−1 g and (H + λ0 I)−1 g being sufficiently close.
Putting all the observations together, and balancing all the parameters, we can obtain a cubic
minimization subroutine (see FastCubicMin in Algorithm 2) that runs in time O(nd + n3/4 d/ε1/4 ).
2
Preliminaries and Main Theorem
We use k · k to denote the Euclidean norm of a vector and the spectral norm of a matrix. For a
symmetric matrix M we denote by λmax (M) and λmin (M) respectively the maximum and minimum
eigenvalues of M. We denote by A B that A − B is positive semidefinite (PSD). For a PSD
matrix M, we denote by M+ its pseudo-inverse if M is not strictly positive definite.
We make the following Lipschitz continuity assumptions for the gradient and Hessian of the
target function f . Namely, there exist L2 , L > 0 such that
∀x, y ∈ Rd :
k∇2 f (x)k ≤ L2
and
k∇2 f (x) − ∇2 f (y)k ≤ Lkx − yk.
(2.1)
Definition 2.1. We assume the following complexity parameters on the access to f (x):
• Let Tg ∈ R∗ be the time complexity to compute ∇f (x) for any x ∈ Rd .
• Let Th ∈ R∗ be the time complexity to compute ∇2 f (x) v for any x, v ∈ Rd .
P
Definition 2.2. We say that f is of finite-sum form if f = n1 ni=1 fi (x) and k∇2 fi (x)k ≤ L2
for each i ∈ [n]. In this case, we define Th,1 to be the time complexity to compute ∇2 fi (x) v for
arbitrary x, v ∈ Rd and i ∈ [n].
Next we define the strict-saddle function for which an ε-approximate local minimum is almost
equivalent to a local minimum [15, 24].
Definition 2.3 (strict saddle). Suppose f (·) : Rd → R is twice differentiable. For α, β, γ ≥ 0, we
say f is (α, β, γ)-strict saddle if every x ∈ Rd satisfies at least one of the following three conditions:
1. k∇f (x)k ≥ α.
2. λmin (∇2 f ) ≤ −β.
3. There exists a local minimum x? that is γ-close to x in Euclidean distance.
We see that if a function is (α, β, γ)-strict saddle, then for ε < min{α, β 2 } an ε-approximate
local minimum is γ-close to some local minimum.
4
Algorithm 1 FastCubic(f, x0 , ε, L, L2 )
Input: f (x) that satisfies (2.1) with L2 and L; a starting vector x0 ; a target accuracy ε.
1/2
1: κ ← 900
.
εL
2: for t = 0 to ∞ do
> 2
3:
mt (h) , ∇f (xt )> h + h ∇ 2f(xt )h + L6 khk3
4:
(λ, v, vmin ) ← FastCubicMin ∇f (xt ), ∇2 f (xt ), L, L2 , κ
min
whichever gives smaller value for mt (h);
h0 ← either v or λv2L
0
6:
Set xt+1 , xt + h
3/2
c is a constant; we proved c = 2.4 ∗ 106 works
7:
if mt (h0 ) > − cε√L then return xt+1 .
8: end for
5:
2.1
Main Results
The finite-sum setting captures much of supervised learning, including Neural Networks and Generalized Linear Models. The main theorem which we show in our paper is as follows:
Theorem 1. FastCubic (Algorithm 1) starts from a point x0 and outputs a point x such that
√
k∇f (x)k ≤ ε and λmin (∇2 f (x)) ≥ − Lε
in total time (denoting by D , f (x0 ) − f (x∗ ))
√
√
1/4
• Õ Dε3/2L · Tg + DLε7/4 L2 · Th , or
√
√
3/4 1/4
• Õ Dε3/2L · Tg + nTh,1 + Dn εL7/4 L2 · Th,1 in the finite-sum setting (see Definition 2.2).
Here Õ hides logarithmic factors in L, L2 , 1/ε, d, and in maxx k∇f (x)k .
Two Known Subroutines. Our running time of FastCubic relies on the following recent results
for approximate matrix inverse and approximate PCA:
Theorem 2.4 (Approximate Matrix Inverse). Suppose matrix M ∈ Rd×d satisfies kMk ≤ L2 and
2
λI + M δI for constants λ, δ, L2 > 0. Let κ , λ+L
δ . Then, we can compute vector x satisfying
(2.2)
x − (λI + M)−1 b ≤ εkbk,
using Accelerated gradient descent (AGD) in O κ1/2 log(κ/ε) iterations, each requiring O(d) time
plus the time needed to multiply M with a vector.
P
Moreover, suppose M = n1 ni=1 Mi where each Mi is symmetric and satisfies kMi k ≤ L2 . If
Mi b can be computed in time O(d0 ) for each i and vector b, then accelerated SVRG [4,
33] computes
a vector x that satisfies equation (2.2) in time O max{n, n3/4 κ1/2 } · d0 · log2 (κ/ε) .
We refer to the running time for this computation as Tinverse (κ, ε) and the algorithm as A.
Above, the SVRG based running time shall be used only towards our finite-sum case in Definition 2.2.
Theorem 2.5 (AppxPCA [3, 13, 14]). Let M ∈ Rd×d be a symmetric matrix with eigenvalues
1 ≥ λ1 ≥ · · · ≥ λd ≥ 0. With probability at least 1−p, AppxPCA produces a unit vector w satisfying
w> M w ≥ (1 − δ× )(1 − ε)λmax (M) . The total running time is Õ(Tinverse (1/δ× , εδ× )).
3
Our Fast Cubic Regularization Algorithm
Recall that the cubic regularization method of Nesterov and Polyak [28] studies the following upper
bound on the change in objective value as we move from a point xt to xt + h: (it follows simply
5
from the Taylor series truncated to the third order)
∀h ∈ Rd :
f (xt + h) − f (xt ) ≤ mt (h) , ∇f (xt )> h +
h> ∇2 f (xt )h L
+ khk3 .
2
6
(3.1)
Denote by h∗ an arbitrary minimizer of mt (h). We propose in this paper a subroutine FastCubicMin
to minimizes mt (h) approximately. Note that FastCubicMin returns two vectors v and vmin . We
min
then choose h0 to be either v or λv2L
, whichever gives a smaller value for mt (h).
Before discussing the details of FastCubicMin, let us first state a main theorem for FastCubicMin:3
Theorem 2 (Guarantees of FastCubicMin). The algorithm FastCubicMin finds a vector h0 that
satisfies
(a) It produces a vector h0 satisfying mt (h0 ) ≤ 0 and
either
3/2
3000mt (h0 ) ≤ mt (h∗ )
ε √
(b) If m(h∗ ) ≥ − 300
, then kh0 k ≤ kh∗ k +
L
√
√ε
4 L
or
and
mt (h∗ ) ≥ −
ε3/2
√ .
800 L
k∇mt (h0 )k ≤
ε
2
.
(c) FastCubicMin runs in time: (using Õ to hide logarithmic factors in L, L2 , 1/ε, d, k∇f (xt )k)
√
2
• Õ (εL)L1/4
· Th where Th is the time to multiply ∇2 f (xt ) to a vector;
√
2
• Õ max n, n3/4 (εL)L1/4
· Th,1 where Th,1 is the time to multiply ∇2 fi (xt ) with a vector.
Above, the first guarantee promises that we are either done (because mt (h∗ ) is close to zero), or
we obtain a 1/3000 multiplicative approximation to mt (h∗ ). Our second guarantee in Theorem 2
promises that when we are done (because mt (h∗ ) is close to zero), the output vector h0 and h∗ are
roughly similar in Euclidean norm and have a small gradient k∇mt (h0 )k. Our third guarantee gives
the time complexity of FastCubicMin.
Now, our final algorithm FastCubic for finding the ε-approximate local minimum of f (x) is
included in Algorithm 1. It simply iteratively calls FastCubicMin to find an approximate minimizer,
3/2
and it then stops whenever mt (h0 ) > − cε√L for some large constant c.
Roadmap. In Section 4 we show why Theorem 2 implies Theorem 1. All the remaining sections
are for the purpose of proving Theorem 2. Because our FastCubicMin is very technical, instead
of stating what the algorithm is right away, we decide to take a different path. In Section 5, we
first state a lemma characterizing “what h∗ looks like”. In Section 6, we provide a set of sufficient
conditions which “look similar” to the characterization of h∗ , and show that as long as these
conditions are met, Theorem 2-a and 2-b follow easily. Finally, in Section 7, we state FastCubicMin
and explain why it satisfies these sufficient conditions and why it runs in the aforementioned time.
4
Theorem 2 implies Theorem 1
In this section, we show that Theorem 2 implies Theorem 1. It relies on the following lemma
(proved in Appendix B) regarding the sufficient condition for us to reach an ε-approximate local
minimum.
3
To present the simplest result, we have not tried to improve the constant dependency in this paper.
6
3/2
ε √
Lemma 4.1. If mt (h∗ ) ≥ − 800
and h0 is an approximate minimizer of mt (h) satisfying
L
√
√ε
4 L
k∇mt (h0 )k ≤ 2ε ,
√
then we have that k∇f (xt + h0 )k ≤ ε and λmin (∇2 f (xt + h0 )) ≥ − Lε.
kh0 k ≤ kh∗ k +
and
3/2
Proof of Theorem 1 from Theorem 2. When FastCubic terminates, we have mt (h0 ) > − cε√L ; there3/2
ε √
fore, it satisfies mt (h∗ ) ≥ − 800
according to Theorem 2-a. Combining this with Theorem 2-b and
L
Corollary 4.1, we conclude that
in the last iteration of FastCubic, our output satisfies k∇f (xt +h0 )k ≤
√
ε and λmin (∇2 f (xt + h0 )) ≥ − Lε. This finishes the proof with respect to the accuracy conditions.
As for the running time, in every iteration except for the last one, FastCubic satisfies mt (h0 ) ≤
3/2
3/2
√
√
. Therefore by (3.1), we must have decreased the objective by at least Ω −ε
in this
−Ω −ε
L
L
√
(f (x0 )−f ∗ ) L
round, and this cannot happen for more than O
iterations. The final running time
ε3/2
of FastCubic follows from this bound together with Theorem 2-c.
Therefore, in the rest of the paper it suffices to study FastCubicMin and prove Theorem 2.
5
Characterization Lemma of the Minimizer h∗
For notational simplicity in this and the subsequent sections we focus on the following problem:
minimize
h> Hh L
+ khk3
2
6
where H is a symmetric matrix with kHk2 ≤ L2 .
m(h) , g > h +
Recall from the previous section that we have denoted by h∗ an arbitrary minimizer of m(h).
We have the following lemma which characterizes h∗ : (a variant of this lemma has appeared in [8],
and we prove it in the appendix for the sake of completeness)
Lemma 5.1. We have h∗ is a minimizer of m(h) if and only if there exists λ∗ ≥ 0 such that
2λ∗
.
H + λ∗ I 0 ,
(H + λ∗ I)h∗ = −g ,
kh∗ k =
L
The objective value in this case is given by
1
2(λ∗ )3
m(h∗ ) = − g > (H + λ∗ I)+ g −
≤0 .
2
3L2
The following corollary comes from Lemma 5.1 and its proof:
Corollary 5.2. The value λ∗ in Lemma 5.1 is unique, and for every λ satisfying H + λI 0, we
have
2λ
2λ
k(H + λI)−1 gk >
⇐⇒ λ∗ > λ and k(H + λI)−1 gk <
⇐⇒ λ∗ < λ .
L
L
In the above characterization, we have a crude upper bound on λ∗ :
p
Proposition 5.3. We have λ∗ ≤ B , max 2L2 + Lkgk, 1 with λ∗ defined in Lemma 5.1.
Proof. We have Lk(H + BI)−1 gk ≤
Corollary 5.2.
Lkgk
λmin (H+BI)
7
≤
Lkgk
B−L2
< 2B and therefore λ∗ ≤ B due to
6
Sufficient Conditions for Theorem 2-a and 2-b
Without worrying about the design of FastCubicMin at this moment, let us first state a set of
sufficient conditions under which the assumptions in Theorem 2-a can be satisfied.
Main Lemma 1. Consider an algorithm that outputs a real λ ∈ [0, 2B], a vector v ∈ Rd , and a
unit vector vmin ∈ Rd . Additionally, suppose numbers κ, ε̃ ≥ 0 satisfying the following conditions:
1
1
(6.1)
ε̃ ≤
10000 (max {κ, L, L2 , kgk, k(H + λI)−1 k, B})20
(H + (λ − Lε̃)I)−1 0
Moreover, suppose that the outputs (λ, v, vmin ) satisfy one of the following two cases:
Case 1: Lk(H + λI)−1 gk ∈ [2λ − 2Lε̃, 2λ + 2Lε̃]
kv + (H + λI)−1 gk ≤ ε̃
and
Case 2: The following conditions are satisfied:
(a) λ ≥ λ∗ and λ + λmin (H) ≤ κ1
(b) Lk(H + λI)−1 gk ≤ 2λ and kv + (H + λI)−1 gk ≤ ε̃
1
> Hv
(c) vmin
min ≤ λmin (H) + 10κ
min
Then, at least one of the two choices h0 ∈ v, λv2L
satisfy
either
m(h∗ ) ≥ 3000m(h0 )
or
(6.2)
m(h∗ ) ≥ −
32
.
κ3 L2
Let us compare such sufficient conditions to the characterization Lemma 5.1.
• In Case 1, up to a very small error ε̃, we have essentially found a vector v that satisfies
∗
v ≈ −(H + λI)−1 g and kvk ≈ 2λ
L . Therefore, this v should be close to h for obvious reason.
(This is the simple case.)
• In Case 2, we have only found a vector v that satisfies v ≈ −(H + λI)−1 g and kvk . 2λ
L .
In this case, we also compute an approximate lowest eigenvector vmin of λmin (H) up to an
additive 1/10κ accuracy (see case 2-c). We will make sure that, as long as the conditions in
min
2-a hold, then either v or λv2L
will be an approximate minimizer for mt (h).
(This is the hard case.)
Proof of Main Lemma 1. We first consider Case 1. According to Corollary 5.2, if ε̃ = 0 then v is a
minimizer of m(h). The following claim extends this argument to the setting when ε̃ > 0:
Claim 6.1. If λ and v satisfy Case 1 and ε̃ satisfies (6.1), then m(v) ≤ m(h∗ ) +
1
250κ3 L2
From the above lemma it follows that either m(h∗ ) ≥ − κ38L2 otherwise m(h∗ ) ≥ 1.1m(v) which
satisfies the conditions of the theorem.
We now consider Case 2, and in this case we make the following two claims:
n
o
min
Claim 6.2. If λmin (H) ≤ − κ1 then m(h∗ ) ≥ 1500 min m(v), m λv2L
− 500κ13 L2 .
Claim 6.3. If λmin (H) ≥ − κ1 then m(h∗ ) ≥ 2m(v) −
16
κ3 L2
.
Lemma 1 now follows from the two claims because we can output the vector h0 which has the
min
lowest value of m(h0 ) amongst the two choices h0 ∈ v, λ v2d
. This satisfies either m(h∗ ) ≥
3000m(h0 ) or m(h∗ ) ≥ − κ332L2 .
The missing proofs of the three claims are deferred to Appendix D.
8
The next main lemma shows that, under the same sufficient conditions as Main Lemma 1, we
also have that Theorem 2-b holds. (Its proof is contained in Appendix E.)
3/2
ε √
Main Lemma 2. In the same setting as Main Lemma 1, suppose m(h∗ ) ≥ − 300
. Then the
L
output vector v satisfies the following conditions:
3
ε
15
kvk ≤ kh∗ k +
and k∇m(v)k ≤ + 2 .
κL
4 κ L
7
Main Algorithms for Theorem 2
We are now ready to state our main algorithm FastCubicMin and sketch why it satisfies the sufficient
conditions in Main Lemma 1. As described in Algorithm 2, our algorithm starts with a very large
choice λ0 ← 2B and decreases it gradually. At each iteration i, it computes an approximate inverse
v satisfying kv + (H + λi I)−1 gk ≤ ε̃ with respect to the current λi . Then there are three cases,
depending on whether Lkvk is approximately equal to, larger than, or smaller than 2λi . At a high
level, if it is “equal”, then we have met Case 1 in Main Lemma 1; if it is “larger”, then we can
binary search the correct value of λ∗ in the interval [λi , λi−1 ]; and if it is “smaller”, then we need
to compute an approximate eigenvector and carefully choose the next point λi+1 .
We state our main lemma below regarding the correctness and running time of FastCubicMin.
Main Lemma 3. FastCubicMin in Algorithm 2 outputs a real λ ∈ [0, 2B], a vector v ∈ Rd , and
a unit vector vmin ∈ Rd satisfying one of the two sufficient conditions in Main Lemma 1. We also
have that the procedure can be implemented in a total running time of
√
• Õ κL2 · Th if Accelerated Gradient Descent is used in Theorem 2.4 to invert matrices.
√
• Õ max{n, n3/4 κL2 }·Th,1 if we use accelerated SVRG as the subprocedure A in Theorem 2.4.
Here Õ hides logarithmic factors in L, L2 , κ, d, B.
We prove the correctness half of Main Lemma 3, and defer its running time analysis to Appendix G.
7.1
Correctness Half of Main Lemma 3
We will now establish the correctness of our algorithm. We first observe that the BinarySearch
subroutine returns (λ, v, ∅) that satisfies Case 1 of Main Lemma 1.
Fact 7.1. BinarySearch outputs a pair λ and v such that
Lk(H + λI)−1 gk ∈ [2λ − 2Lε̃, 2λ + 2Lε̃]
and
kv + (H + λI)−1 gk ≤ ε̃ .
Proof. The latter is guaranteed by line 3 in BinarySearch, and the former is implied by the latter
because
Lk(H + λI)−1 gk ∈ Lkvk − Lε̃/2, Lkvk + Lε̃/2 ⊆ 2λ − 2Lε̃, 2λ + 2Lε̃ .
We also establish the following invariants regarding the values λi . (Proof in Appendix F.)
Lemma 7.2. The following statements hold for all i until FastCubicMin terminates
(a) λi ∈ [0, 2B], λi + λmax (H) ≤ 3B
(b) λi + λmin (H) ≥
(c) λi+1 + λmin (H)
3
10κ
≤ 34 (λi
+ λmin (H)) unless λi+1 = 0
Moreover when FastCubicMin terminates at Line 20 we have λi + λmin (H) ≤ κ1 .
We now prove the output (λ, v, vmin ) of FastCubicMin satisfies the sufficient conditions of Main Lemma 1.
9
Algorithm 2 FastCubicMin(g, H, L, L2 , κ) (main algorithm for cubic minimization)
Input: g a vector, H a symmetric matrix, parameters κ, L and L2 which satisfies −L2 I H L2 I.
Output: (λ, v,p
vmin )
1
1: B ← L2 + Lkgk + κ
.
20
2: ε̃ ← 1/ 10000 max L, kgk, 3κ
,
B,
1
10
3: λ0 ← 2B.
4: for i = 0 to ∞ do
5:
Compute v such that kv + (H + λi I)−1 gk ≤ ε̃.
6:
if Lkvk ∈ [2λi − Lε̃, 2λi + Lε̃] then
7:
return (λi , v, ∅).
8:
else if Lkvk > 2λi + Lε̃ then
9:
return BinarySearch(λ1 = λi−1 , λ2 = λi , ε̃).
10:
else if Lkvk < 2λi − Lε̃ then
11:
Let Power Method find vector w that is 9/10-appx leading eigenvector of (H + λi I)−1 :
9
10 λmax ((H
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
+ λi I)−1 ) ≤ w> (H + λi I)−1 w ≤ λmax ((H + λi I)−1 ) .
1
Compute a vector w̃ such that kw̃ − (H + λi I)−1 wk ≤ ε̂ , 60B
.
1
∆ ← 21 w̃> w−ε̂
.
1
if ∆ > 2κ
then
λ̃i+1 ← λi − ∆
2.
if λ̃i+1 > 0 then λi+1 ← λ̃i+1 else λi+1 ← 0
else
> Hv
Use AppxPCA to find any unit vector vmin such that vmin
min ≤ λmin (H) +
>
Flip the sign of vmin so that g vmin ≤ 0.
return (λi , v, vmin ).
end if
end if
end for
Algorithm 3 BinarySearch(λ1 , λ2 , ε̃) (binary search subroutine)
Input: λ1 ≥ λ2 , Lk(H + λ1 I)−1 gk ≤ 2λ1 , Lk(H + λ2 I)−1 gk ≥ 2λ2 , λ2 + λmin (H) > 0
Output: (λ, v, ∅)
1: for t = 1 to ∞ do
2
2:
λmid ← λ1 +λ
2
3:
Compute vector v such that kv + (H + λmid I)−1 gk ≤ ε̃/2
4:
if Lkvk ∈ [2λmid − Lε̃, 2λmid + Lε̃] then
5:
return (λmid , v, ∅)
6:
else if Lkvk + Lε̃ ≤ 2λmid then
7:
λ1 ← λmid
8:
else if Lkvk − Lε̃ ≥ 2λmid then
9:
λ2 ← λmid
10:
end if
11: end for
10
1
10κ .
Correctness Proof of Main Lemma 3. We carefully verify these sufficient conditions:
• Lemma 7.2 implies λi ∈ [0, 2B].
3
from Lemma 7.2 implies k(H + λi I)−1 k ≤ 4κ. It is now immediate that
• λi + λmin (H) ≥ 10κ
the choice of ε̃ on Line 2 satisfies the Condition (6.1) in the assumption of Main Lemma 1.
1
• Since ε̃ ≤ 10κL
and λi + λmin (H) ≥
Condition (6.2) in Main Lemma 1.
3
10κ
it follows that (H + (λi − Lε̃)I)−1 0 which proves
• We now verify Case 1 and 2 in the assumption of Main Lemma 1. At the beginning of the
algorithm, our choice λ0 = 2B ensures (using Proposition 5.3) that Lk(H + λ0 I)−1 gk < 2λ0 .
Let us now consider the various places where the algorithm outputs:
– If FastCubicMin terminates at Line 7, then we have kv + (H + λi I)−1 gk ≤ ε̃ and additionally
Lk(H + λi I)−1 gk ∈ Lkvk − Lε̃, Lkvk + Lε̃ ⊆ [2λi − 2Lε̃, 2λi + 2Lε̃] .
Therefore, the output meets Case 1 requirement of Main Lemma 1 with λ = λi .
– If FastCubicMin terminates at Line 9, then Lk(H + λi I)−1 gk > Lkvk − Lε̃ ≥ 2λi . Obviously,
we must have i ≥ 1 in this case because Lk(H + λ0 I)−1 gk < 2λ0 . Therefore, Line 10 must
have been reached at the previous iteration, so it implies Lk(H + λi−1 I)−1 gk < 2λi−1 .
Together, these two imply that we can call BinarySearch with (λi−1 , λi ). Owing to Fact 7.1,
the subroutine outputs a pair (λ, v) satisfying the Case 1 requirement of Main Lemma 1.
– If FastCubicMin terminates on Line 20, we verify that Case 2 of Main Lemma 1 with λ = λi
holds. We first have
Lk(H + λi I)−1 gk ≤ Lkvk + Lε̃ ≤ 2λi .
By Corollary 5.2, we also have that λi ≥ λ∗ . Lemma 7.2 tells us λi satisfies λi +λmin (H) ≤ κ1 .
1
> Hv
Vector v satisfies kv + (H + λi I)−1 gk ≤ ε̃. Vector vmin satisfies vmin
min ≤ λmin (H) + 10κ .
In sum, we have verified that all the assumptions of Main Lemma 1 hold.
Final Proof of Theorem 2. Theorem 2 is a direct corollary of our main lemmas. Main Lemma 3
ensures that the assumptions of Main Lemma 1 and Main Lemma 2 both hold. Now, using the special choice of κ in FastCubic, Theorem 2-a immediately comes from Main Lemma 1; Theorem 2-b
immediately comes from Main Lemma 2; and Theorem 2-c immediately comes from Main Lemma 3.
This finishes the proof of Theorem 2.
Acknowledgements
We thank Ben Recht for very helpful suggestions and corrections to a previous version. Z. Allen-Zhu
is supported by an NSF Grant, no. CCF-1412958, and a Microsoft Research Grant, no. 0518584.
Any opinions, findings and conclusions or recommendations expressed in this material are those of
the author(s) and do not necessarily reflect the views of NSF or Microsoft.
Appendix
A
Computing Hessian-Vector Product in Linear Time
In this section we sketch the intuition regarding why Hessian-vector products can be computed in
linear time in many interesting (especially machine learning) problems. We start by showing that
the gradient can be computed in linear time. The algorithm is often referred to as back-propagation,
11
which dates back to Werbos’s PhD thesis [35], and has been popularized by Rumelharte et al. [31]
for training neural networks.
Claim A.1 (back-propagation, informally stated). Suppose a real-valued function f : Rd → R can
be evaluated by a differentiable circuit of size N . Then, the gradient ∇f can be computed in time
O(N + d) (using a circuit of size O(N + d)). 4
The claim follows from simple induction and chain-rule, and is left to the readers. In the training
of neural networks, often the size of circuits that computes the objective f is proportional to (or
equal to) the number of parameters d. Thus the gradient ∇f can be computed in time O(d) using
a circuit of size d.
Next, we consider computing ∇2 f (x) · v where v ∈ Rd . Let g(x) := h∇f (x), vi be a function
from Rd to R. Then, we see that if suffices to compute the gradient of g, since
∇2 f (x) · v = ∇g(x) .
We observe that g(x) can be evaluated in linear time using circuit of size O(d) since we’ve shown
∇f (x) can. Thus, using Claim A.1 again on function g, 5 we conclude that ∇g(x) can also be
computed in linear time.
B
Proof of Lemma B.1 and Corollary 4.1
Lemma B.1. For all h0 ∈ Rd , it satisfies
k∇f (xt +h0 )k ≤ Lkh0 k2 +k∇mt (h0 )k
and
λmin (∇2 f (xt +h0 )) ≥ −
3L2 max{0,−mt (h∗ )}
2
1/3
−Lkh0 k .
Proof of Lemma B.1. Let us denote by g = ∇f (xt ) and H = ∇2 f (xt ) in this proof. We begin by
proving the first order condition. Note that we have
∇mt (h) = g + Hh + L2 khkh .
Recall h∗ is a minimizer of argmin mt (h). The characterization result in Lemma 5.1 shows H +
Lkh∗ k
2 I 0, and thus
L
g > h∗ + (h∗ )> Hh∗ + kh∗ k3 = ∇mt (h∗ )> h∗ = 0
(B.1)
2
Lkh∗ k
L
3
∗ >
∗ >
∗
(h ) Hh + khk = (h )
H+
I h∗ ≥ 0 .
(B.2)
2
2
They imply
(h∗ )> Hh∗ Lkh∗ k3 ¬ (h∗ )> Hh∗ L ∗ 3
+
=−
− kh k
2
6
2
3
L
L
L
≤ kh∗ k3 − kh∗ k3 = − kh∗ k3
4
3
12
where ¬ uses (B.1) and uses (B.2).
mt (h∗ ) = g > h∗ +
We compute the norm of the gradient at a point xt + h0 for any h0 ∈ Rd :
k∇f (xt + h0 )k ≤ k∇f (xt + h0 ) − ∇mt (h0 )k + k∇mt (h0 )k
Z 1
L
= ∇f (xt ) +
∇2 f (xt + τ h0 )h0 dτ − g + Hh0 + kh0 kh0 + k∇mt (h0 )k
2
0
4
5
Technically, we assume that the gradient of each gate can be computed in O(1) time
We assume here that the original circuits are twice differentiable
12
(B.3)
Z
1
L
∇2 f (xt + τ h0 ) − H h0 dτ + kh0 k2 + k∇mt (h0 )k
2
0
Z 1
®
L
τ dτ + kh0 k2 + k∇mt (h0 )k = Lkh0 k2 + k∇mt (h0 )k
≤ Lkh0 k2
(B.4)
2
0
where ® follows from the Lipschitz continuity on the Hessian (2.1). This proves the first conclusion
of the lemma.
≤
As for the second-order condition, we first note that for all h0 ∈ Rd , by the Lipschitz continuity
on the Hessian (2.1), we have k∇2 f (xt + h0 ) − ∇2 f (xt )k ≤ Lkh0 k. However, this implies
λmin (∇2 f (xt + h0 )) ≥ λmin (∇2 f (xt )) − Lkh0 k .
(B.5)
λmin (∇2 f (xt + h0 )) ≥ −Lkh0 k .
(B.6)
because if two matrices A and B satisfies kA−Bk ≤ p, then it must satisfy λmin (A)−λmin (B) ≤ p
as well. We consider two cases: if λmin (∇2 f (xt )) ≥ 0, then we have
Otherwise, we consider the case where λmin (∇2 f (xt )) = λmin (H) < 0. Let νd be the normalized
eigenvector corresponding to λmin (H), and define
2λmin (H)
h̃ , sign(g > νd ) ·
νd .
L
We calculate mt (h̃) as follows:
h̃> Hh̃ L
4|λmin (H)|3
h̃> Hh̃ L
2(λmin (H))2 >
Hν
+
+ kh̃k3 ≤
+ kh̃k3 =
ν
d
d
2
6
2
6
L2
3L2
3
3
3
4|λmin (H)| 2(λmin (H))
¬ 2(λmin (H))
+
=
,
(B.7)
=
L2
3L2
3L2
where ¬ uses νd> Hνd = λmin (H) < 0, and uses the assumption that λmin (H) < 0. Since by
definition mt (h∗ ) ≤ mt (h̃), we can deduce from inequality (B.7) that
2
1/3
3L |mt (h∗ )|
2
λmin (∇ f (xt )) = λmin (H) ≥ −
.
(B.8)
2
Now we put together inequalities (B.5) and (B.8), and obtain
2
1/3
3L |mt (h∗ )|
2
0
λmin (∇ f (xt + h )) ≥ −
− Lkh0 k .
(B.9)
2
Combining (B.6) and (B.9) we finish the proof of Lemma B.1.
mt (h̃) = g > h̃ +
3/2
ε √
Corollary 4.1. If mt (h∗ ) ≥ − 800
and h0 is an approximate minimizer of mt (h) satisfying
L
√
√ε
4 L
k∇mt (h0 )k ≤ 2ε ,
√
then we have that k∇f (xt + h0 )k ≤ ε and λmin (∇2 f (xt + h0 )) ≥ − Lε.
kh0 k ≤ kh∗ k +
and
3/2
ε √
, along with inequality
Proof of Corollary 4.1. First of all, our assumption that mt (h∗ ) ≥ − 800
L
√
√
(B.3), tells us that kh∗ k ≤ 4√εL . This, together with our assumption on kh0 k, implies kh0 k ≤ 2√εL .
Since we also assume k∇mt (h0 )k ≤ 2ε , we have from Lemma B.1 that
ε ε
k∇f (xt + h0 )k ≤ Lkh0 k2 + k∇mt (h0 )k ≤ + ≤ ε .
4 2
For the second-order condition, we can again apply Lemma B.1 to get
!1/3 √
2
1/3
√
3L max{0, −mt (h∗ )}
Lε
3L3/2 ε3/2
2
0
0
λmin (∇ f (xt + h )) ≥ −
− Lkh k ≥ −
−
≥ − Lε .
2
1600
2
13
C
Proof of Lemma 5.1 and Corollary 5.2
We begin by proving a few lemmas that characterize the system of equations.
Lemma C.1. Consider the following system of equations/inequalities in variables λ, h:
2λ
H + λI 0 ,
(H + λI)h = −g ,
khk =
.
L
The following statements hold for any solution (λ0 , h0 ) of the above system:
(C.1)
• There is a unique value λ0 that satisfies the above equations. λ0 is such that λ0 ≥ −λmin (H).
• If λ0 > −λmin (H), then the corresponding h0 is also unique and is given by h0 = −(H + λI)−1 g.
• If λ0 = −λmin (H) then g > v = 0 for any vector v belonging to the eigenspace corresponding to
λmin (H). Subsequently we also have that the corresponding h0 is of the form
h0 = −(H + λI)+ g + γv
for some γ and v in the lowest eigenspace of H.
Proof of Lemma C.1. Note that H + λI 0 ensures that for any solution λ0 , we have λ0 ≥
−λmin (H). Furthermore, for any λ0 > −λmin (H), the corresponding h is uniquely defined by
h = (H + λI)−1 g since H + λ0 I is invertible. If indeed λ0 = −λmin (H), then we have that the
equation (H − λmin (H)I)h = −g has a solution. This implies that g has no component in the null
space of H − λmin (H)I, or equivalently that it has no component in the eigenspace corresponding
to λmin (H). We also have that every solution of (H − λmin (H)I)h = −g is necessarily of the form
h = −(H − λmin I)+ g + γv
for some γ and v in the lowest eigenspace of H.
We will now prove the uniqueness of λ0 by contradiction. Consider two distinct values of λ1 , λ2
that satisfy the system (C.1). If both λ1 , λ2 > −λmin (H) we get that
2λ2
2λ1
and k(H + λ2 I)−1 gk =
.
k(H + λ1 I)−1 gk =
L
L
Now note that k(H + λI)−1 gk is a strictly decreasing function over the domain λ ∈ (−λmin (H), ∞)
and 2λ
L is strictly increasing over the same domain. Therefore the above two equations cannot be
satisfied for two distinct λ1 , λ2 > −λmin (H) which is a contradiction. Suppose now without loss of
generality that λ1 = −λmin (H). Then we have that the corresponding solution is of the form
h = −(H + λI)+ g + γv
for some γ and v in the lowest eigenspace of H and g has no component in the lowest eigenspace
of H. It follows that k(H − λmin (H)I)+ gk ≥ k(H + λI)−1 gk for any λ > −λmin (H). By a similar
argument as in the first case, we can now see that the following conditions,
2λ1
2λ2
k(H + λ1 I)+ g + γvmin (H)k =
and k(H + λ2 I)−1 gk =
,
L
L
cannot both be satisfied for λ2 > λ1 = −λmin (H), giving us a contradiction. This finishes the proof
of Lemma C.1.
Lemma C.2. Let (λ, h) be a solution of the system (C.1). Then we have that
1
2λ3
m(h) = − g > (H + λI)+ g −
.
2
3L2
14
Proof of Lemma C.2. By the definition of the system (C.1), any solution λ, h to the system should
be such that there exists some γ such that
h = (H + λI)+ g + γv0
where v0 is in the null space of H + λI if it exists; otherwise γ = 0. This gives us the following:
h> Hh L
m(h) = g > h +
+ khk3
2
6
1 >
λ
L
¬
= − h (H + λI)h − khk2 + khk3
2
2
6
3
2λ
1
.
= − g > (H + λI)+ g −
2
3L2
Equality ¬ follows because (H + λI)h = −g. Equality follows because h = (H + λI)+ g + γv0
and khk = 2λ
L.
Lemma 5.1. h∗ is a minimizer of m(h) if and only if there exists λ∗ ≥ 0 such that
2λ∗
.
H + λ∗ I 0 ,
(H + λ∗ I)h∗ = −g ,
kh∗ k =
L
The objective value in this case is given by
1
2(λ∗ )3
m(h∗ ) = − g > (H + λ∗ I)+ g −
≤0 .
2
3L2
Proof of Lemma 5.1. We first compute that
∇m(h) = g + Hh +
L
L
L
khkh and ∇2 m(h) = H + khkI + khk
2
2
2
h
khk
h
khk
>
.
For the forward direction, suppose h∗ is a minimizer of m(h). Let λ∗ = L2 kh∗ k. Then, the necessary
conditions ∇m(h∗ ) = 0 and ∇2 m(h∗ ) 0 can be written as
∗ ∗ > !
h
h
∗
∗
>
∗
∗
g + (H + λ I)h = 0 and w
H+λ I+λ
w ≥ 0, ∀w ∈ Rn .
(C.2)
kh∗ k
kh∗ k
From this we see (H + λ∗ I)h∗ = −g and kh∗ k =
2λ∗
L ,
and the only thing left to verify is H + λ∗ I 0.
Note that if h∗ = 0, then the second inquality in (C.2) directly implies H + λ∗ I 0. Thus, we
only need to focus on h∗ 6= 0. We want to show that w> (H + λ∗ I)w ≥ 0 for every w ∈ Rd . Now,
if w> h∗ = 0 then this trivially follows from (C.2), so it suffices to focus on those w that satisfies
w> h∗ 6= 0.
Since w and h∗ are not orthogonal, there exists γ ∈ R\{0} such that kh∗ + γwk = kh∗ k. (This
can be done by squaring both sides and solving the linear system in λ.) Squaring both sides we
have
γ 2 kwk2
=0 .
(C.3)
(γw)> h∗ +
2
Now we bound the difference
(h∗ + γw)> H(h∗ + γw) h∗ Hh∗
m(h∗ + γw) − m(h∗ ) = g > ((h∗ + γw) − h∗ ) +
−
2
2
∗ + γw)> H(h∗ + γw)
h∗ Hh∗
(h
¬
−
= (h∗ − (h∗ + γw))> (H + λ∗ I)h∗ +
2
2
∗ 2
∗ + γw)> H(h∗ + γw)
(h
h∗ Hh∗
λ γ
=
kwk2 + (h∗ − (h∗ + γw))> Hh∗ +
−
2
2
2
∗ Hh∗
∗ + γw)> H(h∗ + γw)
λ∗ γ 2
h
(h
=
kwk2 +
− (h∗ + γw)> Hh∗ +
2
2
2
15
λ∗ γ 2
γ2
γ2 >
kwk2 + w> Hw =
w (H + λ∗ I)w ,
(C.4)
2
2
2
where ¬ and follow from (C.2) and (C.3), respectively. Since h∗ is a minimizer of m(h), we
immediately have
γ2 >
m(h∗ + γw) − m(h∗ ) =
w (H + λ∗ I)w ≥ 0,
2
and we conclude that (H + λ∗ I) 0.
=
For the backward direction, we will make use Lemma C.1 and Lemma C.2. First we note that
the function m(h) is continuous and bounded from below, and there exists at least one minimizer
h∗ . Suppose now there exists a λ∗ and a corresponding h∗ such that (λ∗ , h∗ ) is a solution to the
system C.1. The backward direction requires us to prove that h∗ must be a minimizer of m(h). By
Lemma C.1 we get the following two cases.
We prove the backward direction by showing that the conditions in Equation C.2 determine the
minimizer up to its norm. To this end we will use Lemma C.1 and Lemma C.2.
First we note that the function m(h) is continuous, bounded from below, and tends to +∞
when khk → ∞, so there exists at least one minimizer h∗ .
Suppose now there exists a λ∗ and a corresponding h∗ such that (λ∗ , h∗ ) is a solution to the
system (C.1). The backward direction requires us to prove that h∗ must be a minimizer of m(h).
By Lemma C.1 we get the following two cases.
• If λ∗ > −λmin (H) then (λ∗ , h∗ ) is the only solution to the system (C.1). By the proof of the
forward direction we see that any minimizer of m(h) must satisfy system (C.1) and therefore
h∗ must be the minimizer.
• If above is not the case, then λ∗ = −λmin (H). Let h0 be any minimizer of m(h). Lemma C.1
and the proof of the forward direction ensures that (λ∗ , h0 ) also satisfies the system (C.1). By
Lemma C.2 we get m(h∗ ) = m(h0 ) and therefore h∗ is a minimizer too.
Corollary 5.2. This value λ∗ is unique, and for every λ satisfying H + λI 0, we have
2λ
2λ
k(H + λI)−1 gk >
⇐⇒ λ∗ > λ and k(H + λI)−1 gk <
⇐⇒ λ∗ < λ .
L
L
Proof of Corollary 5.2. The uniqueness of λ∗ follows from Lemma C.1. To prove the second part
we first make some observations about the function
2y
p(y) ,
− k(H + yI)−1 gk
L
defined on the domain y ∈ (−λmin (H), ∞). Note that p(y) is continuous and strictly increasing
over the domain and p(y) → ∞ as y → ∞.
The corollary requires us to show that
p(λ) < 0 ⇐⇒ λ∗ > λ
and
p(λ) > 0 ⇐⇒ λ∗ < λ .
We begin by showing the first equivalence. To see the backward direction note that if λ∗ > λ >
∗
−λmin (H), by the characterization of λ∗ in Lemma C.1 we have that k(H + λ∗ I)−1 gk = 2λL i.e.
p(λ∗ ) = 0 which implies that p(λ) < 0 as p(y) is a strictly increasing function. For the forward
direction note that since p(y) is continuous and strictly increasing we see that the range of the
function contains [p(λ), ∞). Since p(λ) < 0 there must exist a λ∗ > λ such that p(λ∗ ) = 0 which
by the characterization in Lemma C.1 finishes the proof.
Now we will prove that p(λ) > 0 ⇐⇒ λ∗ < λ. To see the forward direction note that if λ∗ ≥ λ
then p(λ∗ ) = 0 and p(λ) > 0 which contradicts the fact that p(y) is strictly increasing. For the
16
backward direction we consider two cases. Firstly if λ∗ > −λmin (H) the conclusion follows similarly
by the monotonicity of p(y). If λ∗ = −λmin then by Lemma C.1, we have that g has no component
in the lowest eigenspace of H and therefore if we extend p(y) to −λmin (H) by defining
−2λmin (H)
p(−λmin (H)) ,
− k(H − λmin (H)I)+ gk
L
we get that p(y) is increasing in the domain y ∈ [−λmin (H), ∞). Now from the characterization
of the solution in Lemma C.1 we can see that p(−λmin (H)) ≥ 0 and therefore by monotonicity
p(λ) > 0. This finishes the proof.
D
Proof of Main Lemma 1
D.1
Proof of Claim 6.1
Claim 6.1. If λ and v satisfy Case 1 and ε̃ satisfies (6.1), then m(v) ≤ m(h∗ ) +
1
250κ3 L2
Proof of Claim 6.1. Note that by the conditions of the theorem we have that (H+(λ−Lε̃)I)−1 eq
and
Lk(H + (λ − Lε̃)I)−1 gk ≥ 2λ − 2Lε̃ and Lk(H + (λ + Lε̃)I)−1 gk ≤ 2λ − 2Lε̃ ,
according to Corollary 5.2 we must have
λ∗ ∈ [λ − Lε̃, λ + Lε̃]
(D.1)
This also implies (using our assumption on ε̃)
Next, consider the value m(v)
Lkvk ≤ [2λ∗ − 5Lε̃, 2λ∗ + 5Lε̃] .
v > Hv L
v > (H + λI)v
λ Lkvk
+ kvk3 = g > v +
− kvk2
−
.
2
6
2
2
6
We bound the two parts on the right hand side of (D.2) separately. The first part
m(v) = g > v +
(D.2)
v > (H + λI)v
g > (H + λI)−1 g
k(H + λI)−1 kε̃2
≤−
+ kgkε̃ + k(H + λI)−1 gkε̃ +
2
2
2
>
−1
¬
g (H + λI) g
1
≤−
+
(D.3)
2
1000κ3 L2
g > (H + λ∗ I)−1 g
1
+ Lkgk2 k(H + λI)−1 kk(H + (λ + 2Lε̃)I)−1 kε̃ +
≤−
2
1000κ3 L2
>
∗
−1
®
g (H + λ I) g
1
≤−
+
2
500κ3 L2
Above, inequalities ¬ and ® use the assumption on ε̃ in (6.1), and inequality uses
g>v +
−(H + λI)−1 −(H + (λ∗ + Lε̃)I)−1
= −(H + λ∗ I)−1 − Lε̃(H + λ∗ I)−1 (H + (λ∗ + Lε̃)I)−1
Note that (H + λ∗ I)−1 0 by Equations (D.1) and (6.2). The second part of (D.2) can be bounded
as follows
Lkvk
(2λ∗ − 5Lε̃)2 λ∗ − Lε̃ 2λ∗ + 5Lε̃
2 λ
kvk
−
≥
−
2
6
L2
2
6
∗
3
∗
3
¬
2(λ )
2(λ )
1
≥
− 1000ε̃L(λ∗ )2 ≥
−
3L2
3L2
500κ3 L2
17
Above, inequality ¬ uses λ∗ ≤ B (owing to Proposition 5.3) and our assumption on ε̃ from (6.1).
Putting these together we get that
1
m(v) ≤ m(h∗ ) +
.
250κ3 L2
D.2
Proofs for Claims 6.2 and 6.3
For notational simplicity, let us rotate the space into the basis in the eigenspace of H; let the i-th
dimension correspond to the i-th largest eigenvalue λi of H. We have λ1 ≥ λ2 . . . ≥ λd = λmin . Let
gi denote the i-th coordinate of g in this basis.
Lemma 5.1 implies
m(h∗ ) = −
where we denote by
S1 = −
1 X gi2
2(λ∗ )3
2(λ∗ )3
−
=:
S
+
S
−
.
1
2
2
λi + λ∗
3L2
3L2
(D.4)
i
X
1
i:λi +λ∗ ≥ κ
gi2
λi + λ∗
From Corollary 5.2 we can also obtain
X
i:λi +λ∗ >0
Now the assumption k(H + λI)−1 gk ≤
S2 = −
X
1
i:0<λi +λ∗ ≤ κ
gi2
λi + λ∗
gi2
4(λ∗ )2
≤
.
(λi + λ∗ )2
L2
(D.5)
2λ
L
is equivalent to
X
gi2
4λ2
≤
(λi + λ)2
L2
(D.6)
i
We begin with a few auxiliary claims.
Claim D.1. If λmin (H) ≤ − κ1 then S2 ≥ 1000 · m
Proof of Claim D.1. We compute that
X
gi2
S2 = −
=−
λi + λ∗
∗ 1
i:0<λi +λ ≤ κ
X
1
i:0<λi +λ∗ ≤ κ
λvmin
2L
gi2 (λi + λ∗ )
1
≥−
∗
2
(λi + λ )
κ
X
1
i:0<λi +λ∗ ≤ κ
gi2
(λi + λ∗ )2
4
|λmin |3
∗ 2
(λ
)
≥
−16
.
(D.7)
κL2
L2
Above, ¬ uses (D.5), and follows because we have λmin (H) ≤ − κ1 in the assumption and have
λ∗ ≤ −λmin (H) + κ1 in the assumption of Case 2 of Main Lemma 1.
¬
≥−
min
Let us now consider the value of the vector λv2L
. We have that
> Hv
λvmin
λg > vmin λ2 vmin
λ3 ¬ λg > vmin λ2 λmin
λ3
min
m
=
+
+
≤
+
+
2L
2L
8L2
48L2
2L
16L2
48L2
λg > vmin
λ2 λmin λ2 λmin
λg > vmin λ2 λmin
≤
+
−
≤
+
2L
16L2
24L2
2L
48L2
1
Above, ¬ is because our assumption λmin (H) ≤ − κ and assumption vmin Hvmin ≤ λmin (H) +
1
1
together imply vmin Hvmin ≤ λmin
2 . follows from λmin (H) ≤ − κ and λ ≤ −λmin (H) + κ .
1
10κ
Now, recall that the sign of vmin is chosen so g > vmin is non-positive, and therefore by our
18
assumptions λmin (H) ≤ − κ1 and λ ≤ −λmin (H) + κ1 , we get the following inequality:
λvmin
|λmin |3
m
≤−
2L
48L2
Putting inequalities (D.8) and (D.7) together finishes the proof of Claim D.1.
(D.8)
We also show the following lemma, the proof of which can be seen from inequality (D.3), as
part of the proof of Claim 6.1 above.
Lemma D.2. If we have λ, v such that
Lk(H + λI)−1 gk ≤ 2λ
and
with ε̃ satisfying condition (6.1) then we have that
g>v +
Claim D.3. S1 ≥ 4m(v) −
kv + (H + λI)−1 gk ≤ ε̃
v > (H + λI)v
g > (H + λI)−1 g
1
≤−
+
2
2
1000κ3 L2
1
250κ3 L2
Proof of Claim D.3. We have that
v > (H + λI)v λ
L
m(v) = g > v +
− kvk2 + kvk3
2
2
6
>
−1
g (H + λI) g
L
1
¬
2 λ
=−
− kvk
− kvk +
2
2
6
1000κ3 L2
g > (H + λI)−1 g
2λ − 3Lε̃ 2 λ Lε̃
1
≤−
−
+
+
2
L
6
3
1000κ3 L2
®
2λ3
1
g > (H + λI)−1 g
−
+
≤−
2
3L2 500κ3 L2
g > (H + λI)−1 g
1
≤−
+
(D.9)
2
500κ3 L2
Above, ¬ is due to Lemma D.2; uses our condition on v which gives Lkvk ∈ [2λ − 3Lε̃, 2λ + 3Lε̃];
® uses our condition (6.1) on ε̃.
We now bound S1 . For this purpose first we note that if λi + λ∗ ≥
2(λi + λ∗ ) ≥ 1/κ + λi + λ∗ ≥ λi + λ .
Therefore, the sum S1 satisfies
X
gi2
S1 = −
≥ −2
λi + λ∗
∗ 1
i:λi +λ ≥ κ
X
1
i:0<λi +λ∗ ≤ κ
1
κ
and λ − λ∗ ≤
1
κ
then
gi2
1
≥ −2(g > (H + λI)−1 g) ≥ 4m(v) −
(λi + λ)
250κ3 L2
(Note that we have H + λI 0.) This finishes the proof of Claim D.3.
n
o
min
Claim 6.2. If λmin (H) ≤ − κ1 then m(h∗ ) ≥ 1500 min m(v), m λv2L
−
1
500κ3 L2
Proof of Claim 6.2. We derive that
2(λ∗ )3 1
16|λmin |3
¬ 1
m(h∗ ) = (S1 + S2 ) −
≥
(S
+
S
)
−
1
2
2
2
3L2
2
3L
®
1
λvmin
16|λmin |3
≥ 2m(v) −
+
500
·
m
−
500κ3 L2
2L
3L2
¯
1
λvmin
≥ 2m(v) −
+ 1500 · m
3
2
500κ L
2L
19
1
λvmin
−
≥ 1500 min m(v), m
2L
500κ3 L2
Above, ¬ uses equation (D.4), inequality follows because we have λmin (H) ≤ − κ1 in the assumption and have λ∗ ≤ −λmin (H) + κ1 in the assumption of Case 2 of Main Lemma 1; inequality ® uses
Claim D.3 and Claim D.1; and inequality ¯ uses (D.8). This finishes the proof of Claim 6.2.
Claim 6.3. If λmin (H) ≥ − κ1 then m(h∗ ) ≥ 2m(v) −
16
κ3 L2
Proof of Claim 6.3. This time we lower bound S2 slightly differently:
¬
4
16
S2 ≥ − 2 (λ∗ )2 ≥ − 3 2
(D.10)
κL
κ L
where ¬ comes from the second to last inequality from (D.7) and comes from λ∗ ≤ λ ≤
−λmin (H) + κ1 ≤ κ2 using our assumption in Case 2 of Main Lemma 1.
Putting these together we get that
2(λ∗ )3
1
15
16
¬ 1
≥ 2m(v) −
− 3 2 ≥ 2m(v) − 3 2 .
m(h∗ ) = (S1 + S2 ) −
2
3
2
2
3L
500κ L
κ L
κ L
Above, ¬ comes from (D.4), uses Claim D.3, lower bound (D.10) and
E
2(λ∗ )3
3L2
≤
16
3κ3 L2
Proof of Main Lemma 2
3/2
ε √
. Then the
Main Lemma 2. In the same setting as Main Lemma 1, suppose m(h∗ ) ≥ − 300
L
output vector v satisfies the following conditions:
3
ε
15
kvk ≤ kh∗ k +
and k∇m(v)k ≤ + 2 .
κL
4 κ L
Proof of Main Lemma 2. Let’s first note that from the value given in Lemma 5.1,
(λ∗ )3 ≤
L3/2 ε3/2
3L2 |m(h∗ )|
≤
.
2
200
(E.1)
If Case 1 occurs, we have
® 2λ∗
¯
2λ + 2Lε̃
1
+ ε̃ ≤
+ 5ε̃ ≤ kh∗ k +
.
L
L
20κL
Above, inequalities ¬ and both use the assumptions of Case 1; inequality ® uses the fact that
λ∗ ∈ [λ − Lε̃, λ + Lε̃] which again follows from the assumptions of Case 1 (see (D.1)); inequality ¯
∗
uses kh∗ k = 2λL from Lemma 5.1 as well as our assumption (6.1) on ε̃.
¬
kvk ≤ k(H + λI)−1 gk + ε̃ ≤
As for the quantity k∇m(v)k, we bound it as follows
¬
Lkvk
k∇m(v)k = g + Hv +
v ≤ kg + (H + λI)vk + λkvk + Lkvk2
2
®
λ(2λ + 3Lε̃) + (2λ + 3Lε̃)2
≤ kH + λIkε̃ + λkvk + Lkvk2 ≤ (L2 + 2B)ε̃ +
L
2
∗
2
¯
6λ
6(λ + Lε̃)
= (L2 + 2B)ε̃ +
+ 15ε̃λ + 9Lε̃2 ≤
+ (L2 + 32B)ε̃ + 9Lε̃2
L
L
° 6(λ∗ )2
± ε
15
≤
+ (L2 + 56B)ε̃ + 15Lε̃2 ≤ + 2 .
L
4 κ L
Above, inequality ¬ uses triangle inequality; inequality uses kv + (H + λI)−1 gk ≤ ε̃; inequality ®
uses kH + λIk ≤ L2 + 2B and Lkvk ≤ 2λ + 3Lε̃ which comes from our upper bound on kvk above;
¯ uses the fact that λ∗ ∈ [λ − Lε̃, λ + Lε̃] which again follows from the assumptions of Case 1 (see
20
(D.1)); inequality ° uses λ∗ ≤ 2B; and inequality ± uses (E.1) together with our assumption (6.1)
on ε̃.
If Case 2 occurs, we have
® 2(λ∗ + 1/κ)
2λ
3
+ ε̃ ≤
+ ε̃ ≤ kh∗ k +
.
(E.2)
L
L
κL
Above, inequalities ¬ and both use the assumptions of Case 2; inequality ® uses λ ≤ −λmin (H)+
1/κ from our assumption of Case 2 as well as −λmin (H) ≤ λ∗ which comes from Lemma 5.1;
∗
inequality ¯ uses kh∗ k = 2λL from Lemma 5.1 as well as our assumption (6.1) on ε̃.
¬
kvk ≤ k(H + λI)−1 gk + ε̃ ≤
The quantity k∇m(v)k can be bounded in an analogous manner as Case 1:
λ(2λ + Lε̃) + (2λ + Lε̃)2
L
∗ + 1 )2
∗
2
¬ 6λ2
6(λ
1
1
12(λ )
15 ® ε
15
κ
+
+
≤
+ 2 ≤ + 2 .
≤
≤
2
2
L
10κ L
L
10κ L
L
κ L
4 κ L
Above, inequality ¬ uses our assumption (6.1) on ε̃; inequality uses λ ≤ λ∗ + κ1 which appeared
in (E.2); inequality ® uses (E.1).
k∇m(v)k ≤ kH + λIkε̃ + λkvk + Lkvk2 ≤ (L2 + 2B)ε̃ +
F
Proof of Lemma 7.2
Lemma 7.2. The following statements hold for all i until FastCubicMin terminates
(a) λi ∈ [0, 2B], λi + λmax (H) ≤ 3B
(b) λi + λmin (H) ≥
(c) λi+1 + λmin (H)
3
10κ
≤ 34 (λi
+ λmin (H)) unless λi+1 = 0
Moreover when FastCubicMin terminates at Line 20 we have λi + λmin (H) ≤ κ1 .
Proof of Lemma 7.2. The lemma follows via induction.
To see (a) and (b) at the base case i = 0, recall that the definitions of B and L2 together ensure
3
λ0 + λmax (H) ≤ 3B and λ0 + λmin (H) ≥ 10κ
. Also λ0 ∈ [0, 2B].
Suppose now for some i ≥ 0 properties (a) and (b) hold. It is easy to check that λi ≤ λi−1
and thus we have λi + λmax (H) ≤ 2B and λi ≤ 2B. This implies property (a) at iteration i + 1
also hold. We now proceed to show property (c) at iteration i and property (b) at iteration i + 1.
Recall that the algorithm ensures
9
λmax ((H + λi I)−1 ) ≤ w> (H + λi I)−1 w ≤ λmax ((H + λi I)−1 ) ,
10
and by the definition of w̃ we have
9
λmax ((H + λi I)−1 ) − 2ε̂ ≤ w̃> w − ε̂ ≤ λmax ((H + λi I)−1 ) .
(F.1)
10
3
≤ λi + λmin (H) ≤ 3B from the inductive assumption, it follows from the choice of
Now, since 10κ
ε̂ that
1
1
λmax ((H + λi I)−1 )
2ε̂ ≤
≤
=
.
(F.2)
30B
10(λi + λmin (H))
10
Plugging Equation (F.2) into Equation (F.1) we get
8
1
8
1
= λmax ((H + λi I)−1 ) ≤ w̃> w − ε̂ ≤ λmax ((H + λi I)−1 ) =
.
10 λi + λmin (H)
10
λi + λmin (H)
21
Inverting this chain of inequalities, we have
λi + λmin (H)
5(λi + λmin (H))
≤∆≤
.
(F.3)
2
8
From this we derive the following implications:
1
1
∆≤
=⇒ (λi + λmin (H)) ≤
(F.4)
2κ
κ
1
4
∆>
=⇒ (λi + λmin (H)) >
(F.5)
2κ
5κ
If Condition (F.4) happens, our algorithm FastCubicMin outputs on Line 20; in such a case (F.4)
implies our desired inequality λi + λmin (H) ≤ κ1 . If Condition (F.5) happens, our choice λ̃i+1 ←
λi − ∆
2 and Equation (F.3) together imply that
3
11
(λi + λmin (H)) ≥ λ̃i+1 + λmin (H) ≥ (λi + λmin (H))
4
16
Combining this with (F.5) we get that
3
11 4
3
(λi + λmin (H)) ≥ λ̃i+1 + λmin (H) ≥
≥
.
4
16 5κ
10κ
Therefore, we conclude that property (c) at iteration i holds and property (b) at iteration i + 1
hold because λi+1 ≥ λ̃i+1 . This finishes the proof of Lemma 7.2.
G
Proof of Main Lemma 3: Running Time Half
Having proven the correctness of the algorithm, we now aim to bound the overall running time
of FastCubicMin, completing the proof of Main Lemma 3. We prove in Appendix H the following
lemma:
2 )B
Lemma G.1. If λ2 +λmin (H) ≥ c1 ∈ (0, 1) then BinarySearch ends in O log( (λc11−λ
·L·ε̃ ) iterations.
3
Since in our FastCubicMin algorithm, it satisfies λi ≤ 2B and λi +λmin (H) ≥ 10κ
(see Lemma 7.2),
taken together with our choice of ε̃ we have:
Claim G.2. Each invocation of BinarySearch ends in O log(1/ε̃) iterations.
Claim G.3. FastCubicMin ends in at most O(log(Bκ)) outer loops.
Proof. According to Lemma 7.2 we have 43 (λi−1 + λmin (H)) ≥ λi + λmin (H) so the quantity λi +
λmin (H) decreases by a constant factor per iteration (except possibly λi = 0 the last outer loop
in which case we shall terminate in one more iteration). On one hand, we have began with λ0 +
3
according to Lemma 7.2.
λmin (H) ≤ 3B. On the other hand, we always have λi + λmin (H) ≥ 10κ
Therefore, the total number of outer loops is at most O(log(Bκ)).
G.1
Matrix Inverse
Since the key component of the running time is the computation of (H+λi I)−1 b for different vectors
b we will first bound the condition number of the matrix (H + λi I)−1 via the following lemma
Claim G.4. Through out the execution of FastCubicMin and BinarySearch whenever we compute
λi +L2
(H + λi I)−1 b for some vector b it satisfies λi +λ
≤ 10κL2 .
min (H)
Proof of Claim G.4. We first focus on Line 5 and Line 11 of FastCubicMin. There are two cases. If
λi +L2
λi ≥ 2L2 , then according to −L2 I H L2 I we can bound λi +λ
≤ 3 because the left hand
min (H)
22
side is the largest when λi = 2L2 . If λi < 2L2 , then by Lemma 7.2 we know λi + λmin (H) ≥
λi +L2
This implies λi +λ
≤ 10κL2 .
min (H)
3
10κ .
We now focus on Line 3 of BinarySearch. We claim that all values λmid iterated over BinarySearch
3
3
also satisfy λmid + λmin (H) ≥ 10κ
(because the values λmid ≥ λi and λi satisfies λi + λmin (H) ≥ 10κ
according to Lemma 7.2). Therefore, the same case analysis (with respect to λmid ≥ 2L2 and
λi +L2
λmid < 2L2 ) also gives λi +λ
≤ 10κL2 .
min (H)
Claim G.5. Line 5 of FastCubicMin and Line 3 of BinarySearch runs in time Õ(Tinverse (κL2 , ε̃)).
Proof. Whenever we compute (H + λi I)−1 b for some vector v it satisfies kbk ≤ 1/ε̃; therefore to
find v satisfying kv + (H + λi I)−1 bk ≤ ε̃ it suffices to find kv + (H + λi I)−1 bk ≤ ε̃2 kbk. This costs
a total running time Õ(Tinverse (κL2 , ε̃)) according to Theorem 2.4.
Therefore by Theorem 2.4, every time we need to multiply a vector v to (H + λI)−1 to error
δ, the time required to approximately solve such a system is Tinverse (O(κL2 ), δ). We will state our
running time with respect to Tinverse as it is the dominant operation in the algorithm.
G.2
Power Method
We now bound the running time of Power Method in Line 11 of FastCubicMin. It is a folklore (cf. [3,
Appendix A]) that getting any constant multiplicative approximation to the leading eigenvector
of any PSD matrix M ∈ Rd×d requires only O(log d) iterations, each computing Mb for some
vector b. In our case, we have M = (H + λi I)−1 so we cannot compute Mb exactly. Fortunately,
folklore results on inaccurate power method suggests that, as long as each Mb is computed to a
very good accuracy such as ε̃−Ω(log d) , then we can still get a constant multiplicative approximate
leading eigenvector that satisfies Line 11 of FastCubicMin. Ignoring all the details (which are quite
standard and can be found for instance in [3, Appendix A]), we claim that
Claim G.6. Line 11 of FastCubicMin runs in time Õ Tinverse κL2 , ε−Θ(log(d)) = Õ (Tinverse (κL2 , ε)).
G.3
Lowest Eigenvector
We will now focus on the running time for the computation of the lowest eigenvector of the Hessian
which is required in Line 18. We recall Theorem 2.5 from Section 2 which uses Shift and Invert to
compute the largest eigenvalue of a matrix.
Since we are concerned with the lowest eigenvector of H and by assumption −L2 I H L2 I,
2I
we can equivalently compute the largest eigenvector of M , I − H+L
which satisfies 0 M I.
2L2
Note that computing Mv is of the same time complexity as computing Hv. By setting ε = δ× = 0.01
κL2
in Theorem 2.5 and running AppxPCA, we obtain a unit vector w such that
¬
λmin (H) + L2
w> Hw + L2
= w> Mw ≥ (1 − 2δ× )λmax (M) ≥ λmax (M) − 2δ× ≥ 1 −
− 2δ×
2L2
2L2
Above, ¬ uses λmax (M) ≤ 1. Rearranging the terms we obtain w> Hw ≤ λmin (H) + 0.05κ as
desired. In sum,
1−
Claim G.7. The approximate lowest eigenvector computation on Line 18 runs in time Õ (Tinverse (κL2 , ε̃)).
G.4
Putting It All Together
Running-Time Proof of Main Lemma 3. Putting together our bounds in Claim G.2 and Claim G.2
which bound the number of iterations, as well as our bounds in Claim G.6, Claim G.5, and
Claim G.7 for power method, matrix inverse, and lowest eigenvectors, we conclude that our total
23
running time of FastCubicMin is at most Õ (Tinverse (κL2 , ε̃)), where Õ contains factors polylogarithmic in κ, L, L2 , B, d.
By putting together our choice of ε̃ in Line 2 as well as the running time of either accelerated
gradient descent or accelerated SVRG from Theorem 2.4 into formula O(κL2 , ε̃), we finish the proof
of the running time part for Main Lemma 3.
H
Proof of Lemma G.1
2 )B
Lemma G.1. If λ2 +λmin (H) ≥ c1 ∈ (0, 1) then BinarySearch ends in O log( (λc11−λ
·L·ε̃ ) iterations.
Proof of Lemma G.1. We first note that in all iterations of BinarySearch it always satisfies
Lk(H + λ1 I)−1 gk ≤ 2λ1
and
Lk(H + λ2 I)−1 gk ≥ 2λ2 .
(H.1)
This is true at the beginning. In each of the follow-up iterations, if we have set λ1 ← λmid then it
must satisfy Lkvk + Lε̃ ≤ 2λmid but this implies Lk(H + λmid I)−1 gk ≤ 2λmid according to triangle
inequality and kv + (H + λmid I)−1 gk ≤ Lε̃; similarly, if we have set λ2 ← λmid then it must satisfy
Lk(H + λmid I)−1 gk ≥ 2λmid .
2
Suppose now the loop has run for at least log2 ( λ1 −λ
ε̂ ) iterations where ε̂ ,
satisfy λ1 − λ2 ≤ ε̂. At this point, we compute
and therefore
Lε̃c1
40B .
Then, it must
(H + λ1 I)−1 = (H + λ2 I)−1 − (λ1 − λ2 )(H + λ2 I)−1 (H + λ1 I)−1
Lk(H + λ1 I)−1 gk ≥ Lk(H + λ2 I)−1 gk − Lk(λ1 − λ2 )(H + λ2 I)−1 (H + λ1 I)−1 gk
¬
≥ 2λ2 − ε̂k(H + λ2 I)−1 k · 2λ1 ≥ 2λ1 − 2ε̂ − ε̂k(H + λ2 I)−1 k · 2λ1
Above, inequality ¬ uses (H.1) and λ1 − λ2 ≤ ε̂; inequality uses again λ1 − λ2 ≤ ε̂.
Now, we notice that k(H + λ2 I)−1 k ≤ c11 and λ1 ≤ 2B because λ2 only increases and λ1 only
ε̃c1
decreases through the execution of the algorithm. Therefore by the choice of ε̂ = 40B
, we get
Lk(H + λ1 I)−1 gk ≥ 2λ1 − Lε̃/5 .
A completely analogous argument also shows that
Lk(H + λ2 I)−1 gk ≤ 2λ2 + Lε̃/5 .
Therefore, in the immediate next iteration when picking λmid ← (λ1 + λ2 )/2, it must satisfy
2λmid − Lε̃/2 ≤ 2λ − Lε̃/5 ≤ Lk(H + λmid I)−1 gk ≤ 2λ2 + Lε̃/5 ≤ 2λmid + Lε̃/2 .
Then, at this iteration when v is computed to satisfy kv + (H + λmid I)−1 gk ≤ ε̃/2, we also have
2λmid − Lε̃ ≤ Lkvk ≤ 2λmid + Lε̃
which means BinarySearch will stop in this iteration. In sum, we have concluded that there will be
2 )B
no more than O log( (λc11−λ
·L·ε̃ ) iterations.
Acknowledgements
We thank Ben Recht for helpful suggestions and corrections to a previous version.
References
[1] Naman Agarwal, Brian Bullins, and Elad Hazan. Second order stochastic optimization for
machine learning in linear time. arXiv preprint arXiv:1602.03943, 2016.
24
[2] Zeyuan Allen-Zhu and Elad Hazan. Variance Reduction for Faster Non-Convex Optimization.
In ICML, 2016.
[3] Zeyuan Allen-Zhu and Yuanzhi Li. Even Faster SVD Decomposition Yet Without Agonizing
Pain. In NIPS, 2016.
[4] Zeyuan Allen-Zhu and Yang Yuan. Improved SVRG for Non-Strongly-Convex or Sum-of-NonConvex Objectives. In ICML, 2016.
[5] Afonso S Bandeira, Nicolas Boumal, and Vladislav Voroninski. On the low-rank approach
for semidefinite programs arising in synchronization and community detection. arXiv preprint
arXiv:1602.04426, 2016.
[6] S. Bhojanapalli, B. Neyshabur, and N. Srebro. Global Optimality of Local Search for Low
Rank Matrix Recovery. ArXiv e-prints, May 2016.
[7] Yair Carmon, John C. Duchi, Oliver Hinder, and Aaron Sidford. Accelerated methods for
non-convex optimization. arXiv preprint 1611.00756, 2016.
[8] Coralia Cartis, Nicholas IM Gould, and Philippe L Toint. Adaptive cubic regularisation methods for unconstrained optimization. part i: motivation, convergence and numerical results.
Mathematical Programming, 127(2):245–295, 2011.
[9] Coralia Cartis, Nicholas IM Gould, and Philippe L Toint. Adaptive cubic regularisation methods for unconstrained optimization. part ii: worst-case function-and derivative-evaluation complexity. Mathematical Programming, 130(2):295–319, 2011.
[10] Anna Choromanska, Mikael Henaff, Michael Mathieu, Gérard Ben Arous, and Yann LeCun.
The loss surfaces of multilayer networks. In AISTATS, 2015.
[11] Yann N Dauphin, Razvan Pascanu, Caglar Gulcehre, Kyunghyun Cho, Surya Ganguli, and
Yoshua Bengio. Identifying and attacking the saddle point problem in high-dimensional nonconvex optimization. In Advances in neural information processing systems, pages 2933–2941,
2014.
[12] John Duchi, Elad Hazan, and Yoram Singer. Adaptive subgradient methods for online learning
and stochastic optimization. The Journal of Machine Learning Research, 12:2121–2159, 2011.
[13] Dan Garber and Elad Hazan. Fast and simple PCA via convex optimization. ArXiv e-prints,
September 2015.
[14] Dan Garber, Elad Hazan, Chi Jin, Sham M. Kakade, Cameron Musco, Praneeth Netrapalli,
and Aaron Sidford. Robust shift-and-invert preconditioning: Faster and more sample efficient
algorithms for eigenvector computation. In ICML, 2016.
[15] Rong Ge, Furong Huang, Chi Jin, and Yang Yuan. Escaping from saddle points—online
stochastic gradient for tensor decomposition. arXiv:1503.02101, 2015.
[16] Rong Ge, Furong Huang, Chi Jin, and Yang Yuan. Escaping from saddle points—online
stochastic gradient for tensor decomposition. In Proceedings of the 28th Annual Conference
on Learning Theory, COLT 2015, 2015.
25
[17] Rong Ge, Furong Huang, Chi Jin, and Yang Yuan. Escaping from saddle points - online
stochastic gradient for tensor decomposition. In Proceedings of The 28th Conference on Learning Theory, COLT 2015, Paris, France, July 3-6, 2015, pages 797–842, 2015.
[18] Rong Ge, Jason Lee, and Tengyu Ma. Matrix Completion has No Spurious Local Minimum.
ArXiv e-prints, May 2016.
[19] Rong Ge and Tengyu Ma. On the optimization landscape of tensor decompositions, 2016.
[20] Saeed Ghadimi and Guanghui Lan. Accelerated gradient methods for nonconvex nonlinear
and stochastic programming. Mathematical Programming, pages 1–26, feb 2015.
[21] I. J. Goodfellow, O. Vinyals, and A. M. Saxe. Qualitatively characterizing neural network
optimization problems. ArXiv e-prints, December 2014.
[22] Elad Hazan and Tomer Koren. A linear-time algorithm for trust region problems. Mathematical
Programming, pages 1–19, 2015.
[23] Christopher J. Hillar and Lek-Heng Lim. Most tensor problems are np-hard. J. ACM, 60(6):45,
2013.
[24] Jason D. Lee, Max Simchowitz, Michael I. Jordan, and Benjamin Recht. Gradient descent only
converges to minimizers. In Proceedings of the 29th Conference on Learning Theory, COLT
2016, New York, USA, June 23-26, 2016, pages 1246–1257, 2016.
[25] Katta G Murty and Santosh N Kabadi. Some np-complete problems in quadratic and nonlinear
programming. Mathematical programming, 39(2):117–129, 1987.
[26] Yurii Nesterov. A method of solving a convex programming problem with convergence rate
O(1/k 2 ). In Doklady AN SSSR (translated as Soviet Mathematics Doklady), volume 269, pages
543–547, 1983.
[27] Yurii Nesterov. Introductory Lectures on Convex Programming Volume: A Basic course, volume I. Kluwer Academic Publishers, 2004.
[28] Yurii Nesterov and Boris T Polyak. Cubic regularization of newton method and its global
performance. Mathematical Programming, 108(1):177–205, 2006.
[29] Barak A Pearlmutter. Fast exact multiplication by the hessian. Neural computation, 6(1):147–
160, 1994.
[30] Herbert Robbins and Sutton Monro. A stochastic approximation method. The annals of
mathematical statistics, pages 400–407, 1951.
[31] David E Rumelhart, Geoffrey E Hinton, and Ronald J Williams. Learning representations by
back-propagating errors. Cognitive modeling, 5(3):1, 1988.
[32] Mark Schmidt, Nicolas Le Roux, and Francis Bach. Minimizing finite sums with the stochastic
average gradient. arXiv preprint arXiv:1309.2388, pages 1–45, 2013. Preliminary version
appeared in NIPS 2012.
[33] Shai Shalev-Shwartz. SDCA without Duality, Regularization, and Individual Convexity. In
ICML, 2016.
26
[34] Jonathan Richard Shewchuk. An introduction to the conjugate gradient method without the
agonizing pain, 1994.
[35] Paul Werbos. Beyond regression: New tools for prediction and analysis in the behavioral
sciences. 1974.
27
| 8cs.DS
|
Improved Techniques for Training GANs
arXiv:1606.03498v1 [cs.LG] 10 Jun 2016
Tim Salimans
[email protected]
Ian Goodfellow
[email protected]
Wojciech Zaremba
[email protected]
Alec Radford
[email protected]
Vicki Cheung
[email protected]
Xi Chen
[email protected]
Abstract
We present a variety of new architectural features and training procedures that we
apply to the generative adversarial networks (GANs) framework. We focus on two
applications of GANs: semi-supervised learning, and the generation of images
that humans find visually realistic. Unlike most work on generative models, our
primary goal is not to train a model that assigns high likelihood to test data, nor do
we require the model to be able to learn well without using any labels. Using our
new techniques, we achieve state-of-the-art results in semi-supervised classification on MNIST, CIFAR-10 and SVHN. The generated images are of high quality
as confirmed by a visual Turing test: our model generates MNIST samples that
humans cannot distinguish from real data, and CIFAR-10 samples that yield a human error rate of 21.3%. We also present ImageNet samples with unprecedented
resolution and show that our methods enable the model to learn recognizable features of ImageNet classes.
1
Introduction
Generative adversarial networks [1] (GANs) are a class of methods for learning generative models
based on game theory. The goal of GANs is to train a generator network G(z; θ (G) ) that produces
samples from the data distribution, pdata (x), by transforming vectors of noise z as x = G(z; θ (G) ).
The training signal for G is provided by a discriminator network D(x) that is trained to distinguish
samples from the generator distribution pmodel (x) from real data. The generator network G in turn
is then trained to fool the discriminator into accepting its outputs as being real.
Recent applications of GANs have shown that they can produce excellent samples [2, 3]. However,
training GANs requires finding a Nash equilibrium of a non-convex game with continuous, highdimensional parameters. GANs are typically trained using gradient descent techniques that are
designed to find a low value of a cost function, rather than to find the Nash equilibrium of a game.
When used to seek for a Nash equilibrium, these algorithms may fail to converge [4].
In this work, we introduce several techniques intended to encourage convergence of the GANs game.
These techniques are motivated by a heuristic understanding of the non-convergence problem. They
lead to improved semi-supervised learning peformance and improved sample generation. We hope
that some of them may form the basis for future work, providing formal guarantees of convergence.
All code and hyperparameters may be found at:
improved_gan
1
https://github.com/openai/
2
Related work
Several recent papers focus on improving the stability of training and the resulting perceptual quality
of GAN samples [2, 3, 5, 6]. We build on some of these techniques in this work. For instance, we
use some of the “DCGAN” architectural innovations proposed in Radford et al. [3], as discussed
below.
One of our proposed techniques, feature matching, discussed in Sec. 3.1, is similar in spirit to
approaches that use maximum mean discrepancy [7, 8, 9] to train generator networks [10, 11].
Another of our proposed techniques, minibatch features, is based in part on ideas used for batch
normalization [12], while our proposed virtual batch normalization is a direct extension of batch
normalization.
One of the primary goals of this work is to improve the effectiveness of generative adversarial
networks for semi-supervised learning (improving the performance of a supervised task, in this case,
classification, by learning on additional unlabeled examples). Like many deep generative models,
GANs have previously been applied to semi-supervised learning [13, 14], and our work can be seen
as a continuation and refinement of this effort.
3
Toward Convergent GAN Training
Training GANs consists in finding a Nash equilibrium to a two-player non-cooperative game.
Each player wishes to minimize its own cost function, J (D) (θ (D) , θ (G) ) for the discriminator and
J (G) (θ (D) , θ (G) ) for the generator. A Nash equilibirum is a point (θ (D) , θ (G) ) such that J (D) is at a
minimum with respect to θ (D) and J (G) is at a minimum with respect to θ (G) . Unfortunately, finding Nash equilibria is a very difficult problem. Algorithms exist for specialized cases, but we are not
aware of any that are feasible to apply to the GAN game, where the cost functions are non-convex,
the parameters are continuous, and the parameter space is extremely high-dimensional.
The idea that a Nash equilibrium occurs when each player has minimal cost seems to intuitively motivate the idea of using traditional gradient-based minimization techniques to minimize each player’s
cost simultaneously. Unfortunately, a modification to θ (D) that reduces J (D) can increase J (G) , and
a modification to θ (G) that reduces J (G) can increase J (D) . Gradient descent thus fails to converge
for many games. For example, when one player minimizes xy with respect to x and another player
minimizes −xy with respect to y, gradient descent enters a stable orbit, rather than converging to
x = y = 0, the desired equilibrium point [15]. Previous approaches to GAN training have thus
applied gradient descent on each player’s cost simultaneously, despite the lack of guarantee that this
procedure will converge. We introduce the following techniques that are heuristically motivated to
encourage convergence:
3.1
Feature matching
Feature matching addresses the instability of GANs by specifying a new objective for the generator
that prevents it from overtraining on the current discriminator. Instead of directly maximizing the
output of the discriminator, the new objective requires the generator to generate data that matches
the statistics of the real data, where we use the discriminator only to specify the statistics that we
think are worth matching. Specifically, we train the generator to match the expected value of the
features on an intermediate layer of the discriminator. This is a natural choice of statistics for the
generator to match, since by training the discriminator we ask it to find those features that are most
discriminative of real data versus data generated by the current model.
Letting f (x) denote activations on an intermediate layer of the discriminator, our new objective for
the generator is defined as: ||Ex∼pdata f (x) − Ez∼pz (z) f (G(z))||22 . The discriminator, and hence
f (x), are trained in the usual way. As with regular GAN training, the objective has a fixed point
where G exactly matches the distribution of training data. We have no guarantee of reaching this
fixed point in practice, but our empirical results indicate that feature matching is indeed effective in
situations where regular GAN becomes unstable.
2
3.2
Minibatch discrimination
One of the main failure modes for GAN is for the generator to collapse to a parameter setting where
it always emits the same point. When collapse to a single mode is imminent, the gradient of the
discriminator may point in similar directions for many similar points. Because the discriminator
processes each example independently, there is no coordination between its gradients, and thus no
mechanism to tell the outputs of the generator to become more dissimilar to each other. Instead,
all outputs race toward a single point that the discriminator currently believes is highly realistic.
After collapse has occurred, the discriminator learns that this single point comes from the generator,
but gradient descent is unable to separate the identical outputs. The gradients of the discriminator
then push the single point produced by the generator around space forever, and the algorithm cannot
converge to a distribution with the correct amount of entropy. An obvious strategy to avoid this type
of failure is to allow the discriminator to look at multiple data examples in combination, and perform
what we call minibatch discrimination.
The concept of minibatch discrimination is quite general: any discriminator model that looks
at multiple examples in combination, rather than in isolation, could potentially help avoid collapse of the generator. In fact, the successful application of batch normalization in the discriminator by Radford et al. [3] is well explained from this perspective. So far, however, we
have restricted our experiments to models that explicitly aim to identify generator samples that
are particularly close together. One successful specification for modelling the closeness between
examples in a minibatch is as follows: Let f (xi ) ∈ RA denote a vector of features for input xi , produced by some intermediate layer in the discriminator. We then multiply the vector
f (xi ) by a tensor T ∈ RA×B×C , which results in a matrix Mi ∈ RB×C . We then compute
the L1 -distance between the rows of the resulting matrix Mi across samples i ∈ {1, 2, . . . , n}
and apply a negative exponential (Fig. 1): cb (xi , xj ) = exp(−||Mi,b − Mj,b ||L1 ) ∈ R.
The output o(xi ) for this minibatch layer for a sample xi
is then defined as the sum of the cb (xi , xj )’s to all other
samples:
n
X
cb (xi , xj ) ∈ R
o(xi )b =
j=1
h
i
o(xi ) = o(xi )1 , o(xi )2 , . . . , o(xi )B ∈ RB
o(X) ∈ Rn×B
Next, we concatenate the output o(xi ) of the minibatch
layer with the intermediate features f (xi ) that were its Figure 1: Figure sketches how miniinput, and we feed the result into the next layer of the batch discrimination works. Features
discriminator. We compute these minibatch features sep- f (x ) from sample x are multiplied
i
i
arately for samples from the generator and from the train- through
a tensor T , and cross-sample
ing data. As before, the discriminator is still required to distance is computed.
output a single number for each example indicating how
likely it is to come from the training data: The task of the discriminator is thus effectively still to
classify single examples as real data or generated data, but it is now able to use the other examples in
the minibatch as side information. Minibatch discrimination allows us to generate visually appealing
samples very quickly, and in this regard it is superior to feature matching (Section 6). Interestingly,
however, feature matching was found to work much better if the goal is to obtain a strong classifier
using the approach to semi-supervised learning described in Section 5.
3.3
Historical averaging
Pt
When applying this technique, we modify each player’s cost to include a term ||θ − 1t i=1 θ[i]||2 ,
where θ[i] is the value of the parameters at past time i. The historical average of the parameters can
be updated in an online fashion so this learning rule scales well to long time series. This approach is
loosely inspired by the fictitious play [16] algorithm that can find equilibria in other kinds of games.
We found that our approach was able to find equilibria of low-dimensional, continuous non-convex
games, such as the minimax game with one player controlling x, the other player controlling y, and
value function (f (x) − 1)(y − 1), where f (x) = x for x < 0 and f (x) = x2 otherwise. For
3
these same toy games, gradient descent fails by going into extended orbits that do not approach the
equilibrium point.
3.4
One-sided label smoothing
Label smoothing, a technique from the 1980s recently independently re-discovered by Szegedy et.
al [17], replaces the 0 and 1 targets for a classifier with smoothed values, like .9 or .1, and was
recently shown to reduce the vulnerability of neural networks to adversarial examples [18].
Replacing positive classification targets with α and negative targets with β, the optimal discriminator
data (x)+βpmodel (x)
becomes D(x) = αppdata
(x)+pmodel (x) . The presence of pmodel in the numerator is problematic
because, in areas where pdata is approximately zero and pmodel is large, erroneous samples from
pmodel have no incentive to move nearer to the data. We therefore smooth only the positive labels to
α, leaving negative labels set to 0.
3.5
Virtual batch normalization
Batch normalization greatly improves optimization of neural networks, and was shown to be highly
effective for DCGANs [3]. However, it causes the output of a neural network for an input example
x to be highly dependent on several other inputs x0 in the same minibatch. To avoid this problem
we introduce virtual batch normalization (VBN), in which each example x is normalized based on
the statistics collected on a reference batch of examples that are chosen once and fixed at the start
of training, and on x itself. The reference batch is normalized using only its own statistics. VBN is
computationally expensive because it requires running forward propagation on two minibatches of
data, so we use it only in the generator network.
4
Assessment of image quality
Generative adversarial networks lack an objective function, which makes it difficult to
compare performance of different models.
One intuitive metric of performance can be
obtained by having human annotators judge the visual quality of samples [2].
We
automate this process using Amazon Mechanical Turk (MTurk), using the web interface in figure Fig. 2 (live at http://infinite-chamber-35121.herokuapp.com/
cifar-minibatch/), which we use to ask annotators to distinguish between generated data
and real data. The resulting quality assessments of our models are described in Section 6.
Figure 2: Web interface given to annotators. Annotators are asked to distinguish computer generated images from
real ones.
A downside of using human annotators is that the metric
varies depending on the setup of the task and the motivation of the annotators. We also find that results change
drastically when we give annotators feedback about their
mistakes: By learning from such feedback, annotators are
better able to point out the flaws in generated images, giving a more pessimistic quality assessment. The left column of Fig. 2 presents a screen from the annotation process, while the right column shows how we inform annotators about their mistakes.
As an alternative to human annotators, we propose an automatic method to evaluate samples, which
we find to correlate well with human evaluation: We apply the Inception model1 [19] to every
generated image to get the conditional label distribution p(y|x). Images that contain meaningful
objects should have a conditional label distribution p(y|x)
with low entropy. Moreover, we expect
R
the model to generate varied images, so the marginal p(y|x = G(z))dz should have high entropy.
Combining these two requirements, the metric that we propose is: exp(Ex KL(p(y|x)||p(y))), where
we exponentiate results so the values are easier to compare. Our Inception score is closely related
to the objective used for training generative models in CatGAN [14]: Although we had less success
using such an objective for training, we find it is a good metric for evaluation that correlates very
1
We use the pretrained Inception model from http://download.tensorflow.org/models/
image/imagenet/inception-2015-12-05.tgz. Code to compute the Inception score with this
model will be made available by the time of publication.
4
well with human judgment. We find that it’s important to evaluate the metric on a large enough
number of samples (i.e. 50k) as part of this metric measures diversity.
5
Semi-supervised learning
Consider a standard classifier for classifying a data point x into one of K possible classes. Such
a model takes in x as input and outputs a K-dimensional vector of logits {l1 , . . . , lK }, that can
exp(lj )
be turned into class probabilities by applying the softmax: pmodel (y = j|x) = PK exp(l
. In
k)
k=1
supervised learning, such a model is then trained by minimizing the cross-entropy between the
observed labels and the model predictive distribution pmodel (y|x).
We can do semi-supervised learning with any standard classifier by simply adding samples from
the GAN generator G to our data set, labeling them with a new “generated” class y = K + 1, and
correspondingly increasing the dimension of our classifier output from K to K + 1. We may then
use pmodel (y = K + 1 | x) to supply the probability that x is fake, corresponding to 1 − D(x) in
the original GAN framework. We can now also learn from unlabeled data, as long as we know that
it corresponds to one of the K classes of real data by maximizing log pmodel (y ∈ {1, . . . , K}|x).
Assuming half of our data set consists of real data and half of it is generated (this is arbitrary), our
loss function for training the classifier then becomes
L = −Ex,y∼pdata (x,y) [log pmodel (y|x)] − Ex∼G [log pmodel (y = K + 1|x)]
= Lsupervised + Lunsupervised , where
Lsupervised = −Ex,y∼pdata (x,y) log pmodel (y|x, y < K + 1)
Lunsupervised = −{Ex∼pdata (x) log[1 − pmodel (y = K + 1|x)] + Ex∼G log[pmodel (y = K + 1|x)]},
where we have decomposed the total cross-entropy loss into our standard supervised loss function
Lsupervised (the negative log probability of the label, given that the data is real) and an unsupervised
loss Lunsupervised which is in fact the standard GAN game-value as becomes evident when we substitute D(x) = 1 − pmodel (y = K + 1|x) into the expression:
Lunsupervised = −{Ex∼pdata (x) log D(x) + Ez∼noise log(1 − D(G(z)))}.
The optimal solution for minimizing both Lsupervised and Lunsupervised is to have
exp[lj (x)] = c(x)p(y=j, x)∀j<K+1 and exp[lK+1 (x)] = c(x)pG (x) for some undetermined scaling function c(x). The unsupervised loss is thus consistent with the supervised loss in
the sense of Sutskever et al. [13], and we can hope to better estimate this optimal solution from
the data by minimizing these two loss functions jointly. In practice, Lunsupervised will only help if
it is not trivial to minimize for our classifier and we thus need to train G to approximate the data
distribution. One way to do this is by training G to minimize the GAN game-value, using the
discriminator D defined by our classifier. This approach introduces an interaction between G and
our classifier that we do not fully understand yet, but empirically we find that optimizing G using
feature matching GAN works very well for semi-supervised learning, while training G using GAN
with minibatch discrimination does not work at all. Here we present our empirical results using this
approach; developing a full theoretical understanding of the interaction between D and G using this
approach is left for future work.
Finally, note that our classifier with K + 1 outputs is over-parameterized: subtracting a general
function f (x) from each output logit, i.e. setting lj (x) ← lj (x) − f (x)∀j, does not change the
output of the softmax. This means we may equivalently fix lK+1 (x) = 0∀x, in which case Lsupervised
becomes the standard supervised loss function of our original
classifier with K classes, and our
PK
Z(x)
discriminator D is given by D(x) = Z(x)+1
, where Z(x) = k=1 exp[lk (x)].
5.1
Importance of labels for image quality
Besides achieving state-of-the-art results in semi-supervised learning, the approach described above
also has the surprising effect of improving the quality of generated images as judged by human
annotators. The reason appears to be that the human visual system is strongly attuned to image
statistics that can help infer what class of object an image represents, while it is presumably less
sensitive to local statistics that are less important for interpretation of the image. This is supported
5
by the high correlation we find between the quality reported by human annotators and the Inception
score we developed in Section 4, which is explicitly constructed to measure the “objectness” of a
generated image. By having the discriminator D classify the object shown in the image, we bias it to
develop an internal representation that puts emphasis on the same features humans emphasize. This
effect can be understood as a method for transfer learning, and could potentially be applied much
more broadly. We leave further exploration of this possibility for future work.
6
Experiments
We performed semi-supervised experiments on MNIST, CIFAR-10 and SVHN, and sample generation experiments on MNIST, CIFAR-10, SVHN and ImageNet. We provide code to reproduce the
majority of our experiments.
6.1
MNIST
The MNIST dataset contains 60, 000 labeled
images of digits. We perform semi-supervised
training with a small randomly picked fraction
of these, considering setups with 20, 50, 100,
and 200 labeled examples. Results are averaged
over 10 random subsets of labeled data, each
chosen to have a balanced number of examples
from each class. The remaining training images
are provided without labels. Our networks have
5 hidden layers each. We use weight normalization [20] and add Gaussian noise to the output Figure 3: (Left) samples generated by model durof each layer of the discriminator. Table 1 sum- ing semi-supervised training. Samples can be
marizes our results.
clearly distinguished from images coming from
Samples generated by the generator during MNIST dataset. (Right) Samples generated with
semi-supervised learning using feature match- minibatch discrimination. Samples are coming (Section 3.1) do not look visually appealing pletely indistinguishable from dataset images.
(left Fig. 3). By using minibatch discrimination
instead (Section 3.2) we can improve their visual quality. On MTurk, annotators were able to distinguish samples in 52.4% of cases (2000 votes total), where 50% would be obtained by random
guessing. Similarly, researchers in our institution were not able to find any artifacts that would allow them to distinguish samples. However, semi-supervised learning with minibatch discrimination
does not produce as good a classifier as does feature matching.
Model
20
DGN [21]
Virtual Adversarial [22]
CatGAN [14]
Skip Deep Generative Model [23]
Ladder network [24]
Auxiliary Deep Generative Model [23]
Our model
Ensemble of 10 of our models
Number of incorrectly predicted test examples
for a given number of labeled samples
50
100
1677 ± 452
1134 ± 445
221 ± 136
142 ± 96
333 ± 14
212
191 ± 10
132 ± 7
106 ± 37
96 ± 2
93 ± 6.5
86 ± 5.6
200
90 ± 4.2
81 ± 4.3
Table 1: Number of incorrectly classified test examples for the semi-supervised setting on permutation invariant MNIST. Results are averaged over 10 seeds.
6.2
CIFAR-10
CIFAR-10 is a small, well studied dataset of 32 × 32 natural images. We use this data set to study
semi-supervised learning, as well as to examine the visual quality of samples that can be achieved.
For the discriminator in our GAN we use a 9 layer deep convolutional network with dropout and
weight normalization. The generator is a 4 layer deep CNN with batch normalization. Table 2
summarizes our results on the semi-supervised learning task.
6
Model
1000
Ladder network [24]
CatGAN [14]
Our model
Ensemble of 10 of our models
21.83±2.01
19.22±0.54
Test error rate for
a given number of labeled samples
2000
4000
19.61±2.09
17.25±0.66
20.40±0.47
19.58±0.46
18.63±2.32
15.59±0.47
8000
17.72±1.82
14.87±0.89
Table 2: Test error on semi-supervised CIFAR-10. Results are averaged over 10 splits of data.
Figure 4: Samples generated during semi-supervised training on CIFAR-10 with feature matching
(Section 3.1, left) and minibatch discrimination (Section 3.2, right).
When presented with 50% real and 50% fake data generated by our best CIFAR-10 model, MTurk
users correctly categorized 78.7% of images correctly. However, MTurk users may not be sufficiently familiar with CIFAR-10 images or sufficiently motivated; we ourselves were able to categorize images with > 95% accuracy. We validated the Inception score described above by observing
that MTurk accuracy drops to 71.4% when the data is filtered by using only the top 1% of samples
according to the Inception score. We performed a series of ablation experiments to demonstrate that
our proposed techniques improve the Inception score, presented in Table 3. We also present images
for these ablation experiments—in our opinion, the Inception score correlates well with our subjective judgment of image quality. Samples from the dataset achieve the highest value. All the models
that even partially collapse have relatively low scores. We caution that the Inception score should be
used as a rough guide to evaluate models that were trained via some independent criterion; directly
optimizing Inception score will lead to the generation of adversarial examples [25].
6.3
SVHN
For the SVHN data set, we used the same architecture and experimental setup as for CIFAR-10.
Model
Percentage of incorrectly predicted test examples
for a given number of labeled samples
500
1000
2000
DGN [21]
Virtual Adversarial [22]
Auxiliary Deep Generative Model [23]
Skip Deep Generative Model [23]
Our model
Ensemble of 10 of our models
36.02±0.10
24.63
22.86
16.61±0.24
8.11 ± 1.3
5.88 ± 1.0
18.44 ± 4.8
6.16 ± 0.58
Figure 5: (Left) Error rate on SVHN. (Right) Samples from the generator for SVHN.
7
Samples
Model
Score ± std.
Real data
11.24 ± .12
Our methods
8.09 ± .07
-VBN+BN
7.54 ± .07
-L+HA
6.86 ± .06
-LS
6.83 ± .06
-L
4.36 ± .04
-MBF
3.87 ± .03
Table 3: Table of Inception scores for samples generated by various models for 50, 000 images.
Score highly correlates with human judgment, and the best score is achieved for natural images.
Models that generate collapsed samples have relatively low score. This metric allows us to avoid
relying on human evaluations. “Our methods” includes all the techniques described in this work,
except for feature matching and historical averaging. The remaining experiments are ablation experiments showing that our techniques are effective. “-VBN+BN” replaces the VBN in the generator
with BN, as in DCGANs. This causes a small decrease in sample quality on CIFAR. VBN is more
important for ImageNet. “-L+HA” removes the labels from the training process, and adds historical
averaging to compensate. HA makes it possible to still generate some recognizable objects. Without
HA, sample quality is considerably reduced (see ”-L”). “-LS” removes label smoothing and incurs a
noticeable drop in performance relative to “our methods.” “-MBF” removes the minibatch features
and incurs a very large drop in performance, greater even than the drop resulting from removing the
labels. Adding HA cannot prevent this problem.
6.4
ImageNet
We tested our techniques on a dataset of unprecedented scale: 128 × 128 images from the
ILSVRC2012 dataset with 1,000 categories. To our knowledge, no previous publication has applied a generative model to a dataset with both this large of a resolution and this large a number
of object classes. The large number of object classes is particularly challenging for GANs due to
their tendency to underestimate the entropy in the distribution. We extensively modified a publicly
available implementation of DCGANs2 using TensorFlow [26] to achieve high performance, using
a multi-GPU implementation. DCGANs without modification learn some basic image statistics and
generate contiguous shapes with somewhat natural color and texture but do not learn any objects.
Using the techniques described in this paper, GANs learn to generate objects that resemble animals,
but with incorrect anatomy. Results are shown in Fig. 6.
Figure 6: Samples generated from the ImageNet dataset. (Left) Samples generated by a DCGAN.
(Right) Samples generated using the techniques proposed in this work. The new techniques enable
GANs to learn recognizable features of animals, such as fur, eyes, and noses, but these features are
not correctly combined to form an animal with realistic anatomical structure.
2
https://github.com/carpedm20/DCGAN-tensorflow
8
7
Conclusion
Generative adversarial networks are a promising class of generative models that has so far been
held back by unstable training and by the lack of a proper evaluation metric. This work presents
partial solutions to both of these problems. We propose several techniques to stabilize training
that allow us to train models that were previously untrainable. Moreover, our proposed evaluation
metric (the Inception score) gives us a basis for comparing the quality of these models. We apply
our techniques to the problem of semi-supervised learning, achieving state-of-the-art results on a
number of different data sets in computer vision. The contributions made in this work are of a
practical nature; we hope to develop a more rigorous theoretical understanding in future work.
References
[1] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, et al. Generative adversarial nets. In NIPS, 2014.
[2] Emily Denton, Soumith Chintala, Arthur Szlam, and Rob Fergus. Deep generative image models using a
laplacian pyramid of adversarial networks. arXiv preprint arXiv:1506.05751, 2015.
[3] Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434, 2015.
[4] Ian J Goodfellow. On distinguishability criteria for estimating generative models.
arXiv:1412.6515, 2014.
arXiv preprint
[5] Daniel Jiwoong Im, Chris Dongjoo Kim, Hui Jiang, and Roland Memisevic. Generating images with
recurrent adversarial networks. arXiv preprint arXiv:1602.05110, 2016.
[6] Donggeun Yoo, Namil Kim, Sunggyun Park, Anthony S Paek, and In So Kweon. Pixel-level domain
transfer. arXiv preprint arXiv:1603.07442, 2016.
[7] Arthur Gretton, Olivier Bousquet, Alex Smola, and Bernhard Schölkopf. Measuring statistical dependence with hilbert-schmidt norms. In Algorithmic learning theory, pages 63–77. Springer, 2005.
[8] Kenji Fukumizu, Arthur Gretton, Xiaohai Sun, and Bernhard Schölkopf. Kernel measures of conditional
dependence. In NIPS, volume 20, pages 489–496, 2007.
[9] Alex Smola, Arthur Gretton, Le Song, and Bernhard Schölkopf. A hilbert space embedding for distributions. In Algorithmic learning theory, pages 13–31. Springer, 2007.
[10] Yujia Li, Kevin Swersky, and Richard S. Zemel. Generative moment matching networks. CoRR,
abs/1502.02761, 2015.
[11] Gintare Karolina Dziugaite, Daniel M Roy, and Zoubin Ghahramani. Training generative neural networks
via maximum mean discrepancy optimization. arXiv preprint arXiv:1505.03906, 2015.
[12] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing
internal covariate shift. arXiv preprint arXiv:1502.03167, 2015.
[13] Ilya Sutskever, Rafal Jozefowicz, Karol Gregor, et al. Towards principled unsupervised learning. arXiv
preprint arXiv:1511.06440, 2015.
[14] Jost Tobias Springenberg. Unsupervised and semi-supervised learning with categorical generative adversarial networks. arXiv preprint arXiv:1511.06390, 2015.
[15] Ian Goodfellow, Yoshua Bengio, and Aaron Courville. Deep Learning. 2016. MIT Press.
[16] George W Brown. Iterative solution of games by fictitious play. Activity analysis of production and
allocation, 13(1):374–376, 1951.
[17] C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens, and Z. Wojna. Rethinking the Inception Architecture for
Computer Vision. ArXiv e-prints, December 2015.
[18] David Warde-Farley and Ian Goodfellow. Adversarial perturbations of deep neural networks. In Tamir
Hazan, George Papandreou, and Daniel Tarlow, editors, Perturbations, Optimization, and Statistics, chapter 11. 2016. Book in preparation for MIT Press.
[19] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens, and Zbigniew Wojna. Rethinking
the inception architecture for computer vision. arXiv preprint arXiv:1512.00567, 2015.
[20] Tim Salimans and Diederik P Kingma. Weight normalization: A simple reparameterization to accelerate
training of deep neural networks. arXiv preprint arXiv:1602.07868, 2016.
[21] Diederik P Kingma, Shakir Mohamed, Danilo Jimenez Rezende, and Max Welling. Semi-supervised
learning with deep generative models. In Neural Information Processing Systems, 2014.
9
[22] Takeru Miyato, Shin-ichi Maeda, Masanori Koyama, Ken Nakae, and Shin Ishii. Distributional smoothing
by virtual adversarial examples. arXiv preprint arXiv:1507.00677, 2015.
[23] Lars Maaløe, Casper Kaae Sønderby, Søren Kaae Sønderby, and Ole Winther. Auxiliary deep generative
models. arXiv preprint arXiv:1602.05473, 2016.
[24] Antti Rasmus, Mathias Berglund, Mikko Honkala, Harri Valpola, and Tapani Raiko. Semi-supervised
learning with ladder networks. In Advances in Neural Information Processing Systems, 2015.
[25] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, et al. Intriguing properties of neural networks.
arXiv preprint arXiv:1312.6199, 2013.
[26] Martı́n Abadi, Ashish Agarwal, Paul Barham, et al. TensorFlow: Large-scale machine learning on heterogeneous systems, 2015. Software available from tensorflow.org.
10
| 9cs.NE
|
Transitivity of Commutativity for Linear Time-Varying Analog Systems
Mehmet Emir KOKSAL
Department of Mathematics, Ondokuz Mayis University, 55139 Atakum, Samsun, Turkey
[email protected]
Abstract: In this contribution, the transitivity property of commutative first-order linear timevarying systems is investigated with and without initial conditions. It is proven that transitivity
property of first-order systems holds with and without initial conditions. On the base of impulse
response function, transitivity of commutation property is formulated for any triplet of
commutative linear time-varying relaxed systems. Transitivity proves are given for some
special combinations of first and second-order linear time-varying systems which are initially
relaxed.
Keywords: Commutativity, Differential equations, Initial conditions, Time-varying systems,
linear systems, Impulse response.
AMS Subject Classification: 93C05, 93C15, 93A30
1.
Introduction
As the main branches of applied mathematics, differential (and integral) equations arise in
many areas of sciences and engineering including acoustics, electromagnetic, electrodynamics,
fluid dynamics etc. There is a great deal of papers on the theory, technique and applications of
differential equations. Especially, they are used as a major tool in order to achieve many
developments in real engineering problems by modelling, analyzing and solving naturel
problems. For example, an interdisciplinary branch of applied mathematics and electricelectronics engineering, they play a pioneering role in system and control theory, that deal with
the behavior of dynamical systems with inputs, and how their behavior is modified by different
combinations such as cascade and feedback connections which is. When the cascade
connection in system design is considered, the commutativity concept places a prominent role
to improve different system performances.
1
Consider a system 𝐴 described by a linear time-varying differential equation of the form
𝑛𝐴
𝐴: ∑ 𝑎𝑖 (𝑡)
𝑖=0
𝑑𝑖
𝑦 (𝑡) = 𝑥𝐴 (𝑡);
𝑑𝑡 𝑖 𝐴
(1)
(𝑥 − 𝑦) represents input-output pair of the system at any time 𝑡 ∈ 𝑅, 𝑎𝑖 (𝑡) are time-varying
(𝑖)
coefficients with 𝑎𝑛 (𝑡) ≠ 0, 𝑛𝐴 ≥ 0 is the oder of system; and 𝑦𝐴 (𝑡0 ) ∈ 𝑅, 𝑖 = 0, 1, ⋯ , 𝑛𝐴 −
1 are the initial conditions at the initial time 𝑡0 ∈ 𝑅.
When two systems of this type are interconnected sequentially so that the output of the former
feeds the input of the later, it is said that they are connected in cascade [1]. If the order of
connection does not affect the input-output relation of the combined system 𝐴𝐵 or 𝐵𝐴, it is
said that systems 𝐴 and 𝐵 are commutative.
Figure 1: Cascade connection of the differential system 𝐴 and 𝐵
If the combined system has an overall input-output relation invariant with the sequence of
connection, it is said that these systems are commutative [2]. In [2], J. E. Marshall has proven
that “for commutativity, either both systems are time-invariant or both systems are timevarying”. In addition, he has revealed the commutativity conditions of first-order systems.
Later Marshall’s work, a great deal of researches has been done on commutativity. In [3-5],
the necessary and sufficient conditions for commutativity of second-order systems are
presented. Koksal has presented the general commutativity conditions for time-varying
systems of any order and reformulated the previous results obtained for seccond-order systems
in the format of general conditions [6]. In this work, the general conditions are used to show
that any system with constant forward and feedback gains is commutative with the system
itself, which is an important fact for the feedback control theory.
In 1985, Koksal prepared a technical report which is a survey on commutativity [7]. In this
2
report, an iterative formula is derived and an explicit formula is given for the entries of the
coefficient matrix expressing the first set of commutativity conditions and the second set of
commutativity conditions, respectively. Hence, the theorem stating these conditions is
formally proved. Morover, explicit commutativity results for fourth-order systems are
obtained. Finally, commutativity of Euler’s system is proved.
The content of the published but undistributed work [7] can be found in the exhaustive journal
paper of M. Koksal introduced the basic fundamentals of the subject [8]. His paper covers
almost all the previous results except the ones related with initial conditions and sensitivity.
This paper is the first tutorial paper that has appeared in the literature.
More than one decade no publication had been appeared in the literature until the work in 2011
[9]. This reference is the second basic journal publication after the first appeared in 1988 [8].
In [9], another generic paper by the same author has presented explicit commutativity
conditions of fifth-order systems in addition to reviews of commutativity of systems with nonzero initial conditions, commutativity and system disturbance, commutativity of Euler
systems.
Research on commutativity has not confined to analog systems only; there has been some
literature on the subject for discrete-time systems as well [10, 11]. Hence, the research is
continuing on both digital and analog area [12]. In [12], explicit results for finding all the
second order commutative pairs of a first-order linear time-varying system have been given
and the derived theoretical results have been verified by an example. About the commutativity
of continuous time linear time-varying systems, [13] has been the last paper appearing in the
literature; it deals with driving necessary and sufficiently conditions for the decomposition of
a second order linear time-varying system into two cascade connected commutative first-order
linear time-varying subsystems.
After a short introduction of the literature in this section, the transitivity property of
commutativity is introduced in Section 2. Transitivity property of commutative first-order
systems with and without initial conditions is studied in Section 3. In Section 4, transitivity
property of commutativity for relaxed systems of any order is formulated on the base of
impulse response function. Section 5 covers the verification of the general results formulated
in Section 4 on the base of impulse response function for first-order systems. Section 6
illustrates the results presented in Sections 4 and 5. Section 7 includes transitivity proves for
some combinations of first and second-order relaxed linear time-varying systems. Finally, the
paper finishes with conclusions and future work which compose the last section, Section 8.
3
2.
Transitivity Property of Commutativity
Logically or mathematically, transitivity is a property of a binary relation such that whenever
one element is related to a second element, and the second element is related to a third element,
then the first element is also related to the third element.
Let 𝐴, 𝐵, 𝐶 be dynamical systems described by linear time-varying differential equations of the
form (1) with their special input-output variables, orders of complexity, and coefficients. If 𝐴
and 𝐵 are commutative among themselves, further, 𝐵 and 𝐶 are also commutative amoung
themselves, what can it be said about the commutativity of systems 𝐴 and 𝐶. If 𝐴 and 𝐶 are
also commutative among themselves, this property is called transitivity property of
commutativity, in other words commutativity is a transitive relation.
For
commutative
systems
𝐴, 𝐵, 𝐶
satisfiying
transitivity,
all
the
triplets
𝐴𝐵𝐶, 𝐴𝐶𝐵, 𝐵𝐴𝐶, 𝐵𝐶𝐴, 𝐶𝐴𝐵, 𝐶𝐵𝐴 yield the same functionally equivalent cascade connected
system and hence the same input-output property. The preference of an individual system
connection depends on its relative performance characteristics such as sensitivity, disturbance,
robustness, and etc., with respect to the others.
It is well-known (or trivial to show) that linear scalar systems, that is systems of order 0, are
always commutative among themselves whether time-invariant or time-varying; and
transitivity always hold for such systems. Further, initially relaxed time-invariant systems are
always commutative. When initial conditions exist, which is valid for non-scalar (1st or higherorder) systems, the commutativity of such systems (even including a scalar system) is not
automatic and requires some conditions. Hence, the following discussions are devoted mainly
to commutativity of systems at least one of them is of order one or higher.
3.
Transitivity for First-order Systems
Let 𝐴, 𝐵, 𝐶 be first-order linear time-varying systems defined by a differential equation of the
form (1). Let 𝑎1 , 𝑎0 ; 𝑏1 , 𝑏0 ; and 𝑐1 , 𝑐0 are time-varying (in general) coefficients of these
systems. Assume also that (𝐴, 𝐵) and (𝐵, 𝐶) are commutative pairs.
The first commutativity conditions [7-9] for (𝐴, 𝐵) and (𝐵, 𝐶) yields
𝑏
𝑎
[ 1] = [ 1
𝑎
𝑏0
0
4
0 𝑘1
] [ ],
1 𝑘0
(2a)
𝑐1
𝑏
[𝑐 ] = [ 1
𝑏0
0
0 ℓ1
] [ ],
1 ℓ0
(2b)
respectively; where 𝑘1 , 𝑘0 , ℓ1 , ℓ0 are some constants.
Inserting the values of 𝑏1 , 𝑏0 as expressed in (2a) into Eq. (2b) and rearranging, we obtain
𝑐1
𝑎
[𝑐 ] = [ 1
𝑎0
0
0
𝑘1 ℓ1
][
].
1 𝑘0 ℓ1 + ℓ0
(2c)
This equation already satisfies the first commutativity condition for Systems 𝐴 and 𝐶; hence,
𝐴 and 𝐶 are commutative pairs under zero initial conditions.
For the case of existence of non-zero initial conditions, the second commutativity condition
[9] for (𝐴, 𝐵) and (𝐵, 𝐶) yields
𝑦𝐴 (𝑡0 ) = 𝑦𝐵 (𝑡0 ) ≠ 0,
(3a)
1 − 𝑎0 (𝑡0 ) 1 − 𝑏0 (𝑡0 )
=
;
𝑎1 (𝑡0 )
𝑏1 (𝑡0 )
(3b)
𝑦𝐵 (𝑡0 ) = 𝑦𝐶 (𝑡0 ) ≠ 0,
(4a)
1 − 𝑏0 (𝑡0 ) 1 − 𝑐0 (𝑡0 )
=
,
𝑏1 (𝑡0 )
𝑐1 (𝑡0 )
(4b)
respectively. Hence, Eqs. (3) and (4) lead to write
𝑦𝐴 (𝑡0 ) = 𝑦𝐶 (𝑡0 ) ≠ 0,
(5a)
1 − 𝑎0 (𝑡0 ) 1 − 𝑐0 (𝑡0 )
=
.
𝑎1 (𝑡0 )
𝑐1 (𝑡0 )
(5b)
The result in Eq. (5) clearly shows the satisfaction of the second commutativity condition for
Subsystems 𝐴 and 𝐶. Hence, 𝐴 and 𝐶 are also commutative when initial conditions exist, and
transitivity property is valid for non-relaxed first-order linear time-varying systems as well.
In fact, Eq. (3b) is not satisfied for all systems 𝐴 and 𝐵 satisfying commutativity conditions
(2a) under relaxed conditions. Using (2a) we see that Eq. (3b) requires
1 − 𝑎0 (𝑡0 ) 1 − 𝑏0 (𝑡0 ) 1 − 𝑘1 𝑎0 (𝑡0 ) − 𝑘0
=
=
=
𝑎1 (𝑡0 )
𝑏1 (𝑡0 )
𝑘1 𝑎1 (𝑡0 )
5
1 − 𝑘0
− 𝑎0 (𝑡0 )
𝑘1
;
𝑎1 (𝑡0 )
(6a)
Or equivalently
𝑘0 = 1 − 𝑘1 .
(6b)
Similarly, using (2b), we see that Eq. (4b) requires
1 − 𝑏0 (𝑡0 ) 1 − 𝑐0 (𝑡0 ) 1 − ℓ1 𝑏0 (𝑡0 ) − ℓ0
=
=
=
𝑏1 (𝑡0 )
𝑐1 (𝑡0 )
ℓ1 𝑏1 (𝑡0 )
1 − ℓ0
− 𝑏0 (𝑡0 )
ℓ1
,
𝑏1 (𝑡0 )
ℓ0 = 1 − ℓ1 .
(7a)
(7b)
Using (2c) with (6b) and (7b), we can easily show that (5b) is satisfied; That is:
1 − 𝑐0 (𝑡0 ) 1 − [𝑘1 ℓ1 𝑎0 (𝑡0 ) + 𝑘0 ℓ1 + ℓ0 ]
=
=
𝑐1 (𝑡0 )
𝑘1 ℓ1 𝑎1 (𝑡0 )
1 − 𝑘0 ℓ1 − ℓ0
− 𝑎0 (𝑡0 )
𝑘1 ℓ1
𝑎1 (𝑡0 )
1 − (1 − 𝑘1 )ℓ1 − (1 − ℓ1 )
− 𝑎0 (𝑡0 ) 1 − 𝑎 (𝑡 )
𝑘1 ℓ1
0 0
=
=
,
𝑎1 (𝑡0 )
𝑎1 (𝑡0 )
(8)
which is exactly Eq. (5b); hence, 𝐴 and 𝐶 are commutative under non-zero conditions as well
if the constants 𝑘1 , ℓ1 , 𝑘0 , , ℓ0 satisfy (6b) and (7b). These are explicit commutativity conditions
under non-zero initial states equivalently replacing by (3b) and (4b) in addition to (2a) and (2b)
which are necessary and sufficient for relaxed case. With these conditions 5b is also satisfied,
so that transitivity is valid.
We state the results obtained so far by two theorems:
Theorem 1: The necessary and sufficient conditions that two first-order linear time-varying
systems 𝐴 and 𝐵 described by differential equations of the form (1) are commutative under
zero initial conditions are that
i.
The coefficients of System 𝐵 must be expressible in terms of the coefficients of those of
𝐴 as in Eq. (2a) where 𝑘1 ≠ 0 and 𝑘0 are arbitrary constants.
ii.
Further, Systems 𝐴 and 𝐵 are commutative under arbitrary non-zero initial conditions
as well if and only if Condition (i) is satisfied with 𝑘0 = 1 − 𝑘1 (that is 𝑘1 ≠ 0 and 𝑘0
cannot be chosen arbitrarily), and their non-zero initial conditions must be equal.
Note that this theorem is the first theorem expressing the commutativity conditions for
unrelaxed systems by using the arbitrary constants relating the coefficients of two systems for
their commutativity in relaxed conditions. Note also that Eqs. (6b) and (7b) are independent of
6
the initial time. So, commutativity does not depend on the initial time 𝑡0 for first-order nonrelaxed systems.
Theorem 2: The transitivity property of commutativity for first-order linear time-varying
systems is always valid for both relaxed and un-relaxed commutative systems that is for
systems with zero and non-zero initial conditions.
4.
Transitivity for High-order Systems
In this section, higher-order linear time-varying systems with zero initial conditions are
considered; scalar (0-order) and first-order systems are also included in the scope. To prove
the transitivity property of commutativity, the impulse response function is used. The results
are general in the sense the subsystems of each commutative pair may have a different order
than its partner.
Let 𝐴 be a liear time-varying initially relaxed system having an impulse response ℎ𝐴 (𝑡, 𝜏); by
definition this is the response to a unit impulse occurring at time 𝜏 and observed at time 𝑡 ≥ 𝜏
[14]. And the response of 𝐴 for an arbitrary input 𝑢𝐴 (𝑡) is expressed by superposition integral
𝑡
𝑦𝐴 (𝑡) = ∫ ℎ𝐴 (𝑡, 𝛾)𝑢𝐴 (𝛾)𝑑𝛾 .
(9a)
𝑡0
Consider now the cascade connection of 𝐴 and 𝐵 as shown in Fig. 2 and denote this connection
by 𝐴𝐵. Smilar to Eq. (9a), for the output of 𝐵 we write
𝑡
𝑦𝐵 (𝑡) = ∫ ℎ𝐵 (𝑡, 𝛾)𝑢𝐵 (𝛾)𝑑𝛾 .
(9b)
𝑡0
It is obvious form Fig. 1, and using Eq. (9a)
𝜏
𝑢𝐵 (𝜏 ) = ∫ ℎ𝐵 (𝜏, 𝛾)𝑢(𝛾)𝑑𝛾 .
𝑡0
Figure 2: Cascade connection 𝐴𝐵 of Systems 𝐴 and 𝐵.
7
(9c)
Finally, inserting Eq. (9c) in Eq. (9b), and observing 𝑦𝐵 (𝑡) in Fig. 2, we obtain
𝑡
𝜏
𝑦(𝑡 ) = ∫ ℎ𝐵 (𝑡, 𝜏) [ ∫ ℎ𝐴 (𝜏, 𝛾)𝑢(𝛾)𝑑𝛾] 𝑑𝜏 .
𝑡0
(9d)
𝑡0
This is the response of 𝐴𝐵 for any input 𝑢(𝑡) applied for 𝑡 ≥ 𝑡0 .To find the impulse response
ℎ𝐴 (𝑡, 𝑡0 ) of the connection 𝐴𝐵, we substitute 𝑢(𝑡) = 𝛿(𝑡 − 𝑡0 ) in Eq. (9d) and arrive
𝑡
𝑡
𝜏
ℎ𝐴𝐵 (𝑡, 𝑡0 ) = ∫ ℎ𝐵 (𝑡, 𝜏) [∫ ℎ𝐴 (𝜏, 𝛾)𝛿(𝛾 − 𝑡0 )𝑑𝛾] 𝑑𝜏 = ∫ ℎ𝐵 (𝑡, 𝜏)ℎ𝐴 (𝜏, 𝑡0 )𝑑𝜏 , (10a)
𝑡0
𝑡0
𝑡0
where the second equality results from the property of impulse function. In a similar result, the
impulse response ℎ𝐵𝐴 (𝑡, 𝜏) of the connection 𝐴𝐵 can be written as
𝑡
ℎ𝐵𝐴 (𝑡, 𝑡0 ) = ∫ ℎ𝐴 (𝑡, 𝜏)ℎ𝐵 (𝜏, 𝑡0 )𝑑𝜏 .
(10b)
𝑡0
Since equal impulse responses yield equal input-output pairs, for commutativity of 𝐴 and 𝐵 it
must be true that
ℎ𝐴𝐵 (𝑡, 𝑡0 ) = ℎ𝐵𝐴 (𝑡, 𝑡0 );
∀𝑡 ≥ 𝑡0 , ∀𝑡0 ∈ 𝑅.
(11a)
Using Eqs. (10a, b); (11a) can be written as
𝑡
𝑡
∫ ℎ𝐵 (𝑡, 𝜏)ℎ𝐴 (𝜏, 𝑡0 )𝑑𝜏 = ∫ ℎ𝐴 (𝑡, 𝜏)ℎ𝐵 (𝜏, 𝑡0 )𝑑𝜏 ;
𝑡0
∀𝑡 ≥ 𝑡0 , ∀𝑡0 ∈ 𝑅,
(11b)
∀𝑡 ≥ 𝑡0 , ∀𝑡0 ∈ 𝑅.
(11c)
𝑡0
or equivalently
𝑡
∫[ℎ𝐵 (𝑡, 𝜏)ℎ𝐴 (𝜏, 𝑡0 ) − ℎ𝐴 (𝑡, 𝜏)ℎ𝐵 (𝜏, 𝑡0 )] 𝑑𝜏 = 0;
𝑡0
In fact, either of Eqs. (11b) or Eq. (11c) express necessary and sufficient conditions in terms
of impulse responses for relaxed systems 𝐴 and 𝐵 to be commutative.
On the other hand, for commutativity of 𝐵 and 𝐶 under zero inintial conditions, the neccessary
and suffifcient condition can be obtained from (11) by changing 𝐴 → 𝐵 and 𝐵 → 𝐶, the result
is
8
ℎ𝐵𝐶 (𝑡, 𝑡0 ) = ℎ𝐶𝐵 (𝑡, 𝑡0 );
𝑡
∀𝑡 ≥ 𝑡0 , ∀𝑡0 ∈ 𝑅,
(12a)
𝑡
∫ ℎ𝐶 (𝑡, 𝜏)ℎ𝐵 (𝜏, 𝑡0 )𝑑𝜏 = ∫ ℎ𝐵 (𝑡, 𝜏)ℎ𝐶 (𝜏, 𝑡0 )𝑑𝜏 ;
𝑡0
∀𝑡 ≥ 𝑡0 , ∀𝑡0 ∈ 𝑅,
(12b)
𝑡0
or equivalently,
𝑡
∫[ℎ𝐶 (𝑡, 𝜏)ℎ𝐵 (𝜏, 𝑡0 ) − ℎ𝐵 (𝑡, 𝜏)ℎ𝐶 (𝜏, 𝑡0 )] 𝑑𝜏 = 0;
∀𝑡 ≥ 𝑡0 , ∀𝑡0 ∈ 𝑅.
(12c)
𝑡0
For the proof of the commutativity of 𝐶 and 𝐴, that is the transitivity property of commutativity,
it is required to show
ℎ𝐶𝐴 (𝑡, 𝑡0 ) = ℎ𝐴𝐶 (𝑡, 𝑡0 );
𝑡
∀𝑡 ≥ 𝑡0 , ∀𝑡0 ∈ 𝑅, that is,
(13a)
𝑡
∫ ℎ𝐴 (𝑡, 𝜏)ℎ𝐶 (𝜏, 𝑡0 )𝑑𝜏 = ∫ ℎ𝐶 (𝑡, 𝜏)ℎ𝐴 (𝜏, 𝑡0 )𝑑𝜏 ;
𝑡0
∀𝑡 ≥ 𝑡0 , ∀𝑡0 ∈ 𝑅,
(13b)
∀𝑡 ≥ 𝑡0 , ∀𝑡0 ∈ 𝑅.
(13c)
𝑡0
or equivalently,
𝑡
∫[ℎ𝐴 (𝑡, 𝜏)ℎ𝐶 (𝜏, 𝑡0 ) − ℎ𝐶 (𝑡, 𝜏)ℎ𝐴 (𝜏, 𝑡0 )] 𝑑𝜏 = 0;
𝑡0
In the light of the presence of equalities in (11b) and (12b), the work left for the proof of
transitivity is to show the validity of (13b); if so then (13a) holds. This process will end with
the proof of transitivity of linear time-varying relaxed systems of higher orders.
It worth’s to remark that Eqs. (10a, b) are first used by J. E. Marshall to prove his assertion
mentioned in Introduction (Eqs. (2, 3) in [2]). Later, they are used by the author for studying
commutativity of Euler systems (Eqs. (34a, b) in [7]).
When we consider a scalar system
𝐴: 𝑎0 (𝑡)𝑦𝐴 (𝑡) = 𝑥𝐴 (𝑡)
(14a)
for which its response ℎ𝐴 (𝑡, 𝑡0 ) to a unit impulse 𝛿(𝑡 − 𝑡0 ) is obviously
ℎ𝐴 (𝑡, 𝑡0 ) = 𝑦𝐴 (𝑡) =
1
1
𝑥𝐴 (𝑡) =
𝛿(𝑡 − 𝑡0 ).
𝑎0 (𝑡)
𝑎0 (𝑡0 )
9
(14b)
Then, let 𝐵 be a time-varying system of 1st order or higher with impulse response ℎ𝐵 (𝑡, 𝑡0 ).
Eqs. (10a) and (10b) lead to
𝑡
ℎ𝐴𝐵 (𝑡, 𝑡0 ) = ∫ ℎ𝐵 (𝑡, 𝜏)
𝑡0
𝑡
ℎ𝐵𝐴 (𝑡, 𝑡0 ) = ∫
𝑡0
𝛿(𝜏 − 𝑡0 )
ℎ𝐵 (𝑡, 𝑡0 )
𝑑𝜏 =
,
𝑎0 (𝑡0 )
𝑎0 (𝑡0 )
𝛿(𝑡 − 𝜏)
ℎ𝐵 (𝑡, 𝑡0 )
ℎ𝐵 (𝜏, 𝑡0 )𝑑𝜏 =
.
𝑎0 (𝜏)
𝑎0 (𝑡)
(15a)
(15b)
From these equations, it is obvious that a scalar system that can be commutative with a 1st or
higher order system must be a time-invariant system; that is 𝑎0 (𝑡) = 𝑎0 (𝑡0 ) = 𝑐𝑜𝑛𝑠𝑡𝑎𝑛𝑡 for
all 𝑡 ≥ 𝑡0 . In other words, a scalar time-varying system cannot commutative with any order of
higher order (1st, 2nd …) time-varying system. The only commutative pairs of it are the scalar
systems. Though commutativity of a constant gain system (scalar time-invariant system) is
always valid with time-varying system of any order, if the initial conditions exist the
commutativity is possible only if the constant gain is 1; that is System 𝐴 is an identity.
The validation of transitivity using the equivalence of impulse responses ℎ𝐴𝐶 (𝑡, 𝑡0 ) and
ℎ𝐶𝐴 (𝑡, 𝑡0 ) for first-order systems are considered in Section 5. But the extension of this result
for systems at least one is 2nd order or higher is not straight forward and appears to be an
unsolved problem. Instead of jumping this problem fully, we consider the special case that
transitivity holds conditionally. Hence, the following discussion is not restricted by order
limitation.
Let 𝐴, 𝐵, 𝐶 be linear time-varying systems with impulse responses ℎ𝐴 , ℎ𝐵 , ℎ𝐶 satisfying
ℎ𝐵 (𝑡, 𝜏)ℎ𝐴 (𝜏, 𝑡0 ) − ℎ𝐴 (𝑡, 𝜏)ℎ𝐵 (𝜏, 𝑡0 ) = 0,
(16a)
ℎ𝐶 (𝑡, 𝜏)ℎ𝐵 (𝜏, 𝑡0 )−ℎ𝐵 (𝑡, 𝜏)ℎ𝐶 (𝜏, 𝑡0 ) = 0.
(16b)
Eq. (16a) implies that Eqs. (11a, b, c) are satisfied, hence (𝐴, 𝐵) is a comutative pair. In a
similar manner (16b) implies that (𝐵, 𝐶) is also a commutative pair.
Multiplying (16b) by ℎ𝐴 (𝜏, 𝑡0 ) and inserting the equivalence of ℎ𝐵 (𝑡, 𝜏)ℎ𝐴 (𝜏, 𝑡0 ) as obtained
from (16a), we find
ℎ𝐶 (𝑡, 𝜏)ℎ𝐵 (𝜏, 𝑡0 )ℎ𝐴 (𝜏, 𝑡0 ) − ℎ𝐵 (𝑡, 𝜏)ℎ𝐶 (𝜏, 𝑡0 )ℎ𝐴 (𝜏, 𝑡0 ) = 0,
(17𝑎)
ℎ𝐶 (𝑡, 𝜏)ℎ𝐵 (𝜏, 𝑡0 )ℎ𝐴 (𝜏, 𝑡0 ) − ℎ𝐶 (𝜏, 𝑡0 )ℎ𝐴 (𝑡, 𝜏)ℎ𝐵 (𝜏, 𝑡0 ) = 0,
(17𝑏)
10
[ℎ𝐶 (𝑡, 𝜏)ℎ𝐴 (𝜏, 𝑡0 ) − ℎ𝐴 (𝑡, 𝜏)ℎ𝐶 (𝜏, 𝑡0 )]ℎ𝐵 (𝜏, 𝑡0 ) = 0.
(17𝑐)
Since ℎ𝐵 (𝜏, 𝑡0 ) ≢ 0, it follows that
ℎ𝐶 (𝑡, 𝜏)ℎ𝐴 (𝜏, 𝑡0 ) − ℎ𝐴 (𝑡, 𝜏)ℎ𝐶 (𝜏, 𝑡0 ) = 0.
(18𝑎)
This is sufficient to write
𝑡
𝑡
∫ ℎ𝐶 (𝑡, 𝜏)ℎ𝐴 (𝜏, 𝑡0 )𝑑𝜏 − ∫ ℎ𝐴 (𝑡, 𝜏)ℎ𝐶 (𝜏, 𝑡0 )𝑑𝜏 = 0.
𝑡0
(18b)
𝑡0
Comparing with Eqs. (10a) and (10b), the first integral in (18b) is the impulse response of 𝐴𝐶
and the second one is that of 𝐶𝐴; hence, (18b) yields
ℎ𝐴𝐶 (𝑡, 𝑡0 ) = ℎ𝐴𝐶 (𝑡, 𝑡0 ).
(19)
So, that 𝐴 and 𝐶 are commutative. This proves the transitivity proper of higher-order systems
under the set assumptions.
5.
Verifications for first-order systems
This section verifies the transitivity property of first-order linear time-varying systems in their
relaxed modes as presented in Section 3 by using the general results obtained for systems of
any order in Section 4.
Consider a first-order linear time-varying system 𝐴 defined by
𝑎1 (𝑡)
𝑑
𝑦 (𝑡) + 𝑎0 (𝑡)𝑦𝐴 (𝑡) = 𝑥𝐴 (𝑡),
𝑑𝑡 𝐴
(20)
with zero initial conditions. The impulse response ℎ𝐴 (𝑡, 𝑡0 ) of this system to an impulse
𝛿(𝑡 − 𝑡0 ) occurring at 𝑡 = 𝑡0 is obviously expressed as [14, 15]
ℎ𝐴 (𝑡, 𝑡0 ) =
𝑡 𝑎 (𝛾)
1
1
∫𝑡 − 0 (𝛾)𝑑𝛾
𝑒 0 𝑎1
=
𝑒 𝑓𝐴(𝑡)− 𝑓𝐴(𝑡0 ) ,
(𝑡
)
(𝑡
)
𝑎1 0
𝑎1 0
(21a)
where
𝑓𝐴 (𝛾) = ∫ −
𝑎0 (𝛾)
𝑑𝛾 ,
𝑎1 (𝛾)
𝑑
𝑎0 (𝛾)
𝑓𝐴 (𝛾) = −
.
𝑑𝛾
𝑎1 (𝛾)
11
(21b)
(21c)
Hence, for Systems 𝐵 and 𝐶 of the same type of 𝐴, it can be written that
ℎ𝐵 (𝑡, 𝑡0 ) =
𝑡 𝑏 (𝛾)
1
1
∫ − 0 𝑑𝛾
𝑒 𝑡0 𝑏1 (𝛾) =
𝑒 𝑓𝐵 (𝑡)− 𝑓𝐵 (𝑡0 ) ,
𝑏1 (𝑡0 )
𝑏1 (𝑡0 )
𝑓𝐵 (𝛾) = ∫ −
𝑑
𝑓 (𝛾)
𝑑𝛾 𝐵
𝑡
𝑏0 (𝛾)
𝑑𝛾 ,
𝑏1 (𝛾)
(22b)
𝑏 (𝛾)
= − 𝑏0 (𝛾) ,
(22c)
1
𝑐 (𝛾)
1
∫𝑡 −𝑐0 (𝛾)𝑑𝛾
0
1
𝑒
1 (𝑡0 )
ℎ𝐶 (𝑡, 𝑡0 ) = 𝑐
𝑡
𝑓𝐶 (𝛾) = ∫ −
𝑡0
=𝑐
(22a)
1
1 (𝑡0 )
𝑒 𝑓𝐶 (𝑡)− 𝑓𝐶 (𝑡0 ) ,
𝑐0 (𝛾)
𝑑𝛾 ,
𝑐1 (𝛾)
𝑑
𝐶0 (𝛾)
𝑓𝐶 (𝛾) = −
.
𝑑𝛾
𝐶1 (𝛾)
(23a)
(23b)
(23c)
For commutativity of 𝐴 and 𝐵, consider the verification of Eq. (11a) (or equivalently Eqs. (11b,
c)). Using the relations in Eq. (2a) between the coefficients of 𝐴 and 𝐵 in Eq. (22b) and then
recognizing (21b), we obtain
𝑓𝐵 (𝛾) = ∫ −
𝑏0 (𝛾)
𝑘1 𝑎0 (𝛾) + 𝑘0
𝑎0 (𝛾)
𝑘0
1
𝑑𝛾 = ∫ −
𝑑𝛾 = ∫ −
𝑑𝛾 − ∫
𝑑𝛾
𝑏1 (𝛾)
𝑘1 𝑎1 (𝛾)
𝑎1 (𝛾)
𝑘1 𝑎1 (𝛾)
= 𝑓𝐴 (𝛾) −
𝑘0
1
∫
𝑑𝛾 .
𝑘1 𝑎1 (𝛾)
(24a)
Using this result in Eq. (22a), we obtain
𝑘
𝑘
1
1
1
𝑓𝐴 (𝑡)− 0 ∫
𝑑𝑡−𝑓𝐴 (𝑡0 )+ 0 ∫
𝑑𝑡0
𝑘
(𝑡)
𝑘
(𝑡
𝑎
𝑎
1
1
1
1
0)
ℎ𝐵 (𝑡, 𝑡0 ) =
𝑒
,
𝑘1 𝑎1 (𝑡0 )
=
1
1
1
𝑒 𝑓𝐴(𝑡)− 𝑓𝐴(𝑡0 ) 𝑒 𝑔(𝑡0 )− 𝑔(𝑡) =
𝑘1 𝑎1 (𝑡0 )
𝑘1
𝑒 𝑔(𝑡0 )− 𝑔(𝑡) ℎ𝐴 (𝑡, 𝑡0 ),
(24b)
where
𝑔(𝑡) =
𝑘0
1
∫
𝑑𝑡 .
𝑘1 𝑎1 (𝑡)
By using (15b) in (7a), we obtain
12
(24c)
𝑡
1 𝑔(𝜏)− 𝑔(𝑡)
𝑒
ℎ𝐴 (𝑡, 𝜏)ℎ𝐴 (𝜏, 𝑡0 )𝑑𝜏 .
𝑘1
ℎ𝐴𝐵 (𝑡, 𝑡0 ) = ∫
𝑡0
(25)
Before proceeding further, we show the transitivity property of ℎ𝐴 (•,•); in fact using (21a)
ℎ𝐴 (𝑡, 𝜏)ℎ𝐴 (𝜏, 𝑡0 ) =
=
1
1
𝑒 𝑓𝐴(𝑡)− 𝑓𝐴(𝜏)
𝑒 𝑓𝐴(𝜏)− 𝑓𝐴(𝑡0 )
𝑎1 (𝜏)
𝑎1 (𝑡0 )
1
1
1
𝑒 𝑓𝐴(𝑡)− 𝑓𝐴 (𝑡0 ) =
ℎ (𝑡, 𝑡0 ).
𝑎1 (𝜏) 𝑎1 (𝑡0 )
𝑎1 (𝜏) 𝐴
(26)
Same property holds for the remaining first-order systems 𝐵 and 𝐶 as well. Another relation
we need in the future proves is arrived as follows:
𝑡
∫
𝑡0
𝑡
𝑔(𝜏)
𝑒
𝑑𝜏 = ∫
𝑎1 (𝜏)
𝑡0
𝑘0
1
∫ (𝜏)𝑑𝜏
𝑘
𝑎
1
1
𝑒
𝑎1 (𝜏)
𝑑𝜏 =
𝑘0
1
𝑘1 𝑘1
𝑑𝜏 𝑡
∫
𝑒 𝑎1 (𝜏) |
𝑘0
𝑡0
𝑘1 𝑔(𝜏) 𝑡
𝑘1
𝑒
| = [𝑒 𝑔(𝑡) − 𝑒 𝑔(𝑡0 ) ].
𝑘0
𝑡0 𝑘0
(27a)
𝑒 −𝑔(𝜏)
𝑘1
∫
𝑑𝜏 = [𝑒 −𝑔(𝑡0 ) − 𝑒 −𝑔(𝑡) ].
𝑎1 (𝜏)
𝑘0
(27b)
=
Similarly, we obtain
𝑡
𝑡0
Using (26) and (27a) in (25), we have
𝑡
𝑡
𝑡0
𝑡0
1 𝑔(𝜏)− 𝑔(𝑡)
1
𝑒 − 𝑔(𝑡)
𝑒 𝑔(𝜏)
ℎ𝐴𝐵 (𝑡, 𝑡0 ) = ∫
𝑒
ℎ (𝑡, 𝑡0 )𝑑𝜏 =
ℎ (𝑡, 𝑡0 ) ∫
𝑑𝜏
𝑘1
𝑎1 (𝜏) 𝐴
𝑘1 𝐴
𝑎1 (𝜏)
=
𝑒 −𝑔(𝑡)
𝑘1
ℎ𝐴 (𝑡, 𝑡0 )
𝑘1
𝑘0
1
[𝑒 𝑔(𝑡) − 𝑒 𝑔(𝑡0 ) ] = 𝑘 ℎ𝐴 (𝑡, 𝑡0 ) [1 − 𝑒 𝑔(𝑡0 )−𝑔(𝑡) ].
0
Using (24b) in (10b) and using the formula (26) and (27b), we obtain sequentially
𝑡
ℎ𝐵𝐴 (𝑡, 𝑡0 ) = ∫ ℎ𝐴 (𝑡, 𝜏)
𝑡0
𝑡
= ∫
𝑡0
1 𝑔(𝑡 )− 𝑔(𝜏)
𝑒 0
ℎ𝐴 (𝜏, 𝑡0 )𝑑𝜏
𝑘1
1 𝑔(𝑡 )− 𝑔(𝜏)
1
𝑒 0
ℎ (𝑡, 𝑡0 )𝑑𝜏
𝑘1
𝑎1 (𝜏) 𝐴
13
(28)
𝑡
𝑒 𝑔(𝑡0 )
𝑒 −𝑔(𝜏)
𝑒 𝑔( 𝑡0 )
𝑘1 −𝑔(𝑡 )
0 − 𝑒 −𝑔(𝑡) ]
(𝑡,
)
=
ℎ𝐴 𝑡0 ∫
𝑑𝜏 =
ℎ𝐴 (𝑡, 𝑡0 )
[𝑒
(𝜏)
𝑘1
𝑎1
𝑘1
𝑘0
𝑡0
1
= 𝑘 ℎ𝐴 (𝑡, 𝑡0 ) [1 − 𝑒 𝑔(𝑡0 )−𝑔(𝑡) ].
(29)
0
Obviously, 𝐴𝐵 and 𝐵𝐴 have the same impulse response as seen in Eqs. (28) and (29),
respectively; so 𝐴 and 𝐵 are commutative.
Remark 1: Although Eq. (11c), a property of impulse response of the commutative systems 𝐴
and 𝐵, seems to mislead to the conclusion of the integrant being zero; that is,
∆( 𝑡0 , 𝜏, 𝑡) ≜ ℎ𝐵 (𝑡, 𝜏)ℎ𝐴 (𝜏, 𝑡0 ) − ℎ𝐴 (𝑡, 𝜏)ℎ𝐵 (𝜏, 𝑡0 ) = 0.
(30)
This is strictly wrong. In fact, using (24b) for ℎ𝐴 (𝑡, 𝑡0 ), and (26) later, we have
∆( 𝑡0 , 𝜏, 𝑡) =
1 𝑔(𝜏)− 𝑔(𝑡)
1 𝑔(𝑡 )− 𝑔(𝜏)
𝑒
ℎ𝐴 (𝑡, 𝜏)ℎ𝐴 (𝜏, 𝑡0 ) − ℎ𝐴 (𝑡, 𝜏)
𝑒 0
ℎ𝐴 (𝜏, 𝑡0 ) = 0
𝑘1
𝑘1
=
1
ℎ (𝑡, 𝜏)ℎ𝐴 (𝜏, 𝑡0 ) [𝑒 𝑔(𝜏)−𝑔(𝑡) − 𝑒 𝑔(𝑡0 )−𝑔(𝜏) ]
𝑘1 𝐴
=
[𝑒 𝑔(𝜏)−𝑔(𝑡) − 𝑒 𝑔(𝑡0 )−𝑔(𝜏) ]
1
ℎ𝐴 (𝑡, 𝑡0 )
= 0.
𝑘1
𝑎1 (𝜏)
(31a)
Since, ℎ𝐴 (𝑡, 𝑡0 ) ≢ 0 and 𝑎1 (𝜏) ≢ 0, this implies
𝑒 𝑔(𝜏)−𝑔(𝑡) = 𝑒 𝑔(𝑡0 )−𝑔(𝜏) ,
(31b)
or equivalently
1
[𝑔(𝑡) + 𝑔( 𝑡0 )].
2
(31c)
1
1
1
1
𝑑𝜏 = [∫
𝑑𝑡 + ∫
𝑑 𝑡 ],
𝑎1 (𝜏)
2
𝑎1 (𝑡)
𝑎1 (𝑡0 ) 0
(31d)
𝑒 2𝑔(𝜏) = 𝑒 𝑔(𝑡)+𝑔( 𝑡0 ) ; 𝑔(𝜏) =
With (24c), the last equation is equivalent to
∫
which is impossible for an arbitrary function 𝑎1 (•) And for general values of 𝑡0 ≤ 𝜏 ≤ 𝑡.
14
Similar result obtained in Eqs. (25-29) can be obtained for commutative pair (𝐵, 𝐶); either by
the parallel derivations made for (𝐴, 𝐵), or by replacing 𝐴 → 𝐵, 𝐵 → 𝐶, so 𝑎1 → 𝑏1 and 𝑘1 , 𝑘0
by ℓ1 , ℓ0 , respectively.
For the commutativity of 𝐴 and 𝐶, to show that ℎ𝐴𝐶 (𝑡, 𝑡0 ) = ℎ𝐶𝐴 (𝑡, 𝑡0 ), the replacements 𝐵 →
𝐶, (𝑘1 , 𝑘0 ) → (𝑘1 ℓ1 , 𝑘0 ℓ1 + ℓ0 ) in Eqs. (25-29) complete the verification of transitivity for
initially relaxed first-order linear time- varying systems by using impulse responses.
6.
Example
This example is given for illustration of the results presented in Sections 4 and 5. Consider
system 𝐴 describe by
(𝑡 + 1)𝑦𝐴́ + (𝑡 + 2)𝑦𝐴 = 𝑥𝐴 , −1 < 𝑡0 ≤ 𝑡,
(32)
where 𝑎1 (𝑡) = 𝑡 + 1, 𝑎0 (𝑡) = 𝑡 + 2; 𝑓(𝑡) for 𝐴 is computed by using (21b) as
𝑓𝐴 (𝑡) = ∫ −
𝑡+2
𝑑𝑡 = −𝑡 − ln(𝑡 + 1).
𝑡+1
(33)
Then, by (21a), its impulse response is computed as
𝑡0 +1
𝑒 −𝑡−ln(𝑡+1)+𝑡0 +ln(𝑡0 +1) 𝑒 −𝑡+𝑡0 +ln 𝑡+1
1
ℎ𝐴 (𝑡, 𝑡0 ) =
=
=
𝑒 −𝑡+𝑡0 .
𝑡0 + 1
𝑡0 + 1
𝑡+1
(34)
Let subsystem 𝐵 be described by
2(𝑡 + 1)𝑦̇ 𝐵 + (2𝑡 + 5)𝑦𝐵 ; −1 < 𝑡0 ≤ 𝑡,
(35)
where 𝑏1 (𝑡) = 2(𝑡 + 1), 𝑏0 (𝑡) = 2𝑡 + 5 which can be obtained from 𝑎1 and 𝑎0 by using Eq.
(2a) with 𝑘1 = 2, 𝑘0 = 1; hence, (𝐴, 𝐵) is a commutative pair. Using (22b), 𝑓(𝑡) for 𝐵 can
be obtained as
𝑓𝐵 (𝑡) = ∫ −
2𝑡 + 5
𝑑𝑡 = −𝑡 − 1.5 ln(𝑡 + 1).
2(𝑡 + 1)
(36)
Then, by (22a), the impulse response of 𝐵 is computed as
𝑡0 +1
𝑒 −𝑡−1.5 ln(𝑡+1)+𝑡0 +1.5 ln(𝑡0 +1) 𝑒 −𝑡+𝑡0 +1.5 ln 𝑡+1
ℎ𝐵 (𝑡, 𝑡0 ) =
=
2(𝑡0 + 1)
2(𝑡0 + 1)
15
=
(𝑡0 + 1)1.5 (𝑡0 + 1)0.5 −𝑡+𝑡
1
0.
𝑒 −𝑡+𝑡0
=
𝑒
(𝑡 + 1)1.5
2(𝑡0 + 1)
2(𝑡 + 1)1.5
(37)
For computations of ℎ𝐴𝐵 (𝑡, 𝑡0 ) and ℎ𝐵𝐴 (𝑡, 𝑡0 ), first consider the integrands in Eqs. (10a), (10b)
and (11b).
ℎ𝐵 (𝑡, 𝜏)ℎ𝐴 (𝜏, 𝑡0 ) =
(𝜏 + 1)0.5 −𝑡+𝜏 1
1
𝑒
𝑒 −𝜏+𝑡0 =
𝑒 −𝑡+𝑡0 , (38a)
1.5
1.5
2(𝑡 + 1)
𝜏+1
2(𝑡 + 1) (𝜏 + 1)0.5
(𝑡 +1)0.5
1
(𝑡 +1)0.5
0
0
ℎ𝐴 (𝑡, 𝜏)ℎ𝐵 (𝜏, 𝑡0 ) = 𝑡+1 𝑒 −𝑡+𝜏 2(𝜏+1)
𝑒 −𝜏+𝑡0 = 2(𝑡+1)(𝜏+1)
𝑒 −𝑡+𝑡0 .
1.5
1.5
(38b)
Obviously, these integrands are not identical and Eq. (30) in connection to Remark 1 is not
satisfied for the general values of 𝑡0 ≤ 𝜏 ≤ 𝑡 ∈ 𝑅. The impulse response of 𝐴𝐵 and 𝐵𝐴 are
computed as in (10a) and (10b), that is the integral of the integrands in (38a) and (38b),
respectively. The results are as follows:
𝑡
𝑡
ℎ𝐴𝐵 (𝑡, 𝑡0 ) = ∫ ℎ𝐵 (𝑡, 𝜏)ℎ𝐴 (𝜏, 𝑡0 )𝑑𝜏 = ∫
𝑡0
𝑡0
𝑒 −𝑡+𝑡0
1
𝑑𝜏
1.5
2(𝑡 + 1) (𝜏 + 1)0.5
𝑒 −𝑡+𝑡0 (𝜏 + 1)0.5 𝑡
=
|
2(𝑡 + 1)1.5
0.5
𝑡0
=
𝑒 −𝑡+𝑡0
𝑒 −𝑡+𝑡0
𝑡0 + 1 0.5
0.5
0.5 ]
[(𝑡
(𝑡
+
1)
−
+
1)
=
[1
−
(
) ],
0
(𝑡 + 1)1.5
𝑡+1
𝑡+1
𝑡
ℎ𝐵𝐴 (𝑡, 𝑡0 ) = ∫
𝑡0
=
(39a)
(𝑡0 + 1)0.5 𝑒 −𝑡+𝑡0
(𝑡0 + 1)0.5 𝑒 −𝑡+𝑡0 (𝜏 + 1)−0.5 𝑡
1
𝑑𝜏
=
|
(𝜏 + 1)1.5
2(𝑡 + 1)
2(𝑡 + 1)
−0.5
𝑡0
(𝑡0 +1)0.5 𝑒 −𝑡+𝑡0
(𝑡+1)
[(𝑡
0
1
+1)0.5
1
− (𝑡+1)0.5 ] =
𝑒 −𝑡+𝑡0
𝑡+1
𝑡 +1 0.5
0
[1 − ( 𝑡+1
)
].
(39b)
Obviously, ℎ𝐴𝐵 (𝑡, 𝑡0 ) ≡ ℎ𝐵𝐴 (𝑡, 𝑡0 ).
When System 𝐶 is obtained from 𝐵 by (2b) with ℓ1 = −0.5, ℓ0 = 3.5, its coefficients are found
as 𝑐1 = −(𝑡 + 1), 𝑐0 = −(2𝑡 + 5)(−0.5) + 3.5 = −𝑡 + 1. Then, 𝑓(𝑡) for 𝐶 is computed from
(23b) as
𝑓𝑐 (𝑡) = ∫ −
−𝑡 + 1
𝑑𝑡 = −𝑡 + 2ln(𝑡 + 1).
−𝑡 − 1
Using (23a), the impulse response of 𝐶 is now obtained as
16
(40a)
ℎ𝐶 (𝑡, 𝑡0 ) =
𝑒 −𝑡+2 ln(𝑡+1)+𝑡0 −2 ln(𝑡0 +1)
1
𝑡+1 2
=−
𝑒 −𝑡+𝑡0 (
)
−(𝑡0 + 1)
𝑡0 + 1
𝑡0 + 1
=−
(𝑡 + 1)2 −𝑡+𝑡
0.
𝑒
(𝑡0 + 1)3
(40b)
The impulse response of the connection 𝐴𝐶 is computed by Eq. (10a) modified by replacing
𝐵 by 𝐶, and using Eqs. (34) and (40b) for 𝐴 and 𝐶; we obtain
𝑡
𝑡
ℎ𝐴𝐶 (𝑡, 𝑡0 ) = ∫ ℎ𝐶 (𝑡, 𝜏)ℎ𝐴 (𝜏, 𝑡0 )𝑑𝜏 = ∫ −
𝑡0
𝑡0
𝑡
= −(𝑡 + 1)2 𝑒 −𝑡+𝑡0 ∫
𝑡0
=
(𝑡 + 1)2 −𝑡+𝜏
1
𝑒
𝑒 −𝜏+𝑡0 𝑑𝜏
3
(𝜏 + 1)
𝜏+1
(𝑡 + 1)2 𝑒 −𝑡+𝑡0
1
1
𝑡
𝑑𝜏
=
|
4
3
(𝜏 + 1)
(𝜏 + 1) 𝑡0
3
(𝑡 + 1)2 𝑒 −𝑡+𝑡0
1
1
[
−
]
(𝑡 + 1)3 (𝑡0 + 1)3
3
𝑒 −𝑡+𝑡0
𝑡+1 3
=
[1 − (
) ].
3(𝑡 + 1)
𝑡0 + 1
(41a)
Note that the impulse response of the connection 𝐶𝐴 can be obtained similarly by using Eq.
(10a) by replacing 𝐵 by 𝐶 and using Eqs. (40b) and (34) for 𝐶 and 𝐴; the result is
𝑡
𝑡
ℎ𝐶𝐴 (𝑡, 𝑡0 ) = ∫ ℎ𝐴 (𝑡, 𝜏)ℎ𝐶 (𝜏, 𝑡0 )𝑑𝜏 = − ∫
𝑡0
𝑡0
1
(𝜏 + 1)2 −𝜏+𝑡
−𝑡+𝜏
0 𝑑𝜏
𝑒
𝑒
(𝑡0 + 1)3
𝑡+1
𝑡
𝑒 −𝑡+𝑡0
𝑒 −𝑡+𝑡0
𝑡
2
(𝜏 + 1)3 |
=−
∫(𝜏
+
1)
𝑑𝜏
=
−
3
3
(𝑡 + 1)(𝑡0 + 1)
3(𝑡 + 1)(𝑡0 + 1)
𝑡0
𝑡0
=
𝑒 −𝑡+𝑡0
[(𝑡 + 1)3 − (𝑡 + 1)3 ]
3(𝑡 + 1)(𝑡0 + 1)3 0
=
𝑒 −𝑡+𝑡0
𝑡+1 3
[1 − (
) ].
3(𝑡 + 1)
𝑡0 + 1
(41b)
This is the same impulse response expressed in (41a). Hence, it is verified that (𝐴, 𝐶) is a
commutative pair.
7.
Transitivity for Some 1st and 2nd Order Relaxed Systems
17
Let 𝐴, 𝐵, 𝐶 be linear time-varying systems of orders 2, 1, 1; respectively. Let they be
represented by the general form of (1) with coefficients 𝑎2 , 𝑎1 , 𝑎0 ; 𝑏1 , 𝑏0 ; 𝑐1 , 𝑐0 ; inputs by
𝑥𝐴 , 𝑥𝐵 , 𝑥𝐶 ; output by 𝑦𝐴, 𝑦𝐵 , 𝑦𝐶 . It is true that 𝐴 has 2nd or lower order commutative pairs 𝐵
if and only if [8-10]
−𝑎20.5
𝑑
4𝑎12 + 3𝑎̇ 22 − 8𝑎1 𝑎̇ 2 + 8𝑎̇ 1 𝑎2 − 4𝑎2 𝑎̈ 2
[𝑎0 −
] 𝑘1 = 0,
𝑑𝑡
16𝑎2
𝑎2
𝑏2
𝑎
[ 𝑏1 ] = 1
𝑏0
[𝑎0
0
0
𝑘2
0.5
𝑎2
0
[ 𝑘1 ],
𝑎2−0.5 (2𝑎1 − 𝑎̇ 2 )
1] 𝑘0
4
(42a)
(42b)
where 𝑘2 , 𝑘1 , 𝑘0 are arbitrary constants. We assume that 𝐵 is a finite order commutative pair
of 𝐴; hence 𝑘2 = 0 and
𝑎20.5
0
𝑏1
𝑘
−0.5
[ ] = [𝑎2 (2𝑎1 − 𝑎̇ 2 )
] [ 1 ].
𝑏0
𝑘
0
1
4
(43)
On the other hand, it is true that 𝐵 has 1st or lower order commutative pairs 𝐶, if [8]
𝑐1
𝑏
[𝑐 ]=[ 1
𝑏0
0
0 ℓ1
] [ ],
1 ℓ0
(44)
where ℓ1 and ℓ0 are arbitrary constants. Inserting the values of 𝑏1 and 𝑏0 from Eq. (43),) into
Eq. (44), and rearranging, we obtain
𝑎20.5
𝑐1
[𝑐 ] = [𝑎2−0.5 (2𝑎1 − 𝑎̇ 2 )
0
4
0
𝑘1 ℓ1
][
].
1 𝑘0 ℓ1 + ℓ0
(45)
It is obviously true that this equation is in the form of (43). So 𝐶 is among the first-order
commutative pairs of 𝐴. Hence, it has been proven that if (𝐴, 𝐵) and (𝐵, 𝐶) are commutative
pairs so is (𝐴, 𝐶). Hence, transitivity is valid.
Let’s get one step up and consider transitivity property between two second order and one firstorder systems. Let 𝐴, 𝐵, 𝐶 be 2nd, 1st, 2nd order systems, respectively, with coefficient
𝑎2 , 𝑎1 , 𝑎0 ; 𝑏1 , 𝑏0 , 𝑐2 , 𝑐1 , 𝑐0 . Assume (𝐴, 𝐵) and (𝐵, 𝐶) are commutative pairs and show (𝐴, 𝐶)
is also commutative. Eq. (43) is valid since (𝐴, 𝐵) is a commutative pair. Similar equation
18
𝑐20.5
0
𝑏1
ℓ
−0.5
[ ] = [𝑐2 (2𝑐1 − 𝑐̇2 )
] [ 1]
𝑏0
ℓ
1 0
4
(46)
is also valid since (𝐵, 𝐶) is a commutative pair. Further, (42a) is valid and the similar equation
for 𝐶 is written as
−𝑐20.5
𝑑
4𝑐12 + 3𝑐̇22 − 8𝑐1 𝑐̇2 + 8𝑐̇1 𝑐2 − 4𝑐2 𝑐̈2
[𝑐 −
] ℓ1 = 0.
𝑑𝑡 0
16𝑐2
(47)
Since 𝐵 is of order 1, 𝑘1 in (43) and ℓ1 in (46) are not zero; hence (42a) and (47) yield
𝑎0 = [𝐴0 +
4𝑎12 + 3𝑎̇ 22 − 8𝑎1 𝑎̇ 2 + 8𝑎̇ 1 𝑎2 − 4𝑎2 𝑎̈ 2
],
16𝑎2
(48a)
4𝑐12 + 3𝑐̇22 − 8𝑐1 𝑐̇2 + 8𝑐̇1 𝑐2 − 4𝑐2 𝑐̈2
],
16𝑐2
(48b)
𝑐0 = [𝐶0 +
respectively, where 𝐴0 and 𝐶0 are arbitrary constants.
From Eq. (43), the first line, we obtain
2(𝑏̇12 + 𝑏1 𝑏̈1 )
𝑎̈ 2 =
.
𝑘12
𝑏12
2𝑏1 𝑏̇1
𝑎2 = 2 → 𝑎̇2 =
,
𝑘1
𝑘12
(49a)
Using these results in Eq. (43), the second line, we obtain
𝑎1 =
𝑏1 (2𝑏0 − 2𝑘0 + 𝑏̇1 )
𝑏̇1 (2𝑏0 − 2𝑘0 + 𝑏̇1 ) + 𝑏1 (2𝑏̇0 + 𝑏̈1 )
→
𝑎̇
=
.
1
𝑘12
𝑘12
(49b)
Inserting the values of 𝑎2 , 𝑎̇ 2 , 𝑎̈ 2 , 𝑎1 , 𝑎̇ 1 in Eq. (49) into Eq. (48a) and simplifying, we obtain
𝑏0 − 𝑘0 2 𝑏1 𝑏̇0
𝑎0 = 𝐴0 + [
] + 2 .
𝑘1
𝑘1
(49c)
Similar results can be obtained starting from (45) and using (48b):
𝑐2 =
𝑐1 =
𝑏12
,
ℓ12
𝑏1 (2𝑏0 −2ℓ0 +𝑏̇1 )
ℓ21
19
(50a)
,
(50b)
𝑐0 = 𝐶0 + [
𝑏0 − ℓ0 2 𝑏1 𝑏̇0
] + 2 .
ℓ1
ℓ1
(50c)
To prove the commutativity of 𝐴 and 𝐶, it is sufficient to show that
𝑎2
𝑐2
𝑎
[ 𝑐1 ] = 1
𝑐0
[𝑎 0
0
0
𝑚2
0.5
𝑎2
0
[ 𝑚1 ],
𝑎2−0.5 (2𝑎1 − 𝑎̇ 2 )
1 ] 𝑚0
4
(51)
which is similar equation to (42b), where 𝑚2 , 𝑚1 , 𝑚0 are some constants. It is true that, the
coefficients of 𝐶 can be expressed as in (51) by proper choice of constants 𝑚2 , 𝑚1 , 𝑚0 ; in fact,
inserting in coefficients 𝑎𝑖 in Eq. (49) and coefficient 𝑐𝑖 in Eq. (50) into Eq. (51), we find that
𝑘12
,
ℓ12
(52a)
2𝑘1 (𝑘0 − ℓ0 )
,
ℓ12
(52b)
𝑚2 =
𝑚1 =
𝑚0 = 𝐶0 − 𝐴0
𝑘12
ℓ21
+[
𝑘0 −ℓ0 2
ℓ1
] ,
(52c)
which are all constants. Hence, Eq. (51) together with Eq. (42) imply that 𝐶 is a second-order
commutative pair of 𝐴, that is 𝐴 and 𝐶 are also commutative. This proves the transitivity
property of (𝐴, 𝐵) and (𝐵, 𝐶) to (𝐴, 𝐶).
We express the result by a theorem:
Theorem 3: The commutativity property between three subsystems of which at most 2 of them
second order and the other(s) are first-order always satisfy the transitivity property under zero
initial conditions.
8.
Conclusions
Transitivity property of commutativity is defined for commutative linear time-varying systems
of any order. The transitivity results are presented for first-order systems with and without
initial conditions. The study is carried on two approaches; one is based on use of impulse
response function and the other depends on use of general conditions set by the author and
others for commutativity of linear time-varying systems.
20
One illustrative example is included to show the validity of the results. However, general
transitivity property of commutative high order systems have not been proved or disproved
and this remains as an unsolved problem yet.
REFERENCES
[1]
R. Boylestad and L. Nashelsky, Electronic Devices and Circuit Theory, Prentice Hall,
New Jersey, 2013.
[2]
E. Marshal, Commutativity of time varying systems, Electronics Letters, 13 (1977)
539-540.
[3]
M. Koksal, Commutativity of second order time-varying systems, International Journal
of Control, 36 (1982) 541-544.
[4]
S. V. Salehi, Comments on ‘Commutativity of second-order time-varying systems’,
International Journal of Control, 37 (1983) 1195-1196.
[5]
M. Koksal, Corrections on ‘Commutativity of second-order time-varying systems’,
International Journal of Control, 38 (1983) 273-274.
[6]
M. Koksal, General conditions for the commutativity of time-varying systems,
IASTED International Conference on Telecommunication and Control (TELCONi84),
1984, pp. 223-225.
[7]
M. Koksal, A Survey on the Commutativity of time-varying systems, METU,
Gaziantep Engineering Faculty, Technical. Report no: GEEE CAS-85/1, 1985.
[8]
M. Koksal, An exhaustive study on the commutativity of time-varying systems,
International Journal of Control, 47 (1988) 1521-1537.
[9]
M. Koksal and M. E. Koksal, Commutativity of linear time-varying differential
systems with non-zero initial conditions: A review and some new extensions,
Mathematical Problems in Engineering, 2011 (2011) 1-25.
[10]
M. E. Koksal and M. Koksal, Commutativity of cascade connected discrete time linear
time-varying systems, 2013 Automatic Control National Meeting TOK’2013, (2013)
p.1128-1131.
[11]
M. Koksal and M. E. Koksal, Commutativity of cascade connected discrete-time linear
time-varying systems, Transactions of the Institute of Measurement and Control, 37
(2015) 615-622.
[12]
M. E. Koksal, The Second order commutative pairs of a first-order linear time-varying
system, Applied Mathematics and Information Sciences, 9 (2015) 1-6.
[13]
M. E. Koksal, Decomposition of a second-order linear time-varying differential system
21
as the series connection of two first-order commutative pairs, Open Mathematics, 14
(2016) 693-704.
[14]
C. A. Desoer, Notes For A Second Course On Linear Systems, Van Nostrand
Rheinhold, New York, 1970.
[15]
Chi-Tsong Chen, Linear System Theory and Design, New York Oxford, Oxford
University Press, 1999.
22
| 3cs.SY
|
A Beginner’s Guide to Systems Simulation in
Immunology
Grazziela P. Figueredo, Peer-Olaf Siebers, Uwe Aickelin, and Stephanie Foan
arXiv:1307.1597v2 [cs.CE] 8 Jul 2013
Intelligent Modelling and Analysis Research Group, School of Computer Science,
The University of Nottingham, NG8 1BB, UK
gzf, uxa, pos, [email protected]
Abstract.
Some common systems modelling and simulation approaches for immune
problems are Monte Carlo simulations, system dynamics, discrete-event
simulation and agent-based simulation. These methods, however, are still
not widely adopted in immunology research. In addition, to our knowledge, there is few research on the processes for the development of simulation models for the immune system. Hence, for this work, we have two
contributions to knowledge. The first one is to show the importance of
systems simulation to help immunological research and to draw the attention of simulation developers to this research field. The second contribution is the introduction of a quick guide containing the main steps for
modelling and simulation in immunology, together with challenges that
occur during the model development. Further, this paper introduces an
example of a simulation problem, where we test our guidelines.
1
Introduction
Some important advances in immunology were facilitated by the joint work of
immunologists and mathematicians [1]. Many concepts existing in theoretical
immunology are the result of mathematical models. Inumerous existing models
in immunology are based on sets of ordinary differential equations (ODEs) [2, 3].
This approach for immunology, however, in practice limits the modelling effort
to simpler dynamics involving fewer immune elements such as cells or molecules
and it only allows analysis at an aggregate level. Moreover, it is not trivial
to model problems involving individual localisation, memory of past events (or
states) and emerging properties mathematically [1]. Hence, systems simulation
emerged as a complement of mathematics that allows to overcome some of these
limitations. Moreover, systems simulation modelling methods are closer to the
natural description of the system, without the need of an in depth understanding
of mathematics [4].
Sauro et al. [5] debates the usefulness of simulation in contrast to reductionism in biology. In reductionism, the dynamics of complex systems can be understood from studying the properties of their parts [6]. In contrast to reductionism,
in holism (in its methodological version) the properties of the parts contribute to
our understanding of the whole, but the properties can only be fully understood
through the dynamics of the whole [6]. Systems simulation is, therefore, based
on holism. For biology, Sauro et al. [5] state that “reductionism has proven to be
a highly successful strategy and has enabled us to uncover the molecular details of
biological systems in unprecedented detail”. The success of reductionist methods
raised some scepticism as to the need for alternative approaches, such as systems
biology. The challenge for simulation is, therefore, to generate novel insights that
cannot be uncovered just by looking at a phenomena using reductionism. Examples of successful simulation approaches that helped advance immunological
research were introduced in [7]. The models reviewed simulate interactions of immune cells and chemical substances, humoral responses and drug testing. With
these simulations it is possible to observe emergent behaviour in the systems,
which is not feasible with reductionism.
As a first objective of this study, therefore, we want to show that there is a
distinct place for simulation in the tool set that aids advances in immunology.
Moreover we want to show that there is a wide range of problems in immunology
to be tackled by computer scientists and simulation developers.
As there few examples on the methodology for constructing immune systems
simulations [8–10], the second objective is to introduce general guidances for conducting simulation studies in immunology and outline the challenges that might
be encountered during the development of a simulation model. These guidances
adapted from the work developed by Robinson [11] for operational research, considering the characteristics observed in immune simulations. We complement the
current methodologies for constructing immune systems simulations by presenting a framework containing a life-cycle with the main steps to be followed by
any developer, independent of the simulation modelling method chosen.
The remainder of the paper is organized as follows. In Section 2 we review the
main characteristics of the simulation methods used in immunology. In Section 3
we present the main steps and propose a life-cycle for conducting a simulation
study for immune problems. In Section 4 we present a case study where we develop a system dynamics simulation model from the steps outlined. In Section 5,
we present our conclusions and future work.
2
Simulation Methods
The choice of a modelling technique for a problem is driven by the resources available such as experimental data, an understanding of the mechanisms involved,
the hypothesis to be tested and the level of abstraction needed to test the hypothesis. Once the conceptual model is defined, a simulation method needs to
be chosen. There is a wide spectrum of simulation methods used in immunology.
These methods are classified as static or dynamic, stochastic or deterministic,
continuous or discrete. These methods model the system using either top-down
or bottom-up perspectives. Static models help understand connections between
system components, without explicitly representing time [12]. Dynamic models aid in understanding dynamic implications and consequences over time of
a system structure [13]. Deterministic models do not contain any probabilistic
components. Stochastic models, on the other hand, consider random components [14]. The characteristic of being continuous or discrete determine whether
time and variables of the model change continuously or discretely. For example
the age of an individual changes continuously in time, whereas the number of
immune cells that die with age is a discrete value. Top-down approaches focus on
the system at an aggregate level, while bottom-up approaches split the system
in individual parts that will interact giving rise to the behaviour of the system
as a whole [15].
Sauro et al. [5] mentions that the construction of models of biochemical and
cellular behaviour has been traditionally carried out through a bottom-up approach, which combines laboratory data and knowledge of a reaction network to
produce a dynamic model. This process, however, requires the reaction network
to be known and the possibility to carry out the various laboratory experiments.
Furthermore, the modelling relies on the fact that data from laboratory experimentation matches real-world phenomena, which is not always correct. Samples
can be compromised during collection or during the experimentation process. In
addition, although bottom-up approaches are very useful for immunology, there
are circumstances where they can not be applied. Examples include when the
reaction network process is not well understood, or laboratory experiments are
known not to be able to reproduce the real-world reactions (for instance, given
the environmental differences such as temperature). In addition, the authors
argue that “top-down modelling strategies are closer to the spirit of systems biology exactly because they make use of systems-level data, rather than having
originated from a more reductionist approach of molecular purification”. The
conclusion reached in their study was that there is no best approach as it is
preferable to view them as complementary. Their ideas match other studies in
biology and other research areas, which investigate the merits of each approach
and their combination for simulation [16]. To our knowledge, the most common system simulation approaches for immunology are Monte Carlo simulation,
system dynamics, discrete-event simulation, cellular automata and agent-based
simulation.
Monte Carlo simulations [17] are largely used in molecular theoretical immunology. These techniques generate random numbers and observe that fraction of the numbers that obey a certain property (of properties). These methods
are suitable for obtaining numerical solutions for problems as an alternative to
their analytical solution. The disadvantage of these methods are that they do
not provide information of how the elements of the system change during the
simulation (dynamics of the system). Instead, they focus on the determination
of the system outcome given a certain input, not taking time into account.
System dynamics (SD) is a top-down modelling technique that uses stocks,
flows and feedback loops as concepts to model the behaviour of complex systems
in a stock and flow diagram. It is an aspect of systems theory that is initially
applied in order to understand complex aggregate behaviours in industry [18].
Currently, SD is applied to any complex system characterized by interdependency, mutual interaction, information feedback and circular causality. System
dynamics simulation (SDS) is a continuous simulation for an SD model. It consists of a set of ODEs that are solved for a certain time interval. These ODEs,
however, are implicit in the system’s structure and the relationships between the
elements modelled can be established with experimental data. Some examples
of SD applied to the immune system are found in [19, 20].
Discrete-event simulation is also a top-down approach that models a system
as a set of entities being processed and evolving over time according to the
availability of resources and the triggering of events. The simulator maintains an
ordered queue of events [21]. Each event occurs at an instant in time and marks
a change of state in the system. It is process-oriented and the entities involved
are passive [11], with no pro-activity. The entities are individually represented
and can be tracked throughout the system simulation. The models are stochastic
and outputs usually represent average values [22]. Examples of applications for
the immune system are found in [23–25].
Cellular automata is a discrete model consisting of two main components. The
first component is an infinite regular grid of cells, which constitutes the universe
or space of the cellular automata. In computer simulations, however, due to
space limitations, the cellular automata space is predetermined and finite. The
second component is a finite automaton (or cell). Each cell from the grid contains
a finite number of states and a predefined set of cells called neighbourhood.
The communication of a cell with other cells within its neighbourhood is local,
deterministic, uniform and synchronous [26]. Each cell is initialized with an initial
state at time t = 0. As time advances, the cells are updated according to a fixed
rule, which is, in general, a mathematical function. This rule defines the next
state of each cell according to its current state and its neighbourhood states.
Examples of several applications for the immune system are found in [7].
Agent-based simulation is a technique that employs autonomous agents that
interact with each other [27]. The agents’ behaviour is described by rules that determine how they learn, interact with each other and adapt. The overall system
behaviour arises from the agents’ individual dynamics and their interactions [21].
For immunology, it can amalgamate in vitro data on individual interactions between cells and molecules of the immune system to build an impression of the
system as a whole. Cellular automata and agent-based simulation have some
similarities such as individual rules and interactions between the individuals entities. Moreover, both are bottom-up approaches capable of representing emergent behaviour in the system. The entities in cellular automata, however, do not
have memory and only interact with individuals from the predefined neighbourhood, as their location does not change. In agent-based simulation, on the other
hand, the agents are individual entities with memory and are capable to interact
with any other agent in the system. Several examples of agent-based models in
immunology are found in the european virtual human immune system project
(ImmunoGrid) [28].
By reviewing the simulation methods and some of their applications to immune problems, such as those from [7, 23–25, 29], it is possible to outline the
benefits of simulation to immunology and, therefore, achieve our first objec-
tive of this work. Compared with real-world experimentation, simulation is time
and cost-effective. Most laboratory experiments are expensive and have to be in
agreement with ethical specifications. Furthermore, in a simulation environment,
it is possible to systematically generate different scenarios, conduct and replicate
experiments.
In the following section we fulfil our second research goal of this paper, which
is to introduce a descriptive guide for the development of a simulation model in
immunology and the challenges that might be encountered during this process.
These guidances are kept general and can be applied to all simulation methods
in immunology.
3
Steps in a Simulation Study
As there is not much done to established guidance to develop simulations for
the immune system, we studied those developed by [11] for simulation in operational research problems and adapted them for simulations of the immune
system. The adaptation was performed by studying several simulations under
different approaches developed for the immune system, as mentioned in the previous section. In order to adapt the guidelines developed by [11], we observed the
similarities and differences with operational research and outlined general steps
for building immune simulations. These steps represent a life-cycle of a simulation and therefore the method we present is iterative. Furthermore, we discuss
the pitfalls that might be encountered during the process, as shown below. In
some of the steps, extra efforts specific to immunology were not added as we
believe they are generic for any type of simulation.
1. Define the Objectives. Overall, the objectives are either to investigate
a theory or propose a “what-if” scenario with no concern to ethics restrictions.
The scenario proposed can either be based on experimental data or defined as
an intuition of what might happen in reality. Furthermore, there are also cases
where actual models do not match real-world experimentation and they need to
be further investigated (in a simulation model). In addition, new hypotheses and
research questions may be defined together with immunologists as simulation
goals. The objectives come from real-world observation. We assume, however,
that real-world observation and experimentation has been previously performed
by immunologists.
2. Describe the system. In this step, it is necessary to use documents
(immunology books and articles, transcripts of interviews with experts, etc.)
describing how the immune elements to be simulated work and interact. The
description of the system is based on knowledge acquired by theoretical work,
real-world observation and laboratory experimentation. Due to the complexity
of the elements and processes in the immune system, however, this knowledge
is scarce. The immune system is far from being fully fathomable, and the descriptions found in literature are only partial representations and assumptions
of what occurs in reality.
3. Investigate existing theories and established models. In order to
build a new simulation model, it is common to look at the existing models
and investigate their hypotheses, objectives, validation process and limitations.
For some cases, these established models have somehow been validated against
experimental data. With this practice it is possible to build a new model as an
improvement of what has already been established in order to further investigate
a certain immune process.
4. Use experimental data. Currently, most simulation models are built
based on real-world observation and experimentation. There are some models,
however, where there is no data available (for example, when Jerne’s network
theory was conceptualized [30]). These models are based purely on theoretical
assumptions with the purpose of providing more insights about what happens
in the real world. Furthermore, in the field of immunology, the non-existence of
data can be due to the lack of understanding of a process, or a difficulty or even
impossibility in collecting information with current technology. In other cases, a
hypothesis is first formulated requiring experimental data to confirm it. There is,
therefore, the need to collect this data. For instance, Foan et al. [20] implemented
a system dynamics simulation of T cell subsets throughout a person’s lifetime
based on an established mathematical model developed by Balcheva [31]. The
authors conclude that further validation of this model is necessary and so a
novel data set should be collected as there are arguably more specific markers
that could help to gain further insights from the model.
5. Build conceptual model. The conceptual model of a problem is an
abstraction intended to contain the principal aspects observed in the real world,
considering the necessary level of details [32]. In this step we formally define the
model scope, the objectives previously outlined, the inputs and outputs and the
simplifications. The process of creating a conceptual model evolves with decisions
regarding the model scope and level of detail [11]. The acceptance of the conceptual model should be agreed with immunologists. According to Ulgen et al. [33],
“rigorous validation procedure for the conceptual model is as important as the
verification and validation of the model because it saves time and redirects the
simulation developers in the right direction before time is wasted in the study”.
Due to the limitations of a immune simulation, it is important to abstract the
relevant real-world features and build a simple model. According to Kotiadis and
Robinson [32], the importance of model abstraction “relies on the fact that there
is no need to model all that is known about the real problem. Simpler models are
developed and run faster, they are flexible, require less data and results are easier to be interpreted”. The nature of the immune problems thus implies that the
model should be developed in order to address a few objectives, within a limited
scope. The description of the system (and definition of the conceptual model)
should therefore focus on the parts of the immune system (scope, elements, information available, assumptions, hypotheses) relevant to achieve the simulation
goals. Daigle discusses the challenges of modelling immunology [34]. As it is a
field in which information is still being gathered, simulations have to be updated
frequently to suit new findings. Moreover, current computational resources and
modelling techniques are in development. It is still thereby impossible to represent computationally an entire pool of cells of a typical immune response (around
1012 cells). In addition, immunological systems are mostly hierarchical, involving
several layers and complex interactions between the elements of these layers.
6. Identify elements, parameters, aggregates, etc. already established in theory and real-world data. The study of the conceptual model
provides a means to understand the problem and the best way to represent the
elements of the system. This stage influences the choice of the most appropriate
simulation approach. For example, if the in the conceptual model it is established
the interactions of the simulation will occur at a cellular population level rather
than an individual cell level, this might indicate that a top-down simulation
approach would be more suitable to build the model.
7. Decide on the most appropriate simulation approach. This decision
is made based on the characteristics of the problems, the research questions to be
addressed, the scope, the level of aggregation and the experimental data available. Some of the most common approaches used in immunology are agent-based
modelling and simulation, discrete-event modelling and simulation, cellular automata and system dynamics. Cellular automata is used for problems involving
autonomous individual interactions within a neighbourhood placed in a lattice
and emergent behaviour. Agent-based simulation is suitable for problems involving autonomous individual behaviour, elements spacial localization, memory and
emergence. Discrete-event simulation tackles problems that are process-oriented,
which have passive individual entities and chronological sequence of events. Furthermore, each event occurs at an instant in time and marks a change of state in
the system. It can be used for any experiment where there is no need for continuous time. SD defines a system at a high level of aggregation and, therefore, it
should be used when the research question involves patterns of behaviours and
feed-back interactions between the aggregates. This approach is very useful to
simulate dynamics of populations and interactions between different populations
overtime. For example, interactions between tumours and populations of effector
cells, populations of viruses and T cells, etc.
8. Represent elements, parameters, etc. using the appropriate simulation approach. Once the simulation approach is chosen, the elements defined in the conceptual model need to be translated into their correspondent
implementation used by each approach, for instance, stocks, flows, parameters
and information for SD or agents and rules for ABMS. This step is part of the
construction of the simulation model, defined in the next step.
9. Build the simulation model. This stage includes the development of
the computational implementation of the model in a simulation tool. The implementation is a software representation of the requisites defined in the conceptual
model. The computational model is the final product to be used by the immunologists.
10. Verify the model. The model verification is the process of ensuring that
the model design has been transformed into a computer model with sufficient
accuracy [11].
11. Validate the model with existing theories and, if available, realworld data. Validation ensures that the model is sufficiently accurate for the
purpose at hand. For immunology it is acknowledged that models are not intended to be completely accurate for a number of reasons: (1) there is no real
world data to compare against, (2) there is little data, (3) real-world data is
inaccurate, (4) even if the data is accurate, the real world data is only a sample,
which in itself creates inaccuracy. Verification and validation are continuous and
iterative processes performed throughout the life cycle of a simulation study [11].
12. Experimental design. The experimental design improves the experimentation process by the definition of sound experiments with trustworthy results. In this stage, the experimental factors that are most likely to lead to significant improvements are identified. This process is developed using data analysis,
expert knowledge, preliminary experimentation and sensitivity analysis.
13. Experimentation. Experimentation is conducted following the experimental design guidelines. It can make use of multiple simulation replications;
interactive experimentation (observing the simulation and making changes to
the model to see the effects); batch experimentation (setting experimental factors and leaving the models to run for a pre-defined run length); comparing
alternatives (where there is a limited number of scenarios to be compared) and
search experimentation (when there is no predefined number of scenarios).
14. Result Analysis. Plots and statistics are collected during the simulation. The result analysis is the process that interprets results and the best way
to present them.
15. Report Findings. After results are interpreted, there is the need to
report the findings from the simulations. For immunology, it can be new insights,
verification of a theory, etc.
16. Validate and add more requisites with immunologists. Building an immune simulation is an iterative process. Generally the model is built
together with immunologists, and, in every step of the framework, the model
elements should be verified with them.
The process of simulation is iterative, as shown in Figure 1. During the model
development, additional data might become available, which changes the system description/objectives and impacts on every step of the process. Moreover,
as validation occurs throughout the whole process, if any of the stages is not
validated (data available, real world understanding and description, conceptual
model, computer model, experimental design, etc), there is the need to go back
and rethink the invalid state, which impacts on the subsequent steps.
4
Case Study
The objective of our case study is to exemplify how the steps defined previously
can be employed to help in the development of a simulation for the immune
system. The problem to be investigated is how the population of peripheral naive
T cells evolve during the course of an individual’s lifetime. The development of
the simulation model is depicted below.
Fig. 1. Process of simulation study: the life cycle.
1. Define Objectives. The simulation goal is concerned with establishing
an understanding of the peripheral naive T cell repertoire dynamics over time.
2. Describe the System. In the human body there is a type of white blood
cell, namely the naive T cell, which plays an important role in the immune system
by responding to new infections in the organism. Before an individual reaches
the age of 20, the set of naive T cells is sustained primarily from thymic output.
In middle age, however, there is a change in the source of naive T cells: as the
thymus involutes, there is a considerable shrinkage in its T cell output, which
means that new T cells are mostly produced by peripheral expansion. There is
also a belief that some memory T cells have their phenotype reverted back to
the naive proliferation cells type [35]. Furthermore, memory cells are originated
from active T cells.
3. Investigate existing theories and models. Immunologists found out
that thymic contribution in an individual are quantified by the level of a biological marker called ‘T cell receptors excision circle’ (TREC). TREC is circular
DNA originated during the formation of the T-cell receptor. The percentage of T
cells possessing TRECs decays with shrinkage of thymic output, activation and
reproduction of naive T cells [35]. This means that naive T cells originating from
the thymus have a greater percentage of TREC than those originating through
other proliferation and with time there is a depletion of naive T cells from thymus in the organism. There is an existing model proposed by Murray [35] that
investigates the thymic output and decay of these cells mathematically with the
use of an ODE system.
4. Experimental data. TREC data collected by immunologists is presented
by Murray et al. [35], which also develops an ODE system model for the dynamics
of peripheral naive T cells. Furthermore, the authors provided us with data on
active cells and total naive T cells in individuals with age ranging from 1 to 55
years. If we assume that this data has been validated by the immunologists and
expresses what occurs in reality, we can use this information for the continuation
of our investigations on naive cells from peripheral proliferation.
5. Conceptual model. The population dynamics of the model is shown
in Figure 2. We have four main populations of cells: naive from thymus (in
the figure naive), naive from proliferation, active cells and memory cells. Naive
and memory cells are sources of naive cells from proliferation. Active cells are
sources of memory cells. The scope of the simulation, therefore, is limited to the
dynamics of these four populations. The number of naive from thymus and active
cells are given by real-world data collected by immunologists. The conceptual
model and data used are the same as the those from the existing model proposed
by Murray [35]. The objective is to determine what are the dynamics of the naive
cells from peripheral proliferation with age under a systems simulation approach
different from the ODEs simulation. Another goal is to determine the rates in
which the naive cells from thymus and memory cells become naive cells from
proliferation; and the rates in which active cells become memory cells.
Fig. 2. Dynamics of Naive T cells.
6 and 7. Identify elements, parameters, etc. and decide on the most
appropriate simulation approach. As the investigations regard populations
dynamics at a high level of aggregation, as defined in the conceptual model, we
decided to build the simulation using the system dynamics approach, where the
aggregates will be each different cell population and the feedback loops are those
represented by the arrows in the conceptual model of figure 2.
8. Represent elements, parameters, etc. using the appropriate simulation approach. As stated before, system dynamics models consist of stock,
flows, information and feedback loops. The stocks of the simulation are the
variables that accumulate over time; flows modify the stocks by adding or sub-
tracting elements. For the simulation, therefore, the stocks will be the number of naive cells from proliferation and memory cells. The stock of naive cells
from proliferation is modified by flows such as addition of new cells from thymus, peripheral proliferation and death, as shown in Figure 2 from the conceptual model. The memory cells stock is changed by the flows: active cells
reversing into memory and memory cells death. Both naive and active cells are
represented in the simulation as look up tables. Moreover, parameters representing rates in which flows modify the stocks need to be incorporated. These
parameters are: N aiveT hymusP rolif erationRate, N aiveP rolif erationRate,
N aiveP rolif erationDeathRate, M emoryT oN P Rate, M emoryDeathRate and
ReversionT oM emoryRate.
9. Build the simulation model. The final SD model implemented is shown
in Figure 3.
Fig. 3. SD built from naive T cells data.
In the figure, the stocks are represented by the box
are represented by the hourglass
, flow
, the flow variables
, parameters
and infor-
mation
. Information indicates that the stock value is used in the flow
calculation. RealNaives and RealActives are look-up tables containing the experimental data.
9. Verify the model. As the model studied is quite simple, we verified
our implementation against the conceptual model. Further verification was performed during the experimental stage.
10. Validate the model with existing theories and, if available, realworld data. As we mentioned, the validation process is performed throughout
all simulation development. We validate our results against the data set provided
for the total number of naive T cells in the organism, as shown in the result
analysis (step 12).
11. Experimental design and experimentation. For this case study we
will run one experiment in which we adjust the parameters to fit the original
data. The simulation was run for a period correspondent to sixty years.
12. Result Analysis and findings. We calibrated the simulation parameters against the data provided and the results obtained are shown in Figure 4(a).
Figure 4(b) shows the results obtained in [35]. For both approaches, the total
number of naive cells for the simulation are different for the first eighteen years
and this difference as well as the parameters obtained need further investigation
with immunologists. After twenty years, the numbers for the real data observed
and the outcomes obtained are very close. Moreover, it is possible to observe how
the population of naive cells from proliferation grows with time and is prevalent
in the total naive cells cohort after fifty years.
Results for the SD model
100
90
80
80
70
60
50
60
50
40
40
30
30
20
20
10
10
0
Total Naive % CD4+ data
Log10 percentage T cells from thymus data
Total Naive %CD4+ from ODE simulation
Naive from proliferation from ODE simulation
90
Naive %
Naive %
70
Results for the ODE model
100
Total Naive % CD4+ data
Log10 percentage T cells from thymus data
Total T cells from simulation
Naive from proliferation from simulation
0
10
20
30
Years
40
50
60
0
0
10
20
30
Years
40
50
60
(a)
(b)
Fig. 4. Simulation results. For the SD simulation, the parameter values used are:
NaiveThymusProliferationRate = 0.025, NaiveProliferationDeathRate = 0.017, MemoryToNPRate = 0.001, ReversionToMemoryRate = 0 and MemoryDeathRate = 0.05.
13. Validate and add more requisites with immunologists. Results
(and parameters calibration) validation needs to be done with immunologists, as
well as new requisites do be added in the simulation model. There is the need to
investigate whether the simulation models are informative and relatively accurate. Further research also needs to be done to explain the outcome differences
for both models.
5
Conclusions
Although there are examples showing the success of simulation aiding advances
in immunology, this set of methodologies is not popular among immunologists.
The overall objective of our research, therefore, is to outline the potential contribution of simulation methods to help immunological studies and invite experts
(simulation developers, computer scientists, etc) to build solutions in this field.
For this paper, we had two research goals. The first goal was to show that there is
a distinct place for simulation in the tool set used by immunologists and present
the most common simulation approaches. As there are no general guidelines for
the development of immune simulations, our second objective was to introduce
our own guidances for conducting simulation studies in immunology and outline
the pitfalls that might be encountered during the development of a simulation
model. We achieved our first objective by arguing that popular methodologies
used such as ODEs and reductionist methods have limitations that are overcome
by simulation. Moreover, we argued that for many simulation methods, the problem representation is closer to the systems natural description. For our second
goal, we studied several simulations using different approaches in the literature
and outlined their common features. This helped us to develop a simulation lifecycle with several steps to be followed. These steps encompass common aspects
to be considered during the development of immune simulations, independent of
the simulation approach adopted. As future work we want to improve the lifecycle introduced and develop a decision framework that further helps with the
choice of a simulation approach according to the problem presented. In addition,
we aim at making our set of guidances more specific according to the simulation
method used.
References
1. Louzoun, Y.: The evolution of mathematical immunology. Imm. Rev. 216 (2007)
9–20
2. Gruber, J. Models of Immune Systems: The Use of Differential Equations Available at: http://www.lymenet.de/literatur/immundif.htm (Last accessed
13 Jun 2011)
3. Eftimie, R., Bramson, J.L., Earn, D.J.: Interactions between the immune system
and cancer: A brief review of non-spatial mathematical models. Bull Math Biol.
(2010)
4. Bonabeau, E.: Agent-based modeling: Methods and techniques for simulating human systems. In: Proc. of the National Academy of Sciences of the United States
of America. Volume 99. (2002) 7280–7287
5. Sauro, H.M., Harel, D., Kwiatkowska, M., Shaffer, C.A., Uhrmacher, A.M., Hucka,
M., Mendes, P., Strömback, L., Tyson, J.J.: Challenges for modeling and simulation
methods in systems biology, Winter Simulation Conference (2006) 1720–1730
6. ∅streng, W.: Reductionism versus Holism Contrasting Approaches? In: Consilience. Interdisciplinary Communications 2005/2006. Centre for Advanced Study,
Oslo (2007) 11–14
7. Fachada, N., Lopes, V., Rosa, A.: Agent-based modelling and simulation of the
immune system: a review. In: EPIA 2007 - 13th Portuguese Conference on Artificial
Intelligence. (2007)
8. Kitano, H.: Computational systems biology. Nature 420 (2002) 206–210
9. Kitano, H.: Systems biology: A brief overview. Science 295 (2002) 1662–1664
10. Andrews, P.S., Polack, F.A.C., Sampson, A.T., Stepney, S., Timmis, J.: The CoSMoS process version 0.1: A process for the modelling and simulation of complex
systems. Technical Report YCS-2010-453, Department of Computer Science, University of York (2010)
11. Robinson, S.: Simulation: The Practice of Model Development and Use. John
Wiley and sons, Ltd (2004)
12. Silva, P.S., Trigo, A., Varajo, J., Pinto, T.: Simulation concepts and applications.
In Lytras, M.D., Ordonez de Pablos, P., Ziderman, A., Roulstone, A., Maurer,
H., Imber, J.B., eds.: Organizational, Business, and Technological Aspects of the
Knowledge Society. Volume 112 of Communications in Computer and Information
Science. Springer Berlin Heidelberg (2010) 429–434
13. Babulak, E., Wang, M.: 1. In: Discrete Event Simulation: State of the Art. InTech
(2010) 1–8
14. Banks, J., Carson, J., Nelson, B., Nicol, D.: Discrete-event system simulation. 4th
edn. Pearson (2005)
15. Macal, C.M.: To agent-based simulation from system dynamics. In: Proc. of the
2010 Winter Simulation Conference. (2010)
16. Schieritz, N., Milling, P.M.: Modeling the forrest or modeling the trees: A comparison of system dynamics and agent based simulation. In: Proc. of the XXI Int.
Conference of the System Dynamics society. (2003)
17. Metropolis, N., Ulam, S.: The Monte Carlo method. J. Amer. Stat. Assoc. 44
(1949) 335–341
18. Forrester, J.W.: Industrial Dynamics. Pegasus Communications (1961)
19. Figueredo, G.P., Aickelin, U., Siebers, P.O.: Systems dynamics or agent-based
modelling for immune simulation? In: Proc. of the Int. Conf. on AIS. (2011)
20. Foan, S.J., Jackson, A.M., Spendlove, I., Aickelin, U.: Simulating the dynamics of
T cell subsets throughout the lifetime. In: Proc. of the Int. Conf. on AIS. (2011)
21. Siebers, P.O., Aickelin, U.: Introduction to Multi-Agent Simulation. In: Encyclopaedia of Decision Making and decision support technologies. (2007) 554–564
22. Tako, A.A., Robinson, S.: Comparing model development in discrete event simulation and system dynamics. In Rossetti, M.D., Hill, R., Dunkin, A., Ingalls, R.G.,
eds.: Proc. of the 2009 winter simulation conference. (2009) 979–990
23. Look, A.T., Schriber, T.J., Nawrocki, J.F., Murphy, W.H.: Computer simulation
of the cellular immune response to malignant lymphoid cells: logic of approach,
model design and laboratory verification. Immunology 43 (1981) 677–690
24. Martin S. Zand, Benjamin J. Briggs, A.B., Vo, T.: Discrete event modeling of
CD4+ memory T cell generation. The Journal of Immunology 173 (2004) 3763–
3772
25. Figge, M.T.: Stochastic discrete event simulation of germinal center reactions.
Phys. Rev. E 71 (2005) 051907
26. Wolfram, S.: Statistical mechanics of cellular automata. Rev. Mod. Phys. 5 (1983)
601–644
27. Macal, C.M., North, M.J.: Tutorial on agent-based modeling and simulation. In:
Proc. of the 2005 Winter Simulation Conference. (2005)
28. ImmunoGrid. The European Virtual Immune System Project Available at:
www.immunogrid.eu (Last accessed 15 Jul 2012)
29. Thorne, B.C., Bailey, A.M., Pierce, S.M.: Combining experiments with multi-cell
agent-based modeling to study biological tissue patterning. Briefings in bioinformatics 8 (2007) 245–257
30. Jerne, N.K.: Towards a network theory of the immune system. Ann. Immunol
(Inst. Pasteur) 125C (1974) 73–89
31. Baltcheva, I., Codarri, L., Pantaleo, G., Boudec, J.Y.L.: Lifelong Dynamics of
Human CD4+ CD25+ Regulatory T Cells: Insights from in vivo Data and Mathematical Modeling. Journal of Theoretical Biology 266 (2010) 307–322
32. Kotiadis, K., Robinson, S.: Conceptual modelling: Knowledge acquisition and
model abstraction. In Madon, S.J., Hill, R.R., Mnch, L., Rose, O., Jefferson, T.,
Fowler, J.W., eds.: Proc. of the 2008 Winter Simulation Conference. (2008) 951–958
33. Ulgen, O.M., Black, J.J., Johnsonbaugh, B., Klungle, R.: Simulation methodology
- a practitioner’s perspective. Int. Journal of Industrial Engineering, Applications
and Practice 1 (1994)
34. Daigle, J.: Human immune system simulation: A survey of current approaches.
Georgia State University (2006)
35. Murray, J.M., Kaufmann, G.R., Hodgkin, P.D., Lewin, S.R., Kelleher, A.D., Davenport, M.P., Zaunders, J.: Naive T cells are maintained by thymic output in early
ages but by proliferation without phenotypic change after twenty. Immunology and
Cell Biology 81 (2003) 487–495
| 5cs.CE
|
Output Feedback Control of a Class of
Uncertain Systems Under
Control-derivative dependent disturbances
arXiv:1610.05847v1 [cs.SY] 19 Oct 2016
M. Alamir ∗ J. Dobrowolski ∗∗ A. T. Mohammed ∗
∗
CNRS, University of Grenoble Alpes
(mazen.alamir(amgad.mohamed)@gipsa-lab.grenoble-inp.fr).
∗∗
Schneider Electric Industrie. 37 Quai Paul Louis Merlin. 38000
Grenoble ([email protected])
Abstract: This paper addresses the design of robust dynamic output feedback control for highly
uncertain systems in which the unknown disturbance might be excited by the derivative of the
control input. This context appears in many industrial problems such as the speed control of the
hydraulic turbines and the frequency stabilization in micro grids to cite but few examples. A key
feature that has to be carefully addressed in this context is that too agressive feedback might
lead to the loss of controllability and/or a significant drop in the closed-loop performance. The
paper formulates the problem, underlines its relevance and gives a rigorous solution in which
a dynamic output feedback is given together with a realistic set of sufficient conditions on
the controller’s parameters that enable to assess the behavior of the closed-loop under different
circumstances. A numerical example is given to illustrate the relevance of the proposed successful
design characterization.
Keywords: Robust Control, control-dependent uncertainties, low gain feedback.
1. INTRODUCTION
This paper addresses the problem of output feedback design of dynamical systems showing the following features:
(1) There is a simple although uncertain relationship
between u and y (y is of relative degree 0). This simplicity suggests the use of appealingly simple feedback
design.
(2) However, a cautionless design of such a simple law
might excite some internal dynamics of the system
leading to oscillations and/or lack of controllability.
(3) In spite of this issue, the use of advanced observerbased nonlinear controller that would be designed
using the entire model in order to master the internal
dynamics is not welcomed by practitioners because
of its high complexity and the need to derive a tight
models of complex behaviors that are not always well
mastered. Moreover, the observer design is not always
easy nor the system is always observable.
These reasons make the framework addressed in the
present contribution appealing (when possible), namely,
keep simple control structure but set the parameters of
the simple control law in a sound and theoretically assessed way so that the above mentioned coupling issue is
appropriately addressed.
To the best of the authors knowledge, existing works
do not tackle the case of output feedback of uncertain
systems with control derivative-dependent uncertainties.
Typical classes of uncertainties that are commonly handled
are parametric uncertainties [Kar 1999], additive bounded
state-dependent uncertainties [Cheng 1998, Nunes et al.
2009] or nonlinear uncertainties with affine state dependent bounds [Li et al. 2016] to cite but few representative
references.
Works belonging to the model-free approach [Fliess and
Join 2013], while addressing higher relative degrees, are
also based on a representation of the unknown term to be
estimated that is not dependent on the control behavior
in a way that might enhance instable coupling.
This paper is organized as follows: First, the problem is
stated in section 2 and motivated through the description
of two industrial example (control of hydraulic turbines
and frequency stabilization in micro grids). Section 3 discusses the working assumptions. The main results regarding the successful feasibility bounds are given in Section
4. An illustrative example is given in Section 5. Finally,
Section 6 concludes the paper and gives some hints for
further investigation.
In the sequel, given a time interval I := [t1 , t2 ] and a
variable x, the boldfaced symbol x is used to denote the
profile x(t) t∈I of x over I when the later is known without ambiguity from the context. Otherwise the explicit
notation x(I) is used instead. Let xt := x(t) and if x
is scalar, let kxk∞ := supt∈I |x(t)|. The set of possible
profiles w of the exogenous uncertainty involved in (2) is
denoted by W.
2. PROBLEM STATEMENT
We consider uncertain dynamic systems of the form:
h
i
ẏ(t) = α(t) u(t) − g(t) + ℓ(t, η(t)))
(1)
η̇(t) = E(η [t−τ,t] , u̇[t−τ,t] , w[t−τ,t] )
(2)
where y ∈ R and u ∈ [u, u] are the regulated output and
the control input respectively. g(·) is an unknown term
satisfying g(t) ∈ [g, g] with known bounds g and g. The
vector η ∈ Rnη represents internal states with dynamics
given by (2) in which E : Rnη × R× Rnw → Rnη is possibly
unknown map. the vector w ∈ Rnw represents exogenous
disturbances. Finally α(t) ∈ [α, ᾱ] is an unknown scalar
that might be time-varying, with known bounds α and ᾱ.
Let ∆u := ū − u and ∆g = ḡ − g denotes the admissible
excursion of the control input and the unknown term g
respectively.
Remark 1. Equation (2) is a rather general representation
that can encompasses time delayed systems, general partial differential equations to cite but few examples.
A careful examination of (1) clearly shows that the controllability of y can be guaranteed if one can enforce the
following inclusion:
h := g − ℓ(η) ∈ [u + ρ, u − ρ] ⊂ [u, ū]
(3)
for some ρ > 0 and if |u̇| can be taken arbitrarily high
since in this case, u can instantaneously dominate h and
consequently enforce any desired sign to ẏ. However, the
fact that the dynamic (2) of η depends on u̇ makes this
second condition unrealistic as a strong dynamics on u
(high u̇) can steer the system to regions where (3) is
violated leading to the loss of controllability and even
instability. This is precisely the issue that is investigated
in this contribution. Note that the absence of precise
knowledge of ℓ(·, ·), E(·, ·, ·), g(·) and α makes the problem
even more challenging as it prevents, for instance, the
reconstruction of η (from the only measured y) and the use
of the estimated values in the necessarily careful control
design. Moreover, the observability itself may not hold in
many situations.
In this paper, the following simple dynamic output feedback is considered
u = S λ(yd − y) + z
(4)
ż = λf (u − z)
(5)
where S is the saturation function that projects the argument inside [u, u] and yd is the constant set-point for
the regulated output y. This feedback is defined up to
the choice of the two positive parameters λf , λ > 0. The
problem addressed in this paper can be formulated as
follows:
Problem
Find a high level characterization of the
uncertainties and maps involved in (1)(2) and associated conditions on the
control parameters (λ, λf ) so that the
behaviour of the closed-loop system can
be conveniently/quantifiably assessed.
More precisely, the assessment of the closed-loop properties should be understood in the following sense:
(1) Asymptotic stability should be obtained in the absence of disturbance (w ≡ 0, g ≡ 0 and ℓ ≡ 0)
(2) A quantifiable bound can be given on the asymptotic
tracking error in the case where persistent dynamic g
and non vanishing disturbance w are present.
Before addressing the problem stated above, the following section underlines the relevance of the framework described by (1)-(2) by showing that it describes at least two
concrete industrial problems.
2.1 Relevance of the framwork
The framework described above fits at least two important
industrial problems that are briefly discussed here.
Control of hydraulic turbines
Efficient control of hydraulic turbines in the context of
Power Storage Plant (PSP) is gaining a renewed interest.
In this context, the hydraulic plant is used as a mean to
compensate for the intermittency of renewable sources of
energy [Connolly et al. 2012]. As such, the response time
of speed control is a key parameter that determines the
success of the whole power management system. In this
context, the regulated variable y is the rotational speed
Ω, the control u is the guiding vane’s opening γ while the
internal state η is the state of the penstock (flow rate and
pressure values inside the penstock) that drives the water
flow from the water reservoir to the turbine’s level. All
practitioners know [Zhou 2011] that the major obstacle
that limits high rotational acceleration of the turbine lies
in the induced oscillating pressure waves in the penstock.
When referring to the internal dynamic (2), the time laps
τ represent twice the time needed for the pressure wave
to travel the penstock’s length. These pressure oscillations
generate disturbing torque variations at the turbine’s level
that can go beyond what the controlled torque can compensate if the dynamics of the control is not monitored
appropriately. That is the reason why simple control laws
that are not carefully tuned against this phenomenon fail
to produce satisfactory results.
Distributed frequency-control in isolated micro grids
This is another example that falls into the framework
studied in this paper. Here the regulated variable is the
grid frequency f . The control input u is the sum of produced powers by the Gensets and the renewable sources
while the uncertainty w is the derivative of the load power
consumption. The internal dynamics is induced by the
flexibility of power transmission between Gensets following
sudden change in the load power demand. The need for a
distributed control design (in which only the frequency
measurement is used) emphasizes the need for simple
control laws that cannot afford the task of model-based
reconstruction of the internal dynamic (which needs both
observability and total knowledge of the model’s details).
This internal dynamics disturbs the commonly used [Zhao
et al. 2016, Xialing et al. 2014] power balance equation
f˙ = α(P − PL ) by an additional term such that:
f˙ = α(u − PL + Cη)
(6)
η̇ = Aη + B1 u̇ + B2 ṖL
(7)
which obviously takes the form (1)-(2) with w = ṖL .
Note that a 4th order dynamics is typically needed to
correctly represent the phenomena when only two Gensets
are simply considered.
contexts described in Section 2.1, the inequalities (9)-(10)
become true after a finite time 1 with:
Z t
eA(t−σ) B2 wσ dσ
(11)
c0 := ǫ + sup sup C
c1 := sup C
t≥0
The first unavoidable assumption is called the Control Authority Margin (CAM) assumption. It simply states that
u can ultimately dominates the unknown term g ∈ [g, ḡ]
alone with some margins, denoted by ̺+ > 0 and ̺− :
Assumption 3.1. (Control Authority Margin).
There exist two scalars ̺− > 0 and ̺+ > 0 such that:
(8)
ū ≥ ḡ + ̺+ and u ≤ g − ̺−
These different positive constants are shown in Figure 1.
w∈W t≥0
d1 := |CB1 | + sup CA
t≥0
u
g
̺−
ḡ
∆g
ū
̺+
Fig. 1. Definition of the different positive constants involved in the characterization of the actuator and the
unknown term g.
The next assumption characterizes the behavior of the
additional term ℓ involved in (1) under the dynamic (2)
for all possible disturbance profiles w ∈ W and all profiles
u:
Assumption 3.2. There exist positive scalars c0 , d0 > 0
and c1 , d1 > 0 such that, for all w ∈ W and all admissible
input profiles u (inside [u, ū]), the following inequality
holds:
|ℓ(t, η(t))| ≤ c0 + c1 ku̇[t−τ,t] k∞
(9)
|ℓ̇(t, η(t))| ≤ d0 + d1 ku̇[t−τ,t]k∞
(10)
Moreover, the evolution of the state η remains bounded
for any bounded profiles in kuk∞ and ku̇k∞ .
This assumption can be better understood if one considers
the case where the dynamics (2) is given by (7) in which
the matrix A is Hurwitz (open-loop stable internal dynamics). Indeed in this case, which holds for both industrial
B1 dσ
(12)
Z
t
eA(t−σ) B2 wσ dσ
(13)
eA(t−σ) B1 dσ
(14)
0
Z
t
0
In the case of hydraulic turbines, asymptotic stability
of the penstock dynamics comes from the fact that this
system is theoretically marginally stable but becomes
asymptotically stable by introducing a pre-compensator
in the definition of the flow rate feedback. The control u
discussed in this paper represent then the definition of the
remaining feed-froward term. As for the micro grids case,
the asymptotic stability comes from the use of the droop
frequency control as a first rapid loop on the top of which,
the control we are interested in amounts at translating the
droop curve vertically. This is obviously equivalent to an
action on the total produced power u.
In the sequel, the short notation ku̇k∞ is used for
ku̇[t−τ,t]k∞ as t is removed and τ is supposed to be fixed
once for all.
Note that without this assumption, there is obviously
no means to guarantee bounded behavior of the output
tracking error and this, regardless of the control law being
used.
∆u
e
0
d0 := ǫ + sup sup CA
3. WORKING ASSUMPTIONS
In this section, the high level characterization that are
mentioned in the problem’s statement (Section 2) are
introduced. These characterizations delimit the class of
unknown dynamics and quantities for which the simple law
(4)-(5) can be appropriate provided that some conditions
on the choice of λ and λf are fulfilled. These conditions
are established in Section 4.
w∈W t≥0
0
Z t
A(t−σ)
4. SUFFICIENT CONDITIONS ON CONTROL
PARAMETERS AND ASSOCIATED BOUNDS ON
THE TRACKING ERROR
Let us first of all examine some properties of the signals
involved in the feedback laws (4)-(5).
Lemma 4.1. Under the feedback law (4)-(5), for all λ > 0
satisfying:
1
λ<
(15)
ᾱc1
the following inequality holds:
λᾱ β + c0 + λf ∆u
ku̇k∞ ≤ δu :=
(16)
1 − λᾱc1
where
(17)
β := ∆g + max ̺+ , ̺−
Proof. By the very definition (4), inequality (16) has to
be checked only inside the admissible domain as otherwise
u̇ = 0 and the inequality obviously holds. Now, when
the saturation constraint is not active, one obviously has
according to (4):
|u̇| ≤ λ|ẏ| + |ż|
(18)
but according to (1), it comes that:
h
i
|ẏ| ≤ ᾱ |u − g| + |ℓ|
(19)
an examination of Figure 1 clearly shows that:
o
n
|u − g| ≤ β := ∆g + max ̺+ , ̺−
therefore, using (20) and (9) in (19) gives:
h
i
|ẏ| ≤ ᾱ β + c0 + c1 ku̇k∞
1
(20)
(21)
beyond which the influence of the initial state that prevailed at the
beginning of the system’s first operation time becomes lower than ǫ.
≤ λf ∆u
z
(2) any dynamics starting in A− enters A0 in finite time.
(3) A0 is invariant.
z − λ(y − yd ) = ū
θ
≥ α(̺+ − |ℓ|)
A+
A0
z − λ(y − yd ) = u
≥ α(̺− − |ℓ|)
≤ λf ∆u
y − yd
A−
Fig. 2. Definition of the sets A+ , A− and A0 and disposition of the vector fields.
on the other hand, since z is a first order filtered version
of u, it comes that z ∈ [u, ū] and hence:
|ż| ≤ λf ∆u
(22)
using (20)-(21) in (18) leads to:
i
h
(23)
|u̇| ≤ λ ᾱβ + c0 + c1 ku̇k∞ + λf ∆u
which can be rewritten equivalently as follows:
(1 − λᾱc1 )ku̇k∞ ≤ λᾱ(β + c0 ) + λf ∆u
which gives obviously (16) if (15) is satisfied.
(24)
✷
In order to go further in the analysis of the closed-loop
behavior, the following sets are defined (see Figure 2):
A+ := {(y, z) | z − λ(y − yd ) ≥ ū}
(25)
(26)
A− := {(y, z) | z − λ(y − yd ) ≤ u}
A0 := {(y, z) | z − λ(y − yd ) ∈ (u, ū)}
(27)
Note that by definition, if (y, z) ∈ A+ then u = ū, if
(y, z) ∈ A− then u = u while if (y, z) ∈ A0 then the
control is not saturated, namely u = λ(yd − y) + z.
Based on these definitions, the following result can be
stated:
Lemma 4.2. (A0 is attractive and invariant).
If the following conditions hold:
• λ < 1/(ᾱc1 )
• λf and λ are such that
λf < λϕ(λ)
(28)
where:
c1 α i−1 h
(β + c0 )c1 ᾱ i
̺ − c0 − λ
α
1 − λᾱc1
1 − λᾱc1
(29)
where ̺ := min{̺− , ̺+ }, then the set A0 is attractive
and invariant for the closed-loop dynamics associated to
the control law (4)-(5) and to any constant desired value
yd .
h
ϕ(λ) := ∆u 1 +
Proof. The steps of the proof are the following:
(1) any dynamics starting in A+ enters A0 in finite time.
proof of (1). Assume that (y, z) ∈ A+ (see Figure 2.
Examination of the vector field (see Figure 2) suggests
that the result would be proved if one can prove that the
following inequality holds:
|ż| < λ|ẏ|
(30)
On the other hand we know that |ż| ≤ λf ∆u while |ẏ| ≥
α(̺+ − |ℓ|). Now using the inequality (9), the inequality
(30) would be satisfied if the following condition holds:
λf ∆u < α ̺+ − (c0 + c1 ku̇k∞ ) λ
(31)
and since λ < 1/(ᾱc1 ) the inequality (16) of Lemma 4.1
holds, therefore, (30) would be satisfied if the following
inequality holds:
##
"
"
λᾱ β + c0 + λf ∆u
+
λf ∆u < α ̺ − c0 + c1
λ (32)
1 − λᾱc1
which is equivalent to (after straightforward manipulations):
h
c1 α i−1 h +
(β + c0 )ᾱλ i
λf < ∆u 1 +
̺ − c0 − c1
αλ
1 − λαc1
1 − λᾱc1
(33)
proof of (2). Following exactly the same arguments, it
can be shown that A0 is attractive for any initial condition
in A− provided that the following condition holds:
h
c1 α i−1 h −
(β + c0 )ᾱλ i
λf < ∆u 1 +
̺ − c0 − c1
αλ
1 − λαc1
1 − λᾱc1
(34)
Combining the more restrictive of the two conditions (33)
and (34) obviously leads to (29).
Proof of (3). This simply comes from the fact that the
flows ẋ and ż are continuous (in y and z) and it has
just been shown that the vector fields on the intersection
boundaries A+ ∩ A0 and A− ∩ A0 does strictly enter inside
the set A0 . Therefore A0 is invariant.
✷
Now regarding the existence of λ, λf > 0 that satisfy the
conditions of Lemma 4.2, one has the following result:
Lemma 4.3. (Existence of λ, λf ). If the following condition holds:
c0 < ̺ := min{̺− , ̺+ }
(35)
then there exists λ∗ > 0 such that for all λ < λ∗ , there
exists λf > 0 such that the pair (λ, λf ) satisfies the
requirements of Lemma 4.2. More precisely, λ∗ is solution
of:
1
|
inf ϕ(λ) ≥ 0
(36)
λ∗ = sup σ ≤
ᾱc1
λ∈(0,σ)
Proof. This is because, thanks to (35), ϕ(·) satisfies:
min{̺− , ̺+ } − c0
ϕ(0) =
>0
(37)
∆u (1 + c1 )
and since ϕ is a continuous function of λ inside (0, ᾱc1 1 ),
there exists λ∗ ∈ (0, ᾱc1 1 ) such that for all λ ∈ (0, λ∗ ),
ϕ(λ) > 0 hence there exists λf > 0 satisfying (29).
✷
Based on the result of Lemma 4.2, it comes out that the
asymptotic behavior of the closed-loop is only determined
by the behavior of the system under the non-saturated
control law u = λ(yd −y)+z. But this feedback law leads to
a dynamics (inside A0 ) that can be written in the following
form:
−αλ α y − yd
ẏ
(38)
=
−λf λ 0 z − h
ż
where h is the unknown dynamics given by:
h = g − ℓ(η)
(39)
and using the following notation:
−αλ α
(40)
ey := y − yd ; ez := z − h ; A0 :=
−λf λ 0
the dynamics (38) becomes:
0
e
ėy
= A0 y −
ez
ėz
ḣ
(41)
Based on this observation, the main result of the paper
can be derived:
Proposition 4.1. (Main result).
If the following conditions are satisfied:
(1) Assumptions 3.1-3.2 hold,
(2) c0 < min{̺− , ̺+ },
(3) λ ∈ (0, λ∗ ) and λf < λϕ(λ) where λ∗ solves (36) in
which ϕ is given by (29).
(4) the unknown term g satisfies |ġ| ≤ δg
then the tracking error ey satisfies the following inequality:
δg + d0 + d1 δu
lim |y(t) − yd | ≤
(42)
t→∞
λλf
where δu > 0 is the constant given by (16).
Proof. Under the Assumption of the proposition, the conditions of Lemma 4.2 (by definition of λ∗ ) hold. Therefore,
the closed-loop trajectories lie ultimately inside A0 and the
dynamics (41) persistently prevails after a finite time. This
means that one has asymptotically:
Z t
0
ey (t)
A0 (t−σ)
dσ
(43)
e
→
ez (t)
ḣ(σ)
0
Focusing on ey and taking the worst case leads to:
Z t
lim |ey (t)| ≤
C1 eA0 (t−σ) C2 dσ max |ḣ(σ)| (44)
t→∞
σ∈[0,t]
0
T
where C1 := (1, 0) and C2 = (0, 1) . This gives, because
A0 is invertible [see (40)]:
lim |ey (t)| ≤ C1 A−1
0 C2 max |ḣ(σ)|
t→∞
σ∈[0,t]
(45)
Now by definition (39) of h, the fact that |ġ| ≤ δg and (16)
of Lemma 4.1:
max |ḣ(σ)| ≤ δg + d0 + d1 δu
(46)
σ∈[0,t]
Moreover, simple computations show that C1 A−1
0 C2 =
1/(λλf ) which obviously gives the result.
✷
5. ILLUSTRATIVE EXAMPLE
where A, B1 , B2 and C are given by:
" #
#
" #
"
0
0.1
−3.4 −1 −0.5
A = 0.25 −1.7 0.5 ; B1 = 0 ; B2 = 0.2
0
0
1.2 2.75 −3.6
0
1]
C = [1
while the delay τ = 10 is used. The unknown signal g(t)
is given by:
g(t) := 1 − 0.3 cos(0.1t) − 0.1 sin(0.2t + π/6);
(47)
Therefore, the bounds g = 0.5 and ḡ = 1.5 can be used
(∆g = 1). Moreover, the bound on the derivative of g is
δg = 0.05. Therefore, considering control saturations given
by u = −1 and ū = 3 leads to ̺− = 1.5 and ̺+ = 1.5,
̺ = 1.5 and ∆u = 4. The bounds α = 0.1, ᾱ = 0.3 are
supposed to hold while a true value α = 0.2 is used in the
simulations. We consider as admissible set W to be the
set of all disturbances that are bounded by w̄ = 0.1. For
simulation, the disturbance signal w(t) = w̄ sin(t) is used.
Computation of (11)-(14) gives:
c0 = 0.05 c1 = 0.07 ; d0 = 0.2 ; d1 = 0.42
(48)
Using these values in the definition of ϕ(·) and λ∗ leads
to the evolution of ϕ(λ), the bound on the asymptotic
tracking error and the bound λϕ(λ) on the values of λf
are depicted in Figure 3 where it can be inferred that
λ∗ = 17.18. Note that the upper bound 1/(ᾱc1 ) ≈ 47.4 in
this example meaning that in this example, λ∗ is limited
by the need for a positive sign of ϕ.
Figure 4 shows the behavior of the closed-loop system
when the value λ(1) = 17 ∈ (0, λ∗ ) is adopted together
with λf = 0.95λϕ(λ) ≈ 0.0102 satisfying (28). This Figure
shows quite good closed-loop behavior with the tracking
error asymptotically below its theoretical bound a while
after the set-point gets constant.
On the other hand, Figure 5 shows the behavior of the
closed-loop system when λ(2) = 10λ(1) is used (with the
same value of λf . The closed-loop clearly shows that the
trajectories are no more adequately attracted to a small
neighborhood of the desired values. This is mainly due to
the high excursion of the internal state η which leads to
high value of ℓ that breaks the attractiveness of the tight
neighborhood of yd . This excursion has to be compared to
the rather moderate one which can be observed on Figure
4. This clearly shows that a rationalized choice of (λ, λf )
is a key step in the success of the simple control law (4)-(5)
for this particular class of systems.
6. CONCLUSION AND FUTURE WORKS
In this paper, the problem of dynamic output feedback
of systems with control-derivative dependent uncertainties
is analyzed. High level characterization of the involved
map is given so that a successful tuning of a simple
feedback law is given together with an upper bound on the
output tracking error. A natural extension of this work is
a concrete application to the realistic examples described
in the paper.
Let us consider the uncertain system given by (1)-(2) in
which the internal dynamic (2) is given by:
REFERENCES
η̇(t) = Aη(t) + B1 ku̇[t−τ,t]k∞ + B2 w
ℓ(η) = Cη
Cheng, C.F. (1998). Output feedback stabilization for
uncertain systems: constrained riccati approach. IEEE
Transactions on Automatic Control, 43(1), 81–84.
4
·10
101
100
10
2
0
−2
10−1
5
0
5
10
15
20
Asymptotic bound on |ey | for λf = 0.95λ ∗ ϕ(λ)
10−2
10−3
0
0
100 200 300 400
4
100
0
5
10
15
Upper bound λ ∗ ϕ(λ) on λf
20
0
0
−2
−0.2
5
10
15
20
λ
Fig. 3. (top) Evolution of ϕ(λ) given by (29). (Middle)
Asymptotic bound on the output tracking error ey [see
(42)]. (Bottom) Upper bound λϕ(λ) on the control
parameter λf [see (33)].
|ey |
y and yd
101
100
10
10−1
5
10−2
10−3
0
100 200 300 400
10−4
0
time
u(t)
4
100 200 300 400
time
ℓ(t)
8
6
2
4
0
2
0
100 200 300 400
time
0
0
100 200 300 400
time
Fig. 4. Behavior of the closed-loop system for the choice
(λ(1) , λf ) = (17, 0.0102) meeting the theoretical
bounds.
time
ℓ(t)
6
4
2
0
100 200 300 400
time
0
100 200 300 400
8
2
0.2
−2
0
time
u(t)
102
0
ey
y and yd
Evolution of ϕ(λ) vs λ
−2
0
0
100 200 300 400
time
Fig. 5. Behavior of the closed-loop system for the choice
(λ(2) , λf ) = (170, 0.0102) that violate the prescribed
bounds.
Connolly, D., Lund, H., Mathiesen, B., Pican, E., and
Leahy, M. (2012). The technical and economic implications of integrating fluctuating renewable energy using
energy storage. Renewable Energy, 43, 47–60.
Fliess, M. and Join, C. (2013). Model-free control.
arXiv:1305.7085v2.
Kar, I. (1999). Design of static output feedback controller
for uncertain systems. Automatica, 35(1), 169 – 175.
Li, G., Lin, Y., and Zhang, X. (2016). Adaptive output
feedback control for a class of nonlinear uncertain systems with quantized input signal. International Journal
of Robust and Nonlinear Control.
Nunes, E.V.L., Hsu, L., and Lizarralde, F. (2009). Global
exact tracking for uncertain systems using outputfeedback sliding mode control. IEEE Transactions on
Automatic Control, 54(5), 1141–1147.
Xialing, X., Jianghui, X., Youping, X., Ruyu, B., Hui, L.,
Tao, L., and Jing, Y. (2014). Research on the dynamic
frequency characteristic of large-scale power grid considering the action of power system splitting and load
shedding. In Proceedings of the 2014 international conference on power systems technology (powercon 2014).
Zhao, J., Lyu, X., Fu, Y., Hu, X., and Li, F. (2016).
Coordinated microgrid frequency regulation based on
dfig variable coefficient using virtual inertia and primary frequency control. IEEE Transactions on Energy
Conversion, 31(3), 833–845.
Zhou, J.X. (2011). Analytical study on possible selfexcited oscillation in s-shaped regions of pump-turbines.
Proc. IMech, 225, 1132–1142.
| 3cs.SY
|
Protein Folding Optimization using Differential Evolution
Extended with Local Search and Component Reinitialization
arXiv:1710.07031v1 [cs.AI] 19 Oct 2017
Borko Bošković · Janez Brest
October 20, 2017
Abstract This paper presents a novel differential evolution algorithm for protein folding optimization that
is applied to a three-dimensional AB off-lattice model.
The proposed algorithm includes two new mechanisms.
A local search is used to improve convergence speed and
to reduce the runtime complexity of the energy calculation. For this purpose, a local movement is introduced
within the local search. The designed evolutionary algorithm has fast convergence and, therefore, when it
is trapped into local optimum or a relatively good solution is located, it is hard to locate a better similar
solution. The similar solution is different from the good
solution in only a few components. A component reinitialization method is designed to mitigate this problem.
Both the new mechanisms and the proposed algorithm
were analyzed on well-known amino-acid sequences that
are used frequently in the literature. Experimental results show that the employed new mechanisms improve
the efficiency of our algorithm and the proposed algorithm is superior to other state-of-the-art algorithms.
It obtained a hit ratio of 100 % for sequences up to 18
monomers within a budget of 1011 solution evaluations.
New best-known solutions were obtained for most of the
sequences. The existence of the symmetric best-known
solutions is also demonstrated in the paper.
Keywords Protein folding optimization · Threedimensional AB off-lattice model · Differential
evolution · Local search · Component reinitialization
B. Bošković
Faculty of Electrical Engineering and Computer Science,
University of Maribor, SI-2000 Maribor, Slovenia
E-mail: [email protected]
J. Brest
Faculty of Electrical Engineering and Computer Science,
University of Maribor, SI-2000 Maribor, Slovenia
E-mail: [email protected]
1 Introduction
The protein structure prediction represents the problem of how to predict the native structure of a protein from its amino-acid sequence. This problem is one
of the more important challenges of this century [17]
and, because of its nature, it attracts scientists from
different fields, such as Physics, Chemistry, Biology,
Mathematics, and Computer Science. Within the protein structure prediction, the Protein Folding Optimization (PFO) represents a computational problem for simulating the protein folding process and to find a native
structure. Most proteins must fold into unique threedimensional structure, known as a native structure, to
perform their biological function [2]. Protein function
is determined by its structure. Inability of a protein to
form its native structure prevents a protein from fulfilling its function correctly, and this may be the basis of
various human diseases [1, 25, 28, 29].
The PFO belongs to the class of NP-hard problems [11] and, with current algorithms and computational resources, it is possible to predict the native structures of relatively small proteins. The reason for that is
the huge and multimodal search space. For example, a
polypeptide that has only 18 amino-acids, will have 31
angles within a simplified AB model (see Section 3). Using uniform discretization with only 10 values for each
angle, there would be 1031 possible configurations. To
evaluate and select the correctly folded conformation
among all these conformations in the time elapsed since
the Big Bang, we need the huge computational speed
of 1031 /(4.32 · 1017 ) = 2.31 · 1013 conformation evaluations per second. This is much faster than the speed
obtained within our experiment, where we can evaluate
only 5.73 · 105 conformations per second. From these
numbers, we can see that the search space is huge, even
2
in the simplified model, which makes this problem very
hard. However, in reality, the proteins fold into their
native conformation on a time scale of seconds, and
this contradiction is known as Levinthal’s paradox [7].
An optimization algorithm can give good results of a
PFO problem only if it can locate good solutions and
evaluate solutions efficiency. Here, the approximation
techniques, such as heuristic and metaheuristic, with
efficient data structures, become the only viable alternatives as the problem size increases.
There exist some simplified protein models, such
as HP models within different lattices [5, 21] and the
AB off-lattice model [30, 31]. Simplified protein models
were designed for development, testing, and comparison
of different approaches. The AB off-lattice model was
used in the paper for demonstrating the efficiency of
the proposed algorithm. This model takes into account
the hydrophobic interactions which represent the main
driving forces of a protein structure formation and, as
such, still imitates its main features realistically [14].
Although this model is incomplete, it allows the development, testing, and comparison of various search
algorithms, and offers a global perspective of protein
structures. It can be helpful in confirming or questioning important theories [3].
In our recent work [4], we proposed a Differential
Evolution (DE) [32,33] algorithm for PFO that includes
DE /best/1 /bin strategy, self-adaptive mechanism [6] of
main control parameters, random reinitialization and
temporal locality mechanism [10, 38]. This algorithm
belongs to the ab-initio PFO methods, which optimize
structures from scratch and do not require any information about related sequences. It showed a very fast convergence, and it was capable of obtaining significantly
better results than other state-of-the-art algorithms.
In this paper, we propose two new mechanisms, that
additionally, improve the efficiency of our algorithm. A
new local search mechanism was designed in order to
improve convergence speed and to reduce runtime complexity of the algorithm. A similar idea was already used
within the HP model [5], where it is applied to the cubic
lattice. Using a simple local search mechanism where
only one solution’s component is changed can produce
a structure whereby a lot of monomers are moved. This
means their positions must be recalculated and efficient energy calculation is not possible. In contrast to
simple local search, our mechanism improves the quality of conformations using the local movements within
the three-dimensional AB off-lattice model. We define
a local movement as a transformation of conformation
whereby only two consecutive monomers are moved locally in such a way that the remaining monomers remain in their positions. In such a way, the designed
Borko Bošković, Janez Brest
local movement allows efficient evaluation of neighborhood solutions. With the fast convergence, the algorithm can locate good solutions quickly, but it has a
problem locating good similar solutions. For that purpose, a component reinitialization was designed and incorporated within our algorithm. This mechanism is
employed when the local best solution is detected. It
produces similar solutions that are different from the
local best solution in only a few components.
We called the proposed algorithm DElscr and it was
tested on two sets of amino-acid sequences that were
used frequently in the literature. The first set included
18 real peptide sequences, and the second set included
4 well-known artificial Fibonacci sequences with different lengths. Experimental results show that the proposed mechanisms improve the efficiency of the algorithm, and the algorithm is superior to other stateof-the-art algorithms. Its superiority is especially evident for larger sequences. With the proposed algorithm,
that is stochastic, we cannot prove optimality of the
obtained conformations. However, we can infer about
them according to the observed hit ratio. The experimental results show that our algorithm obtained hit
a ratio of 100 % for sequences that contain up to 18
monomers. For all larger sequences, we can only report
the best-known conformations that are almost surely
not optimal. Based on these observations, the main contributions of this paper are:
1. The proposed new DE algorithm for the PFO on a
three-dimensional AB off-lattice model.
2. The local search mechanism that improves convergence speed and reduces runtime complexity of solution evaluations within the neighborhood.
3. The component reinitialization which increases the
likelihood of finding good similar solution.
4. With the observed hit ratios we show how difficult
the PFO is, even in a simplified model, and that,
with the current algorithm, we can confirm solutions
with a hit ratio of 100 % only for sequences that have
up to 18 monomers.
5. An approach for determining the algorithm’s asymptotic average-case performances.
6. The existence of two best-known (potentially global
best) structures that are symmetrical for all sequences
with up to 25 monomers.
7. The new best-known conformations for most of the
sequences.
The remainder of this paper is organized as follows.
A related work for the PFO on a three-dimensional AB
off-lattice model is described in Section 2. The threedimensional AB off-lattice model is described in Section 3. A description of the introduced algorithm with
the emphasis on new mechanisms is given in Section 4.
Protein Folding Optimization using DE with Local Search and Component Reinitialization
The experimental setup and numerical results are presented in Section 5. Section 6 concludes this paper.
2 Related work
Over the years, different algorithms have been applied
successfully to the PFO on a three-dimensional AB offlattice model. In [12] the low energy configurations are
optimized using the pruned-enriched-Rosenbluth method (PERM). This method was applied to the lattice
model quite successfully [34]. The conformational space
annealing was studied using Fibonacci sequences in [20]
and compared with nPERMis (new PERM with importance sampling) [13]. Next, the algorithm that was compared with the nPERM algorithm was proposed in [8].
In this work, the problem is converted from a nonlinear constraint-satisfied problem to an unconstrained optimization problem which can be solved by the wellknown gradient method. The statistical temperature
molecular dynamics based algorithm ”statistical temperature annealing” was applied to an AB model in [18].
The efficiency of an improved tabu search algorithm
was analyzed in [39]. The following hybrid algorithms
were also developed for the AB model: A hybrid algorithm that combines the genetic algorithm and tabu
search algorithm [35, 36, 40], particle swarm optimization and levy flight [9], the particle swarm optimization,
genetic algorithm, and tabu search algorithm [41], and
improved genetic algorithm and particle swarm optimization algorithm with multiple populations [42]. An
artificial bee colony algorithm [22–24, 37] and an improved harmony search algorithm that is combined with
dimensional mean based perturbation strategy [15] were
also analyzed on the AB model.
The authors in [16] determined the structural features of the PFO using Fitness Landscape Analysis
(FLA) techniques based on the generated landscape
path. From the results of FLA, it has been shown that
the PFO has a highly rugged landscape structure containing many local optima and needle-like funnels, with
no global structure that characterizes the PFO complexity. The obtained results also show that the artificial bee colony algorithm outperforms all other algorithms significantly in all instances for the three-dimensional AB off-lattice model.
In our recent work [4], we proposed a differential
evolution algorithm that is adapted to PFO on a threedimensional AB off-lattice model. Within this algorithm,
we incorporated a self-adaptive mechanism, a mutation
strategy for the fast convergence and a temporal locality. The obtained results of this algorithm show that it
is superior to the algorithms from the literature, including the artificial bee colony algorithm, and significantly
3
lower free energy values were obtained for longer AB
sequences.
3 Three-dimensional AB off-lattice model
The basic building blocks of proteins are amino-acids.
The linear chain of amino acids is a polypeptide, and
a protein contains at least one long polypeptide. Each
polypeptide can be represented with a unique amino
acid sequence. The polypeptide must fold into a specific three-dimensional native structure before it can
perform its biological function(s) [27]. Thus, all information necessary for folding must be contained in the
amino-acid sequence, and this is known as the Anfinsenhypothesis [7].
From the amino-acid sequence it is possible to generate different conformations, which is also dependent
on the used model. In general, there exist two types
of simplified models: Off-lattice and lattice. The lattice
model maps each position of amino-acid to the point
on a discrete lattice. In contrast to the lattice model,
the off-lattice model allows any position and, as such,
is more accurate. The simplified three-dimensional AB
off-lattice model was proposed in [31]. Instead of 20
standard amino-acids, this model uses only two different types of amino-acids: A – hydrophobic and B – hydrophilic. Thus, an amino-acid sequence is represented
as a string s = {s1 , s2 , ..., sL }, si ∈ {A, B}, where A
represents a hydrophobic, B a hydrophilic amino-acid
and L the length of the sequence. The three-dimensional
structure of an AB sequence is defined by bond angles
θ = {θ1 , θ2 , ..., θL−2 }, torsional angles β = {β1 , β2 , ...,
βL−3 } and the unit-length chemical bond between two
consecutive amino-acids (see Fig. 1).
Different energy calculations can be used within different models. Within an AB model, the free energy
value is calculated using a simple trigonometric form of
backbone bend potentials E1 (θ) and a species-dependent
Lennard-Jones 12,6 form of non-bonded interactions
E2 (s, θ, β) as shown in the following equation [31]:
E(s, θ, β) = E1 (θ) + E2 (s, θ, β)
E1 (θ) =
L−2
1 X
[1 − cos(θi )]
4 i=1
E2 (s, θ, β) = 4
L−2
X
L
X
(1)
[d(pi , pj )–12 –c(si , sj ) · d(pi , pj )–6 ]
i=1 j=i+2
where pi = {xi , yi , zi } represents the position of the i-th
amino-acid within the three-dimensional space. These
positions are determined as shown in Fig. 1 and by the
following equation:
4
Borko Bošković, Janez Brest
(a)
(b)
Fig. 1 A schematic diagram of a sequence ABAB. (a) Projection of a structure with θ1 = 30, θ2 = −60 and β1 = 0 onto the
XY-plane. (b) Projection of a structure with θ1 = 30, θ2 = −60 and β1 = 45 onto the ZY-plane.
4.1 Proposed algorithm
{0, 0, 0}
{0, 1, 0}
{cos(θ ), 1 + sin(θ ), 0}
1
1
pi =
{xi−1 + cos(θi−2 ) · cos(βi−3 ),
yi−1 + sin(θi−2 ) · cos(βi−3 ),
z
i−1 + sin(βi−3 )}
if i = 1,
if i = 2,
if i = 3,
(2)
if 4 ≤ i ≤ L.
In Eq. (1) d(pi , pj ) denotes the Euclidean distance between positions pi and pj , while c(si , sj ) determines
the attractive, weak attractive or weak repulsive nonbonded interaction for the pair si and sj as shown in
the following equation:
1
c(si , sj ) = 0.5
−0.5
if si = A and sj = A,
if si = B and sj = B,
if si 6= sj .
The objective of PFO within the context of an AB offlattice model is to simulate the folding process and to
find the angles’ vector or conformation that minimizes
the free-energy value:
{θ ∗ , β ∗ } = arg min E(s, θ, β).
4 Proposed algorithm for PFO
In this paper, we extend our differential evolution algorithm [4] with two new mechanisms. The first mechanism is a local search that improves convergence speed
and reduces runtime complexity for solution evaluation
within the specific neighborhood. The second mechanism is component reinitialization, which allows the algorithm to locate good similar conformations according
to the local best solution.
Hereinafter, we will describe briefly the DElscr algorithm that is shown in Fig. 2. The lines that represent improvements according to the previous version
are highlighted with gray background. The optimization process begins with initialization (line 2). Each iteration of while loop (line 3) represents one generation
of the evolutionary process. A mutation, crossover, and
selection are performed for each population’s individual
{x1 , x2 , ..., xNp } within one generation. The DE /best/1
mutation strategy and binary crossover (lines 7 – 18)
are used for creating a trial individual u. The values
of mutation F , and crossover Cr control parameters
are set with the self-adaptive jDE mechanism (lines 5
and 6) [6]. The trial individual is evaluated in line 19
using Eq. (1). From this line, we can see that the individuals are D-dimensional vectors that contain real
coded bond θ and torsional β angles:
xi = {xi,1 , xi,2 , ..., xi,D } =
= {θ1 , θ2 , ..., θL−2 , β1 , β2 , ..., βL−3 }
xi,j ∈ [−π, π];
D =2·L−5
i = 1, 2, ..., Np;
j = 1, 2, ..., D
The selection mechanism, temporal locality and local
search are shown in lines 20 – 40. If the trial vector
is better than the corresponding vector from the population (eu < ei ), then the second trial vector u∗ is
generated using temporal locality [38] and the better
vector replaces population vector xi . The next mechanism is a local search. Within this mechanism, a local
movement is used for improving the population best
individual xpb according to each pair {θn−1 , βn−2 }. The
values of {θn−1 , βn−2 } represent the angles that specify
the position of the (n + 1)-th monomer according to
the position of the n-th monomer. After each generation, either random or component reinitialization will
Protein Folding Optimization using DE with Local Search and Component Reinitialization
1: procedure DElscr (s, Np)
2:
Initialize a population P
{xi , Fi = 0.5, Cr i = 0.9, ei = E(s,xi )} ∈ P
xi,j = −π + 2 · π · rand [0,1]
i = 1, 2, ..., Np; j = 1, 2, ..., D; D = 2 · length(s) − 5
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
24:
25:
26:
27:
28:
29:
30:
31:
32:
33:
p
{xb , eb } = {xlb , elb } = {xp
b , eb } = BEST(P )
while stopping criteria is not met do
for i = 1 to Np do
if rand [0,1] < 0.1 then F = 0.1 + 0.9 · rand [0,1]
else F = Fi end if
if rand [0,1] < 0.1 then Cr = rand [0,1]
else Cr = Cr i end if
do r1 =rand {1,Np} while r1 =i end do
do r2 =rand {1,Np} while r2 =i or r2 =r1 end do
jrand = rand {1,D}
for j = 1 to D do
if rand [0,1] < Cr or j = jrand then
uj = xb,j + F · (xr1 ,j − xr2 ,j )
if uj ≤ -π then uj = 2 · π + uj end if
if uj > π then uj = 2 · (-π) + uj end if
else
uj = xi,j
end if
end for
eu = E(s,u) // Energy calculation
if eu ≤ ei then
// Temporal locality
for j = 1 to D do
u∗
j = xb,j + 0.5 · (uj − xi,j )
∗
∗
if u∗
j ≤ -π then uj = 2 · π + uj end if
∗
if u∗
>
π
then
u
=
2
·
(-π)
+
u∗
j
j
j end if
end for
∗
e∗
u = E(s,u )
if e∗
u ≤ eu then
{xi , Fi , Cri , ei } = {u∗ , F, Cr , e∗
u}
else
{xi , Fi , Cri , ei } = {u, F, Cr , eu }
end if
// Local Search
34:
35:
for n = 2 to L − 1 do
θn-1 = rand [0,1] · (xp
b,n-1 − xi,n-1 )
3:
4:
5:
6:
36:
βn-2 = rand [0,1] · (xp
b,n+(L-4) − xi,n+(L-4) )
37:
{v, ev } = LOCAL MOVEMENT(xp
b , n, θn-1 , βn-2 )
38:
39:
40:
41:
42:
p
if ev ≤ eb then {xp
b , ei } = {v, ev } end if
end for
end if
end for
p
{xp
b , eb } = BEST(P )
43:
p p
if ep
b ≤ eb then {xb , eb } = {xb , eb } end if
p
l
l
REINITIALIZATION({xp
44:
b , eb },{xb , eb },P )
45:
end while
46:
return {xb , eb }
47: end procedure
Fig. 2 The proposed DElscr algorithm.
be performed if the reinitialization criteria are satisfied
(line 44). At the end of the evolutionary process, the
algorithm returns the best obtained solution xb and its
energy value eb , as shown in line 46. The local search
5
and reinitialization are described in more detail in the
following subsections. For a detailed description of the
rest mechanism, we refer readers to [4].
4.2 Local search
If the trial vector is better than the corresponding population vector, then the temporal locality and local search
are performed. The local search calculates the values of
pair {θn−1 , βn−2 } for each monomer, with the exception of the first two monomers. Positions of these two
monomers are fixed within the AB-model, see Eq. (2).
For the rest monomers, the angle values are defined
with a randomly scaled difference between the population best individual xb and current individual xi , as
shown in the following equation:
θn−1 = rand [0,1] · (xpb,n−1 − xi,n−1 )
βn−2 = rand [0,1] · (xpb,n+(L−4) − xi,n+(L−4) )
i = 1, ..., Np;
n = 2, ..., L − 1;
A meticulous reader may notice that the θ has index
n − 1, while β has index n − 2. This means that, for
n = 2 the value of β0 is calculated using the values of
θ components. The reason for that is in the position of
the third monomer, which is dependent only on the θ1 ,
see Eq. (2). Therefore, within the local movement for
n = 2 the value of β0 is ignored, and local movement
takes into account only the value of θ1 .
The local movement is a transformation of conformation whereby only two consecutive monomers are
moved locally in such a way that the remaining monomers remain in their positions. There is only one exception, for the last monomer only one monomer is moved,
while all the remaining monomers remain in their positions. Fig. 3 shows an example of two monomers’ local
movement. The polygon that is defined with points P1 ,
P2 , P3 , P4 represents the section of original conformation where points represent the monomer positions. The
local movement moves the point P2 to the point X2 according to the pair {θn−1 , βn−2 } while the position of
point X3 is calculated using Eqs. (3) – (6). In these
calculations, the point X3 is the nearest to the point
P3 in such a way that the new polygon P1 , X2 , X3 ,
P4 must end at the point P4 .
The calculation begins with the determination of
point C whose position is in the middle of points P4
and X2 :
C = X2 +
P4 − X2
.
2
(3)
The length L between points C and X3 is calculated by
using the triangle P4 , C, X3 and Pythagoras’s theorem:
6
Borko Bošković, Janez Brest
Fig. 3 A schematic diagram of a local movement. The polygon P1 , P2 , P3 , P4 is transformed to polygon P1 , X2 , X3 , P4
in such a way that only the two consecutive points P2 and P3 are moved to points X2 and X3 , while the remaining points
P1 and P4 stay unchanged in their positions.
∆e2 =
s
L=
2
1−
kP4 − X2 k
.
4
P4 − X2
.
kP4 − X2 k
(4)
(P3 − C) − CN
· L.
k(P3 − C) − CN k
(5)
(6)
The created polygon contains unchanged points P1 and
P4 , which means only monomers X2 and X3 are moved
locally, while the remaining monomers stay in their
unchanged positions. This feature allows us to design
the fast conformation evaluation within the local movement. Two additional data structures E1 and E2 were
used for this purpose. The values of elements within
these data structures are determined according to Eq.
(1), as follows:
E1i = 1 − cos(θi )
E2i,j = d(pi , pj )−12 − c(si , sj ) · d(pi , pj )−6
i ∈ {1, ..., L − 2};
j ∈ {i + 2, ..., L}.
Using these data structures for the population best individual and its energy value, we can calculate the energy of the conformation created by local movement
efficiently, as is shown in Eqs. (7) – (9) where n represents the variable that was sent to the local movement
procedure, as shown in Fig. 2 (see line 37).
∆e1 = E1n−1 − (1 − cos(θn−1 )) +
E1n − (1 − cos(θn )) +
E1n+1 − (1 − cos(θn+1 ))
E2i,j −
d(pi , pj )−12 − c(si , sj ) · d(pi , pj )−6
n+3
X
i h
X
i
E2j,i −
+
(8)
i=n+1 j=1
At the end, point X3 is calculated by scaling of vector
(P3 − C) − CN as follows:
X3 = C +
L
h
X
i=n+1 j=n+2
The vector projection of (P3 − C) onto line X2 , P4 is
calculated with the following equation:
CN = (P3 − C) ·
n+3
X
(7)
d(pi , pj )−12 − c(si , sj ) · d(pi , pj )−6
i
∆e1
+ 4 · ∆e2 )
(9)
4
From Eqs. (7) – (9) we can observe that the time com2
plexity of energy calculation is reduced from L2 to 2L.
In this way, the designed local movement allows faster
evaluation of neighborhood solutions, and its usage within local search improves convergence speed.
ev = epb − (
4.3 Reinitialization
After each generation reinitialization will be performed
if the reinitialization criteria are satisfied. In our previous work, random reinitialization will be performed
if the best individual stays unchanged within the evolution process for more than 105 evaluations. In this
work we design a new reinitialization mechanism. For
that purpose, the algorithm includes three types of best
individuals and three new control parameters, as shown
in Fig. 4. The individuals xpb , xlb , xb represent the population best, local best and global best individuals. The
population best individual is the best individual in the
current population, the local best individual is the best
individual among all similar individuals, and the global
best individual is the best individual obtained within
the evolutionary process. From this description of the
best individuals, we can see that reinitialization has to
control evolutionary process to:
Protein Folding Optimization using DE with Local Search and Component Reinitialization
p
l
l
1: procedure REINITIALIZATION({xp
b , eb },{xb , eb },P )
is
unchanged
for
at
least
P
·
D
evaluations
then
2:
if xp
b
b
p p
l
l
l
3:
if ep
b ≤ eb then {xb , eb } = {xb , eb } end if
4:
if xlb is unchanged for Lb · D reinitializations then
5:
// Random reinitialization
6:
xi = RANDOM() ; i = 1, 2, ..., Np
p
7:
{xlb , elb } = {xp
b , eb } = BEST(P )
8:
else
9:
// Component reinitialization
10:
xi = RANDOM(xlb , C); i = 1, 2, ..., Np
p
11:
{xp
b , eb } = BEST(P )
12:
end if
13:
end if
14: end procedure
Fig. 4 The reinitialization mechanism.
• Locate the best individual by using the current population,
• Locate the best similar individual by using component reinitialization, and
• Locate the global best individual by using random
reinitialization.
The following new control parameters are introduced
to the reinitialization mechanism: Pb , Lb and C. The
Pb defines how long the best population individual can
stay unchanged within the evolutionary process (line 2).
For example, with Pb = 100 and dimension 21, the
reinitialization is required if the best population individual stayed in the evolutionary process for at least
Pb · D = 100 · 21 = 2100 evaluations. When this condition is satisfied, the algorithm performs random or component reinitialization according to the parameter Lb . If
the local best individual is not changed for Lb · D reinitializations, then random reinitialization is performed,
and component reinitialization otherwise (line 4). The
last parameter C determines the number of components that are different between the local best individual and individuals generated by component reinitialization (line 10). Within this operation, the C components of each population individual xi are selected, and
their values are replaced with random values on the interval [−π, π], while all the remaining components get
the values from the local best individual.
5 Experiments
The DElscr algorithm was compiled with a GNU C++
compiler 4.6.3 and executed using an Intel Core i7 computer with 2.93 GHz CPU and 8 GB RAM under Linux
Mint 13 Maya and a grid environment (Slovenian Initiative for National Grid1 ). In order to evaluate the
efficiency of the proposed algorithm, we used a set of
1
Available at http://www.sling.si/sling/
7
amino-acid sequences as shown in Table 1. This set includes 5 Fibonacci sequences and 18 real peptide sequences from the Protein Data Bank database2 . The
K-D method is used to transform real peptide sequences
to the AB sequences. In this method the amino acids I,
V, P, L, C, M, A, and G are transformed to hydrophobic ones (A) and amino acids D, E, H, F, K, N, Q, R,
S, T, W, and Y to hydrophilic ones (B) [26]. The selected sequences have different lengths, which enabled
us to analyze the algorithm according to different problem dimensions and, because they were used frequently
in literature, they enabled us to compare the proposed
algorithm with different algorithms. In order to analyze the efficiency of the introduced mechanisms and
algorithm, we measured the following statistics:
• The mean obtained energy value for all runs:
PN
i=1
Emean =
Ei
N
where Ei denotes the obtained energy value for the
i-th run and N the number of all runs.
• The best obtained energy value among all runs:
Ebest = max{E1 , E2 , ..., EN }.
Note that all energy values within our experiments
are multiplied by -1, which means that all energy
values are positive and higher values are better.
• The standard deviation of energy values for all runs:
s
PN
− Emean )2
N −1
i=1 (Ei
Estd =
• The hit ratio or percentage of runs during which
the best solution has equal or better energy value
according to the target value (target):
hitr =
Nh
N
where Nh denotes the number of runs where the best
obtained solution has good enough energy value eb
according to the target value (eb ≥ target).
• The mean number of solution evaluation for all Nh
runs:
PNh
NSE i
NSE mean = i=1
Nh
where NSE i represents the number of solution evaluations for the i-th run.
2
Available at https://www.rcsb.org/pdb/home/home.do
8
Borko Bošković, Janez Brest
Table 1 Details of amino-acid sequences used in experiments.
Label
1BXP
1CB3
1BXL
1EDP
2ZNF
1EDN
2H3S
1ARE
2KGU
1TZ4
1TZ5
1AGT
1CRN
2KAP
1HVV
Length
13
13
16
17
18
21
25
29
34
37
37
38
46
60
75
D
21
21
27
29
31
37
45
53
63
69
69
71
87
115
145
1GK4
84
163
1PCH
88
171
2EWH
98
191
F13
F21
F34
F55
F89
13
21
34
55
89
21
37
63
105
173
Sequence
ABBBBBBABBBAB
BABBBAABBAAAB
ABAABBAAAAABBABB
ABABBAABBBAABBABA
ABABBAABBABAABBABA
ABABBAABBBAABBABABAAB
AABBAABBBBBABBBABAABBBBBB
BBBAABAABBABABBBAABBBBBBBBBBB
ABAABBAABABBABAABAABABABABABAAABBB
BABBABBAABBAAABBAABBAABABBBABAABBBBBB
AAABAABAABBABABBAABBBBAABBBABAABBABBB
AAAABABABABABAABAABBAAABBABAABBBABABAB
BBAAABAAABBBBBAABAAABABAAAABBBAAAAAAAABAAABBAB
BBAABBABABABABBABABBBBABAABABAABBBBBBABBBAABAAABBABBABBAAAAB
BAABBABBBBBBAABABBBABBABBABABAAAAABBBABAABBABBBABBAABBABBAABBBB
BAABBBBBABBB
ABABAABABBBBABBBABBABBBBAABAABBBBBAABABBBABBABBBAABBABBBBBAABAB
AAABABAABBBBAABABBBBA
ABBBAAABBBAAABABAABAAABBABBBBBABAAABBBBABABBAABAAAAAABBABBABABA
BABBABBAABAABBBAABBAAABA
AABABAAAAAAABBBAAAAAABAABAABBAABABAAABBBAAAABABAAABABBAAABAAABA
AABAABBAABAAAAABAAABABBBABBAAABAABA
ABBABBABABBAB
BABABBABABBABBABABBAB
ABBABBABABBABBABABBABABBABBABABBAB
BABABBABABBABBABABBABABBABBABABBABBABABBABABBABBABABBAB
ABBABBABABBABBABABBABABBABBABABBABBABABBABABBABBABABBABABBABBAB
ABBABBABABBABABBABBABABBAB
• The standard deviation of solution evaluations for
all Nh runs:
s
PNh
2
i=1 (NSE i − NSE mean )
NSE std =
Nh − 1
• The mean runtime for all runs:
PN
ti
tmean = i=1
N
where ti represents the runtime of the i-th run.
• The mean speed for all runs:
PN
vi
vmean = i=1
N
where vi represents the speed (the number of solution evaluations per second) of the i-th run.
The listed statistics were measured within the context
of the following stopping conditions:
• The maximum number of solution evaluations
NSElmt : NSE i ≥ NSE lmt .
• The runtime limit tlmt : ti ≥ tlmt .
• The energy value of the best obtained solution target:
eb ≥ target.
Our algorithm belongs to stochastic algorithms, therefore, all the reported results of the proposed algorithm
within this work are based on N = 100 independent
runs. The described statistics, the defined stopping criteria and the determined number of independent runs
were used to analyze the influence of new parameters
and mechanisms on the algorithm’s efficiency. The algorithm was also compared with the state-of-the art
algorithms. The asymptotic average-case performances
were determined for the 6 shortest sequences and an
analysis of the obtained conformations was also performed and will be given in the continuation of the paper.
5.1 Parameter settings
The influence of the new control parameters (Pb , Lb , C)
on the algorithm’s efficiency was analyzed by using Fibonacci sequences. In this analysis, the stopping condition was the maximum number of solution evaluations
NSElmt = 1010 . For each sequence, we started with the
following setting: Pb = 50, Lb = 10, and C = 5. Using 6 settings, where only one value of each setting was
changed to the nearest higher or lower value, we tried
to get better settings. The parameter values are used
from the following sets:
Pb ∈ {10, 25, 50, 100}
Lb ∈ {1, 2, 5, 10, 20, 50}
C ∈ {2, 5, 10, 20}
For the started setting the following 6 settings were
used:
{ Pb = 10, Lb = 10, C = 5}, {Pb = 25, Lb = 10, C = 5},
{ Pb = 50, Lb = 5, C = 5}, {Pb = 50, Lb = 20, C = 5},
{ Pb = 50, Lb = 10, C = 2}, {Pb = 50, Lb = 10, C = 10}.
Protein Folding Optimization using DE with Local Search and Component Reinitialization
9
Table 2 The analysis of the new control parameters (Pb , Lb , C). N = 100 independent runs were performed for each setting
and the stopping conditions were the maximum number of solution evaluations NSElmt = 1010 and target value.
Pb Lb C
50 2 5
100 2 5
25 2 5
50 5 5
50 2 10
50 1 5
50 10 5
50 2 2
Emean
6.9961
6.9961
6.9961
6.9961
6.9961
6.9961
6.9961
6.9847
Estd
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.03
hitr
NSE mean
NSE std
100 6.41E+07 5.59E+07
100 6.88E+07 5.96E+07
100 7.52E+07 7.01E+07
100 9.07E+07 8.06E+07
100 9.52E+07 9.67E+07
100 9.59E+07 9.55E+07
100 9.77E+07 9.48E+07
90 3.15E+09 2.67E+09
Pb
25
25
50
50
10
25
25
25
Lb C
20 5
10 5
10 5
20 5
20 5
50 5
20 10
20 2
Emean
16.4500
16.4492
16.4432
16.4415
16.4307
16.4254
16.4037
15.4393
(a) F13, target = 6.9961
Pb
50
50
25
50
100
50
50
Lb
10
20
10
5
10
10
10
C
5
5
5
5
5
10
2
Emean
30.0670
30.0596
30.0519
29.9108
29.9034
29.3722
24.2650
(c) F34
Estd
0.45
0.40
0.47
0.38
0.47
0.35
1.94
Estd hitr
NSE mean
NSE std
0.09 21 4.91E+09 2.75E+09
0.08 19 4.89E+09 2.81E+09
0.08 17 3.92E+09 2.98E+09
0.10 22 4.86E+09 3.17E+09
0.09 15 3.93E+09 2.51E+09
0.12 33 4.84E+09 2.85E+09
0.08
2 7.94E+09 2.53E+09
0.45
0
-
(b) F21, target = 16.5544
Pb
25
25
10
50
25
25
25
Lb
5
10
5
5
2
5
5
C
10
10
10
10
10
5
20
Emean
49.0262
49.0233
49.0148
48.9379
48.9192
47.8458
47.4250
Estd
0.78
1.26
1.05
1.19
1.03
1.74
0.88
(d) F55
We repeated this process until the new better setting
was found. The results of the least iterations, together
with recommended settings, are shown in Table 2. For
clarity, the recommended settings and their results are
shown in bold typeface. The displayed results show that
each sequence has its own optimal setting, but it is still
possible to select settings that can be used for any unknown sequence. We define these settings according to
the dimension of the problem, as follows:
(
{50, 10, 5} if n < 45
{Pb , Lb , C} =
{25, 5, 10} otherwise.
These settings are used in all the following experiments,
because they can provide a good hit ratio for small sequences and good energy values for larger sequences.
The search space for larger sequences is huge, which
means the algorithm cannot provide optimal or suboptimal solutions. Therefore, for these sequences, the algorithm has to perform more reinitializations, and the
component reinitialization has to change more components randomly.
The shown results confirm, additionally, that the
variable NSE have near-exponential or near-geometric
distribution (NSE mean ≈ NSE std ). Under such distributions, given the Nh = 100 runs in all of our experiments, a reliable rule-of-thumb estimates the 95% confidence interval:
1.96
1.96
CI 95 ≈ [(1 − √ ) · NSE mean , (1 + √ ) · NSE mean ]
Nh
Nh
≈ [0.8 · NSE mean , 1.2 · NSE mean ].
Pb
50
25
50
25
50
100
50
50
Lb
2
2
5
5
1
2
2
2
C
10
10
10
10
10
10
5
20
Emean
76.8608
76.6879
76.5090
76.4541
76.3478
76.3275
75.1975
75.0143
Estd
1.64
1.89
1.88
1.93
1.40
1.71
2.62
1.52
(e) F89
5.2 Local search
The local search within our algorithm was designed to
increase the speed of algorithm convergence and speed
of neighborhood solution evaluations. In order to demonstrate these advantages, the algorithm was analyzed
with (DElscr ) and without (DEwls
lscr ) local search. Within
this analysis, the algorithms were compared using the
following stopping conditions: NSE lmt = 107 and tlmt =
tmean (DEwls
lscr ), as shown in Table 3. With the first stopping condition, we show that the local search improves
mean energy value Emean on 16 out of 23 sequences
and reduces the mean runtime tmean for all sequences.
Using the second stopping condition, both algorithms
were limited with the same runtime and, in this case,
local search improves Emean in 22 out of 23 sequences.
These values are marked in bold typeface within the
table. Results also show that speed up factor cv was
1.46 for the smallest sequences (F13 and 1BXP) and
3.91 for the largest sequence 2EWH, while the Emean
is worse by 0.0171 for sequence 1CB3 and better for
all other sequences. For the largest sequence 2EWH,
the improvement of Emean was huge, from 182.2880 to
205.3507 or by 23.0627.
From the obtained results, we can conclude that the
local search improves the convergence speed of the algorithm for most of the sequences, while the speed of solution evaluation is increased for all sequences. Somebody
would expect better speed up factor, but note that some
conditions must be satisfied for local search and, there-
10
Borko Bošković, Janez Brest
Table 3 The influence of the local search to the algorithm’s efficiency according to two algorithms: DElscr - with local
search and DEwls
lscr - without local search. Two different comparisons were made according to two different stopping conditions:
NSE lmt = 107 and tlmt = tmean (DEwls
lscr ). The reported mean speed vmean represents the mean number of function evaluations
per second, cv represents the speed up factor vmean (DElscr )/vmean (DEwls
lscr ) and the mean runtime tmean is given in seconds.
NSE lmt = 107
Label
1BXP
1CB3
1BXL
1EDP
2ZNF
1EDN
2H3S
1ARE
2KGU
1TZ4
1TZ5
1AGT
1CRN
2KAP
1HVV
1GK4
1PCH
2EWH
F13
F21
F34
F55
F89
DEwls
lscr
DElscr
tmean
12.54
12.18
15.02
16.00
17.44
20.27
24.85
29.37
34.89
37.47
36.66
38.81
51.66
76.79
104.64
122.54
130.19
162.03
12.42
20.90
35.89
71.27
140.58
vmean
Emean
Estd
7.98E+05
4.7280 0.24
8.22E+05
7.9643 1.02
6.67E+05 16.2149 0.51
6.25E+05 13.7281 1.07
5.74E+05 16.1670 1.86
4.94E+05 17.9565 2.51
4.03E+05 15.1685 2.36
3.41E+05 19.3024 1.92
2.87E+05 43.6622 3.46
2.68E+05 28.7054 4.77
2.73E+05 34.3793 4.36
2.58E+05 52.9353 4.99
1.94E+05 78.8070 4.74
1.31E+05 54.6804 6.60
9.62E+04 57.5815 6.55
8.23E+04 72.3524 7.15
7.76E+04 103.4913 13.35
6.24E+04 189.5316 14.33
8.05E+05
6.0591 0.94
4.79E+05 12.0495 1.66
2.79E+05 20.0652 2.59
1.41E+05 35.2611 3.55
7.19E+04 55.4025 5.06
tmean
18.35
18.00
24.81
27.39
29.78
38.05
50.64
63.85
84.42
95.07
95.11
100.39
139.22
220.23
320.20
394.79
438.63
546.97
18.08
37.84
82.90
191.88
455.38
vmean
Emean
Estd
5.45E+05
4.6772 0.26
5.56E+05 8.1511 0.65
4.03E+05 16.2338 0.66
3.65E+05 13.3930 1.80
3.36E+05 14.4350 3.01
2.63E+05 16.5951 3.10
1.98E+05 15.1545 2.79
1.57E+05 18.6434 2.48
1.19E+05 40.6789 4.66
1.05E+05 25.3528 5.30
1.05E+05 32.9423 5.08
9.97E+04 47.6395 5.86
7.19E+04 78.8601 5.69
4.54E+04 55.3668 6.08
3.12E+04 57.3717 6.56
2.53E+04 72.8575 7.87
2.28E+04 101.7607 11.42
1.83E+04 182.2880 16.48
5.53E+05
6.0951 1.09
2.64E+05 11.3298 1.75
1.21E+05 18.3942 2.29
5.21E+04 35.3539 3.02
2.20E+04 55.0803 5.55
tmean
18.35
18.00
24.81
27.39
29.78
38.05
50.64
63.85
84.42
95.07
95.11
100.39
139.22
220.23
320.20
394.79
438.63
546.97
18.08
37.84
82.90
191.88
455.38
tlmt = tmean (DEwls
lscr )
DElscr
vmean
cv
Emean
Estd
7.98E+05 1.46
4.7831 0.26
8.22E+05 1.48
8.1340 0.92
6.72E+05 1.66 16.3452 0.48
6.07E+05 1.66 14.1388 0.54
5.73E+05 1.70 16.7171 1.24
4.97E+05 1.89 18.9328 1.85
4.06E+05 2.06 16.1873 2.34
3.44E+05 2.20 20.2815 1.80
2.92E+05 2.46 46.1607 2.28
2.74E+05 2.61 31.7309 4.26
2.81E+05 2.67 36.7141 4.66
2.66E+05 2.67 56.5688 3.80
2.02E+05 2.82 82.8484 2.14
1.37E+05 3.01 59.9772 5.93
1.03E+05 3.29 63.5027 6.29
9.10E+04 3.59 78.8594 6.17
8.71E+04 3.82 116.0248 11.99
7.14E+04 3.91 205.3507 10.62
8.07E+05 1.46
6.3373 0.83
4.82E+05 1.82 12.8967 1.70
2.87E+05 2.38 21.4521 2.62
1.46E+05 2.80 37.9751 2.44
7.92E+04 3.61 59.1793 4.40
Table 4 The influence of the component reinitialization to the algorithm’s efficiency according to two algorithms: DElscr - with
11
component reinitialization and DEwcr
lscr - without component reinitialization. Two stopping conditions were used: NSE lmt = 10
and target values that were set to the best known energy values (see Table 8). The shown CNSE represents the reduction of
NSE mean : DElscr (NSE mean ) / DEwcr
lscr (NSE mean ).
Label
Length
D
target
1BXP
1CB3
1BXL
1EDP
2ZNF
F13
13
13
16
17
18
13
21
21
27
29
31
21
5.6104
8.4589
17.3962
15.0092
18.3402
6.9961
NSE mean
1.56E+09
3.61E+07
1.24E+10
4.58E+09
2.10E+09
8.92E+07
DElscr
CNSE
NSE std
0.38 1.68E+09
0.18 4.26E+07
< 0.37 1.24E+10
< 0.22 4.21E+09
< 0.05 1.92E+09
0.08 8.52E+07
fore, the speed up factor is dependent on the relationship between the number of solution evaluations inside
and outside the local search. However, using the local
search, the algorithm is capable of obtaining better energy values for almost all sequences, and this improvement of energy values increases for larger sequences.
5.3 Component reinitialization
The main goal of the component reinitialization is to
redirect the evolutionary process in such a way that a
similar good solution can be located according the local best solution. To demonstrate the influence of this
mechanism on the algorithm’s efficiency, the algorithm
was analyzed with (DElscr ) and without (DEwcr
lscr ) component reinitialization. Within this analysis, the algorithms were compared using the target values of the
hitr
100
100
100
100
100
100
NSE mean
4.07E+09
1.99E+08
3.36E+10
2.14E+10
4.31E+10
1.14E+09
DEwcr
lscr
NSE std
4.20E+09
1.77E+08
2.71E+10
2.10E+10
2.75E+10
1.27E+09
hitr
100
100
32
96
50
100
best-known energy values and NSE lmt = 1011 , as shown
in Table 4. The best values of NSE mean and hit r are
marked in bold typeface. From the shown results, we
can find that the algorithm that uses component reinitialization is capable of reaching the best-known energy
value in all runs, and for that it required significantly
less solution evaluations (NSE ). On the other hand,
the algorithm without component reinitialization was
not capable of reaching hit r = 100 within the budget of NSE lmt = 1011 for sequences 1BXL, 1EDP and
2ZNF. From these observations, we can conclude that
the proposed component reinitialization allows the algorithm to locate good similar solutions and to reach
the best-known energy values. This is shown clearly in
Table 4, with CNSE which represents the relationship
between the obtained values of NSE mean for both algorithms. The component reinitialization reduces the
Protein Folding Optimization using DE with Local Search and Component Reinitialization
11
Table 5 Asymptotic average-case performances for DElscr . Values marked with the * are obtained by using the grid environmean
ment. In these cases tmean is calculated as follows: tmean = NSE
, where vmean represents the obtained mean speed of three
vmean
independent runs on our test computer in a given period of time (tlmt = 3600 seconds). All other results are obtained on our
test computer.
Label
1BXP
1CB3
1BXL
1EDP
2ZNF
F13
target
tmean
NSE mean
target
tmean
NSE mean
target
tmean
NSE mean
target
tmean
NSE mean
target
tmean
NSE mean
target
tmean
NSE mean
1
1.8013
0.16
2.45e+05
1.9174
0.06
1.03e+05
11.1862
167.93
8.96e+07
6.3823
677.33
6.74e+08
9.3228
4153.21
3.67e+09
1.8225
0.42
6.62e+05
2
2.0063
10.52
1.38e+07
1.9786
1.01
1.41e+06
13.8397
1472.79
7.07e+08
8.9122
437.82
3.88e+08
12.1166
270.26
2.18e+08
2.0453
3.34
4.52e+06
3
2.6838
395.94
4.62e+08
2.3884
1.67
2.06e+06
13.6386
611.86
2.68e+08
8.7042
1280.42
1.04e+09
11.9772
488.05
3.65e+08
4.6082
4.48
5.20e+06
4
3.1846
1349.36
1.42e+09
4.0429
4.22
4.82e+06
14.0105
508.47
2.14e+08
9.1152
776.01
5.97e+08
12.3307
755.42
5.50e+08
4.6858
8.03
8.43e+06
NSE mean from 0.38 to less than 0.05. Finally, it is obvious that the reinitialization mechanism is responsible
for the DElscr obtaining hit r = 100 for all the sequences
shown in Table 4.
Results show additionally that the value of NSE mean
is not only dependent on the sequence length or problem dimension, but also on the sequence itself. For example, for DElscr the value of NSE mean is 5.9 times
smaller for sequence 2ZNF in comparison with sequence
1BXL, although the dimension of the first sequence is
greater than the dimension of the second sequence.
5.4 Asymptotic average-case performance
In this section, we introduce an approach to determine
asymptotic average-case performance of the algorithm
for small sequences. The condition for this is the ability of the algorithm to obtain the best known solution
with hitr = 100. Until now, only our algorithm is able
to fulfill this condition for the 6 shortest sequences.
Six subsequences are generated for each of these sequences. The first subsequence has removed the last
monomer, the second subsequence has removed the last
two monomers, etc. This means that the length of each
next subsequence is decreased by 1. For example, for sequence 1CB3 (BABBBAABBAAAB) the following six
subsequences are generated:
1.
2.
3.
4.
5.
6.
BABBBAABBAAA (L = 12, target = 8.4088),
BABBBAABBAA (L = 11, target = 6.0209),
BABBBAABBA (L = 10, target = 4.0429),
BABBBAABB (L = 9, target = 2.3884),
BABBBAAB (L = 8, target = 1.9786),
BABBBAA (L = 7, target = 1.9174).
5
4.0191
12.39
1.17e+07
6.0209
4.70
4.52e+06
16.8991
25965.63 *
9.93e+09 *
11.5309
3494.96
2.47e+09
14.6296
562.56
3.77e+08
5.0428
10.47
9.74e+06
6
4.9321
125.70
1.09e+08
8.4088
80.85
7.11e+07
16.9404
113126.9 *
4.05e+10 *
11.7522
922.22
6.27e+08
14.6733
41844.24 *
1.32e+10 *
6.8092
70.98
6.19e+07
7
5.6104
1965.08
1.56e+09
8.4589
44.47
3.61e+07
17.3962
39706.68 *
1.34e+10 *
15.0092
7272.60
4.58e+09
18.3402
7437.15
2.10e+09
6.9961
110.54
8.92e+07
Asymptotic model
0.0015 · 2.8911L
4589.5644 · 2.5970L
0.0001 · 2.8662L
344.4917 · 2.5502L
0.0030 · 2.7931L
6240.7027 · 2.5976L
10.7641 · 1.4097L
2.32e+07 · 1.3103L
16.6030 · 1.3505L
3.19e+07·1.2642L
0.0019 · 2.3266L
6138.4492 · 2.0846L
The best-known or target values are determined for
all subsequences. In our case, we performed one run
for each subsequence with tlmt = 4 days and the best
reached energy value is used as a target value. Using
these target values as a stopping condition, it is possible to calculate asymptotic average-case performance.
The original sequence is also included within this calculation. This means the asymptotic average-case performance is determined by using 7 sequences.
Table 5 and Figure 5 display the target values, obtained mean values and asymptotic average-case performances for DElscr . From the shown results, we observe
that the best asymptotic average-case performances was
obtained for sequence 2ZNF while the worst for sequences 1BXP and 1BXL. We can again observe that
the value of NSEmean and tmean are not only dependent
on the sequence length. Only one monomer can influence these values significantly. For example, DElscr requires less solution evaluations (NSE ) and runtime (t)
to reach the target value for the subsequence of sequence 1EDP that has a length of 16 in comparison
with a subsequence of length 13. From these results, we
can conclude that the structure of the sequence has a
big influence on the difficulty of the problem.
5.5 Comparison with other algorithms
In this section, our algorithm is compared with other algorithms according to two stopping conditions NSElmt
and tlmt , and according to the best obtained energy
values.
The obtained results for stopping conditions NSElmt
that were set according to the literature [4,22] and three
12
11
13
14
15
16
17
12
13
14
20000
5000
observed t mean
model (t mean ) : 16.603 ⋅ 1.3505L
observed NSE mean
model (N S E mean ) : 31907895 ⋅ 1.2642L
●
●
2000
Runtime (t ) [s]
16
●
●
●
●
●
●
7
15
1000
5e+07
2e+07
5e+06
Number of solution evaluations (NSE )
50.0
5.0 10.0
2.0
Runtime (t ) [s]
●
●
1e+06
5e+08
●
●
●
1.0
●
●
●
500
5e+09
2e+09
1e+09
2000
●
observed t mean
model (t mean ) : 0.0019 ⋅ 2.3266L
observed NSE mean
model (N S E mean ) : 6138.449 ⋅ 2.0846L
5e+10
●
10
●
●
●
8
9
10
11
12
13
1e+10
2e+09
5e+08
1e+08
100
13
Number of solution evaluations (NSE )
10000
2000
2e+06
Runtime (t ) [s]
500
1e+07
5e+00
5e+05
11
(c) 1BXL
0.5
5000
10
(b) 1CB3
1000
Runtime (t ) [s]
9
(a) 1BXP
●
500
8
Length (L )
observed t mean
model (t mean ) : 10.7641 ⋅ 1.4097L
observed NSE mean
model (N S E mean ) : 23203404 ⋅ 1.3103L
12
●
7
Length (L )
●
11
●
Length (L )
●
●
●
5e+09
13
●
12
13
14
15
16
Length (L )
Length (L )
Length (L )
(d) 1EDP
(e) F13
(f) 2ZNF
2e+09
12
●
●
5e+08
11
●
●
17
18
Number of solution evaluations (NSE )
10
●
●
●
observed t mean
model (t mean ) : 0.003 ⋅ 2.7931L
observed NSE mean
model (N S E mean ) : 6240.703 ⋅ 2.5976L
2e+08
9
Number of solution evaluations (NSE )
8
●
Number of solution evaluations (NSE )
5e+01
●
7
●
●
1e+05
5e+05
Runtime (t ) [s]
5e+06
5e+00
●
●
observed t mean
model (t mean ) : 1e−04 ⋅ 2.8662L
observed NSE mean
model (N S E mean ) : 344.4917 ⋅ 2.5502L
5e−01
5e+07
5e+01
●
5e−01
Runtime (t ) [s]
●
5e−02
5e+08
5e+02
●
Number of solution evaluations (NSE )
●
●
observed t mean
model (t mean ) : 0.0015 ⋅ 2.8911L
observed NSE●mean
model (N S E mean ) : 4589.564 ⋅ 2.597L
50000
Borko Bošković, Janez Brest
5e+07
12
Fig. 5 Asymptotic average-case performances for DElscr .
algorithms DElscr , DEpfo and BE-ABC are shown in
Table 6. The best obtained energy values are marked in
bold typeface. It can be observed that DElscr and DEpfo
are comparable, and both outperform BE-ABC. Results
that take into account speed up factor (see Table 3) are
shown in the last column of the table. In this case, both
algorithms DElscr and DEpfo spend approximately the
same amount of runtime, and DElscr outperforms DEpfo
on most sequences.
Within this comparison, the value of NSElmt was
relatively small. This means the reinitialization mechanism within DEpfo did not have a significant impact on
the obtained results. Therefore, DEpfo and DElscr were
also compared according to the tlmt that was set to 4
days. A grid environment was used within this comparison and results are shown in Table 7. In this comparison, DElscr obtained better values of Emean , Ebest and
hitr in most sequences, and equal values for the shortest
sequences. DElscr obtained hitr of 100, 100, 19 and 65
for shorter sequences 1BXL, 1EDP, 2H3S and F21. For
the same sequences, DEpfo obtained hitr of 6.67, 13.33, 0
and 0, respectively. Significant improvement was obtained for longer sequences too. For example, the best
energy values were improved from 90.914, 131.7787 and
225.0968 to 106.419, 156.525 and 245.519 for sequences
1GK4, 1PCH and 2EWH. The energy values were improved by 15.505, 24.7463 and 20.4222. Note that DElscr
obtained the new best known solutions for all sequences
with L ≥ 18, the hitr = 100 for 6 shortest sequences
and hitr > 1 for 9 sequences by using tlmt = 4 days.
The most important results in this paper are shown
in Table 8, which collects the best energy values from all
experiments that were described in previous sections,
and the best-known energy values from the literature.
It is evident that DElscr confirmed the best known energy values for the 3 shortest sequences, and reached the
new best-known energy values for all other sequences,
except for sequence F13. Solutions for the best energy
values reached by DElscr are shown in Tables 9 and 10.
In [19] an efficient global optimization method is ap-
Protein Folding Optimization using DE with Local Search and Component Reinitialization
13
Table 6 Comparison of the DElscr algorithm with state-of-the-art algorithms.
Label
M
1CB3
1BXL
1EDP
2H3S
2KGU
1TZ4
1TZ5
1AGT
1CRN
1HVV
1GK4
1PCH
2EWH
F13
F21
F34
F55
20
20
20
20
20
20
20
20
20
20
20
80
80
4
4
12
20
Ebest
7.7450
16.2618
13.1764
17.1724
41.0221
34.5265
37.8896
49.9861
74.7849
45.0054
49.9316
121.0579
193.8143
4.9533
11.1304
19.9550
29.5163
DElscr
Emean
4.5108
12.5045
8.1986
11.5310
33.6539
21.6863
25.9996
39.1897
62.2668
35.9335
42.0261
87.5748
162.3482
3.0907
6.5538
13.3057
22.4019
Estd
2.13
2.17
2.78
2.45
3.99
3.62
4.12
5.21
7.60
4.92
4.77
11.42
16.60
0.78
1.53
2.47
3.58
NSElmt = M · 104
DEpfo [4]
Ebest
Emean
Estd
8.3690
5.5884
1.96
16.3443
12.6104
2.53
13.5620
8.6666
2.56
16.5030
10.6767
2.75
44.3369
35.3850
4.70
30.9211
20.4361
5.28
38.1868
27.3412
4.08
50.6311
39.0268
5.34
74.4068
60.2444
7.58
44.7264
34.8059
5.29
52.0651
44.8591
4.72
103.1776
79.4878
8.85
171.6390 144.9060 12.84
5.7290
3.6040
0.66
11.2211
7.9567
1.53
19.3529
14.0749
2.09
31.9554
24.6243
3.57
BE-ABC [22, 23]
Ebest
Emean
Estd
– 5.9417 0.78
– 11.6942 1.13
–
8.0500 0.93
– 10.4618 1.13
– 22.7195 2.01
– 14.9436 2.22
– 17.4859 1.37
– 25.6024 2.34
– 42.3083 2.96
– 21.5386 3.53
– 27.0410 3.24
– 51.6674 3.50
– 94.5785 5.70
3.3945
2.8196 0.38
6.9065
5.2674 0.76
10.4224
8.3239 0.92
18.8385 14.4556 1.56
NSElmt = M · 104 · cv
DElscr
Ebest
Emean
Estd
7.7450
4.5929
2.16
16.7137
13.1940
2.22
13.1895
8.5313
2.81
17.4858
11.9565
2.48
44.0110
36.4642
4.39
35.3505
24.9569
4.55
40.0161
28.9335
3.60
54.0897
43.4210
5.45
82.5999
68.3890
7.28
57.1990
46.3685
5.61
69.5798
56.6853
5.16
128.4882
99.3441 14.52
210.7021 181.5912 17.49
4.9704
3.1977
0.79
11.7522
7.6885
1.75
21.0345
15.4491
2.85
33.1788
26.8111
3.34
Table 7 The obtained results for DElscr and DEpfo within a runtime limit of 4 days.
Label
L
1BXP
1CB3
1BXL
1EDP
2ZNF
1EDN
2H3S
1ARE
2KGU
1TZ4
1TZ5
1AGT
1CRN
2KAP
1HVV
1GK4
1PCH
2EWH
F13
F21
F34
F55
F89
13
13
16
17
18
21
25
29
34
37
37
38
46
60
75
84
88
98
13
21
34
55
89
DElscr , number of independent runs N=100
Ebest
Emean
Estd
hitr
5.6104
5.6104 0.0000
100.00
8.4589
8.4589 0.0000
100.00
17.3962
17.3962 0.0000
100.00
15.0092
15.0092 0.0000
100.00
18.3402
18.3402 0.0000
100.00
21.4703
21.3669 0.0431
7.00
21.1519
20.9956 0.0995
19.00
25.2800
24.5444 0.1718
1.00
52.7165
51.7233 0.3829
1.00
43.0229
41.8734 0.4285
1.00
49.3868
48.6399 0.3292
1.00
65.1990
64.1285 0.4173
1.00
92.9853
89.8223 0.6514
1.00
85.5099
83.1503 1.0041
1.00
95.4475
91.4531 1.9215
1.00
106.4190
99.6704 3.0377
1.00
156.5250 153.1003 2.7117
1.00
245.5190 240.2247 2.1421
1.00
6.9961
6.9961 0.0000
100.00
16.5544
16.5304 0.0329
65.00
31.3455
30.4913 0.3458
1.00
51.9030
49.5009 0.8817
1.00
81.5297
76.4804 2.0603
1.00
plied to the sequence F89. Within this work, 32,200
distinct conformations were obtained, and the best obtained energy was 73.1065. DElscr improves this energy
by 10.4695.
5.6 Analysis of the obtained structures
For most of the sequences, the best conformations were
obtained by using tlmt = 4 days. Within this experiment, 100 solutions were generated with 100 independent runs. Distribution of the Root-Mean-Square Error
(RMSE) values as a function of energy for all these solutions according to the best-known conformation for
DEpfo [4], number of independent runs N=30
Ebest
Emean
Estd
hitr
–
–
–
–
8.4589
8.4589 0.0000
100.00
17.3962
17.1916 0.0878
6.67
15.0092
14.9423 0.0471
13.33
–
–
–
–
–
–
–
–
20.0979
19.6147 0.2699
0.00
–
–
–
–
50.2960
49.1661 0.6334
0.00
39.7340
37.8329 0.9983
0.00
47.1513
43.9959 1.4087
0.00
62.8951
60.4175 1.0439
0.00
89.2001
86.0390 1.4529
0.00
–
–
–
–
82.1427
68.8332 4.0852
0.00
90.9140
84.6836 3.3356
0.00
131.7787 117.7603 6.2617
0.00
225.0968 203.6813 7.1844
0.00
6.9961
6.9961 0.0000
100.00
16.2250
15.8894 0.1849
0.00
28.2509
25.6602 1.0523
0.00
45.0942
41.8670 1.4693
0.00
–
–
–
–
selected sequences is shown in Fig. 6. Note that the
RMSE is calculated by using the superposition between
matched pairs. From Fig. 6a, we can see that only two
different solutions were reached for sequence 2ZNF. Similar graphs with only two different solutions were obtained for 6 sequences where hitr = 100 (see Table 7).
Three-dimensional representations of these solutions are
displayed in Fig. 7. As is shown, for each of these sequences two solutions are symmetrical according to the
XY-plane. This can also be seen from Tables 9 and 10.
Two reported solutions for one sequence are very similar. They are different in some components that belong
to β torsional angles (marked in bold typeface), and
14
Borko Bošković, Janez Brest
Table 8 Comparisons of the best energy values reported in the literature and the best energy values obtained by DElscr . The
solution vectors obtained by DElscr are shown in Tables 9 and 10.
Label
DElscr
1BXP
1CB3
1BXL
1EDP
2ZNF
1EDN
2H3S
1ARE
2KGU
1TZ4
1TZ5
1AGT
1CRN
2KAP
1HVV
1GK4
1PCH
2EWH
F13
F21
F34
F55
F89
5.6104
8.4589
17.3962
15.0092
18.3402
21.4703
21.1519
25.2800
52.7165
43.0229
49.3868
65.1990
92.9853
85.5099
95.4475
106.4193
156.5252
245.5193
6.9961
16.5544
31.3459
52.0558
83.5761
DEpfo
[4]
–
8.4589
17.3962
15.0092
–
–
20.0979
–
50.2960
39.7340
47.1513
62.8951
89.2001
–
82.1427
90.9140
131.779
225.097
6.9961
16.2250
28.2509
45.0942
–
ImHS
[15]
4.498
–
15.200
–
15.056
17.721
15.340
17.416
40.696
–
–
40.300
61.426
44.972
–
–
–
–
–
–
–
–
–
BE-ABC
[22, 23]
2.8930
8.4580
15.9261
13.9276
5.8150
7.6890
18.3299
10.2580
28.1423
39.4901
45.3233
51.8019
54.7253
27.1400
47.4484
49.4871
91.3508
146.8231
6.9961
15.6258
28.0516
42.5814
–
I-PSO
[9]
–
–
–
–
–
–
–
–
20.9633
–
–
–
28.7591
15.9988
–
–
46.4964
–
–
–
–
–
–
their values represent angles with opposite directions.
For a little bit longer sequences more different solutions
were reached with different energy values, as shown in
Figs. 6b and 6c while, for the longest sequences, all
100 solutions are different with different energy values.
For example, this is illustrated in Fig. 6d for sequence
2EWH. From these results, we can conclude that all
reported symmetrical solutions could be optimal, especially those obtained with hitr = 100, and all other
solutions with hitr = 1 are almost surely not optimal.
6 Conclusions
In this paper, we presented a novel differential evolution
algorithm for protein folding optimization. To improve
its efficiency, the algorithm is extended with a component reinitialization and local search that includes a
local movement. The component reinitialization is designed to redirect the search process to similar solutions
that are different from the already found good solution in only a few components. Thus, the search space
around good solutions is explored thoroughly and, consequently, the algorithm can find better solutions. We
also designed the local movement for a three-dimensional
AB-off lattice model in such a way that only a two
consecutive monomers are moved locally while all the
remaining monomers remain in their positions. With
additional data structure this type of movement allows
us to reduce the runtime complexity of the energy cal-
PGATS
[41]
–
–
–
–
–
–
–
–
32.2599
–
–
–
49.6487
28.1052
–
–
49.5729
–
–
–
–
–
–
MPGPSO
[42]
–
–
–
–
–
–
–
–
–
–
–
–
43.9339
18.9513
–
–
38.2766
–
–
–
–
–
–
ABC
[24]
–
–
–
–
–
–
–
–
31.9480
–
–
–
52.3249
30.3643
–
–
63.4272
–
–
–
–
–
–
GATS
[35, 36]
–
8.2515
15.8246
13.7769
–
–
18.1640
–
–
39.3444
45.3019
46.0842
–
25.1003
–
–
–
–
6.9539
14.7974
27.9897
42.4746
–
2
C-ABC
[37]
–
–
–
–
–
–
–
–
–
–
–
–
–
–
–
–
–
–
7.0025
14.9570
28.0055
42.2769
–
culation within the local search from L2 to 2L. The 23
sequences are used in the experiments to analyze the
proposed algorithm and its mechanisms.
From the results of the algorithms with and without
local search, it is evident that the local movement with
additional data structure reduces the runtime complexity of the energy calculation, or increases the number of
function evaluations per second by factor 1.46 for the
shortest sequences, and by factor 3.91 for the longest sequence. This speed up is dependent on sequence length
and the relationship between the number of solution
evaluations inside and outside the local search. The local search also improves the algorithm’s convergence
speed for most of the sequences. Because of both advantages, the local search improves the efficiency of the
algorithm, and this improvement is greater for longer
sequences.
Using the best-known energy values as a stopping
condition, we demonstrated the usefulness of component reinitialization. It reduces the required mean number of solution evaluations to reach the best-known energy value from 0.38 to less than 0.05. This indicates
that the component reinitialization redirects the search
process successfully to similar solutions, and allows the
algorithm to locate the best-known solutions efficiently.
Our algorithm is the first algorithm that is capable of obtaining a hit ratio of 100 % for 6 shorter sequences within the budget of 1011 function evaluations.
Therefore, we introduce an approach for determining
Protein Folding Optimization using DE with Local Search and Component Reinitialization
●
0.0
0.2
0.4
0.6
RMSE
●
●
(a) 2ZNF
1.0
●
●
●
●
the best obtained solution
the obtained soluton
●
0.0
0.5
1.0
1.5
RMSE
(b) 1EDN
244
●
●
●
●●
● ●
●
●
●
0.0
the best obtained solution
the obtained soluton ● ●
0.2
0.4
0.6 0.8
RMSE
(c) 2H3S
●
●
●
●
●
●●
Energy (E)
240
242
●
●
●
0.8
●
●
20.90
21.30
●
the best obtained solution
the obtained soluton
21.25
●
●
●●
●
●
●
238
●
●
●
●●● ●
236
●
●
●
Energy (E)
21.00
●
20.80
●
●
●
21.10
21.45
●
Energy (E)
21.35 21.40
●
15
Energy (E)
20
25
●
15
●
1.0
●
●
●
●
●
●
●
● ●
● ●
● ●
●
●
●
● ● ●●
●
●
●●
● ●●● ●● ● ●
●
● ● ●● ●●
●
●
●●●
●
●
●
●
●
●
●
●
●
●● ● ● ●
●
● ●
●
●
● ●● ●●
● ●
● ●
●
●●
●●
●● ●
● ●●
●
●
●
●
●
●
●● ● ●
●
●
the best obtained solution
the obtained soluton
●
1.2
●
●
1.4
0.0
0.5
1.0
1.5
RMSE
2.0
2.5
(d) 2EWH
Fig. 6 Distribution of the Root-Mean-Square Error (RMSE) values as a function of energy for all 100 obtained conformations
within runtime limit of 4 days, calculated from the best-known conformation.
asymptotic average-case performances. Our algorithm
obtained the best runtime asymptotic average-case performance for sequence 2ZNF (16.6030 · 1.3505L ) and
the worst for 1BXP (0.0015 · 2.8911L ). This approach
shows additionally that the difficulty of the problem is
not only dependent on sequence length, but also on the
sequence itself.
The proposed algorithm was also compared with recently published state-of-the-art algorithms for PFO. It
outperforms all competitors, and the obtained energy
values improve the best-known energy values from the
literature for all sequences with L ≥ 18. For example,
the best energy value of sequence 1PCH was improved
from 131.7787 to 156.5250 or by 24.7463.
The structure of the best obtained solutions was also
analyzed. We figured out that there exist two symmetric best-known solutions for sequences with L ≤ 25.
For these sequences, our algorithm obtained a hit ratio
equal to or greater than 7 %. The solutions of these
sequences could be optimal, especially those with a hit
ratio of 100 %, and solutions for all other sequences are
almost surely not optimal.
In the future work we will try to improve the algorithm further by using knowledge about symmetric
solutions. This knowledge can be integrated within the
evaluation function, or used to reduce the size of the
search space. Additionally, we will try to design an algorithm that will reduce the likelihood of the exploration
of already explored search space.
Acknowledgements
Our work was supported by the Slovenian Research
Agency under Program P2-0041 (Computer Systems,
Methodologies, and Intelligent Services).
References
1. Alberts, B., Bray, D., Hopkin, K., Johnson, A.D., Lewis,
J., Raff, M., Roberts, K., Walter, P.: Essential Cell Biology, 4 edn. Garland Science (2013)
2. Balchin, D., Hayer-Hartl, M., Hartl, F.U.: In vivo aspects
of protein folding and quality control. Science 353(6294)
(2016). doi:10.1126/science.aac4354
3. Bazzoli, A., Tettamanzi, A.G.B.: A Memetic Algorithm
for Protein Structure Prediction in a 3D-Lattice HP
Model. In: G.R. Raidl, S. Cagnoni, J. Branke, D.W.
Corne, R. Drechsler, Y. Jin, C.G. Johnson, P. Machado,
E. Marchiori, F. Rothlauf, G.D. Smith, G. Squillero (eds.)
Applications of Evolutionary Computing, Lecture Notes
in Computer Science, vol. 3005, pp. 1–10. Springer Berlin
Heidelberg (2004). doi:10.1007/978-3-540-24653-4 1
4. Bošković, B., Brest, J.: Differential evolution for protein
folding optimization based on a three-dimensional AB
off-lattice model. Journal of Molecular Modeling 22, 1–
15 (2016). doi:10.1007/s00894-016-3104-z
5. Bošković, B., Brest, J.: Genetic Algorithm with Advanced Mechanisms Applied to the Protein Structure
Prediction in a Hydrophobic-Polar Model and Cubic
Lattice. Applied Soft Computing 45, 61–70 (2016).
doi:10.1016/j.asoc.2016.04.001
6. Brest, J., Greiner, S., Bošković, B., Mernik, M., Žumer,
V.: Self-Adapting Control Parameters in Differential Evolution: A Comparative Study on Numerical Benchmark
Problems. IEEE Trans. Evol. Comput 10(6), 646–657
(2006). doi:10.1109/TEVC.2006.872133
7. Buxbaum, E.: Fundamentals of Protein Structure and
Function. Springer US (2007)
8. Chen, M., Huang, W.q.: Heuristic algorithm for offlattice protein folding problem.
Journal of Zhejiang University SCIENCE B 7(1), 7–12 (2006).
doi:10.1631/jzus.2006.B0007
9. Chen, X., Lv, M., Zhao, L., Zhang, X.: An Improved Particle Swarm Optimization for Protein Folding Prediction. International Journal of Information
Engineering and Electronic Business 3(1), 1–8 (2011).
doi:10.5815/ijieeb.2011.01.01
10. Denning, P.J.: The Locality Principle. Commun. ACM
48(7), 19–24 (2005). doi:10.1145/1070838.1070856
11. Fraenkel, A.S.: Complexity of protein folding. Bulletin of Mathematical Biology 55(6), 1199–1210 (1993).
doi:10.1007/BF02460704
16
Borko Bošković, Janez Brest
Table 9 The best solutions obtained by the DElscr algorithm.
Label
1BXP
1BXP
Solution vector in degrees
43.2915, 2.88166, -48.728, 0.0655009, 12.6242, 66.0927, -6.40805, 8.96332, 8.80015, 2.23544, 74.0763, -6.62061, 1.31798, -104.099, 160.341,
-177.384, -20.6892, 26.8003, 127.789, 166.27, 10.2979
43.2915, 2.88166, -48.728, 0.0654961, 12.6242, 66.0927, -6.40805, 8.96332, 8.80016, 2.23544, 74.0763, 6.62061, -1.31798, 104.099, -160.341,
177.384, 20.6892, -26.8003, -127.789, -166.27, -10.2979
1CB3
-14.0758, 25.2546, -38.7359, -9.58086, 21.0366, 14.7617, -0.998265, 21.5393, 71.2738, -27.6012, -5.16526, -19.1483, -149.775, 172.54, 178.086,
178.164, 91.6772, 4.85452, -31.1093, 28.9806, 3.41538
1CB3
-14.0758, 25.2546, -38.7359, -9.58086, 21.0366, 14.7617, -0.998265, 21.5393, 71.2738, -27.6012, -5.16526, 19.1483, 149.775, -172.54, -178.086,
-178.164, -91.6772, -4.85451, 31.1093, -28.9806, -3.41538
1BXL
-22.4292, -32.2737, -16.9254, 5.81295, 15.6175, 26.9979, -38.2372, 52.8361, -48.2442, -24.0736, 49.3335, -36.1178, 13.9215, 12.5486, 1.91872,
55.1452, 147.302, -127.63, 168.592, -62.9624, -27.0891, 28.7221, 27.4283, 152.122, -177.152, 67.7357, -5.21217
1BXL
-22.4292, -32.2737, -16.9254, 5.81296, 15.6175, 26.9979, -38.2372, 52.8361, -48.2442, -24.0736, 49.3335, -36.1178, 13.9215, 12.5486, -1.91872,
-55.1452, -147.302, 127.63, -168.592, 62.9624, 27.0891, -28.7221, -27.4283, -152.122, 177.152, -67.7357, 5.21217
1EDP
-22.6336, 7.26974, 60.7674, 23.936, -50.4261, 4.41672, 11.4886, 46.499, 13.2306, -12.2668, 22.7087, 4.07035, 30.6245, -69.1251, 16.9542, 26.0209,
124.911, -155.575, -61.088, 1.55078, 53.7379, 159.421, -162.592, -156.44, -170.499, -85.1224, 2.36332, -25.7677, 67.3571
1EDP
-22.6336, 7.26974, 60.7674, 23.9359, -50.4261, 4.41671, 11.4886, 46.499,13.2306, -12.2668, 22.7087, 4.07036, 30.6245, -69.1251, 16.9542, -26.0209,
-124.911, 155.575, 61.088, -1.55077, -53.7379, -159.421, 162.592, 156.44, 170.499, 85.1224, -2.36333, 25.7677, -67.3571
2ZNF
-22.512, 7.71692, -75.1038, 26.0695, 35.539, 19.645, 6.73951, 21.8104, -57.4641, 1.6924, 6.15567, 3.08902, 9.89786, 23.8155, -48.9192, -4.31387,
78.7078, 2.66583, -114.943, -148.187, -162.564, -79.1176, 8.87759, -178.428, 42.9368, 15.8392, -18.6691, -104.193, 166.46, 12.876,
140.107
2ZNF
-22.512 7.71691, -75.1038, 26.0694, 35.539, 19.645, 6.73952, 21.8104, -57.4641, 1.6924, 6.15567, 3.08903, 9.89787, 23.8155, -48.9192, -4.31386,
-78.7078, -2.66582, 114.943, 148.187, 162.564, 79.1176, -8.87759, 178.428, -42.9368, -15.8392, 18.6691, 104.193, -166.46, -12.876,
-140.107
1EDN
-23.2048, 31.2208, 46.764, 48.9339, -43.6867, -28.0164, -17.6723, -38.3711, -25.1772, 10.6263, 9.07757, 33.5364, -4.83762, -6.09916, 25.0581, -81.151,
15.5944, -3.62479, -36.6783, 41.0025, 127.461, -147.732, -53.6249, -22.4102, -68.6344, -166.972, 147.028, -171.451, -155.381, 121.71,
29.6786, 131.144, 15.2983, 24.5428, -54.7787, -83.2637, -29.6805
1EDN
-23.2048, 31.2207, 46.764, 48.9338, -43.6867, -28.0164, -17.6723, -38.3711, -25.1772, 10.6262, 9.07754, 33.5364, -4.83764, -6.09917, 25.058, -81.151,
15.5944, -3.62479, -36.6783, -41.0025, -127.461, 147.732, 53.6249, 22.4102, 68.6344, 166.972, -147.028, 171.451, 155.381, -121.71,
-29.6786, -131.144, -15.2983, -24.5428, 54.7786, 83.2637, 29.6805
2H3S
30.6395, -51.1361, 34.4028, -0.410148, -32.439, -10.4102, -2.09416, 12.4798, -5.74202, -60.0842, 12.6704, -8.68551, -36.5963, -14.4828, -17.9172,
13.0794, 0.148026, 17.7335, -6.06512, 1.46386, -69.7022, 3.0363, 36.2347, -57.1061, -174.679, 173.256, -170.68, -156.724, 142.58, 40.6316,
22.5668, -1.4454, 175.849, -114.818, -61.1893, -4.11275, -27.6809, 84.4735, 144.867, 176.731, 161.605, -97.3254, -158.173, 113.225,
54.3451
2H3S
30.6395, -51.1361, 34.4028, -0.410146, -32.439, -10.4102, -2.09415, 12.4798, -5.74202, -60.0842, 12.6704, -8.68551, -36.5963, -14.4828, -17.9172,
13.0794, 0.148029, 17.7335, -6.06512, 1.46387, -69.7022, 3.0363, 36.2347, 57.1061, 174.679, -173.256, 170.68, 156.724, -142.58, -40.6317,
-22.5668, 1.44539, -175.849, 114.818, 61.1893, 4.11273, 27.6808, -84.4735, -144.867, -176.731, -161.605, 97.3254, 158.173,
-113.225, -54.3451
1ARE
-11.5547, -1.31888, -15.2569, -42.6589, 17.4763, 5.06218, 14.6442, 24.251, 2.4936, -13.6942, 28.8183, -37.0023, -1.8785, -0.867767, -5.02831, 35.1061,
-45.2208, -7.89329, 3.88011, -1.06756, -41.5237, 42.6134, 14.078, 1.71866, -70.4096, 19.7351, 23.088, -23.8173, -48.341, -27.8178, -175.825, 102.429,
138.697, -140.149, -46.4258, -148.917, -22.129, -165.228, 134.775, 48.8091, -12.7742, -50.1159, -163.389, 154.413, 126.397, -131.045, -60.8471,
167.884, 102.299, 52.433, -15.9838, -113.979, -58.9094
2KGU
-156.228, 84.3317, -1.89424, -22.9614, 4.96104, -10.8986, 42.0037, -54.9878, -4.36371, -80.394, 6.84565, -4.01855, -29.0786, 38.404, -24.9304, 51.317,
-53.2373, 15.7134, -51.9703, 1.34405, 37.6371, 36.5939, 35.6007, -52.9444, 32.6405, -108.259, -56.7621, 71.7249, 5.9403, 4.99762, 0.0626093, 8.48403,
-161.728, -140.31, 137.06, 46.113, 21.1367, 45.0214, -27.4148, 37.097, -8.18763, -148.71, 107.671, -141.471, -176.445, 152.171, -23.7168, -63.0744,
-154.472, 9.04166, -89.3673, 21.6149, -71.4051, 41.2427, -22.0274, 113.616, 22.7052, 159.166, -13.0884, -8.78814, 19.7018, 51.7085, 100.664
1TZ4
-13.2782, 2.2117, -21.4873, 13.5614, -50.9456, -18.6314, 58.273, 35.906, -51.557, 43.4606, 14.4093, 26.9361, -9.90087, 51.937, 12.5408, -12.0182,
-39.4559, -3.12819, -37.7837, 39.5619, 14.5525, -105.659, -2.39298, 23.2026, 13.2624, 7.00485, -63.913, 21.5608, -2.32347, -4.49988, 14.2846, -2.28795,
25.5405, -52.7743, -3.52791, 88.4618, 172.62, 63.4655, 167.01, -112.19, -129.059, 165.903, 161.114, -13.1829, -29.6599, -142.278, -118.354, -17.9561,
62.8846, 132.227, 150.392, 59.2519, -21.0203, -51.4332, 27.3873, 6.73164, 10.4224, -36.4599, -134.654, -177.842, -46.6888, 152.495, 60.1706, -2.02398,
-21.7416, 80.9012, 145.281, -5.47398, -93.0592
1TZ5
19.2916, -25.4743, 37.9748, -2.24909, -71.5607, -62.5534, -1.20914, 60.5119, -37.8385, 56.051, -23.4795, 88.5824, -23.4208, 9.88257, -27.3279, 2.64366,
4.72144, -32.8121, -37.9781, 28.6777, -1.79099, -1.45295, -1.55722, -28.7841, 53.5811, -7.1834, 28.1114, -34.8815, -63.6796, -10.3914, 19.0723,
-11.8679, -27.892, -37.0612, 62.2441, 52.8494, -172.157, 163.821, 57.7857, -56.1207, -158.293, 168.377, -23.9122, -20.1987, 4.84453, -2.90382,
-36.4027, -130.461, 158.501, 160.429, -146.987, -129.731, 128.236, 34.5188, 28.0746, -55.2047, -137.136, -167.33, 32.699, 35.1309, -64.3505, 35.5609,
22.6136, -27.8717, -112.09, 160.134, -131.465, -173.819, -165.595
1AGT
26.1664, -11.7847, -37.6783, -3.62257, 75.6542, -40.6386, 24.7245, -92.2268, -13.2317, 23.7859, -94.496, -2.2422, 17.0125, 1.77619, -39.6216, 113.613,
89.2441, -4.72246, -98.3805, -65.8427, 44.1904, -17.7537, -84.2295, -2.33273, -55.9952, -46.6065, -1.4073, 40.7682, -24.0458, 37.5641, 44.8005,
-1.59772, 9.28805, 12.0621, -52.5228, 21.2213, 120.34, -179.746, -63.7778, 18.8875, 10.4973, -6.76493, -21.1687, 57.2819, 176.355, 27.0186, -122.011,
-33.6896, -59.127, -178.6, 14.4172, -4.50249, -165.323, -6.2693, 12.6379, -54.0356, -62.0563, 7.63468, 127.685, 19.1897, 133.256, -157.687, -140.601,
-147.971, -19.8046, 48.3979, 166.316, 54.8275, 156.099, 3.00416, -140.869
1CRN
36.2044, -2.8726, -58.3456, -109.168, 67.0176, 63.8303, -34.4202, -9.13991, 27.0952, 6.87578, -175.904, -10.0323, -14.0642, 169.202, 139.54, 37.4324,
-30.2161, 3.47033, 120.567, 8.05841, 74.7893, -51.1755, -78.2244, -6.56336, -37.0077, -0.404044, 22.8391, -11.2627, -2.90337, -113.857, -122.645,
-5.62342, 80.5936, -19.4761, 87.9704, 12.3212, -4.15216, 3.25955, 39.0676, 30.3454, 61.9086, 12.9802, -97.5976, 8.44839, -76.9107, 30.3312, 41.3065,
24.8161, -34.1801, 20.2024, 33.4227, -14.1966, 80.1988, 28.2782, -166.287, -129.879, -11.8488, 19.3045, -66.9439, -48.3743, -164.938, -12.7803, 7.1584,
29.2233, 12.4003, 39.0045, -58.1443, 52.0717, 43.5443, -0.577219, -103.851, -146.583, -9.20535, -12.6377, -128.769, 27.4049, -32.4672, 16.7912,
-135.646, -149.836, 98.1608, 22.2995, 23.6538, 11.9513, 104.699, -3.17593, 35.9215
2KAP
46.469, 9.17062, -12.987, -39.265, -23.0536, 170.718, 7.46538, -139.561, 9.6654, -109.874, 39.7501, -77.1224, -8.16555, 82.4941, -21.6873, 93.4429,
-10.6347, 10.7423, 20.8323, -4.45369, -11.4409, -5.20606, 147.482, 172.455, -52.8927, 8.19366, 92.0005, -44.9143, -45.2074, -1.91882, -16.8158,
4.77317, 17.6662, 124.018, 11.9037, -1.64667, 74.9571, 15.2233, -5.48327, -140.99, 19.2716, 15.4203, -48.7429, -34.6525, 5.87344, -6.16017, 41.6324,
-16.8426, 49.2516, -28.6507, 29.562, -13.1308, 17.3443, -61.6342, 8.11077, -104.361, 26.4602, 4.19116, 18.9505, 67.8863, 154.188, -116.346, 19.3865,
-84.7317, -27.0054, -31.7152, -24.4805, -34.1621, -13.632, 63.0618, 8.2678, -2.60225, 37.1772, -9.33353, -46.953, -117.05, -167.457, 155.733, -103.59,
35.8403, 44.471, -168.41, -18.09, 28.324, 22.7177, -44.4081, -37.3318, -3.25116, 42.16, 57.0729, -4.27289, -157.026, 158.794, -19.8247, 37.3666,
142.492, -5.84281, 168.076, -163.22, -4.06088, -38.3009, 6.22285, -20.4278, 52.41, 156.856, 11.6928, 124.907, -164.531, 100.594, -172.86, -66.3312,
-55.802, 13.087, 38.8668, 61.5511
Protein Folding Optimization using DE with Local Search and Component Reinitialization
17
Table 10 The best solutions obtained by the DElscr algorithm.
1HVV
123.296, 175.071, -104.639, -8.91165, 17.898, 3.71452, 16.7371, -6.98533, -76.7403, -9.21776, 75.2042, 124.487, 46.9457, 18.594, -46.3806, 145.54,
124.742, -1.01034, 2.75056, 10.5349, -7.0204, 118.317, 75.4052, -30.5719, 69.7023, -72.9042, 176.87, -87.9407, -48.7546, 88.4046, -11.7349, 44.1027,
-5.14578, -65.6298, 20.3053, -20.9559, -72.1726, 94.427, 45.6623, -45.1943, 55.0226, 121.96, 34.0003, 6.44231, 140.467, 17.4955, 85.266, -21.9109,
-95.2265, -24.295, 2.64738, -17.4667, 155.115, -5.99421, -158.933, -6.23019, -43.4673, -28.7354, -11.6413, -19.2361, -10.0374, 34.4593, -178.973,
-160.129, -42.443, 5.40211, -87.2981, -36.4155, 10.9084, 64.5935, -5.14139, 28.5554, -18.3091, -3.52734, 45.1155, 9.74914, 114.875, 75.0862, 26.4021,
22.7921, -63.2735, -138.233, -64.6958, -1.96622, 99.8458, -16.8097, 71.6329, 28.1128, -43.4581, -82.0519, 173.534, -96.2117, -16.7565, -32.4898,
48.5607, 11.8749, -64.2088, 1.38784, -57.9374, 25.3469, 54.5768, 34.603, -28.9913, -102.93, -64.0207, -23.3654, 157.688, 118.979, 5.08255, 61.9001,
-18.6156, -45.7361, -10.7387, 47.8304, -26.517, -104.738, 22.5679, 115.444, -6.44774, 116.951, 14.5305, -0.364305, 25.7291, 114.23, -22.2869, 97.1472,
0.163606, -93.1095, -14.0922, 67.6929, 103.708, 12.5582, -62.8467, -103.687, -55.8599, 11.7904, 63.4791, 74.9204, 26.812, -72.0632, -116.517, -46.752,
39.4218, 63.2466, 101.367
1GK4
-156.276, -11.1754, -77.6765, -33.983, -9.25845, 27.2496, -22.3371, 55.198, -12.6002, 30.4017, 35.9669, -18.5758, -87.0091, 63.2376, 163.731,
-40.6359, 102.397, 146.119, -16.2788, 0.676412, 169.967, -18.769, -43.9974, -62.1645, -51.3378, -148.518, 144.287, -170.774, -25.6785, -22.1326,
27.6078, -8.93322, -75.4448, -68.3403, -16.5373, -128.752, 53.7389, -140.548, -115.598, -1.7515, -106.07, -14.6299, 82.7281, -28.3651, 36.138, 37.748,
-23.152, -11.6747, 20.0652, -55.0055, -150.457, -49.5769, 23.078, 65.4316, 111.217, 34.2085, 8.50149, 21.775, -107.598, 139.532, -69.7933, 68.8643,
88.8784, -166.542, -174.708, 31.9934, 137.315, -3.2063, -88.0773, -6.44718, 23.1364, 95.4164, -34.9923, 2.63385, 13.9679, -23.3478, -54.8455, 43.5591,
-4.20436, -43.6293, -144.761, 18.6453, 179.845, -29.3419, 34.6059, 116.989, -2.74396, 108.058, 157.394, 13.5671, -34.673, -71.5358, 29.1635, -12.3351,
-58.4977, -28.7525, -42.495, -47.6578, 29.0888, -101.645, -92.0769, 10.8024, 101.05, 70.2237, 17.3966, -48.5046, 27.4408, 61.4589, 71.0815, 77.6142,
36.2921, -18.009, -121.503, -122.087, -32.2391, 9.22399, 11.065, 70.686, 48.1796, -38.5792, -91.5736, 15.5793, -173.479, -25.563, 108.153, 54.9784,
-50.888, -86.9754, -52.2761, -23.2842, 20.5012, 16.6481, -23.1703, -40.3118, -64.6739, 26.9541, 177.66, 154.029, 125.888, 2.61904, -11.91, -67.2757,
7.45577, 48.5355, -2.84037, -62.5508, -1.94643, 7.24978, 48.2981, 62.3802, 127.482, 80.9534, -6.93496, -105.181, -59.8353, -45.6067, -113.32, 1.58149,
-33.4263, -8.31888, 28.5097, -8.75938, 151.901
1PCH
41.1106, -75.4478, -108.35, -88.2808, 52.6905, 63.9346, 98.575, -146.011, -103.251, -81.3915, -7.97873, -5.67391, -138.034, 77.0572, -3.43666,
-168.458, -44.1355, -51.9614, -0.548426, -11.6916, -13.85, -90.7331, 162.157, -85.119, -80.3963, 2.41355, -4.42424, 54.2753, 89.5966, -147.035, 78.814,
99.0202, -89.7149, -126.891, -85.1937, 16.9404, -73.4292, 66.5445, 160.832, 40.1368, 145.935, 17.6023, -17.2377, 45.6224, 136.733, -49.815, -43.4151,
-102.65, 8.11703, -109.171, -24.407, -47.2699, -132.484, -42.5279, -119.761, 147.911, 73.3968, 3.65598, 74.2541, 19.778, -2.2315, -111.056, -25.0304,
-86.2442, 0.146651, 83.5428, -29.3382, 64.0093, 147.03, 73.9582, 29.3652, 154.906, 6.25221, 112.96, 165.03, 11.5624, -48.3887, -130.92, 158.099,
-6.47314, -26.6561, -18.5458, -79.5949, 16.5753, 68.4287, 54.4969, 8.59587, -21.8077, -38.0325, -14.1827, -48.149, -7.71342, 5.32628, 34.3591, 4.97453,
-19.9588, -29.6581, -39.8474, 152.45, -7.02004, -71.166, 40.6025, 90.9133, 41.1172, 10.2277, -41.2978, -21.4705, -9.53513, 33.3237, -170.287, -147.144,
-21.8209, 27.7537, 19.2982, 29.7286, -13.7392, -55.8712, -28.3766, -8.00291, 125.209, 161.205, 19.9805, -46.2199, -2.73427, -12.1052, 106.535, 20.4942,
-28.1138, 4.59628, 62.9273, -0.534962, 6.33028, -7.44623, 29.7984, -42.8088, 5.32914, 17.9307, 174.593, -36.0277, 3.2934, 24.3011, 174.231, -1.94325,
107.108, -4.82491, 44.1005, -38.1958, 47.1931, 0.689977, -48.5599, -2.29967, -16.9223, -9.36993, -13.2723, 23.7414, 17.0389, 157.955, 52.6113,
14.0859, -100.053, -42.2958, -9.25703, 37.9416, 58.9846, -28.9415, -56.1025, 37.1207, 50.4847, -55.062
2EWH
151.436, -92.5903, -3.75076, -9.668, -0.975246, 10.9844, 84.1865, -57.1686, 38.7048, 72.7755, 64.7086, 50.2283, 29.5956, -111.649, 163.46, -92.508,
-148.929, 61.6607, 124.085, 171.298, -75.5506, -52.2331, -1.37142, -19.9511, 67.108, 14.27, -1.03143, 0.0222875, -116.144, -32.3217, -125.421,
-102.86, 145.85, 108.298, -60.4006, -54.4374, -88.1894, -14.6692, 121.958, 125.926, 167.02, -74.5737, -130.938, 62.1354, 106.58, 35.1401, 93.7873,
162.441, 14.5873, -4.93593, 6.27325, -7.31527, 10.8045, -53.8271, -132.021, -37.8517, 30.6805, -89.4307, -61.2859, -31.1104, 90.5217, 118.145,
-4.46292, -52.5412, 113.612, 159.141, -3.43442, 62.945, 12.7417, -19.7111, -15.1737, 30.368, -27.5934, -138.544, 7.81197, -59.2248, 9.7981, 122.127,
-164.755, 36.2949, 27.611, -37.779, 39.5707, -22.2883, 38.9922, 4.30684, 71.8969, -21.7556, -128.587, -76.5211, -0.480596, 65.3271, 8.48192, 158.34,
107.055, -66.142, 18.5806, -129.465, -18.4294, -11.7806, 50.9098, 132.39, 101.434, 0.100595, -28.2695, 0.622333, 52.5959, 150.715, -11.0556, -45.8273,
-39.1155, 8.30602, 56.8165, -46.0845, 17.9186, 11.1667, 19.5044, 89.7449, -23.746, -0.808862, 84.8186, 0.457484, -89.925, -21.9539, 26.9255, -59.9087,
45.3057, -9.74849, -70.462, -14.0544, 18.7807, 59.6121, -176.746, -30.6629, -54.0245, -8.68261, -14.3952, 43.302, 23.9289, 59.5115, -35.76, -58.2711,
2.27009, -152.857, -35.6967, -137.17, -49.407, 36.9083, 58.3246, -26.7135, -71.9911, 26.5305, 53.0114, 52.0147, -21.6699, -14.1031, 12.009, -63.6288,
-54.0997, -10.6967, 5.20745, -124.069, -19.2082, 55.6145, -3.07942, 12.0508, 131.359, -177.867, 32.7929, 24.0475, 25.0758, -0.0780254, 5.44473,
36.0356, 28.5948, 22.5261, -174.132, -3.91424, -11.5903, -132.095, -172.313, -25.055, -156.073, 11.1264, 29.3397, -64.3192, -38.8408, -177.009,
22.2676, 11.9142, 56.3727
F13
7.66522, -83.448, 13.0886, 0.55134, 29.1616, -47.908, 2.75327, -31.0327, -31.3119, -46.3918, 0.276218, 9.04884, -29.5745, -116.199, 160.508,
0.890189, 129.381, 24.5074, 113.38, -161.672, 98.7127
F13
7.66522, -83.448, 13.0886, 0.551338, 29.1616, -47.908, 2.75327, -31.0327, -31.3119, -46.3918, 0.276222, -9.04884, 29.5745, 116.199, -160.508,
-0.890189, -129.381, -24.5074, -113.38, 161.672, -98.7127
F21
-5.70817, -70.6345, 12.6013, -78.4561, 5.14012, 2.49148, 57.5974, -25.416, 27.2287, -35.8677, -5.33428, -13.9895, 3.02158, 19.9054, 74.4006, -31.0707,
4.76465, -19.1022, -32.9492, 155.506, -16.0013, -169.101, 162.893, -94.9124, 155.503, -140.891, 153.332, 40.6752, 137.563, 48.1957,
-35.2245, 66.7533, -37.5734, 137.909, -144.521, -52.7295, -156.871
F21
-5.70816, -70.6345, 12.6014, -78.4561, 5.14014, 2.49149, 57.5974, -25.416, 27.2287, -35.8677, -5.33427, -13.9895, 3.0216, 19.9055, 74.4006, -31.0707,
4.76466, -19.1022, -32.9492, -155.506, 16.0013, 169.101, -162.893, 94.9124, -155.503, 140.891, -153.332, -40.6752, -137.563, -48.1957,
35.2245, -66.7533, 37.5734, -137.909, 144.521, 52.7295, 156.871
F34
12.3298, -83.1718, 20.1532, 8.42606, 37.8998, -37.8448, 9.33408, -77.8143, 7.4245, -73.1774, 26.15, -80.0668, 46.3843, 6.49943, -29.8816, 51.2622,
-33.6564, 38.6885, -67.9543, 46.7986, -10.4886, -27.9647, -10.0583, -39.8364, -49.6972, -25.641, 44.7456, -59.6061, 18.6305, -20.9127, 25.4877,
13.4228, 1.77009, 42.1284, 129.207, -149.941, 1.89517, -120.166, 18.4003, 159.01, -168.548, 143.358, 151.62, -49.9323, -164.471, -44.6816, 177.501,
-32.6178, 2.86468, -2.00479, -22.1516, -57.0231, -143.09, 131.37, -127.956, 147.157, 57.657, -21.2642, 27.2822, -52.9505, 17.7835, 119.254, 18.7327
F55
-15.6437, 97.9193, 1.00666, 95.3815, 1.86855, -64.0331, -141.452, -2.83476, 104.146, 8.21281, -162.93, -74.3953, 1.96392, 7.65968, -29.2495, 52.5953,
52.8264, -0.624594, 137.07, -4.89079, 0.957561, 150.771, 19.388, 7.34186, 59.4269, 8.22775, -64.6383, -54.8633, -8.8461, 59.752, 162.033, 13.6066,
-78.2664, 13.0242, 102.375, 3.23899, -2.60196, -16.3626, 36.9652, -37.8734, 30.0569, 3.86882, -34.6667, -22.5344, 25.3408, 89.0776, 16.5037, -17.6911,
-91.108, 4.84917, -9.27247, -4.88184, 6.63221, 4.31554, -28.558, -8.46761, 171.776, -66.1542, 29.7446, -114.307, -0.574113, 83.8969, -5.34284, 64.1091,
6.16523, 112.965, -5.20239, 68.6776, -17.6844, -113.527, 31.5684, -100.963, 152.521, -49.1181, -26.2181, -129.399, -24.1061, 42.4841, -45.0845,
67.3054, -35.1296, 15.741, 34.8919, 33.4461, 7.88754, 179.833, 41.9367, -135.639, 118.444, -156.824, -55.2589, 169.967, 0.172651, 91.1955, -4.02574,
-81.307, 23.9152, -29.0528, -154.072, -133.971, 16.3457, -93.6495, -5.71876, 80.2119, -21.5927
F89
179.886, -95.54, -17.2583, -27.4664, -23.3,38.1298, 19.5766, 34.4132, 8.84695, -117.344, -20.5596, -131.943, -11.8962, -48.6976, 129.133, -107.932,
7.29644, 89.0642, 21.8161, 56.2631, -33.6006, -22.3247, -47.5945, -48.8527, -51.6802, 36.4829, 85.1461, 3.54137, 107.149, -0.955333, -51.2341,
65.8004, 13.8676, -69.7918, -0.860014, -134.319, -37.8356, -2.74527, -12.5366, -93.6285, -28.5384, -105.157, 19.0699, 81.6706, 6.93831, 0.887398,
116.484, 23.1153, 132.738, 10.4558, 73.2237, -15.4114, 6.88586, -109.859, -3.3155, -82.7065, 2.76043, -42.4804, 82.5479, -18.3209, -29.9615, -74.7318,
-24.0277, 60.1736, -26.3071, -15.531, -14.9412, -79.5093, -1.99245, -48.5295, -70.5006, 58.6443, -42.9465, 50.0326, -70.0616, -9.55698, -109.482,
-2.75044, -87.6997, -8.8569, 86.5537, -22.4479, 72.1052, 14.0501, -27.4652, -8.8744, 56.3962, 0.863411, -142.989, -54.5377, 29.0611, -61.1795, 50.3774,
-12.8387, 73.4752, 12.0947, -39.6898, -28.42, 143.035, 28.1471, 39.6651, 10.0519, -140.34, -2.35037, 123.344, 3.62448, 125.741, 132.141, 71.1956,
-36.3432, -36.7204, -39.5973, 57.8245, -31.8281, 13.6268, -143.946, -36.4178, -6.53297, 34.1645, 12.4669, -82.0619, 14.2377, -32.9623, 49.1945,
137.212, -16.0272, -178.526, 12.3581,69.7334, -1.88293, 147.327, 145.168, 27.4845, -35.3688, 8.48146, -81.2594, -5.25881, 119.388, -139.654, 57.454,
159.88, 27.509, -0.616009, 114.258, -13.2053, -39.494, 65.2585, -42.1881, 0.311959, -22.2436, -162.899, -54.8573, -20.7022, -14.7214, 128.993, 5.42026,
114.069, -21.3562, -46.7891, 18.6434, 15.3431, 121.287, -3.95967, -82.6683, -9.4711, -120.912, -1.33882, -28.956, 43.9338, -42.9642, -139.445, 137.938,
4.62324
18
Borko Bošković, Janez Brest
(a) 1BXP
(b) 1BXP
(c) 1CB3
(d) 1CB3
(e) 1BXL
(f) 1BXL
(g) 1EDP
(h) 1EDP
(i) 2ZNF
(j) 2ZNF
(k) 1EDN
(l) 1EDN
(m) 2H3S
(n) 2H3S
(o) F13
(p) F13
(q) F21
Fig. 7 The best obtained conformations that could be optimal.
(r) F21
Protein Folding Optimization using DE with Local Search and Component Reinitialization
12. Hsu, H.P., Mehra, V., Grassberger, P.: Structure optimization in an off-lattice protein model. Phys. Rev. E
68 (2003). doi:10.1103/PhysRevE.68.037703
13. Hsu, H.P., Mehra, V., Nadler, W., Grassberger, P.:
Growth algorithms for lattice heteropolymers at low temperatures. The Journal of Chemical Physics 118(1), 444–
451 (2003). doi:10.1063/1.1522710
14. Huang, W., Liu, J.: Structure optimization in a threedimensional off-lattice protein model.
Biopolymers
82(2), 93–98 (2006). doi:10.1002/bip.20400
15. Jana, N.D., Sil, J., Das, S.: An Improved Harmony Search
Algorithm for Protein Structure Prediction Using 3D OffLattice Model, pp. 304–314. Springer Singapore, Singapore (2017). doi:10.1007/978-981-10-3728-3 30
16. Jana, N.D., Sil, J., Das, S.: Selection of appropriate metaheuristic algorithms for protein structure prediction in
{AB} off-lattice model: a perspective from fitness landscape analysis. Information Sciences 391392, 28–64
(2017). doi:10.1016/j.ins.2017.01.020
17. Kennedy, D., Norman, C.: Editorial: So much
more to know.
Science 309, 78–102 (2005).
doi:10.1126/science.309.5731.78b
18. Kim, J., Straub, J.E., Keyes, T.: Structure optimization
and folding mechanisms of off-lattice protein models using statistical temperature molecular dynamics simulation: Statistical temperature annealing. Phys. Rev. E 76
(2007). doi:10.1103/PhysRevE.76.011913
19. Kim, S.Y.: Three-dimensional off-lattice ab model protein with the 89-residue fibonacci sequence. Chaos,
Solitons & Fractals 90(Supplement C), 111–117 (2016).
doi:10.1016/j.chaos.2016.04.010
20. Kim, S.Y., Lee, S.B., Lee, J.: Structure optimization by conformational space annealing in an offlattice protein model.
Phys. Rev. E 72 (2005).
doi:10.1103/PhysRevE.72.011916
21. Lau, K.F., Dill, K.A.: A lattice statistical mechanics
model of the conformational and sequence spaces of
proteins. Macromolecules 22(10), 3986–3997 (1989).
doi:10.1021/ma00200a030
22. Li, B., Chiong, R., Lin, M.: A balance-evolution artificial bee colony algorithm for protein structure optimization based on a three-dimensional AB off-lattice model.
Computational Biology and Chemistry 54, 1–12 (2015).
doi:10.1016/j.compbiolchem.2014.11.004
23. Li, B., Lin, M., Liu, Q., Li, Y., Zhou, C.: Protein folding
optimization based on 3D off-lattice model via an improved artificial bee colony algorithm. Journal of Molecular Modeling 21(10), 261 (2015). doi:10.1007/s00894015-2806-y
24. Li, Y., Zhou, C., Zheng, X.: The Application of Artificial Bee Colony Algorithm in Protein Structure Prediction. In: L. Pan, G. Pun, M. Prez-Jimnez, T. Song
(eds.) Bio-Inspired Computing - Theories and Applications, Communications in Computer and Information
Science, vol. 472, pp. 255–258. Springer Berlin Heidelberg (2014). doi:10.1007/978-3-662-45049-9 42
25. Márquez-Chamorroa,
A.E.,
Asencio-Cortésa,
G.,
Santiesteban-Tocab, C.E., Aguilar-Ruiza, J.S.: Soft
Computing Methods for the Prediction of Protein
Tertiary Structures: A Survey. Applied Soft Computing
35, 398–410 (2015). doi:10.1016/j.asoc.2015.06.024
26. Mount, D.W.: Bioinformatics: Sequence and Genome
Analysis. New York: Cold Spring Harbor Laboratory
Press (2001)
27. Petsko, G., Ringe, D.: Protein Structure and Function.
Primers in biology. New Science Press (2004)
19
28. PJ, T., BH, Q., PL, P.: Defective protein folding as a
basis of human disease 20, 456–459 (1995)
29. Selkoe, D.J.: Folding proteins in fatal ways. Nature pp.
900–904 (2003). doi:10.1038/nature02264
30. Stillinger, F.H., Head-Gordon, T.: Collective aspects of
protein folding illustrated by a toy model. Phys. Rev. E
52, 2872–2877 (1995). doi:10.1103/PhysRevE.52.2872
31. Stillinger, F.H., Head-Gordon, T., Hirshfeld, C.L.: Toy
model for protein folding. Phys. Rev. E 48, 1469–1477
(1993). doi:10.1103/PhysRevE.48.1469
32. Storn, R., Price, K.: Differential Evolution - a simple
and efficient adaptive scheme for global optimization over
continuous spaces. Tech. Rep. TR-95-012, Berkeley, CA
(1995)
33. Storn, R., Price, K.: Differential Evolution - A Simple and
Efficient Heuristic for Global Optimization Over Continuous Spaces. J. Glob. Optim. 11(4), 341–359 (1997)
34. Thachuk, C., Shmygelska, A., Hoos, H.H.: A replica
exchange monte carlo algorithm for protein folding in
the hp model. BMC Bioinformatics 8(1), 342 (2007).
doi:10.1186/1471-2105-8-342
35. Wang, T., Zhang, X.: 3D Protein Structure Prediction
with Genetic Tabu Search Algorithm in Off-Lattice AB
Model. In: Knowledge Acquisition and Modeling, 2009.
KAM ’09. Second International Symposium on, vol. 1,
pp. 43–46 (2009). doi:10.1109/KAM.2009.2
36. Wang, T., Zhang, X.: A case study of 3D protein structure prediction with genetic algorithm and Tabu search.
Wuhan University Journal of Natural Sciences 16(2),
125–129 (2011). doi:10.1007/s11859-011-0723-1
37. Wang, Y., Guo, G., Chen, L.: Chaotic Artificial Bee
Colony algorithm: A new approach to the problem of minimization of energy of the 3D protein
structure. Molecular Biology 47(6), 894–900 (2013).
doi:10.1134/S0026893313060162
38. Wong, K.C., Wu, C.H., Mok, R.K.P., Peng, C., Zhang,
Z.: Evolutionary multimodal optimization using the
principle of locality. Inf. Sci. 194, 138–170 (2012).
doi:10.1016/j.ins.2011.12.016
39. Zhang, X., Cheng, W.: Protein 3D Structure Prediction by Improved Tabu Search in Off-Lattice AB Model.
In: Bioinformatics and Biomedical Engineering, 2008.
ICBBE 2008. The 2nd International Conference on, pp.
184–187 (2008). doi:10.1109/ICBBE.2008.50
40. Zhang, X., Wang, T., Luo, H., Yang, J.Y., Deng, Y.,
Tang, J., Yang, M.Q.: 3D Protein structure prediction
with genetic tabu search algorithm. BMC Systems Biology 4, 1–9 (2010). doi:10.1186/1752-0509-4-S1-S6
41. Zhou, C., Hou, C., Wei, X., Zhang, Q.: Improved hybrid
optimization algorithm for 3D protein structure prediction. Journal of Molecular Modeling 20(7), 2289 (2014).
doi:10.1007/s00894-014-2289-2
42. Zhou, C., Hu, T., Zhou, S.: Protein Structure Prediction Based on Improved Multiple Populations and GAPSO. In: L. Pan, G. Pun, M. Prez-Jimnez, T. Song
(eds.) Bio-Inspired Computing - Theories and Applications, Communications in Computer and Information
Science, vol. 472, pp. 644–647. Springer Berlin Heidelberg (2014). doi:10.1007/978-3-662-45049-9 105
| 2cs.AI
|
1
A Bayesian Approach to Policy Recognition
and State Representation Learning
arXiv:1605.01278v4 [stat.ML] 4 Aug 2017
Adrian Šošić, Abdelhak M. Zoubir and Heinz Koeppl
Abstract—Learning from demonstration (LfD) is the process of building behavioral models of a task from demonstrations provided by
an expert. These models can be used e.g. for system control by generalizing the expert demonstrations to previously unencountered
situations. Most LfD methods, however, make strong assumptions about the expert behavior, e.g. they assume the existence of a
deterministic optimal ground truth policy or require direct monitoring of the expert’s controls, which limits their practical use as part of a
general system identification framework. In this work, we consider the LfD problem in a more general setting where we allow for
arbitrary stochastic expert policies, without reasoning about the optimality of the demonstrations. Following a Bayesian methodology,
we model the full posterior distribution of possible expert controllers that explain the provided demonstration data. Moreover, we show
that our methodology can be applied in a nonparametric context to infer the complexity of the state representation used by the expert,
and to learn task-appropriate partitionings of the system state space.
Index Terms—learning from demonstration, policy recognition, imitation learning, Bayesian nonparametric modeling, Markov chain
Monte Carlo, Gibbs sampling, distance dependent Chinese restaurant process
F
1
L
I NTRODUCTION
(LfD) has become a
viable alternative to classical reinforcement learning as
a new data-driven learning paradigm for building behavioral models based on demonstration data. By exploiting
the domain knowledge provided by an expert demonstrator,
LfD-built models can focus on the relevant parts of a system’s state space [1] and hence avoid the need of tedious exploration steps performed by reinforcement learners, which
often require an impractically high number of interactions
with the system environment [2] and always come with the
risk of letting the system run into undesired or unsafe states
[3]. In addition to that, LfD-built models have been shown
to outperform the expert in several experiments [4], [5], [6].
However, most existing LfD methods come with strong
requirements that limit their practical use in real-world scenarios. In particular, they often require direct monitoring of
the expert’s controls (e.g. [5], [7], [8]) which is possible only
under laboratory-like conditions, or they need to interact
with the target system via a simulator, if not by controlling
the system directly (e.g. [9]). Moreover, many methods are
restricted to problems with finite state spaces (e.g. [10]), or
they compute only point estimates of the relevant system
parameters without providing any information about their
level of confidence (e.g. [9], [11], [12]). Last but not least, the
expert is typically assumed to follow an optimal deterministic policy (e.g. [13]) or to at least approximate one, based
on some presupposed degree of confidence in the optimal
•
•
•
EARNING FROM DEMONSTRATION
Adrian Šošić is a member of the Signal Processing Group and an associate
member of the Bioinspired Communication Systems Lab, Technische Universität Darmstadt, Germany. E-mail: [email protected]
Abdelhak M. Zoubir is the head of the Signal Processing Group, Technische
Universität Darmstadt, Germany. E-mail: [email protected]
Heinz Koeppl is the head of the Bioinspired Communication Systems Lab
and a member of the Centre for Cognitive Science, Technische Universität
Darmstadt, Germany. E-mail: [email protected]
behavior (e.g. [14]). While such an assumption may be
reasonable in some situations (e.g. for problems in robotics
involving a human demonstrator [1]), it is not appropriate
in many others, such as in multi-agent environments, where
an optimal deterministic policy often does not exist [15].
In fact, there are many situations in which the assumption
of a deterministic expert behavior is violated. In a more
general system identification setting, our goal could be, for
instance, to detect the deviation of an agent’s policy from
its known nominal behavior, e.g. for the purpose of fault or
fraud detection (note that the term “expert” is slightly misleading in this context). Also, there are situations in which
we might not want to reason about the optimality of the
demonstrations; for instance, when studying the exploration
strategy of an agent who tries to model its environment (or
the reactions of other agents [16]) by randomly triggering
different events. In all these cases, existing LfD methods
can at best approximate the behavior of the expert as they
presuppose the existence of some underlying deterministic
ground truth policy.
In this work, we present a novel approach to LfD in order
to address the above-mentioned shortcomings of existing
methods. Central to our work is the problem of policy
recognition, that is, extracting the (possibly stochastic and
non-optimal) policy of a system from observations of its
behavior. Taking a general system identification view on the
problem, our goal is herein to make as few assumptions
about the expert behavior as possible. In particular, we consider the whole class of stochastic expert policies, without
ever reasoning about the optimality of the demonstrations.
As a result of this, our hypothesis space is not restricted to a
certain class of ground truth policies, such as deterministic
or softmax policies (c.f. [14]). This is in contrast to inverse
reinforcement learning approaches (see Section 1.2), which
interpret the observed demonstrations as the result of some
preceding planing procedure conducted by the expert which
2
they try to invert. In the above-mentioned case of fault
detection, for example, such an inversion attempt will generally fail since the demonstrated behavior can be arbitrarily
far from optimal, which renders an explanation of the data
in terms of a simple reward function impossible.
Another advantage of our problem formulation is that
the resulting inference machinery is entirely passive, in the
sense that we require no active control of the target system
nor access to the action sequence performed by the expert.
Accordingly, our method is applicable to a broader range of
problems than targeted by most existing LfD frameworks
and can be used for system identification in cases where
we cannot interact with the target system. However, our
objective in this paper is twofold: we not only attempt to
answer the question how the expert performs a given task
but also to infer which information is used by the expert to
solve it. This knowledge is captured in the form of a joint
posterior distribution over possible expert state representations and corresponding state controllers. As the complexity
of the expert’s state representation is unknown a priori,
we finally present a Bayesian nonparametric approach to
explore the underlying structure of the system space based
on the available demonstration data.
1.1 Problem statement
Given a set of expert demonstrations in the form of a system
trajectory s = (s1 , s2 , . . . , sT ) ∈ S T of length T , where
S denotes the system state space, our goal is to determine
the latent control policy used by the expert to generate the
state sequence.1 We formalize this problem as a discretetime decision-making process (i.e. we assume that the expert
executes exactly one control action per trajectory state) and
adopt the Markov decision process (MDP) formalism [17]
as the underlying framework describing the dynamics of
our system. More specifically, we consider a reduced MDP
(S, A, T , π) which consists of a countable or uncountable
system state space S , a finite set of actions A containing
|A| elements, a transition model T : S × S × A → R≥0
where T (s0 | s, a) denotes the probability (density) assigned
to the event of reaching state s0 after taking action a in
state s, and a policy π modeling the expert’s choice of
actions.2 In the following, we assume that the expert policy
is parametrized by a parameter ω ∈ Ω, which we call
the global control parameter of the system, and we write
π(a | s, ω), π : A × S × Ω → [0, 1], to denote the expert’s
local policy (i.e. the distribution of actions a played by the
expert) at any given state s under ω . The set Ω is called the
parameter space of the policy, which specifies the class of
feasible action distributions. The specific form of Ω will be
discussed later.
Using a parametric description for π is convenient as
it shifts the recognition task from determining the possibly
infinite set of local policies at all states in S to inferring the
posterior distribution p(ω | s), which contains all information that is relevant for predicting the expert behavior,
Z
p(a | s∗ , s) =
π(a | s∗ , ω)p(ω | s) dω.
Ω
1. The generalization to multiple trajectories is straightforward as
they are conditionally independent given the system parameters.
2. This reduced model is sometimes referred to as an MDP\R (see
e.g. [9], [18], [19]) to emphasize the nonexistence of a reward function.
Herein, s∗ ∈ S is some arbitrary query point and p(a | s∗ , s)
is the corresponding predictive action distribution. Since the
local policies are coupled through the global control parameter ω as indicated by the above integral equation, inferring
ω means not only to determine the individual local policies
but also their spatial dependencies. Consequently, learning
the structure of ω from demonstration data can be also
interpreted as learning a suitable state representation for
the task performed by the expert. This relationship will be
discussed in detail in the forthcoming sections. In Section 3,
we further extend this reasoning to a nonparametric policy
model whose hypothesis class finally covers all stochastic
policies on S .
For the remainder of this paper, we make the common
assumptions that the transition model T as well as the
system state space S and the action set A are known. The
assumption of knowing S follows naturally because we
already assumed that we can observe the expert acting in S .
In the proposed Bayesian framework, the latter assumption
can be easily relaxed by considering noisy or incomplete trajectory data. However, as this would not provide additional
insights into the main principles of our method, we do not
consider such an extension in this work.
The assumption of knowing the transition dynamics T is
a simplifying one but prevents us from running into model
identifiability problems: if we do not constrain our system
transition model in some reasonable way, any observed state
transition in S could be trivially explained by a corresponding local adaptation of the assumed transition model T
and, thus, there would be little hope to extract the true expert policy from the demonstration data. Assuming a fixed
transition model is the easiest way to resolve this model
ambiguity. However, there are alternatives which we leave
for future work, for example, using a parametrized family of
transition models for joint inference. This extension can be
integrated seamlessly into our Bayesian framework and is
useful in cases where we can constrain the system dynamics
in a natural way, e.g. when modeling physical processes.
Also, it should be mentioned that we can tolerate deviations
from the true system dynamics as long as our model T is
sufficiently accurate to extract information about the expert
action sequence locally, because our inference algorithm
naturally processes the demonstration data piece-wise in
the form of one-step state transitions {(st , st+1 )} (see algorithmic details in Section 2 and results in Section 4.2). This
is in contrast to planning-based approaches, where small
modeling errors in the dynamics can accumulate and yield
consistently wrong policy estimates [8], [20].
The requirement of knowing the action set A is less
stringent: if A is unknown a priori, we can still assume a potentially rich class of actions, as long as the transition model
can provide the corresponding dynamics (see example in
Section 4.2). For instance, we might be able to provide a
model which describes the movement of a robotic arm even
if the maximum torque that can be generated by the system
is unknown. Figuring out which of the hypothetical actions
are actually performed by the expert and, more importantly,
how they are used in a given context, shall be the task of
our inference algorithm.
3
1.2
Related work
The idea of learning from demonstration has now been
around for several decades. Most of the work on LfD has
been presented by the robotics community (see [1] for a
survey), but recent advances in the field have triggered developments in other research areas, such as cognitive science
[21] and human-machine interaction [22]. Depending on the
setup, the problem is referred to as imitation learning [23],
apprenticeship learning [9], inverse reinforcement learning
[13], inverse optimal control [24], preference elicitation [21],
plan recognition [25] or behavioral cloning [5]. Most LfD
models can be categorized as intentional models (with
inverse reinforcement learning models as the primary example), or sub-intentional models (e.g. behavioral cloning
models). While the latter class only predicts an agent’s
behavior via a learned policy representation, intentional
models (additionally) attempt to capture the agent’s beliefs
and intentions, e.g. in the form of a reward function. For this
reason, intentional models are often reputed to have better
generalization abilities3 ; however, they typically require a
certain amount of task-specific prior knowledge in order
to resolve the ambiguous relationship between intention
and behavior, since there are often many ways to solve
a certain task [13]. Also, albeit being interesting from a
psychological point of view [21], intentional models target
a much harder problem than what is actually required in
many LfD scenarios. For instance, it is not necessary to
understand an agent’s intention if we only wish to analyze
its behavior locally.
Answering the question whether or not an intentionbased modeling of the LfD problem is advantageous, is out
of the scope of this paper; however, we point to the comprehensive discussion in [26]. Rather, we present a hybrid
solution containing both intentional and sub-intentional elements. More specifically, our method does not explicitly capture the expert’s goals in the form of a reward function but
infers a policy model directly from the demonstration data;
nonetheless, the presented algorithm learns a task-specific
representation of the system state space which encodes the
structure of the underlying control problem to facilitate the
policy prediction task. An early version of this idea can
be found in [27], where the authors proposed a simple
method to partition a system’s state space into a set of socalled control situations to learn a global system controller
based on a small set of informative states. However, their
framework does not incorporate any demonstration data
and the proposed partitioning is based on heuristics. A more
sophisticated partitioning approach utilizing expert demonstrations is shown in [11]; yet, the proposed expectationmaximization framework applies to deterministic policies
and finite state spaces only.
The closest methods to ours can be probably found in
[19] and [10]. The authors of [19] presented a nonparametric
inverse reinforcement learning approach to cluster the expert data based on a set of learned subgoals encoded in the
form of local rewards. Unfortunately, the required subgoal
3. The rationale behind this is that an agent’s intention is always
specific to the task being performed and can hence serve as a compact
description of it [13]. However, if the intention of the agent is misunderstood, then also the subsequent generalization step will trivially fail.
assignments are learned only for the demonstration set and,
thus, the algorithm cannot be used for action prediction
at unvisited states unless it is extended with a non-trivial
post-processing step which solves the subgoal assignment
problem. Moreover, the algorithm requires an MDP solver,
which causes difficulties for systems with uncountable state
spaces. The sub-intentional model in [10], on the other hand,
can be used to learn a class of finite state controllers directly
from the expert demonstrations. Like our framework, the
algorithm can handle various kinds of uncertainty about
the data but, again, the proposed approach is limited to
discrete settings. In the context of reinforcement learning,
we further point to the work presented in [28] whose
authors follow a nonparametric strategy similar to ours, to
learn a distribution over predictive state representations for
decision-making.
1.3
Paper outline
The outline of the paper is as follows: In Section 2, we
introduce our parametric policy recognition framework and
derive inference algorithms for both countable and uncountable state spaces. In Section 3, we consider the policy recognition problem from a nonparametric viewpoint and provide insights into the state representation learning problem.
Simulation results are presented in Section 4 and we give
a conclusion of our work in Section 5. In the supplement,
we provide additional simulation results, a note on the
computational complexity of our model, as well as an indepth discussion on the issue of marginal invariance and
the problem of policy prediction in large states spaces.
2
2.1
PARAMETRIC P OLICY R ECOGNITION
Finite state spaces: the static model
First, let us assume that the expert system can be modeled
on a finite state space S and let |S| denote its cardinality.
For notational convenience, we represent both states and
actions by integer values. Starting with the most general
case, we assume that the expert executes an individual
control strategy at each possible system state. Accordingly,
we introduce a set of local control parameters or local con|S|
trollers {θ i }i=1 by which we describe the expert’s choice of
actions. More specifically, we model the executed actions
as categorical random variables and let the j th element of
θ i represent the probability that the expert chooses action
j at state i. Consequently, θ i lies in the (|A| − 1)-simplex,
which we denote by the symbol ∆ for brevity of notation, i.e.
θ i ∈ ∆ ⊆ R|A| . Summarizing all local control parameters in
a single matrix, Θ ∈ Ω ⊆ ∆|S| , we obtain the global control
parameter of the system as already introduced in Section 1.1,
which compactly captures the expert behavior. Note that
we denote the global control parameter here by Θ instead
of ω , for reasons that will become clear later. Each action a
is thus characterized by the local policy that is induced by
the control parameter of the underlying state,
π(a | s = i, Θ) = C AT(a | θ i ).
For simplicity, we will write π(a | θ i ) since the state information is used only to select the appropriate local controller.
4
Considering a finite set of actions, it is convenient to
place a symmetric Dirichlet prior on the local control parameters,
pθ (θ i | α) = D IR(θ i | α · 1|A| ),
which forms the conjugate distribution to the categorical
distribution over actions. Here, 1|A| denotes the vector of
all ones of length |A|. The prior is itself parametrized by a
concentration parameter α which can be further described
by a hyperprior pα (α), giving rise to a Bayesian hierarchical
model. For simplicity, we assume that the value of α is fixed
for the remainder of this paper, but the extension to a full
Bayesian treatment is straightforward. The joint distribution
of all remaining model variables is, therefore, given as
p(s, a, Θ | α) = p1 (s1 )
... ×
TY
−1
t=1
|S|
Y
i=1
pθ (θ i | α) . . .
(1)
T (st+1 | st , at )π(at | θ st ),
where a = (a1 , a2 , . . . , aT −1 ) denotes the latent action
sequence taken by the expert and p1 (s1 ) is the initial state
distribution of the system. Throughout the rest of the paper,
we refer to this model as the static model. The corresponding
graphical visualization is depicted in Fig. 1.
2.1.1
Gibbs sampling
Following a Bayesian methodology, our goal is to determine
the posterior distribution p(Θ | s, α), which contains all
information necessary to make predictions about the expert behavior. For the static model in Eq. (1), the required
marginalization of the latent action sequence a can be
computed efficiently because the joint distribution factorizes
over time instants. For the extended models presented in
later sections, however, a direct marginalization becomes
computationally intractable due to the exponential growth
of latent variable configurations. As a solution to this problem, we follow a sampling-based inference strategy which
is later on generalized to more complex settings.
For the simple model described above, we first approximate the joint posterior distribution p(Θ, a | s, α) over both
controllers and actions using a finite number of Q samples,
and then marginalize over a in a second step,
X
p(Θ | s, α) =
p(Θ, a | s, α)
(2)
≈
X
a
st−1
st
st+1
at−1
at
at+1
α
z
Θ
Fig. 1: Graphical model of the policy recognition framework. The underlying dynamical structure is that of an
MDP whose global control parameter Θ is treated as a
random variable with prior distribution parametrized by α.
The indicator node z is used for the clustering model in
Section 2.2. Observed variables are highlighted in gray.
Herein, a−t and Θ−i refer to all actions/controllers except
at and θ i , respectively. The latter of the two expressions
reveals that, in order to sample θ i , we need to consider
only those actions played at the corresponding state i.
Furthermore, the first expression shows that, given Θ, all
actions {at } can be sampled independently of each other.
Therefore, inference can be done in parallel for all θ i . This
can be also seen from the posterior distribution of the global
control parameter, which factorizes over states,
p(Θ | s, a, α) ∝
|S|
Y
i=1
pθ (θ i | α)
Y
t:st =i
π(at | θ i ).
(3)
From the conjugacy of pθ (θ i | α) and π(at | θ i ), it follows
that the posterior over θ i is again Dirichlet distributed with
updated concentration parameter. In particular, denoting by
φi,j the number of times that action j is played at state i for
the current assignment of actions a,
X
φi,j :=
1(at = j),
(4)
t:st =i
and by collecting these quantities in the form of vectors, i.e.
φi := [φi,1 , . . . , φi,|A| ], we can rewrite Eq. (3) as
p(Θ | s, a, α) =
|S|
Y
i=1
D IR(θ i | φi + α · 1|A| ).
(5)
a
Q
Q
X
1
1 X
δ {q} {q} (Θ, a) =
δ {q} (Θ),
Q q=1 Θ a
Q q=1 Θ
where (Θ{q} , a{q} ) ∼ p(Θ, a | s, α), and δx (·) denotes
Dirac’s delta function centered at x. This two-step approach
gives rise to a simple inference procedure since the joint
Q
samples {(Θ{q} , a{q} )}q=1 can be easily obtained from a
Gibbs sampling scheme, i.e. by sampling iteratively from
the following two conditional distributions,
p(at | a−t , s, Θ, α) ∝ T (st+1 | st , at )π(at | θ st ),
Y
p(θ i | Θ−i , s, a, α) ∝ pθ (θ i | α)
π(at | θ i ).
t:st =i
2.1.2 Collapsed Gibbs sampling
Choosing a Dirichlet distribution as prior model for the
local controllers is convenient as it allows us to arrive at
analytic expressions for the conditional distributions that
are required to run the Gibbs sampler. As an alternative,
we can exploit the conjugacy property of pθ (θ i | α) and
π(at | θ i ) to marginalize out the control parameters during
the sampling process, giving rise to a collapsed sampling
scheme. Collapsed sampling is advantageous in two different respects: first, it reduces the total number of variables
to be sampled and, hence, the number of computations
required per Gibbs iteration; second, it increases the mixing
speed of the underlying Markov chain that governs the
sampling process, reducing the correlation of the obtained
5
samples and, with it, the variance of the resulting policy
estimate.
Formally, collapsing means that we no longer approximate the joint distribution p(Θ, a | s, α) as done in Eq. (2),
but instead sample from the marginal density p(a | s, α),
X
p(Θ | s, α) =
p(Θ | s, a, α)p(a | s, α)
a
≈
X
a
1
p(Θ | s, a, α)
Q
Q
Q
X
q=1
1 X
=
p(Θ | s, a{q} , α),
Q q=1
δa{q} (a)
(6)
where a{q} ∼ p(a | s, α). In contrast to the previous
approach, the target distribution is no longer represented
by a sum of Dirac measures but described by a product of
Dirichlet mixtures (compare Eq. (5)). The required samples
{a{q} } can be obtained from a collapsed Gibbs sampler with
Z
p(s, a, Θ | α) dΘ
p(at | a−t , s, α) ∝
|S|
∆
Z
Y
∝ T (st+1 | st , at )
pθ (θ st | α)
π(at0 | θ st ) dθ st .
∆
t0 :st0 =st
It turns out that the above distribution provides an easy
sampling mechanism since the integral part, when viewed
as a function of action at only, can be identified as the conditional of a Dirichlet-multinomial distribution. This distribution is then reweighted by the likelihood T (st+1 | st , at )
of the observed transition. The final (unnormalized) weights
of the resulting categorical distribution are hence given as
p(at = j | a−t , s, α) ∝ T (st+1 | st , at = j) · (ϕt,j + α), (7)
where ϕt,j counts the number of occurrences of action j
among all actions in a−t played at the same state as at (that
is, st ). Explicitly,
X
ϕt,j :=
1(at0 = j).
t0 :st0 =st
t0 6=t
Note that these values can be also expressed in terms of the
sufficient statistics introduced in the last section,
ϕt,j = φst ,j − 1(at = j).
As before, actions played at different states may be sampled
independently of each other because they are generated by
different local controllers. Consequently, inference about Θ
again decouples for all states.
2.2
Towards large state spaces: a clustering approach
While the methodology introduced so far provides a means
to solve the policy recognition problem in finite state spaces,
the presented approaches quickly become infeasible for
large spaces as, in the continuous limit, the number of
parameters to be learned (i.e. the size of Θ) will grow
unbounded. In that sense, the presented methodology is
prone to overfitting because, for larger problems, we will
never have enough demonstration data to sufficiently cover
the whole system state space. In particular, the static model
S
θ1
θ2
θ3
θ4
θ5
Fig. 2: Schematic illustration of the clustering model. The
state space S is partitioned into a set of clusters {Ck }, each
governed by its own local control parameter θ k .
makes no assumptions about the structure of Θ but treats
all local policies separately (see Eq. (5)); hence, we are not
able to generalize the demonstrated behavior to regions of
the state space that are not directly visited by the expert.
Yet, we would certainly like to predict the expert behavior
also at states for which there is no trajectory data available. Moreover, we should expect a well-designed model to
produce increasingly accurate predictions at regions closer
to the observed trajectories (with the precise definition of
“closeness” being left open for the moment).
A simple way to counteract the overfitting problem, in
general, is to restrict the complexity of a model by limiting
the number of its free parameters. In our case, we can avoid
the parameter space to grow unbounded by considering
only a finite number of local policies that need to be shared
between the states. The underlying assumption is that, at
each state, the expert selects an action according to one
of K local policies, with corresponding control parameters {θ k }K
k=1 . Accordingly, we introduce a set of indicator
|S|
or cluster assignment variables, {zi }i=1 , zi ∈ {1, . . . , K},
which map the states to their local controllers (Fig. 1).
Obviously, such an assignment implies a partitioning of the
state space (Fig. 2), resulting in the following K clusters,
Ck := {i : zi = k},
k ∈ {1, . . . , K}.
Although we motivated the clustering of states by the
problem of overfitting, partitioning a system’s space is not
only convenient from a statistical point of view; mapping
the inference problem down to a lower-dimensional space
is also reasonable for practical reasons as we are typically
interested in understanding an agent’s behavior on a certain
task-appropriate scale. The following paragraphs discuss
these reasons in detail:
• In practice, the observed trajectory data will always be
noisy since we can take our measurements only up to a
certain finite precision. Even though we do not explicitly
consider observation noise in this paper, clustering the data
appears reasonable in order to robustify the model against
small perturbations in our observations.
• Considering the LfD problem from a control perspective,
the complexity of subsequent planning steps can be potentially reduced if the system dynamics can be approximately
described on a lower-dimensional manifold of the state
space, meaning that the system behavior can be well represented by a smaller set of informative states (c.f. finite state
6
controllers [29], control situations [27]). The LfD problem
can then be interpreted as the problem of learning a (nearoptimal) controller based on a small set of local policies that
together provide a good approximation of the global agent
behavior. What remains is the question how we can find
such a representation. The clustering approach described
above offers one possible solution to this problem.
• Finally, in any real setup, it is reasonable to assume that
the expert itself can only execute a finite-precision policy
due to its own limited sensing abilities of the system state
space. Consequently, the demonstrated behavior is going
to be optimal only up to a certain finite precision because
the agent is generally not able to discriminate between
arbitrary small differences of states. An interesting question
in this context is whether we can infer the underlying state
representation of the expert by observing its reactions to the
environment in the form of the resulting state trajectory. We
will discuss this issue in detail in Section 3.
By introducing the cluster assignment variables {zi }, the
joint distribution in Eq. (1) changes into
p(s, a, z, Θ | α) = p1 (s1 )
... ×
TY
−1
t=1
K
Y
k=1
(8)
pθ (θ k | α) . . .
T (st+1 | st , at )π(at | θ zst )pz (z),
where z = (z1 , z2 , . . . , z|S| ) denotes the collection of all
indicator variables and pz (z) is the corresponding prior
distribution to be further discussed in Section 2.2.3. Note
that the static model can be recovered as a special case of
the above when each state describes its own cluster, i.e. by
setting K = |S| and fixing zi = i (hence the name static).
In contrast to the static model, we now require both the
indicator zi and the corresponding control parameter θ zi
in order to characterize the expert’s behavior at a given
state i. Accordingly, the global control parameter of the
model is ω = (Θ, z) with underlying parameter space
Ω ⊆ ∆K ×{1, . . . , K}|S| (see Section 1.1), and our target distribution becomes p(Θ, z | s, α). In what follows, we derive
the Gibbs and the collapsed Gibbs sampler as mechanisms
for approximate inference in this setting.
2.2.1 Gibbs sampling
As shown by the following equations, the expressions for
the conditional distributions over actions and controllers
take a similar form to those of the static model. Here, the
only difference is that we no longer group the actions by
their states but according to their generating local policies
or, equivalently, the clusters {Ck },
p(at | a−t , s, z, Θ, α) ∝ T (st+1 | st , at ) · π(at | θ zst ),
Y
p(θ k | Θ−k , s, a, z, α) ∝ pθ (θ k | α)
π(at | θ k )
t:zst =k
= pθ (θ k | α)
Y
t:st ∈Ck
π(at | θ k ).
The latter expression again takes the form of a Dirichlet
distribution with updated concentration parameter,
p(θ k | Θ−k , s, a, z, α) = D IR(θ k | ξ k + α · 1|A| ),
where ξ k := [ξk,1 , . . . , ξk,|A| ], and ξk,j denotes the number
of times that action j is played at states belonging to cluster
Ck in the current assignment of a. Explicitly,
X
X X
ξk,j :=
1(at = j) =
1(at = j),
(9)
i∈Ck t:st =i
t:zst =k
which is nothing but the sum of the φi,j ’s of the corresponding states,
X
ξk,j =
φi,j .
i∈Ck
In addition to the actions and control parameters, we now
also need to sample the indicators {zi }, whose conditional
distributions can be expressed in terms of the corresponding
prior model and the likelihood of the triggered actions,
Y
p(zi | z −i , s, a, Θ, α) ∝ p(zi | z −i )
π(at | θ zi ). (10)
t:st =i
2.2.2 Collapsed Gibbs sampling
As before, we derive the collapsed Gibbs sampler by
marginalizing out the control parameters,
Z
p(s, a, z, Θ | α) dΘ
(11)
p(zi | z −i , s, a, α) ∝
∝ p(zi | z −i )
∝ p(zi | z −i )
∝ p(zi | z −i )
∝ p(zi | z −i )
Z
∆K
K
Y
∆K k=1
Z
K
Y
∆K k=1
Z
K
Y
∆K k=1
K
Y
k=1
pθ (θ k | α)
pθ (θ k | α)
pθ (θ k | α)
TY
−1
t=1
|S|
π(at | θ zst ) dΘ
Y Y
i0 =1 t:st =i0
Y
π(at | θ zi0 ) dΘ
Y
i0 :zi0 =k t:st =i0
π(at | θ k ) dΘ
Z
Y
pθ (θ k | α)
π(at | θ k ) dθ k .
∆
t:st ∈Ck
Here, we first grouped the actions by their associated states
and then grouped the states themselves by the clusters
{Ck }. Again, this distribution admits an easy sampling
mechanism as it takes the form of a product of Dirichletmultinomials, reweighted by the conditional prior distribution over indicators. In particular, we observe that all
actions played at some state i appear in exactly one of
the K integrals of the last equation. In other words, by
changing the value of zi (i.e. by assigning state i to another
cluster), only two of the involved integrals are affected: the
one belonging to the previously assigned cluster, and the
one of the new cluster. Inference about the value of zi can
thus be carried out using the following two sets of sufficient
statistics:
•
•
φi,j : the number of actions j played at state i,
ψi,j,k : the number of actions j played at states assigned to cluster Ck , excluding state i.
The φi,j ’s are the same as in Eq. (4) and their definition is
repeated here just as a reminder. For the ψi,j,k ’s, on the other
hand, we find the following explicit expression,
X X
ψi,j,k :=
1(at = j),
i0 ∈Ck t:st =i0
i0 6=i
7
which can be also written in terms of the statistics used for
the ordinary Gibbs sampler,
ψi,j,k = ξk,j − 1(i ∈ Ck ) · φi,j .
By collecting these quantities in a vector, i.e. ψ i,k :=
[ψi,1,k , . . . , ψi,|A|,k ], we end up with the following simplified expression,
p(zi = k | z −i , s, a, α) ∝ p(zi = k | z −i ) . . .
K
Y
... ×
D IR M ULT(ψ i,k0 + 1(k 0 = k) · φi | α).
k0 =1
Further, we obtain the following result for the conditional
distribution of action at ,
p(at | a−t , s, z, α) ∝ T (st+1 | st , at ) . . .
Z
Y
π(at0 | θ zst ) dθ zst .
... ×
pθ (θ zst | α)
∆
t0 :zs 0 =zst
t
By introducing the sufficient statistics {ϑt,j }, which count
the number of occurrences of action j among all states that
are currently assigned to the same cluster as st (i.e. the
cluster Czst ), excluding at itself,
X
ϑt,j :=
1(at = j),
t0 :zs 0 =zst
t
t0 6=t
we finally arrive at the following expression,
p(at = j | a−j , s, z, α) ∝ (ϑt,j + α) · T (st+1 | st , at = j).
As for the static model, we can establish a relationship between the statistics used for the ordinary and the collapsed
sampler,
ϑt,j = ξzst ,j − 1(at = j).
2.2.3 Prior models
In order to complete our model, we need to specify a prior
distribution over indicator variables pz (z). The following
paragraphs present three such candidate models:
Non-informative prior
The simplest of all prior models is the non-informative
prior over partitionings, reflecting the assumption that, a
priori, all cluster assignments are equally likely and that the
indicators {zi } are mutually independent. In this case, pz (z)
is constant and, hence, the term p(zi | z −i ) in Eq. (10) and
Eq. (11) disappears, so that the conditional distribution of
indicator zi becomes directly proportional to the likelihood
of the inferred action sequence.
Mixing prior
Another simple yet expressive prior can be realized by the
(finite) Dirichlet mixture model. Instead of assuming that
the indicator variables are independent, the model uses a set
of mixing coefficients q = [q1 , . . . , qK ], where qk represents
the prior probability that an indicator variable takes on
value k . The mixing coefficients are themselves modeled by
a Dirichlet distribution, so that we finally have
γ
· 1K ),
q ∼ D IR(q |
K
zi | q ∼ C AT(zi | q),
where γ is another concentration parameter, controlling the
variability of the mixing coefficients. Note that the indicator
variables are still conditionally independent given the mixing
coefficients in this model. More specifically, for a fixed q , the
conditional distribution of a single indicator in Eq. (10) and
Eq. (11) takes the following simple form,
p(zi = k | z −i , q) = qk .
If the value of q is unknown, we have two options to include
this prior into our model. One is to sample q additionally to
the remaining variables by drawing values from the following conditional distribution during the Gibbs procedure,
|S|
Y
γ
· 1K )
C AT(zi | q)
K
i=1
γ
∝ D IR(q | ζ +
· 1K ),
K
where ζ := [ζ1 , . . . , ζK ], and ζk denotes the number of
variables zi that map to cluster Ck ,
p(q | s, a, z, Θ, α) ∝ D IR(q |
ζk =
|S|
X
1(zi = k).
i=1
Alternatively, we can again make use of the conjugacy property to marginalize out the mixing proportions q during the
inference process, just as we did for the control parameters
in previous sections. The result is (additional) collapsing in
q . In this case, we simply replace the factor p(zi = k | z −i )
in the conditional distribution of zi by
γ
(−i)
(12)
p(zi = k | z −i , γ) ∝ (ζk + ),
K
(−i)
where ζk
is defined like ζk but without counting the
current value of indicator zi ,
(−i)
ζk
:=
|S|
X
i0 =1
i0 6=i
1(zi = k) = ζk − 1(zi = k).
A detailed derivation is omitted here but follows the same
style as for the collapsing in Section 2.1.2.
Spatial prior
Both previous prior models assume (conditional) independence of the indicator variables and, hence, make no specific
assumptions about their dependency structure. However,
we can also use the prior model to promote a certain type of
spatial state clustering. A reasonable choice is, for instance,
to use a model which preferably groups “similar” states
together (in other words, a model which favors clusterings
that assign those states the same local control parameter).
Similarity of states can be expressed, for example, by a
monotonically decreasing decay function f : [0, ∞) → [0, 1]
which takes as input the distance between two states. The
required pairwise distances can be, in turn, defined via some
distance metric χ : S × S → [0, ∞).
In fact, apart from the reasons listed in Section 2.2, there
is an additional motivation, more intrinsically related to the
dynamics of the system, why such a clustering can be useful:
given that the transition model of our system admits locally
smooth dynamics (which is typically the case for real-world
8
systems), the resulting optimal control policy often turns
out to be spatially smooth, too [11]. More specifically, under
an optimal policy, two nearby states are highly likely to
experience similar controls; hence, it is reasonable to assume
a priori that both share the same local control parameter.
For the policy recognition task, it certainly makes sense to
regularize the inference problem by encoding this particular
structure of the solution space into our model. The Potts
model [30], which is a special case of a Markov random
field with pairwise clique potentials [31], offers one way to
do this,
pz (z) ∝
|S|
Y
i=1
exp
!
|S|
βX
f (di,j )δ(zi , zj ) .
2 j=1
j6=i
Here, δ denotes Kronecker’s delta, di,j := χ(si , sj ), i, j ∈
{1, . . . , |S|}, are the state similarity values, and β ∈ [0, ∞)
is the (inverse) temperature of the model which controls
the strength of the prior. From this equation, we can easily
derive the conditional distribution of a single indicator
variable zi as
p(zi | z −i ) ∝ exp β
|S|
X
j=1
j6=i
!
f (di,j )δ(zi , zj ) .
(13)
This completes our inference framework for finite spaces.
2.3
Countably infinite and uncountable state spaces
A major advantage of the clustering approach presented in
the last section is that, due to the limited number of local
policies to be learned from the finite amount of demonstration data, we can now apply the same methodology to
state spaces of arbitrary size, including countably infinite
and uncountable state spaces. This extension had been
practically impossible for the static model because of the
overfitting problem explained in Section 2.2. Nevertheless,
there remains a fundamental conceptual problem: a direct
extension of the model to these spaces would imply that
the distribution over possible state partitionings becomes an
infinite-dimensional object (i.e., in the case of uncountable
state spaces, a distribution over functional mappings from
states to local controllers), requiring an infinite number of
indicator variables. Certainly, such an object is non-trivial to
handle computationally.
However, while the number of latent cluster assignments grows unbounded with the size of the state space,
the amount of observed trajectory data always remains
finite. A possible solution to the problem is, therefore, to
reformulate the inference task on a reduced state space
Se := {s1 , s2 , . . . , sT } containing only states along the observed trajectories. Reducing the state space in this way
means that we need to consider only a finite set of indicator
variables {zt }Tt=1 , one for each expert state s ∈ Se, which
always induces a model of finite size. Assuming that no
state is visited twice, we may further use the same index
set for both variable types.4 In order to limit the complexity
of the dependency structure of the indicator variables for
larger data sets, we further let the value of indicator zt
depend only on a subset of the remaining variables z −t as
defined by some neighborhood rule N . The resulting joint
distribution is then given as
!
T
Y
β X
pz (z | s) ∝
exp
f (dt,t0 )δ(zt , zt0 ) ,
2 t0 ∈N
t=1
t
which now implicitly depends on the state sequence s
through the pairwise distances dt,t0 := χ(st , st0 ), t, t0 ∈
{1, . . . , T } (hence the conditioning on s).
The use of a finite number of indicator variables along
the expert trajectories obviously circumvents the abovementioned problem of representational complexity. Nevertheless, there are some caveats associated with this approach. First of all, using a reduced state space model raises
the question of marginal invariance [32]: if we added a new
trajectory point to the data set, would this change our belief
about the expert policy at previously visited states? In particular, how is this different from modeling that new point
together with the initial ones in the first place? And further,
what does such a reduced model imply for unvisited states?
Can we still use it to make predictions about their local
policies? These questions are, in fact, important if we plan
to use our model to generalize the expert demonstrations to
new situations. For a detailed discussion on this issue, the
reader is referred to the supplement. Here, we focus on the
inferential aspects of the problem, which means to identify
the system parameters at the given trajectory states.
Another (but related) issue resulting from the reduced
modeling approach is that we lose the simple generative
interpretation of the process that could be used to explain
the data generation beforehand. In the case of finite state
spaces, we could think of a trajectory as being constructed
by the following, step-wise mechanism: first, the prior pz (z)
is used to generate a set of indicator variables for all states.
Independently, we pick some value for α from pα (α) and
sample K local control parameters from pθ (θ k | α). To
finally generate a trajectory, we start with an initial state
s1 , generated by p1 (s1 ), select a random action a1 from
π(a1 | s1 , θ zs1 ) and transition to a new state s2 according to
T (s2 | s1 , a1 ), where we select another action a2 , and so on.
Such a directed way of thinking is possible since the finite
model naturally obeys a causal structure where later states
depend on earlier ones and the decisions made there. Furthermore, the cluster assignments and the local controllers
could be generated in advance and isolated from each other
because they were modeled marginally independent.
For the reduced state space model, this interpretation no
longer applies as the model has no natural directionality. In
fact, its variables depend on each other in a cyclic fashion:
altering the value of a particular indicator variable (say, the
one corresponding to the last trajectory point) will have
4. Note that we make this assumption for notational convenience
only and that it is not required from a mathematical point of view.
Nonetheless, for uncountable state spaces the assumption is reasonable
since the event of reaching the same state twice has zero probability
for most dynamic models. In the general case, however, the indicator
variables require their own index set to ensure that each system state is
associated with exactly one cluster, even when visited multiple times.
9
an effect on the values of all remaining indicators due to
their spatial relationship encoded by the “prior distribution”
pz (z | s). Changing the values of the other indicators,
however, will influence the actions being played at the
respective states which, in turn, alters the probability of
ending up with the observed trajectory in the first place and,
hence, the position and value of the indicator variable we
started with. Explaining the data generation of this model
using a simple generative process is, therefore, not possible.
Nevertheless, the individual building blocks of our
model (that is, the policy, the transition model, etc.) together
form a valid distribution over the model variables, which
can be readily used for parameter inference. For the reasons
explained above, it makes sense to define this distribution in
the form of a discriminative model, ignoring the underlying
generative aspects of the process. This is sufficient since we
can always condition on the observed state sequence s,
p(a, Θ, z | s, α) =
K
Y
1
pz (z | s)
pθ (θ k | α) . . .
Zs
k=1
. . . × p1 (s1 )
TY
−1
t=1
T (st+1 | st , at )π(at | θ zst ).
Herein, Zs is a data-dependent normalizing constant. The
structure of this distribution is illustrated by the factor graph
shown in the supplement (Fig. S-1), which highlights the
circular dependence between the variables. Note that, for
any fixed state sequence s, this distribution indeed encodes
the same basic properties as the finite model in Eq. (8).
In particular, the conditional distributions of all remaining
variables remain unchanged, which allows us to apply the
same inference machinery that we already used in the finite
case. For a deeper discussion on the difference between the
two models, we again point to the supplement.
3
N ONPARAMETRIC P OLICY R ECOGNITION
In the last section, we presented a probabilistic policy recognition framework for modeling the expert behavior using
a finite mixture of K local policies. Basically, there are two
situations when such a model is useful:
•
•
either, we know the true number of expert policies,
or, irrespective of the true behavioral complexity, we
want to find an approximate system description in
terms of at most K distinct control situations [27]
(c.f. finite state controllers [29]).
In all other cases, we are faced with the non-trivial problem of choosing K . In fact, the choice of K should not
just be considered a mathematical necessity to perform
inference in our model. By selecting a certain value for
K we can, of course, directly control the complexity class
of potentially inferred expert controllers. However, from a
system identification point of view, it is more reasonable to
infer the required granularity of the state partitioning from
the observed expert behavior itself, instead of enforcing a
particular model complexity. This way, we can gain valuable
information about the underlying control structure and state
representation used by the expert, which offers a possibility
to learn a state partitioning of task-appropriate complexity
directly from the demonstration data. Hence, the problem
of selecting the right model structure should be considered
as part of the inference problem itself.
From a statistical modeling perspective, there are two
common ways to approach this problem. One is to make use
of model selection techniques in order to determine the most
parsimonious model that is in agreement with the observed
data. However, choosing a particular model complexity still
means that we consider only one possible explanation for
the data, although other explanations might be likewise
plausible. For many inference tasks, including this one, the
more elegant approach is to keep the complexity flexible
and, hence, adaptable to the data. Mathematically, this can
be achieved by assuming a potentially infinite set of model
parameters (in our case controllers) from which we activate
only a finite subset to explain the particular data set at hand.
This alternative way of thinking opens the door to the rich
class of nonparametric models, which provide an integrated
framework to formulate the inference problem over both
model parameters and model complexity as a joint learning
problem.
3.1
A Dirichlet process mixture model
The classical way to nonparametric clustering is to use
a Dirichlet process mixture model (DPMM) [33]. These
models can be obtained by starting from a finite mixture
model and letting the number of mixture components (i.e.
the number of local controllers) approach infinity. In our
case, we start with the clustering model from Section 2.2,
using a mixing prior over indicator variables,
q ∼ D IR(q |
γ
K
· 1K )
θ k ∼ D IR(θ k | α · 1|A| )
s1 ∼ p1 (s1 )
zi | q ∼ C AT(zi | q)
at | st , Θ, z ∼ π(at | θ zst ) (14)
st+1 | st , at ∼ T (st+1 | st , at ) .
From these equations, we arrive at the corresponding nonparametric model as K goes to infinity. For the theoretical
foundations of this limit, the reader is referred to the more
general literature on Dirichlet processes, such as [33], [34].
In this paper, we restrict ourselves to providing the resulting
sampling mechanisms for the policy recognition problem.
In a DPMM, the mixing proportions q of the local
parameters are marginalized out (that is, we use a collapsed sampler). The resulting distribution over partitionings is described by a Chinese restaurant process (CRP) [35]
which can be derived, for instance, by considering the limit
K → ∞ of the mixing process induced by the Gibbs update
in Eq. (12),
(
(−i)
ζ
if k ∈ {1, . . . , K ∗ },
p(zi = k | z −i , γ) ∝ k
(15)
γ
if k = K ∗ + 1.
Here, K ∗ denotes the number of distinct entries in z −i
which are represented by the numerical values {1, . . . , K ∗ }.
In this model, a state joins an existing cluster (i.e. a group of
states whose indicators have the same value) with probability proportional to the number of states already contained in
that cluster. Alternatively, it may create a new cluster with
probability proportional to γ .
From the model equations (14) it is evident that, given a
particular setting of indicators, the conditional distributions
of all other variable types remain unchanged. Effectively, we
10
only replaced the prior model pz (z) by the CRP. Hence, we
can apply the same Gibbs updates for the actions and controllers as before and need to rederive only the conditional
distributions of the indicator variables under consideration
of the above defined process. According to Eq. (15), we
herein need to distinguish whether an indicator variable
takes a value already occupied by other indicators (i.e. it
joins an existing cluster) or it is assigned a new value (i.e.
∗
it creates a new cluster). Let {θ k }K
k=1 denote the set of
control parameters associated with z −i . In the first case
(k ∈ {1, . . . , K ∗ }), we can then write
S
c2
s2
c8
s3
s4
c3
s1
c1
s7
c4
c6
s6
s5
c5
s8
c7
c9
s9
Fig. 3: Schematic illustration of the ddCRP-based clustering
applied to the reduced state space model in Section 2.3.
p(zi = k |
Each trajectory state is connected to some other state of
= p(zi = k | z −i , {at }t:st =i , θ k , α, γ)
the sequence. The connected components of the resulting
∝ p(zi = k | z −i , θ k , α, γ)p({at }t:st =i | zi = k, z −i , θ k , α, γ) graph implicitly define the state clustering. Coloring of the
∝ p(zi = k | z −i , γ)p({at }t:st =i | θ k )
background illustrates the spatial cluster extrapolation (see
Y
(−i)
Section A in the supplement). Note that the underlying
∝ ζk ·
π(at | θ k ).
decision-making process is assumed to be discrete in time;
t:st =i
the continuous gray line shown in the figure is only to
In the second case (k = K ∗ + 1), we instead obtain
indicate the temporal ordering of the trajectory states.
∗
K∗
p(zi = K + 1 | z −i , s, a, {θ k }k=1 , α, γ)
∗
z −i , s, a, {θ k0 }K
k0 =1 , α, γ)
= p(zi = K ∗ + 1 | z −i , {at }t:st =i , α, γ)
∝ p(zi = K ∗ + 1 | z −i , α, γ) . . .
. . . × p({at }t:st =i | zi = K ∗ + 1, z −i , α, γ)
∝ p(zi = K ∗ + 1 | z −i , γ)p({at }t:st =i | zi = K ∗ + 1, α)
Z
∝γ·
p({at }t:st =i | θ K ∗ +1 )pθ (θ K ∗ +1 | α) dθ K ∗ +1
Z∆ Y
π(at | θ K ∗ +1 )pθ (θ K ∗ +1 | α) dθ K ∗ +1
∝γ·
∆ t:s =i
t
∝ γ · D IR M ULT(φi | α).
If a new cluster is created, we further need to initialize the
corresponding control parameter θ K ∗ +1 by performing the
respective Gibbs update, i.e. by sampling from
∗
p(θ K ∗ +1 | z, s, a, {θ k }K
k=1 , α, γ)
= p(θ K ∗ +1 | {at }t:zst =K ∗ +1 , α)
∝ pθ (θ K ∗ +1 | α)p({at }t:zst =K ∗ +1 | θ K ∗ +1 )
Y
∝ pθ (θ K ∗ +1 | α) ·
π(at | θ K ∗ +1 )
t:zst =K ∗ +1
∝ D IR(θ K ∗ +1 | ξ K ∗ +1 + α · 1|A| ).
Should a cluster get unoccupied during the sampling process, the corresponding control parameter may be removed
from the stored parameter set {θ k } and the index set for
k needs to be updated accordingly. Note that this sampling
mechanism is a specific instance of Algorithm 2 described in
[33]. A collapsed variant can be derived in a similar fashion.
3.2 Policy recognition using the distance-dependent
Chinese restaurant process
In the previous section, we have seen that the DPMM can
be derived as the nonparametric limit model of a finite
mixture using a set of latent mixing proportions q for the
clusters. Although the DPMM allows us to keep the number
of active controllers flexible and, hence, adaptable to the
complexity of the demonstration data, the CRP as the underlying clustering mechanism does not capture any spatial
dependencies between the indicator variables. In fact, in the
CRP, the indicators {zi } are coupled only via their relative
frequencies (see Eq. (15)) but not through their individual
locations in space, resulting in an exchangeable collection
of random variables [35]. In fact, one could argue that the
spatial structure of the clustering problem is a priori ignored.
The fact that DPMMs are nevertheless used for spatial
clustering tasks can be explained by the particular form
of data likelihood models that are typically used for the
mixture components. In a Gaussian mixture model [36], for
instance, the spatial clusters emerge due to the unimodal
nature of the mixture components, which encodes the locality property of the model needed to obtain a meaningful
spatial clustering of the data. For the policy recognition
problem, however, the DPMM is not able to exploit any
spatial information via the data likelihood since the clustering of states is performed at the level of the inferred action
information (see Eq. (10)) and not on the state sequence
itself. Consequently, we cannot expect to obtain a smooth
clustering of the system state space, especially when the
expert policies are overlapping (i.e. when they share one or
more common actions) so that the action information alone
is not sufficient to discriminate between policies. For uncountable state spaces, this problem is further complicated
by the fact that we observe at most one expert state transition
per system state. Here, the spatial context of the data is the
only information which can resolve this ambiguity.
In order to facilitate a spatially smooth clustering, we
therefore need to consider non-exchangeable distributions
over partitionings. More specifically, we need to design our
model in such a way that, whenever a state s is “close” to
some other state s0 and assigned to some cluster Ck , then,
a priori, s0 should belong to the same cluster Ck with high
probability. In that sense, we are looking for the nonparametric counterpart of the Potts model. One model with such
properties is the distance-dependent Chinese restaurant
11
process (ddCRP) [32].5 As opposed to the traditional CRP,
the ddCRP explicitly takes into account the spatial structure
of the data. This is done in the form of pairwise distances
between states, which can be obtained, for instance, by
defining an appropriate distance metric on the state space
(see Section 2.2.3). Instead of assigning states to clusters as
done by the CRP, the ddCRP assigns states to other states
according to their pairwise distances. More specifically, the
probability that state i gets assigned to state j is defined as
(
ν
if i = j,
p(ci = j | D, ν) ∝
(16)
f (di,j ) otherwise,
where ν ∈ [0, ∞) is called the self-link parameter of the process, D denotes the collection of all pairwise state distances,
and ci is the “to-state” assignment of state i, which can be
thought of as a directed edge on the graph defined on the
set of all states (see Fig. 3). Accordingly, i and j in Eq. (16)
can take values {1, . . . , |S|} for the finite state space model
and {1, . . . , T } for our reduced state space model. The state
clustering is then obtained as a byproduct of this mapping
via the connected components of the resulting graph (see
Fig. 3 again).
Replacing the CRP by the ddCRP and following the
same line of argument as in [32], we obtain the required
conditional distribution of the state assignment ci as
j = i,
ν
p(ci = j | c−i ,s,a, α,D, ν) ∝ f (di,j )
no clusters merged,
f (di,j )·L Czi and Czj merged,
where we use the shorthand notation
L=
D IR M ULT(ξ zi + ξ zj | α)
D IR M ULT(ξ zi | α)D IR M ULT(ξ zj | α)
for the data likelihood term. The ξk,j ’s are defined as in
Eq. (9) but are based on the clustering which arises when we
ignore the current link ci . The resulting Gibbs sampler is a
collapsed one as the local control parameters are necessarily
marginalized out during the inference process.
4
S IMULATION R ESULTS
In this section, we present simulation results for two types of
system dynamics. As a proof-of-concept, we first investigate
the case of uncountable state spaces which we consider the
more challenging setting for reasons explained earlier. To
compare our framework with existing methods, we further
provide simulation results for the standard grid world
benchmark (see e.g. [9], [11], [19]). It should be pointed
out, however, that establishing a fair comparison between
LfD models is generally difficult due to their different
working principles (e.g. reward prediction vs. action prediction), objectives (system identification vs. optimal control),
requirements (e.g. MDP solver, knowledge of the expert’s
discount factor, countable vs. uncountable state space), and
5. Note that the authors of [32] avoid calling this model nonparametric since it cannot be cast as a mixture model originating from a random
measure. However, we stick to this term in order to make a clear
distinction to the parametric models in Section 2, and to highlight the
fact that there is no parameter K determining the number of controllers.
Fig. 4: Schematic illustration of the expert policy used in
Section 4.1, which applies eight local controllers to sixteen
distinct regions. A sample trajectory is shown in color.
assumptions (e.g. deterministic vs. stochastic expert behavior). Accordingly, our goal is rather to demonstrate the
prediction abilities of the considered models than to push
the models to their individual limits. Therefore, and to reduce the overall computational load, we tuned most model
hyper-parameters by hand. Our code is available at https:
//github.com/AdrianSosic/BayesianPolicyRecognition.
4.1
Example 1: uncountable state space
As an illustrative example, we consider a dynamical system which describes the circular motion of an agent on
a two-dimensional state space. The actions of the agent
correspond to 24 directions that divide the space of possible
angles [0, 2π) into equally-sized intervals. More specifically,
action j corresponds to the angle (j − 1) 2π
24 . The transition
model of the system is defined as follows: for each selected
action, the agent first makes a step of length µ = 1 in
the intended direction. The so-obtained position is then
distorted by additive zero-mean isotropic Gaussian noise of
variance σ 2 . This defines our transition kernel as
T (st+1 | st , at = j) = N (st+1 | st + µ · ej , σ 2 I),
(17)
where st , st+1 ∈ R2 , ej denotes the two-dimensional unit
vector pointing in the direction of action j , and I is the
two-dimensional identity matrix. The overall goal of our
agent is to describe a circular motion around the origin in
the best possible manner allowed by the available actions.
However, due to limited sensory information, the agent is
not able to observe its exact position on the plane but can
only distinguish between certain regions of the state space,
as illustrated by Fig. 4. Also, the agent is unsure about the
optimal control strategy, i.e. it does not always make optimal
decisions but selects its actions uniformly at random from a
subset of actions, consisting of the optimal one and the two
actions pointing to neighboring directions (see Fig. 4 again).
To increase the difficulty of the prediction task, we further
let the agent change the direction of travel whenever the
critical distance of r = 5 to the origin is exceeded.
Having defined the expert behavior, we generate 10
sample trajectories of length T = 100. Herein, we assume
a motion noise level of σ = 0.2 and initialize the agent’s
12
0
0
χ(s, s ) = ||s − s ||2 .
(18)
The corresponding similarity values are calculated using a
Gaussian-shaped kernel. More specifically,
!
d2
fPotts (d) = exp − 2
σf
for the Potts model and
fddCRP (d) = (1 − )fPotts (d) +
for the ddCRP model, with σf = 1 and a constant offset of
= 0.01 which ensures that states with large distances can
still join the same cluster. For the Potts model, we further
use a neighborhood structure containing the eight closest
trajectory points of a state. This way, we ensure that, in
principle, each local expert policy may occur at least once
in the neighborhood of a state. The concentration parameter
for the local controls is set to α = 1, corresponding to a
uniform prior belief over local policies.
A major drawback of the Potts model is that posterior
inference about the temperature parameter β is complicated
due to the nonlinear effect of the parameter on the normalization of the model. Therefore, we manually selected
a temperature of β = 1.6 based on a minimization of
the average policy prediction error (discussed below) via
parameter sweeping. As opposed to this, we extend the
inference problem for the ddCRP to the self-link parameter ν
as suggested in [32]. For this, we use an exponential prior,
pν (ν) = E XP(ν | λ),
with rate parameter λ = 0.1, and applied the independence
Metropolis-Hastings algorithm [37] using pν (ν) as proposal
distribution with an initial value of ν = 1. In all our
simulations, the sampler quickly converged to its stationary
distribution, yielding posterior values for ν with a mean of
0.024 and a standard deviation of 0.023.
To locally compare the predicted policy with the ground
truth at a given state, we compute their earth mover’s
distance (EMD) [38] with a ground distance metric measuring the absolute angular difference between the involved
actions. To track the learning progress of the algorithms,
we calculate the average EMD over all states of the given
trajectory set at each Gibbs iteration. Herein, the local policy
predictions are computed from the single Gibbs sample of
the respective iteration, consisting of all sampled actions, indicators and – in case of non-collapsed sampling – the local
control parameters. The resulting mean EMDs and standard
deviations are depicted in Fig. 5. The inset further shows
the average EMD computed at non-trajectory states which
are sampled on a regular grid (depicted in the supplement),
reflecting the quality of the resulting spatial prediction.
As expected, the finite mixture model (using the true
number of local policies, a collapsed mixing prior, and
1.4
1.4
1.2
1
0.8
0.6
0.4
10 0
1.2
average EMD
position uniformly at random on the unit circle. An example
trajectory is shown in Fig. 4. The obtained trajectory data is
fed into the presented inference algorithms to approximate
the posterior distribution over expert controllers, and the
whole experiment is repeated in 100 Monte Carlo runs.
For the spatial models, we use the Euclidean metric to
compute the pairwise distances between states,
1
0.8
0.6
0.4
10 0
10 1
10 2
mixing prior
Potts (complete)
Potts (collapsed)
ddCRP
10 1
10 2
Gibbs iteration
Fig. 5: Average policy prediction error at the simulated
trajectory states (main figure) and at non-trajectory states
(inset). Shown are the empirical mean values and standard
deviations, estimated from 100 Monte Carlo runs.
γ = 1) is not able to learn a reasonable policy representation
from the expert demonstrations since it does not explore the
spatial structure of the data. In fact, the resulting prediction
error shows only a slight improvement as compared to
an untrained model. In contrast to this, all spatial models
capture the expert behavior reasonably well. In agreement
with our reasoning in Section 2.1.2, we observe that the
collapsed Potts model mixes significantly faster and has a
smaller prediction variance than the non-collapsed version.
However, the ddCRP model gives the best result, both
in terms of mixing speed (see [32] for an explanation of
this phenomenon) and model accuracy. Interestingly, this is
despite the fact that the ddCRP model additionally needs to
infer the number of local controllers necessary to reproduce
the expert behavior. The corresponding posterior distribution, which shows a pronounced peak at the true number,
is depicted in the supplement. There, we also provide additional simulation results which give insights into the learned
state partitioning and the resulting spatial policy prediction
error. The results reveal that all expert motion patterns can
be identified by our algorithm.
4.2
Example 2: finite state space
In this section, we compare the prediction capabilities of our
model to existing LfD frameworks, in particular: the maximum margin method in [9] (max-margin), the maximum
entropy approach in [12] (max-ent), and the expectationmaximization algorithm in [11] (EM). For the comparison,
we restrict ourselves to the ddCRP model which showed
the best performance among all presented models.
As a first experiment, we compare all methods on a finite
version of the setting in Section 4.1, which is obtained by
discretizing the continuous state space into a regular grid
S = {(x, y) ∈ Z2 : |x|, |y| ≤ 10}, resulting in a total of 441
states. The transition probabilities are chosen proportional
to the normal densities in Eq. (17) sampled at the grid points.
Here, we used a noise level of σ = 1 and a reduced number
of eight actions. Probability mass “lying outside” the finite
grid area is shifted to the closest border states of the grid.
Figure 6a delineates the average EMD over the number
of trajectories (each of length T = 10) provided for training.
13
1.5
1.5
average EMD
average EMD
2
1
0.5
0
10 1
max-ent
max-margin
EM
ddCRP
10 2
10 3
10 4
number of trajectories
(a) learning curves (circular policy)
1.5
max-ent
max-margin
EM
ddCRP
1
1
0.5
0
10 1
max-ent
max-margin
EM
ddCRP
0.5
10 2
10 3
10 4
0
0
0.2
0.4
0.6
0.8
1
number of trajectories
(b) learning curves (MDP policy)
(c) model robustness (MDP policy)
Fig. 6: Average EMD values for the prediction task described in Section 4.2: (a) circular policy (b,c) MDP policy. Shown are
the empirical mean values and standard deviations, estimated from 100 Monte Carlo runs. The EMD values are computed
based on (a) the predicted action distributions and (b,c) the predicted next-state distributions. Note that the curves of
max-ent (purple) and max-margin (yellow) in subfigure (a) lie on top of each other.
We observe that neither of the two intentional models (maxent and max-margin) is able to capture the demonstrated expert behavior. This is due to the fact that the circular expert
motion cannot be explained by a simple state-dependent
reward structure but requires a more complex state-action
reward model, which is not considered in the original formulations [9], [12]. While the EM model is indeed able to
capture the general trend of the data, the prediction is less
accurate as compared to that of the ddCRP model, since it
cannot reproduce the stochastic nature of the expert policy.
In fact, this difference in performance will become even
more pronounced for expert policies which distribute their
probability mass on a larger subset of actions. Therefore,
the ddCRP model outperforms all other models since the
provided expert behavior violates their assumptions.
To analyze how the ddCRP competes against the other
models in their nominal situations, we further compare all
algorithms on a standard grid world task where the expert
behavior is obtained as the optimal response to a simple
state-dependent reward function. Herein, each state on the
grid is assigned a reward with a chance of 1%, which is then
drawn from a standard normal distribution. Worlds which
contain no reward are discarded. The discount factor 0.9,
which is used to compute the expert policy (see [17]), is
provided as additional input for the intentional models.
The results are shown in Figure 6b, which illustrates that
the intention-based max-margin method outperforms all
other methods for small amounts of training data. The subintentional methods (EM and ddCRP), on the other hand,
yield better asymptotic estimates and smaller prediction
variances. It should be pointed out that the three reference
methods have a clear advantage over the ddCRP in this case
because they assume a deterministic expert behavior a priori
and do not need to infer this piece of information from the
data. Despite this additional challenge, the ddCRP model
yields a competitive performance.
Finally, we compare all approaches in terms of their
robustness against modeling errors. For this purpose, we
repeat the previous experiment with a fixed number of 1000
trajectories but employ a different transition model for inference than used for data generation. More specifically, we utilize an overly fine-grained model consisting of 24 directions,
assuming that the true action set is unknown, as suggested
in Section 1.1. Additionally, we perturb the assumed model
by multiplying (and later renormalizing) each transition
probability with a random number generated according to
f (u) = tan( π4 (u + 1)), with u ∼ U NIFORM(−η, η) and
perturbation strength η ∈ [0, 1]. Due the resulting model
mismatch, a comparison to the ground truth policy based
on the predicted action distribution becomes meaningless.
Instead, we compute the Euclidean EMDs between the
true and the predicted next-state distributions, which we
obtain by marginalizing the actions of the true/assumed
transition model with respect to the true/learned policy.
Figure 6c depicts the resulting prediction performance for
different perturbation strengths η . The results confirm that
our approach is not only less sensitive to modeling errors as
argued in Section 1.1; also, the prediction variance is notably
smaller than those of the intentional models.
5
C ONCLUSION
In this work, we proposed a novel approach to LfD by
jointly learning the latent control policy of an observed
expert demonstrator together with a task-appropriate representation of the system state space. With the described
parametric and nonparametric models, we presented two
formulations of the same problem that can be used either
to learn a global system controller of specified complexity,
or to infer the required model complexity from the observed expert behavior itself. Simulation results for both
countable and uncountable state spaces and a comparison
to existing frameworks demonstrated the efficacy of our
approach. Most notably, the results showed that our method
is able to learn accurate predictive behavioral models in
situations where intentional methods fail, i.e. when the
expert behavior cannot be explained as the result of a simple
planning procedure. This makes our method applicable to
a broader range of problems and suggests its use in a more
general system identification context where we have no such
prior knowledge about the expert behavior. Additionally,
the task-adapted state representation learned through our
framework can be used for further reasoning.
14
R EFERENCES
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
B. D. Argall, S. Chernova, M. Veloso, and B. Browning, “A survey
of robot learning from demonstration,” Robotics and Autonomous
Systems, vol. 57, no. 5, pp. 469–483, 2009.
M. P. Deisenroth, D. Fox, and C. E. Rasmussen, “Gaussian processes for data-efficient learning in robotics and control,” IEEE
Transactions on Pattern Analysis and Machine Intelligence, vol. 37,
no. 2, pp. 408–423, 2015.
P. Abbeel and A. Y. Ng, “Exploration and apprenticeship learning
in reinforcement learning,” in Proc. 22nd International Conference on
Machine Learning, 2005, pp. 1–8.
D. Michie, M. Bain, and J. Hayes-Miches, “Cognitive models from
subcognitive skills,” IEE Control Engineering Series, vol. 44, pp. 71–
99, 1990.
C. Sammut, S. Hurst, D. Kedzier, and D. Michie, “Learning to fly,”
in Proc. 9th International Workshop on Machine Learning, 1992, pp.
385–393.
P. Abbeel, A. Coates, and A. Y. Ng, “Autonomous helicopter aerobatics through apprenticeship learning,” The International Journal
of Robotics Research, 2010.
D. A. Pomerleau, “Efficient training of artificial neural networks
for autonomous navigation,” Neural Computation, vol. 3, no. 1, pp.
88–97, 1991.
C. G. Atkeson and S. Schaal, “Robot learning from demonstration,” in Proc. 14th International Conference on Machine Learning,
vol. 97, 1997, pp. 12–20.
P. Abbeel and A. Y. Ng, “Apprenticeship learning via inverse
reinforcement learning,” in Proc. 21st International Conference on
Machine Learning, 2004.
A. Panella and P. J. Gmytrasiewicz, “Nonparametric Bayesian
learning of other agents’ policies in interactive POMDPs,” in
Proc. International Conference on Autonomous Agents and Multiagent
Systems, 2015, pp. 1875–1876.
A. Šošić, A. M. Zoubir, and H. Koeppl, “Policy recognition via expectation maximization,” in Proc. 41st IEEE International Conference
on Acoustics, Speech and Signal Processing, 2016.
B. D. Ziebart, A. L. Maas, J. A. Bagnell, and A. K. Dey, “Maximum entropy inverse reinforcement learning.” in Proc. 23rd AAAI
Conference on Artificial Intelligence, 2008, pp. 1433–1438.
A. Y. Ng and S. J. Russell, “Algorithms for inverse reinforcement
learning.” in Proc. 17th International Conference on Machine Learning,
2000, pp. 663–670.
D. Ramachandran and E. Amir, “Bayesian inverse reinforcement
learning,” Proc. 20th International Joint Conference on Artical Intelligence, vol. 51, pp. 2586–2591, 2007.
S. D. Parsons, P. Gymtrasiewicz, and M. J. Wooldridge, Game theory
and decision theory in agent-based systems.
Springer Science &
Business Media, 2012, vol. 5.
K. Hindriks and D. Tykhonov, “Opponent modelling in automated
multi-issue negotiation using Bayesian learning,” in Proc. 7th
International Joint Conference on Autonomous Agents and Multiagent
Systems, 2008, pp. 331–338.
R. S. Sutton and A. G. Barto, Reinforcement learning: An introduction.
MIT press, 1998.
U. Syed and R. E. Schapire, “A game-theoretic approach to apprenticeship learning,” in Advances in Neural Information Processing
Systems, 2007, pp. 1449–1456.
B. Michini and J. P. How, “Bayesian nonparametric inverse reinforcement learning,” in Machine Learning and Knowledge Discovery
in Databases. Springer, 2012, pp. 148–163.
C. G. Atkeson and J. C. Santamaria, “A comparison of direct and
model-based reinforcement learning,” in International Conference
on Robotics and Automation, 1997.
C. A. Rothkopf and C. Dimitrakakis, “Preference elicitation and inverse reinforcement learning,” in Machine Learning and Knowledge
Discovery in Databases. Springer, 2011, pp. 34–48.
O. Pietquin, “Inverse reinforcement learning for interactive systems,” in Proc. 2nd Workshop on Machine Learning for Interactive
Systems, 2013, pp. 71–75.
S. Schaal, “Is imitation learning the route to humanoid robots?”
Trends in Cognitive Sciences, vol. 3, no. 6, pp. 233–242, 1999.
K. Dvijotham and E. Todorov, “Inverse optimal control with
linearly-solvable MDPs,” in Proc. 27th International Conference on
Machine Learning, 2010, pp. 335–342.
E. Charniak and R. P. Goldman, “A Bayesian model of plan
recognition,” Artificial Intelligence, vol. 64, no. 1, pp. 53–79, 1993.
[26] B. Piot, M. Geist, and O. Pietquin, “Learning from demonstrations:
Is it worth estimating a reward function?” in Machine Learning and
Knowledge Discovery in Databases. Springer, 2013, pp. 17–32.
[27] M. Waltz and K. Fu, “A heuristic approach to reinforcement
learning control systems,” IEEE Transactions on Automatic Control,
vol. 10, no. 4, pp. 390–398, 1965.
[28] F. Doshi-Velez, D. Pfau, F. Wood, and N. Roy, “Bayesian nonparametric methods for partially-observable reinforcement learning,” IEEE Transactions on Pattern Analysis and Machine Intelligence,
vol. 37, no. 2, pp. 394–407, 2015.
[29] N. Meuleau, L. Peshkin, K.-E. Kim, and L. P. Kaelbling, “Learning
finite-state controllers for partially observable environments,” in
Proc. 15th Conference on Uncertainty in Artificial Intelligence, 1999,
pp. 427–436.
[30] R. B. Potts, “Some generalized order-disorder transformations,”
in Mathematical Proceedings of the Cambridge Philosophical Society,
vol. 48, no. 01, 1952, pp. 106–109.
[31] D. Koller and N. Friedman, Probabilistic Graphical Models: Principles
and Techniques. MIT press, 2009.
[32] D. M. Blei and P. I. Frazier, “Distance dependent Chinese restaurant processes,” The Journal of Machine Learning Research, vol. 12,
pp. 2461–2488, 2011.
[33] R. M. Neal, “Markov chain sampling methods for Dirichlet process
mixture models,” Journal of Computational and Graphical Statistics,
vol. 9, no. 2, pp. 249–265, 2000.
[34] T. S. Ferguson, “A Bayesian analysis of some nonparametric
problems,” The Annals of Statistics, pp. 209–230, 1973.
[35] D. J. Aldous, Exchangeability and related topics. Springer, 1985.
[36] C. E. Rasmussen, “The infinite Gaussian mixture model,” in Advances in Neural Information Processing Systems. MIT Press, 2000,
pp. 554–560.
[37] S. Chib and E. Greenberg, “Understanding the MetropolisHastings algorithm,” The American Statistician, vol. 49, no. 4, pp.
327–335, 1995.
[38] Y. Rubner, C. Tomasi, and L. J. Guibas, “A metric for distributions
with applications to image databases,” in 6th International Conference on Computer Vision, 1998, 1998, pp. 59–66.
Adrian Šošić is a member of the Signal Processing Group and an associate member of
the Bioinspired Communication Systems Lab at
Technische Universität Darmstadt. Currently, he
is working towards his Ph.D. degree under the
supervision of Prof. Abdelhak M. Zoubir and
Prof. Heinz Koeppl. His research interests center
around topics from machine learning and (inverse) reinforcement learning, with a focus on
probabilistic inference, multi-agent systems, and
Bayesian nonparametrics.
Abdelhak M. Zoubir is professor at the Department of Electrical Engineering and Information Technology at Technische Universität Darmstadt, Germany. His research interest lies in statistical methods for signal processing with emphasis on bootstrap techniques, robust detection
and estimation, and array processing applied
to telecommunications, radar, sonar, automotive
monitoring and biomedicine.
Heinz Koeppl is professor at the Department of
Electrical Engineering and Information Technology at Technische Universität Darmstadt, Germany. His research interests include Bayesian
inference methods for biomolecular data and
methods for reconstructing large-scale biological
or technological multi-agent systems from observational data.
1
A Bayesian Approach to Policy Recognition
and State Representation Learning
(Supplement)
Adrian Šošić, Abdelhak M. Zoubir and Heinz Koeppl
F
A
IN
M ARGINAL I NVARIANCE & P OLICY P REDICTION
L ARGE S TATE S PACES
z2
When we extended our reasoning to large state spaces
in Section 2.3 using a reduced state space model (see
Fig. S-1), we inevitably arrived at the following questions:
By modeling the expert behavior only along observed
trajectories, what does the resulting model imply for the
remaining states of the state space? Can we still use it for
predicting their local policies? The purpose of this section
is to provide an in-depth discussion on the implications of
this reduced modeling approach in the context of policy
prediction.
When investigating the above-mentioned questions from a
probabilistic perspective (i.e. by analyzing the induced joint
distribution of our model), it turns out that they are strongly
related to what is known as marginal invariance [1] (sometimes also referred to as marginalization property or simply
consistency [2]). This property states that a model is consistent in the sense that it always provides the same marginal
distributions for any subset of its variables, irrespective of
the initial model size. In other words, a marginally invariant
policy model yields the same answer for the given trajectory
points, even if we include additional states into our reduced
set Se for which we have not observed any demonstrations.
For our spatial models, that is, the Potts model and
the ddCRP, it can be shown that this consistency property
is indeed lacking (see [1] for a detailed discussion). This
means that we cannot expect to get compatible results when
conducting our reduced model inference on two data sets
of different sizes. On the contrary, making predictions for
new states would always require to rerun our Gibbs sampler
on the augmented data set, including all additional states.
This brings us to the following practical dilemma: imagine
an on-line policy recognition scenario where we observe an
expert controlling our system. After a certain period of time,
we are asked to take over control, using the experience we
have acquired during the observation period. Each control
command, whether performed by the expert or by us, will
trigger a new state transition, meaning that new data points
arrive sequentially one after another. Consequently, it is
impossible to decide in advance which states to include in
S
s2
s3
z3
s1
z4
s7
s4
z1
z5
s5
s8
z8
z7
s6
z6
z9
s9
(a) reduced state space model
zt
zt+1
zt+2
pz
st
st+1
T
at
at+1
st+2
T
at+2
π
π
π
α
pθ
Θ
(b) corresponding factor graph
Fig. S-1: (a) Illustration of the reduced state space model,
which operates on the space Se = {s1 , s2 , . . . , sT } of visited
trajectory states. Note that the underlying decision-making
process is assumed to be discrete in time; the continuous
gray line shown in the figure is only to indicate the temporal
ordering of the trajectory states. (b) Corresponding factor
graph, highlighting the circular dependence between the
variables. The factors are defined by the same building
blocks that are used for the finite state space model. Observed variables are shaded in gray.
2
-15
1.6
-10
1.2
0
1
5
5
-5
0
5
10
-10
7
15
-15 -10
(a) mean prediction error
6
5
0
4
5
3
10
2
15
1
0.2
15
0.4
-15 -10
0.4
10
0.6
15
0.6
0
0.8
10
8
-5
-5
EMD
-5
-15
0.8
1.4
EMD
-10
1
controller
1.8
-15
-5
0
5
10
15
-15 -10
(b) standard deviation
-5
0
5
10
15
(c) sample partitioning
0.4
probability
probability
0.6
0.3
0.2
0.1
5
10
15
0.2
0
0
0
0
0.4
20
5
number of controllers
10
action 15
s
(d) inferred model complexity
20
25
8
7
6 rs
5
4 rolle
3
2 cont
1
(e) posterior sample of the local controllers
Fig. S-2: Simulation results for the ddCRP model on the continuous state space task described in Section 4.1. (a) Mean
values of the spatial policy prediction error. (b) Standard deviations of the spatial policy prediction error. (c) Example
partitioning of the state space, based on the local controllers depicted in sub-figure (e). The expert partitioning is shown
in Fig. 4 of the main paper. (d) Posterior distribution of the number of local controllers. (e) Posterior sample of the local
controllers found by the model. The results in (a,b,d) are based on 100 Monte Carlo runs while (c,e) are obtained from a
single posterior sample. The figures in the top row were rendered using a spatial resolution of 2000x2000.
our reduced space Se and which not. A rigorous approach in
the above-described sense would thus require to recalibrate
the model after each state transition – a costly operation.
However, it is evident that the resulting data set is
naturally divided into two disjoint parts, namely the expert
demonstrations and the subsequent states reached during
execution of the learned policy. Clearly, transitions occurring
after the learning phase should by no means affect our
belief about the expert policy and, hence, they should be
completely discarded from the model. The easiest way to
achieve this is, indeed, to “freeze” the model after the
demonstration phase and to use the learned parameters to
extrapolate the gathered policy information to surrounding
states. This can be done, for instance, by retaining the
structure of the involved spatial prior model to compute the
resulting maximum a posteriori estimates for the extrapolated
indicators of the new states, based on the inferred model
parameters. In the case of the ddCRP, this coincides with
the nearest-neighbor estimate (see Eq. (16)),
ĉnew = arg max f (dt,new ) = arg min dt,new .
t∈{1,...,T }
(S-1)
t∈{1,...,T }
Herein, ĉnew is the estimate for the indicator of the new
state and dt,new denotes the distance of that state to the tth
trajectory point.
Now, one could argue that the comfort of retaining a
finite model structure for modeling inference problems on
countably infinite or uncountable state spaces comes at the
cost of not being able to provide a consistent posterior
predictive distribution. However, the reduced state space
approach allows us to incorporate the spatial information
of the data in a fairly natural manner (i.e. in the form of
pairwise distances), providing an easy way to model the
expert behavior. Furthermore, our results demonstrate that
the reduced model is able to capture the relevant spatial
properties of a policy sufficiently accurate in order to make
profound predictions about unseen states (see also subsequent section). Whether there exist alternative tractable
models with similar properties remains to be seen.
B
A DDITIONAL S IMULATION R ESULTS
In this section, we provide additional simulation results
for the ddCRP model on the continuous state space task
described in Section 4.1.
Figure S-2a visualizes the spatial EMD prediction errors
of the trained model in the form of a heat map, which
compares the ground truth expert policy at non-trajectory
points with the mean prediction provided by our model.
The test points are placed on a regular grid of size 2000x2000
centered around the origin. The required indicator variables
3
at the interpolated states are computed according to Eq. (S1). In line with our expectation, the prediction error reaches
its maximum at the policy boundaries but is comparably
small within each policy region, indicating a good model
fit. Note that the “windmill shape” of the error can be
explained as a result of the reduced state space approach in
combination with the inherent asymmetry of the used data
generation scheme: regions of the state space containing trajectory endings are locally underrepresented in the data set
(see example trajectory in Fig. 4 in the paper); this increases
the chance of assigning the end points of a trajectory to the
cluster of the preceding region, resulting in a smearing of
the previous cluster into the next region.
Also, we can observe that the variance of the error
(Fig. S-2b) reaches its maximum at the transition regions
and generally grows with the distance to the supporting
trajectory data, reflecting the increasing prediction uncertainty at cluster boundaries and regions far from the expert
demonstrations. Both figures were computed based on the
learned policy representations of 100 Monte Carlo runs. Figure S-2c illustrates an example state partitioning of one such
experiment, using the inferred local controllers depicted in
Fig. S-2e. The result reveals that all expert motion patterns
could be identified by our model. Note, however, that the
two figures correspond to a single Gibbs sample of the
process, which is not representative for the whole posterior
distribution. Averaging over several experiments as done in
Fig. S-2a and Fig. S-2b is not possible at the sample level due
to the varying dimensionality of the corresponding policy
representations (i.e. the number of learned controllers). Even
taking averages over samples of equal dimensionality is
not meaningful due to the multimodality of the posterior
distribution, which arises from the inherent symmetry of
the representation (i.e. interchanging two local controllers
together with their corresponding indices yields the same
model). Hence, averaging samples is possible only at the
prediction level.
Finally, Fig. S-2d depicts the posterior distribution of the
number of local controllers used by the model, which shows
a pronounced peak at the true number used by the expert.
C
C OMPUTATIONAL C OMPLEXITY
The overall computational cost of performing inference in
our model depends largely on two factors: the complexity
per Gibbs iteration and the mixing speed of the underlying
Markov chain. Each Gibbs iteration consists of up to three
stages: 1) sampling T categorical action variables {at } from
the set {1, . . . , |A|}, where T is the size of the demonstration set; 2) ddCRP model: sampling NS categorical state
assignments {ci } from the set {1, . . . , NS }, where NS is
e ); remaining models:
the number of states (i.e. |S| or |S|
sampling NS categorical partition assignments {zi } from
the set {1, . . . , K}, where K is the number of local controllers; 3) for non-collapsed models: sampling K Dirichletdistributed control parameters {θ k } on the (|A| − 1)simplex.
Collapsing the control parameters generally improves
the mixing speed of the chain (see Fig. 5 in the paper) but
requires that action variables belonging to the same cluster
be updated sequentially; hence, a non-collapsed strategy can
be advantageous for larger data sets. Sampling the variables
{at }, {θ k } and {zi } is computationally cheap because the
involved action likelihoods {T (s0 | s, a)} as well as the
neighborhood structure N (Potts model) and the similarity
values {f (di,j )} can be pre-computed. The most demanding
operation is the update of {ci }, which requires tracking
the connected components of the underlying ddCRP graph.
Using an appropriate graph representation, this can be done
in polylogarithmic worst case time [3].
R EFERENCES
[1] D. M. Blei and P. I. Frazier, “Distance dependent Chinese restaurant
processes,” The Journal of Machine Learning Research, vol. 12, pp.
2461–2488, 2011.
[2] C. E. Rasmussen, “Gaussian processes for machine learning.” MIT
Press, 2006.
[3] B. M. Kapron, V. King, and B. Mountjoy, “Dynamic graph connectivity in polylogarithmic worst case time,” in Proc. 24th Annual
ACM-SIAM Symposium on Discrete Algorithms, 2013, pp. 1131–1142.
| 3cs.SY
|
Hyper-heuristics Can Achieve Optimal Performance
for Pseudo-Boolean Optimisation∗
arXiv:1801.07546v1 [cs.NE] 23 Jan 2018
Andrei Lissovoi, Pietro S. Oliveto, and John Alasdair Warwicker
Department of Computer Science
University of Sheffield
Sheffield, S1 4DP
United Kingdom
Abstract
Selection hyper-heuristics are randomised search methodologies which choose and execute heuristics
from a set of low-level heuristics. Recent research for the LeadingOnes benchmark function has shown
that the standard Simple Random, Permutation, Random Gradient, Greedy and Reinforcement Learning
selection mechanisms show no effects of learning. The idea behind the learning mechanisms is to continue to exploit the currently selected heuristic as long as it is successful. However, the probability that a
promising heuristic is successful in the next step is relatively low when perturbing a reasonable solution
to a combinatorial optimisation problem. In this paper we generalise the ‘simple’ selection-perturbation
mechanisms so success can be measured over some fixed period of time τ , rather than in a single iteration.
We present a benchmark function where it is necessary to learn to exploit a particular low-level heuristic,
rigorously proving that it makes the difference between an efficient and an inefficient algorithm. For
LeadingOnes we prove that the Generalised Random Gradient, and the Generalised Greedy Gradient
hyper-heuristics achieve optimal performance, while Generalised Greedy, although not as fast, still outperforms Random Local Search. The performance of the former two hyper-heuristics improves as the
number of operators to choose from increases, while that of the Generalised Greedy hyper-heuristic does
not. Experimental analyses confirm these results for realistic problem sizes and shed some light on the
best choices of the parameter τ in various situations.
1
Introduction
Many successful applications of randomised search heuristics to real-world optimisation problems have been
reported. Despite these successes, it is still difficult to decide which particular search heuristic is a good
choice for the problem at hand and what parameter settings should be used. In particular, while it is well
understood that each heuristic will be efficient on some classes of problems and inefficient on others [23],
very little guidance is available explaining how to choose an algorithm for a given problem. The high level
idea behind the field of hyper-heuristics is to overcome this difficulty by evolving the search heuristic for the
problem rather than choosing one in advance (or applying several arbitrary ones until a satisfactory solution
is found). The overall goal is to automate the design and the tuning of the algorithm and parameters for
the optimisation problem, hence achieving a more generally applicable system.
Hyper-heuristics are usually classified into two categories: selection hyper-heuristics and generation hyperheuristics [5]. The former repeatedly select from a set of low-level heuristics to create new solutions for the
problem, while the latter generate new heuristics from components of existing heuristics. The low-level
heuristics can be further classified as either construction heuristics, which build a solution incrementally,
or perturbation heuristics, which start with a complete solution and try to iteratively improve the current
solution.
An extended abstract of this manuscript has appeared at the 2017 Genetic and evolutionary Computation Conference
(GECCO 2017) [15].
∗
1
Despite the numerous successes for NP-hard optimisation problems, including scheduling [6, 7, 10],
timetabling [21], vehicle routing [3], cutting and packing [16], the theoretical understanding of selection
hyper-heuristics is very limited. Some insights into the behaviour of selection heuristics have been achieved
via landscape analyses [17, 18, 19]. Concerning their performance, [14] presented the first runtime analysis of
selection-perturbation hyper-heuristics. They considered a Simple Random hyper-heuristic [6] that at each
step randomly chooses between low-level heuristics and presented an example benchmark function class,
called GapPath, where it is necessary to use more than one low-level heuristic to optimise the problem.
Similar results have also been presented by [11].
A comparative time-complexity analysis of selection hyper-heuristics has recently been presented by [1].
They considered several common selection mechanisms, namely Simple Random, Permutation, Random
Gradient and Greedy [6, 7] and analysed their performance on the standard LeadingOnes benchmark
function when using a low-level heuristic set consisting of a 1-bit flip and a 2-bit flip operator (i.e., the
same set previously considered by [14]). The runtime analysis results show that the four ‘simple’ selection
mechanisms have the same asymptotic runtime, while experimental trials indicate that the runtimes are
practically equivalent. Recently, additive Reinforcement Learning selection was, under mild assumptions,
also shown to often have asymptotically equivalent performance to Simple Random selection, including for
the same problem setting (i.e., LeadingOnes selecting between 1-bit and 2-bit mutation operators) [2].
In particular, the results indicate that selection mechanisms such as additive Reinforcement Learning and
Random Gradient do not learn to exploit the more successful low-level heuristics and end up having the
same performance as Simple Random selection.
The main idea behind the learning mechanisms considered above is to continue to exploit the currently
selected heuristic as long as it is successful. Unlike construction heuristics, where iterating a greedy move
on a currently successful heuristic may work for several consecutive construction steps, the probability that
a promising heuristic is successful in the next step is relatively low when perturbing a reasonable solution to
a combinatorial optimisation problem.
In this paper, we generalise the ‘simple’ selection-perturbation learning mechanisms analysed by [1]
so success can be measured over some fixed period of time τ (i.e., Exploitation Stage), rather than in a
single iteration. Since the Simple Random and Permutation mechanisms do not use any feedback from the
optimisation process (i.e., do not attempt to learn), we only generalise the Greedy and the Random Gradient
learning mechanisms.
We first show that the learning period τ of the generalised hyper-heuristics is necessary, by presenting
an example function, Generalised GapPath with Traps (GGPT), inspired by the GapPath function of [14].
We initially show that the generalised hyper-heuristics are efficient on GapPath, before generalising the
function such that it is necessary to learn to prefer one operator over another. Then we add points of high
non-optimal fitness throughout the function, and show that the ‘simple’ mechanisms fail on the function
with overwhelming probability, while proving that the generalised hyper-heuristics find the global optimum
of the function with overwhelming probability, given a sufficiently large learning period τ = Ω(n3 ).
Afterwards we use the LeadingOnes (LO) benchmark function to show that the hyper-heuristics can
be surprisingly fast. We first give exact runtime results for the ‘simple’ mechanisms on LO, proving that
2
2
they all have expected runtime ln(3)
2 n + o(n ), confirming what was suggested by the experimental analysis
of [1]. This result indicates that all the ‘simple’ mechanisms essentially choose operators at random in each
iteration, and perform worse than the single operator that always flips 1-bit (i.e., RLS) which has a runtime
on LO of 12 n2 . We then prove upper bounds on the same function for the Generalised Random Gradient
(GRG) and Generalised Greedy (GG) hyper-heuristics. We rigorously prove that both the generalised hyperheuristics have a better runtime than RLS up to lower order terms for appropriately chosen values for the
parameter τ . Furthermore, the GRG hyper-heuristic can achieve the proven best possible runtime up to
lower order terms for an algorithm using just two mutation operators, 1+ln(2)
n2 + o(n2 ), for τ that satisfies
4
1
the GG hyper-heuristic,
we show that for a
both τ = ω(n) and τ ≤ ( 2 − ǫ)n ln(n), for ǫ = Θ(1). Concerning
arctan(2)
n2 + o(n2 ) ≈ 0.47797n2 + o(n2 ).
learning period τ ∈ ω(n) ∩ o(n2 ), its runtime is at most ln(5)
8 +
4
The tightness of the upper bound is shown by proving a matching runtime for an ideal Greedy algorithm
which performs a Decision Stage ‘for free’ in every iteration. This proves that GG is not as fast as GRG for
LO, but is still faster than either of its low-level heuristics applied alone.
Next we introduce a new hyper-heuristic, that we call Generalised Greedy Gradient (GGG), which com-
2
Algorithm 1 Simple Selection Hyper-heuristic [6, 7, 1]
1: Choose s ∈ S uniformly at random
2: while stopping conditions not satisfied do
3:
Choose h ∈ H according to the learning mechanism
4:
s′ ← h(s)
5:
if f (s′ ) > f (s) then
6:
s ← s′
bines the learning ideas of both the GRG and GG hyper-heuristics. In particular, it uses the Decision Stage
of GG and the Exploitation Stage of GRG. We show that it is just as efficient as the other generalised
hyper-heuristics on GapPath, and GGPT for τ = Ω(n3 ) w.o.p.. Furthermore it requires lower learning
periods τ to match the performance of GRG for LO.
Since all the presented results consider hyper-heuristics that choose between only two operators, we
present an analysis of the hyper-heuristics for LeadingOnes when they are allowed to choose between
k = o(n) operators. We first show that including more operators is detrimental to the performance of the
‘simple’ mechanisms. On the other hand, we prove that the performance of the GRG hyper-heuristic improves
as the number of available operators to choose from increases. In particular, when using k operators, GRG
is faster than the best possible performance achievable by using only k − 1 operators, for k = o(n).
We conclude the paper with an experimental analysis of the hyper-heuristics for realistic problem sizes.
The experiments confirm that all generalised hyper-heuristics outperform the single operators and that GRG
and GGG are faster than GG on LO. For two operators, it is shown how proper choices for the parameter τ can
lead to the near optimal performance, while a relationship between the performance of the hyper-heuristics
and the problem size n is suggested. For hyper-heuristics that are able to choose between more than two
operators, the experiments show that the performance of GRG and GGG improves with more operators
with appropriate choices of τ . The performance of the GG hyper-heuristic is worsened by including more
operators, yet it still outperforms RLS.
The rest of the paper is structured as follows. In Section 2, we formally introduce the hyper-heuristic
framework together with the ‘simple’ selection hyper-heuristics and the generalised ones. In Section 3 we
present the runtime results for the ‘simple’ and generalised hyper-heuristics on the GapPath and GGPT
functions. In Section 4 we analyse the ‘simple’ and generalised hyper-heuristics on the LeadingOnes
benchmark function. In Section 5 we introduce and theoretically analyse the Generalised Greedy Gradient
hyper-heuristic. In Section 6 we present the results for hyper-heuristics that choose between more than two
operators. Section 7 presents the experimental analysis. In the conclusion we present a discussion and some
avenues for future work.
Compared to its conference version [15] this manuscript has been considerably extended. The results of
Section 4 have been generalised for any values of τ less than ( 21 − ǫ)n ln n, for ǫ = Θ(1) and n2 for GRG and
GG respectively. In general, the results have been considerably strengthened. In particular, compared to the
extended abstract, in the present manuscript we prove that GRG and GGG achieve optimal runtimes for
LO. Section 5 and Section 6 are completely new additions, while a more comprehensive set of experiments
is included in Section 7.
2
Preliminaries
Let S be a finite search space, H a set of low-level heuristics and f : S → R+ a cost function. Algorithm 1
shows the pseudocode representation for a ‘simple’ selection hyper-heuristic as used in previous theoretical
analyses [1].
The following learning mechanisms have been commonly used in the literature to solve combinatorial
optimisation problems [6, 7]: Simple Random, which selects a low-level heuristic independently with
probability ph in each iteration (usually ph = 1/|H|, i.e., uniformly at random); Permutation, which
generates a random ordering of low-level heuristics and returns them in that sequence when called by the
hyper-heuristic; Greedy, which applies all available low-level heuristics in parallel and returns the best
found solution; Random Gradient, which randomly selects a low-level heuristic, and keeps using it as long
3
as it obtains improvements.
[1] analysed the runtime of these four ‘simple’ mechanisms1 for the LeadingOnes benchmark function,
choosing between flipping either one, or two, randomly chosen bits of the bit string (H = {1BitFlip, 2BitFlip}).
We derive the exact runtimes in Section 4. It was shown experimentally by [1] that all mechanisms have the
same performance as just choosing low-level heuristics at random. By making a heuristic selection decision in
every iteration without taking past performance into account, the ‘simple’ mechanisms do not have enough
time to learn which operator is preferable at the current optimisation stage.
In this paper we generalise the ‘simple’ mechanisms to allow a longer time period to decide whether a
low-level heuristic is currently successful or not, aiming to maintain the intrinsic ideas of the ‘simple’ learning
mechanisms while generalising sufficiently to allow learning to take place. In particular, we consider:
Generalised Greedy (GG), where all low-level heuristics are tested on the same candidate solution
until an improvement is found (Decision Stage). The improvement (chosen uniformly at random if there
are multiple ones) is then accepted, and the corresponding operator is run for a fixed period of time τ
(Exploitation Stage).
Generalised Random Gradient (GRG), where a low-level heuristic is chosen uniformly at random
(Decision Stage) and run for a period of fixed time τ . If an improvement is found before the end of the
period, then a new period of time τ is immediately initialised (Exploitation Stage). If the chosen operator
fails to provide an improvement in τ iterations, a new operator is chosen at random.
3
Learning is Necessary
In this section we consider the GapPath function (GP) previously introduced by [14] as an example to show
that applying multiple low-level heuristics may be necessary. We prove that the generalised hyper-heuristics
are also efficient on the problem. We then generalise the GP function to make one operator useful more
often than the other, and add points of very high non-optimal fitness along the path that ‘trap’ algorithms
that have not learned to prefer one operator to the other. The result will be that the ‘simple’ mechanisms
fail w.o.p., while the generalised hyper-heuristics can find the global optimum efficiently w.o.p..
3.1
GapPath
The GapPath function is defined as follows [14]:
(
ZM(x)
if Ridge(x) mod 3 = 1
GP(x) :=
ZM(x) + 2nRidge(x) otherwise
Pn
where ZM(x) := i=1 (1 − xi ), and Ridge(x) := i if x = 1i 0n−i , and 0 otherwise.
The function consists of a short path which corresponds to all search points of the form 1i 0n−i . The path
contains gaps where the function values are inferior to the rest of the path. Such gaps consist of all path
points where i ≡ 1 (mod 3). As a result of these gaps, an algorithm that only uses either the 1BitFlip or
the 2BitFlip operator will have infinite expected runtime, since it is necessary to alternate the operators
in order to make progress on the path. [14] prove that the runtime of the Simple Random mechanism,
initialised on the 0n bit
and usingonly the 1BitFlip and 2BitFlip operators with probability p and
string
3
n2
−3n2
for any probability p ∈ (0, 1). In the following two theorems we
+ 3(1−p)p
1 − p respectively, is n6(1−p)
show that the generalised hyper-heuristics are also efficient on GapPath when choosing between the same
two operators.
Theorem 1 The expected runtime of the Generalised Greedy hyper-heuristic, when initialised on the 0n bit
string, for GapPath is n3 /3 + 2n2 /3 + 2τ n/3 − τ .
Proof: The GP function is optimised by requiring multiple repetitions of 1-bit and 2-bit consecutive
improvements. There is only one possible successful move at each point on the path and only the use of the
correct operator will lead to an improvement.
1 Throughout
the paper, bounds on the runtimes of the considered ‘simple’ mechanisms refer to the number of fitness
evaluations performed by Algorithm 1 using these mechanisms.
4
If a 1-bit improvement is necessary, the success probability of the 1BitFlip operator is 1/n, implying
n expected 1-bit mutations before an improvement is found (by standard waiting time arguments). During
this time, n 2-bit mutations will be performed in parallel, as the Generalised Greedy hyper-heuristic applies
both operators during the Decision Stage.
When a 2-bit improvement is necessary, the success probability of the 2BitFlip operator is 1/n · 1/n · 2
(since two specific bits must be flipped, but can be chosen in either order) implying n2 /2 expected 2-bit
mutations before an improvement is found. The same number of 1-bit mutations will be evaluated in parallel.
After an operator has succeeded, the Generalised Greedy hyper-heuristic will deterministically apply
the successful operator (on its own) for another τ iterations (the Exploitation Stage), during which no
improvement can occur (due to the nature of the GP function), before starting the next Decision Stage.
Initially, a 2-bit improvement is required. The expected number of fitness function evaluations needed
to reach the first point on the path 110n−2 from the 0n bit string is 2 · n2 /2 = n2 , which will be followed by
τ non-improving steps. The next improvement will then occur by the 1BitFlip operator after an expected
2n steps, followed by τ unsuccessful 1-bit flips. Hence, the expected time to optimise the first three bits is
n2 + τ + 2n + τ = n2 + 2n + 2τ .
This three bit pattern occurs a total of n/3 times before the optimum is reached, and since the global
optimum is reached just before the final Exploitation Stage starts, the total expected runtime is E(TGG ) =
n/3 · (n2 + 2(n + τ )) − τ = n3 /3 + 2n2 /3 + 2τ n/3 − τ .
Theorem 2 The expected runtime of the Generalised Random Gradient hyper-heuristic, when initialised on
the 0n bit string, for GapPath is at most n3 /3 + 2n2 /3 + 4τ n/3 − τ .
Proof: The Generalised Random Gradient hyper-heuristic chooses a mutation operator uniformly at random, and applies it for a period of τ iterations. If an improvement is produced, a new period of τ iterations
using the same operator is started immediately.
Let T1 and T2 respectively be the number of iterations before a 1-bit or a 2-bit improvement is constructed
when it is possible to do so. If the appropriate operator i is chosen (i.e., with probability 1/2), it may succeed
with probability pi in each iteration, or it may fail during all τ iterations with probability (1 − pi )τ in which
case the random choice is repeated. If the other operator is chosen, τ iterations are wasted before the random
choice is repeated. Combined, using qi := 1 − pi for brevity:
!
!
τ
X
1
1
k−1
τ
kqi pi + qi (τ + E(Ti )) + (τ + E(Ti ))
E(Ti ) =
2
2
k=1
!
τ
X
=
kqik−1 p + qiτ E(Ti + τ ) + τ
(1)
=
k=1
∞
X
j=0
=
∞
X
k=1
τ
X
(k +
jτ )qijτ +k−1 pi
k=1
kqik−1 pi
!
+τ +
!
+
τ qijτ
!
τ qiτ
1
τ qiτ
=
+
τ
+
.
1 − qiτ
pi
1 − qiτ
(2)
(3)
A simple rearrangement yields (1), and applying it recursively to the qijτ E(Ti + jτ ) terms yields (2). As
the inner sums are contiguous, a rearrangement in (3) simplifies the expression to an expectation of a
geometrically-distributed variable and a sum of an infinite geometric series. As the expected number of
τ qiτ
qiτ
periods of time τ where the correct operator is applied but does not produce an improvement is 1−q
τ , 1−q τ is
i
i
a lower bound on the number of iterations where a correct operator is applied before a success (i.e., 1/pi ) and
τ qτ
τ + 1−qi τ is an upper bound on 1/pi . Substituting these bounds into Eq. (3) yields 2/pi ≤ E(Ti ) ≤ 2/pi + τ .
i
After an improvement is constructed, the successful operator will run for an additional τ iterations while
not being able to construct the next improvement. To reach the optimum from 0n , improvements must be
constructed by alternating 2-bit and 1-bit mutations n/3 times, while the final τ -iteration Exploitation Stage
is unnecessary. The total expected optimisation time is therefore:
n
(E(T1 ) + E(T2 ) + 2τ ) − τ ≤ n3 /3 + 2n2 /3 + 4τ n/3 − τ.
3
5
3.2
Generalising GapPath
Having shown that the generalised hyper-heuristics are efficient on GP, we now introduce a function class
on which it is necessary for the hyper-heuristics to learn which operator to use, if they are to be efficient.
We will prove that the proposed generalised hyper-heuristics outperform the ‘simple’ mechanisms.
Let n be of the form n = d(2k + 1) for some d, k ∈ N.
(
ZM(x)
if Ridge(x) ∈ Sk
GGPk (x) :=
ZM(x) + 2nRidge(x)
otherwise
where Sk = {c(2k + 1) + 1 − 2β | c, β ∈ N : c ≤ d, β ≤ k}, and ZM(x) and Ridge(x) defined as in Section 3.
This function is similar to the GapPath function, but requires k consecutive 2-bit successes after the 1-bit
improvement is successfully executed at search points of the form 1(c−1)(2k+1) 0n−((c−1)(2k+1)) for 0 ≤ c ≤ d.
Hence learning to use 2-bit flips often leads to improved performance.
We modify the GGPk function by inserting a ‘trap’ point (point of high non-optimal fitness) at a random
position within each group of k 2-bit flips. These ‘trap’ points can be reached by a 1-bit mutation, while
the next path point remains reachable only by a 2-bit mutation, and are local optima for both mutation
operators. If a ‘trap’ point is ever constructed, the algorithm will not be able to reach the global optimum.
Since there is a finite probability for any mechanism to construct a ‘trap’ point, the expected runtimes will
all be infinite by the law of total probability. Hence, we consider the success probabilities for the ‘simple’
mechanisms and generalised hyper-heuristics on this modified function (denoted GGPT).
Theorem 3 Starting at 0n , all of the ‘simple’ selection mechanisms will fail to find the global optimum on
GGPT in finite time, with probability at least 1 − n−Ω(n) .
Proof: We will show that when each mechanism is at a point from which a ‘trap’ point is accessible, it will
construct the ‘trap’ point before the next path point with probability 1−O(1/n). If this holds, as there are at
least n/(2k + 1) = Ω(n) points from which a ‘trap’ point is accessible, the ‘simple’ mechanisms will therefore
construct a ‘trap’ point before reaching the global optimum with probability 1 − O(1/n)Ω(n) = 1 − n−Ω(n) .
The Greedy mechanism applies both mutation operators in one iteration. The probability that a ‘trap’
point is constructed via a 1-bit mutation is 1/n, while the probability that the next path point is constructed
is 2/n2 . If both are constructed, the mechanism accepts the ‘trap’ point as it has higher fitness. The
1/n
probability that a ‘trap’ point is constructed before the next path point is therefore at least 1/n+2/n
2 =
2
1 − n+2 = 1 − O(1/n).
The Permutation mechanism, following a two-bit success leading to a point from which a ‘trap’ point
is accessible, will perform 1-bit and 2-bit mutations in this order. The probability that a ‘trap’ point is
constructed over two iterations is thus 1/n, while the probability that the next path point is constructed
over two iterations is (1 − 1/n) · 2/n2 . The probability that a ‘trap’ point is constructed before the next path
1/n
1/n
2
point is therefore 1/n+(1−1/n)2/n
2 = 1/n+2/n2 + n2 +4n+2−4/n = 1 − O(1/n).
The Simple Random mechanism constructs a ‘trap’ point with probability 1/(2n), and the next path
point with probability 1/n2 . The probability that a ‘trap’ point is constructed before the next path point is
1/(2n)
2
therefore at least 1/(2n)+1/n
2 = 1 − n+2 = 1 − O(1/n).
The Random Gradient mechanism behaves equivalently to the Simple Random mechanism, unless it
manages to construct an improvement by a two-bit mutation immediately following another two-bit improvement. The probability that an individual ‘trap’ point is skipped in this manner is 1 − 2/n2 = O(1/n),
and thus the probability that the Random Gradient constructs the ‘trap’ point before the next path point
is (1 − O(1/n))2 = 1 − O(1/n).
On the other hand, the following theorem shows that with sufficiently large τ , the generalised hyper-heuristics
can avoid the traps w.o.p..
Theorem 4 Starting at 0n for τ = Ω(n3 ), the Generalised Random Gradient hyper-heuristic will find the
kn3
4τ n
+ 2(2k+1)
+ O(n2 ) steps with probability at least
global optimum of GGPT (with constant k) in at most 2k+1
6
1 − 2−Ω(n) . However, for τ = O(n2 ), the Generalised Random Gradient hyper-heuristic will fail to reach the
global optimum of GGPT in finite time with probability at least 1 − 2−Ω(n) .
Proof: The Generalised Random Gradient hyper-heuristic avoids making a decision inside the two-bit
region if it is able to construct k−1 two-bit improvements within τ iterations each. The probability of success
for a 2-bit mutation is 2/n2 ; the probability of no successes in τ steps is (1 − 2/n2 )τ and the probability of
(at least) one success in a period of τ steps is this value subtracted from 1. Thus, the probability that all
k−1
k − 1 two-bit improvements are found in time is 1 − (1 − 2/n2 )τ
.
cn2 k−1
For τ = cn2 , where c is a constant, the probability of the required k−1 successful periods is 1 − 1 − 2/n2
=
(1 − e−2c − o(1))k−1 = Θ(1), and so the probability of failing to do so is at least constant. If another random
Decision Stage is performed inside the region of two-bit improvements, a ‘trap’ point is accessible via a 1-bit
mutation from the point at which the decision occurs with probability at least 1/(k − 1) = Ω(1) (as traps
are placed uniformly at random inside these regions), and is found by the Generalised Random Gradient
2
hyper-heuristic with probability at least 1/2 · (1 − 1/n)cn ≥ 1/4. Thus, the probability that the no ‘trap’
point is constructed in a single region of k two-bit improvements is at most a constant smaller than one,
which means that over the n/(2k + 1) such regions, the probability that a ‘trap’ point is constructed is at
least 1 − 2−Ω(n) .
For τ = cn3 , where c is a constant, the probability that all T = (k−1)·n/(2k+1) = Θ(n) two-bit successes
cn3 T
are constructed within τ iterations is at least 1 − 1 − 2/n2
= (1−e−2cn)T ≥ 1−T e−2cn = 1−e−Ω(n)
by applying a union bound. Thus, with overwhelming probability, the algorithm is not trapped, and the
expected optimisation time is
k
n
4τ n
kn3
n + 2τ + n2 + 2τ =
E(TGRG | no trap) ≤
+
+ O(n2 ),
2k + 1
2
2k + 1 2(2k + 1)
with the 2τ terms corresponding to using the wrong operator for, in expectation, one period of τ iterations
at the start of the phase, and using the successful operator for τ iterations while the next improvement can
only be constructed by the other operator.
The following theorem can be proven using a similar approach.
Theorem 5 Starting at 0n for τ = Ω(n3 ), the Generalised Greedy hyper-heuristic will find the global optin3
2τ n
+ 2k+1
+ O(n2 ) steps with probability at least 1 − 2−Ω(n) .
mum of GGPT (with constant k) in at most 2k+1
2
However, for τ = O(n ), the Generalised Greedy hyper-heuristic will fail to reach the global optimum of
GGPT in finite time with probability at least 1 − 2−Ω(n) .
Proof: The Generalised Greedy hyper-heuristic avoids making a decision inside the two-bit region if and
only if it can produce k − 1 two-bit improvements within τ iterations following the first two-bit improvement.
Let X be a binomially-distributed random variable with parameters p = 2/n2 (the success probability of the
two-bit mutation operator) and τ trials, and thus E(X) = 2τ /n2 . We note that the probability distribution
of X up to k − 1 is identical to that of the number of two-bit improvements constructed during the Decision
Stage, while X ≥ k − 1 corresponds to constructing all remaining k − 1 improvements.
For τ = cn2 , we note that the probability of constructing exactly zero improvements during the τ
2
iterations is at least (1 − 2/n2 )cn ≥ e−2c − O(1/n2 ) = Ω(1). If another Decision Stage is performed inside
the region of two-bit improvements, a ‘trap’ point is accessible via a 1-bit mutation from the point at which
the decision occurs with probability at least 1/(k − 1) = Ω(1) (as traps are placed uniformly at random inside
these regions), and is found by the Greedy mechanism with probability at least (1/n)/(1/n + 2/n2 ) ≥ 1/4.
Thus, the probability that the Generalised Greedy hyper-heuristic does not construct a ‘trap’ point in a
single region of k two-bit improvements is at most a constant smaller than one, which means that over the
n/(2k + 1) such regions, the probability that a ‘trap’ point is constructed is at least 1 − 2−Ω(n) .
For τ = cn3 , where c is a constant, E(X) = 2cn. Applying a Chernoff bound, we note that the probability
of observing fewer than cn successes in τ steps is at most e−Ω(n) , and if n is sufficiently large, cn > k. Thus,
with overwhelming probability, a decision step does not occur inside a single two-bit region, and by a union
7
bound, does not occur inside any of the n/(2k + 1) two-bit regions in the bit string. Thus, with overwhelming
probability, the algorithm is not trapped, and the expected optimisation time is
E(TGG | no trap) ≤
n
2τ n
n3
2n + τ + n2 + τ =
+
+ O(n2 ).
2k + 1
2k + 1 2k + 1
We point out that, if the function values were inverted such that the ‘trap’ points had optimal fitness, and
the 1n bit string did not, then the generalised hyper-heuristics would fail to find any of the global optima
(w.o.p), while the ‘simple’ mechanisms would succeed (w.o.p). It is not surprising that learning mechanisms
fail on a function especially designed to deceive them.
4
Hyper-Heuristics Are Faster
In this section we show that the generalised hyper-heuristics can outperform low-level heuristics, even
when the latter are efficient for the problem at hand. We consider the LeadingOnes function, LO(x) :=
Pn Q i
i=1
j−1 xj , which counts the number of consecutive one-bits in a bit string before the first zero-bit.
It is well known that the standard (1+1)-EA and RLS algorithms on LeadingOnes have expected
2
2
2
runtimes of e−1
2 n − o(n ) and 0.5n fitness function evaluations respectively [4]. In the following we show
that the generalised hyper-heuristics are more efficient.
We first introduce a lower bound on the runtime of all algorithms which use only the 1BitFlip and
2BitFlip operators. Since the 1BitFlip operator has a success probability of n1 while the 2BitFlip
· 2 = (2n − 2i − 2)/n2 where i = LO(x), it is easy to see that for the first n/2 leading
operator of n1 · n−i−1
n
ones the 2BitFlip operator is more effective, while the 1BitFlip operator is preferable afterwards. Hence,
the expected runtime of an algorithm using these operators in such a way gives a trivial lower bound on all
stochastic unbiased algorithms using only the same two operators. This can be easily proved following the
approach used for the (1+1)-EA by [4].
Theorem 6 ([4]) The expected time needed to find the optimum of LeadingOnes given random initialisation is
n
1X
An−i ,
E(T ) = T n =
2 i=1
where An−i is the expected time needed to find an improvement given a solution with fitness i.
Theorem 7 The best possible expected runtime on LeadingOnes for a stochastic, unbiased algorithm using
only the 1BitFlip and 2BitFlip operators is 1+ln(2)
n2 + O(n) ≈ 0.42329n2 + O(n).
4
Proof: We partition the analysis into two phases: Phase 1, where the 2BitFlip operator is used for the
first half of the search, and Phase 2, where the 1BitFlip operator is used for the second half of the search.
Phase 1 The 2BitFlip operator is stochastic and unbiased, so the result from Theorem 6 holds for this
phase.
To find the expected runtime for this phase, we can consider every point LO(x) ≥ n2 to be optimal. That
n2
, we can consider, for i ≥ n/2, An−i = 0.
is, while for 0 ≤ i < n/2, An−i = 2n−2i−2
Hence, we note that for Phase 1,
E(T1 ) =
n/2−1
n−1
n−1
1X
n2
1 X
1 X n2
An−i =
=
2 i=0
2 i=0 2n − 2i − 2
2
2k
k=n/2
=
n2
4
n−1
X
k=n/2
1
n2
Hn−1 − Hn/2−1 ,
=
k
4
8
where Hx is the xth Harmonic number, which can be bounded by
ln(n + 1) + γ −
1
1
≤ Hn ≤ ln(n + 1) + γ −
,
n+1
2(n + 1)
and γ is the Euler-Mascheroni constant (γ ≈ 0.57722 . . . ). Hence,
3 n2
n2
·
≤ E(T1 ) ≤ ln(2) +
,
ln(2) ·
4
2n
4
i.e.,
ln(2) 2
n + O(n).
E(T1 ) =
4
Phase 2 The runtime for Phase 2 can be calculated using Theorem 6 with An−i = n for i ≥ n/2. Similarly
to Phase 1, we state An−i = 0 for i < n/2.
n/2
n−1
1
1X
1X
n2
n = · n · n/2 =
E(T2 ) =
An−i =
.
2 i=0
2 j=1
2
4
Hence, the total expected runtime is:
E(T ) = E(T1 ) + E(T2 ) =
ln(2) 2
1
1 + ln(2) 2
n + O(n) + n2 =
n + O(n).
4
4
4
The rest of the section is structured as follows. In Subsection 4.1 we show that the ‘simple’ hyper-heuristic
mechanisms all perform equivalently to Simple Random for LeadingOnes, up to lower order terms. In
Subsection 4.2 we show that the generalised hyper-heuristics are much faster.
4.1
‘Simple’ Mechanisms
In this section we show that the standard ‘simple’ selection mechanisms all have the same runtime on LeadingOnes. The following theorem derives the runtime for the Simple Random mechanism. The subsequent
corollary extends the result to all the other ‘simple’ mechanisms.
Theorem 8 The expected runtime of the Simple Random mechanism on LeadingOnes for p ∈ (0, 1) is
1
2
2
4(1−p) ln ((2 − p)/p) n + o(n ). If p = 0 the expected runtime is infinite. If p = 1, the expected runtime is
2
0.5n .
Proof: If p = 0, only the 2BitFlip operator is used. There is a finite probability of reaching the point
1n−1 0, which cannot be improved by using the 2BitFlip operator. Hence, by the law of total probability,
the expected runtime is infinite.
2
If p = 1, only
Pnthe 1BitFlip operator is used, resulting in exactly RLS, which has expected runtime 0.5n .
E(Tp ) = 12 i=1 An−i , where An−i is the expected time needed to find an improvement given a solution
with fitness i, and the algorithm is initialised with a search point chosen uniformly at random. In each iteration, the 1BitFlip operator is chosen with probability p; the 2BitFlip operator is chosen with probability
1 − p. Hence, (An−i )−1 = p · n1 + (1 − p) · 2n−2i−2
, and
n2
An−i =
1
p
n
+
(1−p)(2n−2i−2)
n2
=
n2
2(1 − p)(n − i − 1) + np
The total expected runtime is
n−1
n
n2
n2 X
1
1X
=
2 i=0 2(1 − p)(n − i − 1) + np
2
(2p − 2)k + (2 − p)n
k=1
!
!
n + (2−p)n
n2
n2 ln((2 − p)/p)
2p−2
ln
+ o(n2 ),
=
+
o(1)
=
(2−p)n
2(2p − 2)
4(1 − p)
E(T ) =
2p−2
9
(4)
(5)
where Equation 4 hbecomes
Equation
5i via the fact that
Pn
n+bn/a
1
1
+ o(1) with a = 2p − 2 and b = 2 − p.
k=1 a·k+b·n = a ln
bn/a
When p = 0.5 (i.e., equal chance of choosing each operator in each iteration), the standard Simple Random
2
2
mechanism has expected runtime ln(3)
2 n + o(n ). The runtime improves with increasing p, hence the optimal
choice is p = 1.
Corollary 9 The expected runtime of the Permutation, Greedy and Random Gradient mechanisms on
2
2
LeadingOnes is ln(3)
2 n + o(n ).
Proof: For the Greedy and Permutation mechanisms, let pi = 1/n + (1 − 1/n)(2n − 2i − 2)/n2 be the
probability that at least one improvement is constructed in respectively one or two iterations. To upper bound
the optimisation time, we note that the difference between 2/pi (i.e., the waiting time for an improvement
to be constructed in terms of fitness evaluations) and the An−i waiting times used to prove Theorem 8 is
at most constant, and thus the difference between these mechanisms and Simple Random expected times is
limited to lower-order terms. We note that with probability at most 2/n2 , both mutations considered are
improvements; in such a case, we can consider an improvement step to be made in 0 iterations. As this
occurs at most a constant number of times in expectation, and the maximum waiting time for any improving
step is O(n), the lower-bound differs from the upper bound by at most an O(n) term. Thus, the expected
runtime of these mechanisms is also ln(3)/2 · n2 + o(n2 ).
For the Random Gradient mechanism, we note that the probability that an operator, when repeated
following a success, is successful again is at most 2/n. Since there are at most n improvements to be made
throughout the search space, the expected number of repeats which produce an improvement is at most
2. When not repeating a successful operator, the Random Gradient mechanism behaves identically to the
Simple Random mechanism. Its expected runtime is therefore at least the runtime of the Simple Random
mechanism less an O(n) term, and at most the runtime of the Simple Random mechanism plus n (as there
are at most n iterations where the mechanisms differ in operator selection). Thus, its expected runtime is
also ln(3)/2 · n2 + o(n2 ).
4.2
Generalised Hyper-heuristics
In this section we present a rigorous theoretical analysis of the generalised hyper-heuristics on LeadingOnes.
We start by introducing some important tools which will be used in the analyses.
Theorem 10 Wald’s Equation [22, 9] Let T be a random variable with bounded expectation and let
X1 , X2 , . . . be non-negative random variables with E(Xi |T ≥ i) ≤ C. Then
#
" T
X
Xi ≤ E[T ] · C.
E
i=1
Theorem 11 Additive Drift Theorem [12, 20] Let {Xt }t≥0 be a Markov process over a finite set of
states S, and g : S → R+
0 a function that assigns a non-negative real number to every state. Let the time to
reach the optimum be T := min{t ≥ 0 : g(Xt ) = 0}. If there exists δ > 0 such that at any time step t ≥ 0
and at any state Xt ≥ 0 the following condition holds:
E(∆(t)|g(Xt ) > 0) = E(g(Xt ) − g(Xt+1 )|g(Xt ) > 0) ≥ δ,
then
E(T |X0 ) ≤
and
E(T ) ≤
g(X0 )
δ
E(g(X0 ))
.
δ
Theorem 12 Variable Drift Theorem [13] Let (Xt )t≥0 be a stochastic process over some state space S ⊆
{0} ∪ [xmin , xmax ], where xmin > 0. Let h(x) be an integrable, monotone increasing function on [xmin , xmax ]
10
such that E(Xt − Xt+1 | Xt ≥ xmin ) ≥ h(Xt ). Then it holds for the first hitting time T := min{t | Xt = 0}
that
Z X0
xmin
1
E(T | X0 ) ≤
+
dx.
h(xmin )
xmin h(x)
4.2.1
Generalised Random Gradient
The following theorem considers the Generalised Random Gradient hyper-heuristic. The theorem allows us
to show that the expected runtime of the hyper-heuristic for the function is the optimal runtime that may
be achieved by using two operators. This result will be highlighted in Corollary 15 and depicted in Figure 1.
Theorem 13 The
expected runtime of the Generalised Random Gradient hyper-heuristic on LeadingOnes
with τ ≤ 12 − ǫ n ln(n), for ǫ = Θ(1), is at most
τ
τ
0.5
w 2τ + e n M (1) + e2 n (1−(j−1)/w) M
2
X
n
1−j/w
n
+ o(n2 )
τ
τ
2 j=1
(e n + e2 n (1−j/w) − 2)w
where M (x) := min( nτ , x), for any w ∈ Ω(1) ∩ o
n
exp(2τ /n)
.
Proof: For the purpose of this proof, we partition the optimisation process into w chunks based on the
value of the LeadingOnes fitness function: during chunk j, the LO value of the current solution is at least
(j − 1)n/w and less than j · n/w. After all the w chunks have been completed, the global optimum, with an
LO value of n, will have been found.
We upper-bound the expectation of the runtime T of the Generalised Random Gradient hyper-heuristic
on LO by the sum of the expected values of Tj , the runtimes on each chunk of the optimisation process,
plus the times taken to begin a chunk. As our analysis
Pw of E(Tj ) requires the chunk to start with a random
choice of mutation operator, we bound E(T ) ≤ j=1 (E(Tj ) + E(Sj+1 )) where Sj+1 is a random variable
denoting the expected number of iterations between the first solution in chunk j +1 being constructed and the
following random operator choice. We will later show that with proper parameter choices, the contribution
of the Sj+1 terms can be bounded by o(n2 ), and therefore does not affect the leading constant in the overall
bound.
Let us now consider Tj , the number of iterations the hyper-heuristic spends in chunk j. Recall that
a mutation operator is selected uniformly at random, and is allowed to run until it fails to produce an
improvement for τ sequential iterations. Let Nj be a random variable denoting the number of random
operator choices the hyper-heuristic performs during chunk j, and Xj,1 , . . . , Xj,Nj be the number of iterations
PNj
Xj,k , and by applying Wald’s
the hyper-heuristic runs each chosen operator for. In which case, Tj = k=1
equation (Theorem 10), using E(Xj ) to denote an upper bound on all E(Xj,k ) in chunk j:
E(Tj ) =
PNj
k=1 E(Xj,k )
≤ E(Nj )E(Xj ).
(6)
To bound E(Nj ), the expected number of times the random operator selection is performed during chunk
j, we lower-bound the expected number of improvements found following the operator selection, and apply
the Additive Drift Theorem [12] to find the expected number of random operator selections occurring before
a sufficient number of improvements have been found to enter the next chunk.
Let F1 and F2 denote the events that the 1BitFlip and 2BitFlip operators fail to find an improvement
during τ iterations. For the 1BitFlip operator, this event occurs with probability P (F1 ) = (1 − 1/n)τ
τ
τ
throughout the process, which is within [e− n − 1/n, e− n ]. For the 2BitFlip operator, recall that during
chunk j, the ancestor individual has at most jn/w − 1, and at least (j − 1)n/w, one bits, and thus:
τ
τ
P (F2 ) ≤ (1 − 2 · (1/n) · (n − (jn/w − 1) − 1)/n) ≤ e−2 n (1−j/w) ,
τ
P (F2 ) ≥ (1 − 2(1/n)(n − (j − 1)n/w − 1)/n)
τ
2(1 − (j − 1)/w)
2τ
> 1−
≥ e− n (1−(j−1)/w) − 1/n.
n
11
We note that a geometric distribution with parameter p = P (F1 ) can be used to model the number of
improvements that the 1BitFlip operator finds prior to failing to find an improvement for τ iterations; the
expectation of this distribution is (1 − p)/p = 1/p − 1. Combined over both operators, the expected number
of improvements Dj produced following a single random operator selection during chunk j, E(Dj ), is greater
than:
1
2τ
1
1
1
τ
−1 +
− 1 ≥ e n /2 + e n (1−j/w) /2 − 1,
2 P (F1 )
2 P (F2 )
by inserting the upper bounds on P (F1 ) and P (F2 ). We use this expectation as the drift in the Additive
Drift Theorem to upper-bound E(Nj ). Recall that each chunk consists of advancing through at most n/w
fitness values; as bits beyond the leading ones prefix and the first zero bit remain uniformly distributed, at
most n/(2w) improvements by mutation are required in expectation. If each step of a random process in
expectation contributes E(Dj ) improvements by mutation, then the expected number of steps required to
complete chunk j is at most:
E(Nj ) ≤
n
n/(2w)
≤ τ
2τ
(1−j/w)
E(Dj )
n
n
(e + e
− 2)w
(7)
by the Additive Drift Theorem.
To bound E(Xj ), the expected number of iterations before a selected mutation operator fails to produce
an improvement for τ iterations, we apply Wald’s equation: let S be the number of improvements found
by the operator before it fails, and W1 , . . . , WS be the number of iterations it took to find each of those
improvements; then, once selected, the 1BitFlip operator runs for:
E(Xj | 1-bit) = τ +
S
X
E(Wk ) = E(S)E(W1 | S ≥ 1, 1-bit)
k=1
where τ accounts for the iterations immediately before failure, and the sum for the iterations preceding each
constructed improvement. Recall that E(S) = 1/P (F1 ) − 1 by the properties of the geometric distribution,
and observe that E(W1 | S ≥ 1) = E(W1 | W1 ≤ τ ) ≤ min(τ, E(W1 )). Using a waiting time argument for
E(W1 ) being equal to the reciprocal of the improvement probability for the 1BitFlip operator, we get (with
a similar argument for the two-bit mutation operator):
E(Xj | 1-bit) ≤ τ + (1/P (F1 ) − 1) min(τ, n),
E(Xj | 2-bit) ≤ τ + (1/P (F2 ) − 1) min (τ, n/(2(1 − j/w))) .
Combining these conditional expectations with lower bounds on P (F1 ) and P (F2 ) yields
E(Xj ) ≤ 1/2 · E(Xj | 1-bit) + 1/2 · E(Xj | 2-bit)
≤τ+
min(τ, n/(2 − 2j/w))
min(τ, n)
,
+
τ
2τ
2(e− n − 1/n) 2(e− n (1−(j−1)/w) − 1/n)
(8)
which can be simplified to:
min( nτ , 1/(2 − 2j/w))
min( nτ , 1)
n 2τ
+
E(Xj ) ≤
+ −τ
2τ
2 n
e n − 1/n e− n (1−(j−1)/w) − 1/n
τ
τ
τ
2τ
n 2τ
+ O(1)
+ e n min
, 1 + e n (1−(j−1)/w) min
, (2 − 2j/w)−1
≤
2 n
n
n
using a/(b − c/n) = a/b + ac/(b2 n − bc) = a/b + O(1/n), where a, b, and c are constants with respect to n,
to limit the contributions of the −1/n terms in denominators to lower-order terms.
Finally, we can bound the overall runtime of the hyper-heuristic. Recall that Sj denotes the number
of iterations the hyper-heuristic spends in chunk j while using the random operator chosen during chunk
j − 1, and as E(Sj ) ≤ max (E(Xj | 1-bit), E(Xj | 2-bit)) < 2E(Xj ) = O (n exp(2τ /n)), and since w =
12
P
w
n
2
o exp(2τ
j=1 E(Sj+1 ) = o(n ). Substituting the bounds (7) and (8) into (6) yields the theorem
/n) ,
statement:
w
w
X
X
E(Nj )E(Xj ) + o(n2 )
(E(Tj ) + E(Sj+1 )) ≤
E(T ) ≤
j=1
j=1
w
n2 X
≤ o(n2 ) +
×
2
j=1
2τ
n
2τ
τ
n
+ e min( nτ , 1) + e n (1−(j−1)/w) min
τ
2τ
(e n + e n (1−j/w) − 2)w
τ
0.5
n , 1−j/w
.
We can use this result to get a tight upper bound on the runtime of the simple Random Gradient mechanism.
We know the bound is tight by Theorem 8.
Corollary 14 The expected runtime of the Generalised Random Gradient hyper-heuristic on LeadingOnes,
1
with τ = 1 (i.e., ǫ = 21 − n ln
n ), is at most
ln(3) 2
n + o(n2 ).
2
Proof: We constrain w = ω(log(n)) ∩ o(n). Note that for τ = 1, we have min(τ /n, 1) = min(τ /n, 0.5/(1 −
j/w)) = 1/n. For w = ω(log(n)), we have e2/w = 1 + o(1).
If we plug in τ = 1 into our result, we obtain
1
1
w
n2 X n2 + e n · n1 + e2 n (1−(j−1)/w) · n1
E(T ) ≤
+ o(n2 )
1
1
2 j=1
(e n + e2 n (1−j/w) − 2)w
2
1
w
n X e n + e2/w · e n (1−j/w) − 2 + 4
+ o(n2 )
=
1
2
2 j=1
(e n + e n (1−j/w) − 2)w
w
4
1
n X 1
+ o(n2 )
+ ·
=
2 j=1 w w e n1 + e n2 (1−j/w) − 2
exp(1/n)+2
w
ln exp(2/n)
X
4
n
1
+ o(n2 ) =
n2 + o(n2 )
=
· 1
2
2 j=1 w e n + e n (1−j/w) − 2
2 − exp(1/n)
=
ln(3) 2
n + o(n2 ).
2
We can also find parameter settings leading to the best-possible runtime, up to lower order terms, using
only the 1BitFlip and 2BitFlip operators, from Theorem 7. In [15], the authors looked at the parameter
setting τ = cn for constant c, and hence could not prove the optimal runtime.
Corollary 15 The expected runtime of the Generalised
Random Gradient hyper-heuristic on LeadingOnes,
with τ that satisfies both τ = ω(n) and τ ≤ 21 − ǫ n ln(n), for ǫ = Θ(1), is at most
1 + ln(2) 2
n + o(n2 ).
4
n
o(1)
.
exp(2τ /n) ∩ n
0.5
M ( 1−j/w ). Since we have
Proof: We constrain w ∈ ω(log n) ∩ o
Consider first the terms M (1) and
τ = ω(n), it is easy to see that M (1) = 1. For
0.5
0.5
0.5
M ( 1−j/w
), it is important to note that for the first half of the search, M ( 1−j/w
) = 1−j/w
, and in the second
13
0.5
0.5
half of the search, M ( 1−j/w
) = O(w) = no(1) (excluding the case when j = w in which M ( 1−j/w
)=
n
o(1)
). Note that
τ
n
2τ /n·1/w
τ
n
=
1/w
= ω(1) and e
= o(n ) = 1 + o(1). Hence:
2τ
τ
0.5
w 2τ + e n M (1) + e n (1−(j−1)/w) M
2
X
n
1−j/w
n
+ o(n2 )
E(T ) ≤
τ
τ
2 j=1
(e n + e2 n (1−j/w) − 2)w
τ
2τ
0.5
w 2τ + e n + e2τ /n·1/w · e n (1−j/w) M
2
X
n
1−j/w
n
+ o(n2 )
=
τ
τ
2
(1−j/w)
n
n
2 j=1
(e + e
− 2)w
τ
2τ
0.5
(1−j/w)
w 2τ
M 1−j/w
n2 X n + e n + e n
+ o(n2 )
=
τ
τ
2 j=1
(e n + e2 n (1−j/w) − 2)w
2τ
τ
0.5
w
n + e n (1−j/w) M
2τ
e
2
−
2
X
1−j/w
n
1
n +2
+
=
τ
τ
τ
τ
2 j=1 w e n + e2 n (1−j/w) − 2
e n + e2 n (1−j/w) − 2
+ o(n2 )
τ
2τ
0.5
w
n + e n (1−j/w) M
2
e
X
1−j/w − 2
n
1
+ o(n2 ).
=
τ
τ
2 j=1 w
e n + e2 n (1−j/w) − 2
The exponential terms will dominate the summation term (since nτ = ω(1)). When j > w2 , we note that
w
exp(τ /n) > exp( 2τ
n (1 − j/w)) and this term will dominate the other. Vice versa for j < 2 , and at j = w/2,
the two values are equal. Hence, we can split the sum into two sections:
τ
2τ
0.5
w
n + e n (1−j/w) M
e
−
2
2
X
1−j/w
1
n
+ o(n2 )
E(T ) ≤
τ
τ
2 j=1 w
e n + e2 n (1−j/w) − 2
!
2τ
τ
w/2−1
0.5
(1−j/w)
w
2
2
X
· 1−j/w
n
n
n2 X 1 e n
1
e
n
+
=
+
τ
2τ
2
w
2w
2
w en
e n (1−j/w)
j=1
j=w/2+1
2
+ o(n )
w/2
2
X
n
1 n2
1 + ln(2)
=
+
+ o(n2 ) =
n2 + o(n2 ).
4 j=1 w − j
4
4
Note that at j = w, the respective ‘dominating terms’ are exp(τ /n) and exp(2τ /n(1 − w/w)) = 1, and
0.5
) = nτ < 12 ln(n). Hence, exp(τ /n) will dominate.
M ( 1−j/w
4.2.2
Generalised Greedy
The following results regarding the runtime of the Generalised Greedy hyper-heuristic can be proved using
similar techniques and calculations to those used in the analysis of the Generalised Random Gradient hyperheuristic but applying the Variable Drift Theorem [13] instead of the Additive Drift Theorem.
To simplify the presentation of the proofs, we will use 2 as a bound on the expected increase in fitness
value following a successful mutation throughout the process. As the bits following the first zero-bit in any
individual remain uniformly distributed throughout the process, this expectation would be correct if the bit
string was of infinite length. We note that the law of total expectation can be used to prove that this overestimation will only affect lower-order terms in the derived runtimes, e.g. by using that the free-riders are
overwhelmingly unlikely to span to the end of the bit string while the parent individual has fitness at most
n − log2 n, and that a trivial upper bound for Generalised Greedy on the last k ≤ n bits is k(2n + τ ) = o(n2 ).
We begin by stating a lemma on the expected time and progress for a combined Decision and Exploitation
Stage.
14
Lemma 16 Let i = LeadingOnes(x(t) ) < n − 1, and Y (t) = n − i be the distance of the Generalised Greedy
hyper-heuristic’s current solution at the start of the t’th Decision Stage to the LeadingOnes optimum, and
let τ = o(n2 ). Then,
E(Y (t) − Y (t+1) | Yt ≥ 2) =
2 · τ · (5n2 − (8i + 8)n + 4(1 + i)2 )
+ 2 − o(1),
n2 (3n − 2i − 2)
and in expectation, the next Decision Stage begins after
τ+
2n2
3n − 2i − (4 −
2(i+1)
n )
fitness evaluations.
Proof: We look at the expected number of fitness evaluations for the Decision and Exploitation Stage,
given the starting position LO(x) = i.
During the Decision Stage, the hyper-heuristic will run both operators until one of them is successful
(i.e., makes an improvement in the fitness). The expected number of fitness evaluations for this to happen
is given by:
−1
1
2n − 2i − 2
E(TDSi ) = 2 · 1 − 1 −
1−
n
n2
2n2
2n3
=
.
=
3n2 − 2in + 2i − 4n + 2
3n − 2i − (4 − 2(i+1) )
n
The probability of the Decision Stage choosing each operator, given by the ‘First Hit Probability’ is
1/n
(2n−2i−2)/n2
n
2n−2i−2
1/n+(2n−2i−2)/n2 = 3n−2i−2 for the 1BitFlip operator and 1/n+(2n−2i−2)/n2 = 3n−2i−2 for the 2BitFlip
operator. It is possible for both operators to succeed in one iteration; for our analysis, we note that this
= O(1) times throughout the optimisation process, and opt to ignore
happens an expected O(n2 )· n1 · 2n−2i−2
n2
any progress made by such Decision and Exploitation Stages, accounting for this by adding a lower-order
O(τ + n) term to the expected runtime.
In the Exploitation Stage, the operator which succeeded during the Decision Stage runs for τ steps.
Throughout the two stages, the function value will increase. We can upper-bound the total increase by
keeping i constant throughout (i.e., using the value from the Decision Stage), and lower-bound it by applying
a high-probability bound on the number of observed LeadingOnes improvements. The highest success
2
2
probability occurs when using the 2BitFlip operator at i = 0; i.e 2n−2
n2 < n . Since τ = o(n ), we can use
Chernoff bounds to bound the number of successes in τ steps as o(n) with probability 1 − 2−Ω(n) (i.e., the
probability of Ω(n) successes in τ steps is at most 2−Ω(n) ). Hence, we can bound the progress from below
by considering i + o(n). The probability of each operator being chosen will remain the same, since this is
based on the original value of i at the Decision Stage.
Hence, the expected progress of the Exploitation Stage in τ iterations is at most
1
2n − 2i − 2 2n − 2i − 2
n
· +
·
2·τ ·
3n − 2i − 2 n 3n − 2i − 2
n2
2 · τ · (5n2 − (8i + 8)n + 4(1 + i)2 )
=
n2 (3n − 2i − 2)
and at least
1
2n − 2(i + o(n)) − 2
2n − 2i − 2
n
· +
·
2·τ ·
3n − 2i − 2
n
3n − 2i − 2
n2
2
2
2 · τ · (5n − (8i + 8)n + 4(1 + i) )
− o(1).
=
n2 (3n − 2i − 2)
using 2 to bound the expected fitness change of an accepted mutation.
15
Summing the expected progress and expected number of fitness evaluations performed in both stages
yields the Lemma statement.
We further note that for LO(x) = n − 1, the decision step takes an expected 2n iterations and will
conclude the optimisation process.
As the number of iterations spent in the decision step is not deterministic and not identically distributed
over the optimisation process, we cannot directly apply the Variable Drift Theorem [13] to bound the expected
runtime for the Generalised Greedy hyper-heuristic. However, we can upper bound the runtime through an
application of Wald’s Equation (Theorem 10). By splitting the analysis into w chunks, we can upper bound
the expected times for a Decision and Exploitation Stage within each chunk. We will then use the Variable
Drift Theorem to bound the expected number of Decision and Exploitation Stages required to optimise each
segment, and combine the two results to upper bound the total expected runtime.
The significance of the following Theorem will be highlighted in Corollary 19 and depicted in Figure 1.
Theorem 17 The expected runtime of the Generalised Greedy hyper-heuristic on LeadingOnes with τ =
o(n2 ) is at most
Z kn
w
X
w
2w
1
τ
di + o(n2 )
n
+
2
2
(k−1)n 2τ ·(5n −(8i+8)n+4(1+i) )
3w − 2k n
+2
2
w
k=1
n (3n−2i−2)
2
for any w = o min n, nτ
.
Proof: Let LeadingOnes(x) = i denote the current state. We divide the analysis into w chunks of length
2
n/w (for w = o( nτ )). We can bound the expected time for a Decision and Exploitation Stage (using the
result from Lemma 16) in the kth chunk, where (k − 1)n/w ≤ i ≤ kn/w − 1, by
τ
2n2
2w
+τ ≤
· n.
+
3w − 2k n
3n − 2i − 4 − 2(i+1)
n
We can make this claim since we know w = o(n).
Within each chunk, we will use the Variable Drift Theorem to bound the number of expected
Decision
2w
and Exploitation Stages, by looking at the drift that happens in a ‘D+E’ Stage, i.e., at most 3w−2k + nτ · n
steps. We note that there may be some minor overlap between chunks, although at most, this will result in
the addition of τ · w = o(n2 ) iterations throughout the search.
We have seen that,
E(Yt − Yt+1 | Yt ) =
2 · τ · (5n2 − (8i + 8)n + 4(1 + i)2 )
+ 2 − o(1) = h(n − i) = h(x).
n2 (3n − 2i − 2)
We consider Xt = 0 to be the end of the chunk. Let xmin = 1 represent the point kn
w − 1, and we note
xmin
n
=
O(
).
that h(x
)
τ
min
Hence, by the Variable Drift Theorem, the expected runtime of the kth chunk is
τ
2w
n·
+
E(Tk ) ≤
3w − 2k n
Z kn
w −1
xmin
1
+ o(1) di +
2τ ·(5n2 −(8i+8)n+4(1+i)2 )
(k−1)n
h(x
min )
+
2
w
n2 (3n−2i−2)
Z kn
w
τ
1
2w
n
di + O(max{τ, n}).
+
≤
2
2
(k−1)n 2τ ·(5n −(8i+8)n+4(1+i) )
3w − 2k n
+2
w
n2 (3n−2i−2)
16
The total expected runtime, by Wald’s equation will be at most equal to the sum of
expected
the
runtimes
n2
of each chunk, plus some lower-order terms amounting to at most O(max{τ, n}) · o min n, τ
= o(n2 ).
Thus,
!
w
X
E(T ) ≤
E(Tk ) + o(n2 ).
k=1
The upper bound delivered by the theorem holds for any subquadratic value of τ . This will indicate that
the Generalised Greedy hyper-heuristic is not very sensitive to the parameter τ for LeadingOnes (as seen
in Corollary 19). As a result, the parameter should be easy to set in practice as confirmed by experiments
in Section 7.
We can now use this result to tightly upper bound the runtime of the simple Greedy mechanism (when
τ = 0, the two mechanisms act equivalently). We know this bound is tight by Corollary 9.
Corollary 18 For τ = 0, the expected runtime of the Generalised Greedy hyper-heuristic is at most
ln(3) 2
n + o(n2 ).
2
Proof: We constrain w ∈ ω(1) ∩ o(n). Inserting the relevant parameters into the equation will give:
Z kn
w
X
w
2w
1
τ
di + o(n2 )
E(T ) ≤
n
+
2
2
(k−1)n 2τ ·(5n −(8i+8)n+4(1+i) )
3w − 2k n
+2
w
k=1
n2 (3n−2i−2)
!
kn
Z
w
X
w
1
2w
n
=
di + o(n2 )
(k−1)n 2
3w − 2k
w
k=1
w
X
ln(3) 2
1
2
+ o(n2 ) =
n + o(n2 )
=n ·
3w − 2k
2
k=1
for w = ω(1).
We can also find parameter settings leading to the best runtime that can be achieved for the Generalised
Greedy hyper-heuristic.
Corollary 19 For τ ∈ ω(n) ∩ o(n2 ) the expected runtime of the Generalised Greedy hyper-heuristic is at
most
ln(5) arctan(2)
n2 + o(n2 ).
+
8
4
Proof: We constrain w = o(n). For τ ∈ ω(n) ∩ o(n2 ), we have:
τ
2w
τ
τ
+ = +o
.
3w − 2k n
n
n
τ
2τ · (5n2 − (8i + 8)n + 4(1 + i)2 )
2τ (5n2 − 8in + 4i2 )
.
+
2
=
+
o
n2 (3n − 2i − 2)
n2 (3n − 2i − 2)
n
17
Hence, we can simplify the expected runtime of the hyper-heuristic:
Z kn
w
X
w
2w
1
τ
E(T ) ≤
di + o(n2 )
+
n
2
2
(k−1)n 2τ ·(5n −(8i+8)n+4(1+i) )
3w − 2k n
+
2
2
w
k=1
n (3n−2i−2)
!
Z kn
w
2
X
w
n (3n − 2i − 2)
=
τ
di + o(n2 )
(k−1)n 2τ (5n2 − 8in + 4i2 )
w
k=1
!
Z kn
w
2 X
w
n
3n − 2i − 2
n2 1
1
2
=
di
+
o(n
)
=
ln(5)
+
arctan(2)
+ o(n2 )
(k−1)n 5n2 − 8in + 4i2
2
2
4
2
w
k=1
ln(5) arctan(2)
=
n2 + o(n2 ).
+
8
4
We now analyse an ‘ideal’ algorithm which will provide a theoretical lower bound for the Generalised Greedy
hyper-heuristic. This hyper-heuristic is sped-up by getting the result of the Decision Stages for ‘free’ in every
iteration, and uses this result to decide which operator to apply.
Theorem 20 The ideal greedy algorithm which performs the Decision Stage ‘for free’ in every iteration has
runtime:
ln(5) arctan(2)
n2 + o(n2 ).
+
E(T ) =
8
4
Proof: The probability of the Decision Stage choosing each operator is
2
(2n−2i−2)/n
1/n+(2n−2i−2)/n2
1/n
1/n+(2n−2i−2)/n2
=
n
3n−2i−2
for
2n−2i−2
3n−2i−2
for the 2BitFlip operator. It is possible for both
=
the 1BitFlip operator and
operators to succeed in one iteration; for our analysis, we note that this happens an expected O(n2 ) · n1 ·
2n−2i−2
= O(1) times throughout the optimisation and we constrain this to an o(n2 ) term. Using Theorem
n2
6, we have:
−1
n−1
n
1 X
1
2n − 2i − 2 2n − 2i − 2
E(T ) = ·
· +
·
2 i=0 3n − 2i − 2 n 3n − 2i − 2
n2
n−1
ln(5) arctan(2)
1 X (3n − 2i − 2)n2
n2 + o(n2 ).
=
+
= ·
2 i=0 (2n − 2i − 2)2 + n2
8
4
Hence, the upper bound provided in Corollary 19 is tight.
5
A New Hyper-heuristic: Generalised Greedy Gradient
We now consider a new mechanism, called Generalised Greedy Gradient (GGG), which combines the Decision
Stage idea of the Generalised Greedy hyper-heuristic i.e., where all operators are run in parallel and the first
to make a fitness improvement is chosen, and applies this operator choice to the gradient-based Exploitation
Stage of the Generalised Random Gradient hyper-heuristic i.e., the chosen operator is run for a period of τ
steps, and each fitness improvement resets the period of τ . The idea is to combine the learning aspects of
the two hyper-heuristics to create a hyper-heuristic that chooses the best operator and exploits this choice.
In this section we will prove that the Generalised Greedy Gradient hyper-heuristic can solve the GapPath
and GGPT efficiently in a similar way to GRG and GG. We also show that the GGG hyper-heuristic can
achieve the same optimal performance on LeadingOnes as GRG by using shorter learning periods τ .
5.1
GapPath and GGPT
We first introduce runtime results for the GGG hyper-heuristic on the GapPath and GGPT functions
considered in Section 3, which show that the hyper-heuristic can learn to use a given operator in a similar
18
way to the other generalised hyper-heuristics.
Theorem 21 Starting at 0n , the expected runtime of the Generalised Greedy Gradient hyper-heuristic on
GapPath is
n3
2n2
2τ n
+
+
− τ.
3
3
3
Proof: The GGG hyper-heuristic will act equivalently to GG on GapPath; that is, it will run the Greedy
operator Decision Stage and deterministically choose the unique operator which will give a success. Then
the mechanism will waste τ iterations, since it cannot enact any improvements in this Exploitation Stage
with the selected operator. Hence, the runtime is equivalent to the result from Theorem 1.
Theorem 22 Starting at 0n for τ = Ω(n3 ), the Generalised Greedy Gradient hyper-heuristic will find the
3
2τ n
2
+ (k+1)n
global optimum of GGPT (with constant k) in at most 2k+1
2(2k+1) + O(n ) steps with probability at least
1 − 2−Ω(n) . However, for τ = O(n2 ), the Generalised Random Gradient hyper-heuristic will fail to reach the
global optimum of GGPT in finite time with probability at least 1 − 2−Ω(n) .
Proof: Initially, a 1BitFlip improvement is necessary before the first 2BitFlip flip. For τ = Ω(n3 ),
similarly to Theorem 5, this will take time 2n + τ to account for the wasted iterations. The GRG hyper2
heuristic will then run the Greedy Decision Stage to choose the 2BitFlip operator in time 2 · n2 = n2 . The
Exploitation Stage will continue with this operator to find the remaining necessary 2BitFlip improvements
and avoid the ‘trap’ point with probability 1 − 2−Ω(n) , as per Theorem 4. The time for the remaining k − 1
2
improvements will be (k−1)n
+ τ to account for the wasted iterations at the end.
2
n
times to account for each run of 2BitFlip flips necessary to find
This process must be repeated 2k+1
the global optimum. Hence, the expected running time will be
k−1 2
n
2
2n + τ + n +
n +τ
2k + 1
2
which gives the result.
However, if τ = O(n2 ), a similar argument to Theorem 4 can be made. That is, for τ = cn2 , where
cn2 k−1
=
c is a constant, the probability of the required k − 1 successful periods is 1 − 1 − 2/n2
(1 − e−2c − o(1))k−1 = Θ(1), and so the probability of failing to do so is at least constant. If another
random Decision Stage is performed inside the region of two-bit improvements, a ‘trap’ is accessible via a
1-bit mutation from the point at which the decision occurs with probability at least 1/(k − 1) = Ω(1) (as
traps are placed uniformly at random inside these regions), and is found by the Random Gradient mech2
anism with probability at least 1/2 · (1 − 1/n)cn ≥ 1/4. Thus, the probability that the no ‘trap’ point is
constructed in a single region of k two-bit improvements is at most a constant smaller than one, which means
that over the n/(2k+1) such regions, the probability that a ‘trap’ point is constructed is at least 1−2−Ω(n) .
We have seen that the Generalised Greedy Gradient hyper-heuristic performs similarly to the two previous
generalised hyper-heuristics on the GapPath series of functions. By learning to favour 2-bit flips it performs
efficiently.
5.2
LeadingOnes
We now prove that the Generalised Greedy Gradient mechanism can lead to improved performance on the
LeadingOnes benchmark function compared to the Generalised Random Gradient hyper-heuristic. The
result will be highlighted in Corollary 25 and a comparative performance with the other hyper-heuristics will
be depicted in Figure 1.
Theorem
23 The runtime of the Generalised Greedy Gradient hyper-heuristic on LeadingOnes, with τ ≤
1
2 − ǫ n ln(n), ǫ = Θ(1), is at most
X
w
3w − 2j + 2
n2
·
E(T ) ≤ o(n2 ) +
2
3w − 2j
j=1
19
τ
2τ
0.5
· (3w − 2j) + w · e n · M (1) + (2(w − j + 1)) · e n (1−(j−1)/w) · M ( 1−j/w
)
2τ
τ
w w · e n + 2(w − j) · e n (1−j/w)
n
where M (x) := min( nτ , x), for w = o exp(2τ
/n) .
2w +
τ
n
Proof: We partition the optimisation process into w chunks based on the value of the LeadingOnes
fitness function: during chunk j, the LO value of the current solution is at least (j − 1)n/w and less than
j · n/w.
We follow the proof for the Generalised Random Gradient on LeadingOnes, in which we upper bounded
the final runtime by the sum of the expected values of Tj , the runtime of each chunk of the process, plus the
times taken
P to begin a chunk. Since the analysis of E(Tj ) requires an operator choice, we bound the runtime
E(T ) ≤ w
j=1 (E(Tj ) + E(Sj+1 )), where Sj+1 is a random variable denoting the expected number of fitness
evaluations between the first solution in chunk j + 1 being constructed and the first operator
decision of
1
chunk j beginning.
We
note
again
(with
the
same
parameter
combinations
i.e.,
τ
≤
−
ǫ
n
ln(n),
ǫ = Θ(1),
2
n
2
w = o exp(2τ
/n) ), the contribution of the Sj+1 terms can be bounded by o(n ).
We let Nj denote the number of operator decisions within chunk j, and Xj,1 , . . . , Xj,NJ the number of
fitness function evaluations within the Decision Stage and the Exploitation Stage of the chosen operator.
PNj
Xj,k , and by Wald’s equation, using E(Xj ) to denote an upper bound on all E(Xj,k ) in
Then Tj = k=1
chunk j,
E(Tj ) ≤ E(Nj )E(Xj ).
We can upper-bound E(Nj ), the number of Decision Stages in chunk j, by lower-bounding the expected
number of improvements found from the Decision and Exploitation Stage of the chosen operator, and using
the Additive Drift Theorem to find the expected number of operator selections required to enter the next
chunk.
Recall from the proof of Theorem 13 that, denoting F1 and F2 as the events that the 1BitFlip and
2BitFlip operators fail to find an improvement during τ iterations:
τ
e− n −
τ
1
≤ P (F1 ) ≤ e− n ;
n
2τ
e− n (1−(j−1)/w) −
2τ
1
≤ P (F2 ) ≤ e− n (1−j/w) .
n
After each Decision Stage, an operator will be chosen. We use the First Hit Probability to look at the
probability of each operator being chosen:
P (C1 ) =
Note that in chunk j, with
1
n
+
(j−1)n
w
1
n
2(n−i−1)
n2
=
−1≤i≤
n
;
3n − 2i − 2
jn
w
P (C2 ) =
2n − 2i − 2
.
3n − 2i − 2
− 1, we can bound the two probabilities:
w
w
≤ P (C1 ) ≤
;
3w − 2j + 2
3w − 2j
2w − 2j
2w − 2j + 2
≤ P (C2 ) ≤
.
3w − 2j
3w − 2j + 2
We look at finding a lower bound for E(Dj ), the expected number of improvements Dj produced following
an operator Decision and Exploitation Stage. Since we are looking for a lower bound, we use the lower bound
for the probability of choosing each operator, and an upper bound for the number of improvements given
the chosen operator. Note there is 1 improvement from the Decision Stage.
1
1
− 1 + P (C2 ) ·
−1
E(Dj ) ≥ 1 + P (C1 ) ·
P (F1 )
P (F2 )
w
2w − 2j
1
1
≥1+
−
1
+
−
1
·
3w − 2j + 2
exp(− nτ )
3w − 2j exp(− 2τ
n (1 − j/w))
τ
w
2(w − j) 2τ (1−j/w)
· en +
·en
≥
3w − 2j + 2
3w − 2j
τ
2τ
2(w − j)
w
· en +
· e n (1−j/w) .
≥
3w − 2j + 2
3w − 2j + 2
20
In each chunk, in expectation, we require n/(2w) fitness improvements to exit the chunk. If each random
operator decision leads to more E(Dj ) improvements, by the Additive Drift Theorem, the expected number
of steps required to complete chunk j, E(Nj ) satisfies:
E(Nj ) ≤
≤
n/(2w)
n
≤
·
E(Dj )
2w
1
w
3w−2j+2
n(3w − 2j + 2)
·
2w
τ
n
·e +
2(w−j)
3w−2j+2
2τ
· e n (1−j/w)
.
2τ
(1−j/w)
1
τ
n
w · e + 2(w − j) · e n
We now look at bounding E(Xj ), an upper bound on the number of fitness function evaluations for a
Decision and Exploitation Stage within chunk j.
The worst case expected time for the duration of the Decision Stage in chunk j is
−1
2(w − j)
2n2 w
1
1−
=
E(TDSj ) ≤ 2 · 1 − 1 −
n
nw
(3w − 2j)n − 2(w − j)
2nw
=
+ o(n).
3w − 2j
The small order term will be omitted for brevity, but note that it makes no difference to the leading constant
on the asymptotically significant term.
We can also upper-bound E(Xj |C{1,2} ), the expected iterations before a selected operator fails to produce
an improvement within τ iterations. Using, for each operator,
E(Xj |C1 ) ≤ τ + (1/P (F1 ) − 1) · min(τ, n)
E(Xj |C2 ) ≤ τ + (1/P (F2 ) − 1) · min(τ, 0.5n/(1 − j/w)).
Hence:
E(Xj ) ≤ E(TDSj ) + P (C1 ) · E(Xj |C1 ) + P (C2 ) · E(Xj |C2 )
2w − 2j + 2 min(τ, n/(2 − 2j/w))
w
min(τ, n)
2nw
+
+τ +
·
·
τ
3w − 2j
3w − 2j e− n − 1/n 3w − 2j + 2 e− 2τ
n (1−(j−1)/w) − 1/n
2w − 2j + 2 min(τ, n/(2 − 2j/w))
w
min(τ, n)
2nw
.
+
+τ +
· −τ
· − 2τ (1−(j−1)/w)
≤
n
3w − 2j
3w − 2j e
3w − 2j
− 1/n
− 1/n
e n
≤
which can be simplified to
E(Xj ) ≤ o(n) +
n
·
3w − 2j
τ
τ
τ
τ
2τ
0.5
(1− j−1
)
n
n
w
2w + (3w − 2j) + we min
min
.
, 1 + (2w − 2j + 2)e
,
n
n
n 1 − j/w
Finally, we bound the overall expected runtime of the hyper-heuristic. We have:
w
w
X
X
(E(Tj ) + E(Sj+1 )) ≤
E(Nj )E(Xj ) + o(n2 )
E(T ) ≤
j=1
and hence
j=1
X
w
3w − 2j + 2
n2
·
E(T ) ≤ o(n ) +
2
3w − 2j
j=1
2
2τ
τ
2w + nτ (3w − 2j) + we n min( nτ , 1) + (2(w − j + 1))e n (1−
τ
2τ
w w · e n + 2(w − j) · e n (1−j/w)
j−1
w )
0.5
)
min( nτ , 1−j/w
.
When we set τ = 0, the hyper-heuristic reduces to the simple Greedy mechanism, and the following
corollary provides a tight runtime upper bound. We know it is tight from Corollary 9.
21
Corollary 24 When τ = 0 (i.e., ǫ =
heuristic is at most
1
2 ),
the expected runtime of the Generalised Greedy Gradient hyperln(3) 2
n + o(n2 ).
2
Proof: We constrain w = ω(1) ∩ o(n2 ). When τ = 0, we have
w
w
n2 X 2(3w − 2j + 2)
2w
n2 X 3w − 2j + 2
=
·
·
·
E(T ) ≤
2 j=1 3w − 2j
w (w + 2(w − j))
2 j=1 (3w − 2j)2
=
ln(3) 2
n + o(n2 ).
2
The following corollary shows that we can find parameter values such that the GGG hyper-heuristic can
match the optimal performance of the GRG hyper-heuristic (Corollary 15) with the best possible runtime
from Theorem 7.
Corollary 25 For τ that satisfies both τ = ω(n) and τ ≤ ( 21 − ǫ)n ln(n), for ǫ = Θ(1), the expected runtime
of the Generalised Greedy Gradient hyper-heuristic is at most
1 + ln(2) 2
n + o(n2 ).
4
n
exp(2τ /n) .
0.5
Consider first the terms M (1) and M ( 1−j/w
). Since we have τ = ω(n), we have nτ = ω(1); it is
0.5
), it is important to note that for the first half of the search,
easy to see that M (1) = 1. For M ( 1−j/w
0.5
0.5
0.5
M ( 1−j/w ) = 1−j/w , and in the second half of the search, M ( 1−j/w
) = O( nτ ) = O(ln n) for any constant
0.5
τ
c > 0 (there is a special case when j = w in which M ( 1−j/w ) = n = O(ln n)). Note also that nτ = ω(1),
e2τ /n·1/w = o(n1/w ) = 1 + o(1), 3w−2j+2
3w−2j = 1 + o(1).
Proof: We constrain w ∈ ω(ln n) ∩ o
Hence, the runtime will become:
E(T ) ≤ o(n2 ) +
2w +
τ
n
w
n2 X
(1 + o(1))·
·
2 j=1
2τ
τ
0.5
· (3w − 2j) + w · e n + (2(w − j + 1)) · e2τ /n·1/w e n (1−j/w) · M ( 1−j/w
)
τ
2τ
w w · e n + 2(w − j) · e n (1−j/w)
n2
·
= o(n2 ) +
2
w
X
2w +
τ
2τ
0.5
· (3w − 2j) + w · e n + (2(w − j + 1)) · e n (1−j/w) · M ( 1−j/w
)
2τ
τ
n + 2(w − j) · e n (1−j/w)
w
w
·
e
j=1
w
n2 X 1
= o(n2 ) +
·
·
2 j=1 w
!
2τ
τ
0.5
)
w · e n + 2(w − j + 1) · e n (1−j/w) · M ( 1−j/w
2w + nτ · (3w − 2j)
+
τ
τ
2τ
2τ
w · e n + 2(w − j) · e n (1−j/w)
w · e n + 2(w − j) · e n (1−j/w)
!
2τ
τ
0.5
w
w · e n + 2(w − j + 1) · e n (1−j/w) · M ( 1−j/w
)
n2 X 1
2
·
·
.
= o(n ) +
2τ
τ
2 j=1 w
w · e n + 2(w − j) · e n (1−j/w)
τ
n
22
Runtime (/n2 )
0.60
GG
GRG
GGG
RLS
0.50
2Opt
0.40
0
5
10
15
20
25
30
35
40
45
50
τ (/n)
Figure 1: The leading constants in the theoretical upper bounds on the average number of fitness function
evaluations required by the Generalised Greedy, Generalised Random Gradient, and Generalised Greedy
Gradient hyper-heuristics to find the LeadingOnes optimum (w = 100, 000).
Note that the exponential terms will dominate the summation term (since nτ = ω(1)). When j > w2 , we note
that w · exp(τ /n) = ω(2(w − j + 1) · exp( 2τ
n (1 − j/w))) and this term will dominate the other. Vice versa for
j < w2 . We consider j = w/2 as a special case.
!
τ
2τ
0.5
w
2 X
w · e n + 2(w − j + 1) · e n (1−j/w) · M ( 1−j/w
)
1
n
·
·
E(T ) ≤ o(n2 ) +
τ
2τ
2 j=1 w
w · e n + 2(w − j) · e n (1−j/w)
2τ
w/2−1
0.5
n (1−j/w) ·
2
X
e
1 2(w − j + 1)
n
1−j/w
·
·
·
= o(n2 ) +
2τ
2
w
2(w
−
j)
n (1−j/w)
e
j=1
2τ
τ
w
2
2
X
n
n
n
(2w + 2)e
1 e
n
+
·
·
· τ
+
2τ
2w
2
w en
2w · e n
j=w/2+1
w/2−1
w
2
2
2
X 1
X
n
0.5 n
n
1
= o(n2 ) +
+
·
·
+
·
2
w
1
−
j/w
2w
2
w
j=1
j=w/2+1
=
since
2(w−( w
2 )+1)
2(w−( w
2 ))
=
w/2
X
n2
·
4 j=1
2w+2
2w
1
w−j
+
n2
+ o(n2 ) =
4
= 1 + o(1).
1 + ln(2)
4
n2 + o(n2 ).
Figure 1 shows the relevant theoretical upper bounds for the Generalised Greedy, Generalised Random
Gradient and Generalised Greedy gradient hyper-heuristics from Theorem 13, Theorem 17, and Theorem
23. The results for the Generalised Greedy Gradient are better than the results for the Generalised Random
Gradient at each value of τ , although they are essentially indistinguishable within the context of Figure 1.
The two hyper-heuristics seem to quickly tend towards the optimal results of Corollary 15 and Corollary 25,
and this would be seen by considering τ values that satisfy both τ = ω(n) and τ ≤ ( 12 −ǫ)n ln(n), for ǫ = Θ(1).
23
6
Increasing the Choice of Low-level Heuristics
Runtime analyses of hyper-heuristics are easier if the algorithms can only choose between two operators.
However, in realistic contexts, hyper-heuristics have to choose between many more operators.
In this section, we consider the previously analysed generalised hyper-heuristics, yet extend the set H of
low-level heuristics to be of size k = o(n) i.e., |H| = k. We extend upon the previous analysis by considering
the k operators as k different mutation operators, each flipping between 1 and k indistinct bits uniformly
at random respectively. This would more accurately represent hyper-heuristic approaches employed for
real-world problems, as opposed to setting k = 2 for the purposes of the analysis.
As before, we consider the LeadingOnes benchmark function. We will rigorously show that the performance of the ‘simple’ mechanisms deteriorates as the number of operators (k) increases. On the other hand,
we prove decreasing upper bounds on the runtime of the Generalised Random Gradient hyper-heuristic as the
number of operators increases. Since the Decision Stage of the Generalised Greedy hyper-heuristic (and the
Generalised Greedy Gradient) hyper-heuristic makes the analysis very complex, we present an experimental
analysis of these hyper-heuristics in Section 7.
We first derive the best possible runtime achievable by any hyper-heuristic using k = o(n) operators.
Before we prove this result we introduce the following two helpful lemmata.
Lemma 26 The probability of improvement from an operator which flips m indistinct bits in a bit-string on
the LeadingOnes benchmark function, at the state LO(x) = i, is
m−1
n−i−1
1
+ O(n−2 ).
P (impm ) = m · ·
n
n
Proof: In the context of the LeadingOnes benchmark function, the way to make an improvement is to
flip the first 0-bit after the leading 1-bits, whilst keeping the prefix leading 1-bits unflipped. Although it is
possible to make improvements by flipping and re-flipping the same bit in some combination (with probability
Θ( n12 )), the operators are likely to make a fitness improvement by flipping the first 0-bit (with probability
1/n, which can occur within any of the m bit-flips), and using the remaining m − 1 flips within the suffix
)m−1 ). Hence we can say that the probability of an improvement when
n − i − 1 bits (with probability ( n−i−1
n
flipping k bits is
m−1
1
n−i−1
P (impm ) = m · ·
+ O(n−2 ).
n
n
We can leave the formula in this state since all analyses will consider the use of the 1BitFlip operator.
Since this always has an improvement probability of 1/n on LeadingOnes, any lower-order terms in other
operators will not affect the leading constants if it is chosen sufficiently often. We also note that through a
simple calculation, the 1BitFlip operator is always the best choice in the second half of the search space
n
2 ≤ i < n.
Lemma 27 Consider two indistinct bitflip operators flipping a and b bits respectively, with b > a. Then
operator a outperforms b (excluding i = n − 1) when
1/(b−a) !
b
− 1.
i<n 1−
a
This follows simply from Lemma 26. In particular, an operator flipping m bits outperforms an operator
n
flipping m − 1 bits when i < m
− 1.
Theorem 28 The best-possible runtime on LeadingOnes for a mechanism using k = o(n) operators,
(1, . . . , k) bitflips, is
n/k−1
k−1
X
X n/x−1
1
1
1 X
± o(n2 ).
+
E(Tk,opt ) =
1
n−i−1 x−1
1
n−i−1 k−1
2
x
·
·
k
·
·
x=1 i=n/(x+1)
i=0
n
n
n
n
24
n
Proof: From Lemma 27, we can say the mBitFlip operator is optimal during the time when m+1
≤i≤
n
n
−
1
(unless
m
=
k,
in
which
case
it
is
optimal
for
0
≤
i
≤
−
1).
Since
the
1BitFlip
operator
has
m
k
success probability O(n−1 ), and is chosen with the same probability as all other operators, the leading term
in any success probability is O((nk)−1 ) = ω(n−2 ) and any O(n−2 ) terms are insignificant. Thus, they can
be grouped into a lower order o(n2 ) term.
Hence, by standard waiting time arguments, coupled with Theorem 6, we can show that the optimal
runtime when using k operators is
n/x−1
n/k−1
k−1
X
X
X
1
1
1
± o(n2 ).
+
E(Tk,opt ) =
n−i−1 x−1
1
n−i−1 k−1
1
2
x· ·
·
x=1
i=0 k ·
n
i=n/(x+1)
n
n
n
1+ln(2) 2
1 2
2
n
,
E(T
)
=
n
≈
0.42329n
,
In particular,
taking
limits
as
n
→
∞,
we
have
E(T
)
=
2,opt
1,opt
2
4
ln(2)
ln(3)
ln(2)
ln(3)
3721
1
2
2
2
2
n ≈ 0.40525n , E(T5,opt ) = 11520 + 2 − 4
n ≈ 0.39492n .
E(T3,opt ) = 3 + 2 − 4
This result gives a theoretical lower bound on the expected runtime of any algorithm using the k operators
on LeadingOnes. We can see that including further operators can improve the runtime.
6.1
‘Simple’ Mechanisms
We will now see how the simple learning mechanisms (Simple Random, Permutation, Greedy and Random Gradient) perform when tasked with choosing between k operators (i.e., (1, . . . , k) bitflips). We will
show that incorporating more operators is detrimental to the performance of the ‘simple’ mechanisms on
LeadingOnes.
We again start by stating the expected runtime of the Simple Random mechanism, and use this as a
basis for the other three mechanisms.
Recall that the standard Simple Random mechanism chooses each operator uniformly at random in each
iteration (i.e., with probability 1/k when using k operators).
Theorem 29 The runtime of the Simple Random mechanism on LeadingOnes with k = o(n) operators,
(1, . . . , k) bitflips, is
n−1
1X
1
E(TSR|k ) =
− o(n2 ).
1
n−i−1 m−1
2 i=0 1 · Pk
m=1 m · n ·
k
n
Proof: We have seen previously that the probability of an improvement when flipping m indistinct bits is
m−1
n−i−1
1
+ O(n−2 ).
P (impm ) = m · ·
n
n
We can use this claim again since, with a sub-linear number of operators, so long as the 1BitFlip operator
is used, any u.a.r. choice of operators will give an improvement probability in each iteration of ω(n−2 ).
The probability of an improvement in a single iteration is
!
m−1
k
1 X
1
n−i−1
−1
−2
(An−i ) = ·
m· ·
+ O(n )
k m=1
n
n
!
m−1
k
1
1
1 X
1
n−i−1
−2
x· ·
+ O(n ) = Ω
=
+ ·
kn k m=2
n
n
n
and hence
An−i =
1
1
k
·
We can use Theorem 6 to give the result:
Pk
m=1 m ·
n−1
n−1
1X
1X
E(T ) =
An−i =
2 i=0
2 i=0
1
n
·
n−i−1 m−1
n
− o(n).
1
1
k
·
Pk
m=1
25
m·
1
n
·
n−i−1 m−1
n
− o(n2 ).
Corollary 30 The expected runtime of the Permutation, Greedy and Random Gradient mechanisms on
LeadingOnes with k = o(n) operators, (1, . . . , k) bitflips, is
E(Tk ) =
n−1
1X
2 i=0
1
1
k
·
Pk
m=1
m·
1
n
·
Proof: For the Greedy and Permutation mechanisms, let
n−i−1 m−1
n
pi = P (imp1 ) + (1 − P (imp1 )) · P (imp2 ) + · · · +
=
k
X
m=1
=
k
X
m=1
m−1
Y
1
·
n
n−i−1
n
j−1 !
1
(1 − o(1)) · m · ·
n
n−i−1
n
m−1
j=1
1−j·
k−1
Y
j=1
± o(n2 ).
(1 − P (impj )) · pk
·m· 1 ·
n
!
n−i−1
n
m−1
be the probability that at least one improvement is constructed in respectively 1, . . . , k iterations. To move
from the second equation to the third, we note that Pimpj = O( n1 ) and
Qm−1
1 o(n)
= 1 − o(1). To upper bound the optimisation time, we note that
j=1 (1 − Pimpj ) = (1 − O( n ))
the difference between k/pi (i.e., the waiting time for an improvement to be constructed in terms of fitness
evaluations) and the An−i waiting times used to prove Theorem 29 is at most constant, and thus the difference
between these mechanisms and Simple Random expected times is limited to lower-order terms. We note that
with probability at most k(k − 1)/n2 , at least two mutations (of the k) considered are improvements; in such
a case, we can consider all but one of the improvement steps to be made in 0 iterations. As this occurs at
most a sublinear number of times in expectation, and the maximum waiting time for any improving step is
O(n), the lower-bound differs from the upper bound by at most an o(n2 ) term. Thus, the expected runtime
Pn−1
1
± o(n2 ).
of these mechanisms is also 21 i=0 1 Pk
n−i−1 m−1
1
)
m=1 m· n ·(
k·
n
For the Random Gradient mechanism, we note that the probability that an operator, when repeated
following a success, is successful again is at most k/n. Since there are at most n improvements to be made
throughout the search space, the expected number of repeats which produce an improvement is at most
k = o(n). When not repeating a successful operator, the Random Gradient mechanism behaves identically
to the Simple Random mechanism. Its expected runtime is therefore at least the runtime of the Simple
Random mechanism less an o(n2 ) term, and at most the runtime of the Simple Random mechanism plus n
(as there are at most n iterations where the mechanisms differ in operator selection). Thus, its expected
Pn−1
1
runtime is also 12 i=0 1 Pk
± o(n2 ).
n−i−1 m−1
1
)
m=1 m· n ·(
k·
n
Note that the expected runtimes from Theorem 29 and Corollary 30 increase as k increases, suggesting that
incorporating more operators is detrimental to the performance of the ‘simple’ mechanisms.
In particular,
√
√ the
2
3 2
2
2
results for using 1, 2 and 3 operators respectively are 12 n2 , ln(3)
n2 ≈
n
≈
0.54931n
and
arctan
2
4
2
0.65281n2.
6.2
Generalised Random Gradient
We present a rigorous theoretical analysis of the runtime of the Generalised Random Gradient hyper-heuristic
on LeadingOnes using k ≥ 2 operators. The derived upperbound on the expected runtime decreases with
k.
26
Theorem 31 The expected runtime of the Generalised Random Gradient hyper-heuristic on LeadingOnes
for τ ≤ k1 − ǫ n ln(n), ǫ = Θ(1) with k = o(n) operators, (1, . . . , k) bitflips, is at most:
i
hP
τ
k
1
τ
mn
(1−(j−1)/w)m−1
w
· M m(1−j/w)
m−1
m=1 e
n2 X k · n +
± o(n2 ).
i
hP
·
τ
m−1
k
2 j=1
(1−j/w)
m
n
−k
w
m=1 e
n
where M (x) := min( nτ , x), for any w ∈ Ω(1) ∩ o exp(kτ
/n) .
Proof: We follow the proof idea for the Generalised Random Gradient hyper-heuristic for two operators
(Theorem 13). That is, we partition the optimisation process into w chunk based on the value of the
LeadingOnes fitness function: during chunk j, the LO value of the current solution is at least (j − 1)n/w
and less than j · n/w. After all w chunks have been completed, the global optimum, with an LO value of n,
will have been found.
We upper-bound the expectation of the runtime T of the Generalised Random Gradient hyper-heuristic
on LO by the sum of the expected values of Tj , the runtimes on each chunk of the optimisation process,
plus the times taken to begin a chunk. As our analysis
of E(Tj ) requires the chunk to start with a random
P
choice of mutation operator, we bound E(T ) ≤ w
(E(T
j ) + E(Sj+1 )) where Sj+1 is a random variable
j=1
denoting the expected number of iterations between the first solution in chunk j +1 being constructed and the
following random operator choice. We will later show that with proper parameter choices, the contribution
of Sj+1 terms can be bounded by o(n2 ), and therefore does not affect the leading constant in the overall
bound.
Let us now consider Tj , the number of iterations the hyper-heuristic spends in chunk j. Recall that
a mutation operator is selected uniformly at random, and is allowed to run until it fails to produce an
improvement for τ sequential iterations. Let Nj be a random variable denoting the number of random
operator choices the hyper-heuristic performs during chunk j, and Xj,1 , . . . , Xj,Nj be the number of iterations
PNj
Xj,k , and by applying Wald’s
the hyper-heuristic runs each chosen operator for, in which case Tj = k=1
equation (Theorem 10), using E(Xj ) to denote an upper bound on all E(Xj,k ) in chunk j:
PNj
E(Xj,k ) ≤ E(Nj )E(Xj ).
(9)
E(Tj ) = k=1
To bound E(Nj ), the expected number of times the random operator selection is performed during chunk
j, we lower-bound the expected number of improvements found following operator selection, and apply the
Additive Drift Theorem [12] to find the expected number of random operator selections occurring before a
sufficient number of improvements have been found to enter the next chunk.
Recall from Lemma 26 that the improvement probability for a mBitFlip operator is
m−1
n−i−1
1
+ O(n−2 ).
P (impm ) = m · ·
n
n
We note again that since the 1BitFlip operator is used, we can omit the O(n−2 ) terms, and they will fall
into a o(n2 ) term in the final result.
Let Fm denote the event that the mBitFlip operator fails to find an improvement during τ iterations.
τ
This will occur with probability P (Fm ) = (1 − P (impm )) . Within chunk j, the ancestor individual has at
most jn/w − 1, and at least (j − 1)n/w , one bits. Hence:
m−1 !τ
m−1 !τ
j
1
1 n − (jn/w − 1) − 1
1−
≤ 1−m·
P (Fm ) ≤ 1 − m ·
n
n
n
w
≤ e−
P (Fm ) ≥
mτ
n
(1−j/w)m−1
1
1−m·
n
≥ e−
mτ
n
;
n − ((j − 1)n/w) − 1
n
(1−(j−1)/w)m−1
−
m−1 !τ
1
.
n
27
≥
1
1−m·
n
m−1 !τ
j−1
1−
w
We note that a geometric distribution with parameter p = P (Fm ) can be used to model the number of
improvements that the mBitFlip operator finds prior to failing to find an improvement for τ iterations; the
expectation of this distribution is (1 − p)/p = 1/p − 1. Combined over all k operators, the expected number
of improvements Dj produced following a single random operator selection during chunk j, E(Dj ), is greater
than:
!
k
k
X
1
1
1 X mτ (1−j/w)m−1
−1
E(Dj ) ≥
e n
−1 ≥
k
P
(F
)
k
m
m=1
m=1
by inserting the upper bounds for P (Fm ). We use this expectation as the drift in the Additive Drift Theorem
to upper-bound E(Nj ). Recall that each chunk consists of advancing through at most n/w fitness values;
as bits beyond the leading ones prefix and the first zero bit remain uniformly distributed, in expectation
n/(2w) improvements by mutation are required. If each step of a random process in expectation contributes
E(Dj ) improvements by mutation, the expected number of steps required to complete chunk j is at most:
E(Nj ) ≤
by the Additive Drift Theorem.
n/(2w)
k n
≤ · · P
k
E(Dj )
2 w
m=1 e
1
mτ
n
(1−j/w)m−1
(10)
−k
To bound E(Xj ), the expected number of iterations before a selected mutation operator fails to produce
an improvement for τ iterations, we apply Wald’s equation: let S be the number of improvements found
by the operator before it fails, and W1 , . . . , WS be the number of iterations it took to find each of those
improvements. Then, once selected, the mBitFlip operator runs for:
E(Xj |m) = τ +
S
X
E(Ws ) = E(S)E(Wm | S ≥ 1)
s=1
where τ accounts for the iterations immediately before failure, and the sum for the iterations preceding each
constructed improvement. Recall that E(S) = 1/P (Fm ) − 1 by the properties of the geometric distribution,
and observe that E(Wm | S ≥ 1) = E(Wm | Wm ≤ τ ) ≤ min(τ, E(Wm )). Using a waiting time argument for
E(Wm ) being equal to the reciprocal of the improvement probability for the mBitFlip operator, we get:
1
1
− 1 · min τ,
E(Xj | m) ≤ τ +
m−1
P (Fm )
m · n1 · n−(jn/w−1)−1
n
!
1
n
=τ+
− 1 · min τ,
mτ
m−1
m(1 − j/w)m−1
e− n (1−(j−1)/w)
− n1
j−1 m−1
mτ
n
+ O(1).
≤ τ + e n (1− w )
· min τ,
m(1 − j/w)m−1
Combining these conditional expectations yields
k
k
X
X
1
n
1 mτ (1− j−1 )m−1
n
w
E(Xj ) ≤
· min τ,
+ O(k)
· E(Xj | m) ≤ τ +
·e
k
k
m(1 − j/w)m−1
m=1
m=1
!
k
X
τ
1 mτ (1− j−1 )m−1
τ
1
w
≤n
· min
+ O(k)
+
·e n
,
n m=1 k
n m(1 − j/w)m−1
!
k
X
j−1 m−1
mτ
1
n kτ
τ
(1−
)
w
e n
+ O(k).
+
,
≤
· min
k
n
n m(1 − j/w)m−1
m=1
28
(11)
Runtime (/n2 )
0.60
2-op
3-op
4-op
5-op
0.50
0.40
0
5
10
15
20
25
30
35
40
45
50
τ (/n)
Figure 2: The leading constants in the theoretical upper bounds on the average number of fitness function
evaluations required by the k-operator Generalised Random Gradient hyper-heuristic to find the LeadingOnes optimum (w = 100, 000).
Finally, we return to bounding the overall runtime of the hyper-heuristic. Recall that Sj denoted the
number of iterations the algorithm spends in chunk j while using the random operator chosen during chunk
j − 1, and as E(Sj ) ≤ max (E(Xj | mBitFlip))m=1,...,k < kE(Xj ) = O (n exp(kτ /n)), and since w =
P
w
n
2
o exp(kτ
j=1 E(Sj+1 ) = o(n ). Substituting the bounds (10) and (11) into (9) yields the theorem
/n) ,
statement:
w
w
X
X
(E(Tj ) + E(Sj+1 )) ≤
E(Nj )E(Xj ) + o(n2 )
E(T ) ≤
j=1
j=1
w
X
1
n
×
·
≤ o(n2 ) +
2
w
j=1
2
kτ
n
+
i
h mτ
j−1 m−1
1
τ
n (1− w )
e
,
·
min
m=1
n m(1−j/w)m−1
P
.
j m−1
mτ
k
n (1− w )
−k
m=1 e
Pk
Figure 2 shows the relevant theoretical upper bounds for the k-operator variant of the Generalised
Random Gradient hyper-heuristic from Theorem 31. The upper bounds found by the hyper-heuristics with
more operators are better than the ones with less operators. In particular, the upper bounds for the any
hyper-heuristic with k-operators outperform the best possible runtime for any hyper-heuristic with less than
k operators. We depict this result for k = 3 and k = 5 in Figure 3.
7
Experimental Results
In the previous sections we have proven that the generalised hyper-heuristics are extremely fast for the
LeadingOnes benchmark function for large enough problem sizes n. In this section we present some
experimental results for the generalised hyper-heuristics in order to shed light on their performance for
realistic problem sizes. All parameter combinations have been simulated 10,000 times.
7.1
Two Low-level Heuristics (k = 2)
We first consider the generalised hyper-heuristics using two operators only (i.e., 1BitFlip and 2BitFlip),
and look at the impact of the parameter τ and problem size n.
Figure 4 shows the runtimes for the three generalised hyper-heuristics on LeadingOnes for n = 10, 000
and n = 50, 000, illustrating the effect τ has on the runtime.
29
0.44
2-op
3-op
0.43
Runtime (/n2 )
Runtime (/n2 )
0.44
2Opt
0.42
0.41
0.40
0
10
20
30
40
50
4-op
5-op
0.42
0.40
0.38
4Opt
0
10
20
τ (/n)
30
40
50
τ (/n)
Figure 3: A comparison of the optimal runtimes of the Generalised Random Gradient hyper-heuristic with
k-operators against the leading constant in the theoretical upper bound of the runtime of the Generalised
Random Gradient hyper-heuristic with k + 1-operators; for k = 2 and k = 4.
Runtime (/n2 )
0.6
0.55
RLS
0.5
Gen. Greedy
Gen. Random Gradient
Gen. Greedy Gradient
0.45
0.4
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
2.2
2.4
2.6
2.8
3
τ (/(n ln(n)))
Figure 4: Average number of fitness function evaluations required by the hyper-heuristics with k = 2
operators to find the LeadingOnes optimum for n = 10, 000 (solid), n = 50, 000 (dashed).
For the Generalised Greedy hyper-heuristic, we see that for the larger problem size, the runtimes improve
slightly. The figure suggests that for a sufficiently large τ the hyper-heuristic outperforms RLS; in particular,
for every value of τ greater than approximately 0.3n ln n (for n = 50, 000 this value is 162, 297). Also, for
every τ ≥ n ln n the runtime of the GG hyper-heuristic does not change much confirming the robustness of
the hyper-heuristic to the choice of τ , indicated by Corollary 19.
The Generalised Random Gradient hyper-heuristic performance seems more dependent on the choice
of τ . It is worth noting that as the problem size increases, for τ ≈ 0.55n ln(n), the runtime seems to be
n2 + o(n2 ) ≈ 0.42329n2). For well
approaching the optimal performance proven in Corollary 15 (i.e., 1+ln(2)
4
chosen τ values, the hyper-heuristic beats the runtime of RLS, and also the experimental runtime for the
recently presented k-bit mutation algorithm with self-adjusting k, of 0.450n2, for the given parameter choices
[8]2 . As τ increases past 0.6n2 , we see a detriment in the performance of the hyper-heuristic; it is worth
noting, however, that for n = 50, 000, it is required that τ > 1.5n ln(n) = 811, 483 to be worse than the RLS
expected runtime of 0.5n2 , indicating that the parameter is very robust.
The Generalised Greedy Gradient hyper-heuristic showcases similar performance reliance on τ as the
Generalised Random Gradient hyper-heuristic. The GGG hyper-heuristic, however, does have improved
performance over the GRG hyper-heuristic for both values of n, suggesting that incorporating the Decision
2 We did not perform an experimental comparison with this algorithm as it is unclear how to set the various parameters for
a fair comparison.
30
τ = 0.25n ln n
τ = 0.5n ln n
τ = 0.6n ln n
τ = 0.75n ln n
Runtime (/n2 )
0.55
0.5
0.45
0.4 2
10
103
104
105
n
106
107
108
Figure 5: Average number of fitness function evaluations required for the Generalised Random Gradient
hyper-heuristic with k = 2 operators to find the LeadingOnes optimum as the problem size n increases.
Runtime (/n2 )
0.60
0.55
RLS
0.50
2-op
3-op
4-op
5-op
0.45
0.40
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
τ (/(n ln(n)))
Figure 6: Average number of fitness function evaluations required by the Generalised Random Gradient
hyper-heuristic to find the LeadingOnes optimum, n = 100, 000.
Stage of the GG hyper-heuristic has a positive effect on the runtime.
Figure 5 shows the effects of increasing the problem size n for a variety of fixed τ values, for the GRG
hyper-heuristic. We can see that an increased problem size leads to faster runtimes. The performance
difference between the τ values decreases with increased n, suggesting that further increasing n would lead
to similar, optimal performance for a large range of values, as suggested by Corollary 15. For n = 108 , the
runtime for τ = 0.6n ln n is ≈ 0.4271n2, only slightly deviated from the optimal value of ≈ 0.4232n2.
7.2
More Than Two Low-level Heuristics (k ≥ 2)
We now consider the generalised hyper-heuristics using k operators, (1, . . . , k) bitflips, and look at the impact
of the parameter τ .
Figure 6 shows the runtimes for the Generalised Random Gradient hyper-heuristic when choosing from
2 to 5 separate mutation operators on LeadingOnes, for a problem size of n = 100, 000.
We see that incorporating more operators can be beneficial to the performance of the GRG hyper31
Runtime (/n2 )
0.60
2-op
3-op
4-op
5-op
0.55
0.50
0.45
0
2
4
6
8
10
12
14
16
18
20
22
24
τ (/n)
Figure 7: Average number of fitness function evaluations required by the Generalised Greedy hyper-heuristic
to find the LeadingOnes optimum, n = 100, 000.
Runtime (/n2 )
0.54
0.52
0.50
0.48
2-op
4-op
0.46
0
50
100
150
200
250
300
350
400
450
500
550
600
650
3-op
5-op
700
750
τ (/n)
Figure 8: Average number of fitness function evaluations required by the Generalised Greedy hyper-heuristic
to find the LeadingOnes optimum, n = 100, 000.
heuristic. Whilst the 2-operator hyper-heuristic achieves a best performance of approximately 0.44n2 , the
5-operator hyper-heuristic achieves a best performance of around 0.425n2. It is, however, important to set
τ optimally to achieve the best performance. For 0 ≤ 0.35n ln(n) ≤ τ and τ ≥ 1.5n ln(n), the 2-operator
hyper-heuristic outperforms the others; however, it is unable to achieve a better optimal performance. The
results suggest that the more operators that are incorporated in the hyper-heuristic, the shorter the range
of τ values for which it performs the best, in comparison with the hyper-heuristics with fewer operators.
Figure 7 shows the runtimes for the Generalised Greedy hyper-heuristic when choosing from 2 to 5
separate mutation operators on LeadingOnes, for a problem size of n = 100, 000.
Unlike the Generalised Random Gradient hyper-heuristic the Generalised Greedy, like the ‘simple’ mechanisms (see Theorem 29, Corollary 30), displays a detrimental performance when more operators are included.
Unlike the GRG hyper-heuristic, which exploits a high-performing operator while running for a shorter time
with low-performing operators, the GG hyper-heuristic runs each chosen operator for a fixed period of time.
Involving more operators means a lower chance of choosing the best operator in the Decision Stage.
Figure 8 shows the runtimes for the Generalised Greedy hyper-heuristic when choosing from 2 to 5 separate mutation operators on LeadingOnes, for a problem size of n = 100, 000, for a much larger range of
32
Runtime (/n2 )
0.60
0.55
RLS
0.50
2-op
3-op
4-op
5-op
0.45
0.40
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
τ (/(n ln(n)))
Figure 9: Average number of fitness function evaluations required by the Generalised Greedy Gradient
hyper-heuristic to find the LeadingOnes optimum, n = 100, 000.
τ values. We can see that the performance of the 2-operator hyper-heuristic remains the best performing
hyper-heuristic. Although the performance of the 5-operator hyper-heuristic is the worst, it is worth noting
than even this operator outperforms the runtime of the single operator algorithm RLS (i.e., 0.5n2 ). For
this hyper-heuristic, the detriment in performance between τ = 50n and τ = 750n is approximately 0.01n2 ,
confirming that the Generalised Greedy hyper-heuristic is robust to the choice of τ .
Figure 9 shows the runtime for the Generalised Greedy Gradient hyper-heuristic on LeadingOnes for
n = 100, 000.
While the hyper-heuristic displays similar performance to the GRG hyper-heuristic, we note that the
best performance of each k outperforms the respective selection for the GRG hyper-heuristic with the same
k. This suggests that, while the Decision Stage can be detrimental in the case of the GG hyper-heuristic,
the Exploitation Stage idea can correct for poor operator choices, leading to better performance. We can
see, for example, that the best performance of the 5-operator variation of the GGG hyper-heuristic, for
τ ≈ 0.62n ln(n), achieves a performance of approximately 0.418n2, outperforming the best possible 2-operator
runtime, shown in Corollary 25, of ≈ 0.423n2 .
8
Conclusion
The field analysing theoretically the performance of hyper-heuristics is growing. Previous results have shown
that the ‘simple’ learning mechanisms applied in combinatorial optimisation applications show no signs of
learning on LeadingOnes and perform equivalently to choosing operators at random, up to lower order
terms. We generalised the existing learning mechanisms to allow success to be measured over a longer period
of time, τ , rather than in a single iteration, and proved that these generalised hyper-heuristics are faster
than all standard evolutionary and local-search algorithms for LeadingOnes.
To apply the generalised hyper-heuristics, a value for the learning period τ is required. Although our
results indicate that τ is a fairly robust parameter (i.e., for n = 10, 000 all generalised hyper-heuristics
achieved faster experimental runtimes than that of the (1+1) EA for all tested values of τ between 1 and
1,000,000), setting it appropriately will lead to optimal performance. Concerning the Generalised Random
Gradient hyper-heuristic, clearly τ must be large enough to have at least a constant expected number of
successes within τ steps, if the hyper-heuristic has to learn about the operator performance. Obviously,
setting large values of τ may lead to large runtimes, since switching operators requires Ω(τ ) steps. Similar
considerations may also be made for the Generalised Greedy hyper-heuristic that experiments indicate is even
more robust to the choice of τ . Concerning the negative effects that may occur for too large values of τ , the
GapPath function can be considered as a worst-case scenario for the generalised hyper-heuristics, because
33
the operator has to change after every success. Nevertheless, we have proved that the hyper-heuristics are
asymptotically just as efficient as the ‘simple’ mechanisms for values of τ as large as O(n2 ).
The Generalised Greedy Gradient hyper-heuristic, which combines Generalised Greedy’s Decision Stage
with Generalised Random Gradient’s Exploitation Stage, can achieve even better performance than either
hyper-heuristic.
We finally have rigorously shown that the performance of the ‘simple’ mechanisms deteriorates as the
choice of operators (choosing from k ≥ 2) increases, while the performance of the Generalised Random
Gradient hyper-heuristic improves with a larger choice.
Several directions can be explored in future work. On one hand, we will analyse the hyper-heuristics’
performance on a broader class of problems, including classical ones from combinatorial optimisation, and
generalise the hyper-heuristics further by allowing the algorithms to automatically learn how to adapt τ . Finally, more sophisticated hyper-heuristics, such as the choice function approach [6] that have shown superior
performance in practical applications, should be analysed.
References
[1] F. Alanazi and P. K. Lehre. Runtime analysis of selection hyper-heuristics with classical learning mechanisms.
In 2014 IEEE Congress on Evolutionary Computation (CEC), pages 2515–2523. Springer, July 2014.
[2] Fawaz Alanazi and Per Kristian Lehre. Limits to learning in reinforcement learning hyper-heuristics. In Evolutionary Computation in Combinatorial Optimization (EvoCOP 2016), pages 170–185, Cham, 2016. Springer
International Publishing.
[3] Shahriar Asta and Ender Özcan. An apprenticeship learning hyper-heuristic for vehicle routing in HyFlex. In
IEEE Symposium on Evolving and Autonomous Learning Systems (EALS 2014), pages 65–72. IEEE, 2014.
[4] Süntje Böttcher, Benjamin Doerr, and Frank Neumann. Optimal fixed and adaptive mutation rates for the
leadingones problem. In Parallel Problem Solving from Nature (PPSN XI), pages 1–10, Berlin, Heidelberg, 2010.
Springer.
[5] Edmund K Burke, Michel Gendreau, Matthew Hyde, Graham Kendall, Gabriela Ochoa, Ender Özcan, and
Rong Qu. Hyper-heuristics: A survey of the state of the art. Journal of the Operational Research Society,
64(12):1695–1724, 2013.
[6] Peter Cowling, Graham Kendall, and Eric Soubeiga. A hyperheuristic approach to scheduling a sales summit.
In Practice and Theory of Automated Timetabling III, pages 176–190, Berlin, Heidelberg, 2001. Springer.
[7] Peter Cowling, Graham Kendall, and Eric Soubeiga. Hyperheuristics: A tool for rapid prototyping in scheduling
and optimisation. In Applications of Evolutionary Computing: EvoWorkshops 2002: EvoCOP, EvoIASP, EvoSTIM/EvoPLAN Kinsale, Ireland, April 3–4, 2002 Proceedings, pages 1–10, Berlin, Heidelberg, 2002. Springer.
[8] Benjamin Doerr, Carola Doerr, and Jing Yang. k-bit mutation with self-adjusting k outperforms standard
bit mutation. In Parallel Problem Solving from Nature (PPSN XIV), pages 824–834. Springer International
Publishing, 2016.
[9] Benjamin Doerr and Marvin Künnemann. How the (1+λ) evolutionary algorithm optimizes linear functions.
In Genetic and Evolutionary Computation Conference, GECCO ’13, Amsterdam, The Netherlands, July 6-10,
2013, pages 1589–1596, 2013.
[10] Jonathon Gibbs, Graham Kendall, and Ender Özcan. Scheduling english football fixtures over the holiday period
using hyper-heuristics. In Parallel Problem Solving from Nature (PPSN XI), pages 496–505, Berlin, Heidelberg,
2010. Springer.
[11] Jun He, Feidun He, and Hongbin Dong. Pure strategy or mixed strategy? In Evolutionary Computation in
Combinatorial Optimization: 12th European Conference (EvoCOP 2012), pages 218–229, Berlin, Heidelberg,
2012. Springer.
[12] Jun He and Xin Yao. Drift analysis and average time complexity of evolutionary algorithms. Artificial Intelligence, 127(1):57–85, 2001.
[13] Daniel Johannsen. Random combinatorial structures and randomized search heuristics. PhD thesis, Universita̋t
des Saarlandes, Postfach 151141, 66041 Saarbra̋cken, 2010.
[14] Per Kristian Lehre and Ender Özcan. A runtime analysis of simple hyper-heuristics: To mix or not to mix
operators. In Proceedings of the Twelfth Workshop on Foundations of Genetic Algorithms XII, FOGA XII ’13,
pages 97–104, New York, NY, USA, 2013. ACM.
34
[15] Andrei Lissovoi, Pietro S. Oliveto, and John Alasdair Warwicker. On the runtime analysis of generalised selection
hyper-heuristics for pseudo-boolean optimisation. In Proceedings of the Genetic and Evolutionary Computation
Conference, GECCO ’17, pages 849–856, New York, NY, USA, 2017. ACM.
[16] Eunice López-Camacho, Hugo Terashima-Marin, Peter Ross, and Gabriela Ochoa. A unified hyper-heuristic
framework for solving bin packing problems. Expert Systems with Applications, 41(15):6876–6889, 2014.
[17] İbrahim Maden, S. Uyar, and Ender Özcan. Landscape analysis of simple perturbative hyperheuristics. In 15th
International Conference on Soft Computing, pages 16–22. Mendel, 2009.
[18] Gabriela Ochoa, Rong Qu, and Edmund K. Burke. Analyzing the landscape of a graph based hyper-heuristic for
timetabling problems. In Proceedings of the 11th Annual Conference on Genetic and Evolutionary Computation,
GECCO ’09, pages 341–348, New York, NY, USA, 2009. ACM.
[19] Gabriela Ochoa, José Antonio Vázquez-Rodrı́guez, Sanja Petrovic, and Edmund Burke. Dispatching rules for
production scheduling: a hyper-heuristic landscape analysis. In IEEE Congress on Evolutionary Computation.,
pages 1873–1880. IEEE, 2009.
[20] P. S. Oliveto and Xin Yao. Runtime analysis of evolutionary algorithms for discrete optimization. In Anne Auger
and Benjamin Doerr, editors, Theory of Randomized Search Heuristics: Foundations and Recent Developments,
pages 21–52. World Scientific, 02 2011.
[21] Ender Özcan, Mustafa Misir, Gabriela Ochoa, and Edmund K. Burke. A reinforcement learning - great-deluge
hyper-heuristic for examination timetabling. Applied Metaheuristic Computing, 1(1):39–59, 2010.
[22] Abraham Wald. On cumulative sums of random variables. Ann. Math. Statist., 15(3):283–296, 09 1944.
[23] David H Wolpert and William G Macready. No free lunch theorems for optimization. IEEE Transactions on
Evolutionary Computation, 1(1):67–82, 1997.
35
| 9cs.NE
|
Generalized Nash Equilibrium Problem by the Alternating Direction
Method of Multipliers
arXiv:1703.08509v1 [cs.GT] 24 Mar 2017
Farzad Salehisadaghiani, and Lacra Pavel
Abstract— In this paper, the problem of finding a generalized
Nash equilibrium (GNE) of a networked game is studied.
Players are only able to choose their decisions from a feasible
action set. The feasible set is considered to be a private linear
equality constraint that is coupled through decisions of the
other players. We consider that each player has his own private
constraint and it has not to be shared with the other players.
This general case also embodies the one with shared constraints
between players and it can be also simply extended to the case
with inequality constraints. Since the players don’t have access
to other players’ actions, they need to exchange estimates of
others’ actions and a local copy of the Lagrangian multiplier
with their neighbors over a connected communication graph.
We develop a relatively fast algorithm by reformulating the
conservative GNE problem within the framework of inexactADMM. The convergence of the algorithm is guaranteed under
a few mild assumptions on cost functions. Finally, the algorithm
is simulated for a wireless ad-hoc network.
I. I NTRODUCTION
The problem of finding a generalized Nash equilibrium
(GNE) has recently drawn many attentions due to its applicability in various networked games with coupling constraints
such as power grids [1], optical networks [2] and wireless
communication networks [3].
In such games, each player aims to minimize his cost
function by taking a proper action in response to other
players. Each player’s feasible decision set is dependent on
other players’ actions. We are interested in seeking a GNE,
which is a point that no player can unilaterally deviate his
local action to minimize his cost function.
Due to similarities between this problem and distributed
consensus optimization problems (DCOPs), we aim to employ an efficient, robust and fast optimization technique
referred to as the alternating direction method of multipliers
(ADMM) to find a GNE of a multi-player game. The key
differences between DCOP and GNE problem is that in a
DCOP each agent desires to minimize a global objective by
controlling a full vector optimization variable. However, in a
GNE problem, we have a set of local optimization problems
assigned to each player who controls only his action (which
is an element of a full vector). This leads to have a suboptimization problem associated to each player such that
each of them is dependent on the other players’ actions.
Related Works. Our work is related to the literature on
(generalized) Nash games such as [4]–[7] and DCOPs such
as [8], [9]. (G)NE seeking in distributed networked games
The authors are with the Department of Electrical and Computer Engineering, University of Toronto, Toronto, ON M5S 3G4, Canada (emails: [email protected],
[email protected]).
has recently attracted an increased interest due to many realworld applications. To name only a few, [10], [11].
In [10], the problem of finding GNE is studied in a
networked game. A gradient-based algorithm is designed
over a complete communication graph. Convergence proof
is analyzed in the presence of delay and dynamic change.
Similar to [1], the problem of seeking a GNE is discussed
in [12] for the games that players have access to all players’
actions on which their cost depends on. Thus, they exchange
only local multipliers not the estimates of others’ actions.
A Nash game with a coupled constraint is considered in
[13]. A variational inequality related approach is used to
compute a GNE of the game. The authors in [14] solve a
quadratic GNE problem with linear shared coupled constraint
using Lemkes method.
Recently, there has been a widespread research on GNE
seeking in aggregative games [1], [13], [14]. Quadratic aggregative games with affine and convex coupling constraints
are studied in [15], [16]. A coordination scheme belonging
to a class of asymmetric projection algorithm is presented
in [15] and its convergence to a GNE is then discussed.
A model-free dynamic control law, based on monotone
operator, is proposed in [16] to ensure the global convergence
to a GNE.
ADMM algorithms, which are in the scope of this paper,
have been developed in 70’s to find an efficient way to
obtain an optimal point of distributed optimization problems
(DOPs) [17], [18]. After the re-introduction of this method
in [18], ADMM has become widely used to locate optimal
points of DOPs as a robust and fast technique [19], [20].
An inexact-ADMM algorithm is proposed in [8] for DCOPs
with an affine equality constraint. In [21], we develop a
methodology to relate a distributed NE problem with no
coupled constraint to a DCOP using augmentation technique.
It is shown that an NE problem can be treated as a set of suboptimization problems with an augmented equality constraint
for the estimates of the players.
Contributions. Motivated by ADMM methods designed
for DCOPs, we develop a relatively fast algorithm for GNE
problems within the framework of inexact-ADMM. Players
are only aware of their own cost functions (which are not
in the form of aggregative but general game), problem data
(which is related to a private coupled equality constraint for
each player) and action set of all players (they are not aware
of the others’ actions). We reformulate the conservative GNE
problem into the corresponding Lagrange dual problem and
then we augment it using local estimates of players’ actions
as well as local copies of Lagrange multipliers. We derive a
set of GNE conditions using the associated KKT conditions
and finally we develop an inexact-ADMM algorithm based
on the augmented Lagrange function which is related to each
player. The convergence proof of the algorithm to a GNE of
the game is then provided under a few mild assumptions on
players’ cost functions.
The paper is organized as follows. The problem statement
and assumptions are provided in Section II. In Section III,
an inexact-ADMM-like algorithm is derived. Convergence of
the proposed algorithm to a GNE of the game is discussed
in Section IV. Simulation results are illustrated in Section V
and concluding remarks are presented in Section VI.
II. P ROBLEM S TATEMENT
Consider V = {1, . . . , N } as a set of N players that seek
a GNE of a networked game with individual linear coupled
constraints. The game is denoted by G and defined as follows:
• Ωi ⊂ R: Action set of player i, ∀i ∈ V ,
Q
N
• Ω=
i∈V Ωi ⊂ R : Action set of all players,
• Ji : Ω → R: Cost function of player i, ∀i ∈ V .
The game G(V, Ωi , Ji ) is defined over the set of players,
V , the action set of player i ∈ V , Ωi and the cost function
of player i ∈ V , Ji .
The players’ actions are denoted as follows:
• x = (xi , x−i ) ∈ Ω: All players actions,
• xi ∈ Ωi : Player i’s action, ∀i ∈ V ,
Q
• x−i ∈ Ω−i :=
j∈V \{i} Ωj :All players’ actions except i.
The local data of player i is given as Ai := [Aii , [Aij ]j6=i ] ∈
Rm×N with Aii ∈ Rm , and bi ∈ Rm . Note that in this work,
we are interested in a more general case in which each player
has access only to his own private constraint which has not
to be shared with the other players.
Assumption 1. 1 There exists a positive semi-definite matrix
B i ∈ Rm×m for all i ∈ V such that,
Aii ∈ ker(B i − Im ), Aij ∈ ker(B i ) j 6= i.
(1)
Player i’s feasible strategy set is then denoted by
Xi (x−i ) := {xi ∈ Ωi |Ai x = bi }2 .
The game is played such that for a given x−i ∈ Ω−i ,
each player i ∈ V aims to minimize his own cost function
selfishly w.r.t. xi subject to a private equality constraint.
(
minimize Ji (xi , x−i )
xi
(2)
s.t.
xi ∈ Xi (x−i ).
Each optimization problem is run by a particular player i at
the same time with other players. Note that (2) can simply
accommodate inequality constraints Ai x ≥ bi by adding a
slack variable y ≥ 0 and rewrite the inequality constraint as
Ai x−y = bi . Note that the constraint y ≥ 0 can be enforced
by a convex indicator function.
1 Derivation of the algorithm is not dependent on Assumption 1. Simulations for various cases for matrix Ai , even the ones that do not satisfy this
assumption, verify that the algorithm works even without considering this
assumption. However, this assumption is used at the convergence proof.
2 This case contains games with shared constraints for which Ax = b for
all i ∈ V and m ≥ N
An NE of a game is defined as follows:
Definition 1. Consider an N -player game G(V, Ωi , Ji ), each
player i minimizing the cost function Ji : Ω → R. A vector
x∗ = (x∗i , x∗−i ) ∈ Ω is called a GNE of this game if
Ji (x∗i , x∗−i ) ≤ Ji (xi , x∗−i ) ∀xi ∈ Xi (x∗−i ), ∀i ∈ V.
(3)
An NE lies at the intersection of all solutions of the set
(2). The key challenges are two-fold. First, each optimization
problem in (2) is dependent on the solution of the other
simultaneous problems; second, the solutions are coupled
through a linear constraint.
Since this game is distributed, no player is aware of the
actions, cost functions and constraints of the other players.
Thus, each player i maintains an estimate of the other
players’ actions. We define a few notations, in the following,
for players’ estimates.
i
i
i
• x = (xi , x−i ) ∈ Ω: Player i’s estimate of all players
actions,
i
• xi ∈ Ωi : Player i’s estimate of his own action which is
i
indeed his action,
Q i.e., xi = xi ∀i ∈ V ,
i
• x−i ∈ Ω−i :=
j∈V \{i} Ωj : Player i’s estimate of all
other players’ actions except his action,
T
1T
, . . . , xN ]T ∈ ΩN : Augmented vector of
• x = [x
estimates of all players’ actions
Note that all players’ actions can be interchangeably represented as x = (xii )i∈V .
We assume that the cost function Ji , the action set Ω and
the problem data Ai , bi are the only information available to
player i. Thus, the players exchange their estimates in order
to update their actions. An undirected communication graph
GC (V, E) is defined with E ⊆ V × V denoting the set of
communication links between the players. (i, j) ∈ E if and
only if players i and j exchange estimates. In the following,
we have a few definitions for GC :
• Ni := {j ∈ V |(i, j) ∈ E}: Set of neighbors of i in GC ,
• H := [hij ]i,j∈V : Adjacency matrix associated with GC
where hij = 1 if (i, j) ∈ E and hij = 0 otherwise,
• D := diag{|N1 |, . . . , |NN |}: Degree matrix associated
with GC .
The following assumption is used for GC .
Assumption 2. GC is a connected graph.
We aim to link game (2) to a set of optimization problems
whose solutions can be based on the ADMM. Game (2) is
equivalently represented as the following problem for i ∈ V .
(
min Ji (xi , x−i )
xi ∈Ωi
(4)
s.t. Ai x = bi .
Let λ ∈ Rm be the joint Lagrange dual variable associated
with the linear constraint Ai x = bi ∀i ∈ V . The Lagrange
dual problem of (4) can be written as follows for i ∈ V .
n
maxm min Ji (xi , x−i ) + λT (Ai x − bi ).
(5)
λ∈R xi ∈Ωi
Note that we are interested in computing a variational
GNE of the game since we assumed a common Lagrange
multiplier λ for each player in (5). Using the estimates
of the actions xi and the local copies of the Lagrange
dual variables λi for i ∈ V , we reformulate (5) so that
the objective function is separable (the estimates are also
interpreted as the local copies of x). Moreover, the slack
variables tls |{l∈V, s∈Nl } ∈ Ω and pls |{l∈V, s∈Nl } ∈ Rm are
employed to enforce that the local copies are equivalent.
T
Ji(xii , xi−i )+IΩi (xii )+λi (Ai xi −bi )
max
min
i
i
m
ls
ls
λ ∈R ,p xi ∈Ωi ,t
s.t.
xl = tls ∀l ∈ V, ∀s ∈ Nl ,
xs = tls
∀l ∈ V, ∀s ∈ Nl ,
l
ls
∀l ∈ V, ∀s ∈ Nl ,
s
ls
∀l ∈ V, ∀s ∈ Nl ,
λ =p
λ =p
(6)
i
0
if
x
∈
Ω
i
i
is an indicator function
where IΩi (xii ) :=
∞ otherwise
of the feasibility constraint xii ∈ Ωi . Note that the solution
of (6) is equivalent to that of (5), i.e., xii = xi , λi = λ
for all i ∈ V . The ADMM is then employed to solve (6)
in a distributed manner. A characterization of the NE for
game (2) could be obtained by finding the KKT conditions on
(6). Let {uls , v ls , wls , z ls }l∈V,s∈Nl with uls , v ls ∈ RN and
wls , z ls ∈ Rm be the Lagrange multipliers associated with
the four constraints in (6), respectively. The corresponding
Lagrange function for player i, ∀i ∈ V is as follows:
Li xii , {tls , uls , v ls }, λi , {pls , wls , z ls }
(
T
:= Ji (xii , xi−i ) + IΩi (xii ) + λi (Ai xi − bi )
XX
T
T
+
uls (xl − tls ) + v ls (xs − tls ),
l∈V s∈Nl
−
XX
T
T
wls (λl − pls ) + z ls (λs − pls ).
(7)
l∈V s∈Nl
∗
∗
∗
∗
∗
∗
Let (xi , λi )i∈V and {uls , v ls , wls , z ls }l∈V, s∈Nl be a
pair
and dual optimal
to (6). Let denote
P of primal
P solution
ij
ji
i
ij
u
+
v
:=
U
and
w
+ z ji := W i for
j∈Ni
j∈Ni
notational simplicity. The KKT conditions are summarized
as follows for i ∈ V, j ∈ Ni :
∗
∗
∗T
∗
∇i Ji (xi ) + ∂IΩi (xii ) + λi Aii + Uii = 0,
Ai xi ∗ − bi − W i ∗ = 0 ,
m
(8)
i∗
j∗
i∗
j∗
x
=
x
,
λ
=
λ
,
ij ∗
∗
∗
∗
u + v ij = 0N , wij + z ij = 0m ,
where ∇i Ji (·) is gradient of Ji w.r.t. xi and ∂i IΩi (·) is a
∗
subgradient of IΩi at xi . By (8) and Assumption 2, x1 =
N∗
∗
1∗
N∗
∗
. . . = x := x and λ = . . . = λ := λ . Then, the GNE
of (2), x∗ , satisfies the following equations ∀i ∈ V, j ∈ Ni :
∗
∗
∗T i
i∗
∇i Ji (x ) + ∂IΩi (xi ) + λ Ai + Ui = 0,
∗
(9)
Ai x∗ − bi − W i = 0m ,
∗
∗
ij
ij
ij ∗
ij ∗
u + v = 0N , w + z = 0m .
In the following, we state a few assumptions for the existence
of a GNE.
Assumption 3. For every i ∈ V ,
• Ωi ⊂ R is compact and convex
• Xi (x−i ) is non-empty for every x−i
1
• Ji (xi , x−i ) is C in xi , jointly continuous in x and
convex in xi , for every x−i .
The convexity of Ωi implies that the indicator function IΩi
is convex. This yields that there exists at least one bounded
subgradient ∂IΩi .
Assumption 4. Let F : ΩN → RN , F (x) := [∇i Ji (xi )]i∈V
be the pseudo-gradient vector (game map) where x :=
T
T
[x1 , . . . , xN ]T ∈ ΩN . F is cocoercive ∀x ∈ ΩN and
y ∈ ΩN , i.e.,
(F (x) − F (y))T (x − y) ≥ σF kF (x) − F (y)k2 , (10)
where σF > 0.
III. D ISTRIBUTED I NEXACT-ADMM A LGORITHM
Our objective is to find an ADMM-like algorithm for
computing a GNE of G(V, Ωi , Ji ) using only imperfect
information over the communication graph GC (V, E).
The algorithm is elaborated in the following steps:
1- Initialization Step: Each player i ∈ V maintains an
initial estimate for all players’ actions, xi (0) ∈ Ω and an
initial Lagrange multiplier, λi (0) ∈ Rm . The initial values
of uij (0), v ij (0), wij (0) and z ij (0) are all set to be zero for
all i ∈ V , j ∈ Ni .
2- Communication Step: At iteration T (k), each player
i ∈ V exchanges his previous estimate of the other players’
actions xi (k − 1) and his dual Lagrange multiplier λi (k − 1)
with his neighbors j ∈ Ni .
3- Action Update Step: At this moment all players update
their actions, estimates and Lagrange multiplier via the
ADMM. For each player i ∈ V , let the augmented Lagrange
function associated with (6) be as follows:
Lai xii , {tls , uls , v ls }, λi , {pls , wls , z ls }
T
:= Ji (xii , xi−i ) + IΩi (xii ) + λi (Ai xi − bi )
XX
T
T
+
uls (xl − tls ) + v ls (xs − tls )
l∈V s∈Nl
−
XX
T
T
wls (λl − pls ) + z ls (λs − pls )
l∈V s∈Nl
cXX
+
(kxl − tls k2 + kxs − tls k2 )
2
l∈V s∈Nl
cXX
−
(kλl − pls k2 + kλs − pls k2 ),
2
(11)
l∈V s∈Nl
where c > 0 is a scalar coefficient. Consider the ADMM
algorithm associated with (6) based on (11):
The dual Lagrange multipliers uij , v ij , wij , z ij update
rules ∀i ∈ V, j ∈ Ni are as follows:
c
uij (k) = uij (k − 1) + xi (k − 1) − xj (k − 1) , (12)
2
c
v ij (k) = v ij (k − 1) + xj (k − 1) − xi (k − 1) , (13)
2
c
ij
ij
w (k) = w (k − 1) + λi (k − 1) − λj (k − 1) , (14)
2
z ij (k) = z ij (k − 1) +
c j
λ (k − 1) − λi (k − 1) .
2
(15)
as follows:
n
xii (k) = arg min
max
Ji (xii , xi−i (k − 1)) + IΩi (xii )
i
i
xi
The update rule for the slack variable tij ∈ RN , ∀i ∈
V, j ∈ Ni , which is based on (11), is as follows:
a
i
ls
ls
ls
tij (k) = arg min
L
i x (k − 1), {t , u (k), v (k)},
tij
λi (k − 1), {pls (k − 1), wls (k), z ls (k)}
n
− (uij (k) + v ij (k))T tij
= arg min
tij
o
c
+ (kxi (k − 1) − tij k2 + kxj (k − 1) − tij k2 )
2
1
1
= (uij (k) + v ij (k)) + (xi (k − 1) + xj (k − 1)).(16)
2c
2
The initial conditions uij (0) = v ij (0) = 0N ∀i ∈ V, j ∈ Ni
along with (12) and (13) suggest that uij (k) + v ij (k) = 0N
∀i ∈ V, j ∈ Ni , k > 0. Then, (16) yields,
tij (k) =
xi (k − 1) + xj (k − 1)
.
2
(17)
Similarly the update rule for pij ∈ Rm ∀i ∈ V, j ∈ Ni is
represented as the following:
λi (k − 1) + λj (k − 1)
.
(18)
2
The local estimate update for all i ∈ V is a min-max
optimization problem based on (11).
a
xii (k) = arg min
max
L
xii , {tls (k), uls (k), v ls (k)}, λi ,
i
xii ∈R λi ∈Rm
{pls (k), wls (k), z ls (k)}
n
= arg min
max
Ji (xii , xi−i (k − 1)) + IΩi (xii )
i
i
pij (k) =
xi
T
T
+Uii (k)T xii − W i (k)T λi
X
xi (k − 1) + xji (k − 1)
+c
xii − i
2
λi −
j∈Ni
λi (k − 1) + λj (k − 1)
2
j∈Ni
iT
2o
∀i ∈ V.
(19)
xii −
λ
xii (k − 1) + xji (k − 1)
2
2
+ Uii (k)T xii
(20)
Aii xii + Ai−i xi−i (k − 1) − bi − W i (k)
o
X
+c
(λi (k − 1) + λj (k − 1)) − c|Ni |kλi k2
+λ
j∈Ni
Substituting back (22) into (21), we obtain,
n
xii (k) = arg min
Ji (xii , xi−i (k − 1)) + IΩi (xii )
i
xi
+c
X
j∈Ni
+
1
4c|Ni |
xii (k − 1) + xji (k − 1) 2
+ Uii (k)T xii
2
X
c
(λi (k − 1) + λj (k − 1))
xii −
j∈Ni
i
−W (k) + Aii xii + Ai−i xi−i (k − 1) − bi
2o
.
(23)
T
nh
Aii i i
∇i Ji (xi (k − 1)) +
A x (k − 1)
2c|Ni |
xi
iT
X
−bi − W i (k) + c
(λi (k − 1) + λj (k − 1))
.
j∈Ni
xi
X
Only the third line of (21) is dependent on λi which is the
maximization variable. Thus, to maximize (21) w.r.t. λi , we
let λi (k) be as follows:
1 i i
λi (k) =
A x (k) + Ai−i xi−i (k − 1) − bi
2c|Ni | i i
X λi (k − 1) + λj (k − 1)
.
(22)
−W i (k) + 2c
2
xii (k) = arg min
i
In the derivation of (19), we used (17) and (18). The next
min-max problem is equivalent in the sense that its solutions
for xii and λi are equivalent to the solutions of (19).
n
xii (k) = arg min
max
Ji (xii , xi−i (k − 1)) + IΩi (xii )
i
i
+c
xii −
2
j∈Ni
X
X
We simplify (23) by using a proximal first-order approximation for Ji (xii , xi−i (k − 1)) and the last term in (23) around
xi (k − 1); thus using inexact ADMM it follows:
−λi bi + λi Aii xii + λi Ai−i xi−i (k − 1)
−c
xii (k − 1) + xji (k − 1) 2
+ Uii (k)T xii
2
j∈Ni
1 h X i
(λ (k − 1) + λj (k − 1))
−c|Ni | λi −
c
2c|Ni |
j∈Ni
i 2
i
i i
i
−W (k) + Ai xi + A−i xi−i (k − 1) − bi
X
1
+
c
(λi (k − 1) + λj (k − 1))
4c|Ni |
j∈Ni
2o
.
(21)
−W i (k) + Aii xii + Ai−i xi−i (k − 1) − bi
+c
λ
T
λ
∀i ∈ V.
j∈Ni
For the last two lines of (20), we complete the squared term
βi i
kx − xii (k − 1)k2 + IΩi (xii )
2 i
o
X
xi (k − 1) + xji (k − 1) 2
+c
xii − i
+ Uii (k)T xii , (24)
2
.(xii
−
xii (k
− 1)) +
j∈Ni
where βi > 0 is a penalty factor for the proximal first-order
approximation for i ∈ V . We obtain an equivalent problem
to (24) in the following such that its gradient w.r.t. xii is
equivalent to that of (24).
n
αi i
−1
i
xii (k) = arg min
I
(x
)
+
x
−
α
βi xii (k − 1)
Ω
i
i
i
2 i
xii
−∇i Ji (xi (k − 1)) − Uii (k) + c
X
(xii (k − 1) + xji (k − 1))
j∈Ni
T
Aii
1:
Ai xi (k − 1) − bi − W i (k)
2c|Ni |
2 o
X
,
λi (k − 1) + λj (k − 1)
+c
−
Algorithm 1 ADMM Algorithm
2:
3:
(25)
4:
j∈Ni
where αi := βi + 2c|Ni |. Let proxag [s] := arg minx {g(x) +
− sk2 } be the proximal operator for the non-smooth
i
function g. Note that proxα
IΩi [s] = TΩi [s] where TΩi : R →
Ωi is an Euclidean projection. Then for each player i, ∀i ∈ V
we obtain,
h
xii (k) = TΩi αi−1 γi xii (k − 1) − ∇i Ji (xi (k − 1)) − Uii (k)
a
2 kx
5:
6:
7:
initialization xi (0) ∈ Ω, λi (0) ∈ Rm , U i (0) = 0N and
W i (0) = 0m ∀i ∈ V
for k = 1, 2, . . . do
for each player i ∈ V do
players i, j ∀j ∈ Ni exchange xi (k − 1),
xj (k − 1), λi (k − 1)Pand λj (k − 1).
U i (k) = U i (k−1)+c j∈Ni (xi (k−1)−xj (k−1))
P
i
j
W i (k) = W ih(k−1)+c
j∈Ni (λ (k−1)−λ (k−1))
xii (k) = TΩi αi−1 γi xii (k −1)−∇i Ji (xi (k −1))
P
−Uii (k) + c j∈Ni xji (k − 1)
T
Aii
Ai−i xi−i (k − 1) − bi − W i (k)
− 2c|N
i|
i
P
+c j∈Ni λi (k − 1) + λj (k − 1)
1
λi (k) = 2c|N
Aii xii (k) + Ai−i xi−i (k − 1) − bi
i|
i
j
P
(k−1)
−W i (k) + 2c j∈Ni λ (k−1)+λ
2
T
Aii i i
A x (k − 1) − bi − W i (k)
2c|Ni | −i −i
j∈Ni
i
X
+c
λi (k − 1) + λj (k − 1)
,
(26)
+c
X
xji (k − 1) −
8:
xi−i (k) =
10:
end for
11: end for
P
9:
j∈Ni
T
1
where γi := βi + c|Ni | − 2c|N
Aii Aii .
i|
Eventually, after updating his action and dual Lagrange
multiplier, each player i takes average of the received information with his estimate and updates his estimate as follows:
1
1 X j
1
xi−i (k) = xi−i (k−1)+
U i (k),
x−i(k−1) −
2
|Ni |
2c|Ni | −i
j∈Ni
|
{z
}
{z
}
|
RECEIVED INFORMATION
PENALTY TERM
(27)
where c > 0 is a scalar coefficient which is also used in (11).
Note that in (27), the penalty term is associated with the
difference between the estimates of the neighboring players
(Equations (12), (13)).
Then the ADMM algorithm is represented as in Algorithm 1.
IV. C ONVERGENCE P ROOF
Theorem 1. Let βi > 0 be player i’s penalty factor of
the approximation in the inexact ADMM Algorithm 1 which
satisfies
1
σF >
,
(28)
kAi k2
2βi − c|Ni i |
where σF is a positive constant for the cocoercive property
of F , Aii is player i’s problem data and Ni is the number of
neighbors of player i. Under Assumptions 1-4, the sequence
{xi (k)} ∀i ∈ V , generated by Algorithm 1 converges to x∗
NE of game (2).
Proof . From step 7 in Algorithm 1, we obtain the following optimality condition:
+2c
X
j∈Ni
xii (k) −
j∈Ni
xj−i (k−1))
|Ni |
−
i
U−i
(k−1)
2c|Ni |
xii (k − 1) + xji (k − 1)
= 0 ∀i ∈ V,
2
Adding and subtracting Aii xii (k) from the second line of (29)
and using (22), it leads to
T
∇i Ji (xi (k − 1)) + ∂i IΩi (xii (k)) + Aii λi (k)
+δi (xii (k) − xii (k − 1)) + Uii (k)
(30)
j
i
X
x (k − 1) + xi (k − 1)
+2c
= 0 ∀i ∈ V,
xii (k) − i
2
j∈Ni
T
1
Aii Aii . Adding the first Nash
where δi := βi − 2c|N
i|
condition (9) and multiplying by (xii (k) − x∗i ), we obtain,
T
∇i Ji (xi (k − 1)) − ∇i Ji (x∗ ) (xii (k − 1) − x∗i )
T
+ ∇i Ji (xi (k − 1)) − ∇i Ji (x∗ ) (xii (k) − xii (k − 1))
T
+ ∂i IΩi (xii (k)) − ∂i IΩi (x∗i ) (xii (k) − x∗i )
T
+ λi (k) − λ∗ Aii (xii (k) − x∗i )
(31)
+δi (xii (k) − xii (k − 1))T (xii (k) − x∗i )
∗ T
+ Uii (k) − Uii
(xii (k) − x∗i )
X
xi (k − 1)+xji (k − 1)T i
+2c
xii (k)− i
(xi (k)−x∗i ) = 0
2
j∈Ni
∇i Ji (xi (k − 1)) + βi (xii (k) − xii (k − 1)) + ∂i IΩi (xii (k))
T
Ai i i
+ i
A x (k − 1) − bi − W i (k)
(29)
2c|Ni |
X
+c
(λi (k − 1) + λj (k − 1)) + Uii (k)
j∈Ni
On the other hand, from (22) we obtain,
0m = −Aii xii (k) − Ai−i xi−i (k − 1) + bi
X
λi (k) + λj (k)
+W i (k) + 2c
λi (k) −
2
j∈Ni
X
+c
(λi (k) + λj (k) − λi (k − 1) − λj (k − 1)),(32)
j∈Ni
i
i
Note that by (14) and (15), W (k + 1) = W (k) +
i
j
P
(k)
2c j∈Ni λi (k) − λ (k)+λ
. Then, We add the second
2
Nash condition (9) to (32). Moreover, using Assumption 1,
there exists a matrix B i such that we multiply into (32) and
then multiply the result by (λi (k) − λ∗ ).
T
0 = −(xii (k) − x∗i )T Aii (λi (k) − λ∗ )
∗ T iT
B (λi (k) − λ∗ )
+ W i (k + 1) − W i
X
+c
(λi (k) + λj (k) − λi (k − 1) − λj (k − 1))T
.B
(λi (k) − λ∗ ).
j∈Ni
iT
i
(λ (k) − λ∗ )
X
(λi (k) + λj (k) − λi (k − 1) − λj (k − 1))T
+c
.B
j∈Ni
(33)
(λi (k) − λ∗ ).
T
X
xi (k) + xj (k) − xi (k − 1) − xj (k − 1)
+c
.B
The second and third terms are bounded as follows:
T
∇i Ji (xi (k − 1))−∇i Ji (x∗ ) (xii (k)−xii (k − 1)) ≥ (37)
ρ
−1
k∇i Ji (xi (k − 1))−∇i Ji (x∗ )k2 − kxii (k)−xii (k − 1)k2 ,
2ρ
2
(∂IΩi (xii (k)) − ∂IΩi (x∗i ))T (xii (k) − x∗i ) ≥ 0.
1
1
kF (x(k − 1)) − F (x∗ )k2 − kx(k) − x(k − 1)k2M1
2ρ
2
+(x(k) − x(k − 1))T diag((δi ei eTi )i∈V )(x(k) − x∗ )
X X ij
∗
ji ∗ T i
∗
+
(ui (k + 1)+viji (k + 1)−uij
i −vi ) (xi (k)−xi)
−
j∈Ni
(λi (k) − λ∗ )
(34)
j
i
X
T
x (k − 1)+xi (k − 1)
(xii (k)−x∗i ) = 0.
+2c
xii (k)− i
2
j∈Ni
The last equation that we need to add it to (34) is the one
associated with xi−i , (27), and multiplied by (xi−i − x∗−i ).
XX
ji
T
i
∗
(uij
−i (k + 1) + v−i (k + 1)) (x−i (k) − x−i )
i∈V j∈Ni
2
T
+ (w(k + 1) − w∗ )T diag((B i )i∈V )(w(k + 1) − w(k))
c
+c(λ(k) − λ(k − 1))T ((D + H) ⊗ Im )
.((D + H) ⊗ IN )(x(k) − x∗ ) ≤ 0,
(39)
∗T
(35)
Adding (35) to (34), and using (12) and (13) for simplification, we obtain,
T
∇i Ji (xi (k − 1)) − ∇i Ji (x∗ ) (xii (k − 1) − x∗i )
T
+ ∇i Ji (xi (k − 1)) − ∇i Ji (x∗ ) (xii (k) − xii (k − 1))
T
+ ∂i IΩi (xii (k)) − ∂i IΩi (x∗i ) (xii (k) − x∗i )
+δi (xii (k) − xii (k − 1))T (xii (k) − x∗i )
X ij
∗
ji ∗ T i
∗
+
(ui (k + 1)+viji (k + 1)−uij
i −vi ) (xi (k)−xi )
j∈Ni
j∈Ni
+
T
j∈Ni
ji
T
i
∗
(uij
−i (k + 1) + v−i (k + 1)) (x−i (k) − x−i )
i∈V j∈Ni
.diag((B i )i∈V )(λ(k) − λ∗ ) + c(x(k) − x(k − 1))T
i
U−i
(k)T (xi−i (k) − x∗−i )
X
xi (k − 1) + xj−i (k − 1) T
+2c
xi−i (k) − −i
2
X
(38)
Using (37) and (38) in (36) and summing over i ∈ V , we
obtain,
T
F (x(k − 1)) − F (x∗ ) (x(k − 1) − x∗ )
iT
+
(36)
for any ρ > 0 ∀i ∈ V . By the convexity of IΩi (Assumption 3), we have for the third term,
+δi (xii (k) − xii (k − 1))T (xii (k) − x∗i )
∗ T
+ Uii (k) − Uii
(xii (k) − x∗i )
∗
T
+ W i (k + 1) − W i )T B i (λi (k) − λ∗ )
X
+c
(λi (k) + λj (k) − λi (k − 1) − λj (k − 1))T .
.(xi−i (k) − x∗−i ) = 0.
iT
j∈Ni
Adding (33) to (31), it yields,
T
∇i Ji (xi (k − 1)) − ∇i Ji (x∗ ) (xii (k − 1) − x∗i )
T
+ ∇i Ji (xi (k − 1)) − ∇i Ji (x∗ ) (xii (k) − xii (k − 1))
T
+ ∂i IΩi (xii (k)) − ∂i IΩi (x∗i ) (xii (k) − x∗i )
.B
∗
∗
(wij (k + 1) + z ji (k + 1) − wij − z ji )T
.(xii (k) − x∗i ) = 0.
j∈Ni
iT
X
+
∗T
where M1 := diag((ρei eTi )i∈V ), x∗ = [x1 , . . . , xN ]T ,
T
∗T
T
∗T
λ = [λ1 , . . . , λN ]T and P
λ∗ = [λ1 , . . . , λN ]T . Moreover, w = (wi )i∈V ∈ Rm i∈V |Ni | and wi =P(wij )j∈Ni ∈
∗
Rm|Ni | and also w∗ = (wi )i∈V ∈ Rm i∈V |Ni | and
∗
∗
wi = (wiij )j∈Ni ∈ Rm|Ni | .
We bound
the first term using
TAssumption 4,
∗
F (x(k − 1)) − F (x ) (x(k − 1) − x∗ )
≥ σF kF (x(k − 1)) − F (x∗ )k2 .
(40)
We also simplify the fourth and the fifth lines in
(39).
GC is P
an undirected
graph, for any {aij },
P Since
P
P
a
=
a
. Then,
ij
ji
i∈V
i∈V
j∈Ni
i
X
X j∈N
∗
ij
ji
ij ∗
(ui (k + 1)+vi (k + 1)−ui −viji )T(xii (k)−x∗i )
i∈V j∈Ni
+
XX
i∈V j∈Ni
ji
T
i
∗
(uij
−i (k + 1) + v−i (k + 1)) (x−i (k) − x−i )
=
XX
+
XX
∗
ij T
i
∗
(uij
i (k + 1) − ui ) (xi (k) − xi )
i∈V j∈Ni
∗
(viij (k + 1) − viij )T (xji (k) − x∗i )
i∈V j∈Ni
+
XX
T
i
∗
uij
−i (k + 1) (x−i (k) − x−i )
i∈V j∈Ni
+
XX
ij
(k + 1)T (xj−i (k) − x∗−i ).
v−i
(41)
i∈V j∈Ni
Note that by (12) and (13) as well as the initial conditions for
Lagrange multipliers uij (0) = v ij (0) = 0N ∀i ∈ V, j ∈ Ni ,
we obtain,
uij (k) + v ij (k) = 0N
∀i ∈ V, j ∈ Ni , k > 0.
(42)
Substituting
and using (9), we obtain,
X X ij(42) into (41)
∗ T
j
i
(ui (k + 1) − uij
i ) (xi (k) − xi (k))
i∈V j∈Ni
+
XX
j
T
i
uij
−i (k + 1) (x−i (k) − x−i (k))
i∈V j∈Ni
=
∗
2 X X ij
T
ij
ij
(u (k + 1) − uij
i ei ) (u (k + 1) − u (k))
c
i∈V j∈Ni
2
:= (u(k + 1) − u∗ )T (u(k + 1) − u(k)),
c
(43)
P
i
ij
where u = (ui )i∈V ∈ RN i∈V |Ni | and
P u = (u )j∈N∗i ∈
|N
|
N
∗
i∗
N |Ni |
i
i∈V
and ui =
R
and also u = (u )i∈V ∈ R
ij ∗
(ui )j∈Ni ⊗ ei ∈ RN |Ni | . Using (40) and (43), for ρ = 2σ1F
(39) becomes,
1
− kx(k) − x(k − 1)k2M1
2
+(x(k) − x(k − 1))T M2 (x(k) − x∗ )
∗
T
+(λ(k) − λ(k − 1)) Q(λ(k) − λ )
2
+ (u(k + 1) − u∗ )T (u(k + 1) − u(k))
c
2
+ (w(k + 1) − w∗ )T B(w(k + 1) − w(k)) ≤ 0, (44)
c
where M2 := diag (δi ei eTi )i∈V + c((D + H) ⊗ IN ), B :=
T
o
1
1
+ ku(k + 1) − u∗ k2 + kw(k + 1) − w∗ k2B ≤
c
c
n1
1
∗ 2
kx(k − 1) − x kM2 + kλ(k − 1) − λ∗ k2Q
2
2
o
1
1
∗ 2
+ ku(k) − u k + kw(k) − w∗ k2B
c
c
1
1
− kx(k) − x(k − 1)k2M2 −M1 − kλ(k) − λ(k − 1)k2Q
2
2
1
1
2
− ku(k + 1) − u(k)k − kw(k + 1) − w(k)k2B . (47)
c
c
By the condition (28), M2 − M1 0. Then (47) implies the
following two results:
1) 12 kx(k) − x∗ k2M2 + 12 kλ(k) − λ∗ k2Q + 1c ku(k + 1) −
u∗ k2 + 1c kw(k + 1) − w∗ k2B → θ, for some θ ≥ 0,
x(k) − x(k − 1) → 0N 2
λ(k) − λ(k − 1) → 0
mN
2)
.
u(k + 1) − u(k) → 0N Pi∈V |Ni |
w(k + 1) − w(k) → 0m Pi∈V |Ni |
Result
1
implies
that
the
sequences
{xi (k)}, {λi (k)}, {uij (k)} and {wij (k)} (similarly
{v ij (k)} and {z ij (k)}) are bounded and have limit points
denoted by x̄i , λ̄i , ūij and w̄ij (v̄ ij and z̄ ij ), respectively.
Then, we obtain,
1
1
θ = kx̄(k) − x∗ k2M2 + kλ̄(k) − λ∗ k2Q
2
2
1
1
∗ 2
+ kū(k + 1) − u k + kw̄(k + 1) − w̄∗ k2B (48)
c
c
Result 2 along with (12) and (14) yield ∀i ∈ V, j ∈ Ni ,
x̄i = x̄j := x̄, λ̄i = λ̄j := λ̄
Moreover, by (42) we arrive at,
ūij + v̄ ij = 0N
∀i ∈ V, j ∈ Ni .
c((D + A) ⊗ IN ) = c((D (2I − LN )D ) ⊗ IN ),
1
1
1
(45)
1
(a(k) − a(k − 1)) M (a(k) − a ) = ka(k) − a∗ k2M
2
1
1
2
+ ka(k) − a(k − 1)kM − ka(k − 1) − a∗ k2M .
(46)
2
2
Then (44) becomes,
n1
1
kx(k) − x∗ k2M2 + kλ(k) − λ∗ k2Q
2
2
∗
i
A x̄ − b −
X
ij
w̄ + z̄ ji = 0m .
(52)
j∈Ni
1
where L := D − H, D 2 , D− 2 and LN := D− 2 LD− 2 are
the Laplacian of GC , the square root and reciprocal square
root of D and the normalized Laplacian of GC , respectively.
1
Since D 0, D− 2 exist, it is shown in that λmax (LN ) ≤ 2.
Then (45) yields that c((D + H) ⊗ IN ) 0. This concludes
M2 0. Moreover, by Assumption 1, B 0. We use the
following inequality in (44) for every {a(k)} and M 0:
T
j∈Ni
i
1
2
(50)
Similarly, w̄ij + z̄ ij = 0m ∀i ∈ V, j ∈ Ni . Result 2 also
implies that by (49), (30) and (32):
X ij
∇i Ji (x̄) + ∂IΩi (x̄i ) +
(ūi + ṽiji ) = 0, (51)
diag((B i )i∈V ) and Q := c((D + H) ⊗ IN )B. Note that,
1
2
(49)
Comparing (49)-(52) with (9), it follows ∀i ∈ V, j ∈ Ni ,
x̄i = x∗ (x̄ = x∗ ), λ̄i = x∗ (λ̄ = λ∗ ),
ij
ij ∗
ū = u
∗
(ū = u ), w̄
ij
=w
ij ∗
(53)
∗
(w̄ = w ).
This completes the proof.
(54)
V. S IMULATION R ESULTS
In this section, we simulate our algorithm for a wireless
ad-hoc network (WANET). Consider a WANET with 16
nodes and 16 multi-hop communication links as in Fig. 1 (a).
There are 15 users who aim to transfer data from a source to
a destination. Solid line represents a link Lj , j ∈ {1, . . . , 16}
and dashed line displays a path Ri , i ∈ {1, . . . , 15} that is
assigned to user i to transfer data. Each link Lj has a positive
1
12
R7 L8 R6
R14 L9
R8
13
3
L10
L15
L11
R10
L12
9
L6
L14
R9
16
14
L4
L13
15
2
11
13
3
4
10
4
R3
5
R11
1
R4
L5
R15
L16
12
R5
11
10
VI. C ONCLUSIONS
2
L7
9
14
5
8
7
6
15 R13
8
R2
L3
L1
R12 L2 R1
6
7
Fig. 1: (a) Wireless Ad-Hoc Network (left). (b) Communication graph GC (right).
9
8
7
User Flow Rate
6
5
4
3
2
1
0
500
1000
1500
2000
Iteration
Fig. 2: Flow rates of users 1, 3, 5, 15. GNE points are
represented by black stars.
capacity Cj > 0 that restricts the users’ data flow . The data
flow of user i is denoted by xi such that 0 ≤ xi ≤ 10. For
each user i, a cost function Ji is defined as in [7]:
X
κ
P
Ji (xi , x−i ) :=
−χi log(xi +1),
Cj − w:Lj ∈Rw xw
j:Lj ∈Ri
where κ > 0 and χi > 0 are network-wide known and userspecific parameters, respectively.
The problem is to find a GNE of the game which is played
over a communication graph GC (depicted in Fig. 1 (b)). We
assume a common constraint x1 + x3 + x5 = 14 between the
players. It is straightforward to check the Assumptions 2,3
and 4 on GC and the cost functions. The users’ flow rate and
the dual Lagrange multipliers (they converge to 0) as well
as the normalized error are shown in Fig. 2 and Fig. 3 for
χi = 15 ∀i ∈ {1, . . . , 15} and Cj = 15 ∀j ∈ {1, . . . , 16}.
2
0
λi
−2
−4
−6
−8
0
500
1000
1500
2000
Iteration
Fig. 3: Dual Lagrange multipliers for users 1, 3, 5, 15.
0
Normalized Error ||x(k)−x*||/||x*||
10
−2
10
−4
10
−6
10
0
500
1000
1500
2000
Iteration
Fig. 4: Normalized error of the users’ data flow.
A distributed GNE seeking algorithm is designed within
the framework of inexact-ADMM. Each player is only aware
of his own cost function, problem data and action set of
all players. Each player exchanges the estimates of other
players’ actions along with the local copy of his Lagrange
multiplier with his communication neighbors. An inexactADMM algorithm is then derived to find a GNE of the game.
Convergence of the algorithm is then provided under mild
conditions on cost functions.
R EFERENCES
[1] M. Zhu and E. Frazzoli, “Distributed robust adaptive equilibrium
computation for generalized convex games,” Automatica, vol. 63, pp.
82–91, 2016.
[2] L. Pavel, “An extension of duality to a game-theoretic framework,”
Automatica, vol. 43, no. 2, pp. 226–237, 2007.
[3] Z. Han, Game theory in wireless and communication networks: theory,
models, and applications. Cambridge University Press, 2012.
[4] T. Alpcan and T. Başar, “Distributed algorithms for Nash equilibria
of flow control games,” in Advances in Dynamic Games. Springer,
2005, pp. 473–498.
[5] F. Facchinei and C. Kanzow, “Generalized Nash equilibrium problems,” Annals of Operations Research, vol. 175, no. 1, pp. 177–211,
2010.
[6] A. Fischer, M. Herrich, and K. Schönefeld, “Generalized Nash equilibrium problems-recent advances and challenges,” Pesquisa Operacional, vol. 34, no. 3, pp. 521–558, 2014.
[7] F. Salehisadaghiani and L. Pavel, “Distributed Nash equilibrium seeking by gossip in games on graphs,” in Decision and Control (CDC),
2016 IEEE 55th Conference on. IEEE, 2016, pp. 6111–6116.
[8] T.-H. Chang, M. Hong, and X. Wang, “Multi-agent distributed optimization via inexact consensus admm,” IEEE Transactions on Signal
Processing, vol. 63, no. 2, pp. 482–497, 2015.
[9] B. Johansson, “On distributed optimization in networked systems,”
Ph.D. Dissertation, 2008.
[10] F. Parise, B. Gentile, S. Grammatico, and J. Lygeros, “Network
aggregative games: Distributed convergence to Nash equilibria,” in
2015 54th IEEE Conference on Decision and Control (CDC). IEEE,
2015, pp. 2295–2300.
[11] F. Salehisadaghiani and L. Pavel, “Distributed Nash equilibrium seeking: A gossip-based algorithm,” Automatica, vol. 72, pp. 209–216,
2016.
[12] P. Yi and L. Pavel, “A distributed primal-dual algorithm for computation of generalized nash equilibria with shared affine coupling constraints via operator splitting methods,” arXiv preprint
arXiv:1703.05388, 2017.
[13] H. Yin, U. V. Shanbhag, and P. G. Mehta, “Nash equilibrium problems
with scaled congestion costs and shared constraints,” IEEE Transactions on Automatic Control, vol. 56, no. 7, pp. 1702–1708, 2011.
[14] D. A. Schiro, J.-S. Pang, and U. V. Shanbhag, “On the solution of
affine generalized Nash equilibrium problems with shared constraints
by lemkes method,” Mathematical Programming, pp. 1–46, 2013.
[15] D. Paccagnan, B. Gentile, F. Parise, M. Kamgarpour, and J. Lygeros,
“Distributed computation of generalized Nash equilibria in quadratic
aggregative games with affine coupling constraints,” in Decision and
Control (CDC), 2016 IEEE 55th Conference on. IEEE, 2016, pp.
6123–6128.
[16] S. Grammatico, “Aggregative control of competitive agents with coupled quadratic costs and shared constraints,” in Decision and Control
(CDC), 2016 IEEE 55th Conference on. IEEE, 2016, pp. 3597–3602.
[17] D. P. Bertsekas and J. N. Tsitsiklis, Parallel and Distributed Computation: Numerical Methods. Athena Scientific, Belmont, Massachusetts,
1997.
[18] S. Boyd, N. Parikh, E. Chu, B. Peleato, and J. Eckstein, “Distributed
optimization and statistical learning via the alternating direction
method of multipliers,” Foundations and Trends R in Machine Learning, vol. 3, no. 1, pp. 1–122, 2011.
[19] B. He and X. Yuan, “On the o(1/n) convergence rate of the douglasrachford alternating direction method,” SIAM Journal on Numerical
Analysis, vol. 50, no. 2, pp. 700–709, 2012.
[20] E. Wei and A. Ozdaglar, “Distributed alternating direction method of
multipliers,” in 2012 IEEE 51st IEEE Conference on Decision and
Control (CDC). IEEE, 2012, pp. 5445–5450.
[21] F. Salehisadaghiani and L. Pavel, “Distributed Nash equilibrium seeking via the alternating direction method of multipliers,” accepted in
IFAC World Congress, 2017.
| 3cs.SY
|
Stochastic Runtime Analysis of a Cross-Entropy
Algorithm for Traveling Salesman Problems
Zijun Wua,1,∗, Rolf H. Möhringb,2,∗∗, Jianhui Laic,
arXiv:1612.06962v2 [cs.DS] 11 Oct 2017
a
Beijing Institute for Scientific and Engineering Computing (BISEC), Pingle Yuan 100,
Beijing, P. R. China
b
Beijing Institute for Scientific and Engineering Computing (BISEC), Pingle Yuan 100,
Beijing, P. R. China
c
College of Metropolitan Transportation, Beijing University of Technology, Pingle Yuan
100, Beijing, P. R. China
Abstract
This article analyzes the stochastic runtime of a Cross-Entropy Algorithm
mimicking an Max-Min Ant System with iteration-best reinforcement. It
investigates the impact of magnitude of the sample size on the runtime to
find optimal solutions for TSP instances.
For simple TSP instances that have a {1, n}-valued distance function and
a unique optimal solution, we show that sample size N ∈ ω(ln n) results in a
stochastically polynomial runtime, and N ∈ O(ln n) results in a stochastically
exponential runtime, where “stochastically” means with a probability of 1 −
n−ω(1) , and n represents number of cities. In particular, for N ∈ ω(ln n),
we prove a stochastic runtime of O(N · n6 ) with the vertex-based random
solution generation, and a stochastic runtime of O(N · n3 ln n) with the edgebased random solution generation. These runtimes are very close to the best
known expected runtime for variants of Max-Min Ant System with best-sofar reinforcement by choosing a small N ∈ ω(ln n). They are obtained for
∗
Principal corresponding author
Corresponding author
Corresponding author
Email addresses: [email protected] (Zijun Wu), [email protected]
(Rolf H. Möhring), [email protected] (Jianhui Lai)
1
The author is also affiliated with the School of Applied Mathematics and Physics at
Beijing University of Technology
2
The author is a professor emeritus of mathematics at Berlin University of Technology
∗∗
Preprint submitted to Theoretical Computer Science
October 12, 2017
the stronger notion of stochastic runtime, and analyze the runtime in most
cases.
We also inspect more complex instances with n vertices positioned on
an m × m grid. When the n vertices span a convex polygon, we obtain
a stochastic runtime of O(n4 m5+ ) with the vertex-based random solution
generation, and a stochastic runtime of O(n3 m5+ ) for the edge-based random
solution generation. When there are k ∈ O(1) many vertices inside a convex
polygon spanned by the other n − k vertices, we obtain a stochastic runtime
of O(n4 m5+ + n6k−1 m ) with the vertex-based random solution generation,
and a stochastic runtime of O(n3 m5+ + n3k m ) with the edge-based random
solution generation. These runtimes are better than the expected runtime
for the so-called (µ+λ) EA reported in a recent article, and again obtained
for the stronger notion of stochastic runtime.
Keywords: probabilistic analysis of algorithms, stochastic runtime analysis
of evolutionary algorithms, Cross Entropy algorithm, Max-Min Ant
System, (µ+λ) EA.
1. Introduction
The Cross Entropy (CE) algorithm is a general-purpose evolutionary algorithm (EA) that has been applied successfully to many N P-hard combinatorial optimization problems, see e.g. the book [1] for an overview. It
was initially designed for rare event simulation by Rubinstein [2] in 1997,
and thereafter formulated as an optimization tool for both continuous and
discrete optimization (see [3]).
CE has much in common with the famous ant colony optimization (ACO,
see [4]) and the estimation of distribution algorithms (EDAs, see [5]). They
all belong to the so-called model-based search paradigm (MBS), see [6]. Instead of only manipulating solutions, which is very typical in traditional
heuristics like Genetic Algorithms [7] and Local Search [8] and others, MBS
algorithms attempt to optimize the solution reproducing mechanism. In each
iteration, they produce new solutions by sampling from a probabilistic distribution on the search space. The distribution is often called a model in the
literature (see e.g. [6] and [9]). This model evolves iteratively by incorporating information from some elite solutions occurring in the search history,
so as to asymptotically model the spread of optimal solutions in the search
2
space. See the recent Thesis [9] for more details on MBS algorithms and
their mathematical properties.
An important issue for MBS algorithms is to determine a suitable size
for the sampling in each iteration. A large sample size makes each iteration
unwieldy, however a small sample size may mislead the underlying search
due to the randomness in the sampling. Sample size reflects the iterative
complexity (computational complexity in each iteration). Whether a large
sample size is harmful depends on the required optimization time (i.e., the
total number of iterations required to reach an optimal solution). This article
aims to shed a light on this issue by theoretically analyzing the relation
between sample size and optimization time for a CE variant that includes
also some essential features of the famous Max-Min Ant System (MMAS
[10]). To this end, a thorough runtime analysis is needed.
The theoretical runtime analysis of EAs has gained rapidly growing interest in recent years, see e.g. [11], [12], [13], [14], [15], [16], [17], [18], [19],
[20], [21], [22], [23], [24], and [25]. In the analysis, an oracle-based view of
computation is adopted, i.e., the runtime of an algorithm is expressed as
the total number of solutions evaluated before reaching an optimal solution.
Since the presence of randomness, the runtime of an EA is often conveyed
in expectation or with high probability. Due to the famous No Free Lunch
Theorem [26], the analysis must be problem-specific. The first steps towards
this type of analysis were made for the so-called (1+1) EA [11] on some test
problems that use pseudo boolean functions as cost functions, e.g., OneMax
[15], LeadingOnes [27] and BinVar [11]. Recent research addresses problems of practical importance, such as the computing a minimum spanning
trees (MST) [28], matroid optimization [29], traveling salesman problem [30],
the shortest path problem [23], the maximum satisfiability problem [31] and
the max-cut problem [32].
Runtime analysis generally considers two cases: expected runtime analysis
and stochastic runtime analysis. Expected runtime is the average runtime of
an algorithm on a particular problem, see, e.g., the runtime results of (1 + 1)
EA reported in [11]. Expected runtime reflects the oracle-based average performance of an algorithm. A mature technique for expected runtime analysis
is the so-called drift analysis [12]. However, this technique requires that the
algorithm has a finite expected runtime for the underlying problem. By [33],
drift analysis is not applicable to the traditional CE [3].
An algorithm with a smaller expected runtime need not be more efficient,
see [34] for details. In contrast, stochastic runtime provides a better under3
standing of the performance of a (randomized) EA. Stochastic runtime is a
runtime result conveyed with an overwhelming probability guarantee (see,
e.g., the classic runtime result of 1-ANT in [15]), where an overwhelming
probability means a probability tending to 1 superpolynomially fast in the
problem size. It therefore reflects the efficiency of an algorithm for most cases
in the sense of uncertainty. This article is concerned with stochastic runtime
analysis, aiming to figure out the relation between stochastic runtime and
magnitude of the sample size.
Runtime analysis of CE algorithms has been initiated in [33], where Wu
and Kolonko proved a pioneering stochastic runtime result for the traditional
CE on the standard test problem LeadingOnes. As a continuation of the
study of [33], Wu et al [34] further investigated the stochastic runtime of
the traditional CE on another test problem OneMax. The runtime results
reported in [33] and [34] showed that sample size plays a crucial role in
efficiently finding an optimal solution. In particular, Wu et al [34] showed
that if the problem size n is moderately adapted to the sample size N , then
4
the stochastic runtime of the traditional CE on OneMax is O(n1.5+ 3 ) for
arbitrarily small > 0 and a constant smoothing parameter ρ > 0, which
beats the best-known stochastic runtime O(n2 ) reported in [13] for the classic
1-ANT algorithm, although 1-ANT employs a much smaller sample size (i.e.,
sample size equals one). Moreover, by imposing upper and lower bounds on
the sampling probabilities as was done in MMAS [10], Wu et al [34] showed
further that the stochastic runtime of the resulting CE can be significantly
improved even in a very rugged search space.
The present article continues the stochastic runtime analysis of [34], but
now in combinatorial optimization with a study of CE on the traveling salesman problem (TSP). We emphasize the impact of the magnitude of N on
the stochastic runtime, put ρ = 1, and consider a CE variant resembling an
MMAS with iteration-best reinforcement under two different random solution generation mechanisms, namely, a vertex-based random solution generation and an edge-based random solution generation.
Stochastic runtime for MMAS with iteration-best reinforcement on simple problems like OneMax has been studied in [20] and [25]. In particular,
Neumann et al [20] showed that to obtain a stochastically polynomial runtime for OneMax, N/ρ ∈ Ω(ln n) is necessary. We shall not only extend
this to TSP for the case of ρ = 1, but also prove that N ∈ ω(ln n) is already
sufficient to guarantee a stochastically polynomial runtime for simple TSP
instances.
4
TSP is a famous N P-complete combinatorial optimization problem. It
concerns finding a shortest Hamiltonian cycle on a weighted complete graph.
Existing algorithms exactly solving TSP generally have a prohibitive complexity. For instance, the Held-Karp algorithm [35] solves the problem with
a complexity of O(n2 2n ). A well-known polynomial time approximation algorithm for metric TSP is the so-called Christofides algorithm [36], which
finds a solution with a cost at most 3/2 times the cost of optimal solutions.
As mentioned in [37], this is still the best known approximation algorithm
for the general metric TSP so far. For Euclidean TSP there exists a famous
polynomial-time approximation scheme (PTAS) by Arora, see [38]. To design a superior approximation algorithm, researchers in recent years tend to
study TSP instances with particular structures, see, e.g., [39].
Due to the prohibitive running time of exact algorithms, heuristics are
frequently employed in practice so as to efficiently compute an acceptable
solution for a TSP problem, e.g., the Lin-Kernighan (LK) algorithm [40].
As a popular heuristic, CE has also been applied in practice to solve TSP
instances, see [41] and [3]. The implementation there shows that CE can also
efficiently compute an acceptable solution.
In view of the high complexity of general TSP, we consider in our analysis
two classes of TSP instances with a particular structure. The first kind of
instances has been used in [19] and [42] for analyzing the expected runtime of
some MMAS variants with best-so-far reinforcement. These TSP instances
have polynomially many objective function values and a unique optimal solution. Moreover, on these TSP instances, solutions containing more edges
from the optimal solution have a smaller cost than those with fewer such
edges. For more details on these instances, see Section 5.
For these simple instances, we prove in Theorem 2 that with a probability
1 − e−Ω(n ) , the runtime is O(n6+ ) with the vertex-based random solution
generation, and O(n3+ ln n) with the edge-based random solution generation,
for any constant ∈ (0, 1) and N ∈ Ω(n ). For the case of N ∈ ω(ln n), we
show that the runtimes (resp., O(n6 N and n3 (ln n)N ) are even smaller with
probability 1 − n−ω(1) , see Corollary 1. These results are very close to the
known expected runtime O(n6 + n lnρ n ) for (1 + 1) MMAA reported in [19],
and the expected runtime O(n3 ln n + nρln ) for MMAS∗Arb reported in [42]
(where ρ ∈ (0, 1) is an evaporation rate), if N ∈ ω(ln n) is suitably small.
But they give the stronger guarantee of achieving the optimal solution in the
respective runtime with an overwhelming probability. Moreover, we show
5
a stochastically exponential runtime for a suitable choice of N ∈ O(ln n),
see Theorem 3. This generalizes the finding in [20] for OneMax to TSP
instances. Therefore, N ∈ Ω(ln n) is necessary, and N ∈ ω(ln n) is sufficient
for a stochastically polynomial runtime for simple TSP instances.
We also inspect more complex instances with n vertices positioned on an
m×m grid, and the Euclidean distance as distance function. These instances
have been employed in [43] and [30] for analyzing the expected runtime of
(µ+λ) EA and randomized local search (RLS). When the n vertices span a
convex polygon without vertices in the interior of the polygon (so they are
the corners of that polygon), we prove a stochastic runtime of O(n4 m5+ )
for the vertex-based random solution generation, and a stochastic runtime
of O(n3 m5+ ) for the edge-based random solution generation, see Theorem
4 for details. Similarly, the in the stochastic runtimes can be removed
by slightly decreasing the probability guarantee, see Corollary 2. When the
vertices span a convex polygon with k ∈ O(1) vertices in the interior, we
show a stochastic runtime of O(n4 m5+ + n6k−1 m ) with the vertex-based
random solution generation, and a stochastic runtime of O(n3 m5+ + n3k m )
with the edge-based random solution generation, see Theorem 5 for details.
These runtimes are better than the expected runtime for the so-called (µ+λ)
EA and RLS reported in the recent paper [30].
The remainder of this paper is arranged as follows. Section 2 defines the
traditional CE and related algorithms, Section 3 defines the traveling salesman problem and provides more details of the used CE variants, Section 4
shows some important facts on the two random solution generation methods,
and Section 5 reports the stochastic runtime results on the TSP instances.
A short conclusion and suggestions for future work are given in Section 6.
Notations for runtime
Our analysis employs some commonly used notations from complexity
theory. We use O(f (n)) to denote the class of functions which are bounded
from above by the function f (n), i.e., those functions g(n) with g(n) ≤ c·f (n)
for large enough n and some constant c ≥ 0 not depending on n. Similarly,
Ω(f (n)) is the class of functions that are bounded from below by f (n), i.e.,
for any g(n) ∈ Ω(f (n)) there exists a constant c > 0 not depending on n
such that g(n) ≥ c · f (n) for large enough n. Class Θ(f (n)) is the intersection of Ω(f (n)) and O(f (n)). Class o(f (n)) is the class of functions g(n)
with g(n)/f (n) → 0 as n → ∞, and class ω(f (n)) is the class of functions
6
g(n) with g(n)/f (n) → +∞ as n → ∞. Obviously, o(f (n)) ⊂ O(f (n)) and
ω(f (n)) ⊂ Ω(f (n)).
2. The general cross entropy algorithm and related algorithms
We now introduce the traditional CE algorithm. The CE variant we will
analyze inherits the framework of this traditional version. To compare our
results with those in the literature, we shall give also details about some
related algorithms.
2.1. The traditional cross entropy algorithm
Algorithm 1 lists the traditional CE that was proposed in [3], adapted to
an abstract notion of combinatorial optimization problems. The algorithm
assumes a combinatorial minimization problem (S, f ), where S is a finite
search space of “feasible” solutions and f is the cost function. Every feasible
solution s ∈ S is composed of elements from a fixed finite set A, the ground
set of the problem, i.e., we assume S ⊆ An for some integer n ∈ N. Furthermore there is a product distribution on the product space An that induces a
distribution on S ⊆ An . The distribution on An can usually be represented as
a vector (or matrix) of real-valued probabilities. The convex combination of
the two distributions in Step 6 of Algorithm 1 then corresponds to a convex
combination of the two vectors (or matrices).
Specific to the TSP, the ground set A can be the set of nodes or edges, n
is the number of nodes, and a feasible solution is a sequence of elements from
A that forms a Hamiltonian cycle. The product distribution for the TSP is
represented as an n×n matrix.
When we consider the set of nodes as our ground set A, each row i
of the matrix is a marginal distribution that specifies choice probabilities
for all nodes following the current node i. A random Hamiltonian cycle
is sequentially constructed from the product distribution by allowing only
nodes not yet visited as continuations in each step, see Algorithm 2 for more
details.
When we consider the set of edges as A, marginals of the product distribution will be represented by the same n×n matrix where the sum of the
(i, j)-th and (j, i)-th entries reflects the probability that the edge {i, j} occurs in a random solution. A random Hamiltonian cycle is still constructed
sequentially and only edges leading to a feasible solution are taken in each
step, see Algorithm 3 for details.
7
Algorithm 1 The general Cross-Entropy algorithm
Input:
an initial distribution Π0 on the solution space, a fixed smoothing parameter ρ ∈ (0, 1], a sample size N ∈ N+ , an elite size M ∈ N+ with M ≤ N
1:
2:
3:
4:
5:
6:
7:
8:
t = 0;
loop
(1)
(N )
independently generate N random solutions Xt , . . . , Xt with the
current distribution Πt ;
[1]
sort these N solutions in non-decreasing order as f (Xt ) ≤ · · · ≤
[N ]
f (Xt ) according to the cost function f ;
learn an empirical distribution Wt from the M best solutions
[1]
[M ]
X t , . . . , Xt ;
set Πt+1 = (1 − ρ)Πt + ρWt ;
t = t + 1;
end loop
Traditionally, CE sets a small elite ratio α ∈ (0, 1) and uses the best
bα · N c solutions in Step 5 to build the empirical distribution Wt . Here, we
use the elite size M instead. This does not intrinsically change the original
algorithm. Steps 3 and 5 depend on the detailed definition of the underlying
problem. We shall give details to them in Subsection 3.2.
Step 6 of Algorithm 1 plays a crucial role in the different theoretical analyses of the algorithm, see, e.g., [44], [33], [45], [9], [34]. The occurrence of
good solutions are probabilistically enforced by incorporating the new information Wt into Πt+1 . This idea, somehow, coincides with the reinforcement
learning in [46]. The smoothing parameter ρ reflects the relative importance
of the new information Wt in the next sampling. It balances global exploration and local exploitation to a certain degree. A larger ρ makes the
algorithm concentrate more on the particular area spanned by the elite solu[1]
[M ]
tions Xt , . . . , Xt , while a smaller ρ gives more opportunities to solutions
outside that area.
However, balancing global exploration and local exploitation through tuning ρ is ultimately limited. Wu and Kolonko [33] proved that the famous
“genetic drift” [47] phenomenon also happens in this algorithmic scheme,
i.e., the sampling (Step 3) eventually freezes at a single solution and that
solution needs not to be optimal. This means that the algorithm gradually
8
loses the power of global exploration.
As a compensation for global exploration, Wu et al [34] proved that a
moderately large sample size N might be helpful. The results there showed
that a moderately large N configured with a large ρ (e.g., ρ = 1) can make the
algorithm very efficient. Although a large N introduces a high computational
burden in each iteration, the total number of iterations required for getting
an optimal solution is considerably reduced.
Wu et al [34] also indicated another way to compensate the global exploration, i.e., imposing a lower bound πmin ∈ (0, 1) and an upper bound
πmax ∈ (0, 1) on the sampling distributions in each iteration. This idea is
originated from MMAS [10]. In each iteration t, after applying Step 6, the
entries of distribution Πt+1 that are out of the range [πmin , πmax ] are reset to
that range by assigning to them the nearest bounds, see (6) in Section 3 for
more details. Wu et al [34] have proved that this can make CE more efficient
even in the case of a rugged search space.
To follow these theoretical suggestions made in [34], we shall in our
stochastic runtime analysis use a CE that modifies the traditional CE (Algorithm 1) accordingly. We shall see that these modifications make the CE
very efficient for the considered TSP instances.
2.2. Related evolutionary algorithms
Related evolutionary algorithms for TSP whose runtime has been extensively studied are RLS [28], (µ + λ) EA [30], and those theoretical abstractions of MMAS [10] including MMAS∗bs [17], (1+1) MMAA [19]. We now
give algorithmic details of them. In order to facilitate the comparison, their
runtimes for TSP instances will be discussed in Section 5.
(µ + λ) EA is an extension of the famous (1 + 1) EA [11]. (µ + λ) EA
randomly chooses µ solutions as the initial population. In each iteration,
(µ+λ) EA randomly chooses λ parents from current population, then produces
λ children by applying randomized mutation to each of the selected parents,
and forms the next population by taking the best µ solutions from these
µ + λ solutions in the end of current iteration. The expected runtime of
(µ + λ) EA on TSP instances is studied in [30], where Sutton et al uses a
Poisson distribution to determine the number of randomized mutations (2opt move or jump operation) should be taken by a selected parent in each
iteration.
RLS is a local search technique [48]. It employs a randomized neighborhood. In each iteration, it randomly chooses a number of components of the
9
best solution found so far and then changes these components. The expected
runtime of RLS for TSP instances is also studied in [30], where the neighborhood is taken to be a k-exchange neighborhood with k randomly determined
by a Poisson distribution.
(1+1) MMAA is a simplified version of the famous MMAS [10], where
the sample size is set to 1 and pheromones are updated only with the best
solution found so far (best-so-far reinforcement) in each iteration. In each
iteration of (1+1) MMAA, the ant which constructed the best solution found
so far deposits an amount πmax of pheromones on the traversed edges, and an
amount πmin of pheromones on the non-traversed edges, and the pheromones
are updated by linearly combining the old and these newly added pheromones
as in Algorithm 1. The expected runtime of (1+1) MMAA on simple TSP
instances is studied in [19]. The expected runtime of its variant MMAS∗Arb
on simple TSP instances is studied in [42].
3. The traveling salesman problem and details of the CE variant
Now, we formally define TSP, and give more details of the CE variant we
will analyze.
3.1. The traveling salesman problem
We consider an undirected
graph G = (V, E) with vertex set V = {1, . . . , n}
and edge set E = {i, j} | i ∈ V, j ∈ V, i 6= j . A Hamiltonian cycle is a
sequence {{il , il+1 } | l = 1, . . . , n} of edges such that
a) i1 = in+1 ;
b) (i1 , . . . , in ) is a permutation of {1, 2, . . . , n}.
This definition actually considers E as the ground set A. As mentioned above,
we can also put A = V and represent Hamiltonian cycles in a more compact way as permutations of V. Note that a Hamiltonian cycle corresponds
to n different permutations, whereas a permutation corresponds to a unique
Hamiltonian cycle. However, the two representations are intrinsically the
same. We shall use them interchangeably in the sequel. To facilitate our
discussion, we shall refer to a Hamiltonian cycle by just referring to one of
the n corresponding permutations, and denote by S the set of all possible
permutations. We employ the convention that two permutations are said
10
to be same iff they form the same underlying Hamiltonian cycle. The notation {k, l} ∈ s shall mean that the edge {k, l} belongs to the underlying
Hamiltonian cycle of the solution (permutation) s.
Once a distance function d : E 7→ R+ is given, the (total traveling) cost
f (s) of a feasible solution s = (i1 , i2 , . . . , in ) ∈ S is then calculated by
f (s) :=
n−1
X
d(ij , ij+1 ) + d(in , i1 ).
(1)
j=1
We denote by S ∗ ⊆ S the set of feasible solutions (Hamiltonian cycles) that
minimize the cost (1).
3.2. Details of the CE variant
The CE variant we consider in the analysis completely inherits the structure of Algorithm 1, and additionally employs a component from MMAS.
We now formalize the sampling distribution, and define Steps 3 and 5 in
more detail. As mentioned, we represent a sampling distribution (a product
distribution on An ) for the TSP by a matrix Π = (πi,j )n×n , such that
Pn
a)
j=1 πi,j = 1, for all i = 1, . . . , n,
b) πi,i = 0 for all i = 1, . . . , n,
c) πi,j = πj,i for each edge {i, j} ∈ E.
For each edge {i, j} ∈ E, πi,j reflects the probability that a Hamilton cycle
continues with vertex j when it is in vertex i. In the sequel, we write the
t
sampling distribution Πt in iteration t as (πi,j
)n×n , where the superscript t of
t
0
πi,j indicates the iteration. The initial distribution Π0 = (πi,j
)n×n is, without
1
0
0
for
loss of generality, set to be the uniform distribution, i.e., πi,j = πj,i
= n−1
all edges {i, j} ∈ E.
We shall consider two random solution generation methods, a vertex-based
random solution generation and an edge-based random solution generation.
Algorithm 2 lists the vertex-based random solution generation method. This
method uses V as the ground set A. A product distribution of An is therefore
represented as a matrix Π = (πi,j )n×n satisfying a)-c) above, i.e., each row
of Π represents a sampling distribution on A = V. Directly sampling from
Π may produce infeasible solutions from An − S. To avoid that, Algorithm
2 starts with a randomly fixed initial node, and then sequentially extends
11
a partial solution with an unvisited vertex until a complete permutation is
obtained. This method is efficient and rather popular in practice, see, e.g.,
[41] and [4]. Here, “s + (v)” means that appends a vertex v to the end of a
partial solution s.
Algorithm 2 Vertex-based random solution generation
Input:
a distribution Π = (πi,j )n×n
Output:
a permutation of 1, 2, . . . , n
1: s = ∅, and Vunivisted = V ;
2: randomly select v from V, s = s + (v), and Vunvisited = Vunvisited − {v};
3: while (|Vunvisited 6= ∅|) do
4:
select a random vertex v 0 from Vunvisited with a probability
P[v 0 | s] = P
πv,v0
k∈Vunvisited
5:
6:
7:
8:
πv,k
;
(2)
set s = s+(v 0 ), Vunvisited = Vunvisited − {v 0 };
v = v0;
end while
return s;
The edge-based random solution generation is listed in Algorithm 3. The
idea is from [42]. This method considers edge set E as the ground set A. A
feasible solution is then a sequence of edges that form a Hamiltonian cycle, i.e.
S ⊆ E n . To unify the notation of feasible solutions, Algorithm 3 translates
its outcomes into permutations. As the actual ground set is E, a product
matrix such that each row is a marginal specifying
distribution is an n× n(n−1)
2
a sampling distribution on E. Algorithm 3 only considers those with identical
marginals, a product distribution can be therefore fully characterized by
one of its marginals and is therefore again represented by an n × n matrix
Π = (πi,j )n×n as above. PAn edge
P {i, j} ∈ E is then sampled from Π with
probability (πi,j + πj,i )/ nk=1 nl=1 πk,l = 2πi,j /n since each row of Π sums
up to 1. A random sequence ∈ E n is generated by independently sampling
from Π n times. To avoid infeasible solutions, Algorithm 3 considers in
every sampling only edges that are admissible by the edges selected before.
Given a set B of edges such that the subgraph (V, B) does neither contain a
12
cycle nor a vertex of degree ≥ 3, an edge e0 ∈ E is said to be admissible by
B if and only if the subgraph (V, B ∪ {e0 }) still does neither contain a cycle
nor a vertex of degree ≥ 3. We denote by Badmissible the set of edges ∈
/ B that
are admissible by B.
Algorithm 3 Edge-based random solution generation
Input:
a distribution Π = (πi,j )n×n
Output:
a permutation of 1, 2, . . . , n
1: B = ∅, Badmissible = E;
2: while (|B| ≤ n − 1) do
3:
select an edge {i, j} from Badmissible with a probability
P[e | s] = P
4:
5:
6:
7:
8:
πi,j + πj,i
;
{k,l}∈Badmissible πk,l + πl,k
(3)
set B = B∪{{i, j}};
update Badmissible ;
end while
let s = (1, i2 , i3 , . . . , in ) with {1, i2 }, {ij , ij+1 } ∈ B for j = 2, . . . , n−1;
return s;
(1)
(N )
The N random solutions Xt , . . . , Xt in iteration t are then generated
by N runs of Algorithm 2 or Algorithm 3 with the current distribution Πt =
t
t
)n×n is then calculated from
)n×n . The empirical distribution Wt = (wi,j
(πi,j
the M elite solutions by setting
PM
({i, j})
k=1 1{e0 ∈E | e0 ∈X[k]
t
t }
wi,j
=
,
(4)
M
where 1A (·) is the indicator function of set A = {e0 ∈ E | e0 ∈ Xt } for each
t+1
{i, j} ∈ E. The next distribution Πt+1 = (πi,j
)n×n is therefore obtained as
[k]
t+1
t
t
πi,j
= (1 − ρ)πi,j
+ ρwi,j
(5)
for each {i, j} ∈ E.
We continue with the suggestions made in [34]. In the CE variant, we
shall use a moderately large N and a large ρ = 1. To fully use the best elite
13
solutions, we take M = 1. To prevent premature convergence (i.e., a possible
stagnation at a non-optimal solution), we employ a feature from MMAS
[10], called max-min calibration, in the construction of Πt+1 ,. We choose
a lower bound πmin ∈ (0, 1) and an upper bound πmax ∈ (0, 1), and, after
applying (5), adjust Πt+1 by
t+1
πmin if πi,j < πmin ,
t+1
t+1
t+1
πi,j
= πi,j
(6)
if πi,j
∈ [πmin , πmax ],
t+1
πmax if πi,j
> πmax ,
for any edge {i, j} ∈ E. Note that the max-min calibration is the only step
that does not occur in the general CE (i.e., Algorithm 1).
This setting turns CE into an MMAS with iteration-best reinforcement,
[1]
i.e., only the iteration-best solution Xt is allowed to change the ‘pheromones’
Πt . Stützle and Hoos [10] indicated in an empirical study that the practical performance of iteration-best reinforcement is comparable to best-so-far
reinforcement for TSP instances. Thus, it should also be worthwhile to compare the theoretical runtime of iteration-best reinforcement with the known
expected runtimes of best-so-far reinforcement for TSP instances presented
in, e.g., [19] and [42].
4. Properties of the random solution generation methods
Before we start with our runtime analysis, we shall discuss some relevant
properties of the two random solution generation methods, which concern the
probability of producing a k-exchange move of the iteration-best solution in
the next sampling.
Formally, a k-exchange move on a Hamiltonian cycle is an operation that
removes k edges from the cycle and adds k new edges to obtain again a cycle.
A k-opt move is a k-exchange move reducing the total travel cost. Figure
1a shows an example of a 2-exchange move, in which edges {i, j}, {k, l} are
removed, and edges {i, l}, {k, j} are added.Figure 1b shows an example of a
3-exchange move.
In our analysis, we shall consider only iteration-best reinforcement with
ρ = 1 and the max-min calibration (6). The empirical distribution Wt =
t
(wi,j
)n×n for each iteration t ∈ N in this particular case therefore satisfies
(
[1]
min{1, πmax } = πmax if edge {i, j} ∈ Xt ,
t+1
t+1
πi,j
= πj,i
=
(7)
max{0, πmin } = πmin otherwise,
14
i
k
i
l
w
v
k
j
l
j
(a) A 2-exchange move
(b) A 3-exchange move
Figure 1: Examples for edge exchange moves
for every edge {i, j} ∈ E and iteration t ∈ N. Furthermore, Πt+1 = Wt .
[1]
Since Πt+1 is biased towards the iteration-best solution Xt , k-exchanges
[1]
of Xt with a large k are unlikely to happen among the N draws from Πt+1
by either of the two generation methods. Thus, an optimal solution is more
likely to be reached by a sequence of repeatedly k-exchange moves with small
k from iteration-best solutions. Therefore, it is necessary to estimate the
[1]
probabilities of producing a k-exchange of Xt in the two generation methods,
especially for the case of small k.
4.1. Probabilities of producing k-exchanges in the vertex-based random solution generation
The probability of producing k-exchanges with k = 2, 3 in the vertexbased random solution generation has been studied in Zhou [19]. With πmin =
1
and πmax = 1 − n1 , Zhou [19] proved for (1 + 1) MMAA that with a
n2
probability of Ω(1/n5 ), Algorithm 2 produces a random solution having more
edges from s∗ than x∗t (the best solution found so far) provided that x∗t is not
optimal. Zhou [19] actually showed that if x∗t 6= s∗ , then there exists either
a 2-opt move or a 3-opt move for x∗t , and Algorithm 2 produces an arbitrary
2-exchange of x∗t with a probability of Ω(1/n3 ), and an arbitrary 3-exchange
of x∗t with a probability of Ω(1/n5 ).
1
Although we use πmin = n(n−2)
and consider iteration-best reinforcement,
a similar result holds in our case. Claim 1 below gives a lower bound on the
probability of producing a k-exchange move of the iteration-best solution in
the next round with the vertex-based random solution generation.
15
[1]
Claim 1. Let M = 1, ρ = 1, and consider a k-exchange move of Xt for some
integer k = 2, 3, . . . , n. Then, Algorithm 2 produces the given k-exchange
move with a probability Ω(1/n2k−1 ) in every of the N draws in iteration t + 1.
Proof. Recall that in Algorithm 2, the probability (2) to select a continuing
t
t
) for
(or, equivalently, πj,i
edge {i, j} is always bounded from below by πi,j
each iteration t ∈ N, since each row of Πt sums up to 1. Given a k-exchange
[1]
move of Xt , one possibility to generate it from Πt+1 = Wt by Algorithm 2
is that one of the new edges is added in the last step. This happens with a
n−k
1 k−1
1
· 1 − n1
≥ e·n2k−1
probability at least n1 · n(n−2)
, where e ≈ 2.71828
1
is Euler’s number, n represents the probability to select the starting vertex,
1
is the common lower bound of the probability to select the remaining
n(n−2)
k − 1 new edges, and 1 − n1 is the common lower bound of the probability to
[1]
select one of the remaining n − k edges from Xt .
[1]
Because of Claim 1, every 2-exchange of Xt is produced from Πt+1 by
Algorithm 2 with a probability Ω(1/n3 ), and every 3-exchange is produced
by Algorithm 2 with a probability Ω(1/n5 ). Note that for any k = 2, 3, . . . ,
[1]
if a k-opt move of Xt occurs among the N draws in the next sampling,
[1]
[1]
then f (Xt+1 ) < f (Xt ) must hold. Thus, if we take a moderately large
sample size, say N = Θ(n5+ ) for some > 0, with a probability 1 − (1 −
[1]
[1]
Ω(n−5 ))Ω(5+) = 1 − e−Ω(n ) , f (Xt+1 ) < f (Xt ) will hold, provided that there
[1]
still exists a 2-opt or 3-opt move from Xt .
Claim 2. Suppose that M = 1, ρ = 1. Then, for iteration t + 1, the probability
[1]
that Algorithm 2 produces a solution with a cost not larger than Xt in one
application is in Ω(1).
[1]
Proof. Observe that the probability that Xt is reproduced in one application
of Algorithm 2 is larger than (1 − 1/n)n−1 ∈ Ω(1), which implies that the
[1]
cost of the generated random solution is not larger than f (Xt ).
[1]
Note that if Xt is reproduced at least once among the N draws in the
[1]
next sampling, then f (Xt+1 )[1] ≤ f (Xt ). Thus, if the sample size N ∈
[1]
[1]
Ω(ln n), then f (Xt+1 ) ≤ f (Xt ) with a probability 1 − (1 − Ω(1))N = 1 −
[1]
[1]
O(1/n). Particularly, when N ∈ Ω(n ) for some > 0, f (Xt+1 ) ≤ f (Xt )
with an overwhelming probability 1 − e−Ω(n ) .
16
4.2. Probabilities of producing k-exchanges in the edge-based random solution
generation
The behavior of the edge-based random solution generation is comprehensively studied in [42]. Kötzing et al [42] proved for MMAS∗Arb and a constant
k ∈ O(1) that, with a probability of Ω(1), Algorithm 3 produces a random
solution that is obtained by a k-exchange move from the best solution found
so far.
t
t
t
t
Recall that in each iteration t, either πi,j
= πj,i
= πmin or πi,j
= πj,i
= πmax
for any edge {i, j} ∈ E. For convenience, we will call an edge {i, j} ∈ E with
t
t
= πmax a high edge, and otherwise a low edge. Kötzing et al [42]
= πj,i
πi,j
showed the probability of the event that Algorithm 3 chooses
√ a high edge in
an arbitrary fixed step conditioned on the event that l ≤ n low edges have
been chosen in some l steps before this step is 1 − O(1/n). Our setting is
1
only slightly different with from theirs, i.e., we use πmin = n(n−2)
but they
1
put πmin = n(n−1) . Thus, the result should also hold here. Claim 3 below
formally asserts this, readers may also refer to [42] for a similar proof.
Claim 3. Assume M = 1, ρ = 1. Then, the probability of choosing√a high
edge at any fixed step in Algorithm 3 is at least 1 − 12/n if at most n low
edges have been chosen before that step and there exists at least one high
admissible edge to be added.
Proof.
√ We now fix a step n−m for some m = 0, 1, . . . , n−1, and assume that
l ≤ n low edges have been chosen before this step. Obviously, we still need
to add m + 1 ≥ 1 edges to obtain a complete solution. We now estimate the
numbers of admissible high and low edges in this step. Note that every of
the l low edges blocks at most 3 of the m+l remaining high edges (at most
two which are incident to the end points of the low edge, and at most one
that may introduce a cycle). So at least m+l − 3l = m−2l ≥ m−3l high
edges are available for adding in this step. Of course, it may happen that
there is no admissible high edges in this step. However, we are not interested
in such a case. We consider only the case that there exists at least one
admissible high edge in this step, i.e. the number of admissible high edges
in this step is at least max{1, m − 3l}. Note also that the n−m edges added
before partition the subgraph of G = (V, E) with vertices V and edges from
the partial solution constructed so far into exactly m connected components
(here, we see an isolated vertex also as a connected component). For any two
of the components, there are at most 4 admissible edges connecting them.
17
Therefore,
there are at most min{4 m2 , n2 } admissible low edges. Observing
√
l ≤ n, the probability of choosing a high edge in this step is bounded from
below by
(
√
2m2
3
min{4 m2 , n2 } πmin
1− (m−3l)n(n−2)
≥ 1− (n−2)
if m > 3 n,
≥
(8)
1−
√
12
max{1, m − 3l} πmax
if m ≤ 3 n,
1− n−2
where the first inequality is obtained by observing that
m
n
n
2
min{4
,
} ≤ min{2m ,
} ≤ 2m2 ,
2
2
2
τmax = 1 − 1/n, πmin =
1
,
n(n−2)
and
2m2
m−3l
≤
2
√
1
− 3 2n
m
m
≤ 3n.
With Claim 3, we can show that, for any t ∈ N and any fixed k ∈ O(1),
[1]
the probability of the event that a k-exchange of Xt is produced by one
application of Algorithm 3 is Ω(1), see Claim 4. Here, we shall use a different
proof from the one presented by Kötzing et al [42], which appears to us as
problematic.
Claim 4. Let M = 1, ρ = 1. For any k ∈ O(1), with probability Ω(1), the
[1]
random solution produced by Algorithm 3 is a k-exchange of Xt .
Proof. Let k ∈ O(1) be arbitrarily fixed, and M be the set of all k-element
subsets of {1, 2, . . . , n/2} (where we assume without loss of generality that
n is even). Obviously, |M| ∈ Θ(nk ) since k ∈ O(1). Let M ∈ M be an arbitrarily fixed k-element subset. The probability of the event that Algorithm 3
selects k new edges (low edges) at steps i ∈ M and n − k edges (high edges)
[1]
from Xt at other steps, is bounded from below by
n−i+1
− (n − i + k + 1))πmin
1
1 n−k Y ( 2
≥ Θ( k ), (9)
1 − O( )
n
n(n − 1)πmin + (n − i + k + 1)πmax
n
i∈M
where 1 − O(1/n) is a lower bound for the probability of selecting an edge
[1]
from Xt . In each step i ∈ M, the edges chosen before partition the graph
into n−i+1 connected components, and for any two of the components there
exists at least 2 edges connecting
them without introducing a cycle. Hence,
n−i+1
there are at least
admissible edges in each step i ∈ M. Notice also
2
that the number of admissible high edges in this case is at most n − i + k + 1
18
(n − i+k+1 is the maximal number of high edges that have not been chosen
((n−i+1
)−(n−i+k+1))πmin
2
of (9) is just the lower
before). Therefore, each factor n(n−1)π
min +(n−i+k+1)πmax
bound of the probability for choosing an admissible edge not belonging to
[1]
Xt in a step i ∈ M.
As a result, the probability of the random event that Algorithm 3 pro[1]
duces a k-exchange of Xt with k ∈ O(1) in any of the N independent draws
in iteration t+1 is bounded from below by |M|·Θ( n1k ) = Θ(nk )·Θ( n1k ) ∈ Ω(1),
since new edges can also be added in steps l ≥ n/2.
Notice that in the edge-based random solution generation, for any k =
[1]
2, 3, . . . , n, any two k-exchanges of Xt are generated with the same probability, since the generation does not require adding the edges in a particular
order. Therefore, by Claim 4, for any k ∈ O(1), any specified k-exchange of
[1]
[1]
Xt will be produced with a probability Θ(1/nk ). Since reproducing Xt can
[1]
be seen as a 0-exchange of Xt , we can thus derive the following conclusion.
Claim 5. Let M = 1, ρ = 1. With probability Ω(1), the random solution
[1]
generated by Algorithm 3 has a cost not larger than that of Xt .
Claim 6 shows that it is unlikely that the random solution generated by
[1]
Algorithm 3 is “very” different from the last iteration-best solution Xt . This
will be fundamental for deriving the runtime lower bound.
Claim 6. Let M = 1, ρ = 1. For any δ ∈ (0, 1], with an overwhelming
min{δ,1/4}/2 )
probability 1 − e−ω(n
, the random solution generated by Algorithm 3
[1]
is a k-exchange move from Xt for some k < nδ .
Proof. Let δ ∈ (0, 1] be arbitrarily fixed, and put γ = min{δ, 1/4}. To prove
the claim, we just need to show that with an overwhelming probability, the
[1]
random solution generated by Algorithm 3 is a k-exchange of Xt for some
k ≤ nγ ≤ n1/4 . This is again implied by the fact that with an overwhelming
γ
probability, at most nγ/2 low edges are chosen within the first T := n − 3n4
γ
steps in Algorithm 3, since the best case nγ/2 + 3n4 is still smaller than nγ .
By Claim 3, for any k ≤ nγ/2 and any m ≤ T, Algorithm 3 chooses
high edges with a probability at least 1 − 12/n at step m if at most k edges
have been chosen before step m, since there exist at least n − m − 3k ≥
3nγ
− 3nγ/2 ≥ 3 admissible high edges at step m.
4
19
Let P denote the probability of the random event that at most nγ/2 low
edges are chosen within T steps, and Q the probability of the random event
that at least nγ/2 + 1 low edges are chosen within the same T steps. Then
P = 1 − Q. We shall bound Q from above, which will give a lower bound for
P.
Let E be the random event that at least nγ/2 + 1 low edges are chosen
within T steps. Then Q = P[E]. For each l = 1, . . . , nγ/2 + 1, we define
a random variable vl denoting the first step m ≤ T such that l low edges
are chosen within m steps. Obviously, E implies the random event E1 that
v1 < v2 < · · · < vnγ/2 +1 ≤ T. Thus, Q ≤ P[E1 ], and P ≥ 1 − P[E1 ].
Observe that
X
P[E1 ] =
P[v1 = a1 , . . . , vnγ/2 +1 = anγ/2 +1 ],
a1 <a2 <···<anγ/2 +1 ≤T
and v1 = a1 , . . . , vnγ/2 +1 = anγ/2 +1 is equivalent to the random event that
before step a1 only high edges are chosen, that at any step between al and
al+1 only high edges are chosen for any l with 1 ≤ l ≤ nγ/2 , and that at steps
a1 , . . . , anγ/2 +1 only low edges are chosen. Thus, we have by Claim 3 that
P[v1 = a1 , . . . , vnγ/2 +1 = anγ/2 +1 ] ≤
12 nγ/2 +1
,
n
since at each step al , there exists at least one admissible high edge and we
do not care about what happens
after step vnγ/2 +1 .
There are at most nγ/2T +1 different combinations for a1 < a2 < · · · <
nγ/2 +1
.
anγ/2 +1 . Therefore, P ≥ 1 − P[E1 ] ≥ 1 − nγ/2T +1 12
n
γ/1
By Stirling’s formula, and observing that n + 1 ∈ o(T ), T ∈ Θ(n), we
nγ/2 +1
γ/2
γ/2
have nγ/2T +1 12
= e−ω(n ) . Hence, P ≥ 1 − e−ω(n ) is overwhelmn
ingly large.
5. Main results
We shall now analyze the stochastic runtime of our two different random
solution generation methods for two classes of TSP instances that have been
well studied in the literature.
20
5.1. Stochastic runtime analysis for simple instances
We first consider a class of simple TSP instances that is defined by the
following distance function d : E → R on a graph with n vertices.
1 if {i, j} = {i, i + 1} for each i = 1, 2, . . . , n − 1,
d({i, j}) = 1 if {i, j} = {n, 1},
(10)
n otherwise.
Obviously, TSP instances with this distance function have a unique optimal
solution s∗ = (1, 2, . . . , n) (in the sense of the underlying Hamiltonian cycle),
and s∗ has a cost of n. The cost of an arbitrary feasible solution s equals
k + (n − k) · n, where k is the number of edges ∈ s that are also in s∗ . We
shall refer to these instances as G1 in the sequel.
The class G1 has been used in [19] and [42] for analyzing the expected
runtime of variants of MMAS. Zhou [19] proved that the (1 + 1) MMAA
algorithm has an expected runtime of O(n6 + n lnρ n ) on G1 in the case of nonvisibility (i.e., without the greedy distance information in the sampling), and
has an expected runtime of O(n5 + n lnρ n ) in the case of visibility (i.e., with
considering the greedy distance information in the sampling). Kötzing et al
[42] continued the study in [19]. They investigated the expected runtime of
(1 + 1) MMAA and its variant MMAS∗Arb on G1 and other TSP instances
on which both (1 + 1) MMAA and MMAS∗Arb have exponential expected
runtime. MMAS∗Arb differs with (1 + 1) MMAA only in the random solution
generation. MMAS∗Arb uses Algorithm 3 as its random solution generation
method, while (1 + 1) MMAA used Algorithm 2. Kötzing et al [42] proved
that MMAS∗Arb has an expected runtime of O(n3 ln n + n lnρ n ) on G1 .
Theorem 1 shows a stochastic runtime of O(n6+ ) for the CE variant with
the add-on, i.e., Algorithm 1 with max-min calibration (6), the vertex-based
random solution generation, and a stochastic runtime of O(n4+ ) for the
edge-based random solution generation. These results are comparable with
the above known expected runtimes. Although we are not able to get strictly
superior runtimes, our results are actually stronger and more informative.
Theorem 1 (Stochastic runtime of Algorithm 1 with max-min calibration
on G1 ). Assume that we set M = 1, ρ = 1, and use Algorithm 1 with the
1
, πmax = 1 − n1 . Then
max-min calibration (6) for the values πmin = n(n−2)
a) if we use the vertex-based random solution generation method (Algorithm 2), and take a sample size N ∈ Ω(n5+ ) for any constant ∈
21
5
(0, 1), then with a probability at least 1 − e−Ω(N/n ) the optimal solution
s∗ can be found within n iterations;
b) if we use the edge-based random solution generation method (Algorithm
3), and take a sample size N ∈ Ω(n3+ ) for a constant ∈ (0, 1), then
3
with a probability at least 1 − e−Ω(N/n ) , the optimal solution can be
found within n iterations.
Proof. We prove the Theorem by showing that the probability of the random
event that before the optimal solution is met, the number of edges shared by
the iteration-best and optimal solution strictly increases is overwhelmingly
large. This implies that the optimal solution is found within n iterations,
since the optimal solution has only n edges. Furthermore, the runtimes
presented in the Theorem hold. We only discuss the case of a), b) follows
with an almost identical argument.
[1]
By [19] (see also proof of Theorem 2), if Xt is not optimal, it has at
least either a 2-opt move or a 3-opt move. Note that for G1 , any k-opt
move of the iteration-best solution increases the number of its edges shared
with the optimal solution. By Claim 1, any 2-opt move is generated by
Algorithm 2 with probability Ω(n−3 ), and any 3-opt move is generated with
[1]
[1]
probability Ω(n−5 ). Thus, if Xt is not optimal, Xt+1 shares more edges with
[1]
the optimal solution than Xt with a probability at least 1 − (1 − n−5 )N =
5
1−e−Ω(N/n ) ∈ 1−e−Ω(n ) if N ∈ Ω(n5+ ) for any > 0. Thus, this repeatedly
happens within polynomially many number of iterations with overwhelming
5
probability 1 − e−Ω(N/n ) . This completes the proof.
The stochastic runtimes of Theorem 1 are derived for a relatively large
sample size, namely N = Ω(n5+ ) and N = Ω(n3+ ). Actually, Theorem 1
may still hold for a smaller sample size. Theorem 2 partially asserts this.
It states that the total number of iterations required to reach the optimal
solution for both generation schemes may increase considerably if a smaller
sample size is used. However, the stochastic runtime does not increase. Interestingly, one can obtain a smaller stochastic runtime with a small sample
size for the edge-based random solution generation.
Theorem 2 (Stochastic runtime of Algorithm 1 on G1 for a small sample
size). Assume the conditions in Theorem 1, but set N ∈ Ω(n ) for any ∈
(0, 1). Then:
22
a) For the vertex-based random solution generation, Algorithm 1 finds the
optimal solution s∗ within n6 iterations with a probability of 1 − e−Ω(N ) .
b) For the edge-based random solution generation, Algorithm 1 finds the
optimal solution s∗ within n3 ln n iterations with a probability of 1 −
e−Ω(N ) .
Proof of Theorem 2. The proof shares a similar idea with that of Theorem 1.
However, we consider here the random event that the number of edges shared
by the iteration-best and optimal solution does not decrease and strictly
increases enough times within a specified polynomial number of iterations.
For a), we shall consider the first n6 iterations. By Claim 2, the number
of edges shared by the iteration-best and optimal solution does not decrease
N
with a probability 1 − 1 − Ω(1)
= 1 − e−Ω(N ) (N ∈ Ω(n )). Therefore,
the number does not decrease within the first n6 iterations with probability
Qn6
−Ω(N )
) = 1−e−Ω(N ) . By Claim 1, for every consecutive n5 iterations,
t=0 (1−e
if the starting iteration-best solution is not optimal, then with probability
n5
1 − (1 − n−5 )N
= 1 − e−Ω(N ) , the number will strictly increase at least
once within these n5 iterations. Therefore, with overwhelming probability
1 − e−Ω(N ) , the optimal solution will be reached within the period of the first
n6 iterations, since there are n many consecutive n5 iterations within that
period.
b) can be proved by a similar way with a). We shall consider the first
3
n ln n iterations. By Claim 4, with probability 1 − (1 − Ω(1))N = 1 − e−Ω(N ) ,
the number of shared edges does not decrease in consecutive two iterations.
To complete the proof, we need an extra fact on 2, 3-exchanges.
Kötzing et al [42] showed for MMAS∗Arb that if the best solution s∗t found
so far has n−k edges from the optimal solution s∗ , then the probability of the
event that s∗t+1 has at least n−k+1 edges from s∗ , is in Ω(k/n3 ). We shall use a
different but simpler proof to show that this alsoTholds in our case of iteration[1]
best reinforcement. And with this fact, if |Xt
s∗ | = n−k for some 0 < k ≤
3
n, then with probability 1 − ((1 − k · n−3 )N )n /k = 1 − e−Ω(N ) , the number
of edges shared by the iteration-best solution and s∗ will strictly increase
at least once within the period [t, t + n3 /k]. This implies that s∗ is sampled
within the first n3 ln n iterations with overwhelming probability 1 − e−Ω(N ) ,
since n3 ln n iterations can be partitioned into n many consecutive phases
[0, n2 ), [n2 , n2 +n3 /(n−1)), [n2 +n3 /(n−1), n2 +n3 /(n−1)+n3 /(n−2)), . . . .
We now prove that fact.
23
[1]
We first show that when |Xt ∩ s∗ | = n − k with k > 0, then there exists a
[1]
2-opt move or a 3-opt move for Xt (see also [19] for a similar proof). Assume
[1]
that Xt contains exactly n − k edges from s∗ for some integer k > 0. Let
[1]
e∗ = {i, i + 1} be an edge in s∗ but not in Xt . Note that each node of the
[1]
graph is exactly incident to two edges of s∗ and Xt , respectively. Therefore
[1]
[1]
there exists an edge e0 ∈ Xt incident to i, an edge e00 ∈ Xt incident to i+1,
and e0 , e00 are not in s∗ . Figure 2 shows an example, where e0 is either {i, u} or
{i, v}, and e00 is either {i+1, w} or {i+1, y}. If e0 = {i, u} and e00 = {i+1, w} or
u
i
e
v
i+1
∗
w
y
e1
[1]
Figure 2: Demonstration of adding a new edge. The solid edges represent the cycle Xt .
[1]
if e0 = {i, v} and e00 = {i+1, y}, then there exists a 2-opt move of Xt which
removes e0 , e00 of distance n and adds e∗ and another edge (either {u, w} or
{v, y}) of distance at most n + 1 together. If e0 = {i, u}, e00 = {i+1, y} or
[1]
e0 = {i, v}, e00 = {i+1, w}, there is a 3-opt move of Xt which removes e0 , e00 ,
and an edge e1 ∈
/ s∗ , and adds edge e∗ and another two edges, this replacing
3 edges of distance n by 3 edges of distance at most 2n+1 together. Here,
[1]
[1]
observe the fact that adding e∗ to Xt and removing e0 , e00 from Xt results
[1]
in graph containing a cycle, and there must be an edge e1 ∈ Xt on that cycle
that does not belong to s∗ . We choose this edge as the edge e1 . Therefore,
[1]
for each e∗ of the k remaining edges in s∗ that are not in Xt , there exists a
[1]
2-opt or 3-opt move of Xt that adds e∗ .
By Claim 4, for any l ∈ O(1), the probability of producing an l-exchange
[1]
of the iteration-best solution Xt by Algorithm 3 in iteration t + 1 is Ω(1).
Since any two l-exchanges are produced with the same probability, the probability of producing a particular l-exchange in iteration t + 1 is Ω(1/nl ). As
[1]
a result, Algorithm 3 produces for each edge e∗ ∈ s∗ − Xt a 2-opt or 3-opt
[1]
move of Xt that adds edge e∗ with probability at least Ω(1/n3 ).
24
Note that the generation of a 2-exchange (or a 3-exchange) with two newly
added edges e2 , e3 by Algorithm 3 includes two mutually exclusive cases (3!
cases for a 3-exchange): e2 is chosen before e3 , or e3 is chosen before e2 . It
is not difficult to see that these two cases (3! cases for 3-exchange) have the
same probability. Therefore, the probability of the event that Algorithm 3
[1]
generates a 2-opt or 3-opt move of Xt that e∗ as one of the newly added edges
and selects e∗ before the other newly added edges, is bounded from below
[1]
by Ω(1/3!n3 ) = Ω(1/n3 ). Since Xt has k such e∗ and the corresponding k
[1]
events are also mutually exclusive, we obtain that the probability that Xt+1
[1]
[1]
has more edges from s∗ than Xt if Xt has exactly n − k edges from s∗ for
a constant k > 0 is Ω(k/n3 )
Corollary 1 further improves the stochastic runtime for an even smaller
sample size. It can be proved by an argument similar to the proof of Theorem
l
1, where we observe that (1 − (1 − p(n))ω(ln n) )n = 1 − n−ω(1) for any constant
l > 0 and probability p(n) ∈ Ω(1), and that 1 − e−ω(ln n) = 1 − n−ω(1) .
Corollary 1. Assume the conditions in Theorem 1, but let N ∈ ω(ln n).
Then:
a) For the vertex-based random solution generation, Algorithm 1 finds the
optimal solution s∗ within n6 iterations with a probability of 1 − n−ω(1) .
Particularly, if N = (ln n)2 , the runtime is n6 (ln n)2 with probability
1 − n−ω(1) .
b) For the edge-based random solution generation, Algorithm 1 finds the
optimal solution s∗ within n3 ln n iterations with a probability of 1 −
n−ω(1) . Particularly, if N = (ln n)2 , the runtime is n3 (ln n)3 with probability 1 − n−ω(1) .
Theorem 2 tells that, for any ∈ (0, 1), a sample size of N ∈ Θ(n ) is
already sufficient for iteration-best reinforcement to efficiently find an optimal
solution of simple TSP instances with an overwhelming probability. Corollary
1 further shows that N ∈ ω(ln n) even leads to a better runtime with a slightly
smaller but still overwhelming probability. Theorem 3 below shows that with
an overwhelming probability, the runtime of iteration-best reinforcement will
be exponential if N ∈ O(ln n), even if the instances are as simple as those in
G1 .
25
1
ln n.
Theorem 3. Assume the conditions of Theorem 1, but set N < 220
1/200
−Ω(n
)
Then, with probability 1 − e
, Algorithm 1 with edge-based solution
1/300 )
generation does not find the optimal solution s∗ within eΘ(n
iterations.
Proof. We prove the Theorem by inspecting the probability of the random
event that, before the optimal solution is found, the cost of the iteration[1]
best solution Xt will oscillate for exponentially many iterations with an
overwhelming probability. We shall consider this in the last stages of the
optimization process.
Let T0 be the first iteration which samples a solution containing at least
n − n1/4 + n1/5 edges from the optimal solution. We show that with an
overwhelming probability, the number of common edges in the iteration-best
and optimal solution will drop below n − n1/4 + n1/5 and the optimal solution
is not sampled before that. This will imply the conclusion of Theorem 3,
since, with an overwhelming probability, this phenomenon can repeatedly
occur exponentially many times before optimal solution is found.
To that end, we need to show the following:
[1]
1) For any 1/4 > δ > 0, if Xt contains at least n − nδ edges from the
optimal solution, then with a probability O( √1n ), the random solution
generated by Algorithm 3 will contain more edges from the optimal
[1]
solution than Xt in iteration t + 1;
[1]
2) For any 1/4 > δ > 0, if Xt contains at least n − nδ edges from
the optimal solution, then with a probability Ω(1) (at least e−5 ), the
random solution generated by Algorithm 3 will contain fewer edges
[1]
from the optimal solution than Xt in iteration t + 1.
However, we first use these two facts and show them afterwards.
1/20
[1]
By Claim 6, with probability 1−e−ω(n ) , XT0 contains at most n−n1/4 +
n1/5 + n1/10 edges from the optimal solution, since the random event that
the number of common edges from the iteration-best and optimal solution
increases more than n1/10 in one iteration implies an occurrence of a Ω(n1/10 )1/200 )
exchange. Similarly, by Claim 6 again, with probability 1 − e−ω(n
, the
iteration-best solution contains k ∈ [n − n1/4 + n1/5 − n1/6+1/100 , n − n1/4 +
n1/5 + n1/10 + n1/6+1/100 ] edges from the optimal solution in each iteration
t ∈ [T0 , T0 + n1/6 ]. This means that the optimal solution is not found in the
period [T0 , T0 + n1/6 ] with an overwhelming probability. With the help of
1) and 2), we are now to show that within this period, the number of edges
26
shared by the iteration-best and optimal solution is significantly reduced with
an overwhelming probability. This will complete the proof.
To facilitate our discussion, we call an iteration a successful iteration
if its iteration-best solution contains more edges from the optimal solution
than the last iteration-best solution, and an iteration a failure iteration if its
iteration-best solution contains fewer edges from the optimal solution than
the last iteration-best solution.
By 1) and the subsequent discussion, the expected number of success1
n
), since N < 220
ln n. Thus,
ful iterations within [T0 , T0 + n1/6 ] is O( nln1/3
1/6
−Ω(n )
by the Chernoff bound, with probability 1 − e
, at most n1/100 successful iterations can occur within [T0 , T0 + n1/6 ]. By 2) and the subsequent discussion, the expected number of failure iterations in [T0 , T0 + n1/6 ]
1
1
1
ln n. By the Chernoff bound, it happens that
is Ω(n 6 − 44 ), since N < 220
−Ω(n1/6 )
with probability 1 − e
, at least n1/7 failure iterations will occur in
[T0 , T0 + n1/6 ]. Since a successful iteration can add at most n1/100 edges from
1/200 )
the optimal solution with probability 1 − e−ω(n
, it totally adds at most
1/100
1/100
1/50
n
×n
=n
edges from the optimal solution to the iteration-best
1/200 )
solution within [T0 , T0 +n1/6 ] with probability 1−e−ω(n
. Note that within
1/6
−Ω(n1/6 )
1/7
[T0 , T0 + n ], with probability 1 − e
, at least n × 1 = n1/7 “good”
edges are removed from the iteration-best solution. Therefore, with over1/200 )
[1]
whelming probability 1 − e−Ω(n
, XT0 +n1/6 will contain at most
n − n1/4 + n1/5 + n1/10 − n1/7 + n1/50 < n − n1/4 + n1/5
[1]
edges from the optimal solution, since XT0 contains at most n − n1/4 + n1/5 +
1/20
n1/10 iterations with probability 1 − e−Ω(n ) . As a result, with probability
1/200 )
1−e−Ω(n
, the number of common edges in the iteration-best and optimal
solution will again be smaller than n − n1/4 + n1/5 in some iteration after T0 ,
and the optimal solution is not found before that. And this will repeatedly
1/300 )
1/200 )
happen eΘ(n
times with probability 1 − e−Ω(n
.
To finish the proof, we now formallyprove 1) and 2). We first consider 2).
n
By taking
√ k = 2 and considering the 2 2-exchanges that happen in the first
n − 3 n steps in the proof of Claim 4, one can show a tighter probability
[1]
lower bound e15 for producing 2-exchanges of Xt by Algorithm 3. Here, √
we
observe that the probability of choosing a high edge at a step before n − 3 n
is at least 1 − 3/(n − 2), see the proof of Claim 3.
Note that if 2-exchanges deleting 2 edges from the optimal solution happen N times in an iteration, then the iteration will be a failure iteration.
27
By the above and the fact that any two k-exchanges happen with the same
probability, a failure iteration then occurs with a probability at least
1
e5
n−nδ
2
n
2
!N
≥
1
e5
n−nδ
2
n
2
1
! 220
ln n
∈ Ω(n−1/44 ),
1
where δ ∈ (0, 1/4) and N < 220
ln n. This asserts 2).
[1]
1) follows with a similar discussion. Since Xt is assumed to contain
at least n − nδ edges from the optimal solution for some δ ∈ (0, 1/4), and
since Ω(nδ )-exchanges happen with an overwhelmingly small probability, we
need to consider only O(nδ )-exchanges when we estimate the probability of a
successful iteration. For each k ∈ Ω(nδ ), the proportion of failure k-exchanges
is bounded from below by
n−nδ
2knδ
−1/2
k
= e− n + o(1) ≥ e−2n
+ o(1),
n
k
since 0 < δ < 1/4, and k-exchanges removing k edges shared by the iterationbest and optimal solution are not “successful” k-exchanges. Since for any
k ∈ Ω(nδ ), any two k-exchanges happen with the same probability, and
since the sum of the probabilities of successful and failure k-exchanges is
smaller than 1, we conclude that successful O(nδ )-exchanges happen with a
−1/2
probability smaller than 1−e2n
∈ O( √1n ). Therefore, a successful iteration
√ n ) since N < 1 ln n.
happens with a probability 1−(1−O( √1n ))N ∈ O( ln
220
n
Theorem 3 generalizes the finding of [20] to simple TSP instances. It
formally states that for ρ = 1, N ∈ Ω(ln n) is necessary to efficiently find
an optimal solution to TSP. By Theorem 3, Theorem 1, Theorem 2 and its
Corollary 1, we have clearly analyzed the impact of the size of N on the
resulting stochastic runtime for the simple TSP instances in the case of that
ρ = 1. N ∈ ω(ln n) is sufficient to find the optimal solution in a stochastically
polynomial runtime, and the degree of the polynomial may increase with N ,
but the probability guaranteeing the runtime is also increasing with N .
5.2. Stochastic runtime analysis for grid instances
Now, we consider more general TSP instances. Herein, the n vertices are
positioned on an m × m grid for some integer m ∈ N+ . The vertices are
positioned in a way that no three of them are collinear. Figure 3 gives an
28
5
4
3
2
1
0
1
2
3
4
5
Figure 3: A grid instance
example of such an instance where m = 5 and n = 8. The weight of an edge
{l, k} ∈ E in this case is defined as the usual Euclidean distance d(l, k)
between vertex l and vertex k for every l, k = 1, . . . , n. In this section, we
shall refer to these TSP instances as grid instances.
Grid instances have been studied in [43] and [30]. Sutton and Neumann
[43] investigated the expected runtime of (1+1) EA and RLS for these instances. As a continuation of [43], Sutton et al [30] further proved that
the more extensive algorithm (µ + λ) EA finds an optimal solution for the
instances expectedly in
O((µ/λ)n3 m5 +nm5 +(µ/λ)n4k (2k−1)!)
iterations if every of the λ selected parents is mutated by taking a random
number of consecutive 2-exchange moves, and expectedly in
O((µ/λ)n3 m5 +nm5 +(µ/λ)n2k (k−1)!)
iterations with a mixed mutation operator, where k denotes the number of
vertices that are not on the boundary of the convex hull of V. Sutton et
al [30] also studied general Euclidean TSP instances (without collinearity)
and showed similar results in terms of the maximum distance value dmax ,
the minimum distance value dmin , k and the minimum angle in the triangles
formed by the vertices.
Before we present our stochastic runtime, we summarize some structural
properties of grid instances (some just follow from properties of general Euclidean instances). We say that two different edges {i, j} and {k, l} intersect
with each other if there exists a point p such that p ∈
/ {i, j, k, l} locates
on both of the two edges, see, e.g., Figure 4a. We say that a solution is
29
intersection-free if the corresponding Hamiltonian cycle does not contain intersections, see, e.g., Figure 4b.
j
k
p
l
j
l
i
(a) intersection
k
i
(b) intersection free
Figure 4: Example for intersections
Obviously, the triangle inequality [49] holds for grid instances. Therefore,
removing an intersection by a (unique) 2-exchange move in a solution strictly
reduces the total traveling cost, see Figure 4a. Lemma 1 states the well known
fact that an optimal solution of grid instances is intersection-free.
Lemma 1. Optimal solutions of grid instances are intersection-free.
i
l
p
k
j
Figure 5: Example for a 2-opt move
We now restrict 2-opt moves to 2-exchange moves that remove an intersection. For example, removing edges {i, j}, {k, l} in Figure 5 and adding
new edges {i, l}, {k, j} form such a 2-opt move. Lemma 2 below says that
for grid instances, removing one intersection may reduce the total traveling
cost Ω(m−4 ) if it is applicable. We omit the simple proof here. Interested
readers may refer to [30] for a proof.
Lemma 2. If a feasible solution to a grid instance contains intersections,
then removing the intersection can reduce the total traveling cost Ω(m−4 ).
30
The convex hull Y(V ) of the vertex set V is the smallest convex set in R2
that contains V . Its boundary is a convex polygon spanned by some vertices
with possibly other vertices in the interior of that polygon. Let V b denote
the set of vertices on the boundary of Y(V ). Figure 6 illustrates this.
Figure 6: Example of a convex hull
Quintas and Supnick [50] proved that if a solution s is intersection-free,
then the solution respects the hull-order, i.e., any two vertices in the subsequence of s induced by the boundary (the outer polygon) of Y(V ) are
consecutive in s if and only if they are consecutive on the boundary of Y(V ).
Therefore, if V b = V, i.e., all of the vertices are on the convex hull, then
every intersection-free solution is optimal.
Theorem 4 below analyzes the stochastic runtime of Algorithm 1 for grid
instances for the case that V = V b . It states that the stochastic runtime is
O(n4 ·m5+ ) for the vertex-based random solution generation, and O(n3 ·m5+ )
for the edge-based random solution generation. Corollary 2 further improves
the runtime by sacrificing the probability guarantee. These stochastic runtimes are close to the expected runtime O(n3 ·m5 ) for RLS reported by Sutton
et al [43] and [30].
Theorem 4. Consider a TSP instance with n vertices located on an m × m
grid such that no three of them are collinear. Assume that V b = V , i.e., every
vertex in V is on the convex hull V b , that we apply the max-min calibration
1
, ρ = 1, M = 1 and N ∈ Ω(m ) for some
(6) with πmax = 1 − n1 , πmin = n(n−2)
constant > 0. Then:
a) With an overwhelming probability of 1 − e−Ω(N ) , Algorithm 1 finds the
optimal solution within at most n4 · m5 iterations with the vertex-based
random solution generation.
b) With an overwhelming probability of 1 − e−Ω(N ) , Algorithm 1 finds an
optimal solution within at most n3 · m5 iterations with edge-based random solution generation.
31
Proof of Theorem 4. Note that under the conditions of Theorem 4, every
intersection free solution is optimal. By Lemma 2, we know that a 2-opt move
reduces the total traveling cost by Ω(m−4 ). Therefore, n·m5 consecutive 2-opt
moves turn a feasible solution into an optimal one, since the worst solution in
this case has a total traveling cost smaller than n·m and the optimal solution
has total traveling cost larger than n. Notice also that m ≥ n/2, since the n
vertices are positioned on the m × m grid and no three of them are collinear.
With these facts, we prove the Theorem by a similar argument to the one
used in the proof of Theorem 2.
Again, we consider the random event that the cost of the iteration best
solution does not increase within a specified period of polynomially many
iterations and strictly decreases sufficiently many times within that period.
For a), we consider the first n4 m5 iterations. For b), we consider the first
n3 m5 iterations.
4 5
For a) : By Claim 2, with probability (1 − (1 − Ω(1))N )n m = 1 − e−Ω(N ) ,
the cost of the iteration-best solution does not increase within n4 m5 iterations. By Claim 1, for a phase consisting of consecutive n3 iterations, with
3
probability 1 − (1 − n−3 )N ·n = 1 − e−Ω(N ) , in at least one iteration of that
phase an intersection is removed from the iteration-best solution, provided
the phase starts with an iteration-best solution containing at least one intersection. Since the first n4 m5 iterations can have nm5 such phases, a)
follows.
b) follows with an almost identical discussion. We therefore omit the
proof.
Corollary 2. Consider a TSP instance with n vertices located on an m × m
grid such that no three of them are collinear. Assume that V b = V , i.e., every
vertex in V is on the convex hull V b , that we apply the max-min calibration
1
, ρ = 1, M = 1 and N ∈ ω(ln m). Then:
(6) with πmax = 1 − n1 , πmin = n(n−2)
a) With probability 1 − m−ω(1) , Algorithm 1 finds the optimal solution
within at most n4 · m5 iterations with the vertex-based random solution generation.
b) With probability 1−m−ω(1) , Algorithm 1 finds an optimal solution within
at most n3 · m5 iterations with the edge-based random solution generation.
Now, we consider the more interesting case that |V | − |V b | = k ∈ O(1),
i.e., k vertices are not on the convex hull. Note that we can turn an arbitrary
32
intersection-free solution to an optimal solution only by rearranging the positions of those k interior points in that solution, and this requires at most k
consecutive jump moves (see [30] for a proof). A jump move δi,j transforms a
solution into another solution by shifting positions i, j as follows. Solution s
is transformed into solution δi,j (s) by moving the vertex at position i into position j while vertices at positions between i and j are shifted appropriately,
e.g.,
δ2,5 (i1 , i2 , i3 , i4 , i5 , i6 , i7 ) = (i1 , i3 , i4 , i5 , i2 , i6 , i7 ) and
δ5,2 (i1 , i2 , i3 , i4 , i5 , i6 , i7 ) = (i1 , i5 , i2 , i3 , i4 , i6 , i7 ).
It is not difficult to see that a jump move δi,j can be simulated by either a
2-exchange move (in the case that |i−j| = 1) or a 3-exchange move (in all
other cases). Therefore, we can actually turn an intersection-free solution
into an optimal one by a sequence of at most k consecutive 2-exchange or
3-exchange moves. Furthermore, a sequence of k consecutive 2-exchange or
3-exchange moves can be simulated by a κ-exchange move with an integer
κ ≤ 3k. This means that any intersection-free solution can be turned into an
optimal solution by a κ-exchange move with κ ≤ 3k. We shall call such a κexchange move in the sequel a 3k-opt move, although κ may be smaller than
1
) by
3k. Recall that a 3k-opt move is produced with a probability of Ω( n6k−1
1
Algorithm 2 (see Claim 1), and with a probability of Ω( n3k ) by Algorithm
3 (see Lemma 6 of [42], or Claim 4) in any of the N independent draws in
[1]
iteration t, if Xt−1 is intersection-free and not optimal. As a result, we obtain
by a similar proof as above Theorem 5 below.
Theorem 5. Consider a TSP instance with n vertices located on an m × m
grid such that no three of them are collinear. Assume that |V | − |V b | = k ∈
O(1) (k vertices are not on the convex hull V b ), that we apply the max-min
1
calibration 6 with πmax = 1 − n1 , πmin = n(n−2)
, and set ρ = 1, M = 1, for
some constant > 0. Then:
a) If we set N ∈ Ω(n3 · m ), then with an overwhelming probability of
3
1 − e−Ω(N/n ) , Algorithm 1 finds an optimal solution within at most n ·
m5 + n6k−4 iterations with the vertex-based random solution generation;
b) If we set N ∈ Ω(n2 · m ), then with an overwhelming probability of
2
1 − e−Ω(N/n ) , Algorithm 1 finds an optimal solution within at most
n·m5 +n3k−2 iterations with the edge-based random solution generation.
33
Proof of Theorem 5. We only prove a). b) can be derived by a very similar
argument. We define two random events as following:
[1]
[1]
E1 : for each t ≤ n · m5 + n6k−4 , f (Xt−1 ) ≥ f (Xt );
[1]
E2 : for each t ≤ n · m5 + n6k−4 , if Xt−1 is not intersection-free, then a 2-opt
move happens in iteration t.
By a similar argument as the one for Theorem 4, we obtain that P[E1 ∩ E2 ] ≥
3
1−e−Ω(N/n ) . Let η be a random variable denoting the number of iterations for
[1]
which Xt−1 is intersection-free. Notice that, conditioned on E1 ∩E2 , η ≥ n·m5
implies that an optimal solution occurs within n · m5 + n6k−4 iterations.
Conditioned on E1 ∩ E2 and η < n · m5 , there are at least Ω(n6k−4 ) itera[1]
tions in which Xt−1 is intersection-free, since each Xt−1 is either intersection[1]
free or not intersection-free. Note also that in each iteration in which Xt−1
[1]
intersection-free and not optimal, a 3k-opt move that turns Xt−1 into an op1
))N . This
timal solution happens with probability of at least 1 − (1 − Ω( n6k−1
[1]
means for any fixed t ∈ N, if Xt−1 is intersection-free, then the probability of
[1]
1
the event that Xt is optimal is bounded from below by 1 − (1 − Ω( n6k−1
))N .
Therefore, for any fixed Ω(n6k−4 ) iterations in which the iteration-best solu[1]
tion Xt−1 is intersection-free and not optimal, the probability of the event
[1]
that the corresponding Ω(n6k−4 ) Xt ’s are still not optimal, is bounded from
6k−4
3
1
above by (1 − Ω( n6k−1
))N ·n
= e−Ω(N/n ) . This means that, conditioned on
E1 ∩ E2 and η < n · m5 , an optimal solution occurs within n · m5 + n6k−4
3
iterations with a probability of 1 − e−Ω(N/n ) .
As a result, an optimal solution occurs within the first n · m5 + n6k−4
3
iterations with a probability of 1 − e−Ω(N/n ) .
Theorem 5 shows a stochastic runtime of n3 m5+ +n6k−1 m for Algorithm
1 equipped with the vertex-based solution generation, and a stochastic runtime of n3 m5+ + n3k m for Algorithm 1 equipped with edge-based solution
generation, in the case of that |V | − |V b | = k ∈ O(1). This is much better
than the expected runtime
O(µ · n3 m5 +nm5 +µ · n4k (2k−1)!)
for (µ+λ) EA with sequential 2-opt mutations reported by Sutton et al [30].
However, we are not able to analyze the stochastic runtime in the case that
34
k ∈ ω(1), since k ∈ ω(1) interior points may require super-polynomially
many iterations to turn an intersection-free solution into an optimal solution
when a polynomial sample size is used.
6. Conclusion
We have analyzed the stochastic runtime of a CE algorithm on two classes
of TSP instances under two different random solution generation methods.
The stochastic runtimes are comparable with corresponding expected runtimes reported in the literature.
Our results show that the edge-based random solution generation method
makes the algorithm more efficient for TSP instances in most cases. Moreover, N ∈ Ω(ln n) is necessary for efficiently finding an optimal solution with
iteration-best reinforcement. For simple instances, N ∈ ω(ln n) is sufficient
to efficiently find an optimal solution with an overwhelming probability, and
N ∈ O(ln n) results in an exponential runtime with an overwhelming probability. However, for more difficult instances, one may need to use a relatively
large sample size.
Our stochastic runtimes are better than the expected runtimes of the
(µ + λ) EA on the grid instances. The EA randomly changes local structures
of some of its current solutions by a Poisson distributed number of consecutive 2-exchange moves in every iteration, while our algorithm refrains from
local operations on current solutions and only refreshes solutions by sampling from an evolving distribution. The solution reproducing mechanism in
the EA stays the same throughout the optimization, only the current solutions in every iteration vary. However, the solution reproducing mechanism
(sampling distribution) of our algorithm also evolves. This is the essential
difference of MBS with traditional EAs. The comparison of our results with
the expected runtimes in [30] therefore show that using a self-adaptive dynamic solution reproducing mechanism is helpful (in efficiently finding an
optimal solution) when the search space becomes rugged. The stochastic
runtimes in Theorem 4 are only valid for instances with a bounded number
of interior points. In the future, it should be interesting to analyze the case
that |V | − |V b | ∈ ω(1). This might also give more insight to the problem of
RP v.s. P [51].
Our analysis is actually a kind of worst-case analysis, which is rather
pessimistic. We analyze the optimization progress by only checking some very
particular random events. This may not only underestimate the probability
35
of finding an optimal solution with our algorithm, but also overestimate the
required number of iterations. In the future, it should be of great interest to
consider a smoothed runtime analysis over an -neighborhood of the n nodes
in the real plane as has been done for the Simplex method by Spielman and
Teng in their famous paper [52].
Acknowledgment
We thank the anonymous reviewers for their numerous useful suggestions
on improving the scientific quality and English presentation of this article.
References
[1] R. Y. Rubinstein, D. P. Kroese, The cross-entropy method: a unified
approach to combinatorial optimization, Monte-Carlo simulation and
machine learning, Springer Science & Business Media, 2004.
[2] R. Y. Rubinstein, Optimization of computer simulation models with rare
events, European Journal of Operational Research 99 (1) (1997) 89–112.
[3] R. Y. Rubinstein, The cross-entropy method for combinatorial and continuous optimization, Methodology and computing in applied probability 1 (2) (1999) 127–190.
[4] M. Dorigo, T. Stützle, Ant colony optimization, Cambridge, Massachusetts: A Bradford Book, MIT Press, 2004.
[5] M. Hauschild, M. Pelikan, An introduction and survey of estimation
of distribution algorithms, Swarm and Evolutionary Computation 1 (3)
(2011) 111–128.
[6] M. Zlochin, M. Birattari, N. Meuleau, M. Dorigo, Model-based search
for combinatorial optimization: A critical survey, Annals of Operations
Research 131 (1-4) (2004) 373–395.
[7] D. Whitley, A genetic algorithm tutorial, Statistics and computing 4 (2)
(1994) 65–85.
[8] H. R. Lourenço, O. C. Martin, T. Stützle, Iterated local search, Springer,
2003.
36
[9] Z. Wu, Model-based heuristics for combinatorial optimization: a mathematical study of their asymptotic behavior, Ph.D. thesis, Institut für
Angewandte Stochastik und Operations Research (IASOR), Technical
University of Clausthal (2015).
[10] T. Stützle, H. H. Hoos, MAX-MIN ant system, Journal of Future Generation Computer Systems 16 (2000) 889–914.
[11] S. Droste, T. Jansen, I. Wegener, On the analysis of the (1+1) evolutionary algorithm, Theoretical Computer Science 276 (1-2) (2002) 51–81.
[12] J. He, X. Yao, Drift analysis and average time complexity of evolutionary
algorithms, Artificial Intelligence 127 (1) (2001) 57–85.
[13] F. Neumann, C. Witt, Runtime analysis of a simple ant colony optimization algorithm, Tech. rep., Departmant of Computer Science, University
of Dortmund, Germany (2006).
[14] C. Witt, Runtime analysis of the (µ +1) ea on simple pseudo-boolean
functions, Evolutionary Computation 14 (1) (2006) 65–86.
[15] F. Neumann, C. Witt, Runtime analysis of a simple ant colony optimization algorithm, Algorithmica 54 (2) (2009) 243–255.
[16] B. Doerr, F. Neumann, D. Sudholt, C. Witt, Runtime analysis of the 1ant ant colony optimizer, Theoretical Computer Science 412 (17) (2011)
1629–1644.
[17] W. J. Gutjahr, G. Sebastiani, Runtime analysis of ant colony optimization with best-so-far reinforcement, Methodology & Computing in Applied Probability 10 (3) (2008) 409–433.
[18] Y. Zhou, J. He, A runtime analysis of evolutionary algorithms for constrained optimization problems, IEEE Transactions on Evolutionary
Computation 11 (5) (2007) 608–619.
[19] Y. Zhou, Runtime analysis of an ant colony optimization algorithm for
tsp instances, Evolutionary Computation IEEE Transactions on 13 (5)
(2009) 1083–1092.
37
[20] F. Neumann, D. Sudholt, C. Witt, A few ants are enough:aco with
iteration-best update, in: Genetic and Evolutionary Computation Conference, GECCO 2010, Proceedings, Portland, Oregon, Usa, July, 2010,
pp. 63–70.
[21] P. S. Oliverto, C. Witt, Improved time complexity analysis of the simple
genetic algorithm, Theoretical Computer Science 605 (15) (2015) 21–41.
[22] D. Sudholt, C. Thyssen, Runtime analysis of ant colony optimization for
shortest path problems, Journal of Discrete Algorithms 10 (10) (2012)
165–180.
[23] A. Lissovoi, C. Witt, Runtime analysis of ant colony optimization on dynamic shortest path problems, Theoretical Computer Science 561 (2015)
73–85.
[24] Y. Chen, X. Zou, Runtime analysis of a multi-objective evolutionary algorithm for obtaining finite approximations of pareto fronts, Information
Sciences 262 (2014) 62–77.
[25] D. Sudholt, C. Witt, Update strength in edas and aco: How to avoid
genetic drift, in: Genetic and Evolutionary Computation Conference,
2016, pp. 61–68.
[26] D. H. Wolpert, W. G. Macready, No free lunch theorems for optimization, IEEE Transactions on Evolutionary Computation 1 (1) (1997) 67–
82.
[27] F. Neumann, D. Sudholt, C. Witt, Analysis of different mmas aco algorithms on unimodal functions and plateaus, Swarm Intelligence 3 (2009)
35–68.
[28] F. Neumann, I. Wegener, Randomized local search, evolutionary algorithms, and the minimum spanning tree problem, Theoretical Computer
Science 378 (2007) 32–40.
[29] J. Reichel, M. Skutella, Evolutionary algorithms and matroid optimization problems, Algorithmica 57 (1) (2010) 187–206.
[30] A. M. Sutton, F. Neumann, S. Nallaperuma, Parameterized runtime
analyses of evolutionary algorithms for the planar euclidean traveling
salesperson problem, Evolutionary Computation 22 (4) (2014) 595–628.
38
[31] A. M. Sutton, J. Day, F. Neumann, A parameterized runtime analysis
of evolutionary algorithms for max-2-sat, in: Conference on Genetic &
Evolutionary Computation, 2012, pp. 433–440.
[32] Y. Zhou, X. Lai, K. Li, Approximation and parameterized runtime analysis of evolutionary algorithms for the maximum cut problem., IEEE
Transactions on Cybernetics 45 (8) (2015) 1491–1498.
[33] Z. Wu, M. Kolonko, Asymptotic properties of a generalized cross entropy
optimization algorithm, IEEE Transactions on Evolutionary Computation 18 (5) (2014) 658 – 673.
[34] Z. Wu, M. Kolonko, R. H. Möhring, Stochastic runtime analysis of the
cross entropy algorithm, IEEE Transactions on Evolutionary Computation, DOI: 10.1109/TEVC.2017.2667713.
[35] M. Held, R. M. Karp, A dynamic programming approach to sequencing
problems, Journal of the Society for Industrial and Applied Mathematics
10 (1) (1962) 196–210.
[36] N. Christofides, Worst-case analysis of a new heuristic for the travelling
salesman problem, Tech. rep., Graduate School of Industrial Administration, CMU (1976).
[37] M. T. Goodrich, R. Tamassia, Algorithm Design and Applications, Wiley, 2015.
[38] S. Arora, Polynomial time approximation schemes for Euclidean traveling salesman and other geometric problems, Journal of the ACM 45 (5)
(1988) 753–782.
[39] J. S. B. Mitchell, A constant-factor approximation algorithm for tsp with
pairwise-disjoint connected neighborhoods in the plane, in: TwentySixth Symposium on Computational Geometry, 2010, pp. 183–191.
[40] S. Lin, B. W. Kernighan, An effective heuristic algorithm for the
traveling-salesman problem, Operations Research 21 (2) (1973) 498–516.
[41] P. T. D. Boer, D. P. Kroese, S. Mannor, R. Y. Rubinstein, A tutorial
on the cross-entropy method, Annals of Operations Research 134 (1)
(2005) 19–67.
39
[42] T. Kötzing, F. Neumann, H. Röglin, C. Witt, Theoretical analysis of two
aco approaches for the traveling salesman problem, Swarm Intelligence
6 (1) (2012) 1–21.
[43] A. M. Sutton, F. Neumann, A parameterized runtime analysis of evolutionary algorithms for the euclidean traveling salesperson problem, in:
Proceedings of the Twenty-Sixth Conference on Artificial Intelligence
(AAAI’12), AAAI press, 2012, pp. 1105–1111.
[44] A. Costa, O. D. Jones, D. Kroese, Convergence properties of the crossentropy method for discrete optimization, Operations Research Letters
35 (5) (2007) 573–580.
[45] Z. Wu, M. Kolonko, Absorption in model-based search algorithms for
combinatorial optimization, in: Evolutionary Computation (CEC), 2014
IEEE Congress on, IEEE, 2014, pp. 1744–1751.
[46] M. Thomas, Machine learning, New Delhi: McGraw Hill Education India, 1997.
[47] H. Asoh, H. Mühlenbein, On the mean convergence time of evolutionary
algorithms without selection and mutation, in: Parallel Problem Solving
from Nature—PPSN III, Springer, 1994, pp. 88–97.
[48] M. Pirlot, General local search methods, European Journal of Operational Research 92 (3) (1996) 493–511.
[49] M. A. Khamsi, W. A. Kirk, An introduction to metric spaces and fixed
point theory, John Wiley,, 2001.
[50] L. V. Quintas, F. Supnick, On some properties of shortest Hamiltonian
circuits, American Mathematical Monthly 72 (9) (1965) 977–980.
[51] W. Gasarch, Classifying problems into complexity classes, Advances in
Computers 95 (2015) 239–292.
[52] D. A. Spielman, S. H. Teng, Smoothed analysis of algorithms: Why the
simplex algorithm usually takes polynomial time, Journal of the Acm
51 (3) (2004) 385–463.
40
| 2cs.AI
|
1
On User Association in Multi-Tier Full-Duplex
Cellular Networks
arXiv:1607.01119v4 [cs.IT] 13 Jun 2017
Ahmed Hamdi Sakr and Ekram Hossain
Abstract
We address the user association problem in multi-tier in-band full-duplex (FD) networks. Specifically, we consider the case of decoupled user association (DUA) in which users (UEs) are not necessarily
served by the same base station (BS) for uplink (UL) and downlink (DL) transmissions. Instead, UEs can
simultaneously associate to different BSs based on two independent weighted path-loss user association
criteria for UL and DL. We use stochastic geometry to develop a comprehensive modeling framework
for the proposed system model where BSs and UEs are spatially distributed according to independent
point processes. We derive closed-form expressions for the mean rate utility in FD, half-duplex (HD)
DL, and HD UL networks as well as the mean rate utility of legacy nodes with only HD capabilities
in a multi-tier FD network. We formulate and solve an optimization problem that aims at maximizing
the mean rate utility of the FD network by optimizing the DL and UL user association criteria. We
investigate the effects of different network parameters including the spatial density of BSs and power
control parameter. We also investigate the effect of imperfect self-interference cancellation (SIC) and
show that it is more severe at UL, where there exist minimum required SIC capabilities for BSs and
UEs for which FD networks are preferable to HD networks; otherwise, HD networks are preferable.
In addition, we discuss several special cases and provide guidelines on the possible extensions of the
proposed framework. We conclude that DUA outperforms coupled user association (CUA) in which
UEs associate to the same BS for both UL and DL transmissions.
Keywords: Multi-tier cellular networks, in-band full-duplex (FD), decoupled user association
(DUA), coupled user association (CUA), cell association, stochastic geometry.
I. I NTRODUCTION
In-band FD communication has recently attracted significant attention as a potential enabler
for 5G networks to support higher data rates and meet the ever-increasing users demand for
broadband wireless services. In contrast to HD communication in which a time-frequency resource block is only used for either transmission or reception, in-band FD communication
2
implies simultaneous transmission and reception of information in the same frequency band
[1]. This, in turn, introduces extra interference between UL and DL networks, which affects the
network performance gains especially for UL transmissions that suffer from excessive DL-to-UL
interference [2]. The performance of FD networks is also limited by the capability of UEs and
BSs to cancel self-interference (SI), which is caused by the transmitter to its collocated receiver.
Fortunately, this technology is becoming feasible thanks to the recent advancements in antenna
and digital baseband technologies where SI can be reduced close to the level of noise floor in
low-power devices [3].
In this paper, we focus on the user association problem in multi-tier cellular networks (i.e.,
networks that consist of different classes of BSs) that support in-band FD communication. We
consider the general case when a cellular UE can be simultaneously served by two different BSs
for UL and DL data transmission, i.e., decoupled user association (DUA) [4]. In addition, we aim
at optimizing user association in order to maximize the mean rate utility offered by FD networks.
In order to evaluate and optimize the system performance, we use a statistical approach based
on stochastic geometry to capture the network randomness [5]. Specifically, in order to derive
closed-form expressions for the mean rate utility in a generic link in the network, due to their
analytical tractability, we use independent Poisson Point Processes (PPPs) to model the locations
of BSs. The results from the analysis enable us to optimize user association to maximize the mean
rate utility and to understand the impact of network parameters (such as spatial density of BSs,
power control, weighting factors, and SI) on the performance to provide insightful guidelines
for system design. We show that DUA is superior to coupled user association in which UEs
associate to only one BS for both UL and DL transmissions. We also show that FD networks
can outperform HD networks in terms of mean rate for sufficiently high SI cancellation (SIC)
capabilities of BSs and UEs.
A. Related Work and Motivations
In the context of DUA in multi-tier cellular networks, the authors in [6], [7] propose a
framework for performance evaluation of multi-tier HD UL and DL cellular networks. In this
model, the locations of BSs in each tier are modeled by independent PPPs where each network tier
differs in the transmit power and spatial density. Using PPP assumption and weighted path-loss
user association, [6] derives expressions for the rate coverage in HD UL and DL networks as well
as the joint UL-DL rate coverage. On the other hand, closed-form expressions for the mean of
3
the logarithm of the transmission rate and spectrum allocation for both UL and DL transmissions
are derived in [7]. Furthermore, utility maximization problems are solved to optimize both user
association and spectrum partitioning. The authors in [8] use stochastic geometry to derive the
achievable capacity in an HD network with DUA where a real-world simulation tool (Atoll)
is used to verify the accuracy of the expressions. In [9], the DUA problem is formulated as a
matching game in which UEs and BSs in a two-tier FD cellular network rank one another based
on some preference metric, which is a function of the achievable signal-to-interference-plus-noise
ratio (SINR), in order to maximize the total throughput.
The authors in [10]–[16] evaluate the performance of FD networks using statistical modeling.
In [10], a hybrid ad-hoc network is considered in which nodes with HD or FD capabilities
are randomly deployed. For an ALOHA MAC protocol, it shows that FD networks achieve a
0−30% higher throughput, compared to HD networks, for practical values of path-loss exponent.
It also considers the effect of imperfect SIC and shows how it affects the relative performance
of FD and HD transmissions. The authors in [11] present the effects of transmission duration,
SIC, and ratio of HD-to-FD nodes on the performance of ad-hoc networks with asynchronous
ALOHA MAC protocol. The paper also highlights different optimal operating regions in which
FD networks outperform HD networks. In [12], the authors propose a system design that controls
the partial overlap between UL and DL channels in order to maximize the overall rate of FD
cellular networks. The paper compares the performances of two realizations for FD networks:
a two-node topology (2NT) with FD UEs and BSs and a three-node topology (3NT) with FD
BSs and HD UEs. It shows that 3NT achieves performance comparable to that of 2NT, which
paves the road to harvest the gains of FD transmissions with HD UE terminals. In [13], a hybrid
multi-tier FD cellular network is considered in which BSs operate either in FD mode or HD DL
mode. PPP assumption is used to derive expressions for the successful transmission probability
and network throughput. The authors in [14] consider a single-tier network under 2NT and 3NT
where BSs and UEs employ directional antennas. The directional antennas are shown to manage
both SI as well as co-channel interference. In [15], the authors derive the success probability
and achievable spectral efficiency of FD networks under a 3NT with multi-antenna BSs and
single-antenna UEs. In [16], a massive MIMO-enabled FD network is considered where BSs
adopt linear zero-forcing with SI-nulling precoding while UEs adopt SI-aware power control.
The authors show that massive MIMO is a viable solution for FD communications.
Although statistical modeling can be used for long-term performance evaluation of FD net-
4
works, it does not necessarily provide sufficient insights on short-term network performance.
Therefore, tools from optimization theory can be used to evaluate short-term performance of
networks and to find optimal parameters that maximize certain objective functions [17]–[20]. For
example, [17] proposes a joint resource management scheme to mitigate the effect of imperfect
SIC in an OFDMA-based two-tier FD cellular network. This is achieved by jointly assigning
UEs and transmit power for each resource block in both UL and DL transmission based on
the level of SI to maximize a total utility sum of the network. The authors in [18] propose
an iterative algorithm to jointly perform subcarrier assignment and power allocation in order to
maximize the sum-rate performance in a single-cell FD network. In [19], the authors propose a
joint UL/DL user scheduling and power allocation algorithm to maximize the system throughput
by investigating the feasibility conditions of FD operation with 3NT. The authors in [20] propose
two distributed power control and interference management methods to manage interference in
FD networks with 3NT. The proposed MAC protocol with transmit power optimization is shown
to outperform its HD counterpart in terms of total throughput.
B. Contributions, Organization, and Notations
In contrast to previous works on FD networks (e.g., [10]–[16]) in which CUA is used to
solve the association problem, we focus on DUA as a more general association criterion for FD
networks. The contributions of the paper can be summarized as follows:
•
Using tools from stochastic geometry, we provide a tractable framework to analyze the performance of multi-tier FD networks with DUA. We derive closed-form expressions for the
association probability, the mean interference received at UEs and BSs under weighted pathloss user association. In addition, we derive the mean rate utility of an FD network.
•
We investigate the effect of FD transmissions on legacy HD terminals that do not support FD
communication and we provide closed-form expressions for the mean rate utility.
•
We formulate an optimization problem to maximize the mean rate utility of FD networks. By
jointly optimizing both UL and DL user association to maximize the mean rate utility, we also
show that DUA is superior to CUA in FD networks.
•
We show how the proposed framework can be extended to different models in the literature
such as traditional HD UL and HD DL networks. In addition, we highlight different tradeoffs
in the network and show the effect of varying network parameters such as densities of BSs,
power control parameters, and user association weighting factors on network performance.
5
TABLE I
L IST OF KEY NOTATIONS
•
Notation
Definition
Notation
Definition
Φk
Pk
Ψk
Pmax
Ui
ψjk
τ UL , τ DL
α, αb , αu
Point process of BSs of k-th tier
Transmit power of BSs of k-th tier
Point process of active UEs of k-th tier
Maximum transmit power of UEs
Weighting factor for UL user association
Joint association probability
Target SINR threshold
Path-loss exponents
λk
ρk
γk , Γk
Di
DL
AUL
k ,Ak
σbk , σu
G, Gb , Gu
Spatial density of BSs of k-th tier
Receiver sensitivity of BS of k-th tier
Transmit power of UEs of k-th tier
Power control factor of UEs
Weighting factor for DL user association
Per-tier association probability
SI cancellation parameters
Path-loss gains
We investigate the effect of SI on the performance of an FD network and show that FD mode of
operation is preferable to HD mode only if the SIC capabilities of UEs and BSs are sufficient
to mitigate SI. Furthermore, we show that the effect of imperfect SIC is more severe at UL.
•
Via Monte Carlo simulation, we validate our analytical results. Also, through numerical results
we show the feasibility of FD communication to increase the mean transmission rate of cellular
networks.
The rest of the paper is organized as follows. System model and assumptions are described
in Section II. In Section III, both joint distance distributions and user association probabilities
are derived for a typical UE. Section IV presents the analysis for FD interference as well as
mean rate utility of UL, DL, and FD transmissions. An optimization problem is formulated in
Section V to maximize the mean rate utility of FD transmission. Finally, numerical results and
discussion are presented in Section VI before the paper is concluded in Section VII.
R∞
R∞
Notation: Γ[s] = 0 xs−1 e−x dx is the gamma function, Γ[s, a] = a xs−1 e−x dx denotes the
Rb
upper incomplete gamma function, γ[s, b] = 0 xs−1 e−x dx denotes the lower incomplete gamma
Rb
function, and Γ[s, a, b] = a xs−1 e−x dx denotes the generalized gamma function. f (·) and
F (·) denote the probability density function (PDF) and cumulative distribution function (CDF),
respectively. Finally, E[·] denotes the expectation operator and 1{·} is the indicator function.
The key mathematical notations used in this paper are summarized in Table I.
II. S YSTEM M ODEL , A SSUMPTIONS ,
AND
M ETHODOLOGY OF A NALYSIS
A. Multi-Tier Network Model
We consider a cellular network that consists of K tiers of BSs. We use an independent
homogeneous PPP Φk = {xi,k : i = 1, 2, . . . } with spatial density λk BS/m2 to model locations
of BSs belonging to the k-th tier where 1 ≤ k ≤ K and xi,k ∈ R2 denotes the location of the
6
i-th BS in that tier. For all BSs in k-th tier, transmit power is fixed and equal to Pk . Locations
of UEs are modeled in R2 according to an arbitrary independent point process ΦU with a spatial
P
density λu >> K
k=1 λk . Saturation condition is assumed where each transmitter (i.e., a BS or a
UE) sends data packets at the beginning of each time slot in a time-slotted transmission scenario.
B. Channel Model
We assume a co-channel deployment where the wireless channels are subject to both largescale (i.e., path-loss) and small-scale fading. Let kx − yk be the propagation distance between a
generic transmitter (i.e., a BS or a UE) located at x and a receiver located at y. The path-loss of
this link is defined as L(x, y) = Ḡkx − ykᾱ , where ᾱ > 2 is the path-loss exponent and Ḡ > 0
is a constant gain. In this work, we assume different path-loss exponents and gains for different
communication links [21]. That is, ᾱ = α and Ḡ = G for BS-UE links, ᾱ = αb and Ḡ = Gb for
BS-BS links, and ᾱ = αu and Ḡ = Gu for UE-UE links.
In addition to the distance-dependent path-loss, the small-scale fading component of a channel
is modeled by Rayleigh fading with unit average power where different links are assumed to
be independent and identically distributed (i.i.d.). Hence, the power gain of all channels is
exponentially-distributed and denoted by h ∼ Exp(1)1 . Channel coherence time is greater than
or equal to the frame duration. SI channel is modeled by Nakagami-m fading with parameters
(mbk , σbk ) and (mu , σu ) for BSs from k-th tier and UEs, respectively [16], [22]. Hence, the power
σ
gain of SI channel follows a Gamma distribution and denoted by hb ∼ Gamma(mbk , mbbk ) for
k
σu
) for UEs, where
BSs and hu ∼ Gamma(mu , m
u
1
σbk
and
1
σu
are SIC capabilities of BSs and UEs,
respectively. There is no intra-cell interference between UL (DL) transmissions where different
UEs in a cell are served in UL (DL) using orthogonal time-frequency resources (e.g. OFDMA).
Hence, there are only one active UE in UL and one active UE in DL per BS at a certain time
slot and channel.
C. Power Control Model
The UEs adopt fractional channel inversion power control where a UE at x adjusts its transmit
power to ρk (Gkx − ykα ) to compensate for the large-scale fading such that the average received
1
This assumption can be relaxed to other scenarios, e.g., Gamma-distributed power envelope assumptions to model multi-
antenna transmissions.
7
TABLE II
M ODES OF OPERATION
Mode
Definition
FD
3NT FD
Legacy DL
Legacy UL
HD DL
HD UL
All BSs and UEs are FD
All BSs are FD and all UEs are HD
A typical HD UE with DL transmissions in an FD network
A network tier of HD BSs with UL transmissions in an FD network
All BSs and UEs are HD with DL transmissions
All BSs and UEs are HD with UL transmissions
signal power at the serving BS at y is equal to ρk (Gkx − ykα )−(1−) . Note that 0 ≤ ≤ 1 is the
power control factor and ρk is the open loop power spectral density (or receiver sensitivity). Here,
we use γk to denote the instantaneous transmit power of a UE transmitting to a BS belonging
to the k-th tier where Γk is the corresponding random variable. UEs have a limited transmit
power budget of Pmax , where the UEs that are unable to perform channel inversion, transmit
with maximum power. Other power control mechanisms such as SI-aware power control [16]
and interference-aware power control [23] can also be adopted.
D. Mode of Operation and User Association
Besides FD links in which a UE simultaneously transmits and receives data in the same
channel, we consider network tiers in which UEs or BSs do not support FD transmissions (i.e.,
HD UEs and HD BSs). That is, we also evaluate the performance of the network when (i) there
is a typical HD UE that can only receive DL transmission from one BS in a channel during a
transmission interval, and (ii) there is a network tier of HD BSs each of which can only receive
UL transmission in a channel from one UE during a transmission interval. A communication
link is referred to as a legacy uplink when the UE-BS channel carries data only from the UE
to the serving HD BS. On the other hand, when the UE-BS channel carries data only from the
BS to the HD UE being served, the link is referred to as a legacy downlink. Furthermore, we
consider the 3NT model, where the network consists of FD BSs each of which serves two HD
UEs: one HD UE in DL and one HD UE in UL. The aforementioned scenarios are summarized
in Table II and will be discussed in details in Sections IV and VI.
For FD UEs, we consider the case where user association in UL is decoupled from that in
DL. Hence, an FD UE is not necessarily served by the same BS in both UL and DL. That is,
a UE may simultaneously receive data from one BS and transmit data to another in the same
channel. To illustrate, without loss of generality, Fig. 1 shows a realization of a two-tier FD
8
3
1
2
Cell Coverage (DL)
Cell Coverage (UL)
Downlink
Uplink
Macro-cell BS
Small-cell BS
FD User
HD User
Fig. 1. A two-tier FD cellular network with macro-cell and small-cell BSs. Solid lines show DL coverage area while dotted
lines show UL coverage of each cell.
cellular network where a macro-cell network tier is overlaid with a denser and lower power
small-cell network tier. It shows that the coverage area of each cell in DL is different from that
in UL. For example, although UE 1 is served by a macro-cell BS in DL, it is located in UL
coverage area of another small-cell BS, hence, served by two different BSs. On the other hand,
UE 2 is located in the coverage area of one small-cell BS in both DL and UL, hence, served
by the same BS. Fig. 1 also shows an example of 3NT realization where BS 3 serves two HD
UEs simultaneously: one HD UE is in DL and another HD UE is in UL.
We assume a weighted path-loss user association criterion similar to [24], [25] where each
UE independently associates with the BS(s) that minimizes the weighted path-loss of UL and/or
DL. That is, for a UE at y, with a slight abuse of notation, let xi , xUL , and xDL denote the BS
with minimum path-loss from i-th tier, serving BS in UL, and serving BS in DL, respectively.
Then, the association criterion for UL and DL can be described, respectively, as follows:
xUL = arg min Ui kx − ykα
(1)
xDL = arg min Di kx − ykα
(2)
x∈{xi }
x∈{xi }
where xi = arg minx∈Φi kx − ykα , i = {1, 2, . . . , K}, and Ui and Di are the weighting factors,
respectively, for UL and DL user association for the i-th tier.
Assumption 1. Let µi =
Ui
Di
and without loss of generality, let us assume that the network
tiers are ordered such that µ1 ≤ µ2 ≤ · · · ≤ µK .
9
Note that Ui and Di are design parameters and are not necessarily equal. Varying thsese
weighting factors can result in different association scenarios. For example, (i) coupled user
association: when Ui = Di , each UE associates to the same BS for both DL and UL, (ii)
minimum-distance user association: when Ui = U (or Di = D), each UE associates to the nearest
BS for UL (or DL), (iii) maximum-received power user association: when Di = Pi−1 , each UE
associates to the BS that offers the strongest received power for DL, and (iv) minimum-transmit
power user association: when Ui = ρi , each UE associates to the BS that requires lowest transmit
power for UL.
E. Methodology of Analysis
Based on the system model described above, we aim at quantifying the performance of a
generic FD link in terms of mean rate in nats/sec/Hz. We first derive the user association
probability and distance distributions based on decoupled and weighted path-loss user association.
Then, we obtain the mean of interference experienced at a typical BS in UL and at a typical
UE in DL. Next, we derive the mean rate utility of the network as well as that of UL and DL
transmissions. Finally, we optimize UL and DL weighting factors such that the mean rate utility
of the FD network is maximized.
III. A NALYSES OF D ISTANCE AND U SER A SSOCIATION P ROBABILITIES IN FD N ETWORKS
A. Analysis of User Association Probabilities
Note that, even with DUA, an FD UE can associate to the same BS in the k-th tier for both
UL and DL transmissions with a certain probability. It is worth mentioning that this scenario is
different from the CUA as it does not necessitate Uk and Dk to be equal (as in the CUA). That
is, although the UE still uses two different decision criteria for UL and DL user association (i.e.,
DUA in (1) and (2)), one BS meets both criteria. This event occurs depending on the network
realization. Let ψjk denote the joint association probability that a UE is served by a BS from
the j-th tier in DL and a BS from the k-th tier in UL. The following lemma characterizes this
probability.
10
Lemma 1. (Joint association probability) The probability that an FD UE with DUA is served
by the j-th tier for DL and k-th tier for UL transmissions is
K
−1
2
P
λj
max {Dji , Uji } α λi
,
j=k
i=1
2
2
α2 j−1
α
P 1
ψjk =
µl+1
µlα
Dj
−
,
k<j
λj λk Uk
2
2
Υ2l (j)
α Ω
1+µl+1
1+µlα Ωl
l=k
l
0,
k>j
P
−1
P
K
K
l
2
P
Uj
Dj
λi
λi
α
where Ujk = Uk , Djk = Dk , Υl (j) =
Dji λi , and Ωl =
.
2
2
Proof: See Appendix A-I.
i=1 Uiα
i=l+1
(3)
α
i=l+1 Di
Note that the joint association probability ψjk is different from the per-tier association probability, which is defined as the probability that a UE is served by a BS belonging to a certain
tier for DL (UL) regardless of the serving UL (DL) BS. This per-tier association probability can
be obtained directly from the joint association probability derived in Lemma 1. The following
lemma provides expressions for this probability for both DL and UL transmissions.
Lemma 2. (Per-tier association probability) The probability that a UE associates to a BS from
the j-th tier for either DL or UL is defined, respectively, as follows:
DL −1
ADL
j = λj (Λj )
where ΛDL
j =
PK
2
UL
α
i=1 Dji λi and Λj =
PK
and
UL −1
AUL
j = λj (Λj )
(4)
2
α
i=1 Uji λi are the effective spatial densities of the j-th
tier for DL and UL transmissions, respectively.
B. Analysis of Distance to Serving BS(s)
Based on the system model and user association criteria described above, for UEs operating
in FD mode, the marginal PDFs of the distances to the serving BSs in DL and UL are presented
in the following lemma the proof of which can be found in [25, Appendix A].
Lemma 3. (Marginal distance distributions) The CDF of the distance between a generic UE
2
associated with the j-th tier for DL (or UL) and its serving BS is P[Rjm ≤ r] = 1−exp −πΛm
r
j
n
2+n
− n2
m
and the n-th moment is ERjm Rj = Γ 2
πΛj
, where m ∈ {DL, UL}.
Furthermore, the joint PDF of the distance to serving BSs for DL and UL transmissions for
FD UEs is presented in the following lemma.
11
Lemma 4. (Joint distance distribution) The joint PDF of the distance(s) from a generic FD
UE to its serving BS(s), when associated to the j-th tier for DL and k-th tier for UL, is
K
2
P
2
α
max {Dji , Uji } λi rj ,
j = k, rk = rj
2πλj rj exp −π
i=1
fR (rj , rk ) =
K
2
P
4π 2 λj λk rj rk exp −π
max Dji rjα , Uki rkα α λi , k < j, (rj , rk ) ∈ A
(5)
i=1
n
o
1
1
α
α
where A = (rj , rk ) : rj ≥ 0, Djk rj < rk < Ujk rj , and their expectations are given by
K
− m2
2
P
ER [Rjm ] = Γ 2
π
max {Dji , Uji } α λi
,
i=1
j−1
−1 j−1
P
P −1
m
Hjl [µl , µl+1 ; 0]
ER [Rj ] =
Hjl [1, 1; m] ,
Ωl
l=k
l=k
−1 j−1
j−1
P
P
n
Hjl [µl , µl+1 ; n] ,
Hjl [µl , µl+1 ; 0]
ER [Rk ] =
2+m
where Hjl [a, b; c] =
c
π2
− 4+c
2
(Υl (j))
Proof: See Appendix A-II.
k<j
(6)
k<j
l=k
l=k
Γ[ 2+c
2 ]
j=k
− 2+c
− 2+c
2
2
2
2
2+c
2+c
α
α
a α 1 + µl Ω l
− b α 1 + µl+1 Ωl
.
C. Analysis of Uplink Transmission Power
As stated above, the UEs served in UL by a BS from k-th tier are assumed to perform fractional
channel inversion with open loop power spectral density ρk . The UEs are also assumed to have
a constraint Pmax on the transmit power. Thus, we define the required amount of transmit power
of a UE when it associates with a BS from the k-th tier as γk = min {ρk G Rkα , Pmax } where
Rk is the distance to the serving BS from tier k for UL transmission and its CDF is given in
Lemma 3. Therefore, following Lemma 3, we derive the CDF of the transmit power where the
α1 i
h
t
UL
· 1{t < Pmax }.
proof follows directly such that P[Γk > t] = P Rk > ρk G
Lemma 5. (Transmit power distribution) The CDF of the transmit power of a UE associated
to tier k in UL and performing fractional channel inversion power control is P[Γk ≤ t] =
h
α2 i
t
1 − exp − πΛUL
· 1{t < Pmax } and its n-th moment is given by
k
ρk G
"
α2 #
n n
nα
P
nαρ
G
max
k
E[Γnk ] =
, πΛUL
.
(7)
nα γ
k
2
2
ρk G
2(πΛUL
)
k
12
IV. A NALYSIS OF R ATE C OVERAGE OF FD T RANSMISSIONS
In this section, we characterize the mean rate utility for a generic FD link. We assume
that during a transmission interval, a receiver (i.e., BS or a UE) using a particular channel
to communicate with its corresponding transmitter experiences interference from all other BSs
as well as UEs reusing the same channel. We start by defining the SINR received at a typical
UE and BS. Then, the mean rate utility is also derived for a generic FD link and several special
cases in the literature.
A. Signal-to-Interference-plus-Noise Ratio
For the system model described above, the SINR can be expressed as follows:
SINRDL
jk =
Pj h
DL
DL
(I + ISI + σ 2 )GkxDL kα
and
SINRUL
jk =
min{ρk G kxUL kα , Pmax }h
2
UL α
(IUL + IUL
SI + σ )Gkx k
(8)
where xDL ∈ Φj and xUL ∈ Φk denote the serving BS for DL and UL transmissions, respectively,
as defined in (1)-(2). σ 2 is the additive noise power, {h} are channel power gains where subscripts
are dropped for simplicity, and IDL and IUL are the interference received, respectively, at the
tagged UE (i.e., DL) and BS (i.e., UL) from all other BSs and UEs sharing the same channel.
Note that, unlike HD networks, interference signals received from DL and UL transmissions are
correlated due to the relative distance between each UE-BS pair and the observation point. In
order to characterize the interference, using the fact that each cell has exactly one DL and one
UL transmissions, we pair each UE and its serving BS in UL together as an interfering pair.
Hence, the aggregate interference at a typical receiver (i.e., a UE or a BS) located at the origin
(0, 0) can be defined as follows2 :
IDL =
K
X
X
Pi h
+
Gkxkα
X
Pi h
+
Gb kxkαb
i=1 x∈Φi
I
UL
=
K
X
\{xDL }
i=1 x∈Φi \{xUL }
X
y∈Ψi
\{u[xUL ]}
X
γi (y)h
Gu kykαu
(9)
γi (y)h
Gkykα
(10)
y∈Ψi \{u[xUL ]}
where u[x] returns the UE served by BS x in UL, γi (y) = min{ρi G ky − u−1 [y]kα , Pmax } is
the transmit power of the UE, and Ψi is the point process that represents the active UEs. Note
that, due to the correlation between the positions of active UEs and BSs resulting from the
orthogonal time-frequency resource allocations (e.g., OFDMA), this point process does is not
2
Due to the homogeneous PPP assumption, interference statistics are independent of the observation point [5].
13
a PPP. Instead, the positions of active UEs can be seen as a Voronoi perturbed lattice process
which is not mathematically tractable [6], [26].
Assumption 2. The active UEs from the i-th tier are assumed to form an arbitrary point
process Ψi that is (i) stationary with spatial density λi , (ii) independent of the point processes
of active UEs from different tiers, and (iii) independent of Φi 3 .
B. Interference Model
For each type of tagged receivers (i.e., BSs or UEs) and interferers (i.e., BSs or UEs), we
define a pair correlation function g(r) which in turn quantifies the spatial density of interfering
nodes from i-th tier as λi g(r).
(i) BSs-to-UEs (DL-to-DL): Using (2), we know that no interfering BS from i-th tier can have
1
less weighted path-loss to a UE than its serving BS, hence, kxk > Djiα kxDL k, where x ∈ Φi
1
and xDL ∈ Φj [5]–[7], [24], [27]. Therefore, g1 (r) = 1{r > Djiα Rj }, where Rj = kxDL k.
(ii) UEs-to-BS (UL-to-UL): There is no exact boundary for the exclusion region around the
tagged BS where interfering UEs can be arbitrarily close. However, using (1), we know
that none of the interfering UEs associates with that BS, hence, for a UE u served by a BS
1
from i-th tier, kxUL − uk > Uikα kx − uk, where x ∈ Φi and xUL ∈ Φk [6], [7], [25], [28].
1
Therefore, g2 (r) = 1{r > Uikα Ri }, where Ri = kx − uk.
(iii) BSs-to-BS (DL-to-UL): Due to the PPP assumption, the BSs can be very close to each other,
however, this is not true in practice due to physical constraints, different antenna heights,
etc. [21]. Here, we use the approximation proposed in [29] to model interfering BSs from
i-th tier as a non-homogeneous point process with g3 (r) = (1 − exp[−π βλbi r2 ])1{r > db }
to model the repulsion between BSs in practical deployments where βb is the repulsion
parameter and db is a constraint on the length of an DL-to-UL interference link.
(iv) UEs-to-UE (UL-to-DL): We use a similar approximation to model the interfering UEs
served by the i-th tier of BSs in UL as a non-homogeneous point process with g4 (r) =
i
(1 − exp[−π AλUL
r2 ])1{r > du }, where AUL
is the repulsion parameter as in [6] and du is
i
i
the minimum length of an UL-to-DL interference link [30].
3
2
Other approximations for Ψi exist in the literature, e.g., non-homogeneous PPP with spatial density λi 1 − exp[−πΛUL
i r ]
[6], or more generally λi g(r), where g(r) is a pair correlation function as proposed in [29]. Another approximation is using
q
1
PPP assumption with exclusion ball of radius
such as in [31] for single-tier networks.
πΛUL
i
14
C. Self-Interference
UL
In (8), IDL
SI is SI resulting from UL transmission at the UE, and ISI is SI resulting from DL
transmission at the BS. Since the SI incurred at a given receiver depends on its own transmit
power, we define the residual SI power after cancellation as follows:
UL α
IDL
SI = min ρk G kx k , Pmax hu
and
IUL
SI = Pk hbk
(11)
where hbk and hu represent SI channels such that σbk = E[hbk ] and σu = E[hu ] are the inverse
of SIC capability of BSs from tier k and UEs, respectively.
D. Mean Rate Utility
For a given FD link, let us define the rate coverage probability as the probability of UL (DL)
DL
transmission rate to be higher than a required target rate threshold RUL
o (Ro ). In other words,
for UL (DL) transmission to achieve the target rate, the SINR received at the BS (UE) must be
greater than a prescribed threshold τ UL (τ DL ) that can be given using Shannon’s formula such
that τ m = exp[Rm
o ] − 1, m ∈ {DL, UL}. Hence, the rate Rjk of a generic FD link measured in
nats/sec/Hz for a UE served by a BS from j-th tier for DL and a BS from k-th tier for UL is
defined as:
DL
UL
Rjk = P(SINRDL
) ln 1 + τ DL + P(SINRUL
) ln 1 + τ UL
jk > τ
jk > τ
|
{z
} |
{z
}
=RDL
jk
(12)
=RUL
jk
DL
UL
which corresponds to a fixed data rate transmission (i.e., RDL
] and RUL
])
o = ln[1+τ
o = ln[1+τ
UL
when the SINRs for DL and UL transmissions (i.e., SINRDL
jk and SINRjk , respectively) exceed
the predefined thresholds; otherwise, the transmission rate is zero.
Furthermore, we define the mean rate utility of a generic FD link in the network as:
R̄ =
K X
K
X
j=1 k=1
UL
ψjk f RDL
jk , Rjk
(13)
UL
where f RDL
is the mean utility of an FD link given that the UE is associated with jjk , Rjk
DL UL
UL
th tier for DL and k-th tier for UL. In this work, we use f RDL
,
R
=
E
ln Rjk Rjk =
jk
jk
UL
E ln RDL
as the mean utility function measured in ln[nats/sec/Hz] [7], [32]. The
jk + E ln Rjk
motivation of considering this utility is to achieve proportional fairness among UEs. This is
achieved by improving the rates of links with low rate coverage probability while saturating
the rates of links with high rate coverage probability. This performance metric is widely used
15
in the literature to solve the problem of radio resource allocations while achieving fairness
via maximizing R̄ [33]. Therefore, hereafter, we derive the following performance metric for a
generic UE operating in the FD mode:
m
m
m
,
E[ln Rm
jk ] = ln Ro + E ln P(SINRjk > τ )
m ∈ {DL, UL}
(14)
where the expectation is with respect to h, R, Φi , and Ψi .
First, we focus on UL rate coverage, which can be obtained as follows:
" "
!##
UL
UL
UL
2
τ
G
I
+
I
+
σ
(a)
SI
UL
) = E ln P g >
E ln P(SINRUL
jk > τ
α
min{ρk G Rk , Pmax }Rk−α
#
"
UL
UL
2
+
E
I
+
σ
E
I
(b)
SI
= −τ UL G ER
min{ρk G Rkα , Pmax }Rk−α
(15)
where (a) follows from (8) and (b) follows from the Rayleigh fading assumption. The mean of
IUL and IUL
SI are presented in the following lemma.
Lemma 6. (Mean of uplink interference) The mean interference power received at a typical
BS belonging to the k-th tier is given by
K
UL X
E I
=
2πλi
i=1
K1 (db , αb , βλbi )
and the mean of SI is E IUL
= σbk Pk , where
SI
and
Gb
2−α
U α K2 (i)
Pi + ik
ρi
G(α − 2)
!
(2<α<4)
α−2
α−2
d2−α
2−α
2−α
1
1
(d→0)
2
2
2
K1 (d, α, Λ) =
Γ
Γ
− (πΛ)
, πΛd
= − (πΛ)
α−2 2
2
2
2
2#
Pmax α
4 − α(1 − )
UL 2
UL
K2 (i) =
Γ
, πΛi do , πΛi
2
ρi G
"
α2 #
α−2 Pmax
4
−
α
P
max
2
+ (πΛUL
Γ
, πΛUL
i )
i
ρi G
2
ρi G
α(1−)−2
4 − α(1 − )
(Pmax →∞)
UL
UL
2
=
(πΛi ) 2 Γ
, πΛi do .
2
(πΛUL
i )
α(1−)−2
2
(16)
(17)
"
(18)
in which do is the minimum distance between UEs and BSs [21].
Proof: See Appendix B-I.
Note that in (18), the first term represents interference from UEs that are able to perform
channel inversion power control without exceeding the power budget where the second term
represents the interference from UEs who are transmitting with the maximum power Pmax .
16
Next, similar to UL, the rate coverage for DL can be obtained as follows:
"
#
DL
DL
2
E
I
+
E
I
+
σ
SI
DL
) = −τ DL G ER
E ln P(SINRDL
jk > τ
Pj Rj−α
(19)
where the mean of IDL and IDL
SI are presented in the following lemma.
Lemma 7. (Mean of downlink interference) The mean interference power received at a typical
UE served in DL by the j-th tier and in UL by the k-th tier is given by
2−α
K
2−α
α
UL
X
Dji Rj
K1 (du , αu , Λi )
2πλi
Pi +
E [Γi ]
E IDL =
G(α
−
2)
G
u
i=1
where E [Γi ] is given in Lemma 5 and the mean of SI power is
"
2#
DL
σu αρk G
Pmax α
α
UL
E ISI =
, πΛk
.
α γ
2
2
ρk G
2(πΛUL
k )
(20)
(21)
Proof: See Appendix B-II.
The following theorem presents closed-form expressions for the mean rate utility in (13) for
a generic FD link when a UE associates with the same tier j with probability ψjj and associates
with tier j for DL transmission and tier k for UL transmission with probability ψjk .
Theorem 1. (Mean rate utility of FD networks) The mean rate utility of the FD network model
described above is given by R̄ = R̄o −
!
!
K X
K
X
A2 σu E[Γk ]
τ UL σbk Pk + A3 (k)
τ DL A1 (j) Γ 2+α
2
+
+
+
K4 (k)
ψjk G
α
α
2
2
Pj
πΛDL
K3 (j, k)
ρk G (πΛUL
(πΛDL
j
j )
k )
j=1 k=1
(22)
UL
where R̄o = ln RDL
o + ln Ro and
2−α
A1 (j) =
K
X
2πλi Djiα
i=1
G(α − 2)
A3 (k) = σ 2 +
and
K
X
Pi ,
2
A2 = σ +
2πλi
1
Γ[ 2+α
2 ]
i
Gu
i=1
K1 (db , αb , βλbi )
Gb
i=1
K
X
2πλi K1 (du , αu , ΛUL )
K
P
2−α
U α K2 (i)
Pi + ik
ρi
G(α − 2)
2
α
α2
max {Dji , Uji } λi
,
K3 (j, k) =
−1 j−1
j−1
P −1
P
Hjl [1, 1; α]
Hjl [µl , µl+1 ; 0] ,
Ωl
π
i=1
l=k
"
2 + α(1 − )
K4 (k) = (πΛUL
, πΛUL
k ) γ
k
2
α
2
!
E [Γi ]
(23)
.
j=k
(24)
k<j
l=k
Pmax
ρk G
α2 #
"
α2 #
ρk G
2+α
P
max
Γ
, πΛUL
+
k
Pmax
2
ρk G
17
(Pmax →∞)
=
α
2
(πΛUL
k ) Γ
2 + α(1 − )
.
2
(25)
Proof: See Appendix C.
Note that in (25), the first term represents the mean of the inverse of the useful signal power
received at a typical BS from a UE being served that is able to perform channel inversion power
control without exceeding the power budget. On the other hand, the second term represents the
mean of the inverse of the useful signal power received from a UE that is transmitting with the
maximum power Pmax .
E. Special Cases
Using the lemmas derived above and Theorem 1, the performance of (i) a network tier that
consists only of legacy HD BSs with UL transmissions and (ii) a typical legacy HD UE with
DL transmissions is presented in the following two corollaries where the proofs follow directly
from (22). It is worth mentioning that if we add SI at FD BSs and FD UEs, the expressions for
mean rate utility in Corollaries 1 and 2, respectively, will be equivalent to the mean rate utility
of UL and DL transmissions in an FD network.
Corollary 1. (Legacy uplink transmissions) The mean rate utility offered by the k-th tier that
UL
consists only of HD BSs in an FD cellular network to its UEs is given by ln RUL
k = ln Ro −
!!
2−α
λi
K
α
X
K
(d
,
α
,
)1{i
=
6
k}
1
b
b
K4 (k)
U
K
(i)
τ UL 1−
2
β
b
G
σ2 +
2πλi
Pi + ik
ρi
(26)
α .
2
ρk
Gb
G(α − 2)
(πΛUL
k )
i=1
Corollary 2. (Legacy downlink transmissions) The mean rate utility offered by an FD network
DL
to a legacy HD UE served by the j-th tier is given by: ln RDL
j = ln Ro −
2−α
2+α
K
2+α
α
UL
DL
X
D
Γ
Γ
K
(d
,
α
,
Λ
)
τ
1 u
u
ji
i
2
2
2
G
2πλi DL
Pi +
E [Γi ]. (27)
α σ +
α
DL
DL
2
Pj
πΛ
G(α
−
2)
(πΛj ) 2
G
(πΛ
)
u
j
j
i=1
In addition, the performance of HD UL networks and HD DL networks can be obtained as
presented in the following corollaries where the proofs follow directly from (22).
Corollary 3. (Half-duplex uplink networks) The mean rate utility of HD UL networks with
weighted path-loss user association is given by
K
X
K
2−α
UL
X 2πλi U α K2 (i)
UL τ
1−
2
ik
R̄ = ln RUL
−
A
G
σ
+
ρi
o
k
ρ
G(α
−
2)
k
i=1
k=1
!
K4 (k)
α
2
(πΛUL
k )
(28)
18
(Pmax →∞)
(σ 2 =0)
=
(Ui =ρ)
(Pmax →∞)
(σ 2 =0)
=
ln RUL
o
ln RUL
o
−
K
X
2−α
K
2 + α(1 − )
4 − α(1 − ) X 2ρi Uikα
Γ
Γ
AUL
(α − 2) i
ρk
2
2
G
i=1
τ
AUL
k
k=1
UL
2 + α(1 − )
4 − α(1 − )
2τ UL
Γ
Γ
.
−
G (α − 2)
2
2
(29)
(30)
Corollary 4. (Half-duplex downlink networks) The mean rate utility of HD DL networks with
weighted path-loss user association is given by
2−α
K
K
2+α
α
DL
X
X
2πλi Dji
Γ 2
τ
2
ADL
G
Pi
R̄ = ln RDL
α σ +
o −
j
DL
DL
2
P
πΛ
G(α
−
2)
(πΛ
)
j
j
j
j=1
i=1
(σ 2 =0)
=
ln RDL
o −
(Di =Pi−1 )
(σ 2 =0)
=
K
X
j=1
ln RDL
o −
ADL
j
K
τ DL X 2Pi Dij DL
A
Pj i=1 α − 2 i
2τ DL
.
α−2
(31)
(32)
(33)
The results presented in (29) and (32) are consistent with the previous results in [7] on user
association in multi-tier HD cellular networks. Furthermore, from (30) and (33), it can be seen
that the performance of an interference-limited network is independent of the spatial density,
receiver sensitivity, and transmit power of BSs, which is consistent with the results in existing
literature [6], [7], [24], [25], [27].
In addition, by comparing the results in Corollaries 1 and 2 with those in Corollaries 3
and 4, respectively, it can be seen that the mean rates of both UL and DL transmissions in
FD networks are lower than those in HD networks. This degradation is due to the additional
interference resulting from FD transmissions and SI. It can be seen that for UL transmissions,
DL-to-UL interference can be reduced by increasing the isolation of BSs. This can achieved by
increasing SIC capability of BSs or using directional antennas where both vertical and horizontal
patterns are designed such that the BSs are not in the main-lobe of each other. For DL, the effect
of UL-to-DL interference can be seen to be less severe because of the low transmit power of
UEs compared to that of BSs.
Furthermore, the performance of FD networks with CUA and FD networks with 3NT are
presented in the following two corollaries where the proofs follow directly from (22).
19
Corollary 5. (Coupled user association) The mean rate utility of FD networks with CUA is
given by: R̄ = R̄o −
DL
K
X
τ
A1 (k)
2 + α σu E[Γk ]+A2
τ UL 1− σbk Pk + A3 (k)
DL
G
+Γ
+
G
K4 (k) . (34)
Ak
α
α
2
2
Pk
πΛDL
2
ρk
(πΛDL
(πΛDL
k
k )
k )
k=1
Corollary 6. (3-node topology) The mean rate utility of FD networks with 3NT for a UE
served by j-th tier in DL in given by (27) where that for a UE served by k-th tier in UL is
UL
given by ln RUL
k = ln Ro −
K
X
τ UL 1−
G
σ 2 + σbk Pk +
2πλi
ρk
i=1
K1 (db , αb , βλbi )
Gb
2−α
U α K2 (i)
Pi + ik
ρi
G(α − 2)
!!
K4 (k)
α .
2
(πΛUL
k )
(35)
Proof: The proof follows from Corollaries 1 and 2 while considering SI affecting UL
transmissions at BSs.
V. R ATE M AXIMIZATION IN FD N ETWORKS BY O PTIMAL U SER A SSOCIATION
In this section, for optimal user association, we derive closed-form expressions for the weighting factors that maximize the mean rate utility of FD cellular networks given in Theorem 1. For
analytical tractability, we assume that (i) all BSs have the same receiver sensitivity (i.e., ρk = ρ),
(ii) there is no constraint on the maximum transmit power of UEs (i.e., Pmax → ∞), and (iii)
SI at BSs is modeled as noise such that σbk Pk = σb . The last assumption implies that the SIC
capability of a BS is proportional to its transmit power (i.e., σbk ∝
1
),
Pk
that is, the cancellation
capability of BSs with higher transmit power such as macrocell BSs is better than that of BSs
with lower transmit power such as picocell or femtocell BSs. Following a similar procedure as
in [7], we first derive the optimal effective spatial density Λ∗UL
and Λ∗DL
j , then we retrieve the
k
PK λ j
α
α
∗
∗DL 2
2
optimal weighting factors Uk∗ ← a1 (Λ∗UL
=1
k ) and Dj ← a2 (Λj ) which satisfy
j=1 Λ∗DL
j
PK λk
and k=1 Λ∗UL = 1 for any positive constants a1 and a2 (cf. Lemma 2).
k
20
After some mathematical manipulations, maximizing (22) is equivalent to solving the following
set of concave optimization sub-problem simultaneously:
P1:
P2:
subject to
min
UL
ΛDL
j ,Λj
min
UL
ΛDL
j ,Λj
K
X
K
X
c1 λ j
DL 2+α
2
i=1
j=1 Pj (Λj )
K
K
X
X
c3 λ j
K
X
λj
ΛDL
j
j=1
2+α
2
Pi λ i
(ΛDL
i )
ρi λi ΛUL
i
ρj (ΛUL
j )
i=1
K
X
λj
= 1,
= 1,
ΛUL
j
j=1
j=1
2−α
2
and
+
α2
K
X
K
X
ρi λi K1 (du , αu , ΛUL )
c2 λ j
DL 2+α
2
j=1 Pj (Λj )
i=1
K
K
X
X
c4 λ j
+
j=1
Pj (ΛDL
j )
UL
ΛDL
j , Λj ≥ 0,
2+α
2
i=1
(ΛUL
i )
α
2
i
ρ i λi
(ΛUL
i )
2+α
2
∀j
where ci is an arbitrary positive constant for i ∈ {1, 2, 3, 4}.
Using Lagrangian relaxation and taking the first order partial derivatives with respect to ΛDL
j
and ΛUL
j , the optimal user association probability can be obtained as follows:
Λ∗DL
j
=
−2
Pj α
K
X
Pi λ i
i=1
which are equivalent to Dj∗ =
D
Pj
2
α
and
Λ∗UL
j
=
K
X
λi
(36)
i=1
and Uj∗ = U for arbitrary positive constants D and U . This
result is presented in the following theorem.
Theorem 2. (Optimal user association in multi-tier FD networks) For an FD network, the
mean rate utility is maximized when user association is based on maximum received power in
DL and on minimum distance in UL.
It is worth mentioning that the result in Theorem 2 shows the importance of DUA in FD
cellular networks in order to maximize the overall mean rate of the network. The reason that a
UE prefers to be served based on the maximum received SINR for DL and the minimum distance
for UL can be explained as follows. From the DL perspective: (i) the received signal power at
the UE from its serving BS is maximized, (ii) the received interference power at the UE from
other BSs is minimized, and (iii) the SI power at the UE is minimized when transmitting to the
nearest BS in UL. From the UL perspective: (i) the received signal power from the UE at its
serving BS is maximized, (ii) the received interference power from other UEs at the serving BS
is minimized when all UEs transmit with the minimum power, and (iii) since all BSs transmit
with fixed power, the network tier of the serving BS for DL (or DL weighting factors) has no
effect on the mean UL rate. The result above also shows the superiority of DUA compared to
CUA in the FD network under consideration. That is, as shown in Theorem 1, two different
association criteria are required for UL and DL user association to maximize the mean rate of
FD networks. Clearly, this is possible only with DUA but not possible in general under any
21
CUA criterion where the UEs are forced to associate with the same BS for both UL and DL.
This result is also intuitive since CUA is a special case of DUA implying that the maximum
performance of CUA can be always achieved by DUA.
VI. N UMERICAL R ESULTS AND D ISCUSSIONS
A. System Parameters
We use the results obtained above in closed-form to evaluate system performance in different
scenarios. We consider the following cases: (i) FD networks with both CUA and DUA, (ii) FD
networks with 3NT, (iii) HD DL networks (i.e., only DL transmissions), (iv) HD UL networks
(i.e., only UL transmissions), (v) legacy DL transmissions (i.e., HD UEs with DL transmissions in
an FD network), and (vi) legacy UL transmissions (i.e., HD BSs with UL transmissions in an FD
network). For simulation and numerical evaluation, unless otherwise stated, we consider a twotier network with spatial densities λ2 = 4λ1 . The transmit powers of BSs are {P1 , P2 } = {37, 33}
dBm and the maximum transmit power of UEs is Pmax = 23 dBm. The path-loss exponents are
{α, αb , αu } = {4, 3.7, 4}, path-loss constants are {G, Gb , Gu } = {0, 30, 0} dB, and the minimum
distances for the pair correlation functions are {do , db , du } = {1, 40, 1} m. For UL transmissions,
the power control factor is = 0.9 and all BSs have the same receiver sensitivity ρi = −40
dBm. The SIC capabilities of BSs and UEs are { σ1b , σ1u } = {70, 70} dB and the noise power is
k
σ = −104 dBm. For the evaluation of mean rate utility, SINR thresholds τ DL and τ UL are set to
UL
0, i.e., RDL
o = Ro = ln(2) nats/sec/Hz.
B. Validation of Analytical Results
We validate the expressions derived in Lemmas 6 and 7 for the mean interference received
at a BS and a UE, respectively. These two expressions are the keys to deriving all the following
results including Theorem 1 and they include all assumptions made throughout the analysis of
the mean rate utility. The curves in Fig. 2 compare the results obtained from simulations and
analysis as a function of the spatial density of tier 1. For Monte Carlo simulations, the locations
of the BSs simulated as independent PPPs over a 20 × 20 km2 area with a typical UE or BS at
the origin. To simulate the active UEs that are causing UL-to-UL and UL-to-DL interference,
the UEs are first dropped as a PPP with high spatial density, then the UEs associate with BSs
based on the defined association criteria for UL. Then, each BS randomly selects one UE for
UL transmission. Hence, Assumption 2 is not retained in the simulation. The mean interference
power for various links is averaged over 10, 000 iterations. The results in Fig. 2 validate the
-40
-25
-45
-30
Mean of Interference (dBm)
Mean of Interference (dBm)
22
-50
-55
-60
-65
-35
-40
-45
-50
Downlink-to-uplink (Analysis)
Uplink-to-uplink (Analysis)
Simulation
-70
Downlink-to-downlink (Analysis)
Uplink-to-downlink (Analysis)
Simulation
-55
0
5
10
15
20
25
0
Spatial density of tier 1 (# 0.52 : BS/km 2 ), 6 1
5
10
15
20
25
Spatial density of tier 1 (# 0.52 : BS/km 2 ), 6 1
(a) Mean of Interference at a typical BS
(b) Mean of Interference at a typical UE
Fig. 2. Analysis (Lemmas 6 and 7) vs. simulation: Mean of interference resulting from DL and UL transmissions at BSs and
UEs.
0.3
0.5
DUA
CUA
0.45
0.2
Downlink (HD Network)
Uplink (HD Network)
Downlink (Legacy & 3NT)
Uplink (Legacy)
Uplink (3NT)
0.15
0.1
0.05
0
10 0
Mean rate utility (nats/sec/Hz)
Mean rate utility (nats/sec/Hz)
0.25
0.4
0.35
0.3
0.25
0.2
10 1
10 2
0.15
10 0
10 1
Spatial density of tier 1 (# 5 (0.52 :) -1 BS/km 2 ), 6 1
Spatial density of tier 1 (# 5 (0.52 :) -1 BS/km 2 ), 6 1
(a) Mean rate utility of DL and UL
(b) Mean rate utility of DUA and CUA
10 2
Fig. 3. Mean rate utility (in nats/sec/Hz) vs. spatial density of tier 1 (in BS/km2 ).
accuracy of our approach to derive the mean interference and show that the assumptions made
above have a minor effect on the accuracy of the proposed analytical model.
C. Effect of Spatial Density
Fig. 3 shows the effect of varying the density of BSs on the mean rate utility for three
different networks: (i) HD DL network, (ii) HD UL network, and (iii) FD network including
both legacy DL and UL transmissions, and (iv) 3NT. For HD DL networks (i.e., Corollary 4)
in Fig. 3a, the mean rate utility of DL transmissions is almost independent of the density of
BSs. This can be explained as increasing λ results in increasing the power of both the useful
signal and interference and the SINR remains unchanged. For HD UL networks (i.e., Corollary
3), the performance results can be explained as follows. In sparse networks with low density,
the distances between UEs and their serving BSs become large and the UEs transmit with their
23
maximum power with high probability. Consequently, the useful received signal power decreases
and the interference power is increased at the serving BS. Increasing the BS density improves
the mean rate by making UEs closer to their serving BSs which in turn increases their ability
to perform channel inversion without exceeding the maximum power budget. This increases the
useful signal power and decreases interference power. With a very high BS density, the network
becomes interference-limited and almost all UEs are able to invert their channel towards the
serving BS, hence, the mean rate becomes independent of the spatial density of BSs.
Fig. 3a shows that increasing the density of FD networks has two different effects on the mean
rate utility of legacy transmissions. For legacy DL transmissions (i.e., Corollary 2), the mean
rate is lower compared to HD DL networks where the difference is due to the extra interference
resulting from UL transmissions sharing the same spectrum in the FD network. However,
increasing the density of BSs (i) decreases the transmit power of UEs, hence, decreases UL-toDL interference, (ii) increases the useful signal power which is counteracted by the increase in
DL-to-DL interference. Overall, this improves the mean rate of legacy DL transmissions and the
performance approaches that of HD DL networks in very dense networks. On the other hand,
the mean rate of legacy UL BSs (i.e., Corollary 1) degrades with increasing spatial density of
BSs. Compared to HD UL networks, the difference is due to the extra interference resulting
from DL transmissions. With increasing spatial density, the BSs become closer to each other
which increases the DL-to-UL interference power received at all BSs. At the same time, the
useful signal power received at any BS is upper bounded by ρ. Therefore, the SINR becomes
very low and the mean UL rate approaches zero with increasing BS density. A similar behavior
is evident for 3NT as shown in Fig. 3a. The mean DL rate is identical to that of legacy DL
because both rates are evaluated at an HD UE. On the other hand, the difference between the
mean UL rate in 3NT with FD BS and legacy UL with HD BS is due to the SI experienced at
the FD BS in 3NT.
Fig. 3b elaborates more on the effect of spatial density in FD networks by showing the mean
rate utility of a generic FD link for different user association criteria (i.e., Theorem 1 and
Corollary 5). Overall, it can be seen that the spatial density should be adjusted to balance the
trade-off between the rates of DL and UL transmissions in FD networks. As shown in Fig. 3a, in
dense FD networks, UL transmissions become more susceptible to high DL-to-UL interference.
On the other hand, low spatial density of BSs degrades the performance of both DL and UL
transmissions, respectively, due to high UL-to-DL interference and the maximum power budget
24
0.5
0.5
DUA (FD Network)
CUA (FD Network)
Downlink (HD Network)
Uplink (HD Network)
Downlink (Legacy)
Uplink (Legacy)
Mean rate utility (nats/sec/Hz)
0.4
0.35
0.4
0.3
0.25
0.2
0.15
0.35
0.3
0.25
0.2
0.15
0.1
0.1
0.05
0.05
0
-50
Mean Rate Utility (FD Network)
Downlink (HD Network)
Uplink (HD Network)
Downlink (Legacy)
Uplink (Legacy)
0.45
Mean rate utility (nats/sec/Hz)
0.45
0
-45
-40
-35
Receiver sensitivity of BSs (;)
-30
-25
(a) Mean rate utility vs. receiver sensitivity of BSs
0
0.2
0.4
0.6
Power control factor (0)
0.8
1
(b) Mean rate utility vs. power control factor
Fig. 4. Mean rate utility (in nats/sec/Hz) vs. UL power control parameters: sensitivity of BS receivers ρ (in dBm) and power
control factor .
constraint. Therefore, as shown in Fig. 3b, as the spatial density of BSs increases, the mean rate
utility of the FD network increases up to a maximum value due to the improvement in both DL
and UL transmissions, then it starts to decrease due to the degradation in the mean UL rate.
D. Effect of Power Control
Fig. 4a shows the effect of varying ρ on the mean rate utility of different networks. In
general, decreasing the sensitivity of the receiver (i.e., increasing ρ) increases the amount of
transmit power required by each UE to perform channel inversion towards the serving BS (i.e.,
Lemma 5). This, in turn, increases the power of useful signal, UL-to-UL interference, and ULto-DL interference (cf., Lemmas 6 and 7). Hence, as shown in Fig. 4a, the mean rate of DL
transmissions in the FD network deteriorates with increasing ρ because of UL-to-DL interference
compared to the HD DL scenario. For UL transmissions in both HD UL and FD networks, the
mean UL transmission rate increases due to increased useful signal power. This happens up to
a maximum value, then the rate starts to decrease because the transmit power of UEs becomes
limited by the power budget Pmax . Overall, there exists an optimal value of ρ that maximizes the
mean rate of FD networks and splits the performance into two regimes. That is, as ρ increases,
the mean rate of FD networks increases up to a maximum value, then it starts to decrease. This
behavior can be explained as follows. When ρ is very low (the left side of the optimal point),
while DL transmissions experience low interference from UL transmissions, the interference at
the BSs is very high compared to the useful signal power and thus the mean UL transmission
rate is almost 0. As ρ increases, the mean DL transmission rate of DL transmissions starts to
25
0.5
0.45
0.45
0.4
Mean rate utility (nats/sec/Hz)
Mean rate utility (nats/sec/Hz)
0.4
FD
0.35
0.3
0.25
0.2
P1 =33 dBm, P2 =23 dBm
0.15
P1 =33 dBm, P2 =26 dBm
0.1
P1 =33 dBm, P2 =30 dBm
0.05
P1 =33 dBm, P2 =33 dBm
HD-DL
0
10 -1
10 0
0.35
FD
; = -40 dBm
; = -43 dBm
Maximum
0.3
0.25
HD-UL
Maximum
10 1
10 2
0.2
10 -1
D2 /D1
10 0
U2 /U1
10 1
(a) Mean rate utility vs. ratio of DL weighting factors
(b) Mean rate utility vs. ratio of UL weighting factors
Fig. 5. Mean rate utility of FD, HD DL, and HD UL networks (in nats/sec/Hz) vs. the ratio of DL and UL weighting factors.
degrade while the mean rate of UL transmissions improves where this improvement dominates
the overall mean rate performance. This happens until achieving the maximum mean rate. After
this point, as ρ increases (the right side of the optimal point), the mean rates of both DL and UL
transmissions start to fall. Hence, the overall mean rate of the FD network start to decrease. A
similar behavior is observed for varying the power control factor which can be justified using
the same line of arguments as that for varying ρ. It is worth mentioning that Figs. 3b and 4a
show that DUA is always superior to CUA for all ranges of λ and ρ, which is in agreement with
the analytical results in Section V.
E. Effect of Weighting Factors
Fig. 5a shows the mean rate utility of FD and HD DL networks as a function of DL weighting
factors for different transmit power settings. For both networks, it can be clearly seen that the
maximum mean rate utility is achieved when the ratio of the weighting factors is equal to
the inverse of the transmit power ratio. For example, when
P2
P1
= 0.2 (i.e., {P1 , P2 } = {33, 26}
dBm), the rate of FD transmissions and HD DL transmissions are maximized when
Similar remark can be made for different cases when the ratio
values of
P2
,
P1
the rate is maximized when
D2
D1
P2
P1
D2
D1
=
P1
P2
= 5.
is 1, 0.5, and 0.1. For these
is set to 1, 2, and 10, respectively. This behavior is
consistent with Theorem 2 and can be explained as follows. For HD DL networks, intuitively,
the power of the useful signal at the UE is maximized while the power of interference from
other BSs is minimized when Dj∗HD = Pj−1 . Consequently, both the SINR and mean rate are
maximized. For FD networks, same argument holds for DL transmissions as shown in Corollary
2. On the other hand, the mean UL transmission rate in an FD network is independent of DL
26
weighting factors as shown in Theorem 1 and Corollary 1. Hence, the mean rate utility of FD
transmissions is maximized by maximizing the rate of DL transmissions by setting Dj∗FD = Pj−1 .
On the other hand, Fig. 5b shows the mean rate utility of FD and HD UL networks as a
function of UL weighting factors for different receiver sensitivity settings. In contrast to Fig.
5a, the maximum rate utility for both FD and HD networks is achieved when the ratio of
the weighting factors is equal 1. This behavior is evident for different values of ρ and can be
explained as follows. For HD UL networks, when the UEs associate with their nearest BSs,
the aggregate interference power is minimized as the UEs transmit with the minimum power
required to perform channel inversion power control. In addition, the useful signal power is
maximized as the probability of a UE to perform channel inversion without exceeding Pmax
increases as the distance to the serving BS decreases. Therefore, both SINR and mean rate
utility are maximum when Uk∗HD = U for some constant U . For FD networks, same argument
holds for UL transmissions as given in Corollary 1. For DL transmissions in an FD network,
UL-to-DL interference is minimized when all the UEs transmit with the minimum transmit power
as given in Theorem 1 and Corollary 2. Hence, the mean rate utility of FD transmissions is
maximized by simultaneously maximizing the rate of UL transmissions and minimizing the ULto-DL interference where both are achieved by setting Uk∗FD = U . Therefore, it is clear that the
maximum rate offered by an FD network is achieved for
Dj
Dk
=
Pk
Pj
and
Uj
Uk
= 1. In other words,
thanks to DUA, the mean rate utility can be maximized by simultaneously optimizing both DL
and UL weighting factors which is not generally possible with CUA.
F. Effect of Imperfect Self-Interference Cancellation
Fig. 6a shows the effect of SI on the performance of FD networks. It shows that the mean
rate of DL and UL transmissions in FD networks are highly impacted with increasing σu and
σb , respectively. It is clear that the effect of SI is more severe for UL transmissions as it occurs
at the FD BS where the transmit power is generally high compared to the power of the signal
received from UL UE. Fig. 6a also shows that, based on SIC capability of UEs and BSs, HD
transmissions may be preferable to FD transmissions. Fig. 6b shows the minimum required SIC
capabilities of the UEs and BSs so that the rate offered by an FD network is higher than that of
its HD counterpart. For example, for the case, λ2 = 4λ1 , the rate offered by an HD network is
higher than that offered by an FD network when the SIC capabilities of UEs and BSs are less
than 40 and 50 dBm, respectively. In addition, SIC requirements are lower for higher spatial
27
0.5
Mean rate utility (nats/sec/Hz)
0.4
70
6 2 =46 1
Mean rate (FD Network)
Downlink rate (HD Network)
Uplink rate (HD Network)
Downlink rate (FD Network)
Uplink rate (FD Network)
SI cancellation capability of BSs (dB)
0.45
0.35
0.3
0.25
0.2
0.15
0.1
6 2 =86 1
65
60
55
FD
FD
HD
HD
50
45
0.05
0
10
20
30
40
50
SI cancellation capability (dB)
60
70
(a) Mean rate utility vs. SIC capability
40
10
20
30
40
50
SI cancellation capability of users (dB)
60
70
(b) Minimum required SIC capability for FD networks
Fig. 6. Mean rate utility (in nats/sec/Hz) vs. SIC capability of BSs and UEs
1
σb
k
and
1
.
σu
density of BSs (e.g., for λ2 = 8λ1 ) due to the decrease of the transmit power and consequently
SI at UEs. From Fig. 6b, it can also be seen that the SIC capability of UEs does not have to
be as high as that of BSs to achieve the same performance. This is mainly because the transmit
power of FD UEs is not very high compared to the transmit power of BSs.
VII. C ONCLUSION
We have presented a comprehensive framework for user association in multi-tier FD cellular
networks. For both UL and DL transmissions, we have considered different user association
criteria including both CUA and DUA. We have used stochastic geometry to model, analyze,
and evaluate the performance of the proposed system in terms of mean rate utility of FD, UL,
and DL transmissions. Using weighted path-loss user association, we have derived the optimal
weighting factors that maximize the mean rate utility of FD transmissions in the presence of
DL-to-DL, DL-to-UL, UL-to-UL, and UL-to-DL interferences. In addition, we have shown that,
in order to maximize the mean rate utility of FD networks, the UEs should associate with their
nearest BSs in UL and to the BSs that result in the maximum received power in DL. This
shows the advantage of using DUA over CUA. We have also shown that FD networks may be
preferable to HD networks based on the level of SIC at UEs and BSs.
28
A PPENDIX A
P ROOF OF A SSOCIATION P ROBABILITY AND D ISTANCE D ISTRIBUTION
I. Proof of Lemma 1
We first consider the case when a UE associates with different BSs for DL and UL. A UE
associates with different BSs (xDL ∈ Φj for DL and xUL ∈ Φk for UL where j 6= k) under the
following four conditions:
(i) xUL meets the criterion in (1) for UL, i.e., Uk kxUL kα < minx∈Φi Ui kxkα
xDL meets the criterion in (2) for DL, i.e., Dj kxDL kα < minx∈Φi Di kxkα
∀i 6= j, k, (ii)
∀i 6= j, k, (iii) xDL
does not meet the criterion defined for UL in (1), i.e., Uj kxDL kα > Uk kxUL kα , and (iv) xUL does
not meet the criterion defined for DL in (2), i.e., Dk kxUL kα > Dj kxDL kα .
Let Rj = kxDL k and Rk = kxUL k, the event that j 6= k can be expressed as:
K
\
1
1
1
α
α α
α
Djk
Rj < Rk < Ujkα Rj .
min kxk > max Dji Rj , Uki Rk
i=1,i6=j,k
Hence,
(a)
"
"
ψjk = E exp −π
(b)
= 4π 2 λj λk
Z
Z
1
α
Ujk
1
α
Djk
K
X
i=1,i6=j,k
∞
Z
1
αu
Ujk
1
αu
Djk
0
= 2λj λk
(37)
x∈Φi
x
K
X
i=1
max Dji Rjα , Uki Rkα
"
uv exp −π
2
α
K
X
i=1
2
α
2
α
#
1
α
1
α
λi Djk Rj < Rk < Ujk Rj
2
α
#
#
max{Dji uα , Uki v α } λi dvdu
λi max Dji , Uki x
2
!−2
dx
(38)
where (a) follows from: (i) the fact that the minimum distance to a BS from a PPP Φi is
Rayleigh-distributed with CDF P [minx∈Φi kxk ≤ t] = 1 − exp[−πλi t2 ] and (ii) independence
assumption for network tiers, (b) follows since the expectation in (a) is with respect to Rj and
Rk which denote the distances to the serving BSs. Note also that ψjk = 0 when k > j due to
ordering the network tiers such that µi < µi+1 (i.e., Assumption 1).
For the case when j = k, using (1) and (2) and following a similar procedure, the probability
of the event, where xDL = xUL = xo given that xo belongs to the j-th tier, can be expressed as:
"
"
##
K
X
2
max{Dji , Uji } α λi Rj2 .
(39)
ψjj = E exp −π
i=1,i6=j
Hence, the result in (3) can be easily verified.
29
II. Proof of Lemma 4
Following the proof in Appendix A-I, the joint CDF of the distance can be obtained by adding
two more conditions such that Rj > rj and Rk > rk . Then, the joint PDF can be obtained by
differentiation. E[Rjm ] and E[Rkn ] follow the definition of the expected value.
A PPENDIX B
P ROOF OF M EAN I NTERFERENCE
I. Proof of Lemma 6
Following the definition in (10), g2 (r), and g3 (r), we have
"
#!
Z
Z ∞
K
2
α
UL (a) X
r
Pi ∞ 1 − exp −πΛDL
min{ρ
G
R
,
P
}
i
max
i
i
=
2πλi
E I
dr + ERi
r1−α dr
1
αb −1
α
G
r
G
b
db
Uik Ri
i=1
!
2−α
K
X
Uikα
K1 (db , αb , ΛDL
i )
Pi +
ERi min {ρi Riα , Pmax } Ri2−α
=
2πλi
(40)
Gb
G(α − 2)
i=1
where (a) follows from (i) the Rayleigh fading assumption of interference channels gain and (ii)
Campbell’s Theorem [5] knowing that the distance R from the tagged BS to the closest interfering
1
UE from tier i is greater than Uikα Ri (i.e., Ui Riα < Uk Rα ) as explained earlier in Section IV-B.
i
In addition, the spatial density of interfering BSs is λi (1 − exp[−π AλDL
r2 ]). Following Lemma
i
3, we obtain ERi min {ρi G Riα , Pmax } Ri2−α = ρi K2 (i).
II. Proof of Lemma 7
Following the definition in (9), g1 (r), and g4 (r), we have
!
Z ∞
Z ∞
K
UL 2
DL (a) X
1
−
exp
−πΛ
r
ERi [Γi ]
Pi
i
r1−α dr +
dr
E I
=
2πλi
1
α
−1
u
G Djiα Rj
Gu
r
du
i=1
(41)
where (a) follows from (i) the Rayaleigh fading assumption of interference channels gain and (ii)
i
the fact that Di Rα > Dj Rjα and the spatial density of interfering UEs is λi (1 − exp[−π AλUL
r2 ]).
i
Moreover, (21) is obtained from (11) such that E IDL
= σu E [Γk ].
SI
A PPENDIX C
P ROOF OF T HEOREM 1
By combining (13), (15), (19), Lemma 6, and Lemma 7 and rearranging all terms, we have
R̄ =
K
K X
X
j=1 k=1
DL
ψjk ER [ln RUL
jk ] + ER [ln Rjk ]
(42)
30
#
"
K
K
α
X
X
σ
P
+
A
(k)
R
bk k
3
k
UL
UL
= ln RDL
G
ψjk ER
o + ln Ro − τ
α
ρk G
min{Rk , Pρkmax
}
G
j=1
k=1
− τ DL G
K
K
X
A1 (j) X
j=1
Pj
K
K
X
1 X
ψjk ER Rj2 − τ DL G
ψjk (σu E[Γk ] + A2 )ER Rjα
Pj k=1
j=1
k=1
(43)
where the expectation is with respect to the distance to the serving BS(s) (i.e., DL and UL).
Hence, using Lemmas 3 and 4, the expressions in Theorem 1 can be verified.
R EFERENCES
[1] A. Sabharwal, P. Schniter, D. Guo, D. Bliss, S. Rangarajan, and R. Wichman,“In-band full-duplex wireless: Challenges
and opportunities” IEEE J. Select. Areas Commun., vol. 32, no. 9, pp. 1637–1652, Sept. 2014.
[2] I. Randrianantenaina, H. ElSawy, and M.-S. Alouini, “Limits on the capacity of in-band full duplex communication in
uplink cellular networks” in Proc. of 2015 IEEE Global Telecommunications Conference (GLOBECOM), 2015.
[3] S. Hong, J. Brand, J. Choi, M. Jain, J. Mehlman, S. Katti, and P. Levis, “Applications of self-interference cancellation in
5G and beyond,” IEEE Commun. Mag., vol. 52, no. 2, pp. 114–121, Feb. 2014.
[4] H. Elshaer, F. Boccardi, M. Dohler, and R. Irmer, “Downlink and uplink decoupling: A disruptive architectural design
for 5G networks,” in Proc. of 2014 IEEE Global Telecommunications Conference (GLOBECOM), 2014, pp. 1798–1803.
[5] M. Haenggi and R. K. Ganti, Interference in Large Wireless Networks. Now Publishers Inc, 2009.
[6] S. Singh, X. Zhang, J. G. Andrews, “Joint rate and SINR coverage analysis for decoupled uplink-downlink biased cell
associations in HetNets,” IEEE Trans. Wireless Commun., vol. 14, no. 10, pp. 5360–373, May 2015.
[7] Y. Lin, W. Bao, W. Yu, and B. Liang, “Optimizing user association and spectrum allocation in HetNets: A utility
perspective,” IEEE J. Select. Areas Commun., vol. 33, no. 6, pp. 1025–1039, June 2015.
[8] K. Smiljkovikj, H. Elshaer, P. Popovski, F. Boccardi, M. Dohler, L. Gavrilovska, and R. Irmer, “Capacity analysis of
decoupled downlink and uplink access in 5G heterogeneous systems,” Available [Online]: arXiv:1410.7270 (2015).
[9] S. Sekander, H. Tabassum, and E. Hossain, “A matching game for decoupled uplink-downlink user association in fullduplex small cell networks,” in Proc. of 2015 IEEE Global Telecommunications Conference (GLOBECOM), 2015.
[10] Z. Tong and M. Haenggi, “Throughput analysis for full-duplex wireless networks with imperfect self-interference
cancellation,” IEEE Trans. Commun., vol. 63, no. 11, pp. 4490–4500, Nov. 2015.
[11] A. Munari, P. Mähönen, and M. Petrova, “A stochastic geometry framework for asynchronous full-duplex networks.”
Available [Online]: arXiv:1512.01478 (2015).
[12] A. AlAmmouri, H. ElSawy, and M.-S. Alouini, “Flexible design for α-duplex communications in multi-tier cellular
networks,” Available [Online]: arXiv:1511.07903 (2015).
[13] J. Lee and T. Q. S. Quek, “Hybrid full-/half-duplex system analysis in heterogeneous wireless networks,” IEEE Trans.
Wireless Commun., vol. 14, no. 5, pp. 2883–2895, May 2015.
[14] C. Psomas, M. Mohammadi, I. Krikidis, and H. A. Suraweera, “Directional antennas for interference management in
full-duplex cellular networks,” Available [Online]: arXiv:1602.02718 (2016).
[15] I. Atzeni, M. Kountouris, “Full-duplex MIMO small-cell networks with interference iancellation,” Available [Online]:
arXiv:1612.07289 (2016).
[16] A. Shojaeifard, K. Wong, M. Di Renzo, G. Zheng, K. A. Hamdi, and J. Tang, “Massive MIMO-enabled full-duplex
cellular networks,” Available [Online]: arXiv:1611.03854 (2016).
31
[17] J. H. Yun, “Intra and inter-cell resource management in full-duplex heterogeneous cellular networks,” IEEE Trans. Mobile
Comput., vol. 15, no. 2, pp. 392–405, Feb. 2016.
[18] C. Nam, C. Joo, and S. Bahk, “Joint subcarrier assignment and power allocation in full-duplex OFDMA networks,” IEEE
Trans. Wireless Commun., vol. 14, no. 6, pp. 3108–3119, June 2015.
[19] S. Goyal, P. Liu, S. Panwar, R. Yang, R. A. DiFazio, and E. Bala, “Full duplex operation for small cells,” Available
[Online]: arXiv:1412.8708 (2015).
[20] W. Choi, H. Lim, and A. Sabharwal, “Power-controlled medium access control protocol for full-duplex WiFi networks,”
IEEE Trans. Wireless Commun., vol. 14, no. 7, pp. 3601–3613, July 2015.
[21] 3GPP, “TR 36.814 v.9.0.0: Further advancements for E-UTRA,” Mar. 2010.
[22] D. Nguyen, L. Tran, P. Pirinen, and M. Latva-aho, “On the spectral efficiency of full-duplex small cell wireless systems,”
IEEE Trans. Wireless Commun., vol. 13, no. 9, pp. 4896–4910, Sep. 2014.
[23] F. J. Martin-Vega, G. Gomez, M. C. Aguayo-Torres, and M. Di Renzo,“Analytical modeling of interference aware power
control for the uplink of heterogeneous cellular networks,” IEEE Trans. Wireless Commun., vol. 15, no. 10, pp. 6742–6757,
Oct. 2016.
[24] H.-S. Jo, Y. J. Sang, P. Xia, and J. Andrews, “Heterogeneous cellular networks with flexible cell association: A
comprehensive downlink SINR analysis,” IEEE Trans. Wireless Commun., vol. 11, no. 10, pp. 3484–3495, Aug. 2012.
[25] A. H. Sakr and E. Hossain, “Analysis of multi-tier uplink cellular networks with energy harvesting and flexible cell
association,” in Proc. of 2014 IEEE Global Telecommunications Conference (GLOBECOM), 2014, pp. 4712–4717.
[26] B. Blaszczyszyn and D. Yogeshwaran, “Clustering comparison of point processes with applications to random geometric
models,” Available [Online]: arXiv:1212.5285 (2014).
[27] A. H. Sakr and E. Hossain, “Location-aware cross-tier coordinated multipoint transmission in two-tier cellular networks,”
IEEE Trans. Wireless Commun., vol. 13, no. 11, pp. 6311–6325, Aug. 2014.
[28] A. H. Sakr and E. Hossain, “Analysis of K-tier uplink cellular networks with ambient RF energy harvesting,” IEEE J.
Select. Areas Commun., vol. 33, no. 10, pp. 2226–2238, Oct. 2015.
[29] M. Haenggi, “User point processes in cellular networks,” Available [Online]: arXiv:1611.08560 (2016).
[30] M. Ni, L. Zheng, F. Tong, J. Pan and L. Cai, “A geometrical-based throughput bound analysis for device-to-device
communications in cellular networks,” IEEE J. Select. Areas Commun., vol. 33, no. 1, pp. 100–110, Jan. 2015.
[31] T. Bai and R. W. Heath, “Analyzing uplink SINR and rate in massive MIMO systems using stochastic geometry,” IEEE
Trans. Commun., vol. 64, no. 11, pp. 4592–4606, Nov. 2016.
[32] F. Baccelli, B. Blaszczyszyn, and C. Singh, “Analysis of a proportionally fair and locally adaptive spatial aloha in Poisson
networks,” in Proc. IEEE Int. Conf. Comput. Commun. (INFOCOM), 2014, pp. 2544–2552.
[33] F. Kelly, A. Maulloo, and D. Tanm “Rate control in communication networks: shadow prices, proportional fairness and
stability,” Journal of the Operational Research Society, vol. 49, no. 3, pp. 237–252, Mar. 1998.
| 7cs.IT
|
Design and Optimisation of the FlyFast Front-end for
Attribute-based Coordination
Diego Latella
Mieke Massink
Consiglio Nazionale delle Ricerche
Istituto di Scienza e Tecnologie dell’Informazione “A. Faedo”
[email protected], [email protected]
Collective Adaptive Systems (CAS) consist of a large number of interacting objects. The design
of such systems requires scalable analysis tools and methods, which have necessarily to rely on
some form of approximation of the system’s actual behaviour. Promising techniques are those based
on mean-field approximation. The FlyFast model-checker uses an on-the-fly algorithm for bounded
PCTL model-checking of selected individual(s) in the context of very large populations whose global
behaviour is approximated using deterministic limit mean-field techniques. Recently, a front-end
for FlyFast has been proposed which provides a modelling language, PiFF in the sequel, for the
Predicate-based Interaction for FlyFast. In this paper we present details of PiFF design and an
approach to state-space reduction based on probabilistic bisimulation for inhomogeneous DTMCs.
1
Introduction
Collective Adaptive Systems (CAS) consist of a large number of entities with decentralised control and
varying degrees of complex autonomous behaviour. They form the basis of many modern smart city
critical infrastructures. Consequently, their design requires support from formal methods and scalable
automatic tools based on solid mathematical foundations. In [29, 27], Latella et al. presented a scalable mean-field model-checking procedure for verifying bounded Probabilistic Computation Tree Logic
(PCTL, [19]) properties of an individual1 in the context of a system consisting of a large number of
interacting objects. The model-checking procedure is implemented in the tool FlyFast2 . The procedure
performs on-the-fly, mean-field, approximated model-checking based on the idea of fast simulation, as
introduced in [8]. More specifically, the behaviour of a generic agent with S states in a system with a
large number N of instances of the agent at given step (i.e. time) t is approximated by K(µ(t)) where
K(m) is the S × S probability transition matrix of an (inhomogeneous) DTMC and µ(t) is a vector of size
S approximating the mean behaviour of (the rest of) the system at t; each element of µ(t) is associated
with a distinct state of the agent, say C, and gives an approximation of the fraction of instances of the
agent that are in state C in the global system, at step t. Note that such an approximation is a deterministic
one, i.e. µ is a function of the step t (the exact behaviour of the rest of the system would instead be a
large DTMC in turn); note furthermore, that the above transition matrix does not depend on N [29, 27].
Recently, modelling and programming languages have been proposed specifically for autonomic
computing systems and CAS [12, 5]. Typically, in such frameworks, a system is composed of a set of
independent components where a component is a process equipped also with a set of attributes describing features of the component. The attributes of a component can be updated during its execution so
1 The
technique can be applied also to a finite selection of individuals; in addition, systems with several distinct types of
individuals can be dealt with; for the sake of simplicity, in the present paper we consider systems with many instances of a
single individual only and we focus in the model-checking a single individual in such a context.
2 http://j-sam.sourceforge.net/
De Vink and Wiklicky (Eds.): QAPL 2017
EPTCS 250, 2017, pp. 92–110, doi:10.4204/EPTCS.250.6
c D. Latella & M. Massink
This work is licensed under the
Creative Commons Attribution License.
D. Latella & M. Massink
93
that the association between attribute names and attribute values is maintained in the dynamic store of
the component. Attributes can be used in predicates appearing in language constructs for component
interaction. The latter is thus typically modelled using predicate-based output/input multicast, originally
proposed in [26], and playing a fundamental role in the interaction schemes of languages like SCEL [12]
and C ARMA [5]. In fact, predicate-based communication can be used by components to dynamically
organise themselves into ensembles and as a means to dynamically select partners for interaction. Furthermore, it provides a way for representing component features, like for instance component location in
space, which are fundamental for systems distributed in space, such as CAS [30].
In [9] we proposed a front-end modelling language for FlyFast that provides constructs for dealing
with components and predicate-based interaction; in the sequel, the language—which has been inspired
by C ARMA— will be referred to as PiFF, which stands for for Predicate-based Interaction for FlyFast.
Components interact via predicate-based communication. Each component consists of a behaviour, modelled as a DTMC-like agent, like in FlyFast, and a set of attributes. The attribute name-value correspondence is kept in the current store of the component. Actions are predicate based multi-cast output and
input primitives; predicates are defined over attributes. Associated to each action there is also an (atomic)
probabilistic store-update. For instance, assume components have an attribute named loc which takes values in the set of points of a space, thus recording the current location of the component. The following
action models a multi-cast via channel α to all components in the same location as the sender, making
it change location randomly: α ∗ [loc = my.loc]hiJump. Here Jump is assumed to randomly update the
store and, in particular attribute loc. The computational model is clock-synchronous, as in FlyFast, but
at the component level. In addition, each component is equipped with a local outbox. The effect of an
output action α ∗ [πr ]hiσ is to deliver output label αhi to the local outbox, together with the predicate πr ,
which (the store of) the receiver components will be required to satisfy, as well as the current store of
the component executing the action; the current store is updated according to update σ . Note that output
actions are non-blocking and that successive output actions of the same component overwrite its outbox.
An input action α ∗ [πs ]()σ by a component will be executed with a probability which is proportional to
the fraction of all those components whose outboxes currently contain the label αhi, a predicate πr which
is satisfied by the component, and a store which satisfies in turn predicate πs . If such a fraction is zero,
then the input action will not take place (input is blocking), otherwise the action takes place, the store of
the component is updated via σ , and its outbox cleared.
Related Work CAS are typically large systems, so that the formal analysis of models for such systems
hits invariantly the state-space explosion problem. In order to mitigate this problem, the so called ‘onthe-fly’ paradigm is often adopted (see e.g. [10, 4, 21, 16]).
In the context of probabilistic model-checking several on-the-fly approaches have been proposed,
among which [13], [28] and [18]. In [13], a probabilistic model-checker is shown for the time bounded
fragment of PCTL. An on-the-fly approach for full PCTL model-checking is proposed in [28] where,
actually, a specific instantiation is presented of an algorithm which is parametric with respect to the
specific probabilistic processes modelling language and logic, and their specific semantics. Finally,
in [18] an on-the-fly approach is used for detecting a maximal relevant search depth in an infinite state
space and then a global model-checking approach is used for verifying bounded Continuous Stochastic
Logic (CSL) [1, 2] formulas in a continuous time setting on the selected subset of states.
An on-the-fly approach by itself however, does not solve the challenging scalability problems that
arise in truly large parallel systems, such as CAS. To address this type of scalability challenges in probabilistic model-checking, recently, several approaches have been proposed. In [20, 17] approximate
94
Designing Attribute-based FlyFast
probabilistic model-checking is introduced. This is a form of statistical model-checking that consists in
the generation of random executions of an a priori established maximal length [25]. On each execution
the property of interest is checked and statistics are performed over the outcomes. The number of executions required for a reliable result depends on the maximal error-margin of interest. The approach relies
on the analysis of individual execution traces rather than a full state space exploration and is therefore
memory-efficient. However, the number of execution traces that may be required to reach a desired accuracy may be large and therefore time-consuming. The approach works for general models, i.e. models
where stochastic behaviour can also be non Markovian and that do not necessarily model populations of
similar objects. On the other hand, the approach is not independent from the number of objects involved.
As recalled above, in [27] a scalable model-checking algorithm is presented that is based on mean-field
approximation, for the verification of time bounded PCTL properties of an individual in the context of
a system consisting of a large number of interacting objects. Correctness of the algorithm with respect
to exact probabilistic model-checking has been proven in [27] as well. Also this algorithm is actually an
instantiation of the above mentioned parametric algorithm for (exact) probabilistic model-checking [28],
but the algorithm is instantiated on (time bounded PCTL and) the approximate, mean-field, semantics of
a population process modelling language. It is worth pointing out that FlyFast allows users to perform
simulations of their system models and to analyse the latter using their exact probabilistic semantics and
exact PCTL model-checking. In addition, the tool provides approximate model-checking for bounded
PCTL, using the model semantics based on mean-field.
The work of Latella et al. [27] is based on mean-field approximation in the discrete time setting;
approximated mean-field model-checking in the continuous time setting has been presented in the literature as well, where the deterministic approximation of the global system behaviour is formalised as
an initial value problem using a set of differential equations. Preliminary ideas on the exploitation of
mean-field convergence in continuous time for model-checking were informally sketched in [23], but
no model-checking algorithms were presented. Follow-up work on the above mentioned approach can
be found in [24] which relies on earlier results on fluid model-checking by Bortolussi and Hillston [6],
later published in [7], where a global CSL model-checking procedure is proposed for the verification
of properties of a selection of individuals in a population, which relies on fast simulation results. This
work is perhaps closest related to [27, 29]; however their procedure exploits mean-field convergence
and fast simulation [11, 15] in a continuous time setting—using a set of differential equations—rather
than in a discrete time setting—where an inductive definition is used. Moreover, that approach is based
on an interleaving model of computation, rather than a clock-synchronous one; furthermore, a global
model-checking approach, rather than an on-the-fly approach is adopted; it is also worth noting that the
treatment of nested formulas, whose truth value may change over time, turns out to be much more difficult in the interleaving, continuous time, global model-checking approach than in the clock-synchronous,
discrete time, on-the-fly one.
PiFF has been originally proposed in [9], where the complete formal, exact probabilistic, semantics
of the language have been defined. The semantics definition consists of three transition rules—one for
transitions associated with output actions, one for those associated with input actions, and one for transitions to be fired with residual probability. The rules induce a transition relation among component states
and compute the relevant probabilities. From the component transition relation, a component one-step
transition probability matrix is derived, the elements of which may depend on the fractions of the components in the system which are in a certain state. The system-wide one-step transition probability matrix is
obtained by product—due to independence assumptions—using the above mentioned component probability matrix and the actual fractions in the current system global state. In [9] a translation of PiFF to the
model specification language of FlyFast has also been presented which makes PiFF an additional front-
D. Latella & M. Massink
95
end for FlyFast extending its applicability to models of systems based on predicate-based interaction. In
the above mentioned paper, correctness of the translation has been proved as well. In particular, it has
been shown that the probabilistic semantics of any PiFF model are isomorphic to those of the translation
of the model. In other words, the transition probability matrix of (the DTMCs of) the two models is the
same. A companion translation of bounded PCTL formulas is also defined [9] and proven correct.
The notion of the outbox used in PiFF is reminiscent of the notion of the ether in PALOMA [14]
in the sense that the collection of all outboxes together can be thought of as a kind of ether; but such a
collection is intrinsically distributed among the components so that it cannot represent a bottleneck in
the execution of the system neither a singularity point in the deterministic approximation.
We are not aware of other proposals, apart from [9], of probabilistic process languages, equipped
both with standard, DTMC-based, semantics and with mean-field ones, that provide a predicate-based
interaction framework, and that are fully supported by a tool for probabilistic simulation, exact and
mean-field model-checking.
We conclude this section recalling that mean-field/fluid procedures are based on approximations of
the global behaviour of a system. Consequently, the techniques should be considered as complementary
to other, possibly more accurate but often not as scalable, analysis techniques for CAS, primarily those
based on stochastic simulation, such as statistical model-checking.
In this paper we present some details of PiFF, a translation to FlyFast which simplifies that proposed in [9] and an approach to model reduction based on probabilistic bisimulation for Inhomogeneous
DTMCs. In Section 2 we briefly present the main ingredients of the PiFF syntax and informal semantics,
and we recall those features of FlyFast directly relevant for understanding the translation of PiFF to the
FlyFast input language proposed in [9]. A revised and simplified version of the translation is described
in Section 3. In Section 4 we introduce a simplified language for the definition of transition-probabilities
in PiFF that allows us to define in Section 5 a model reduction procedure of the translation result, based
on a notion of bisimulation for the kind of IDTMCs of interest, introduced in Section 5 as well. An
example of application of the procedure is presented in Section 6. Some conclusions are drawn in Section 7. A formal proof of decidability of the cumulative probability test for state-space reduction based
on bisimulation is provided in the Appendix.
2
Summary on PiFF and FlyFast
In the following we present the main ingredients of PiFF and the features of FlyFast relevant for the
present paper.
2.1
PiFF
A PiFF system model specification ϒ = (∆ϒ , Fϒ , Σ0 )(N) is a triple where Fϒ is the set of relevant function
definitions (e.g. store updates, auxiliary constants and functions), ∆ϒ is a set of state defining equations,
and Σ0 is the initial system state (an N-tuple of component states, each of which being a 3-tuple (C, γ, O)
of agent state C, store γ and outbox O). We describe the relevant details below referring to [9] for the
formal definition probabilistic semantics of the language.
The PiFF type system consists of floating point values and operations, as in FlyFast, plus simple enumeration types for attributes, declared according to the syntax attype < name > enum < id − list >.
< id − list > is a finite list of identifiers. Of course, attributes can also take floating point values.
In Figure 1 the attribute type Space is defined that consists of four values A, B, C, D modelling four
Designing Attribute-based FlyFast
96
attype Space enum A, B, C, D;
..
.
const H = 0.6;
const L = 1 − H;
const Hdiv2 = H/2;
const Ldiv2 = L/2;
..
.
attribute loc : Space;
..
.
func Hr(x : Space) : Space; x endfunc;
func N(x : Space) : Space; case x of A : A; B : B; C : B; D : A endfunc;
func S(x : Space) : Space; case x of A : D; B : C; C : C; D : D endfunc;
func E(x : Space) : Space; case x of A : A; B : A; C : D; D : D endfunc;
func W(x : Space) : Space; case x of A : B; B : B; C : C; D : C endfunc;
..
.
func pHr(x : Space) : float; case x of A : H; B : L; C : H; D : L endfunc;
func pN(x : Space) : float; case x of A : 0; B : 0; C : Ldiv2; D : Hdiv2 endfunc;
func pS(x : Space) : float; case x of A : Ldiv2; B : Hdiv2; C : 0; D : 0 endfunc;
func pE(x : Space) : float; case x of A : 0; B : Hdiv2; C : Ldiv2; D : 0 endfunc;
func pW(x : Space) : float; case x of A : Ldiv2; B : 0; C : 0; D : Hdiv2 endfunc;
..
.
update Jump
my.loc := Hr(my.loc) with pHr(my.loc);
my.loc := N(my.loc) with pN(my.loc);
my.loc := S(my.loc) with pS(my.loc);
my.loc := E(my.loc) with pE(my.loc);
my.loc := W(my.loc) with pW(my.loc)
endupdate
Figure 1: A fragment of FSI .
D. Latella & M. Massink
97
locations. Some auxiliary constants are defined, using the const construct inherited from FlyFast:
const < name > = < value >.
A PiFF store update definition has the following syntax3 :
update upd
my.a1 := e11 , . . . , my.ak := ek1 with p1 ;
..
.
my.a1 := e1n , . . . , my.ak := ekn with pn
endupdate
where upd is the update name (unique within the system model specification), a1 , . . . , ak are the attribute names of the component, e11 , . . . , ekn and p1 , . . . , pn are attribute/store-probability expressions
respectively, with syntax defined according to the grammars e ::= va | ca | my.a | f na (e1 , . . . , em ) and
p ::= v p | c p | f n p (e1 , . . . , em ). In the above definition of attribute expressions va is an attribute value
(drawn from finite set V of attribute values), ca is an attribute constant in V defined using the const;
a ∈ {a1 , . . . , ak } is an attribute name and f na is an attribute function defined by the user in Fϒ , which,
when applied to attribute expressions e1 , . . . , em returns an attribute value; the syntax for such function
definitions afd is given below:
afd ::= func f na (x1 : T 1, . . . , xm : T m) : T ; afb endfunc
afb ::= e |case (x1 , . . . , xm ) of(va11 , . . . , vam1 ) : e1 ; (va12 , . . . , vam2 ) : e2 ; . . . (va1k , . . . , vamk ) : ek
where f na is the name of the attribute function, x1 : T 1, . . . , xm : T m are its parameters and their relative
types, T is the type of the result of f na ; e, ei are attribute-expressions and vai j are attribute-values.
In Figure 1 attribute functions N, S, E, W are defined for North, South, East, and West, such that Space
models the Cartesian space with four quadrants: A = N(D) = E(B), B = N(C) = W(A), and so on, as shown
diagrammatically in Figure 2 right. Function Hr is the identity on Space.
In the definition of store-probability expressions v p ∈ (0, 1], c p is a store-probability constant in (0, 1]
defined using the FlyFast const construct, and f n p is a store-probability function defined by the user
in Fϒ , which, when applied to attribute expressions e1 , . . . , em returns a probability value.The syntax for
store-probability function definitions pfd is similar to that of attribute functions:
pfd ::= func f n p (x1 : T 1, . . . , xm : T m) : float; pfb endfunc
pfb ::= p |case (x1 , . . . , xm ) of(va11 , . . . , vam1 ) : p1 ; (va12 , . . . , vam2 ) : p2 ; . . . (va1k , . . . , vamk ) : pk
where f n p is the name of the store-probability function, the result type is float (actually the range [0, 1])
x1 : T 1, . . . , xm : T m are its parameters and their relative types, p, pi are store-probability expressions and
vai j are attribute-values. In any store update definition it must be guaranteed that the values of p1 . . . pn
sum up4 to 1. The informal meaning is clear. The store update will make attributes a1 , . . . , ak take the
values of e1i , . . . , eki respectively with probability equal to the value of pi .
In Figure 1 store-probability functions pHr, pN, pS, pE, pW are defined that give the probabilities of
not moving (pHr), or of jumping to North (pN), South (pS), East (pE), West (pW), as functions of the
3 In
[9] a slightly different syntax for store updates has been used.
this version of the translation we allow only flat updates, i.e. the specific probability of each combination of values
assigned to the attributes must be given explicitly. Other possibilities could be defined using combinations of (independent)
probability distributions.
4 In
Designing Attribute-based FlyFast
98
S
:=
frc (I) :: inf∗ [⊥]hiJump.I +
frc (S) :: nsc∗ [⊥]hiJump.S
I
:=
ii :: inf∗ [⊥]hiJump.I +
ir :: rec∗ [⊥]hiJump.S
Figure 2: SI, a behavioural model.
current location.
Example 1 A simplified version of the behaviour of the epidemic process discussed in [9] is shown in
Figure 2 left5 . In Figure 1 we show a fragment of FSI defining store update Jump together with the relevant
type, constant and function definitions as introduced above. The component has just one attribute, named
loc, with values in Space. The effect of Jump executed by a component in which loc is bound to quadrant
` is to leave the value of loc unchanged with probability pHr(`), change it to the quadrant North of ` with
probability pN(`), and so on. Note that H > L and this implies that higher probability is assigned to A and
C and low probability to B and D. This is represented in Figure 2 right where higher probability locations
are shown in green and lower probability ones are shown in red; moreover, the relevant probabilities are
represented as arrows (H/2, L/2) or self-loops (H, L). A susceptible (state S) component becomes infected
(state I) via an inf action which takes place with probability equal to the fraction of components in the
system which are currently infected (i.e. frc(I)); it remains in state S via the self-loop labelled by action
nsc, with probability frc(S) = 1 − frc(I). An infected node (state I) may recover, entering state S with
action rec and probability ir; while infected, it keeps executing action inf, with probability ii. Note
that, for the sake of simplicity, we use only internal actions, modelled by means of output actions with
predicate false (⊥). We assume that in the initial global state all outboxes are non-empty; each contains
the initial store of the specific component (i.e., its initial location), predicate ⊥ and the empty tuple hi).
A PiFF state defining equation has the following (abstract) form: C := ∑ j∈J [g j ]p j :: act j .C j where either
[g j ]p j is the keyword rest or:
• g j is a boolean expression b which may depend on the current store, but not on the current occupancy measure vector: b ::= > | ⊥ | e ./ e | ¬b | b ∧ b and e ::= va | ca |my.a where > (⊥) denotes
the constant true (false), ./ ∈ {≥, >, ≤, <}, va is an attribute value (drawn from finite set V of
attribute values), ca is an attribute constant in V defined using the FlyFast const construct, and a
is the name of an attribute of the component.
• p j is a transition probability expression p ::= v p | c p | frc (C) | frc (π) | ∏i∈I pi | ∑i∈I pi | 1 − p, for
finite I, where v p ∈ (0, 1], c p a constant in (0, 1] defined via the const construct, and π is defined as
b above, but where expressions e can also be attribute names a (i.e. e ::= va | ca | my.a | a); frc (C)
is the fraction of components currently in state C over the total number N; similarly, frc (π) is the
fraction of components the current store of which satisfies π, over the total number N. Note that it
must be guaranteed that ∏i∈I pi ≤ 1 and ∑i∈I pi ≤ 1.
5 We focus only on those features that are most relevant for the present paper. In [9] also other features are shown like, e.g.
the use of (predicate-based) input actions, which are not the main subject of this paper.
D. Latella & M. Massink
99
• act j can be an output action α ∗ [π]hiσ or an input action α ∗ [π]()σ , where π is as above and σ is
the name of a store update. Note that in the case of an input action, π refers to the store of the
partner component in the previous step of the computation.
If [g j ]p j = rest, then act j must be an output action α ∗ [π]hiσ , to be executed with the residual probability.
2.2
FlyFast
FlyFast accepts a specification h∆, A, C0 i(N) of a model of a system consisting of the clock-synchronous
product of N instances of a probabilistic agent. The states of the DTMC-like agent model are specified
by a set of state-defining equations ∆. The (abstract) form of a state defining equation is the following
C := ∑ri=1 ai .Ci where ai ∈ A—the set of FlyFast actions—C,Ci ∈ S—the set of FlyFast states—and,
for i, j = 1, . . . , r ai 6= a j if i 6= j; note that Ci = C j with i 6= j is allowed instead6 . Each action has a
probability assigned by means of an action probability function definition in A of the form a :: exp where
exp is an expression consisting of constants and frc (C) terms. Constants are floating point values or
names associated to such values using the construct const < name > = < value >; frc (C) denotes the
element associated to state C in the current occupancy measure vector7 . So, strictly speaking, ∆ and A
characterise an inhomogeneous DTMC whose probability matrix K(m) is a function of the occupancy
measure vector m such that for each pair of states C,C0 , the matrix element K(m)C,C0 is the probability of
jumping from C to C0 given the current occupancy measure vector m. Letting S∆ be the set of states of the
agent, with |S∆ | = S, and U S = {(m1 , . . . , mS )|m1 + . . . + mS = 1} denote the unit simplex of dimension
S, we have K : U S × S∆ × S∆ → [0, 1]. Matrix K is generated directly from the input specification
h∆, A, C0 i(N) ; the reader interested in the details of how to derive K is referred to [27, 29]. Auxiliary
function definitions can be specified in A. The initial state C0 is a vector of size N consisting of the initial
state of each individual object. Finally, note that in matrix K(m) the information on specific actions is
lost, which is common in PCTL/DTMC based approaches; furthermore, we note that, by construction,
K(m) does not depend on N (see [27, 29] for details).
3
A revised translation
As in [9], we define a translation such that, given a PiFF system specification ϒ = (∆ϒ , Fϒ , Σ0 )(N) , the
translation returns the FlyFast system specification h∆, A, C0 i(N) preserving probabilistic semantics. The
predicate-based FlyFast front-end is then completed with a simple translation at the PCTL level, for
which we refer to [9].
The system model specification translation consists of two phases. In the first phase, each action
in the input system model specification ϒ is annotated with an identifier which is unique within the
specification. We let ℵ(ϒ) denote the resulting specification. These annotations will make action names
unique specification-wide thus eliminating complications which may arise from multiple occurrences
of the same action, in particular when leading to the same state (see [9] for details). Of course, these
annotations are disregarded in the probabilistic semantics, when considering the interaction model of
components. In other words, an output action αhi in outbox (γ, π, αhi) must match with any input
6 The
concrete FlyFast syntax is: state C{a 1.C 1 + a 2.C 2 ...a r.C r}.
7 The occupancy measure vector is a vector with as many elements as the number of states of an individual agent; the element
associated to a specific state gives the fraction of the subpopulation currently in that state over the size of the overall population.
The occupancy measure vector is a compact representation of the system global state.
Designing Attribute-based FlyFast
100
action α() even if αhi would actually correspond to (α, ι)∗ [π]hi and α() would actually correspond to
(α, η)∗ [π 0 ](). Apart from this detail, the probabilistic semantics as defined in [9] remain unchanged.
The second phase is defined by the translation algorithm defined in Figure 5, which is a revised and
simplified version of that presented in [9] and is applied to ℵ(ϒ). We let I (ℵ(ϒ)) denote the result of
the translation, namely the pure FlyFast system specification h∆, A, C0 i(N) .
We recall here some notation from [9]. We let S∆ϒ denote the set of states of ϒ; Γ∆ϒ is the set of
all stores defined over the attributes of ϒ—a store is a finite mapping from the attributes of the component to a finite set of values V , thus Γ∆ϒ is finite—and O∆ϒ the finite set of all outboxes of ϒ. A
ϒ component-state is a triple (C, γ, O) ∈ S∆ϒ × Γ∆ϒ × O∆ϒ = Ω∆ϒ . If the component-state is the target of a transition modelling the execution of an output action, then O = (γ 0 , π, αhi), where γ 0 is the
store of the (component-state) source of the transition, π is the predicate used in the action—actualised
with γ 0 —and αhi the actual message sent by the action. If, instead, the component-state is the target
of a transition for an input action, then O = hi, i.e. the empty outbox. Note that the set of component
states of ℵ(ϒ) is identical to that of ϒ. Also the set of all stores of ℵ(ϒ) is the same as that of ϒ.
In the algorithm of Figure 5 by t ∗ t 0 we mean the syntactical term representing the product of terms
t and t 0 ; the notation is extended to PROD{t|cond(t)}, denoting the syntactical product t1 ∗ . . . ∗ tn if
{t|cond(t) = tt} = {t1 , . . . ,tn } 6= 0/ and 1 otherwise. Similarly, SUM{t|cond(t)} denotes the syntactical
sum t1 + . . . + tn if {t|cond(t) = tt} = {t1 , . . . ,tn } 6= 0/ and 0 otherwise. The translation algorithm uses a
few auxiliary functions which we briefly discuss below:
• IS : Ω∆ϒ → S is a total injection which maps every component state of ℵ(ϒ) to a distinct state
of I (ℵ(ϒ)); we recall that S denotes the set of state names of FlyFast models.
• IA : (S∆ϒ × Γ∆ϒ ) × (Λ∆ϒ × Iℵ ) × Ω∆ϒ → A is a total injection where, as in [9], Λ∆ϒ is the set of
action labels of ϒ and Iℵ is the set of unique identifiers used in the first phase of the translation.
We recall that A is the set of action names of FlyFast. The mapping of actions is a bit more
delicate because we have to respect FlyFast static constraints and, in particular, we have to avoid
multiple probability function definitions for the same action. A first source of potential violations
(i.e. multiple syntactical occurrences of the same action) has been removed by action annotation in
the first phase of the translation. A second source is the fact that the same action can take place in
different contexts (for example with different stores) or leading to different target component states
(maybe with different probabilities). To that purpose, we could distinguish different occurrences
of the same action in different transitions, each characterised by its source component-state and its
target component-state in Ω∆ϒ . In practice, since an action of a component cannot be influenced
by the current outbox of the component, it is sufficient to restrict the first component of the domain
from Ω∆ϒ to (S∆ϒ × Γ∆ϒ ).
• The interpretation functions defined in Figure 3, namely those depending on stores only (and not
on occupancy measure vectors); we assume EL [[·]]γ extended to EL [[ f n]]γ for defined function f n,
in the standard way. In Figure 3 βϒ denotes the constant to value bindings generated by the const
construct in the input model specification ϒ, whereas store update upd is defined as above.
• The translation function IP for transition probability expressions p j , defined in Figure 4.
Output actions are dealt with in step 1 of the algorithm of Figure 5. Let us consider, for example,
(inf, 1)∗ [⊥]hiJump in the definition of state S in Figure 2 (assuming annotations are integer values and
the action has been annotated with 1). We know that the possible values for locations are A,B,C,D,
so that the set of all stores is {loc} → {A,B,C,D}. The algorithm generates 12 actions8 . Let us focus
8 Diagonal
jumps are not contemplated in the model; technically this comes from the actual probability values used in the
D. Latella & M. Massink
101
EL [[>]]γ
= tt
EL [[⊥]]γ
= ff
EL [[e1 ./ e2 ]]γ
= EL [[e1 ]]γ ./ EL [[e2 ]]γ
EL [[¬b]]γ
= ¬EL [[b]]γ
EL [[b1 ∧ b2 ]]γ = EL [[b1 ]]γ ∧ EL [[b2 ]]γ
EL [[va ]]γ
= va
EL [[ca ]]γ
= βϒ (ca )
EL [[v p ]]γ
= vp
EL [[c p ]]γ
= βϒ (c p )
EL [[a]]γ
= a
EL [[my.a]]γ
= γ(a)
EL [[ f na (e1 , . . . , em )]]γ = EL [[ f na ]]γ (EL [[e1 ]]γ , . . . , EL [[em ]]γ )
EL [[ f n p (e1 , . . . , em )]]γ = EL [[ f n p ]]γ (EL [[e1 ]]γ , . . . , EL [[em ]]γ )
EU [[upd]]γ
= λ γ 0 .dom(γ 0 ) 6= {a1 , . . . , ak } → 0;
γ 0 (a1 ) = EL [[e11 ]]γ ∧ . . . ∧ γ 0 (ak ) = EL [[ek1 ]]γ → EL [[p1 ]]γ ;
..
.
ER [[>]]γ
ER [[⊥]]γ
ER [[e1 ./ e2 ]]γ
ER [[¬b]]γ
ER [[b1 ∧ b2 ]]γ
ER [[va ]]γ
ER [[ca ]]γ
ER [[a]]γ
=
=
=
=
=
=
=
=
γ 0 (a1 ) = EL [[e1n ]]γ ∧ . . . ∧ γ 0 (ak ) = EL [[ekn ]]γ → EL [[pn ]]γ ;
otherwise → 0
tt
ff
ER [[e1 ]]γ ./ ER [[e2 ]]γ
¬ER [[b]]γ
ER [[b1 ]]γ ∧ ER [[b2 ]]γ
va
βϒ (ca )
γ(a)
Figure 3: Interpretation functions relevant for the translation
on the action ξ associated to local position A (i.e. γ = [loc 7→ A]) and possible next position B (i.e.
γ 0 = [loc 7→ B]); the algorithm will generate the FlyFast probability function definition ξ :: pW(A) ∗
(frc (I1) + . . . + frc (In))9 as well as a transition leading to (a state which is the encoding, via IS , of)
the component state with I as (proper) state, store γ 0 , and outbox (γ, ⊥, infhi). Since the action is not
depending on the current outbox, in practice a copy of such a transition is generated for each component
state sharing the same proper state S and the same store γ. The translation scheme for input actions
is defined in case 2 and is similar, except that one has also to consider the sum of the fractions of the
possible partners. The translation of the rest case is straightforward. Note that for every ζ :: r ∗ q ∈ Aγ ,
r is a probability value associated to a store update; since any store update characterizes a probability
distribution over stores, assuming the range of such a distribution is {r1 , . . . , rn } if ζi :: ri ∗ q ∈ Aγ , then
also ζ j :: r j ∗ q ∈ Aγ for all j = 1, . . . , n, j 6= i with ∑ni=1 r j = 1. Thus the remaining probability is
qγ = (1 − SUM{q|ζ :: r ∗ q ∈ Aγ }), where q is either a term IP (p j )γ , with p j occurring in a summand of
the state defining equation (see step 1), or a term IP (p j )γ ∗ SUM{frc (IS (Σ))| . . .} (see step 2). It worth
definition of Jump.
9 Here we assume that I ({(C, γ, O) ∈ Ω |C = I}) = {I1, . . . , In} ⊂ S .
∆ϒ
S
Designing Attribute-based FlyFast
102
IP (v p )γ
IP (c p )γ
IP (frc (C))γ
IP (frc (π))γ
IP (∏i∈I pi )γ
IP (∑i∈I pi )γ
=
=
=
=
=
=
vp
βϒ (c p )
SUM{frc (IS ((C0 , γ 0 , O0 ))) | (C0 , γ 0 , O0 ) ∈ Ω∆ϒ and C0 = C}
SUM{frc (IS ((C0 , γ 0 , O0 ))) | (C0 , γ 0 , O0 ) ∈ Ω∆ϒ and ER [[EL [[π]]γ ]]γ 0 = tt}
PROD{IP (pi )γ | i ∈ I}
SUM{IP (pi )γ | i ∈ I}
Figure 4: Transition probability expressions translation function definition
For each state equation C := ∑ j∈J [g j ]p j :: act j .C j in ∆ϒ :
1. For each output action (α, ι)∗ [π]hiσ = actk with k ∈ J and [gk ]pk 6= rest,
for each γ ∈ Γ∆ϒ s.t. EL [[gk ]]γ = tt and (C, γ, O) ∈ Ω∆ϒ for some O ∈ O∆ϒ , for each γ 0 ∈ Γ∆ϒ s.t. (Ck , γ 0 , (γ, EL [[π]]γ , αhi)) ∈ Ω∆ϒ and
EU [[σ ]]γ (γ 0 ) > 0, let ξ = IA ((C, γ), (αhi, ι), (Ck , γ 0 , (γ, EL [[π]]γ , αhi))) be a fresh new action in the FlyFast model specification
I (ℵ(ϒ)) = h∆, A, C0 i(N) and add the following action probability function definition in A: ξ :: EU [[σ ]]γ (γ 0 ) ∗ IP (pk )γ .
Moreover, for each outbox O ∈ O∆ϒ s.t. (C, γ, O) ∈ Ω∆ϒ , the following summand is added to the equation in ∆ for state
IS ((C, γ, O)): ξ . IS ((Ck , γ 0 , (γ, EL [[π]]γ , αhi)));
2. For each input action (α, ι)∗ [π]()σ = actk , with k ∈ J and [gk ]pk 6= rest,
for each γ ∈ Γ∆ϒ s.t. EL [[gk ]]γ = tt and (C, γ, O) ∈ Ω∆ϒ for some O ∈ O∆ϒ , for each γ 0 ∈ Γ∆ϒ s.t. (Ck , γ 0 , hi) ∈ Ω∆ϒ and
EU [[σ ]]γ (γ 0 ) > 0,
let ξ = IA ((C, γ), (α(), ι), (Ck , γ 0 , hi)), be a fresh new action in the FlyFast model specification I (ℵ(ϒ)) = h∆, A, C0 i(N) and
add the following action probability function definition in A:
ξ :: EU [[σ ]]γ (γ 0 ) ∗ IP (pk )γ ∗
∗SUM{frc (IS (Σ))|Σ = (C00 , γ 00 , (γ, π, αhi)) ∈ Ω∆ϒ ∧
∧ ER [[π]]γ = ER [[EL [[π]]γ ]]γ = tt}.
Moreover, for each outbox O ∈ O∆ϒ s.t. (C, γ, O) ∈ Ω∆ϒ , the following summand is added to the equation in ∆ for state
IS ((C, γ, O)): ξ . IS ((Ck , γ 0 , hi));
3. If there exists k ∈ J s.t. [gk ]pk = rest, and actk = (α, ι)∗ [π]hiσ , for each γ ∈ Γ∆ϒ s.t. (C, γ, O) ∈ Ω∆ϒ for some O ∈
O∆ϒ , let Aγ be the set of probability function definitions which has been constructed in steps (1) and (2) above. Let
qγ be defined by qγ = (1 − SUM{q|ζ :: r ∗ q ∈ Aγ }). For all γ 0 ∈ Γ∆ϒ s.t. (Ck , γ 0 , (γ, EL [[π]]γ , αhi)) ∈ Ω∆ϒ , let ξ =
IA ((C, γ), (α, ι)hi, (Ck , γ 0 , (γ, EL [[π]]γ , αhi))) ∈ Ω∆ϒ , be a fresh new action in the FlyFast model specification I (ℵ(ϒ)) =
h∆, A, C0 i(N) and add the following action probability function definition in A: ξ :: EU [[σ ]]γ (γ 0 ) ∗ qγ .
Moreover, for each outbox O ∈ O∆ϒ s.t. (C, γ, O) ∈ Ω∆ϒ , the following summand is added to the equation in ∆ for state
IS ((C, γ, O)): ξ . IS ((Ck , γ 0 , (γ, EL [[π]]γ , αhi)));
4. No other action probability function definition and transition is included and the initial state C0 of I (ϒ) is defined as C0 =
IS (Σ0 ).
Figure 5: The translation algorithm
pointing out here that the translation of Figure 5 is essentially the same as that presented in [9], when the
latter is applied to the sublanguage of PiFF where one requires that each action occurs at most once. The
annotations performed in the first phase of the translation ensure that this requirement is fulfilled; as we
noted above, these annotations are purely syntactical and are disregarded at the semantics level. We also
recall that in probabilistic, pure DTMC process language semantics, actions are in the end dropped and,
for each pair of states, the cumulative probability of such actions is assigned to the single transition from
one of the states to the other one. Consequently, correctness of the translation, proved in [9], is preserved
by the simplified version presented in this paper.
We note that in the algorithm sets Ω∆ϒ , Γ∆ϒ and O∆ϒ are used. Of course, an alternative approach
could be one which considers only the set Ω∆ϒ of component states which are reachable from a given
initial component state and, consequently, the sets Γ∆ϒ and O∆ϒ of used stores and outboxes. In this way,
the size of the resulting FlyFast model specification would be smaller (for example in terms of number of
D. Latella & M. Massink
103
states). On the other hand, this approach might require recompilation for each model-checking session
starting from a different initial component state.
4
A simplified language for Bisimulation-based optimisation
In this section we consider a simplified language for transition probability expressions appearing in
state defining equations that will allow us to perform bisimulation based optimisation of the result
h∆, A, C0 i(N) . The restricted syntax for transition probability expressions p we use in this section is
the following: p ::= e p | e p · frc (C) | e p · frc (π) and e p ::= v p | c p where v p and c p and π are defined as in
Section 2.
By inspection of the FlyFast translation as defined in Section 3, and recalling that the set S∆ of the
states of the resulting FlyFast model, ranged over by z, zi , . . ., has cardinality S, it is easy to see that the
probability action definition in the result of a translation of a generic output action is either of the form
ξ :: k, or it is of the form ξ :: k ∗ SUM{frc (zi )|i ∈ I} where k is a FlyFast constant. Moreover, if p was
of the form e p · frc (C), then index set I ⊆ {1, . . . , S} identifies those states in S∆ that represent (via IS )
component states with proper local state C; if instead, p was of the form e p · frc (π), then I ⊆ {1, . . . , S}
identifies those states in S∆ that represent (via IS ) component states with a store satisfying π in the
relevant store. At the FlyFast semantics level, recalling that frc (zi ) is exactly the i-th component mi of the
occupancy measure vector m = (m1 , . . . , mS ) of the model, we can rewrite10 the above as k or k · ∑i∈I mi .
Similarly, the probability action definition in the result of a translation of a generic
input action
∗
0
0
(α, ι) [π ]() (executed in local store γ ) will necessarily be of the form k · ∑ j∈I 0 m j or of the form
k · ∑ j∈I m j · ∑ j∈I 0 m j , for index sets I as above and I 0 as follows:
I 0 = {i ∈ {1, . . . , S}|∃C, γ, γ, π, s.t.
zi = IS ((C, γ, (γ, π, αhi))) ∧ ER [[π]]γ 0 = ER [[EL [[π 0 ]]γ 0 ]]γ = tt}
An immediate consequence of using the above mentioned restricted syntax for the probability function definitions is that, letting K : U S ×S ×S → [0, 1] be the transition probability matrix for the FlyFast
translation of a model specification, we have that K(m1 , . . . , mS )z,z0 is a polynomial function of degree at
most 2 in variables m1 , . . . , mS .
5
Bisimilarity and State-space Reduction
The following definition generalises standard probabilistic bisimilarity for state labelled DTMCs to the
case in which transition probabilities are functions instead of constant values.
Definition 1 For finite set of states S, with |S| = S, let K : U S × S × S → [0, 1] and, for z ∈ S and
Q ⊆ S, write K(m)z,Q for ∑z0 ∈Q K(m)z,z0 . Let furthermore L : S → 2AP be a state-labelling function,
for a given set AP of atomic propositions. An equivalence relation R ⊆ S × S is called a bisimulation
relation if and only if z1 R z2 implies: (i) L (z1 ) = L (z2 ) and (ii) K(m)z1 ,Q = K(m)z2 ,Q , for all m ∈ U S
and Q ∈ S/R. The bisimulation equivalence on S is the largest bisimulation relation R ⊆ S × S.
We point out that the notion of bisimilarity does not introduce any approximation, and consequently error,
in a model and related analyses. Bisimilarity is only a way for abstracting from details that are irrelevant
10 With
a little notational abuse using k also as the actual value in [0, 1] of the FlyFast constant k.
Designing Attribute-based FlyFast
104
l
h
H
L
H
L
Figure 6: SI in two quadrants
for the specific analyses of interest. In particular, it is useful to remark that bisimilarity preserves also
state labels, which are directly related to the atomic propositions of logic formulas for which modelchecking is performed. Actually, it is well known that probabilistic bisimilarity coincides with PCTL
equivalence, i.e. the equivalence induced on system states by their satisfaction of PCTL formulas, for
finitely branching systems [3].
Note that K(m1 , . . . , mS )z1 ,Q = K(m1 , . . . , mS )z2 ,Q for all (m1 , . . . , mS ) ∈ U S is in general not decidable. If instead we consider only transition probability matrices as in Section 4, we see that each side
of the above equality is a polynomial function of degree at most 2 in variables m1 , . . . , mS and one can
define a normal form for the polynomial expressions in m1 , . . . , mS supported by an ordering relationon
the variable names (e.g. m1 ≺ . . . ≺ mS ) and get expressions of the general form ∑Si=1 ∑Sj≥i hi j · mi · m j +
hi j , hi , h. Actually, such expressions can always be rewritten in the form
∑Si=1 hi · mi + h for suitable
ui j · mi · m j + u for suitable ui j , u since, recalling that ∑Si=1 mi = 1, we get ∑Si=1 hi · mi =
∑Si=1 ∑Sj≥i
S
an expression of the following
∑Si=1 m
i · ∑i=1 hi · mi which,
by simple algebraic manipulation, yields
S
S
S
S
0
form: ∑i=1 ∑ j≥i ui j · mi · m j ; finally, we get ∑i=1 ∑ j≥i ui j · mi · m j + u by letting ui j = hi j + u0i j and
u = h. The following proposition thus establishes decidability of K(m1 , . . . , mS )z1 ,Q = K(m1 , . . . , mS )z2 ,Q
for all (m1 , . . . , mS ) ∈ U S for transition probability matrices as in Section 4:
Proposition 1
Let A(m1 , . . . , mS ) = ∑Si=1 ∑Sj≥i ai j · mi · m j + a and B(m1 , . . . , mS ) = ∑Si=1 ∑Sj≥i bi j · mi · m j + b with
ai j , bi j , a, b ∈ R, where m1 , . . . , mS are variables taking values over R≥0 with ∑Si=1 mi = 1. The following
holds: (∀m1 , . . . , mS .A(m1 , . . . , mS ) = B(m1 , . . . , mS )) ⇔ ((∀i, j = 1, . . . , S with i ≤ j.ai j = bi j ) ∧ a = b).
The above results can be used for reduction of the state-space of the individual agent, i.e. the resulting
FlyFast model specification, after the application of the translation described in Section 3, by using for
instance the standard probabilistic relational coarsest set partition problem algorithm (see e.g. [22], page
227) with slight obvious modifications due to the presence of state-labels and the need of symbolic
computation capabilities required for checking (degree 2) polynomial expressions equality.11 It is worth
mentioning that state aggregation via bisimilarity is effective only if there is some sort of compatibility
between (i) state labelling—and, consequently, the specific PCTL atomic propositions one uses—and
(ii) the way probabilities are assigned to transitions—and, consequently, the cumulative probabilities to
equivalence classes. We will come back on this issue in the following section.
D. Latella & M. Massink
SA
SB
SC
SD
IA
IB
IC
ID
SA
HφS (m)
H
2 φS (m)
0
H
2 φS (m)
Hir
H
2 ir
0
H
2 ir
105
SB
L
2 φS (m)
LφS (m)
L
2 φS (m)
0
L
2 ir
Lir
L
2 ir
0
SC
0
H
2 φS (m)
HφS (m)
H
2 φS (m)
0
H
2 ir
Hir
H
2 ir
SD
L
2 φS (m)
0
L
2 φS (m)
LφS (m)
L
2 ir
0
L
2 ir
Lir
IA
HφI (m)
H
2 φI (m)
0
H
2 φI (m)
Hii
H
2 ii
0
H
2 ii
IB
L
2 φI (m)
LφI (m)
L
2 φI (m)
0
L
2 ii
Lii
L
2 ii
0
IC
0
H
2 φI (m)
HφI (m)
H
2 φI (m)
0
H
2 ii
Hii
H
2 ii
ID
L
2 φI (m)
0
L
2 φI (m)
LφI (m)
L
2 ii
0
L
2 ii
Lii
Figure 7: IDTMC transition probability matrix K(m), for m in U 8 .
6
Example
The application of the translation to the specification of Example 1 generates an agent model with 8
states, say S∆ = {SA, SB, SC, SD, IA, IB, IC, ID}12 with associated IDTMC probability transition matrix
as shown in Figure 7 where mxy represents the fraction of objects currently in state xy for x ∈ {S, I} and
y ∈ {A, B,C, D}—i.e. the components in state x and with loc = y in the original specification of Fig 2—so
that m = (mSA , mSB , mSC , mSD , mIA , mIB , mIC , mID ) is the occupancy measure vector. In Figure 7, functions
φS and φI are used as abbreviations in the obvious way: φS (m) = mSA + mSB + mSC + mSD and φI (m) =
mIA + mIB + mIC + mID . Let us assume now that we are interested in checking PCTL formulas on the
model of Fig 2 which distinguish components located in A or C from those located in B or D, and
those in state S from those in state I, that is we consider atomic propositions Sh, Ih, Sl and Il and a
labelling L such that L (SA) = L (SC) = {Sh}, L (IA) = L (IC) = {Ih}, L (SB) = L (SD) = {Sl},
and L (IB) = L (ID) = {Il}.
Consider relation R on S∆ defined as R = IS∆ ∪ {(SA, SC), (SB, SD), (IA, IC), (IB, ID)} ∪ {(SC, SA),
(SD, SB), (IC, IA), (ID, IB)} where IS∆ is the identity relation on S∆ . It is very easy to show that R is
a bisimulation according to Definition 1. Clearly R is an equivalence relation and its quotient S∆ /R
is the set {QSh , QSl , QIh , QIl } with QSh = {SA, SC}, QSl = {SB, SD}, QIh = {IA, IC}, QIl = {IB, ID}. In
addition, for all z1 , z2 ∈ S∆ , whenever z1 R z2 , we have L (z1 ) = L (z2 ) and K(m)z1 ,Q = K(m)z2 ,Q for all
Q ∈ S∆ /R and for all m, as one can easily check; clearly, R is also the largest bisimulation relation on S∆ .
The relationship between the two occupancy measure vectors is: mQSh = mSA + mSC , mQSl = mSB + mSD ,
mQIh = mIA + mIC , and mQIl = mIB + mID . We can thus use the reduced IDTMC defined by matrix
b given
b Q , mQ , mQ , mQ ) shown in Figure 9. It corresponds to the FlyFast agent specification ∆
K(m
Sh
Sl
Ih
Il
in Figure 8. In a sense, the high probability locations A and C, in the new model, have collapsed into a
single one, namely h and the low probability ones (B and D) have collapsed into l, as shown in Figure 6.
We point out again the correspondence between the symmetry in the space jump probability on one side
and the definition of the state labelling function on the other side. Finally, note that a coarser labelling
like, e.g. L 0 (SA) = L 0 (SC) = L 0 (IA) = L 0 (IC) = {h}, L 0 (SB) = L 0 (SD) = L 0 (IB) = L 0 (ID) = {l}
would make the model collapse into one with only two states, Qh and Ql , with probabilities H : Qh →
Qh , H : Ql → Qh , L : Qh → Ql and L : Qh → Ql where only the location would be modelled whereas
11 For
instance, on page 227 of [22], line 12, v(x, S) = v(y, S) should be replaced with L(x) = L(y) ∧ v(x, S) = v(y, S) and in
line 13, v(x, S) 6= v(y, S) should be replaced with L(x) 6= L(y) ∨ v(x, S) 6= v(y, S), in order to take state labels into consideration
as well. Of course v(x, S) (L, respectively)is to be intended as K(m)z,Q (L , respectively), using the notation we introduced
above for Bisimilarity.
12 Actually the agent resulting from the translation of Figure 5 has a higher number of states due to the different possibilities
for outbox values. Many of these states are unreachable from the initial state since the agent has no input action and we assume
an initial unreachable state pruning has been performed.
Designing Attribute-based FlyFast
106
action
action
action
action
QSh_inf_QIh:
QSh_inf_QIl:
QSl_inf_QIh:
QSl_inf_QIl:
H*(frc(QIh)+frc(QIl));
L*(frc(QIh)+frc(QIl));
H*(frc(QIh)+frc(QIl));
L*(frc(QIh)+frc(QIl));
action
action
action
action
QSh_nsc_QSh:
QSh_nsc_QSl:
QSl_nsc_QSh:
QSl_nsc_QSl:
H*(frc(QSh)+frc(QSl));
L*(frc(QSh)+frc(QSl));
H*(frc(QSh)+frc(QSl));
L*(frc(QSh)+frc(QSl));
action
action
action
action
QIh_inf_QIh:
QIh_inf_QIl:
QIl_inf_QIh:
QIl_inf_QIl:
H*ii;
L*ii;
H*ii;
L*ii;
action
action
action
action
QIh_rec_QSh:
QIh_rec_QSl:
QIl_rec_QSh:
QIl_rec_QSl:
H*ir;
L*ir;
H*ir;
L*ir;
state
state
state
state
QSh{QSh_inf_QIh.QIh
QSl{QSl_inf_QIh.QIh
QIh{QIh_inf_QIh.QIh
QIl{QIl_inf_QIh.QIh
+
+
+
+
QSh_inf_QIl.QIl
QSl_inf_QIl.QIl
QIh_inf_QIl.QIl
QIl_inf_QIl.QIl
+
+
+
+
QSh_nsc_QSh.QSh
QSl_nsc_QSh.QSh
QIh_rec_QSh.QSh
QIl_rec_QSh.QSh
+ QSh_nsc_QSl.QSl}
+ QSl_nsc_QSl.QSl}
+ QIh_rec_QSl.QSl}
+QIl_rec_QSl.QSl}
b
Figure 8: Reduced agent specification ∆
QSh
QSl
QIh
QIl
QSh
H · (mQSh + mQSl )
H · (mQSh + mQSl )
H · ir
H · ir
QSl
L · (mQSh + mQSl )
L · (mQSh + mQSl )
L · ir
L · ir
QIh
H · (mQIh + mQIl )
H · (mQIh + mQIl )
H · ii
H · ii
QIl
L · (mQIh + mQIl )
L · (mQIh + mQIl )
L · ii
L · ii
b
Figure 9: IDTMC transition probability matrix function K(m),
for m in U 4 .
information on the infection status would be lost.
7
Conclusions
PiFF [9] is a language for a predicate-based front-end of FlyFast, an on-the-fly mean-field modelchecking tool. In this paper we presented a simplified version of the translation proposed in [9] together
with an approach for model reduction that can be applied to the result of the translation. The approach
is based on probabilistic bisimilarity for inhomogeneous DTMCs. An example of application of the procedure has been shown. The implementation of a compiler for PiFF mapping the language to FlyFast
is under development as an add-on of FlyFast. We plan to apply the resulting extended tool to more as
well as more complex models, in order to get concrete insights on the practical applicability of the framework and on the actual limitations imposed by the restrictions necessary for exploiting bisimilarity-based
state-space reduction. Investigating possible ways of relaxing some of such restrictions will also be an
interesting line of research.
Acknowledgments Research partially funded by EU Project n. 600708 A Quantitative Approach to Management and Design of Collective and Adaptive Behaviours (QUANTICOL).
References
[1] Adnan Aziz, Kumud Sanwal, Vigyan Singhal & Robert K. Brayton (2000): Model-checking continous-time
Markov chains. ACM Trans. Comput. Log. 1(1), pp. 162–170, doi:10.1145/343369.343402.
[2] Christel Baier, Boudewijn R. Haverkort, Holger Hermanns & Joost-Pieter Katoen (2003): Model-Checking
Algorithms for Continuous-Time Markov Chains. IEEE Trans. Software Eng. 29(6), pp. 524–541,
doi:10.1109/TSE.2003.1205180.
D. Latella & M. Massink
107
[3] Christel Baier & Joost-Pieter Katoen (2008): Principles of model checking. MIT Press.
[4] Girish Bhat, Rance Cleaveland & Orna Grumberg (1995): Efficient On-the-Fly Model Checking for CTL*.
In: Proceedings, 10th Annual IEEE Symposium on Logic in Computer Science, San Diego, California, USA,
June 26-29, 1995, IEEE Computer Society, pp. 388–397, doi:10.1109/LICS.1995.523273.
[5] L. Bortolussi, G. Cabri, G. Di Marzo Serugendo, V. Galpin, J. Hillston, R. Lanciani, M. Massink & D. Tribastone, M. Weyns (2015): Verification of CAS. In J. Hillston, J. Pitt, M. Wirsing & F. Zambonelli, editors: Collective Adaptive Systems: Qualitative and Quantitative Modelling and Analysis, Schloss Dagstuhl Leibniz-Zentrum für Informatik, Dagstuhl Publishing, Germany, pp. 91–102. Dagstuhl Reports. Vol. 4, Issue
12. Report from Dagstuhl Seminar 14512. ISSN 2192-5283.
[6] Luca Bortolussi & Jane Hillston (2012): Fluid Model Checking. In Maciej Koutny & Irek Ulidowski, editors:
CONCUR 2012 - Concurrency Theory - 23rd International Conference, CONCUR 2012, Newcastle upon
Tyne, UK, September 4-7, 2012. Proceedings, Lecture Notes in Computer Science 7454, Springer, pp. 333–
347, doi:10.1007/978-3-642-32940-1 24.
[7] Luca Bortolussi & Jane Hillston (2015): Model checking single agent behaviours by fluid approximation. Inf.
Comput. 242, pp. 183–226, doi:10.1016/j.ic.2015.03.002.
[8] Jean-Yves Le Boudec, David D. McDonald & Jochen Mundinger (2007): A Generic Mean Field Convergence
Result for Systems of Interacting Objects. In: Fourth International Conference on the Quantitative Evaluaiton
of Systems (QEST 2007), 17-19 September 2007, Edinburgh, Scotland, UK, IEEE Computer Society, pp.
3–18, doi:10.1109/QEST.2007.8.
[9] Vincenzo Ciancia, Diego Latella & Mieke Massink (2016): On-the-Fly Mean-Field Model-Checking for
Attribute-Based Coordination. In Alberto Lluch-Lafuente & José Proença, editors: Coordination Models
and Languages - 18th IFIP WG 6.1 International Conference, COORDINATION 2016, Held as Part of the
11th International Federated Conference on Distributed Computing Techniques, DisCoTec 2016, Heraklion,
Crete, Greece, June 6-9, 2016, Proceedings, Lecture Notes in Computer Science 9686, Springer, pp. 67–83,
doi:10.1007/978-3-319-39519-7 5.
[10] Costas Courcoubetis, Moshe Y. Vardi, Pierre Wolper & Mihalis Yannakakis (1992): Memory-Efficient Algorithms for the Verification of Temporal Properties. Formal Methods in System Design 1(2/3), pp. 275–288,
doi:10.1007/BF00121128.
[11] R.W.R. Darling & J.R. Norris (2008): Differential equation approximations for Markov chains. Probability
Surveys 5, pp. 37–79, doi:10.1214/07-PS121.
[12] Rocco De Nicola, Diego Latella, Alberto Lluch-Lafuente, Michele Loreti, Andrea Margheri, Mieke Massink,
Andrea Morichetta, Rosario Pugliese, Francesco Tiezzi & Andrea Vandin (2015): The SCEL Language:
Design, Implementation, Verification. In Martin Wirsing, Matthias M. Hölzl, Nora Koch & Philip Mayer,
editors: Software Engineering for Collective Autonomic Systems - The ASCENS Approach, Lecture Notes
in Computer Science 8998, Springer, pp. 3–71, doi:10.1007/978-3-319-16310-9 1.
[13] Giuseppe Della Penna, Benedetto Intrigila, Igor Melatti, Enrico Tronci & Marisa Venturini Zilli (2004):
Bounded Probabilistic Model Checking with the Muralpha Verifier. In Alan J. Hu & Andrew K. Martin,
editors: Formal Methods in Computer-Aided Design, 5th International Conference, FMCAD 2004, Austin,
Texas, USA, November 15-17, 2004, Proceedings, Lecture Notes in Computer Science 3312, Springer, pp.
214–229, doi:10.1007/978-3-540-30494-4 16.
[14] Cheng Feng & Jane Hillston (2014): PALOMA: A Process Algebra for Located Markovian Agents. In Gethin
Norman & William H. Sanders, editors: Quantitative Evaluation of Systems - 11th International Conference,
QEST 2014, Florence, Italy, September 8-10, 2014. Proceedings, Lecture Notes in Computer Science 8657,
Springer, pp. 265–280, doi:10.1007/978-3-319-10696-0 22.
[15] Nicolas Gast & Bruno Gaujal (2010): A mean field model of work stealing in large-scale systems. In Vishal
Misra, Paul Barford & Mark S. Squillante, editors: SIGMETRICS 2010, Proceedings of the 2010 ACM
SIGMETRICS International Conference on Measurement and Modeling of Computer Systems, New York,
New York, USA, 14-18 June 2010, ACM, pp. 13–24, doi:10.1145/1811039.1811042.
108
Designing Attribute-based FlyFast
[16] Stefania Gnesi & Franco Mazzanti (2011): An Abstract, on the Fly Framework for the Verification of ServiceOriented Systems. In Martin Wirsing & Matthias M. Hölzl, editors: Rigorous Software Engineering for
Service-Oriented Systems - Results of the SENSORIA Project on Software Engineering for Service-Oriented
Computing, Lecture Notes in Computer Science 6582, Springer, pp. 390–407, doi:10.1007/978-3-642-204012 18.
[17] Guillaume Guirado, Thomas Hérault, Richard Lassaigne & Sylvain Peyronnet (2006): Distribution, Approximation and Probabilistic Model Checking. Electr. Notes Theor. Comput. Sci. 135(2), pp. 19–30,
doi:10.1016/j.entcs.2005.10.016.
[18] Ernst Moritz Hahn, Holger Hermanns, Björn Wachter & Lijun Zhang (2009): INFAMY: An Infinite-State
Markov Model Checker. In Ahmed Bouajjani & Oded Maler, editors: Computer Aided Verification, 21st
International Conference, CAV 2009, Grenoble, France, June 26 - July 2, 2009. Proceedings, Lecture Notes
in Computer Science 5643, Springer, pp. 641–647, doi:10.1007/978-3-642-02658-4 49.
[19] Hans Hansson & Bengt Jonsson (1994): A Logic for Reasoning about Time and Reliability. Formal Asp.
Comput. 6(5), pp. 512–535, doi:10.1007/BF01211866.
[20] Thomas Hérault, Richard Lassaigne, Frédéric Magniette & Sylvain Peyronnet (2004): Approximate Probabilistic Model Checking. In Bernhard Steffen & Giorgio Levi, editors: Verification, Model Checking, and
Abstract Interpretation, 5th International Conference, VMCAI 2004, Venice, January 11-13, 2004, Proceedings, Lecture Notes in Computer Science 2937, Springer, pp. 73–84, doi:10.1007/978-3-540-24622-0 8.
[21] Gerard J. Holzmann (2004): The SPIN Model Checker - primer and reference manual. Addison-Wesley.
[22] D. Huynh & L. Tian (1992): On some equivalence relations for probabilistic processes. Fundamenta Informaticae 17, pp. 211–234.
[23] A. Kolesnichenko, A. Remke & P.-T. de Boer (2012): A logic for model-checking of mean-field models.
Technical Report TR-CTIT-12-11, http://doc.utwente.nl/80267/.
[24] Anna Kolesnichenko, Pieter-Tjerk de Boer, Anne Remke & Boudewijn R. Haverkort (2013): A logic for
model-checking mean-field models. In: 2013 43rd Annual IEEE/IFIP International Conference on Dependable Systems and Networks (DSN), Budapest, Hungary, June 24-27, 2013, IEEE Computer Society, pp. 1–12,
doi:10.1109/DSN.2013.6575345.
[25] Kim G. Larsen & Axel Legay (2016): Statistical Model Checking: Past, Present, and Future. In Tiziana Margaria & Bernhard Steffen, editors: Leveraging Applications of Formal Methods, Verification and Validation:
Foundational Techniques - 7th International Symposium, ISoLA 2016, Imperial, Corfu, Greece, October
10-14, 2016, Proceedings, Part I, Lecture Notes in Computer Science 9952, pp. 3–15, doi:10.1007/978-3319-47166-2 1.
[26] D. Latella (1983): Comunicazione basata su proprietà nei sistemi decentralizzati. [Property-based interprocess communication in decentralized systems] Graduation Thesis. Istituto di Scienze dell’Informazione.
Univ. of Pisa, Italy (in italian).
[27] Diego Latella, Michele Loreti & Mieke Massink (2013): On-the-fly Fast Mean-Field Model-Checking. In
Martı́n Abadi & Alberto Lluch-Lafuente, editors: Trustworthy Global Computing - 8th International Symposium, TGC 2013, Buenos Aires, Argentina, August 30-31, 2013, Revised Selected Papers, Lecture Notes in
Computer Science 8358, Springer, pp. 297–314, doi:10.1007/978-3-319-05119-2 17.
[28] Diego Latella, Michele Loreti & Mieke Massink (2014): On-the-fly Probabilistic Model Checking. In Ivan
Lanese, Alberto Lluch-Lafuente, Ana Sokolova & Hugo Torres Vieira, editors: Proceedings 7th Interaction and Concurrency Experience, ICE 2014, Berlin, Germany, 6th June 2014., EPTCS 166, pp. 45–59,
doi:10.4204/EPTCS.166.6.
[29] Diego Latella, Michele Loreti & Mieke Massink (2015): On-the-fly PCTL fast mean-field approximated model-checking for self-organising coordination.
Sci. Comput. Program. 110, pp. 23–50,
doi:10.1016/j.scico.2015.06.009.
[30] Michele Loreti & Jane Hillston (2016): Modelling and Analysis of Collective Adaptive Systems with CARMA
and its Tools. In Marco Bernardo, Rocco De Nicola & Jane Hillston, editors: Formal Methods for the
D. Latella & M. Massink
109
Quantitative Evaluation of Collective Adaptive Systems - 16th International School on Formal Methods for
the Design of Computer, Communication, and Software Systems, SFM 2016, Bertinoro, Italy, June 20-24,
2016, Advanced Lectures, Lecture Notes in Computer Science 9700, Springer, pp. 83–119, doi:10.1007/9783-319-34096-8 4.
Designing Attribute-based FlyFast
110
A
Appendix
Proof of Proposition 1
⇐: Trivial.
⇒:
We first prove that a = b:
∀m1 , . . . , mS .A(m1 , . . . , mS ) = B(m1 , . . . , mS )
⇒
{Logic}
A(0, . . . , 0) = B(0, . . . , 0)
⇒
{Def. of A(m1 , . . . , mS ) and B(m1 , . . . , mS )}
a=b
Now we prove that aii = bii for i = 1, . . . , S:
∀m1 , . . . , mS .A(m1 , . . . , mS ) = B(m1 , . . . , mS )
⇒
{Take the S − tuple (m̄1 , . . . , m̄S ) where m̄k = 1 if k = i and 0 otherwise}
A(m̄1 , . . . , m̄S ) = B(m̄1 , . . . , m̄S )
⇒
{Def. of A(m1 , . . . , mS ) and B(m1 , . . . , mS )}
aii + a = bii + b
⇒
{a = b (see above)}
aii = bii
Finally we prove that ai j = bi j , for i, j = 1, . . . , S, j > i:
∀m1 , . . . , mS .A(m1 , . . . , mS ) = B(m1 , . . . , mS )
⇒
{(m̃1 , . . . , m̃S ) where m̃k = 0.5 if k ∈ {i, j} and 0 otherwise}
A(m̃1 , . . . , m̃S ) = B(m̃1 , . . . , m̃S )
⇒
{Def. of A(m1 , . . . , mS ) and B(m1 , . . . , mS )}
0.25aii + 0.25ai j + 0.25a j j + a = 0.25bii + 0.25bi j + 0.25b j j + b
⇒
{a = b (see above)}
0.25aii + 0.25ai j + 0.25a j j = 0.25bii + 0.25bi j + 0.25b j j
⇒
{Algebra}
aii + ai j + a j j = bii + bi j + b j j
⇒
{aii = bii and a j j = b j j (see above)}
ai j = bi j
•
| 6cs.PL
|
Multivariable Adaptive Harmonic Steady-State Control for
Rejection of Sinusoidal Disturbances Acting on an Unknown System
Mohammadreza Kamaldar and Jesse B. Hoagg
arXiv:1608.02508v2 [cs.SY] 11 May 2017
Department of Mechanical Engineering, University of Kentucky, Lexington, KY, 40506-0503
Abstract— This paper presents an adaptive harmonic steadystate (AHSS) controller, which addresses the problem of
rejecting sinusoids with known frequencies that act on a completely unknown multi-input multi-output linear time-invariant
system. We analyze the stability and closed-loop performance
of AHSS for single-input single-output systems. In this case,
we show that AHSS asymptotically rejects disturbances. 1
I. I NTRODUCTION
The rejection of sinusoidal disturbances is a fundamental
control objective in many active noise and vibration control applications such as noise cancellation [1], helicopter
vibration reduction [2], and active rotor balancing [3].
For an accurately modeled linear time-invariant (LTI)
system, the internal-model principle can be used to design
a feedback controller capable of rejecting sinusoidal disturbances of known frequencies [4]–[6]. In this case, disturbance rejection is accomplished by incorporating copies of
the disturbance dynamics in the feedback loop.
If, on the other hand, an accurate model of the system
is not available, but the open-loop dynamics are asymptotically stable, then adaptive feedforward cancellation can
be used to accomplish disturbance rejection [7], [8]. One
approach for sinusoidal disturbance rejection is harmonic
steady-state (HSS) control [9], which has been used for
helicopter vibration reduction [2] and active rotor balancing
[3]. To discuss HSS control, let Gyu denote the control-toperformance transfer function, and assume that there is a
single known disturbance frequency ω. Then, HSS control
requires an estimate of Gyu (ω). In the SISO case, the estimate of Gyu (ω), which is a single complex number, must
have an angle within 90◦ of ∠Gyu (ω) to ensure closedloop stability. In the MIMO case, closed-loop stability is
ensured provided that the estimate of Gyu (ω) is sufficiently
accurate. If there are multiple disturbance frequencies, then
estimates are required at each frequency.
For certain applications Gyu (ω) can be difficult to estimate or subject to change. To address this uncertainty,
online estimation methods have been combined with HSS
control [10]–[12]. For example, a recursive-least-squares
identifier is used in [10], [11] to estimate Gyu (ω) in real
time; however, an external excitation signal, which degrades
performance, is required to ensure stability.
1 This paper is the preprint version of the following paper: Kamaldar M,
Hoagg JB, “Multivariable adaptive harmonic steady-state control for rejection of sinusoidal disturbances acting on an unknown system,” American
Control Conference (ACC), 2016. IEEE (pp. 1631-1636), which has been
published by IEEE at DOI: 10.1109/ACC.2016.7525150, and is subject to
IEEE copyright ploicy.
In this paper, we present a new adaptive harmonic steadystate (AHSS) controller, which is effective for rejecting
sinusoids with known frequencies that act on a completely
unknown MIMO LTI system. We analyze the stability and
closed-loop performance for SISO systems. We show that
AHSS asymptotically rejects disturbances.
The new AHSS algorithm in this paper is a frequencydomain method, and all computations are with discrete
Fourier transform (DFT) data. The AHSS algorithm including DFT is demonstrated on a simulation of an acoustic
duct.
II. N OTATION
Let F be either R or C. Let x(i) denote the ith element
of x ∈ Fn , and let A(i,j) denote the element in row i and
column j of A ∈ Fm×n . Let k · k be the 2-norm on Fn .
Next, let A∗ denote the complex √conjugate transpose of
A ∈ Fm×n , and define kAkF , tr A∗ A, which is the
Frobenius norm of A ∈ Fm×n .
Let spec(A) , {λ ∈ C : det(λI − A) = 0} denote
the spectrum of A ∈ Fn×n , and let λmax (A) denote the
maximum eigenvalue of A ∈ Fn×n , which is Hermitian
positive semidefinite. Let ∠λ denote the argument of λ ∈ C
defined on the interval (−π, π] rad. Let OLHP, ORHP,
and CUD denote the open-left-half plane, open-right-half
plane, and closed unit disk in C, respectively. Define N ,
{0, 1, 2, · · · } and Z+ , N\{0}.
III. P ROBLEM F ORMULATION
Consider the system
ẋ(t) = Ax(t) + Bu(t) + D1 d(t),
(1)
y(t) = Cx(t) + Du(t) + D2 d(t),
(2)
where t ≥ 0, x(t) ∈ Rn is the state, x(0) = x0 ∈
Rn is the initial condition, u(t) ∈ Rm is the control, y(t) ∈ R` is the measured performance, d(t) ∈
Rp is the unmeasured disturbance, and A ∈ Rn×n
is asymptotically stable. Define the transfer functions
Gyu (s) , C(sI − A)−1 B + D, and Gyd (s) , C(sI −
A)−1 D1 + D2 . Let ω1 , ω2 ,P
· · · , ωq > 0, and consider the
q
tonal disturbance d(t) =
i=1 dc,i cos ωi t + ds,i sin ωi t,
where dc,1 , · · · , dc,q , ds,1 , · · · , ds,q ∈ Rp .
Our objective is to design a control u that reduces or even
eliminates the effect of the disturbance d on the performance
y. We seek to design a control that relies on no model
information of (1) and (2), and requires knowledge of only
the disturbance frequencies ω1 , · · · , ωq .
For simplicity, we focus on the case where d is the singletone disturbance d(t) = dc cos ωt + ds sin ωt. However, the
adaptive controller presented in this paper generalizes to the
case where d consists of multiple tones. We address multiple
tones in Example 3.
For the moment, assume that Gyu , Gyd , dc , and ds
are known, and consider the harmonic control u(t) =
uc cos ωt + us sin ωt, where uc , us ∈ Rm . Define û ,
uc − us , which is the value at frequency ω of the DFT
obtained from a sampling of u. The HSS performance of
(1) and (2) withcontrol û is
yhss (t, û) , Re M∗ û + dˆ cos ωt − Im M∗ û + dˆ sin ωt,
(3)
ˆ
where M∗ , Gyu (ω) ∈ C
and d , Gyd (ω)(dc −
ds ) ∈ C` . The HSS performance yhss is the steady-state
response of y, that is, limt→∞ [yhss (t, û) − y(t)] = 0 [13,
Chap. 12.12]. Consider the cost function
Zt
1
T
yhss
(τ, û)yhss (τ, û) dτ,
(4)
J(û) , lim
t→∞ t
`×m
0
which is the average power of yhss . Define
ˆ
ŷhss (û) , M∗ û + d,
(5)
which is the value at frequency ω of the DFT obtained
from a sampling of yhss . It follows from (3)–(5) that
∗
(û)ŷhss (û). The following result provides an
J(û) = 12 ŷhss
expression for an open-loop control û = u∗ that minimizes
J. The proof is omitted due to space limitations.
Theorem 1. Consider the cost function (4), and assume
rank M∗ = min{`, m}. Then, the following statements
hold:
−1
ˆ
i) Assume ` > m, and define u∗ , − (M∗∗ M∗ ) M∗∗ d.
−1
∗ ˆ
∗
Then, ŷhss (u∗ )
=
I` − M∗ (M∗M∗ ) M∗ d,
ˆ and for all
J(u∗ ) = 12 dˆ∗ I` − M∗ (M∗∗ M∗ )−1 M∗∗ d,
m
û ∈ C \{u∗ }, J(u∗ ) < J(û).
ˆ Then,
ii) Assume ` = m, and define u∗ , −M∗−1 d.
m
ŷhss (u∗ ) = 0, J(u∗ ) = 0, and for all û ∈ C \{u∗ },
J(u∗ ) < J(û).
iii) Assume ` < m, and let u∗ ∈ {−M∗∗ (M∗ M∗∗ )−1 dˆ +
(Im − M∗∗ (M∗ M∗∗ )−1 M∗ )v : v ∈ Cm }. Then,
ŷhss (u∗ ) = 0 and J(u∗ ) = 0.
Theorem 1 provides an expression for a control u∗ that
ˆ
minimizes J, but u∗ requires knowledge of M∗ and d.
In this paper, we consider a sinusoidal control with
frequency ω but where the amplitude and phase are updated
at discrete times. Let Ts > 0 be the update period, and
for each k ∈ Z+ , let uk ∈ Cm be determined from an
adaptive law presented later. Then, for each k ∈ N and for
all t ∈ [kTs , (k + 1)Ts ), consider the control
u(t) = (Re uk ) cos ωt − (Im uk ) sin ωt.
(6)
Let yk ∈ C` denote the value at frequency ω of the DFT
of the sequence obtained by sampling y on the interval [(k−
1)Ts , kTs ). If Ts is sufficiently large relative to the settling
time of Gyu , then yk+1 ≈ ŷhss (uk ). For the remainder of
this paper, we assume yk+1 = ŷhss (uk ), and it follows from
(5) that
ˆ
yk+1 = M∗ uk + d.
(7)
In addition, we assume rank M∗ = min{`, m}.
IV. H ARMONIC S TEADY-S TATE C ONTROL
In this section, we review HSS control, which relies on
knowledge of an estimate Me ∈ C`×m of M∗ . Let ρ > 0,
and for all k ∈ N, consider the control
uk+1 = uk − ρMe∗ yk+1 ,
(8)
m
where u0 ∈ C is the initial condition. It follows from
(7) that yk+1 = M∗ uk + dˆ = M∗ uk + yk − M∗ uk−1 , and
substituting (8) yields the closed-loop dynamics
yk+1 = yk − ρM∗ Me∗ yk ,
(9)
+
ˆ
where k ∈ Z and y1 = M∗ u0 + d. Define Λ ,
spec(Me∗ M∗ )∩spec(M∗ Me∗ ). The following result presents
the stability properties of the closed-loop system (9). The
proof is omitted due to space limitations.
Theorem 2. Consider the closed-loop system (9), which
consists of (7) and (8). Assume that Λ ⊂ ORHP, and
assume that ρ satisfies
2Re λ
.
(10)
0 < ρ < min
λ∈Λ |λ|2
Then, for all u0 ∈ Cm , u∞ , limk→∞ uk exists and
y∞ , limk→∞ yk exists. Furthermore, for all u0 ∈ Cm ,
the following statements hold:
−1
i) If ` > m, then u∞ = − (Me∗ M∗ ) Me∗ dˆ and y∞ =
−1
∗ ˆ
∗
[I` − M∗ (Me M∗ ) Me ]d.
ii) If ` = m, then u∞ = −M∗−1 dˆ and y∞ = 0.
iii) If ` < m, then u∞ = u0 − Me∗ (M∗ Me∗ )−1 (M∗ u0 + d)
and y∞ = 0.
Theorem 2 relies on the condition that Λ ⊂ ORHP. This
condition depends on the estimate Me of M∗ . In the SISO
case, Λ ⊂ ORHP if and only if Me is within 90◦ of M∗ ,
that is, |∠(Me /M∗ )| < π2 . In this case, (10) is satisfied by
a sufficiently small ρ > 0.
If Me = M∗ , then Λ ⊂ ORHP. In this case, (10) is
satisfied if ρ < 2/λmax (M∗∗ M∗ ).
If Λ∩OLHP is not empty, then for all ρ > 0, I` −ρM∗ Me∗
has at least one eigenvalue outside the CUD. In this case,
(9) implies that yk diverges.
V. A DAPTIVE H ARMONIC S TEADY-S TATE C ONTROL
In this section, we present AHSS control, which does
not require any information regarding M∗ . Let µ ∈ (0, 1],
ν1 > 0, and u0 ∈ Cm , and for all k ∈ N, consider the
control
µ
M ∗ yk+1 ,
(11)
uk+1 = uk −
ν1 + kMk k2F k
where Mk ∈ C`×m is an estimate of M∗ obtained from the
adaptive law presented below. Note that (11) is reminiscent
of the HSS control (8) except the fixed estimate Me is
replaced by the adaptive estimate Mk , and the fixed gain
ρ is replaced by the Mk -dependent gain µ/ ν1 + kMk k2F .
To determine the adaptive law for Mk , consider the cost
function J : R`×m × R`×m → [0, ∞) defined by
1
2
J(Mr , Mi ) , k(Mr + Mi )(uk − uk−1 ) − (yk+1 − yk )k .
2
Note that J(Re M∗ , Im M∗ ) = 0, that is, M∗ minimizes J.
Define the complex gradient
∂J(Mr , Mi )
∂J(Mr , Mi )
∇J(Mr , Mi ) ,
+
∂Mr
∂Mi
= [(Mr + Mi )(uk − uk−1 ) − (yk+1 − yk )]
× (uk − uk−1 )∗ ,
(12)
which is the direction of the maximum rate of change of
J with respect to Mr + Mi [14]. Let M0 ∈ C`×m \{0},
γ ∈ (0, 1], and ν2 > 0, and for all k ∈ Z+ , consider the
adaptive law
Mk , Mk−1 − ηk ∇J(Re Mk−1 , Im Mk−1 ),
(13)
where
γ(ν1 + kMk−1 k2F )2
(14)
ηk ,
2.
ν2 µ2 + (ν1 + kMk−1 k2F )2 kuk − uk−1 k
Using (12)–(14), it follows
that, for all k ∈ N,
h
i
Mk = Mk−1 − ηk Mk−1 (uk − uk−1 ) − (yk+1 − yk )
× (uk − uk−1 )∗ .
(15)
Thus, the AHSS control is given by (11), (14), and (15).
The control architecture is shown in Fig. 1. All AHSS
computations are performed using complex DFT signals. At
time kTs , the control u is updated using (6) and the complex
signal uk . Note that uk is calculated using yk , which is
the DFT of y at frequency ω sampled over the interval
[(k − 1)Ts , kTs ), which corresponds to the time between
the k − 1 and k steps.
The update period Ts must be sufficiently large such that
the harmonic steady-state assumption yk+1 ≈ ŷhss (uk ) is
valid. Numerical testing suggests that Ts should be at least
as large as the settling time associated with the slowest mode
of A, that is, Ts > 4/(ζωn ), where ζ and ωn are the damping
ratio and natural frequency of the slowest mode of A.
The AHSS controller parameters are µ ∈ (0, 1], γ ∈
(0, 1], ν1 > 0, and ν2 > 0. The gains µ and γ influence the
step size of the uk and Mk update equations, respectively.
The gain ν1 and ν2 influence the normalization of the uk
and Mk update equations, respectively.
VI. S TABILITY A NALYSIS
The following result provides stability properties of the
estimator (15). The proof follows from direct computation
and is omitted due to space limitation.
Proposition 1. Consider the open-loop system (7), and
the AHSS control (11), (14), and (15), where µ ∈ (0, 1],
γ ∈ (0, 1], ν1 > 0, and ν2 > 0. Then, for all u0 ∈ C
and M0 ∈ C\{0}, the estimate Mk is bounded, and for all
k ∈ Z+ ,
kMk − M∗ k2F − kMk−1 − M∗ k2F
∗
γk(Mk−1 − M∗ )Mk−1
yk k 2
≤−
. (16)
∗ y k2
ν2 + kMk−1
k
Proposition 1 implies that kMk − M∗ k2F is nonincreasing.
We now analyze closed-loop performance under the assumption that the open-loop system is SISO. Define u∗ ,
ˆ ∗ , which exists because M∗ 6= 0. Note that if uk ≡
−d/M
u∗ , then yk ≡ 0. Next, (7) implies that yk+1 = M∗ uk + dˆ =
M∗ uk + yk − M∗ uk−1 , and substituting (11) yields
∗
µM∗ Mk−1
yk
,
(17)
yk+1 = yk −
ν1 + |Mk−1 |2
ˆ Furthermore, (14) and
where k ∈ Z+ and y1 = M∗ u0 + d.
(15) can be written as
γ
∗
Mk = Mk−1 −
Mk−1 Mk−1
yk
ν2 + |Mk−1 |2 |yk |2
Fig. 1: AHSS control architecture.
ν1 + |Mk−1 |2
+
(yk+1 − yk ) yk∗ Mk−1 . (18)
µ
Define M , C\{x ∈ C : |∠x − ∠M∗ | = π}, which is
the set of all complex numbers except those numbers that
are exactly 180◦ from M∗ . The following result provides
the closed-loop SISO stability properties. The proof is in
Appendix A.
Theorem 3. Consider the closed-loop system (17) and
(18), which consists of (7), (11), (14), and (15), where
µ ∈ (0, 1], γ ∈ (0, 1], ν1 > 0, and ν2 > 0. Then,
(yk , Mk ) ≡ (0, M∗ ) is a Lyapunov stable equilibrium of
(17) and (18). Furthermore, for all initial conditions u0 ∈ C
and M0 ∈ M\{0}, Mk is bounded, limk→∞ uk = u∗ and
limk→∞ yk = 0.
VII. N UMERICAL E XAMPLES
Consider the acoustic duct of length L = 2 m shown
in Fig. 2, where all measurements are from the left end of
the duct. A disturbance speaker is at ξd = 0.95 m, while 2
control speakers are at ξψ1 = 0.4 m and ξψ2 = 1.25 m. All
speakers have cross-sectional area As = 0.0025 m2 . The
equation for the acoustic duct is
∂ 2 p(ξ, t)
1 ∂ 2 p(ξ, t)
=
+ ρ0 ψ̇1 (t)δ(ξ − ξψ1 )
2
2
c
∂t
∂ξ 2
˙
+ ρ0 ψ̇2 (t)δ(ξ − ξψ2 ) + ρ0 d(t)δ(ξ
− ξd ),
where p(ξ, t) is the acoustic pressure, δ is the Dirac delta,
c = 343 m/s is the phase speed of the acoustic wave, ψ1 and
ψ2 are the speaker cone velocities of the control speakers, d
is the speaker cone velocity of the disturbance speaker, and
ρ0 = 1.21 kg/m2 is the equilibrium density of air at room
conditions. See [15] for more details.
Using separation of variables and retaining r modes,
the
Pr solution p(ξ, t) is approximated by p(ξ, t) =
pi=0 qi (t)Vi (ξ), where for i = 1, · · · , r, Vi (ξ) ,
c 2/L sin iπξ/L, and qi satisfies the differential equation
(1), where
Rt
T
Rt
x(t) =
,
q (σ)dσ q1 (t) · · ·
q (σ)dσ qr (t)
0 1
0 r
i
h
i
h
0
1
0
1
2
A = diag
, · · · , −ωn2r −2ζr ωnr
,
−ωn
−2ζ1 ωn
1
1
T
ρ0 0 V1 (ξψ1 ) · · · 0 Vr (ξψ1 )
B=
,
As 0 V1 (ξψ2 ) · · · 0 Vr (ξψ2 )
ρ0
T
0 V1 (ξd ) · · · 0 Vr (ξd )
D1 =
,
As
d
"
N/m2
!
y
u (m/s)
−200
5
0
−5
0
4
8
0
4
t (s)
8
t (s)
Fig. 3: For a SISO plant with |∠M0 − ∠M∗ | < 90◦ , both HSS
and AHSS yield y(t) → 0 as t → ∞. Dashed lines show ±|u∗ |.
"
HSS
200
0
ψ2
φ1
φ2
Feedback
microphone
Feedback
microphone
y
ψ1
0
AHSS
!
Disturbance Control
speaker speaker
AHSS
200
−200
u (m/s)
Control
speaker
HSS
N/m2
and for i = 1, · · · , r, ωni , iπc/L is the natural frequency
of the ith mode, and ζi = 0.2 is the assumed damping ratio
of the ith mode.
Two feedback microphones are in the duct at ξφ1 =
0.3 m and ξφ2 = 1.7 m, and they measure the acoustic pressures φ1 (t) = p(ξφ1 , t) and φ2 (t) = p(ξφ2 , t),
respectively. Thus, for i = 1, 2, φi (t) = Ci x(t), where
ρ0
[ 0 V1 (ξφi ) · · · 0 Vr (ξφi ) ]. For all examCi = A
s
ples, r = 5 and x(0) = 0. The DFT is performed using a 1
kHz sampling frequency. The HSS and AHSS parameters
are Ts = 0.1 s, u0 = 0, µ = γ = 0.2, ν1 = ν2 =
0.1kM0 k2F , ρ = µ/(ν1 + kM0 k2F ), and Me = M0 , where
M0 is specified in each example. The following examples
consider the acoustic duct with different control speaker and
feedback microphone configurations. Let ω1 = 251 rad/s
and ω2 = 628 rad/s.
5
0
−5
ξφ1
ξψ1
0
4
8
0
t (s)
ξd
4
8
t (s)
Fig. 4: For a SISO plant with |∠M0 − ∠M∗ | > 90◦ , the response
y with HSS diverges, whereas AHSS yields y(t) → 0 as t → ∞.
Dashed lines show ±|u∗ |.
ξψ2
ξφ2
L
40
Fig. 2: Acoustic duct.
M0
Mk
M∗
M 61
30
20
Im M
Example 1. SISO (m = ` = 1). Let u = ψ1 , y = φ1 ,
ψ2 = 0, and d = sin ω1 t + 2 cos ω1 t. First, consider
the case where M0 is within 90◦ of the M∗ , specifically,
π
M0 = 2e 3 M∗ . Figure 3 shows y and u for HSS and AHSS.
The control is turned on after 1 s. Both HSS and AHSS yield
2π
asymptotic disturbance rejection. Next, let M0 = 2e 3 M∗
◦
which is not within 90 of M∗ . Figure 4 shows y and u
for HSS and AHSS. In this case, y with HSS diverges,
whereas y with AHSS converges to zero. Figure 5 shows
the trajectory of the estimate Mk , which moves toward
M∗ . Proposition 1 states that |Mk − M∗ | is nondecreasing;
however, this result assumes that y reaches harmonic steady
state. Figure 5 shows that Mk − M∗ may increase slightly
in practice but generally decreases.
4
Example 2. Single-input two-output (m = 1 and ` =
2). Let u = ψ1 , y = [ φ1 φ2 ]T , ψ2 = 0, and
d = sin ω1 t + 2 cos ω1 t. First, consider the case where
M0 is selected such that (10) is satisfied, specifically,
π
π
M0 = [ 1.5e 4 (M∗ )(1,1) 0.5e 3 (M∗ )(2,1) ]T . Note that
the optimal control is u∗ = −1.66 + 0.98, which minimizes the average power (4). Figure 6 shows y and u
for HSS and AHSS. The control is turned on after 1 s.
In this case, HSS and AHSS each yield uk → u∗ as
k → ∞. Thus, limk→∞ kyk k is minimized. Next, let M0 =
3π
2π
[ 1.5e 4 (M∗ )(1,1) 0.5e 3 (M∗ )(2,1) ]T , which does not
10
0
−10
−60
−50
−40
−30
−20 −10
Re M
0
10
20
Fig. 5: Trajectory of Mk with AHSS for a SISO plant where
|∠(M0 /M∗ )| > π2 . The dashed line shows the locus of M
such that |∠(M/M∗ )| = π2 , which is HSS stability boundary
for Me . Selection of Me = M0 from the lower region, where
|∠(M/M∗ )| > π2 , results in an unstable response with HSS,
whereas AHSS yields asymptotic disturbance rejection for all
M0 ∈ M.
satisfy (10). Figure 6 shows y and u for HSS and AHSS.
In this case, y with HSS diverges, whereas y with AHSS
converges and uk → u∗ as k → ∞, which implies that
limk→∞ kyk k is minimized.
4
Example 3. MIMO (m = 2 and ` = 2) with a twotone disturbance. Let u = [ ψ1 ψ2 ]T , y = [ φ1 φ2 ]T ,
and d = sin ω1 t + sin ω2 t + cos ω1 t + cos ω2 t, which is
a two-tone disturbance. Define M∗,1 , Gyu (ω1 ), and
M∗,2 , Gyu (ω2 ). Since d has 2 tones, we use 2 copies of
the HSS or AHSS algorithm—one copy at each disturbance
AHSS
!
"
y(1) N/m2
HSS
150
0
!
"
y(2) N/m2
−150
300
0
−300
u(1) (m/s)
!
"
y(2) N/m2
!
"
y(1) N/m2
HSS
u (m/s)
5
0
0
2
4
t (s)
6
0
2
4
t (s)
6
Fig. 6: For a single-input two-output plant satisfying (10), both
HSS and AHSS minimize limk→∞ kyk k. Dashed lines show
±|u∗ |.
u(2) (m/s)
−5
300
0
−300
300
0
−300
1
0
−1
2
0
−2
0
4
8
.
AHSS
16
0
4
HSS
!
"
y(1) N/m2
0
−150
300
−300
0
−5
0
2
4
t (s)
6
0
2
4
t (s)
6
Fig. 7: For a single-input two-output plant that does not satisfy
(10), the response y with HSS diverges, whereas AHSS minimizes
limk→∞ kyk k. Dashed lines show ±|u∗ |.
frequency. Let M1,0 and M2,0 denote the initial estimates
of M∗,1 and M∗,2 . First, consider the case where M1,0 and
M2,0 are such that (10) is satisfied, specifically, M1,0 =
π
π
0.6e 6 M∗,1 , and M2,0 = 0.9e 3 M∗,2 . Figure 8 shows y
and u for HSS and AHSS. The control is turned on after 1 s.
Both HSS and AHSS yield asymptotic disturbance rejection.
π
Next, consider the case where M1,0 = 0.2e 7 M∗,1 , and
π
14
M2,0 = 0.6e M∗,2 , which do not satisfy (10). Figure 9
shows y and u for HSS and AHSS. In this case, y with HSS
diverges, whereas y with AHSS converges to zero.
4
A PPENDIX A: P ROOF OF T HEOREM 3
Proof. Define M̃k , Mk − M∗ , VM (M̃k ) , |M̃k |2 ,
and ∆VM (k) , VM (M̃k ) − VM (M̃k−1 ). It follows from
Proposition 1 that for all k ∈ Z+
γ|Mk−1 |2 |yk |2 |M̃k−1 |2
∆VM (k) ≤ −
.
(19)
ν2 + |Mk−1 |2 |yk |2
u(1) (m/s )
!
"
y(2) N/m2
0
5
u (m/s)
12
t (s)
8
12
t (s)
16
Fig. 8: For a MIMO plant that satisfies (10) with a 2-tone
disturbance, both HSS and AHSS yield y(t) → 0 as t → ∞.
150
u(2) (m/s )
!
"
y(2) N/m2
!
"
y(1) N/m2
HSS
AHSS
AHSS
300
0
−300
300
0
−300
1
0
−1
2
0
−2
0
4
8
12
t (s)
16
0
4
8
12
t (s)
16
Fig. 9: For a MIMO plant that does not satisfy (10) with a 2-tone
disturbance, the response y with HSS diverges, whereas AHSS
yields y(t) → 0 as t → ∞.
Next, define Vy (yk ) , |yk |2 and ∆Vy (k) , Vy (yk+1 ) −
Vy (yk ). Evaluating ∆Vy (k) along the trajectories of (17)
yields
µ|yk |2
∗
∆Vy (k) = −
2Re M∗ Mk−1
ν1 + |Mk−1 |2
µ|M∗ |2 |Mk−1 |2
−
,
(20)
ν1 + |Mk−1 |2
∗
Note that |M̃k−1 |2 = |Mk−1 |2 + |M∗ |2 − 2Re M∗ Mk−1
,
and it follows from (20) that
µ|yk |2
∆Vy (k) = −
|Mk−1 |2 + |M∗ |2 − |M̃k−1 |2
ν1 + |Mk−1 |2
µ|M∗ |2 |Mk−1 |2
.
(21)
ν1 + |Mk−1 |2
Define the Lyapunov function V (yk , M̃k−1 ) , ln(1 +
aVy (yk )) + bVM (M̃k−1 ), where a , (|M0 | + 2|M∗ |)2 /ν2 ,
and b > 0 is provided later. Consider the Lyapunov difference
∆V (k) , V (yk+1 , M̃k ) − V (yk , M̃k−1 ).
(22)
Since for all x > 0, ln x ≤ x − 1, evaluating ∆V along
the trajectories of (17)
and (18) yields
a∆Vy (k)
∆V (k) = ln 1 +
+ b∆VM (k)
1 + aVy (yk )
a∆Vy (k)
+ b∆VM (k).
(23)
≤
1 + aVy (yk )
Substituting (19) and (21) into
(23) yields
−
aµ|yk |2 |M∗ |2 −
∆V (k) ≤ −
µ|M∗ |2 |Mk−1 |2
ν1 +|Mk−1 |2
(1 + a|yk |2 )(ν1 + |Mk−1 |2 )
aµ|yk |2 |M̃k−1 |2
+
(1 + a|yk |2 )(ν1 + |Mk−1 |2 )
|Mk−1 |2 |yk |2 |M̃k−1 |2
.
(24)
− bγ
ν2 + |Mk−1 |2 |yk |2
Since for all k ∈ Z+ , ∆VM (k) ≤ 0, it follows that
|Mk−1 | ≤ |M̃k−1 | + |M∗ | ≤ |M̃0 | + |M∗ | ≤ |M0 | + 2|M∗ |,
and thus |Mk−1 |2 ≤ aν2 . Therefore, since µ ∈ (0, 1], it
follows from (24) that
c1 |yk |2
aµ|yk |2 |M̃k−1 |2
∆V (k) ≤ −
+
1 + a|yk |2
ν1 (1 + a|yk |2 )
2
|Mk−1 | |yk |2 |M̃k−1 |2
,
(25)
− bγ
ν2 + aν2 |yk |2
2
1 |M∗ |
where c1 , aµν
(ν1 +aν2 )2 > 0.
To show that (0, M∗ ) is a Lyapunov stable equilibrium,
define D , C × {x ∈ C : |x| < |M∗ |/2}, and note that for
2
all (yk , M̃k−1 ) ∈ D, |Mk−1 | ≥ |M∗ |/2. Let b , γν4aµν
2,
1 |M∗ |
and it follows from (25) that for all (yk , M̃k−1 ) ∈ D,
∆V (k) ≤ −c1 |yk |2 /(1 + a|yk |2 ), which is nonpositive.
Therefore, (yk , Mk−1 ) ≡ (0, M∗ ) is a Lyapunov stable
equilibrium.
To show convergence of yk and uk , let M0 ∈ M\{0},
and define (
|M∗ |,
if M0 = M∗ ,
c2 ,
(26)
∗
|Im M̃0 M∗ |/|M̃0 |, if M0 6= M∗ .
First, assume M0 = M∗ , and it follows that for all k ∈ N,
Mk = M∗ . In this case, for all k ∈ N, |Mk | = c2 .
Next, assume M0 6= M∗ . Since M0 ∈ M\{0}, it
follows that, ∠M∗ − ∠M0 6= π, which implies that
∠(M0 /M∗ ) 6= π. Thus, Im(M0 /M∗ ) 6= 0, which implies
that Im M̃0 M∗∗ = Im (M̃0 M∗∗ + M∗ M∗∗ ) = Im M0 M∗∗ =
|M∗ |2 Im(M0 /M∗ ) 6= 0, and it follows from (26) that
c2 > 0. Next, note
that it follows2 from2 (17)
and (18) that
γ|Mk−1 | |yk |
M̃k = 1 −
M̃k−1 ,
ν2 + |Mk−1 |2 |yk |2
which has the solution M̃k = βk M̃0 , where βk ,
Qk−1
γ|Mi |2 |yi+1 |2
. Thus, for all k ∈ Z+ ,
1
−
i=0
ν2 + |Mi |2 |yi+1 |2
|Mk |2 = |M̃k + M∗ |2 = |M̃0 βk + M∗ |2 = |M̃0 |2 βk2 +
2(Re M̃0 M∗∗ )βk + |M∗ |2 . Since f (x) , |M̃0 |2 x2 +
2(Re M̃0 M∗∗ )x + |M∗ |2 is quadratic and positive definite
−Re M̃0 M∗∗
in x, it follows that f is minimized at
. Thus,
|M̃0 |2
for all k ∈ N,
2
|M∗ |2 |M̃0 |2 − Re M̃0 M∗∗
= c22 .
|Mk |2 ≥
|M̃0 |2
aµν2
Thus, for all k ∈ N, |Mk | > c2 . Let b , γν
2 , and it
1 c2
follows from (25) that
c1 |yk |2
∆V (k) ≤ −
.
(27)
1 + a|yk |2
Next, since V is positive-definite, and for all k ∈ N,
∆V (k) is nonpositive, it follows from (22) and (27) that
Pk
Pk
c1 |yi |2
0 ≤ limk→∞ i=1 1+a|y
2 ≤ − limk→∞
i=1 ∆V (i) =
i|
V (y1 , M̃0 ) − limk→∞ V (yk , M̃k−1 ) ≤ V (y1 , M̃0 ), where
the upper and lower bounds imply that all the limc1 |yk |2
= 0. Since in addiits exist. Thus, limk→∞ 1+a|y
2
k|
2
c1 |yk |
tion, 1+a|y
2 is a positive-definite function of yk , it folk|
lows that limk→∞ yk = 0. Furthermore, (7) implies that
ˆ
limk→∞ uk = limk→∞ (yk+1 − d)/M
∗ = u∗ .
R EFERENCES
[1] S. J. Elliott and P. A. Nelson. Active noise control. IEEE Sig. Proc.
Mag., 10(4):12–35, 1993.
[2] P. P. Friedmann and T. A. Millott. Vibration reduction in rotorcraft
using active control: A comparision of various approaches. J.
Guidance, Control, and Dynamics, 18(4):664–673, 1995.
[3] C. R. Knospe, R. W. Hope, S. M. Tamer, and S. J. Fedigan.
Robustness of adaptive unbalance control of rotors with magnetic
bearings. J. Vibration Contr., 2(2):33–52, 1996.
[4] E. J. Davison. The robust control of a servomechanism problem
for linear time-invariant multivariable systems. IEEE Trans. Autom.
Contr., 21:25–34, 1976.
[5] B. A. Francis, A. Sebakhy, and W. M. Wonham. Synthesis of
multivariable regulators: The internal model principle. J. Appl. Math.
Optim., 1:64–86, 1974.
[6] J. B. Hoagg, M. A. Santillo, and D. S. Bernstein. Internal model
control in the shift and delta domains. IEEE Trans. Autom. Contr.,
53:1066–1072, 2008.
[7] M. Bodson, A. Sacks, and P. Khosla. Harmonic generation in
adaptive feedforward cancellation schemes. IEEE Trans. Auto. Contr.,
39(9):1939–1944, 1994.
[8] D. S. Bayard. A general theory of linear time-invariant adpative
feedforward systems with harmonic regressors. IEEE Trans. Auto.
Contr., 45(11):1983–1996, 2000.
[9] S. J. Elliott, C. C. Boucher, and P. A. Nelson. Active noise control.
IEEE Trans. Sig. Proc., 40(5):1041–1052, 1992.
[10] D. Patt, J. Chandrasekar, D. S. Bernstein, and P. P. Friedmann.
Higher-harmonic-control algorithm for helicopter vibration reduction
revisted. J. Guidance, Control, and Dynamics, 28(5):918–930, 2005.
[11] J. Chandrasekar, L. Liu, D. Patt, P. P. Friedmann, and D. S. Bernstein.
Adaptive harmonic steady-state control for disturbance rejection.
IEEE Trans. Contr. Sys. Tech., 14(6):993–1007, 2006.
[12] S. Pigg and M. Bodson. Adaptive algorithms for the rejection of
sinusoidal disturbances acting on unknown plants. IEEE Trans. Contr.
Sys. Tech., 18(4):822–836, 2010.
[13] D. S. Bernstein. Matrix Mathematics. Princeton University Press,
2009.
[14] D. H. Brandwood. A complex gradient operator and its application in
adaptive array theory. IEE Proceedings F (Communications, Radar
and Signal Processing), 130(1):11–16, 1983.
[15] J. Hong, J. C. Ackers, R. Venugopal, M. N. Lee, A. G. Sparks,
P. D. Washabaugh, and D. S. Bernstein. Modeling, identification, and
feedback control of noise in an acoustic duct. IEEE Trans. Contr. Sys.
Tech., 4:283–291, 1996.
| 3cs.SY
|
Detection of Faults in Power System Using Wavelet Transform and
Independent Component Analysis
1
1
2
Dept. of Electrical and Electronics Engineering, IIIT, Bhubaneswar, India.
Dept. of Electrical Engineering, SOA University, Bhubaneswar, India.
3
3
4
Prakash K. Ray, 2B. K. Panigrahi, 2P. K. Rout
Asit Mohanty, 4Harishchandra Dubey
Dept. of Electrical Engineering, CET, Bhubaneswar, India.
Dept. of ECE, The University of Texas at Dallas, USA.
ABSTRACT: Uninterruptible power supply is the main motive of power utility companies that motivate them
for identifying and locating the different types of faults as quickly as possible to protect the power system
prevent complete power black outs using intelligent techniques. Thus, the present research work presents a
novel method for detection of fault disturbances based on Wavelet Transform (WT) and Independent
Component Analysis (ICA). The voltage signal is taken offline under fault conditions and is being processed
through wavelet and ICA for detection. The time-frequency resolution from WT transform detects the fault
initiation instant in the signal. Again, a performance index is calculated from independent component analysis
under fault condition which is used to detect the fault disturbance in the voltage signal. The proposed
approach is tested to be robust enough under various operating scenarios like without noise, with 20-dB noise
and variation in frequency. Further, the detection study is carried out using a performance index, energy
content, by applying the existing Fourier transform (FT), short time Fourier transform (STFT) and the
proposed wavelet transform. Fault disturbances are detected if the energy calculated in each scenario is
greater than the corresponding threshold value. The fault detection study is simulated in MATLAB/Simulink
for a typical power system.
1 INTRODUCTION
Modern power utilities require a efficient
protection scheme to protect the system itself as
well as the connected equipments for improving
better performance under normal as well as
abnormal/faulty operating scenarios. Now-a-days
the electromechanical relays are replaced by digital
relays because of their characteristics like faster
operation, accuracy and reliability. The fault
detection through digital relays and fault detector
(FD) is very vital for implementing any real-time
solutions (Phadke 1988, Dash 2000).
In the literature, so may work are for
monitoring and identification of faults in power
system. Fault situation can be identified by
comparing the difference between the value
between two consecutive cycles when higher than a
threshold value based on Phasor (Sidhu et al. 2002,
Sachdev et al., 1991). But, its demerit being the
modeling of fault resistance. Kalman filter is used
for fault detection based on estimation method
(Chowdhury et al. 1991, Zadeh et al. 2010, &
Girgis 1982). Wavelet transform (WT) is being
used for the detection of fault disturbances in power
system (Ukil & R. Živanović 2007) which detects
the changes in the signal. Then, adaptive filters and
wavelet transform in combination was used for fault
identification in power system (Ray et al. 2010).
But, these techniques affected by variation in
frequency, noise etc.
This paper proposes an algorithm for fault
detection using wavelet transform and independent
component analysis. The sudden changes are to be
detected by the detection techniques so that suitable
solutions can be adapted to protect the power
system from any type of disturbances. The proposed
method performances are being analyzed under
different operating scenarios like in presence of
noise, harmonics and frequency variations. Wavelet
transform (WT) and independent component
analysis (ICA) are suitable candidates for detection
of fault disturbances because of efficient timefrequency resolutions and their reliability to extract
the suitable features for identification purpose.
Again, the techniques are considered under different
operating conditions in order to assess the
robustness and accuracies (Pradhan et al. 2005).
This material is presented to ensure timely dissemination of scholarly and technical work. Copyright and all rights therein are retained by the authors or by the respective
copyright holders. The original citation of this paper is:
P. K. Ray, B. K. Panigrahi, P. K. Rout, A. Mohanty, H. Dubey, "Detection of Faults in Power System Using Wavelet Transform and
Independent Component Analysis", First International Conference on Advancement of Computer Communication & Electrical Technology, October 2016,
Murshidabad, India, DOI: 10.13140/RG.2.2.20394.82882
The techniques used for detection analysis is
presented in Section 2 followed by the algorithm of
implementation in Section 3. Then, Section 4
explains the result analysis both qualitatively and
quantitatively, followed by the conclusions in
Section 5.
2 TECHNIQUES FOR FAULT DETECTION
This section describes the techniques used for
identification of different types of faults in wind
system connected to grid power system. The WT
and ICA as detection techniques are presented
briefly with their mathematical modeling. Different
operating scenarios are taken as case studies to test
these methods for fault identifications which help in
assessing both normal as well as faulty operating
conditions. The details of these methods are as
follows.
2.1 Wavelet transforms (WT)
Figure 1: Wavelet decomposition tree
Considering a voltage signal of the power system
as ( ), the continuous wavelet transform (CWT) is
expressed as:
tN
V (t )*
a
M
1
DWT (V , M , N )
1
M 0m
k
n kM 0m
V (k )*
Mm
0
(2)
Where M and N are replaced by M and kM ,
and k, mare integers. Scaling function in one stage
is expressed as sum of that of next stage and can be
given by:
h ( n)
(t )
2 (2t n)
n
(3)
Using the above equation, the original voltage
signal can be written as [8]:
V (t )
a
jo
(k )2 jo / 2 (2 jo t k )
d
j
(k )2 j / 2 (2 j t k )
(4)
Where 0 is the coarse adjustment parameter in
the scaling function. The detail and approximate coefficients are written:
k
k j jo
The wavelet transform is a signal processing
algorithm which is useful in detection of abnormal
operating conditions based on decomposition of the
power signals into different ranges of frequencies
by the help of a series of low-pass and high-pass
filters. This usually provides us a time-frequency
multi-resolution analysis that greatly useful for
identifying any short of abrupt variations in the
electrical parameters such as voltage, phase,
current, frequency etc. Here, Daubechies4 (dB4) is
being used as the mother wavelet basis function for
the fault detection analysis (Ukil & R. Živanović
2007). Usually, the signal is divided into a set of
approximate (a) and detail (d) co-efficient
representing the low-frequency and high frequency
bands respectively. The decomposition is as
presented below in Figure 1.
CWT (V , M , N )
Where
and
are called as dilation and
translation parameters and
is known as the
wavelet basis function. Now WT in discrete form
as:
dt
(1)
a j (k )
a
j 1
(m)h(m 2k )
(5)
m
d j (k )
c
j 1
(m)h1 (m 2k )
m
(6)
2.2 Independent Component Analysis (ICA)
Independent Component Analysis (ICA) is an
advanced or modified version of principal
component analysis (PCA). It uses high order
statistical analysis based de-correlation for the
source input signal and provides important
information regarding the irregularities presence. In
ICA, the input data is modeled as linear coefficients
that are mutually independent to each other. Based
on the blind source separation algorithm, it can
efficiently and accurately transform the input
voltage signal into mutually and statistically
independent components, thereby helps the
detection process (Hyvärinen et al. 2001, Pöyhönen
et al. 2003, Lopez et al. 2011, & Dubey et al. 2011).
The mutual information nothing but the
individual independency measurement of the input
signal and its entropy is Gaussian in nature and can
be written as:
J ( y ) h(vgaussian ) h(v)
(7)
Differential entropy H of an input signal y with
density ( ) is given by:
h(v) ph log pv d
(8)
Here, the negentropy is estimated based on the
estimation of probability functions of input signals
Then, rather the negentropy can be expressed as:
e g w x e g v
J vi J e wiT x
T
i
gaussian
2
(9)
Here, e is the statistical expectation and G is a nonquadratic factor. For n linear mixtures taken as
x , x , … . x for the n independent components, we
can express :
x j a1s1 a2 s2 .... an sn For all
(10)
Here, x is a random mixture and s is a random
independent variable. Assuming x as the random
vector with elements x , x , … . x and random
vector with elements s , s , … . s , if A is the matrix
having elements a , we can take column vectors;
x ,as the transpose of x. Then, we can write in
matrix form:
x As
(11)
Where, A is a column matrix of elements a j and
then model is expressed as:
n
x ai si
(12)
i 1
The ICA mathematical model is an iterative one
where the independent components are called latent
variables. Here the input mix matrix is considered
to be unknown and the components s are
statistically independent. Finally, once the matrix A
is estimated, we can determine the inverse, W, then
can evaluate the independent component by the
following expression:
s Wx
(iv) Then, the principal components (PC) of the
input data are determined.
(v) In the current study, the independent
components are calculated based on fixed point
iteration [12]-[15].
(vi) Then, de-correlating
matrix, W and the
independent component, S of voltage signal is
determined from which a matrix, x is generated.
(vii) And finally, the pre-fault signal is considered
for constructing the matrix, x which can be used to
calculate the performance index and fault is
detected when the index is more than a set threshold
value. The performance index is given by
PI (k) (normal(absolute(Wf (k )* xn (k ) s f (k )))2 ) (14)
4 SIMULATED RESULTS
This section explains the detection results and
corresponding discussions in a wind system
connected to grid. The power system is simulated in
MATLAB/Simulink environment and the grid is of
230 kV, 50 Hz rating consisting of thermal based
power plant. The considered grid-connected power
system is shown in Figure 2 and here, the sampling
frequency is taken as 2 kHz.
(13)
Figure 2: Grid-connected wind power System.
3 FAULT DETECTION METHODOLOGY
The fault detection methodology is explained in
this section for the grid-connected wind power
system as follows:
(i) A grid-connected wind power system is
simulated in MATLAB where different types of
fault are created. The voltage signal is extracted at
PCC under faulty conditions. The signal is then
passed through wavelet transform to get the timefrequency analysis for identification of faulty
situations.
(ii) The voltage signal is processed through to find
the mean and de-correlation. This is the 1st level of
processing. .
(iii) Then, X data size is reduced and filtered for
better data redundancy.
4.1 Fault detection using wavelet transform
The voltage signal at point of common coupling
(PCC) is taken offline under AG fault. The signal is
processed through WT which detects the fault
instant based on filtering through a set of low-pass
and high-pass filters. The voltage signal with its
detection result is shown in Figure 3. Similarly, the
detection result using AB and ABCG fault is shown
in Figure 4 and Figure 5 respectively. These results
clearly show that WT nicely detect the fault
initiation instant nicely. But, as noticed when the
fault is cleared, as shown in Figs., the recovery to
normal waveform is un-noticed
by wavelet
transform.
Figure 3: Detection of phase-to-ground fault in wind system
connected to grid (red- voltage signal; green-WT output)
4.3 Phase-to-phase fault detection using
Independent Component Analysis
Next, a phase-to-phase fault is near to the point of
common coupling and the corresponding voltage
signal at the PCC. Then, the index as explained in
ICA section is calculated under different operating
conditions like without noise, with 20-dB noise and
under variation in frequency condition. The
simulated of the performance index obtained from
ICA is shown in Figure 6 (b). It is observed that, the
fault initiated at near about 0.065 s detected through
sudden increase in the index value. Before the fault,
the index value is observed to be almost zero. Once,
the fault occurs, the value increases suddenly and if
we set a threshold value, then the can be detected.
But in this case, we have not set the threshold value.
Rather, we are assessing the faulty condition based
on sudden increase.
Figure 4: Detection of phase-to-phase fault in wind system
connected to grid (red- voltage signal; green-WT output)
(a)
Figure 5: Detection of three-phase-to-ground fault in wind
system connected to grid (red- voltage signal; green-WT
output)
4.2 Phase-to-ground fault detection using
Independent Component Analysis
A phase-to-ground fault is near to the point of
common coupling and the corresponding voltage
signal at the PCC is taken offline. Then, the index
as explained in ICA section is calculated under
different operating conditions like without noise,
with 20-dB noise and under variation in frequency
condition. The simulated of the performance index
obtained from ICA is shown in Figure 6 (a). It is
observed that, the fault initiated at near about 0.065
s detected through sudden increase in the index
value. Before the fault, the index value is observed
to be almost zero. Once, the fault occurs, the value
increases suddenly and if we set a threshold value,
then the can be detected. But in this case, we have
not set the threshold value. Rather, we are assessing
the faulty condition based on sudden increase.
(b)
Figure 6: Detection of (a) phase-to-ground fault (b) phase-tophase fault in wind system connected to grid
4.4 Detection of fault disturbances using the
performance index (PI)
This sub-section describes the detection
technique using a performance index called energy
content of the processed faulty signal. The voltage
signal at the point of common coupling is extracted
and processed through different signal processing
transforms like Fourier transform (FT), Short time
Fourier transform (STFT) and wavelet transform.
Then the output waveform after processing is used
to calculate energy which is compared with a
threshold value to know the faulty or normal
operating conditions in the power system. It is
observed that proposed WT provides larger energy
value so that a suitable threshold can be chosen to
accurately detect the faulty conditions. The index in
different cases is shown in Table 1.
Table 1 Energy content using FT, STFT,WT
Zadeh, R.A., A. Ghosh, & G. Ledwich (2010). Combination
of Kalman Filter and Least-Error Square Techniques
IEEE Trans. on Power System 25 (4) ,2868-2880.
Girgis A. (1982). A new Kalman filtering based digital
distance relaying IEEE Trans. on Power Apparatus and
Systems 101,3471–3480.
Scenario
AG Fault
BG Fault
CG Fault
AB Fault
BC Fault
ABC Fault
FT
1.3672
1.4525
1.3324
2.2341
2.6342
3.1302
Energy content
STFT
1.7863
2.2414
1.8367
2.3532
3.1230
3.8225
WT
2.1663
2.7352
2.6538
3.2514
3.8724
4.2431
Ukil, A. & R. Živanović (2007). Abrupt change detection in
power system fault analysis using wavelet transform.
International Conference on Power Systems Transients,
Montreal, Canada.June. 19-23.
Ukil, A. and R. Živanović, (2007). Application of Abrupt
Change Detection in Power Systems Disturbance
Analysis and Relay Performance Monitoring.
5 CONCLUSIONS
This paper presented detection of fault
disturbances using Wavelet Transform and
Independent Component Analysis. Voltage signal
extracted at PCC is being processed through the
above techniques under different fault and operating
scenarios. It is observed that WT transform detected
the fault instant very accurately. But, in some cases,
it could not discriminate the disturbance condition
may be because of noise or frequency variation.
Therefore, the faults are accurately detected using
ICA under all operating scenarios
IEEE
Trans. on Power Delivery 22 (1),59-66.
Ray, P. K., H. C. Dubey, S. R.Mohanty, N. Kishor, & K.
Ganesh (2010). Power quality disturbance detection in
grid-connected wind energy system using wavelet and Stransform, IEEE ICPCES,1-4.
Pradhan, A. K. , A.Routray, & S. R.Mohanty (2005). A
Moving Sum Approach for Fault Detection of Power
Systems Electric Power Components and Systems 34 (4),
385 – 399.
Hyvärinen, A., J. Karhunen, & E. Oja (2001). Independent
Component Analysis. A Wiley Interscience Publication,
REFERENCES
Phadke, G. & J. S. Thorp (1988). Computer Relaying for Power
Systems, New York, John Wiley.
Dash, P. K., A. K. Pradhan, & G. Panda (2000). A novel fuzzy
neural network based distance relaying scheme. IEEE
Trans. on Power Delivery. 15, 902–907.
John Wiley & Sons, Inc.
Sanna, Pöyhönen.,
P. Jover, & H. Hyötyniemi (2003).
Independent component analysis of vibrations for fault
diagnosis of an induction motor Proceedings of IASTED
International
Conference
Circuits,
Signals,
and
Systems,19-21.
Lopez, M. G., H. M.Lozano, L. P. Sanchez, &L. N. O.
Sidhu, T. S., D. S. Ghotra, & M. S. Sachdev (2002). An
Moreno (2011). Blind Source Separation of Audio
adaptive distance relay and its performance comparison
Signals Using Independent Component Analysis and
with a fixed data window distance relay. IEEE Trans. on
Wavelets
Power Delivery. 17,691–697.
Communications and Computers (CONIELECOMP)
Sachdev, M. S. & M. A.Nagpal (1991). Recursive least square
21st International Conference on Electrical
Mexico, 152-157.
algorithm for power system relaying and measurement
Dubey, H.C., S. R. Mohanty, & N. Kishore (2011). Abrupt
applications. IEEE Trans. on Power Delivery. 6,1008–
change detection of fault in power system using
1015.
independent
Chowdhury, F. N., J. P. Christensen, & J. L.Aravena (1991).
component analysis
Conference on Signal Processing, Communication,
Power system fault detection and state estimation using
Computing
Kalman filter with hypothesis testing. IEEE Trans. on
(ICSCCN),659 – 664.
Power Delivery. 6, 1025–1030.
International
and
Networking
Technologies
| 3cs.SY
|
A note on the kernel of a pair of linear maps
Miguel V. Carriegos†
Noemı́ DeCastro‡
∗
Ángel Luis Muñoz Castañeda§
arXiv:1603.02730v1 [math.AC] 8 Mar 2016
September 27, 2017
Abstract
The kernel of a pair of linear maps is studied in the framework of commutative
ring theory with application to behavioral perspective of dynamical systems.
Keywords: pair of linear maps, pencil of matrices, convolution code, linear system,
hereditary ring.
2010 MSC: 93B10, 15A21, 13C10.
1
Introduction
This paper is devoted to the study of kernels of pairs of linear maps. This notion extend
some tools used in systems theory, convolutional codes and Boolean networks [7], [9].
In fact if A = RZ is the R-algebra (R a ring) of sequences of elements of R; A ∈ Rn×n ,
B ∈ Rn×m are matrices and σ is the shift operator σ(x(t)) = x(t + 1). Then set
ker(σI − A, B) = {(x(t), u(t)) ∈ An × Am | σ(x(t)) = x(t + 1) = Ax(t) + Bu(t)}
collects the trajectories of linear system. On the other hand, it has also been defined
ker(σI − A | B) = {u(t) ∈ Am | ∃x(t) ∈ An : σ(x(t)) = x(t + 1) = Ax(t) + Bu(t)}
which is central in the definition of convolutional codes because it collects the codewords
of convolutional code defined by linear system (A, B). In order to attack this problem we
are studying kernels of pairs of matrices with entries in polynomial rings R[z].
Now all along this paper R will denote a commutative ring with identity. Usually R
will be a F-algebra or even a field. We develop our results in a general framework and
claim additional properties or structure when necessary.
∗
Partially supported by INCIBE. Ministerio de Industria, Spain.
RIASC, Universidad de León, SPAIN, mail to: [email protected]
‡
Departamento de Matemáticas, Universidad de León, SPAIN, mail to: [email protected]
§
Institut für Mathematik, Freie Universität, Berlin, GERMANY, mail to: [email protected]
†
1
2
Carriegos, DeCastro & Muñoz-Castañeda
Definition 1.1. Let N and Mi , i = 1, 2 be R-modules and fi : Mi → N be R-linear
maps. We define
ker(f1 | f2 ) = {m2 ∈ M2 | ∃m1 ∈ M1 : f1 (m1 ) + f2 (m2 ) = 0}
while we set
ker(f1 , f2 ) = ker(f1 ⊕ f2 : M1 ⊕ M2 → N ) = {(m1 , m2 ) ∈ M1 ⊕ M2 : f (m1 ) + f2 (m2 ) = 0}
as usual.
The paper is organized as follows: Section 2 deals with main properties of ker(f1 | f2 )
which follow from the fact that ker(f1 | f2 ) is the kernel of certain linear map and it is also
the cockernel of another linear map. We also point out several properties of ker(f1 | f2 )
when maps fi have nice properties.
Section 3 is devoted to give a explicit factorization of ker(f1 | f2 ) in terms of usual
kernels ker(f1 , f2 ) and ker(f1 ) when involved R-modules are hereditary or in particular
when commutative ring R is hereditary (that is when submodules of projective R-modules
are again projective). Note that the class of hereditary rings contains several interesting
classes of commutative rings like for instance fields, principal ideal domains, Dedekind
domains, von Neumann regular rings and Boolean rings.
Next we study scalar extensions R → S in section 4. If R-algebra S happens to be
R-flat then scalar extension of ker(f1 | f2 ) is given by the kernel of pair of extended maps
ker(f1 ⊗ IS | f2 ⊗ IS ). This fact is used extensively in section 5 to develop the case of
product rings R ∼
= R1 × · · · × Rs . We conclude by giving some results related to the kernel
of pairs of polynomial matrices which are applicable both in behavioral linear systems and
convolutional codes.
2
The kernel of a pair. Definition and properties
Let R be a commutative ring; M1 , M2 , N are R-modules and let fj : Mj → N be R-linear
maps.
Definition 2.1. We denote by ker(f1 | f2 ) the subset of all m2 ∈ M2 such that there
exists m1 ∈ M1 with the property f1 (m1 ) + f2 (m2 ) = 0; thats to say
ker(f1 | f2 ) = {m2 ∈ M2 | ∃m1 ∈ M1 : f1 (m1 ) + f2 (m2 ) = 0} .
This is a generalization of so-called kernel of a pair of morphisms as given in [7],
[9]. Note that it is quite clear that ker(f1 | f2 ) is a R-submodule of M2 . It is also
straightforward that
3
Kernel of a pair of linear maps
Proposition 2.2. ker(f1 | f2 ) = f2−1 (Im (f1 )).
Proof. If m2 ∈ ker(f1 | f2 ) then f1 (m1 )+f2 (m2 ) = 0 for some m1 ; that is, f2 (m2 ) ∈ Im (f1 )
or m2 ∈ f2−1 (Im (f1 ))
We are also interested in presenting ker(f1 | f2 ) as the kernel of a R-linear map and
as the cokernel of another R-linear map. Let’s denote by pi : N → N/ Im (fi ) the natural
quotient map of identity sending n 7→ n + Im (fi ). Then it is straightforward that
Proposition 2.3. ker(f1 | f2 ) = ker(p1 ◦ f2 : M2 → N/ Im (f1 ))
Proof. x ∈ ker(p1 ◦ f2 : M2 → N/ Im (f1 )) ⇔ f2 (x) ∈ Im (f1 )
On the other hand put the linear map (f1 , f2 ) : M1 ⊕ M2 → N sending (m1 , m2 ) 7→
(f1 , f2 )(m1 , m2 ) = f1 (m1 ) + f2 (m2 ). Note that ker(f1 ) ⊆ ker(f1 , f2 ) because if m ∈
ker(f1 ) then f1 (m) = 0; hence (f1 , f2 )(m, 0) = 0 and consequently (m, 0) ∈ ker(f1 , f2 ).
Moreover one has
Theorem 2.4. ker(f1 | f2 ) = coker[ker(f1 ) ֒→ ker(f1 , f2 )]
Proof. The projection onto the second factor π2 : M1 ⊕ M2 → M2 restricts to the onto
π
2
map ker(f1 , f2 ) −→
ker(f1 | f2 ) whose kernel consist of all pairs satisfying π2 (x1 , x2 ) = 0
and therefore x2 = 0 by definition of π2 . So, a fortiori, f1 (x1 ) = 0, and consequently we
are done because
h
i
π2
ker ker(f1 , f2 ) −→
ker(f1 | f2 ) = ker(f1 )
That is, the following sequence is short exact
0 → ker(f1 ) ֒→ ker(f1 , f2 ) → ker(f1 | f2 ) → 0
m1 7→ (m1 , 0)
(m1 , m2 ) 7→ m2
The following properties of the kernel of a pair of linear maps, may be easily derived.
Proposition 2.5. Consider the R-linear maps fi : Mi → N . Denote by 0 and I respectiveli the zero linear map and the identity map. Then
(i) ker(0 | f2 ) = ker(f2 )
(ii) If f1 is onto then ker(f1 | f2 ) = M2
(iii) ker(0 | I) = 0
4
Carriegos, DeCastro & Muñoz-Castañeda
(iv) ker(f1 | 0) = M2
(v) In general, Im (f2 ) ⊆ Im (f1 ) ⇒ ker(f1 | f2 ) = M2
(vi) If Ψ2 is an automorphism of M2 then mapping m2 7→ Ψ2 (m2 ) is an isomorphism
ker(f1 | f2 Ψ2 ) → ker(f1 | f2 )
(vii) ker(f1 | f2 Ψ2 ) = Ψ−1
2 (ker(f1 | f2 ))
(viii) If Ψ1 is an automorphism of M1 then ker(f1 Ψ1 | f2 ) = ker(f1 | f2 )
(ix) If Ψ is an isomorphism of N then ker(Ψf1 | f2 ) = ker(f1 | Ψ−1 f2 )
Proof. (i), (ii), (iii) are straightforward from Proposition 2.2; (iv), (v) are directly obtained
from Definition 2.1.
Property (vi) is consequence of the fact that Ψ2 maps ker(f1 | f2 Ψ2 ) onto ker(f1 | f2 ).
Since Ψ2 is injective (because it is a restriction an isomorphism) it follows that Ψ2 is itself
an isomorphism and hence the result. Property (vii) follows straightforward from (vi).
Assertion (viii) is clear because ker(f1 Ψ1 | f2 ) = {m2 ∈ M2 | ∃m1 ∈ M1 : f1 (Ψ1 m1 ) +
f2 m2 = 0}
Finally to prove (ix) note that ker(Ψf1 | f2 ) = {m2 ∈ M2 | ∃m1 ∈ M1 : Ψf1 (m1 ) +
f2 m2 = 0}. Defining condition is equivalent to saying that f1 m1 + Ψ−1 f2 m2 = 0 which is
the definition of ker(f1 | Ψ−1 f2 )
3
Hereditary rings
A R-module M is called hereditary if every submodule of M is projective. If every ideal
of R is projective (that is, R is a ring of global dimension one) then every projective Rmodule is hereditary and commutative ring R is called hereditary. A hereditary integral
domain is a Dedekind domain.
Assume that R-modules Mi , N are hereditary R-modules. It follows that all R-modules
in exact sequence of Theorem 2.4
0 → ker(f1 ) ֒→ ker(f1 , f2 ) → ker(f1 | f2 ) → 0
are projective. Therefore
Theorem 3.1. If R-modules Mi and N are hereditary (in particular if R is itself an
hereditary ring) then one has the factorization ker(f1 , f2 ) ∼
= ker(f1 ) ⊕ ker(f1 | f2 ) and,
consequently
ker(f1 | f2 ) ∼
= ker(f1 , f2 )/ ker(f1 )
5
Kernel of a pair of linear maps
Example 3.2. Put a matrix A ∈ Rn×n with no real eigenvalues and B ∈ Rn×m then one
has the isomorphism ker(zI − A | B) ∼
= ker(zI − A, B). This result, translated to the case
of z = σ the shift operator in R-algebra A = RZ , implies that sequences of admissible
controls ~u(t) of linear system ~x(t + 1) = A~x(t) + B~u(t) are in bijective correspondece with
trajectories (solutions) (~x(t), ~u(t)) of linear system.
4
Scalar Extension. Base change
Assume in the sequel that S is a R-algebra with structural morphism ρ : R → S. Scalar
extension of R-module M is the S-module ρ∗ (M ) = M ⊗R S, while scalar extension of
R-linear map f : M → N is given by ρ∗ (f ) = f ⊗ IS : M ⊗R S → N ⊗R S.
Scalar extension functor (− ⊗R S) preserves onto linear mappings, it is right-exact
(see [1]) and hece it preserves epimorphisms. But notice that neither monomorphisms nor
kernels are conserved in general by scalar extensions; for instance take ker[2 : Z → Z] = 0
and observe that ker[2⊗ IZ/2Z : Z ⊗Z Z/2Z → Z ⊗Z Z/2Z] = ker[0 : Z/2Z → Z/2Z] = Z/2Z
A R-module (respectively R-algebra) S is flat (respectively flat algebra) if functor
(− ⊗R M ) happens to be exact; that is it transforms exact sequences into exact sequences.
Examples of R-flat modules are free R-modules and projective R-modules [1],[2], [3] or [8].
Within these conditions one has
Theorem 4.1. If S is a flat R-algebra then
ρ∗ (ker(f1 | f2 )) = ker(ρ∗ (f1 ) | ρ∗ (f2 ))
Proof. Scalar extension functor ρ∗ (−) = − ⊗R S is exact because S is R-flat. Hence exact
sequence
0
ι
// ker(f1 )
π
// ker(f1 , f2 )
// ker(f1 | f2 )
// 0
yields the exact sequence
0
// ker(f1 ) ⊗R S
ι⊗IS
// ker(f1 , f2 ) ⊗R S
π⊗IS
// ker(f1 | f2 ) ⊗R S
// 0
Once again, since S is R-flat it follows the commutative square
0
// ker(f1 ) ⊗R S
0
// ker(f1 ⊗ IS )
ι⊗IS
ι′ //
// ker(f1 , f2 ) ⊗R S
ker(f1 ⊗ IS , f2 ⊗ IS )
and therefore ρ∗ (ker(f1 |f2 )) = ker(f1 |f2 ) ⊗R S = ker(f1 ⊗ IS | f2 ⊗ IS ) = ker(ρ∗ (f1 ) |
ρ∗ (f2 )).
6
Carriegos, DeCastro & Muñoz-Castañeda
5
Product rings
Now we deal with the case of R = R1 × · · · × Rt being a finite direct product of rings.
In this case each factor ring Ri ∼
= ei R where ei = (0, ..., 0, 1, 0, ..., 0) is the ith structrual
∼ e1 R ⊕ · · · ⊕ et R and therefore each factor Ri ∼
idempotent of product. Hence R1 =
= ei R is
a flat R-algebra because it is projective due to it is a direct summand of the free R-module
R1 . This decomposition can be traslated to kernels of pairs of linear maps:
Theorem 5.1. If πi : R1 × · · · × Rt → Ri is the projection onto ith factor, then πi is
also structural R-algebra morphism and, for a given pair of R-linear maps fj : Mj → N ,
j = 1, 2 one has
ker(f1 | f2 ) ∼
=
t
M
ker(πi∗ (f1 ) | πi∗ (f2 ))
i=1
Proof. The universal property of product yields the natural isomorphism Φ which is unique
commutating both triangles in below natural diagram
πi∗
--
ker(πi∗ (f1 ) | πi∗ (f2 ))
p1
❞❞❞❞❞❞22
❞❞❞❞❞❞❞
L
Φ
ker(f1 | f2 ) ❴ ❴ ❴ ❴ ❴ ❴ ❴ ❴// ti=1 ker(πi∗ (f1 ) | πi∗ (f2 ))
❩❩❩❩❩❩p2❩
❩❩❩❩❩❩,,
11
ker(πj∗ (f1 ) | πj∗ (f2 ))
φ∗j
Now we conclude with the case of R being a finite product of fields. This case contains modular rings Z/mZ where m is a square-free integer and would be of interest in
convolutional coding.
Corollary 5.2. Let Ki be a field for each i and consider a ring R = K1 × · · · × Ks . Let
A ∈ Rp×q1 and B ∈ Rp×q2 be matrices. Then
ker(A | B) = e1 · ker(π1 (A) | π1 (B)) + · · · + es · ker(πs (A) | πs (B))
Notice that Theorem 2.4 applies to above case of R beign a finite product of fields,
hence one has the following result.
Corollary 5.3. Let Ki be a field for each i and consider a ring R = K1 × · · · × Ks . Let
A ∈ Rp×q1 and B ∈ Rp×q2 be matrices. Then
ker(A | B) =
ker(πs (A), πs (B))
ker(π1 (A), π1 (B))
⊕ ··· ⊕
ker(π1 (A))
ker(πs (A))
7
Kernel of a pair of linear maps
Above results might be explained with an example. Consider the modular integer ring
Z/30Z ∼
= Z/2× Z/3Z × Z/5Z, where isomorphism is given by Chinese Remainder Theorem
(a, b, c) 7→ unique x such that x = a(mod 2), x = b(mod 3), x = c(mod 5)
Thence structural idempotents are given by:
Z/2 × Z/3Z × Z/5Z
// Z/30
(1, 0, 0)
// 15
(0, 1, 0)
// 10
(0, 0, 1)
// 6
Thus the kernel might be recovered from local data on every factor field by using Corollary
5.2 with above structural idempotents e1 = 15, e2 = 10, and e3 = 6.
Note that this result can also be generalized to the case of polynomial matrices which
would be useful in the behavioral theory of linear systems and in particular to convolutional
codes.
Corollary 5.4. Consider matrices A(z), B(z) of adequate sizes, p × q1 and p × q2 respectively, and entries in R[z] where R = R1 × · · · × Rs is a product ring with structural
idempotents ei ∈ R. Then trajectories can be computed locally and glued together; that is
to say, one has
ker(A(z) | B(z)) = {u(z) | ∃x(z) : A(z)x(z) + B(z)u(z) = 0} =
= e1 ker(π1 (A)(z) | π1 (B)(z)) + · · · + es ker(πs (A)(z) | πs (B)(z))
Proof. Note that R[z] = (R1 × · · · × Rs )[z] ∼
= R1 [z] × · · · × Rs [z] and thus Ri [z] is a flat
R[z]-algebra. It only remains to detect structural morphism, which is a trivial exercise.
Then we conclude by using Theorem 5.1
To conclude note that if K is a field then K[z] is a principal ideal domain and thus
a Dedekind domain and hence hereditary. Thus by applying Theorem 3.1 one has the
following factorization result.
Corollary 5.5. Consider matrices A(z), B(z) of adequate sizes, p × q1 and p × q2 respectively, and entries in R[z] where R = K1 × · · · × Ks is a product of fields. Then
ker((A(z) | B(z)) =
ker(π1 (A)(z) | π1 (B)(z))
ker(π1 (A)(z) | π1 (B)(z))
⊕ ··· ⊕
ker(π1 (A)(z))
ker(π1 (A)(z))
8
Carriegos, DeCastro & Muñoz-Castañeda
Notice that the key point in above results is some kind of ”commutativity” between
the product of rings and extension from a ring to its ring of polynomials; that is to
say,R[z] = (R1 × · · · × Rs )[z] ∼
= R1 [z] × · · · × Rs [z], which assures that Ri [z] is a flat
R[z]-algebra. This phenomenon is also true for the extensions to formal power series
R 7→ R[[z]]; polynomial ring R[z, z −1 ]; formal power series R[[z, z −1 ]]; Puisseux series;
and, if Ri are domains, for extensions to rational fractions R 7→ R(z) and Laurent series
R 7→ R((z)). Thus we conjecture that Theorem 5.1 could be extended to these scenarios.
To conclude it is worth to note that factorization results à la Corollary 5.2 holds on
by assuring that factor rings Ri [z] (or respectively Ri [[z]], ...) are hereditary.
References
[1] M.F. Atiyah, I.G. McDonald, Introduction to Commutative Algebra, Addison-Wesley,
1969.
[2] H. Bass, Algebraic K-theory, Benjamin, 1968.
[3] N. Bourbaki. Éléments de Mathématique, Algébre Conmutative, ch. 1 à 4, Masson,
1985.
[4] M.V. Carriegos, J.A. Hermida-Alonso, T. Sánchez-Giralda, The pointwise feedback
relation for linear dynamical systems over commutative rings, Lin. Algebra App.
279 (1998) 119–134.
[5] N. DeCastro, Clasificación Feedback de Sistemas y Códigos de Convolución. Aplicaciones en Cibernética, Teorı́a de Códigos y Criptografı́a, Ph.D. Thesis, Universidad
de León, 2016.
[6] M. Kuijper, R. Pinto, J.W. Polderman, Kernel representations for behaviors over finite
rings, Proceedings of the 17th International Symposium on Mathematical Theory
of Networks and Systems, 2006, pp. 2494–2503.
[7] J. Rosenthal, On behaviors and convolutional codes, IEEE Transactions on Inform.
Theory 42(6), (1996), 1881–1891.
[8] C.A. Weibel, The K-book, an introduction to Algebraic K-theory, β-edition
(2013/02/23), http://www.math.rutgers.edu/∼weibel/Kbook/Kbook.pdf
[9] J.C. Willems, Paradigms and puzzles in the theory of dynamical systems, IEEE Transactions on Automatic Control 36(3) (1991), 259–294.
[10] E. Zerz, Behavioral systems theory: a survey, Int. J. Appl. Math. Comput. Sci. 18(3),
(2008), 265–270.
| 0math.AC
|
Fixed-Parameter Algorithms for DAG PartitioningI
René van Beverna,b,∗, Robert Bredereckb , Morgan Chopinc , Sepp Hartungb , Falk Hüffnerb , André Nichterleinb , Ondřej
Suchýd,b
a Novosibirsk
State University, Novosibirsk, Russian Federation
für Softwaretechnik und Theoretische Informatik, TU Berlin, Germany
c Institut für Optimierung und Operations Research, Universität Ulm, Germany
d Faculty of Information Technology, Czech Technical University in Prague, Czech Republic
arXiv:1611.08809v1 [cs.DS] 27 Nov 2016
b Institut
Abstract
Finding the origin of short phrases propagating through the web has been formalized by Leskovec et al. [ACM
SIGKDD 2009] as DAG PARTITIONING: given an arc-weighted directed acyclic graph on n vertices and m arcs,
delete arcs with total weight at most k such that each resulting weakly-connected component contains exactly one
sink—a vertex without outgoing arcs. DAG PARTITIONING is NP-hard.
We show an algorithm to solve DAG PARTITIONING in O(2k · (n + m)) time, that is, in linear time for fixed k. We
complement it with linear-time executable data reduction rules. Our experiments show that, in combination, they can
optimally solve DAG PARTITIONING on simulated citation networks within five minutes for k ≤ 190 and m being 107
and larger. We use our obtained optimal solutions to evaluate the solution quality of Leskovec et al.’s heuristic.
We show that Leskovec et al.’s heuristic works optimally on trees and generalize this result by showing that DAG
2
PARTITIONING is solvable in 2O(t ) · n time if a width-t tree decomposition of the input graph is given. Thus, we
improve an algorithm and answer an open question of Alamdari and Mehrabian [WAW 2012].
We complement our algorithms by lower bounds on the running time of exact algorithms and on the effectivity of
data reduction.
Keywords: NP-hard problem, graph algorithms, polynomial-time data reduction, multiway cut, linear-time algorithms,
algorithm engineering, evaluating heuristics
1. Introduction
The DAG PARTITIONING problem was introduced by Leskovec et al. [26] in order to analyze how short, distinctive
phrases (typically, parts or mutations of quotations, also called memes) spread to various news sites and blogs. To
demonstrate their approach, Leskovec et al. [26] collected and analyzed phrases from 90 million articles from the
time around the United States presidential elections in 2008; the results were featured in the New York Times [27].
Meanwhile, their approach has grown up into N IFTY, a system that allows near real-time observation of the rise and
fall of trends, ideas, and topics in the internet [33].
A core component in the approach of Leskovec et al. [26] is a heuristic to solve the NP-hard DAG PARTITIONING
problem. They use it to cluster short phrases, which may undergo modifications while propagating through the web,
with respect to their origins. To this end, they create an arc-weighted directed acyclic graph with phrases as vertices
and draw an arc from phrase p to phrase q if p presumably originates from q, where the weight of an arc represents the
I A preliminary version of this article appeared in the Proceedings of the 8th International Conference on Algorithms and Complexity (CIAC’13) [5].
Besides providing full proof details, this revised and extended version improves our O(2k · n2 )-time algorithm to run in O(2k · (n + m)) time and
provides linear-time executable data reduction rules. Moreover, we experimentally evaluated the algorithm and compared it to known heuristics by
Leskovec et al. [26]. We prove the latter to work optimally on trees.
∗ Corresponding author.
Email addresses: [email protected] (René van Bevern), [email protected] (Robert Bredereck),
[email protected] (André Nichterlein), [email protected] (Ondřej Suchý)
support for this hypothesis: the weight assigned to an arc (p, q) is chosen inversely proportional to the time difference
between p and q and their Levenshtein distance when using words as tokens, whereas it is proportional to the total
number of documents in the corpus that contain the phrase q.1 A vertex without outgoing arcs is called a sink and
can be interpreted as the origin of a phrase. If a phrase has directed paths to more than one sink, its ultimate origin is
ambiguous and, in the model of Leskovec et al. [26], at least one of the “p originates from q” hypotheses is wrong.
Leskovec et al. [26] introduced DAG PARTITIONING with the aim to resolve these inconsistencies by removing a set
of arcs (hypotheses) with least support:
DAG PARTITIONING
Input: A directed acyclic graph G = (V, A) with positive integer arc weights ω : A → N and a positive integer k ∈ N.
Question: Is there a set S ⊆ A with ∑a∈S ω(a) ≤ k such that each weakly-connected component in G0 = (V, A \ S)
has exactly one sink?
Herein, the model of Leskovec et al. [26] exploits that a weakly-connected component of a directed acyclic graph
contains exactly one sink if and only if all its vertices have directed paths to only one sink. We call a set S ⊆ A such
that each weakly-connected component in G0 = (V, A \ S) has exactly one sink a partitioning set.
Leskovec et al. [26] showed that DAG PARTITIONING is NP-hard and presented a heuristic to find partitioning sets
of small weight. Alamdari and Mehrabian [1] showed that, for fixed ε > 0, even approximating the minimum weight of
a partitioning set within a factor of O(n1−ε ) is NP-hard. In the absence of approximation algorithms, exact solutions to
DAG PARTITIONING become interesting for the reason of evaluating the quality of known heuristics alone.
We aim for solving DAG PARTITIONING exactly using fixed-parameter algorithms—a framework to obtain
algorithms to optimally solve NP-hard problems that run efficiently given that certain parameters of the input data are
small [15, 28, 12]. A natural parameter to consider is the minimum weight k of the partitioning set sought, since one
would expect that wrong hypotheses have little support.
Known results. To date, there are only few studies on DAG PARTITIONING. Leskovec et al. [26] showed that DAG
PARTITIONING is NP-hard and present heuristics. Alamdari and Mehrabian [1] showed that, on n-vertex graphs, DAG
PARTITIONING is hard to approximate in the sense that if P , NP, then there is no polynomial-time factor-(n1−ε )
approximation algorithm for any fixed ε > 0, even if the input graph has unit-weight arcs, maximum outdegree three,
2
and only two sinks. Moreover, Alamdari and Mehrabian [1] showed that DAG PARTITIONING can be solved in 2O(t ) · n
time if a width-t path decomposition of the input graph is given.
DAG PARTITIONING is very similar to the well-known NP-hard M ULTIWAY C UT problem [11]: given an undirected
edge-weighted graph and a subset of the vertices called terminals, delete edges of total weight at most k such that each
terminal is separated from all others. DAG PARTITIONING can be considered as M ULTIWAY C UT with the sinks being
terminals and the additional constraint that not all arcs outgoing from a vertex may be deleted, since this would create a
new sink. Xiao [34] gave an algorithm to solve M ULTIWAY C UT in O(2k · min(n2/3 , m1/2 ) · nm) time. Interestingly, in
contrast to DAG PARTITIONING, M ULTIWAY C UT is constant-factor approximable (see, e. g., Karger et al. [23]).
Our results. We provide algorithmic as well as intractability results. On the algorithmic side, we present an O(2k · (n +
m)) time algorithm for DAG PARTITIONING and complement it with linear-time executable data reduction rules. We
experimentally evaluated both and, in combination, they solved instances with k ≤ 190 optimally within five minutes,
the number of input arcs being 107 and larger. Moreover, we use the optimal solutions found by our algorithm to
evaluate the quality of Leskovec et al. [26]’s heuristic and find that it finds optimal solutions for most instances that our
algorithm solves quickly, but performs worse by a factor of more than two on other instances.
2
Also, we give an algorithm that solves DAG PARTITIONING in 2O(t ) · n time if a width-t tree decomposition of
the input graph is given. We thus answer an open question by Alamdari and Mehrabian [1]. Since every width-t path
decomposition is a width-t tree decomposition but not every graph allowing for a width-t tree decomposition allows
2
for a width-t path decomposition, our algorithm is an improvement over the 2O(t ) · n-time algorithm of Alamdari and
Mehrabian [1], which requires a path decomposition as input.
1 Unfortunately, Leskovec et al. [26] neither give a more precise description nor the range of their weights. All our hardness results even hold for
unit weights, while all our algorithms work whenever the weight of each arc is at least one; however, we choose the weights to be positive integers to
avoid representation issues.
2
On the side of intractability results, we strengthen the NP-hardness results of Leskovec et al. [26] and Alamdari and
Mehrabian [1] to graphs of diameter two and maximum degree three and we show that our O(2k ·(n+m)) time algorithm
cannot be improved to O(2o(k) · poly(n)) time unless the Exponential Time Hypothesis fails. Moreover, we show that
DAG PARTITIONING does not admit polynomial-size problem kernels with respect to k unless NP ⊆ coNP/poly.
Organization of this paper. In Section 2, we introduce necessary notation and two basic structural observations for
DAG PARTITIONING that are important in our proofs.
In Section 3, we present our O(2k · (n + m)) time algorithm and its experimental evaluation. With the help of the
optimal solutions computed by our algorithm we also evaluate the quality of a heuristic presented by Leskovec et al. [26].
Moreover, we discuss the limits of parameterized algorithms and problem kernelization for DAG PARTITIONING
parameterized by k.
2
Section 4 presents our 2O(t ) · n time algorithm. It follows that DAG PARTITIONING is linear-time solvable when at
least one of the parameters k or t is fixed. We further show that the heuristic presented by Leskovec et al. [26] works
optimally on trees.
Section 5 then shows that other parameters are not as helpful in solving DAG PARTITIONING: DAG PARTITIONING
remains NP-hard even when graph parameters like the diameter or maximum degree are constants.
2. Preliminaries and basic observations
We consider finite simple directed graphs G = (V, A) with vertex set V (G) := V and arc set A(G) := A ⊆ V ×V , as
well as finite simple undirected graphs G = (V, E) with vertex set V and edge set E(G) := E ⊆ {{u, v} | u, v ∈ V }. For
a directed graph G, the underlying undirected graph G0 is the graph that has undirected edges in the places where G has
arcs, that is, G0 = (V (G), {{v, w} : (v, w) ∈ A(G)}). We will use n to denote the number of vertices and m to denote the
number of arcs or edges of a graph.
For a (directed or undirected) graph G, we denote by G \ A0 the subgraph obtained by removing from G the arcs or
edges in A0 and by G −V 0 the subgraph obtained by removing from G the vertices in V 0 . For V 0 ⊆ V , we denote by
G[V 0 ] := G − (V \V 0 ) the subgraph of G induced by the vertex set V 0 .
The set of out-neighbors and in-neighbors of a vertex v in a directed graph is N + (v) := {u | (v, u) ∈ A} and N − (v) :=
{u | (u, v) ∈ A}, respectively. The outdegree, the indegree, and the degree of a vertex v ∈ V are d + (v) := |N + (v)|,
d − (v) := |N − (v)|, and d(v) := d + (v) + d − (v), respectively. A vertex v is a sink if d + (v) = 0; it is isolated if d(v) = 0.
A path of length ` − 1 from v1 to v` in an undirected graph G is a tuple (v1 , . . . , v` ) ∈ V ` such that {vi , vi+1 } is an
edge in G for 1 ≤ i ≤ ` − 1. An undirected path in a directed graph is a path in its underlying undirected graph. A
directed path of length ` − 1 from v1 to v` in a directed graph G is a tuple (v1 , . . . , v` ) ∈ V ` such that (vi , vi+1 ) is an arc
in G for 1 ≤ i ≤ ` − 1.
We say that u ∈ V can reach v ∈ V or that v is reachable from u in G if there is a directed path from u to v in G. We
say that u and v are connected if there is a (not necessarily directed) path from u to v in G. In particular, u is reachable
from u and connected to u. We use connected component as an abbreviation for weakly connected component, that is, a
maximal set of pairwise connected vertices. The diameter of G is the maximum length of a shortest path between two
different vertices in the underlying undirected graph of G.
Fixed-parameter algorithms. The main idea in fixed-parameter algorithms is to accept the super-polynomial running
time, which is seemingly inevitable when optimally solving NP-hard problems, but to restrict it to one aspect of the
problem, the parameter. More precisely, a problem Π is fixed-parameter tractable (FPT) with respect to a parameter k
if there is an algorithm solving any instance of Π with size n in f (k) · poly(n) time for some computable function f
[15, 28, 12, 10]. Such an algorithm is called fixed-parameter algorithm. Since Suen et al. [33] point out that the input
instances of DAG PARTITIONING can be so large that even running times quadratic in the input size are prohibitively
large, we focus on finding algorithms that run in linear time if the parameter k is a constant. An important ingredient
of our algorithms is linear-time data reduction, which recently received increased interest since data reduction is
potentially applied to large input data [32, 4, 18, 3, 22, 13].
3
Problem kernelization. One way of deriving fixed-parameter algorithms is (problem) kernelization [17, 7]. As a formal
approach of describing efficient data reduction that preserves optimal solutions, problem kernelization is a powerful
tool for attacking NP-hard problems. A kernelization algorithm consists of data reduction rules that, applied to any
instance x with parameter k, yield an instance x0 with parameter k0 in time polynomial in |x| + k such that (x, k) is a
yes-instance if and only if (x0 , k0 ) is a yes-instance, and if both |x0 | and k0 are bounded by some functions g and g0 in k,
respectively. The function g is referred to as the size of the problem kernel (x0 , k0 ).
Note that it is the parameter that allows us to measure the effectiveness of polynomial-time executable data
reduction, since a statement like “the data reduction shrinks the input by a factor α” would imply that we can solve
NP-hard problems in polynomial time.
From a practical point of view, problem kernelization is potentially applicable to speed up exact and heuristic
algorithms to solve a problem. Since kernelization is applied to shrink potentially large input instances, recently the
running time of kernelization has stepped into the focus and linear-time kernelization algorithms have been developed
for various NP-hard problems [32, 4, 18, 3, 22, 13].
Two basic observations.. The following easy to prove structural observations will be exploited in many proofs of our
work. The first observation states that a minimal partitioning set does not introduce new sinks.
Observation 1. Let G be a directed acyclic graph and S be a minimal partitioning set for G. Then, a vertex is a sink
in G \ S if and only if it is a sink in G.
P ROOF. Clearly, deleting arcs from a directed acyclic graph cannot turn a sink into a non-sink. Therefore, it remains to
show that every sink in G \ S was a sink already in G. Towards a contradiction, assume that there is a vertex s that is
a sink in G \ S but not in G. Then, there is an arc (s, v) in G for some vertex v of G. Let Cv and Cs be the connected
components in G \ S containing v and s, respectively, and let sv be the sink in Cv . Then, for S0 := S \ {(s, v)}, the
connected component Cv ∪Cs in G \ S0 has only one sink, namely sv . Thus, S0 is also a partitioning set for G, but S0 ( S,
a contradiction to S being minimal.
The second observation is that each vertex in a directed acyclic graph is connected to exactly one sink if and only if
it can reach that sink.
Observation 2. Let G be a directed acyclic graph. An arc set S is a partitioning set for G if and only if each vertex
in G can reach exactly one sink in G \ S.
P ROOF. If S is a partitioning set for G, then, by definition, each connected component of G \ S contains exactly one
sink. Therefore, each vertex in G \ S can reach at most one sink. Moreover, since G \ S is a directed acyclic graph and
each vertex in a directed acyclic graph can reach at least one sink, it follows that each vertex in G \ S can reach exactly
one sink.
Now, assume that each vertex in G \ S can reach exactly one sink. We show that each connected component of G \ S
contains exactly one sink. For the sake of contradiction, assume that a connected component C of G \ S contains
multiple sinks s1 , . . . , st . For i ∈ [t], let Ai be the set of vertices that reach si . These vertex sets are pairwise disjoint and,
since every vertex in a directed acyclic graph reaches some sink, they partition the vertex set of C.
Since C is a connected component, there are i, j ∈ [t] with i , j and some arc (v, w) in G \ S from some vertex v ∈ Ai
to some vertex w ∈ A j . This is a contradiction, since v can reach si as well as s j .
3. Parameter weight of the partitioning set sought
This section investigates the influence of the parameter “weight k of the partitioning set sought” on DAG PAR First, in Section 3.1, we present an O(2k · (n + m))-time algorithm and design linear-time executable
data reduction rules. In Section 3.2, we experimentally evaluate the algorithm and data reduction rules and—with
the help of the optimal solutions computed by our algorithm—we also evaluate the quality of a heuristic presented
by Leskovec et al. [26]. In Section 3.3, we investigate the question whether the provided data reduction rules might
have a provable shrinking effect on the input instance in form of a polynomial-size problem kernel. We will see that,
despite the fact that data reduction rules work very effectively in experiments, polynomial-size problem kernels for
DAG PARTITIONING do not exist under reasonable complexity-theoretic assumptions. Moreover, Section 3.3 will also
show that the algorithm presented in Section 3.1 is essentially optimal.
TITIONING .
4
Algorithm 1: Compute a partitioning set of weight at most k.
Input: A directed acyclic graph G = (V, A) with arc weights ω and a positive integer k.
Output: A partitioning set S of weight at most k if it exists; otherwise ‘no’.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
(v1 , v2 , . . . , vn ) ← reverse topological order of the vertices of G
L ← array with n entries
searchtree(1, 0)
/
// start with vertex v1 and S = 0/
output ‘no’
// there is no partitioning set of weight at most k
Procedure searchtree(i, S)
// vertex counter i; (partial) partitioning set S
while vi is a sink s or there is a sink s such that ∀w ∈ N + (vi ) : L[w] = s do
L[vi ] ← s
// associate vi with sink s
i ← i+1
// continue with next vertex
if i > n then
if ω(S) ≤ k then output S
// all vertices have been handled, S is a partitioning set
// a partitioning set of weight at most k has been found
else
D ← {L[w] | w ∈ N + (vi )}
// the set of feasible sinks for vi
if |D| − 1 ≤ k − ω(S) then
// check whether we are allowed to delete |D| − 1 arcs
foreach s ∈ D do
// try to associate vi to each feasible sink s
L[vi ] ← s
S0 ← S ∪ {(vi , w) | w ∈ N + (vi ) and L[w] , s}
searchtree(i + 1, S0 )
3.1. Constant-weight partitioning sets in linear time
We now present an algorithm to compute partitioning sets of weight k in O(2k · (n + m)) time. Interestingly, although
both problems are NP-hard, it will turn out that DAG PARTITIONING is substantially easier to solve exactly than the
closely related M ULTIWAY C UT problem, for which a sophisticated algorithm running in O(2k min(n2/3 , m1/2 )nm) time
was given by Xiao [34]. This is in contrast to M ULTIWAY C UT being constant-factor approximable [23], while DAG
PARTITIONING is inapproximable unless P = NP [1].
The main structural advantage of DAG PARTITIONING over M ULTIWAY C UT is the alternative characterization of
partitioning sets given in Observation 2: we only have to decide which sink each vertex v will reach in the optimally
partitioned graph. To this end, we first decide the question for all out-neighbors of v. This natural processing order
allows us to solve DAG PARTITIONING by the simple search tree algorithm shown in Algorithm 1, which we explain
in the proof of the following theorem.
Theorem 1. Algorithm 1 solves DAG PARTITIONING in O(2k · (n + m)) time.
P ROOF. Algorithm 1 is based on recursive branching and computes a partitioning set S of weight at most k for a
directed acyclic graph G = (V, A). It exploits the following structural properties of a minimal partitioning set S: by
Observation 2, a vertex v is connected to a sink s in G \ S if and only if it can reach that sink s in G \ S. Thus, consider a
vertex v of G and assume that we know, for each out-neighbor w of v in G, the sink s that w can reach in G \ S. We call a
sink s of G feasible for v if an out-neighbor of v in G can reach s in G \ S. Let D be the set of feasible sinks for v. Since
v may be connected to only one sink in G \ S, at least |D| − 1 arcs outgoing from v are deleted by S. However, S does
not disconnect v from all sinks in D, since then S would delete all arcs outgoing from v, contradicting Observation 1.
Hence, exactly one sink s ∈ D must be reachable by v in G \ S. For each such sink s ∈ D, the partitioning set S has to
delete at least |D| − 1 arcs outgoing from v. We simply try out all these possibilities, which gives rise to the following
search tree algorithm.
Algorithm 1 starts with S = 0/ and processes the vertices of G in reverse topological order, that is, each vertex is
processed by procedure ”searchtree” after its out-neighbors. The procedure exploits the invariant that, when processing
5
a vertex v, each out-neighbor w of v has already been associated with the sink s that it reaches in G \ S (in terms of
Algorithm 1, L[w] = s). It then tries all possibilities of associating v with a feasible sink and augmenting S accordingly,
so that the invariant also holds for the vertex processed after v.
Specifically, if v is a sink, then line 7 associates v with itself. If all out-neighbors of v are associated with the same
sink s, then line 7 associates v with s. Otherwise, line 12 computes the set D of feasible sinks for v. In lines 14–17,
the algorithm branches into all possibilities of associating v with one of the |D| feasible sinks s ∈ D (by way of
setting L[v] ← s in line 15) and augmenting S so that v only reaches s in G \ S. That is, in each of the |D| branches, it
adds to S the arcs outgoing from v to the out-neighbors of v that are associated with sinks different from s (the weight
of S increases by at least |D| − 1). Then, line 17 continues with the next vertex in the reverse topological order. After
processing the last vertex, each vertex of G \ S can reach exactly one sink, that is, S is a partitioning set. If a branch
finds a partitioning set with weight at most k, line 10 outputs it.
We analyze the running time of this algorithm. To this end, we first bound the total number of times that procedure
“searchtree” is called. To this end, we first analyze the number of terminal calls, that is, calls that do not recursively call
the procedure. Let T (α) denote the maximum possible number of terminal calls caused by the procedure “searchtree”
when called with a set S satisfying ω(S) ≥ α, including itself if it is a terminal call. Note that procedure “searchtree”
calls itself only in line 17, that is, for each sink s of some set D of feasible sinks with 1 ≤ |D| − 1 ≤ k − α, it calls itself
with a set S0 of weight at least α + |D| − 1. Thus, we have
T (α) ≤ |D| · T (α + |D| − 1).
We now inductively show that T (α) ≤ 2k−α for 0 ≤ α ≤ k. Then, it follows that there is a total number of T (0) ≤
2k terminal calls. For the induction base case, observe that T (k) = 1, since, if procedure “searchtree” is called with a set S
0
of weight at least k, then any recursive call is prevented by the check in line 13. Now, assume that T (α 0 ) ≤ 2k−α holds
for all α 0 with α ≤ α 0 ≤ k. We show T (α − 1) ≤ 2k−(α−1) by exploiting 2 ≤ |D| ≤ 2|D|−1 as follows:
T (α − 1) ≤ |D| · T (α − 1 + |D| − 1) ≤ |D| · 2k−(α−1)−(|D|−1) ≤ |D| ·
2k−(α−1)
≤ 2k−(α−1) .
2|D|−1
It follows that there are at most T (0) = 2k terminal calls to procedure “searchtree”.
In order to bound the total number of calls to procedure “searchtree”, observe the following: if each inner node
of a tree has at least two children, then the number of inner nodes in a tree is at most its number of leaves. Now,
since procedure “searchtree” calls itself only in line 17, that is, for each sink s of some set D of feasible sinks
with 1 ≤ |D| − 1 ≤ k − α, each non-terminal call causes at least two new calls. Thus, since there are O(2k ) terminal
calls, there are also O(2k ) non-terminal calls.
It follows that there are O(2k ) total calls of procedure “searchtree”. For each such call, we iterate, in the worst
case, over all out-neighbors of all vertices in the graph in lines 6–12, which works in O(n + m) time. Moreover,
for each call of procedure “searchtree”, we compute a set S0 in line 16 in O(n + m) time. Hence, a total amount
of O(2k · (n + m)) time is spent in procedure “searchtree”. Initially, Algorithm 1 uses O(n + m) time to compute a
reverse topological ordering [9, Section 22.4].
The experimental results in Section 3.2 will show that Algorithm 1 alone cannot solve even moderately large instances.
Therefore, we complement it by linear-time executable data reduction rules that will allow for a significant speedup.
The following data reduction rule is illustrated in Figure 1.
Reduction Rule 1. If there is an arc (v, w) such that w can reach exactly one sink s , w and v can reach multiple sinks,
then
• if there is no arc (v, s), then add it with weight ω(v, w),
• otherwise, increase ω(v, s) by ω(v, w), and
delete the arc (v, w).
Note that, in the formulation of the data reduction rule, both v and w may be connected to an arbitrary number of sinks
by an undirected path. However, we require that w can reach exactly one sink and that v can reach multiple sinks, that
is, using a directed path.
6
v
v
w
2
s1
s2
s1
s2
Figure 1: The left side shows an input graph with unit weights to which Reduction Rule 1 is applicable. The right side shows the same graph to
which Reduction Rule 1 has been applied as often as possible. Since v can reach multiple sinks and its out-neighbors each can reach only one sink,
its arcs got redirected. Unlabeled arcs have weight one.
Lemma 1. Let (G, ω, k) be a DAG PARTITIONING instance and consider the graph G0 with weights ω 0 output by
Reduction Rule 1 applied to an arc (v, w) of G. Then (G, ω, k) is a yes-instance if and only if (G0 , ω 0 , k) is a yes-instance.
P ROOF. First, assume that (G, ω, k) is a yes-instance and that S is a minimal partitioning set of weight at most k
for G. We show how to transform S into a partitioning set of equal weight for G0 . We distinguish two cases: either S
disconnects v from s or not, where s is the only sink that w can reach.
Case 1) Assume that S disconnects v from s. Note that every subgraph of a directed acyclic graph is again a
directed acyclic graph and that every vertex in a directed acyclic graph is not only connected to, but also can reach some
sink. Hence, by Observation 1, S cannot disconnect w from s, since w can only reach s in G and would have to reach
some other, that is, new sink in G \ S. It follows that S contains the arc (v, w). Now, however, S0 := (S \ {v, w}) ∪ {(v, s)}
is a partitioning set for G0 , since G \ S = G0 \ S0 . Moreover, since ω 0 (v, s) = ω(v, s) + ω(v, w), we have ω 0 (S0 ) = ω(S),
where we, for convenience, declare ω(v, s) = 0 if there is no arc (v, s) in G.
Case 2) Assume that S does not disconnect v from s and, for the sake of a contradiction, that S is not a partitioning
set for G0 . Observe that S contains neither (v, w) nor (v, s), because it is a minimal partitioning set and does not
disconnect v from s. Therefore, G0 \ S differs from G \ S only in the fact that G0 \ S does not have the arc (v, w) but an
arc (v, s) that was possibly not present in G \ S. Hence, since S is a partitioning set for G but not for G0 , two sinks are
connected to each other in G0 \ S via an undirected path using the arc (v, s). Thus, one of the two sinks is s and the
undirected path consists of (v, s) and a subpath p between v and some sink s0 . Then, however, s is connected to s0 also
in G \ S via an undirected path between s and w (S cannot disconnect s from w by Observation 1), the arc (v, w) and the
undirected path p from v to s0 . This contradicts S being a partitioning set for G. We conclude that S is a partitioning set
for G0 . Moreover, since S contains neither (v, w) nor (v, s), one has ω(S) = ω 0 (S).
Now, assume that (G0 , ω 0 , k) is a yes-instance and that S is a minimal partitioning set of weight at most k for G0 .
We show how to transform S into a partitioning set of equal weight for G. Again, we distinguish between two cases:
either S disconnects v from s or not.
Case 1) Assume that S disconnects v from s. Then, (v, s) ∈ S. Now, S0 := S ∪ {(v, w)} is a partitioning set for G,
since G \ S0 = G0 \ S. Moreover, since ω 0 (v, s) = ω(v, s) + ω(v, w), we have ω(S0 ) = ω 0 (S), where we assume that
ω(v, s) = 0 if there is no arc (v, s) in G.
Case 2) Assume that S does not disconnect v from s and, for the sake of a contradiction, assume that S is not a
partitioning set for G. Then, since S is minimal, S does not contain (v, s). Now, observe that G \ S and G0 \ S differ only
in the fact that G \ S has an additional arc (v, w) and that, possibly, (v, s) is missing. Hence, since S is a partitioning set
for G0 but not for G, there is an undirected path between two sinks in G \ S through (v, w). Because S by Observation 1
cannot disconnect w from s, one of these sinks is s and the undirected path consists of a subpath between s and w,
the arc (v, w), and a subpath p between v and a sink s0 . Then, however, s and s0 are also connected in G0 \ S via the
arc (v, s) and the subpath p between v and s0 . This contradicts S being a partitioning set for G0 . Finally, since S does not
contain (v, s), one has ω(S) = ω 0 (S).
7
v
v
2
2
s1
s2
s1
s2
Figure 2: On the left side, a graph that cannot be reduced by Reduction Rule 1 is shown. The gray background highlights the non-sink vertices that
can only reach the sink s1 . The right side shows the same graph to which Reduction Rule 2 has been applied as often as possible. Unlabeled arcs
have weight one.
After applying Reduction Rule 1 exhaustively, that is, as often as it is applicable, we apply a second data reduction rule,
which is illustrated in Figure 2.
Reduction Rule 2. If, for some sink s, the set L of non-sink vertices that can reach only s is nonempty, then delete all
vertices in L.
Lemma 2. Let G be a graph that is exhaustively reduced with respect to Reduction Rule 1 and G0 := G − L be the
graph output by Reduction Rule 2 when applied to G for some sink s. Then, any partitioning set for G is a partitioning
set of equal weight for G0 and vice versa.
P ROOF. In order to prove the lemma, we first make three structural observations about the set L.
i) There is no arc (v, w) from a vertex v < L to a vertex w ∈ L in G: for the sake of a contradiction, assume that such
an arc exists. Then, since v < L and v is obviously not a sink, v can reach a sink s0 , s. It follows that v can reach
two sinks: s via w and s0 . This contradicts the assumption that Reduction Rule 1 is not applicable.
ii) There is no arc (v, w) from a vertex v ∈ L to a vertex w < L with w , s in G: for the sake of a contradiction, assume
that such an arc exists. Then, since w < L and w , s, it follows that w can reach a sink s0 different from s. Then,
also v can reach two sinks: s0 via w and s. This contradicts v ∈ L.
iii) A minimal partitioning set S for G does not contain any arc between vertices in L ∪ {s}: this is because, by
Observation 1, no minimal partitioning set S can disconnect any vertex v ∈ L from s, since otherwise v would
reach another, that is, new sink in G \ S.
Now, let S be a minimal partitioning set for G. Then, S is also a partitioning set for G0 , since G0 \ S = (G \ S) − L and
deleting L from G \ S cannot create new sinks.
In the opposite direction, let S be a partitioning set for G0 . Then, S is also a partitioning set for G, since G \ S is just
0
G \ S with the vertices in L and their arcs added. These, however, can reach only the sink s and are only connected to
vertices to which s is connected. Hence, they do not create new sinks or connect distinct components of G0 \ S.
We now show how to exhaustively apply both data reduction rules in linear time. To this end, we apply Algorithm 2: in
lines 1–4, it computes an array L such that, for each vertex v ∈ V , we have L[v] = {s} if v reaches exactly one sink s and
L[v] = 0/ otherwise. It uses this information to apply Reduction Rule 1 in lines 6–10 and Reduction Rule 2 in lines 11
and 12.
Lemma 3. Given a directed acyclic graph G with weights ω, in O(n + m) time Algorithm 2 produces a directed acyclic
graph G0 with weights ω 0 such that G0 is exhaustively reduced with respect to Reduction Rule 1 and Reduction Rule 2.
In particular, (G, ω, k) is a yes-instance if and only if (G0 , ω 0 , k) is a yes-instance.
8
Algorithm 2: Apply Reduction Rules 1 and 2 exhaustively.
Input: A directed acyclic graph G = (V, A) with arc weights ω.
Output: The result of exhaustively applying Reduction Rules 1 and 2 to G.
1 S ← sinks of G
2 L ← array with n entries
3 foreach v ∈ S do L[v] ← {v}
// L[v] = {s} for any sink s ∈ V will mean that v only reaches the sink s.
4 foreach v ∈ V \ S in reverse topological order do
\
5
L[v] ← L[u]
// Invariant: the intersection contains at most one sink.
u∈N + (v)
foreach v ∈ V with L[v] = 0/ do
foreach w ∈ N + (v) with L[w] = {s} for some s ∈ S and w < S do
if (v, s) < A then add (v, s) with ω(v, s) := 0 to A
ω(v, s) ← ω(v, s) + ω(v, w)
delete arc (v, w)
// Application of Reduction Rule 1
foreach v ∈ V \ S such that L[v] = {s} for some s ∈ S do
delete vertex v
// Application of Reduction Rule 2
12
13
return (G, ω)
6
7
8
9
10
11
P ROOF. We first discuss the semantics of Algorithm 2, then its running time. After line 5, L[v] = {s} for some vertex v
if v can reach exactly one sink s and L[v] = 0/ otherwise: this is, by definition, true for all L[v] with v ∈ S. For v ∈ V \ S
it also holds, since v can reach exactly one sink s if and only if all of its out-neighbors u ∈ N + (v) can reach s and no
other sinks, that is, if and only if L[u] = {s} for all out-neighbors u ∈ N + (v) of v.
Hence, the loop in lines 6–10 applies Reduction Rule 1 to all arcs to which Reduction Rule 1 is applicable.
Moreover, Reduction Rule 1 does not change which sinks are reachable from any vertex and, hence, cannot create new
arcs to which Reduction Rule 1 may be applied. Hence, when reaching line 11, the graph will be exhaustively reduced
with respect to Reduction Rule 1 and we do not have to update the array L.
The loop in lines 11 and 12 now applies Reduction Rule 2, which is allowed, since the graph is exhaustively reduced
with respect to Reduction Rule 1. Moreover, an application of Reduction Rule 2 cannot create new vertices to which
Reduction Rule 2 may become applicable or arcs to which Reduction Rule 1 may become applicable. Hence, line 13
indeed returns a graph that is exhaustively reduced with respect to both data reduction rules.
It remains to analyze the running time. Obviously, lines 1–3 of Algorithm 2 work in O(n) time. To execute
line 4 in O(n + m) time, we iterate over the vertices in V \ S in reverse topological order, which can be computed
in O(n + m) time [9, Section 22.4]. Hence, when computing L[v] for some vertex in line 4, we already know the
values L[u] for all u ∈ N + (v). Moreover, L[v] is the intersection of sets with at most one element and, therefore, also
contains at most one element. It follows that we can compute L[v] in O(|N + (v)|) time for each vertex v ∈ V \ S and,
therefore, in O(n + m) total time for all vertices. The rest of the algorithm only iterates once over all arcs and vertices.
Hence, to show that it works in O(n + m) time, it remains to show how to execute lines 8 and 9 in constant time.
Herein, the main difficulty is that an adjacency list cannot answer queries of the form “(v, s) ∈ A?” in constant time.
However, since we earlier required to iterate over all out-neighbors of a vertex v in O(|N + (v)|) time, we cannot just
use an adjacency matrix instead. We exploit a different trick, which, for the sake of clarity is not made explicit in the
pseudo code: assume that, when considering a vertex v ∈ V in line 6, we have an n-element array A such that A[s]
holds a pointer to the value ω(v, s) if (v, s) ∈ A and A[s] = ⊥ otherwise. Then, we could in constant time check in line 8
whether A[s] = ⊥ to find out whether (v, s) ∈ A and, if this is the case, get a pointer to (and increment) the weight ω(v, s)
in constant time in line 9. However, we cannot afford initializing an n-entry array A for each vertex v ∈ V and we cannot
make assumptions on the value of uninitialized entries. Luckily, we access A[s] only if there is a vertex w ∈ N + (v)
with L[w] = {s} for some s. Hence, we can create an n-entry array A once in the beginning of the algorithm and then,
between lines 6 and 7, set up A for v ∈ V as follows: for each w ∈ N + (v) with L[w] = {s}, set Av [s] := ⊥. Then, for
each s ∈ N + (v) with s ∈ S, let Av [s] point to ω(v, s).
9
Lemma 3 shows that we can exhaustively apply the two Reduction Rules 1 and 2 in linear time using Algorithm 2. A
natural approach for evaluating the quality of our preprocessing would be to provide a performance guarantee in terms
of small problem kernels. Unfortunately, in Section 3.3 we show that, under widely accepted complexity-theoretic
assumptions, there is no problem kernel with size polynomial in k for DAG PARTITIONING. Nevertheless, the next
section shows that our data reduction technique performs remarkably well in empirical tests. Furthermore, we show
that the running time of Algorithm 1 is significantly improved when it is applied after Algorithm 2. We achieve the
largest speedup of Algorithm 1 by interleaving the application of Algorithm 2 and the branching; a technique that
generally works well for search tree algorithms [30].
3.2. Experimental evaluation
In this section, we aim for giving a proof of concept of our O(2k · (n + m)) time search tree algorithm presented
in Section 3.1 by demonstrating to which extent instances of DAG PARTITIONING are solvable within a time frame
of five minutes. Moreover, using the optimal solutions found by our algorithm, we evaluate the quality of a heuristic
presented by Leskovec et al. [26], which can be considered a variant of our search tree algorithm (Algorithm 1): The
difference is that, while our search tree algorithm branches into all possibilities of putting a vertex into a connected
component with some sink, the heuristic just puts each vertex into the connected component with the sink it would be
most expensive to disconnect from. This is the strategy described by Leskovec et al. [26] as yielding the best results
and is in more detail described by Suen et al. [33]. The pseudocode of the heuristics (Algorithm 3) can be found in
Section 4.1, where we prove that the heuristic works optimally on trees.
Implementation details. We implemented the search tree algorithm as well as the heuristic in three variants:
1. without data reduction,
2. with initially applying the data reduction algorithm presented in Algorithm 2, and
3. with interleaving the data reduction using Algorithm 2 with the branching of Algorithm 1.
The source code uses about 1000 lines of C++ and is freely available.2 The experiments were run on a computer with a
3.6 GHz Intel Xeon processor and 64 GiB RAM under Linux 3.2.0, where the source code has been compiled using the
GNU C++ compiler version 4.7.2 and using the highest optimization level (-O3).
Data. We tried to apply our algorithm to the data set described by Leskovec et al. [26]; unfortunately, its optimum
partitioning sets have too large weight to be found by our algorithm.3 In order to prove the feasibility of solving large
instances with small minimum partitioning sets, we generated artificial instances. Herein, however, we stick to the
clustering motivation of DAG PARTITIONING and test our algorithm using simulated citation networks: vertices in a
graph represent articles and if an article v cites an article w, there is an arc (v, w). Herein, we consider only directed
acyclic graphs, which model that an article only cites older articles. A partitioning of such a network into connected
components of which each contains only one sink can be interpreted as a clustering into different topics of which we
want to identify the origins.
To simulate citation networks, we employ preferential attachment graphs—a random graph model commonly used
to model citations between articles [31, 2, 21]. Preferential attachment graphs model the natural growth of citation
networks, in which new articles are published over time and with high probability cite the already highly-cited articles.
Indeed, Jeong et al. [21] empirically verified that in this preferential attachment model, the probability of an article
being cited is linear in the number of times the article has been cited in the past.
To create a preferential attachment graph, we first choose two parameters: the number c of sinks to create and the
maximum outdegree d of each vertex. After creating c sinks, n new vertices are introduced one after another. After
introducing a vertex, we add to it d outgoing unit-weight arcs: for each of these d outgoing arcs, the target is chosen
independently at random among previously introduced vertices such that each vertex is chosen as the target with a
2 http://fpt.akt.tu-berlin.de/dagpart/
3 The
exact weights of the optimum partitioning sets remain unknown, since our algorithm could not compute them within several hours.
10
interleaved data reduction
initial data reduction
no data reduction
400
running time [s]
running time [s]
6
4
2
0
160
300
200
100
0
170
180
190
200
weight k of optimal partitioning set
160
170
180
190
200
weight k of optimal partitioning set
Figure 3: Comparisons of the running time of Leskovec et al. [26]’s heuristic (left) with the running time of our search tree algorithm (right). All
graphs have 106 vertices and roughly 18 · 106 arcs and were generated by adding k random arcs between ten connected components, each being a
preferential attachment graph on 105 vertices with outdegree twenty and a single sink. The heuristic solved all 40 instances optimally. On the right
hand side, no instance was solved in less than an hour without data reduction.
probability proportional to its indegree plus one. We do not add an outgoing arc twice, which might result in a vertex
having less than d outgoing arcs if the target nodes of two arcs to be added coincide.
We compared our algorithm to the heuristic of Leskovec et al. [26] on these graphs, but our algorithm could solve
only instances with up to 300 arcs optimally, since the optimum solution weight grows too quickly in the sizes of the
generated graphs. To show the running time behavior of our algorithm on larger graphs with small solution sizes, we
employ an additional approach: we generate multiple connected components, each being a preferential attachment
graph with one sink, and randomly add k additional arcs between these connected components in a way so that the graph
remains acyclic. Then, obviously, an optimal partitioning set cannot be larger than k. We call the set of k randomly
added arcs embedded partitioning set and it can be viewed as noise in data that clusters well.
Experimental results. Figure 3 compares the running time of the heuristic of Leskovec et al. [26] to the running time
of our Algorithm 1 with increasing optimal partitioning set size k. On the left side, it can be seen that using the data
reduction from Algorithm 2 slows down the heuristic. This is not surprising, since the heuristic itself is implemented to
run in linear time and, hence, instead of first shrinking the input instance by Algorithm 2 in linear time, one might
right away solve the instance heuristically. On the right side, one can observe that, as expected, the running time of
Algorithm 1 increases exponentially in k. We only show the running time of the implementations with data reduction:
without data reduction, we could not solve any instance in less than an hour. We can solve instances with k ≤ 190
optimally within five minutes. This allowed us to verify that the heuristic solved all 40 generated instances optimally,
regardless of the type of data reduction applied.
Figure 4 compares the running time of the heuristic of Leskovec et al. [26] to the running time of our Algorithm 1
with increasing graph size. While the heuristic shows a linear increase of running time with the graph size (on the left
side), such a behavior cannot be observed for the search tree algorithm (on the right side). The reason for this can be
seen in Figure 5: the data reduction applied by Algorithm 2 initially shrinks most input instances to about 2000 arcs
in less than ten seconds. Thus, what we observe in the right plot of Figure 4 is, to a large extent, the running time of
Algorithm 1 for constant k = 190 and roughly constant graph size. Our search tree algorithm allowed us to verify that
the heuristic by Leskovec et al. [26] solved all 80 generated instances optimally regardless of the type of data reduction
applied.
Finally, Figure 6 presents instances that could not be optimally solved by Leskovec et al. [26]’s heuristic. In the left
plot, we see that in instances with large embedded partitioning sets of several hundred thousand arcs, the heuristic of
Leskovec et al. [26] does not find the embedded partitioning set but an about 5‰ larger one. In all cases, the heuristic
11
interleaved data reduction
initial data reduction
no data reduction
400
running time [s]
running time [s]
30
20
10
300
200
100
0
20
40
60
input arcs 106
0
80
20
40
60
input arcs 106
80
Figure 4: Comparisons of the running time of Leskovec et al. [26]’s heuristic (left) with the running time of our search tree algorithm (right). Without
interleaved data reduction, the search tree solved no instance in less than 5 minutes. The graphs were generated by adding k = 190 random arcs
between ten connected components, each being a preferential attachment graph with outdegree twenty and a single sink. The heuristic solved all
80 instances optimally.
8
running time [s]
output arcs
6,000
4,000
2,000
6
4
2
20
40
60
6
input arcs 10
0
80
20
40
60
6
input arcs 10
80
Figure 5: Effect (left) and running time (right) of initially running Algorithm 2 for data reduction. The graphs were generated by adding k = 190
random arcs between ten connected components, each being a preferential attachment graph with outdegree twenty and a single sink.
12
1.006
2.5
deviation factor
deviation factor
3
1.004
2
1.5
1
100
300
weight of embedded partitioning set 103
150
200
250
10
20
15
25
weight k of optimum partitioning set
Figure 6: Comparison of partitioning sets found by Leskovec et al. [26]’s heuristic with an embedded partitioning set of size k (left) and an optimal
partitioning set (right). All graphs on the left side have 106 vertices and roughly 18 · 106 arcs and were generated by adding k random arcs between
ten connected components, each being a preferential attachment graph on 105 vertices with outdegree fifteen and a single sink. All graphs on the
right side are preferential attachment graphs with varying number of vertices, two sinks and outdegree three.
found the same partitioning sets regardless of the type of data reduction applied. Note that the plot only gives a lower
bound on the deviation factor, since there might be even better partitioning sets in the instances than the embedded
one; we were unable to compute the optimal partitioning sets in these instances. In the right plot of Figure 6, we used
smaller preferential attachment graphs (this time without embedded partitioning sets) and see that Leskovec et al. [26]’s
heuristic can be off by more than a factor of two from the optimal partitioning set. Data reduction had no effect on the
quality of the partitioning sets found.
Summary. We have seen that solving large instances with partitioning sets of small weight is realistic using our
algorithm. In particular, instances with more than 107 arcs and k ≤ 190 could be solved in less than five minutes. A
crucial ingredient in this success is the data reduction executed by Algorithm 2; without its help, we could not solve
any of our instances in less than five minutes.
However, we also observed that our algorithm works best on those instances that can already be solved mostly optimally by Leskovec et al. [26]’s heuristic and that the data reduction executed by Algorithm 2 slows down the heuristic.
Having seen that the heuristic by Leskovec et al. [26] can be off by more than a factor of two from the optimum on
random preferential attachment graphs diminishes the hope that, in spite of the non-approximability results of DAG
PARTITIONING by Alamdari and Mehrabian [1], the heuristic of Leskovec et al. [26] might find good approximations
on naturally occurring instances. As we see, we do not have to construct adversarial instances to make the heuristic
find solutions far from optimal.
3.3. Limits of data reduction and fixed-parameter algorithms
In Section 3.1, we have seen linear-time data reduction rules for DAG PARTITIONING. Moreover, the experiments
in Section 3.2 have shown that, on all input instancse we tested our algorithm on, the running time of our O(2k · (n +
m)) time Algorithm 1 merely depended on k because Algorithm 2 shrunk our random input instances to roughly the
same size.
Therefore it is natural to ask whether we can provide data reduction rules that provably shrink the size of each
possible input instance to some fixed polynomial in k, that is, whether there is a polynomial-size problem kernel for
DAG PARTITIONING. Unfortunately, in this section, we give a negative answer to this question. Specifically, we
prove that DAG PARTITIONING does not admit problem kernels with size polynomial in k, unless NP ⊆ coNP/poly.
Moreover, we show that the running time O(2k · (n + m)) of Algorithm 1 cannot be improved to 2o(k) poly(n), unless
the Exponential Time Hypothesis fails. Herein, the Exponential Time Hypothesis as well as NP * coNP/poly are
hypotheses stronger than P , NP, but widely accepted among complexity theorists [20, 16].
13
f
t
xt1
x1f
xt2
x2f
x1
x̄1 C1
x2
x̄2
f0
t0
Figure 7: DAG PARTITIONING instance constructed from the formula consisting only of the clause C1 := (x1 ∨ x̄2 ). Heavy arcs are drawn bold;
dotted arcs are a partitioning set that corresponds to the satisfying assignment setting x1 to true and x2 to false. The variable gadgets are drawn on a
gray background.
Towards proving these results, we first recall the polynomial-time many-to-one reduction from 3-S AT to DAG
PARTITIONING given by Alamdari and Mehrabian [1]. The 3-S AT problem is, given a formula ϕ in conjunctive normal
form with at most three literals per clause, to decide whether ϕ admits a satisfying assignment.
Construction 1 (Alamdari and Mehrabian [1]). Let ϕ be an instance of 3-S AT with the variables x1 , . . . , xn and the
clauses C1 , . . . ,Cm . We construct a DAG PARTITIONING instance (G, ω, k) with k := 4n + 2m that is a yes-instance if
and only if ϕ is satisfiable. The weight function ω will assign only two different weights to the arcs: a normal arc has
weight one and a heavy arc has weight k + 1 and thus cannot be contained in any partitioning set of weight k. The
remainder of this construction is illustrated in Figure 7.
We start constructing the directed acyclic graph G by adding the special vertices f , f 0 ,t, and t 0 together with the
heavy arcs ( f , f 0 ) and (t,t 0 ). The vertices f 0 and t 0 will be the only sinks in G. For each variable xi , introduce the
vertices xti , xif , xi and x̄i together with the heavy arcs (t, xti ) and ( f , xif ) and the normal arcs (xti , xi ), (xti , x̄i ), (xif , xi ),
(xif , x̄i ), (xi , f 0 ), (x̄i , f 0 ), (xi ,t 0 ), and (x̄i ,t 0 ). For each clause C j , add a vertex C j together with the heavy arc (t,C j ).
Finally, if some clause C j contains the literal xi , then add the arc (C j , xi ); if some clause C j contains the literal x̄i , then
add the arc (C j , x̄i ).
Alamdari and Mehrabian [1] showed that, given a formula ϕ in 3-C NF with n variables and m clauses, Construction 1
outputs a graph G with arc weights ω such that ϕ is satisfiable if and only if there is a partitioning set S for G that does
not contain heavy arcs.
Since G has only the two sinks t 0 and f 0 , by Observation 1, a minimal such partitioning set has to partition G into two
connected components, one connected component containing the heavy arc (t,t 0 ) and the other containing ( f , f 0 ). Moreover, if ϕ is satisfiable, then such a partitioning set has weight at most 4n + 2m, since for each xi of the n variables of ϕ,
it deletes at most one of two arcs outgoing from each of the vertices xti , xif , xi , and x̄i , and for each C j of the m clauses,
it deletes at most two out of the three arcs outgoing from the clause vertex C j . We thus obtain the following lemma:
Lemma 4. Given a formula ϕ in 3-C NF with n variables and m clauses, Construction 1 outputs a graph G with arc
weights ω such that ϕ is satisfiable if and only if there is a partitioning set S for G that does not contain heavy arcs.
Moreover, if ϕ is satisfiable, then S has weight at most 4n + 2m and partitions G into one connected component
containing the constructed vertices t and t 0 and the other containing f and f 0 .
We will later exploit Lemma 4 to show our hardness results. Next, we show that arcs with non-unit weights in our
constructions can be simulated by arcs with unit weights. This allows us to show stronger hardness results and to keep
our constructions simple.
3.3.1. Strengthening of hardness results to unit-weight graphs
Construction 1 heavily relied on forbidding the deletion of certain arcs by giving them a high weight. The next
lemma shows that we can replace these arcs by a gadget only using unit-weight arcs without changing the weight of the
partitioning set sought.
14
v
5
w
v
w
Figure 8: Replacing an arc of weight 5 on the left by the gadget of unit-weight arcs on the right.
Lemma 5. There is a polynomial-time many-one reduction from DAG PARTITIONING with polynomially bounded
weights to unweighted DAG PARTITIONING that does not change the weight k of the partitioning set sought.
P ROOF. Let (G, ω, k) be an instance of DAG PARTITIONING. We show how to obtain an instance (G0 , ω 0 , k) by
replacing a single arc of weight more than one by arcs of weight one such that (G, ω, k) is a yes-instance if and only
if (G0 , ω 0 , k) is a yes-instance. The replacement will be done as illustrated in Figure 8. The claim then follows by
repeating this procedure for every arc of weight more than one.
Consider an arc a = (v, w) in G with ω(a) > 1. We obtain G0 and ω 0 from G and ω by setting the weight ω 0 (a) = 1,
adding a set X of ω(a) − 1 vertices to G0 , and inserting for each u ∈ X a weight-one arc (v, u) and a weight-one
arc (u, w).
First, assume that (G, ω, k) is a yes-instance and that S is a minimal partitioning set for G. We show how to obtain
a partitioning set of weight k for G0 . Clearly, if a < S, then S is a partitioning set of equal weight for (G0 , ω 0 , k). If a ∈ S,
then we get a partitioning set of equal weight for (G0 , ω 0 , k) by adding the arcs between v and X to S.
Second, assume that (G0 , ω 0 , k) is a yes-instance and that S is a minimal partitioning set for G0 . We show how to
obtain a partitioning set of weight k for G. To this end, we consider two cases: v and w are in a common or in separate
connected components of G0 \ S.
Case 1) If v and w are in one connected component of G0 \ S, then, by minimality, S does not contain a or any arc
incident to vertices in X. Hence, S is a partitioning set of equal weight for G.
Case 2) If v and w are in separate connected components of G0 \ S, then a ∈ S. Moreover, the vertices in X have
only one outgoing arc. Hence, by Observation 1, S does not contain arcs from X to w but, therefore, contains all arcs
from v to X. Removing these arcs from S results in a partitioning set of equal weight for (G, ω, k).
3.3.2. Limits of fixed-parameter algorithms
We now show that DAG PARTITIONING cannot be solved in 2o(k) poly(n) time unless the Exponential Time
Hypothesis fails. Thus, if our search tree algorithm for DAG PARTITIONING can be improved, then only by replacing
the base of the exponential 2k -term by some smaller constant.
The Exponential Time Hypothesis was introduced by Impagliazzo and Paturi [19] and states that n-variable 3-S AT
cannot be solved in 2o(n) poly(n) time. Using the reduction from 3-S AT to DAG PARTITIONING given by Alamdari
and Mehrabian [1] (Construction 1), we can easily show the following:
Theorem 2. Unless the Exponential Time Hypothesis fails, DAG PARTITIONING cannot be solved in 2o(k) poly(n)
time even if all arcs have unit weight.
P ROOF. Construction 1 reduces an instance of 3-S AT consisting of a formula with n variables and m clauses to an
equivalent instance (G, ω, k) of DAG PARTITIONING with k = 4n + 2m. Thus, a 2o(k) poly(n)-time algorithm for DAG
PARTITIONING would yield a 2o(m) poly(m)-time algorithm for 3-S AT. This, in turn, by the so-called Sparsification
Lemma of Impagliazzo et al. [20, Corollary 2], would imply a 2o(n) poly(n) time algorithm for 3-S AT, which contradicts
the Exponential Time Hypothesis. Since the weights used in Construction 1 are polynomial in the number of created
vertices and edges, we can apply Lemma 5 to transfer the result to the unit-weight case.
15
3.3.3. Limits of problem kernelization
We now show that DAG PARTITIONING has no polynomial-size problem kernel with respect to the parameter k—
the weight of the partitioning set sought. It follows that, despite the effectiveness of data reduction observed in
experiments in Section 3.2, we presumably cannot generally shrink a DAG PARTITIONING instance in polynomial
time to a size polynomial in k.
To show that DAG PARTITIONING does not allow for polynomial-size kernels, we first provide the necessary
concepts and techniques introduced by Bodlaender et al. [8].
Definition 1 (Bodlaender et al. [8, Definition 3.3]). For some finite alphabet Σ, a problem L ⊆ Σ∗ (OR-)cross-composes
into a parameterized problem Q ⊆ Σ∗ × N if there is an algorithm (a (OR-)cross-composition) that transforms instances x1 , . . . , xs of L into an instance (x∗ , k) for Q in time polynomial in ∑si=1 |xi | such that
i) k is bounded by a polynomial in maxsi=1 |xi | + log s and
ii) (x∗ , k) ∈ Q if and only if there is an i ∈ {1, . . . , s} such that xi ∈ L.
Furthermore, the cross-composition may exploit that the input instances x1 , . . . , xs belong to the same equivalence class
of a polynomial equivalence relation R ⊆ Σ∗ × Σ∗ , which is an equivalence relation such that
i) it can be decided in polynomial time whether two instances are equivalent and
ii) any finite set S ⊆ Σ∗ is partitioned into poly(maxx∈S |x|) equivalence classes.
The assumption that all instances belong to the same equivalence class of a polynomial equivalence relation can make
the construction of a cross-composition remarkably easier: when giving a cross-composition from 3-S AT, we can, for
example, assume that the input instances all have the same number of clauses and variables.
Cross-compositions can be used to prove that a parameterized problem has no polynomial-size kernel unless
NP ⊆ coNP/poly.
Theorem 3 (Bodlaender et al. [8, Corollary 3.6]). If some problem L ⊆ Σ∗ is NP-hard under polynomial-time manyone reductions and L cross-composes into the parameterized problem Q ⊆ Σ∗ × N, then there is no polynomial-size
problem kernel for Q unless NP ⊆ coNP/poly.
In the following, we show that 3-S AT cross-composes into DAG PARTITIONING parameterized by k, which yields the
following theorem:
Theorem 4. Unless NP ⊆ coNP/poly, DAG PARTITIONING does not have a polynomial-size problem kernel with
respect to the weight k of the partitioning set sought even if all arcs have unit weight.
Although the proof of Theorem 4 is based on the following construction, which requires arc weights, by using Lemma 5
from Section 3.3.1, we obtain that Theorem 4 holds even on graphs with unit weights.
Construction 2. Let ϕ1 , . . . , ϕs be instances of 3-S AT. Since we may assume ϕ1 , . . . , ϕs to be from the same equivalence class of a polynomial equivalence relation, we may assume that each of the formulas ϕ1 , . . . , ϕs has the same
number n of variables and the same number m of clauses. Moreover, we may assume that s is a power of two;
otherwise we simply add unsatisfiable formulas to the list of instances. We now construct a DAG PARTITIONING
instance (G, ω, k) with k := 4n + 2m + 3 log s that is a yes-instance if and only if ϕi is satisfiable for at least one 1 ≤ i ≤ s,
where we use “log” to denote the binary logarithm. As in Construction 1, the weight function ω will only assign two
possible weight values: a heavy arc has weight k + 1 and thus cannot be contained in any partitioning set. A normal arc
has weight one. The remainder of the construction is illustrated in Figure 9.
For each instance ϕi , let Gi be the graph produced by Construction 1. By Lemma 4, Gi can be partitioned with
4n + 2m arc deletions if and only if ϕi is a yes-instance. We now build a gadget that, by means of additional 3 log s arc
deletions, chooses exactly one graph Gi that has to be partitioned.
To distinguish between multiple instances, we denote the special vertices f , f 0 ,t, and t 0 in Gi by fi , fi0 ,ti , and ti0 .
For all 1 ≤ i ≤ s, we add Gi to the output graph G and merge the vertices f1 , f2 , . . . , fs into a vertex f and the vertices f10 , f20 , . . . , fs0 into a vertex f 0 . Furthermore, we add the vertices t,t 0 , and t 00 and the heavy arcs (t,t 0 ) and (t,t 00 ) to G.
16
t 00
t
t1
t2
t10
t20
f
O
t3
t4
t30
t40
f0
t0
I
Figure 9: Cross composition of four formulas φ1 , . . . , φ4 into a DAG PARTITIONING instance. Of each subgraph Gi corresponding to formula φi , only
the vertices ti , ti0 , and their connecting heavy arc are shown. The introduced binary trees O and I are highlighted in gray. Deleting the 3 log 4 = 6 dotted
arcs requires the graph G1 to be partitioned, since its vertices t1 and t10 are in a different connected component than its vertices f1 = f and f10 = f 0 .
The graphs Gi for i > 1 do not have to be partitioned; they are completely contained in one connected component with f and f 0 .
We add a balanced binary tree O rooted in t 00 and its leaves being the vertices t1 , . . . ,ts that is formed by normal
arcs directed from the root to the leaves. That is, O is an out-tree. Moreover, add a balanced binary tree I rooted in t 0
and its leaves being the vertices t10 , . . . ,ts0 that is formed by normal arcs directed from the leaves to the root. That is,
I is an in-tree. For each vertex v , t 0 in I, add a normal arc (v, f 0 ).
Using this construction, we can now prove Theorem 4.
P ROOF ( OF T HEOREM 4). We only have to show that the instance (G, ω, k) constructed by Construction 2 is a yesinstance if and only if at least one of the input formulas ϕi is satisfiable. Then, the theorem for the weighted case
follows from Theorem 3. Since the weights used in Construction 2 are polynomial in the number of created vertices
and edges, we can apply Lemma 5 to transfer the result to the unit-weight case.
First, assume that a formula ϕi is satisfiable for some 1 ≤ i ≤ s. By Lemma 4 it follows that Gi can be partitioned
by k0 := 4n + 2m arc deletions into two connected components Pt and Pf such that Pt contains ti and ti0 and such
that Pf contains fi = f and fi0 = f 0 . We apply these arc deletions to G and delete 3 log s additional arcs from G
as follows. Let L be the unique directed path in O from t 00 to ti . Analogously, let L0 be the unique directed path
in I from ti0 to t 0 . Observe that each of these directed paths has log s arcs. We partition G into the connected
component Pt0 = Pt ∪ {t,t 0 ,t 00 } ∪ V (L) ∪ V (L0 ) with sink t 0 and into the connected component Pf0 = V (G) \ Pt0 with
sink f 0 . To this end, for each vertex v ∈ V (L) \ {ti }, we remove the outgoing arc that does not belong to L. Hence,
exactly log s arcs incident to vertices of O are removed. Similarly, for each vertex v ∈ V (L0 ) \ {ti0 }, we remove the
incoming arc not belonging to L0 . For each vertex v , t 0 of L0 , we remove the arc to f 0 . Hence, exactly 2 log s arcs
incident to vertices of I are removed. Thus, in total, at most k = 4n + 2m + 3 log s normal arcs are removed to partition G
into Pt0 and Pf0 .
Conversely, let S be a minimal partitioning set for G with ω(S) ≤ k. Then, by Observation 1, G \ S has two
connected components, namely Pt0 with sink t 0 and Pf0 with sink f 0 . Since S cannot contain heavy arcs, t and t 00 are in Pt0 .
Hence, t 00 can reach t 0 in G \ S, since they are in the same component of G \ S. As every directed path from t 00 to t 0
goes through some vertices ti and ti0 , it follows that there is an i ∈ {1, . . . s} such that ti and ti0 are in Pt0 . Since f = fi
and f 0 = fi0 are in Pf0 , the partitioning set S ∩ A(Gi ) partitions Gi into two connected components: one containing ti
and ti0 and the other containing f = fi and f 0 = fi0 . Since S does not contain heavy arcs, from Lemma 4 it follows that
ϕi is satisfiable.
17
Algorithm 3: Leskovec et al. [26]’s heuristic to compute small partitioning sets.
Input: A directed acyclic graph G = (V, A) with arc weights ω.
Output: A partitioning set S.
1 (v1 , v2 , . . . , vn ) ← reverse topological order of the vertices of G
2 L ← array with n entries
3 S←0
/
4 for i = 1 to n do
5
if vi is a sink then L[vi ] ← vi
// associate vi with itself
6
else
7
D ← {L[w] | w ∈ N + (vi )}
// the set of feasible sinks for vi
8
for s ∈ D do
9
As ← {(vi , w) ∈ A | w ∈ N + (vi ) ∧ L[w] = s}
// set of arcs to keep if vi is associated with sink s
10
11
12
13
s∗ ← arg maxs∈D ω(As )
L[vi ] ← s∗
S ← S ∪ {(vi , w) | w ∈ N + (vi ) ∧ L[w] , s∗ }
// cheapest sink s∗ to associate vi with
return S
4. Parameter treewidth
In Section 3, we have seen that DAG PARTITIONING is linear-time solvable when the weight k of the requested
partitioning set is constant. Alamdari and Mehrabian [1] asked whether DAG PARTITIONING is fixed-parameter
tractable with respect to the parameter treewidth, which is a measure of the “tree-likeness” of a graph. We will answer
this question affirmatively.
In Section 4.1, we first show that, if the input graph is indeed a tree, then the heuristic by Leskovec et al. [26] solves
the instance optimally in linear time. Afterwards, in Section 4.2, we prove that this result can be generalized to graphs
of bounded treewidth and thus improve the algorithm for pathwidth given by Alamdari and Mehrabian [1], since the
treewidth of a graph is at most its pathwidth.
4.1. Partitioning trees
In this section, we show that the heuristic by Leskovec et al. [26] solves DAG PARTITIONING in linear time on
trees. This result will be generalized in the next section, where we show how to solve DAG PARTITIONING in linear
time on graphs of bounded treewidth. The heuristic by Leskovec et al. [26] is similar to our search tree algorithm
presented in Algorithm 1: instead of trying all possibilities of associating a vertex with one of its feasible sinks, it
associates each vertex with the sink that it would be most expensive to disconnect from. The algorithm is presented in
Algorithm 3.
Theorem 5. Algorithm 3 solves DAG PARTITIONING optimally in linear time if the underlying undirected graph is a
tree.
P ROOF. Algorithm 3 clearly works in linear time: to implement it, we only have to iterate over the out-neighbors of
each vertex vi once. In particular, all sets As in line 9 can be computed by one iteration over each w ∈ N + (vi ) and
adding the arc (vi , w) to AL[w] . Moreover, Algorithm 3 returns a partitioning set: each vertex v is associated with exactly
one sink L[v] of G and the returned set S deletes exactly the arcs (v, w) for which L[v] , L[w].
We show by induction on i that Algorithm 3 computes a minimum-weight partitioning set for G[{v1 , . . . , vi }]. For
the induction base case, observe that v1 is a sink and, thus, v1 only reaches the sink L[v1 ] = v1 in G \ S for all possible
minimum-weight partitioning sets S for G. Now, assume that there is a minimum-weight partitioning set S such that
each v ∈ {v1 , . . . , vi−1 } only reaches the sink L[v] in G \ S. We show that there is a minimum-weight partitioning set S0
such that each v ∈ {v1 , . . . , vi } reaches only the sink L[v] in G \ S0 . If vi only reaches L[vi ] in G \ S, then we are done.
Otherwise, vi reaches some sink s0 , L[vi ] in G \ S and, hence, is not itself a sink. The graph G is partly illustrated in
Figure 10.
18
v
vi
w
Rs0
RL[vi ]
s0
u
L[vi ]
Figure 10: Illustration of the proof of Theorem 5. Straight lines represent arcs. Wavy lines represent directed paths. Dashed arcs and paths represent
arcs and paths that cannot exist since the underlying undirected graph is a tree.
Let Rs be the set of vertices reachable from vi that reach some sink s in G. Since the underlying undirected graph
of G is a tree, vi has exactly one arc into Rs for each sink s reachable from vi . Let (vi , u) be the arc of vi into RL[vi ] and
(vi , w) be the arc of vi into Rs0 . Observe that the arc (vi , u) exists since the algorithm can set L[vi ] only to sinks reachable
from vi . To show that S0 := (S \ {(vi , w)}) ∪ {(vi , u)} is still a partitioning set, we only have to verify that vi and all
vertices reaching vi only reach one sink in G \ S0 . For all other vertices, this follows from S being a partitioning set.
1. The vertex vi only reaches the sink L[vi ] in G \ S0 : this is because u = v j for some j < i which, by induction
hypothesis, reaches exactly one sink in G \ S and, hence, in G \ S0 .
2. A vertex v that reaches vi in G \ S0 reaches only the sink L[vi ] in G \ S0 : otherwise v reaches s0 in G \ S0 since vi
and, therefore v, reaches s0 in G \ S. This, however, means that v has a path to s0 that bypasses vi in G \ S0 and
hence, in G \ S, which contradicts the undirected underlying graph of G \ S being a tree.
It remains to show ω(S0 ) ≤ ω(S), implying that S0 is also a minimum-weight partitioning set. To see this, we analyze
the sets As0 and AL[vi ] computed in line 9 of Algorithm 3. Observe that As0 = {(vi , w)} and that AL[vi ] = {(vi , u)}.
Since ω(As0 ) ≤ ω(AL[vi ] ) because of the choice of L[vi ] in line 10, we conclude that ω(vi , u) ≤ ω(vi , w) and hence,
ω(S0 ) ≤ ω(S).
4.2. Partitioning DAGs of bounded treewidth
We now give an algorithm that solves DAG PARTITIONING in linear time on graphs of bounded treewidth. In
contrast to Section 3, which presented our search tree algorithm and an experimental evaluation thereof, the algorithm
presented below is of rather theoretical interest: Alamdari and Mehrabian [1] asked whether DAG PARTITIONING is
fixed-parameter tractable with respect to the parameter treewidth. With a dynamic programming algorithm, we can prove
the following theorem, which answers their open question and is an improvement of Alamdari and Mehrabian [1]’s
algorithm, since the treewidth of a graph is at most its pathwidth.
Theorem 6. Given a width-t tree decomposition of the underlying undirected graph, DAG PARTITIONING can be
2
solved in 2O(t ) · n time.
We first formally define the tree decomposition of a graph and its width.
Definition 2 (Treewidth, tree decomposition). Let G = (V, A) be a directed graph. A tree decomposition (T, β ) for G
consists of a rooted tree T = (X, E) and a mapping β : X → 2V of each node x of the tree T to a subset Vx := β (x) ⊆ V ,
called bag, such that
i) for each vertex v ∈ V , there is a node x of T with v ∈ Vx ,
ii) for each arc (u, w) ∈ A, there is a node x of T with {u, w} ⊆ Vx ,
19
s4
Vx
s3
Ux
s1
s2
Figure 11: The set S consisting of the dotted arc is not a partitioning set for the shown graph G, but a partial partitioning set for G[Ux ]. Note that
S would not be a partial partitioning set if it would additionally contain the bold arc, since then the sink s1 would be in a connected component with a
vertex of Vx but not reachable from Vx . Also note that S does not separate s2 from s3 , which are sinks in G[Ux ] but not in G.
iii) for each vertex v ∈ V , the nodes x of T for which v ∈ Vx induce a subtree in T .
A tree decomposition is nice if Vr = 0/ for the root r of T and each node x of T is either
• a leaf : then, Vx = 0,
/
• a forget node: then, x has exactly one child node y and Vx = Vy \ {v} for some v ∈ Vy ,
• an introduce node: then, x has exactly one child node y and Vx = Vy ∪ {v} for some v ∈ V \Vy , or
• a join node: then, x has exactly two child nodes y and z such that Vx = Vy = Vz .
The width of a tree decomposition is one less than the size of its largest bag. The treewidth of a graph G is the minimum
width of a tree decomposition for G. For a node x of T , we denote by Ux the union of Vy for all descendants y of the
node x.
For any constant t, it can be decided in linear time whether a graph has treewidth t and the corresponding tree decomposition of width t can be constructed in linear time [6]. Also in O(tn) time, the tree decomposition of width t can
be transformed into a nice tree decomposition with the same width and O(tn) nodes [24]. Hence, we assume without
loss of generality that we are given a nice tree decomposition.
Our algorithm is based on leaf-to-root dynamic programming. That is, intuitively, we start from the leaf nodes of the
tree decomposition and compute possible partial partitioning sets for each bag from the possible partial partitioning sets
for its child bags. Since our algorithm for DAG PARTITIONING on graphs of bounded treewidth is relatively intricate,
we refer the reader that is yet inexperienced with dynamic programming on tree decompositions to introductory chapters
in corresponding text books [28, 12, 10, 24].
We will now precisely define a partial partitioning set and show that any partial partitioning set for the root bag is a
partitioning set for the entire graph. The definition is illustrated in Figure 11.
Definition 3 (Partial partitioning set). A partial partitioning set S for G[Ux ] is an arc set S ⊆ A(G[Ux ]) such that
(i) no connected component of G[Ux ] \ S contains two different sinks of Ux \Vx , and
(ii) every sink in a connected component of G[Ux ] \ S that contains a vertex of Vx can be reached from some vertex
of Vx in G[Ux ] \ S.
Since we assumed to work on a tree decomposition with a root r such that the bag Vr is empty, any partial partitioning
set for G[Ur ] = G will be a partitioning set for the entire graph G. Moreover, Definition 3(i) does not require partial
partitioning sets for G[Ux ] to separate sinks in the bag Vx . This is because vertices in Vx that are sinks in G[Ux ] might be
non-sinks for a supergraph, as illustrated in Figure 11. Thus, it might be unnecessary to separate the vertices in Vx .
However, due to Definition 2(ii and iii) of a tree decomposition, sinks in Ux \Vx are sinks in all supergraphs G[Uq ]
20
Vx
s3
P1
P2
s1
s2
Figure 12: A pattern (R, G, P), where shown are the graph G and a partition P of its vertices into two sets P1 and P2 . If the graph R is the subgraph
of G induced by the vertices Vx , then, in terms of Definition 5, R = Rx (S) and G = Gx (S) for the partial partitioning set S for G[Ux ] shown in Figure 11.
In this case, the partial partitioning set S satisfies the shown pattern. Moreover, a vertex is a sink in this figure if and only if it is a sink in G[Ux ] \ S
shown in Figure 11.
for q being an ancestor node of x. Definition 3(ii), by Observation 2, allows us to ensure that components containing
both a sink in Ux \Vx and a vertex of Vx end up with only one sink in some supergraph G[Uq ]. The precise purpose of
Definition 3(ii) will be explained in more detail after the upcoming Definition 5.
To keep the notation free from clutter, note that Definition 3 implicitly relies on the bag Vx belonging to each set Ux .
Thus, when a tree decomposition has a node x and a child node y such that Vx ( Vy but Ux = Uy , a partial partitioning
set S for G[Uy ] is not necessarily a partial partitioning set for G[Ux ], although G[Uy ] \ S = G[Ux ] \ S.
Now, assume that we want to compute partial partitioning sets for G[Ux ] from partial partitioning sets for child nodes
of x. These partial partitioning sets might, for example, disagree on which arcs to delete in the child bags or which
connected components of the child bags are meant to end up in a common connected component of the entire graph:
for a child node y of x, multiple connected components of G[Uy ] \ S might be one connected component of G[Ux ] \ S.
To prevent such incompatibilities, we only consider those partial partitioning sets for G[Ux ] that agree with partial
partitioning sets for the child nodes of x on certain patterns.
On a high level, our algorithm will store for each node of the tree decomposition a table with one row for each
possible pattern. The value of a row will be the minimum weight of a partial partitioning set satisfying this pattern. To
compute this value, our algorithm will use the rows with corresponding patterns in the tables of the child nodes. In the
following, we first formalize the terms patterns and satisfying partial partitioning sets. Then, we present our algorithm
and we specify the corresponding patterns. We start by formally defining patterns, see Figure 12 for an illustration.
Definition 4 (Pattern). Let x be a node of a tree decomposition T . A pattern for x is a triple (R, G, P) such that
i) R is a directed acyclic graph with the vertices Vx .
ii) G is a directed acyclic graph with the vertices Vx and at most |Vx | additional vertices such that each vertex
in V (G) \Vx is a non-isolated sink, and
iii) P is a partition of the vertices of G such that each connected component of G is within one set Pi ∈ P and such that
each Pi contains at most one vertex of V (G) \Vx .
We will use a pattern (R, G, P) for x to capture important properties of partial partitioning sets for G[Ux ]. Intuitively,
the graph R will describe which arcs between the vertices in the bag Vx a partial partitioning set S for G[Ux ] will not
delete. The graph G will describe which vertices of Vx can reach each other in G[Ux ] \ S and which sinks outside of Vx
they can reach. The partition P describes which vertices are meant to end up in the same connected component of G \ S
for a partitioning set S of the entire graph. For this reason, the sets of the partition P are allowed to contain only one
vertex of V (G) \Vx as these vertices are sinks.
We will now explain precisely what it means for a partial partitioning set to satisfy a pattern. The following
definition is illustrated in Figure 12.
21
Definition 5 (Pattern satisfaction). Let S be a partial partitioning set for G[Ux ]. A sink s in Ux \Vx is bag-reachable
in G[Ux ] \ S if some vertex in Vx can reach s in G[Ux ] \ S. We define a canonical pattern (Rx (S), Gx (S), Px (S)) at x for S,
where
• Rx (S) is G[Vx ] \ S,
• Gx (S) is the directed acyclic graph on the vertices Vx ∪V 0 , where V 0 is the set of bag-reachable sinks in G[Ux ] \ S,
and there is an arc (u, v) in Gx (S) if and only if the vertex u can reach the vertex v in G[Ux ] \ S, and
• Px (S) is the the partition of the vertices of Gx (S) such that the vertices u and v are in the same set of Px (S) if and
only if they are in the same connected component of G[Ux ] \ S.
Let (R, G, P) be a pattern for x. We say that S satisfies the pattern (R, G, P) at x if
i) R = Rx (S),
ii) G = Gx (S), and
iii) for each set P ∈ Px (S) there exists a set P0 ∈ P such that P ⊆ P0 , that is, P is a coarsening of Px (S).
It is easy to verify that a partial partitioning set S for G[Ux ] \ S satisfies its canonical pattern (Rx (S), Gx (S), Px (S)) at
node x: to this end, observe that (Rx (S), Gx (S), Px (S)) is indeed a pattern for x: for the vertex set Vx ∪V 0 of Gx (S), we
have |V 0 | ≤ |Vx | since each vertex in Vx can reach at most one distinct sink in V 0 ⊆ Ux \Vx in G[Ux ] \ S.
Note that, since Gx (S) contains an arc (u, v) if and only if u can reach v instead of requiring them to be merely
connected, a vertex is a sink in Gx (S) if and only if it is a sink in G[Ux ] \ S. Herein, Definition 3(ii) ensures that any
sink s connected to a vertex in Vx is a vertex in Gx (S).
While it might seem more natural to replace the condition (iii) in Definition 5 by simply P = Px (S), we prefer the
current definition, because it allows for several connected components of G[Ux ] \ S becoming a part of one connected
component of the entire graph. This greatly simplifies some parts of the algorithm.
The Algorithm. We now describe a dynamic programming algorithm. Starting from the leaves of the tree decomposition T and working our way to its root, with each node x of T , we associate a table Tabx that is indexed by all possible
patterns for x. Semantically, we want that
Tabx (R, G, P) = minimum weight of a partial partitioning set for G[Ux ] that satisfies the pattern (R, G, P) at x.
Since we have Vr = 0,
/ there is exactly one pattern (R, G, P) for the root r: R = G is the empty graph and P = 0.
/ Thus,
Tabr has exactly one entry and it contains the minimum weight of a partial partitioning set S for G[Ur ], which is
equivalent to S being a partitioning set for G. It follows that once the tables are correctly filled, to decide the DAG
PARTITIONING instance (G, ω, k), it is enough to test whether the only entry of Tabr is at most k.
We now present an algorithm to fill the tables and prove its correctness. First, we initialize all table entries of all
tables by ∞. By updating the entry Tabx (R, G, P) with m we mean setting Tabx (R, G, P) := m if m < Tabx (R, G, P). For
each leaf node x, it is obviously correct to set Tabx (R, G, P) = 0 for the only pattern (R, G, P) at x, which has the empty
graph as R and G and the empty set as P. In the following, for each type of a node x of a tree decomposition, that is,
for forget nodes, introduce nodes, and join nodes, we independently show how to compute the table Tabx given that
we correctly computed the tables for all children of x. To show that the table Tabx is filled correctly, we prove the
following lemma for each node type.
Lemma 6.
(i) There is a partial partitioning set for G[Ux ] satisfying a pattern (R, G, P) at x with weight at most Tabx (R, G, P).
(ii) The minimum weight of a partial partitioning set for G[Ux ] satisfying a pattern (R, G, P) at x is at least Tabx (R, G, P).
We present the algorithm and the proof for Lemma 6 independently for each node type in Sections 4.2.1, 4.2.2, and
4.2.3, respectively, where we assume that all tables Taby for child nodes y of x have been computed correctly.
22
4.2.1. Forget nodes
We use the following procedure to compute the table Tabx of a forget node x under the assumption that the table Taby
for the child node y of x has been computed correctly.
Procedure 1 (Forget node). Let x be a forget node with a single child y. Assume that v is the vertex being “forgotten”,
that is, v is in the child bag Vy but not in the current bag Vx . From the weights of optimal partial partitioning sets
for G[Uy ], we want to compute the weight of optimal partial partitioning sets for G[Ux ].
To this end, for each pattern (R, G, P) for y, we distinguish four cases. In each case, we will construct a pattern (R0 , G0 , P0 ) for x such that a partial partitioning set for G[Uy ] that satisfies (R, G, P) is a partial partitioning set
for G[Ux ] and satisfies (R0 , G0 , P0 ). Then, we update Tabx (R0 , G0 , P0 ) with the value of Taby (R, G, P). Herein, the
following case distinction is not exhaustive. We do not take action for patterns (R, G, P) that do not satisfy any of the
following conditions (for the reasons informally explained in the cases). In all cases, we set R0 := R − {v}.
Case 1) If v is isolated in G and there is a set {v} in P, then we let G0 := G − {v} and P0 := P \ {{v}} and update
Tabx (R0 , G0 , P0 ) with the value of Taby (R, G, P): an isolated vertex that is alone in its part of P can simply be forgotten.
Case 2) If v is a non-isolated sink in G and v ∈ Pi ∈ P such that Pi ⊆ Vy , then we let G0 := G and P0 := P. We update
Tabx (R0 , G0 , P0 ) with the value of Taby (R, G, P): in this case, the sink v “moves” from Vy to V (G0 ) \Vx . To ensure that
(R0 , G0 , P0 ) is a pattern, the part Pi containing v cannot contain any additional sink in V (G) \Vy , thus we require Pi ⊆ Vy .
Case 3) If v is not a sink in G and there is no sink in V (G) \ Vy such that v is its only in-neighbor, then let
G0 := G − {v} and P0 be the partition of the vertices of G0 obtained from P by removing v from the set it is in. Update
Tabx (R0 , G0 , P0 ) with the value of Taby (R, G, P). This the simplest case, where the vertex is somewhat unimportant to
partial partitioning sets satisfying the pattern (R, G, P) at y, so we simply forget it.
Case 4) If there is a sink u ∈ V (G) \ Vy such that v is its only in-neighbor and {u, v} is a set of P, then let
G0 := G − {u, v} and P0 be the partition of the vertices of G0 obtained from P by removing the set {u, v}. Update
Tabx (R0 , G0 , P0 ) with the value of Taby (R, G, P): If there was a sink u in V (G) \Vy only reachable from v, then it would
be unreachable from Vx since v is forgotten. Therefore, if the part Pi of P containing u and v contained more vertices,
then we could not be sure that a partial partitioning set satisfying the pattern (R, G, P) at y is a partial partitioning set
for G[Ux ] at all. Namely, it may break Definition 3(ii).
We show that Procedure 1 fills the table Tabx associated with a forget node x correctly. First, we show that there is a
partial partitioning set for G[Ux ] satisfying a pattern (R, G, P) at x and having weight at most Tabx (R, G, P) as computed
by Procedure 1.
P ROOF ( OF L EMMA 6(i) FOR FORGET NODES ). Let x be a forget node with child node y and let v be the vertex
“forgotten”, that is, v is in the child bag Vy but not in the current bag Vx . For any table entry Tabx (R0 , G0 , P0 ) < ∞, we
show that there is a partial partitioning set S for G[Ux ] satisfying (R0 , G0 , P0 ) and having weight at most Tabx (R0 , G0 , P0 ).
To this end, observe that, since Tabx (R0 , G0 , P0 ) < ∞ there is a pattern (R, G, P) for y from which Procedure 1 generates (R0 , G0 , P0 ) and such that Tabx (R0 , G0 , P0 ) = Taby (R, G, P). Since there is a partitioning set for G[Uy ] that
satisfies (R, G, P) and has weight at most Taby (R, G, P), it is sufficient to show that any partial partitioning set S
for G[Uy ] that satisfies the pattern (R, G, P) at y is also a partial partitioning set for G[Ux ] that satisfies at x the
pattern (R0 , G0 , P0 ) generated in each of the cases (1)–(4) of Procedure 1.
We first argue that S is a partial partitioning set for G[Ux ] if any of the cases (1)–(4) of Procedure 1 applies. We first
verify Definition 3(i). To this end, observe that by Definition 5, a vertex u ∈ Vy is a sink in Gy (S) if and only if it is a
sink in G[Uy ] \ S. Now, assume that there is a connected component of G[Ux ] \ S = G[Uy ] \ S that contains two different
sinks s1 , s2 in Ux \Vx . Then, one of these sinks, say s1 , must be v. Since, then, v ∈ Vy is a sink in G[Ux ] \ S = G[Uy ] \ S,
it is a sink in G = Gy (S) and none of the cases (3) and (4) apply. Moreover, since s2 is connected to v ∈ Vy in G[Uy ] \ S,
by Definition 3(ii), some vertex in Vy can reach s2 , implying that s2 is a vertex of G = Gy (S). Thus, by Definition 5(iii),
s2 is in the same set Pi ∈ P as s1 = v and, hence, (1) does not apply. Since s2 < Vy , also (2) does not apply.
We now verify Definition 3(ii). It can only be violated if v is the only vertex of Vy that can reach some sink u in
the connected component of v in G[Uy ] \ S. However, then, v is the only in-neighbor of u in Gy (S) = G. Hence, only
case (4) might become applicable. When this case applies, however, {u, v} ∈ P implies that no vertex in Vy ⊇ Vx is
connected to v or u in G[Uy ] \ S = G[Ux ] \ S. Thus, Definition 3(ii) is satisfied.
23
It remains to show that S satisfies the generated pattern (R0 , G0 , P0 ), that is, to verify R0 = Rx (S) (Definition 5(i)),
= Gx (S) (Definition 5(ii)) and that P0 is a coarsening of Px (S) (Definition 5(iii)). Herein, R0 = R − {v} = Ry (S) −
{v} = Rx (S) is trivial. To show G0 = Gx (S), we distinguish between the case of Procedure 1 applied.
Case 1) In this case, v is not in Vx and, obviously, not a bag-reachable sink in G[Ux ] \ S. Hence, v is not in Gx (S).
Moreover, v is isolated in G = Gy (S). Therefore, Procedure 1 sets G0 := G − {v} = Gy (S) − {v} = Gx (S).
G0
Case 2) In this case, v is not in Vx but it is a bag-reachable sink in G[Ux ] \ S, since it is not isolated in G. Therefore,
Procedure 1 sets G0 := G = Gy (S) = Gx (S).
Case 3) In this case, v is not a sink in G (and thus also not in G[Ux ] \ S) and, therefore, clearly does not appear
in Gx (S). Moreover, any sink not in Vy is reachable from a vertex in Vy \ {v} = Vx in Gy (S) if and only if it is reachable
in Gy (S) − {v}. Hence, Procedure 1 sets G0 := G − {v} = Gy (S) − {v} = Gx (S).
Case 4) In this case, u was a bag-reachable sink in G[Uy ] \ S but is not bag-reachable in G[Ux ] \ S. Moreover, since
{v, u} ∈ P, no vertex of Vx is connected to v in G[Uy ] \ S = G[Ux ] \ S. Hence, neither v nor u are vertices of Gx (S).
Hence, Procedure 1 sets G0 := G − {u, v} = Gy (S) − {u, v} = Gx (S).
Finally, we verify Definition 5(iii) by showing that P0 is a coarsening of Px (S). Assume the contrary. Then, there
are two vertices u, w of Gx (S) in the same set of Px (S) but in different sets of P0 . By construction of P0 from P, they are
also in different sets of P = Py (S). It follows that u and w lie in the same connected component of G[Ux ] \ S but in
different connected components of G[Uy ] \ S. Since these two graphs are the same, we have a contradiction.
We now show that the minimum weight of a partial partitioning set for G[Ux ] satisfying a pattern (R, G, P) at x is at
least Tabx (R, G, P) as computed by Procedure 1.
P ROOF ( OF L EMMA 6(ii) FOR FORGET NODES ). Let x be a forget node with child node y. Let v be the vertex
“forgotten” that is, v ∈ Vy but v < Vx . Assume that S is a partial partitioning set for G[Ux ] satisfying the pattern (Rx , Gx , Px )
at x. It is sufficient to construct a pattern (R, G, P) that S satisfies at y and from which Procedure 1 generates exactly the
pattern (Rx , Gx , Px ) to update the table Tabx (Rx , Gx , Px ) with Taby (R, G, P). Then, Lemma 6(ii) follows for forget nodes,
because we have Tabx (Rx , Gx , Px ) ≤ Taby (R, G, P) ≤ ω(S). Herein, the last inequality follows from the induction
hypothesis.
We first show that S is a partial partitioning set for G[Uy ], that is, we verify Definition 3. Definition 3(i) is easy
to verify: since S is a partial partitioning set for G[Ux ], each connected component of G[Ux ] \ S = G[Uy ] \ S contains
at most one sink in Ux \ Vx ⊇ Uy \ Vy . It remains to verify Definition 3(ii). Assume, for a contradiction, that there
is a connected component C in G[Uy ] \ S that contains a vertex of Vy such that no vertex of Vy ⊇ Vx can reach some
sink s ∈ C \ Vy . Then, since S is a partial partitioning set for G[Ux ] \ S, the connected component C cannot contain
vertices of Vx and, hence, C ∩Vy = {v}. However, since G[Uy ] \ S is a directed acyclic graph, v reaches some sink in C.
Since v cannot reach s ∈ C, it follows that C contains two sinks. Since C ∩Vx = 0,
/ this contradicts S being a partial
partitioning set for G[Ux ] \ S. It follows that S is a partial partitioning set for G[Uy ].
We now construct a pattern (R, G, P) that S satisfies at y. Consider R := Ry (S) and G := Gy (S). Note that there are
at most two vertices in G that are not in Gx : one of them is v, the possibly other vertex is a sink u in V (G) \Vy only
reachable from v. We define P as a partition of the vertices of G that partitions the set V (Gx ) in the same way as Px . We
add the possibly missing vertices v and u to that partition as follows: if there is a vertex w ∈ Vx in the same connected
component of G[Uy ] \ S = G[Ux ] \ S as v, then we put v and u into same set as w. Otherwise, we add the set {v} or {u, v},
respectively, to P. By choice of R, G, and P, the partial partitioning set S clearly satisfies the pattern (R, G, P) at y.
We have shown that S satisfies the pattern (R, G, P) at y. Moreover, if any of the cases (1)–(4) of Procedure 1
applies to (R, G, P), then it generates a pattern (R0 , G0 , P0 ) with R0 = Rx and G0 = Gx , since we showed in the proof of
Lemma 6(i) for forget nodes that S satisfies the pattern generated by Procedure 1 at x. Hence, it remains to show that
indeed at least one of the cases (1)–(4) of Procedure 1 applies and that in all cases P0 = Px .
Case 1) If v is an isolated sink in G, then no vertex in Vy ⊇ Vx can reach v in G[Uy ] \ S = G[Ux ] \ S. Hence, there
is no vertex of Vx in the same connected component of G[Ux ] \ S as v, as otherwise v would be a sink in Ux \Vx not
reachable from the vertices of Vx . Hence, by construction of P, we have {v} ∈ P and case (1) of Procedure 1 applies. It
sets P0 := P \ {{v}} = Px .
Case 2) If v is a non-isolated sink in G, then v is a bag-reachable sink in G[Ux ] \ S. Hence, it is contained in Gx
and we have V (Gx ) = V (G). By construction of P, we also have Px = P. The set Pi ∈ Px containing v < Vx cannot
24
contain any other vertex in V (Gx ) \ Vx by Definition 4(iii). Thus, Pi ⊆ Vy and case (2) of Procedure 1 applies. It
sets P0 := P = Px .
Case 3) If v is not a sink in G and there is no sink in V (G) \ Vy only reachable from v in G, then case (3) of
Procedure 1 applies. Since the sink s reachable from v is also reachable from some vertex u < {v, s}, and thus, connected
to u in G[Uy ] \ S, the set in P containing v also contains u. Procedure 1 sets P0 to be P with v removed from the set it is
in. This, by construction of P, is exactly Px .
Case 4) Finally, if there is a sink u in V (G) \Vy only reachable from v, then the connected component of G[Uy ] \ S =
G[Ux ] \ S containing the vertex v does not contain any vertex of Vx , since u is not reachable from any vertex of Vx . It
follows that {u, v} is a set of P. Case (4) of Procedure 1 applies. It sets P0 := P \ {{v, u}} = Px .
4.2.2. Introduce nodes
We use the following procedure to compute the table Tabx of an introduce node x under the assumption that the
table Taby for the child node y of x has been computed correctly.
Procedure 2 (Introduce node). Let x be an introduce node with a single child y. Assume that v is the node being
“introduced”, that is, v is not in the child bag Vy but in the current bag Vx . Moreover, let B ⊆ A(G[Ux ]) be the set of arcs
incident to v. By Definition 2(ii and iii) of a tree decomposition, one actually has B ⊆ A(G[Vx ]).
We now try each possible subset B0 ⊆ B and consider it not deleted by a partial partitioning set for the graph G[Ux ].
Similarly as in the case for forget nodes, we will transform each pattern (R, G, P) for y into a pattern (R0 , G0 , P0 ) for x
such that if a partial partitioning set S for G[Uy ] satisfies (R, G, P), then S ∪ (B \ B0 ) is a partial partitioning set for G[Ux ]
and satisfies (R0 , G0 , P0 ). Then, we update Tabx (R0 , G0 , P0 ) with the value of Taby (R, G, P) + ω(B) − ω(B0 ).
For each pattern (R, G, P) for y such that all vertices incident to the arcs in B0 (if any) except for v are contained
in the same set Pi ∈ P, we obtain R0 from R by adding v and the arcs in B0 to R. Similarly, we obtain G0 from G by
adding v and the arcs in B0 to G. Moreover, for each u, w ∈ V (G0 ) such that u can reach w in G0 we add the arc (u, w)
to G0 . For obtaining P0 , we distinguish two cases.
Case 1) If B0 = 0,
/ then we try all possibilities of adding v to a set in P. That is, for every Pi ∈ P, we get a set P0
from P by adding v to Pi and update Tabx (R0 , G0 , P0 ) with Taby (R, G, P) + ω(B). Additionally, for P0 := P ∪ {{v}}, we
update the entry Tabx (R0 , G0 , P0 ) with Taby (R, G, P) + ω(B).
Case 2) If B0 , 0,
/ then let Pi be the set of P that contains all vertices incident to arcs in B0 except v and let P0 be
obtained from P by adding v to the set Pi . We update Tabx (R0 , G0 , P0 ) with Taby (R, G, P) + ω(B) − ω(B0 ).
Note that, since P simulates the connected components of the resulting graph, all arcs incident on v remaining in
the graph must be within one set of P, i.e., their endpoints different from v must be in one set of P.
We show that Procedure 2 fills the table associated with an introduce node x correctly. First, we show that there is a
partial partitioning set for G[Ux ] satisfying a pattern (R, G, P) at x and having weight at most Tabx (R, G, P) as computed
by Procedure 2.
P ROOF ( OF L EMMA 6(i) FOR INTRODUCE NODES ). Let x be an introduce node with child node y and let v be the
vertex “introduced” that is, v is not in the child bag Vy but in the current bag Vx . Let B ⊆ A(G[Vx ]) be the arcs incident
to v, B0 ⊆ B and, finally, (R, G, P) be some pattern for y such that all vertices incident to the arcs in B0 (if any) except
for v are contained in the same set Pi ∈ P. For any partial partitioning set S for G[Uy ] satisfying the pattern (R, G, P)
at y, we show that S0 = S ∪ (B \ B0 ) is a partial partitioning set for G[Ux ] that satisfies the pattern (R0 , G0 , P0 ) constructed
by Procedure 2. From this, since ω(S0 ) = ω(S) + ω(B) − ω(B0 ), Lemma 6(i) follows (as already discussed in the
beginning of the proof of Lemma 6(i) for forget nodes).
We start by showing that S0 is a partial partitioning set for G[Ux ]. First, we verify Definition 3(i). For the sake of a
contradiction, assume that there is a connected component of G[Ux ] \ S0 that contains two distinct sinks s1 , s2 in Ux \Vx .
Since there is no such connected component in G[Uy ] \ S = (G[Ux ] \ S0 ) − {v}, there are vertices s01 , s02 ∈ Vy in the same
connected components of G[Uy ] \ S as s1 and s2 , respectively, that are incident to some arcs in B0 . By Definition 3(ii),
there are vertices in Vy that can reach s1 and s2 in G[Uy ] \ S. Hence, s1 and s2 are bag-reachable and, therefore, in
G = Gy (S). Since s01 , s02 ∈ Pi and S satisfies (R, G, P), by Definition 4(iii), we also have s1 , s2 ∈ Pi . Then, however,
Pi contains the two different vertices s1 and s2 of V (G) \Vy , which contradictions Definition 4(iii).
25
To show that S0 is a partial partitioning set, it remains to verify Definition 3(ii). For the sake of contradiction,
assume that some connected component of G[Ux ] \ S0 contains some sink s ∈ Ux \Vx , some vertex in Vx , but s is not
reachable from any vertex in Vx . Then, s is not reachable from any vertex in Vy ⊆ Vx in the subgraph G[Uy ] \ S either.
Thus, the connected component does not contain any vertex of Vy and, therefore, not of Vx , since the only vertex
in Vx \Vy is v and the added arcs B0 connect only vertices in Vy .
We have shown that S0 is a partial partitioning set for G[Ux ]. We now show that it satisfies the pattern (R0 , G0 , P0 )
generated by Procedure 2; we verify Definition 5. Definition 5(i), that is, R0 = Rx (S0 ) is trivial by the construction
of R0 .
We verify Definition 5(ii), that is, Gx (S0 ) = G0 . First, observe that V (Gx (S0 )) = V (Gy (S)) ∪ {v} = V (G0 ). We have
to show that there is an arc (u, w) in G0 if and only if u can reach w in G[Ux ] \ S0 . Let (u, w) be such an arc in G0 . If
(u, w) is already in G, then u can reach w in G[Uy ] \ S = (G[Ux ] \ S0 ) − {v}. Otherwise, u reaches w in G0 via some
arcs (u0 , v), (v, w0 ) ∈ B0 , that is, u can reach u0 and w0 can reach w in G[Ux ] \ S0 . It follows that u can reach w in
G[Ux ] \ S0 . Now, for the opposite direction, let u, w be vertices of G0 such that u can reach w in G[Ux ] \ S0 . If u can
reach w in (G[Ux ] \ S0 ) − {v} = G[Uy ] \ S, then the arc (u, w) is already present in G. Otherwise, u reaches w via some
arcs (u0 , v), (v, w0 ) ∈ B0 . The arcs (u0 , v) and (v, w0 ) are in G0 , since u0 , w0 ∈ Vx . Moreover, u reaches u0 and w0 reaches w
in G[Ux ] \ S0 . Hence, there are arcs (u, u0 ) and arc (w0 , w) in G0 and u reaches w in G0 via u0 and w0 . By construction
of G0 , it follows that G0 contains the arc (u, w).
Finally, we verify Definition 5(iii); we show that P0 is a coarsening of Px (S0 ). For the sake of a contradiction,
assume that there are two vertices u, w that are in the same set of Px (S0 ) but in different sets of P0 . By construction
of P0 from P, this implies that u and w are in different sets of P and, therefore, in different connected components
of G[Uy ] \ S. Thus, in order for u and w to be connected in G[Ux ] \ S0 , there are vertices u0 , w0 in the same connected
components of G[Uy ] \ S as u and w, respectively, that are incident to arcs in B0 and, hence, u0 , w0 ∈ Pi ∈ P. But then,
also u, w ∈ Pi ∈ P — a contradiction.
We now show that the minimum weight of a partial partitioning set for G[Ux ] satisfying a pattern (R, G, P) at x is at
least Tabx (R, G, P) as computed by by Procedure 2.
P ROOF ( OF L EMMA 6(ii) FOR INTRODUCE NODES ). Let x be an introduce node with child node y. Let v be the vertex
“introduced”, that is, v < Vy but v ∈ Vx . Assume that S is a minimum-weight partial partitioning set for G[Ux ] satisfying
the pattern (Rx , Gx , Px ) at x. Let B be the set of arcs incident to v in G[Ux ] and B00 := B ∩ S. It is sufficient to construct
a pattern (R, G, P) that S \ B00 satisfies at y and from which Procedure 2 generates exactly the pattern (Rx , Gx , Px ) to
update the table Tabx (Rx , Gx , Px ) with Taby (R, G, P) + ω(B) − ω(B0 ), where B0 = B \ B00 . Then, Lemma 6(ii) follows
for introduce nodes, since Tabx (Rx , Gx , Px ) ≤ Taby (R, G, P) + ω(B) − ω(B0 ) ≤ ω(S \ B00 ) + ω(B) − ω(B0 ) = ω(S).
It is easy to verify that S \ B00 is a partial partitioning set for G[Uy ] (Definition 3), since G[Uy ] \ (S \ B00 ) =
(G[Ux ] \ S) − {v} and S is a partial partitioning set for G[Ux ]; to this end, observe that, by Definition 2(ii and iii) of a
tree decomposition, v only has arcs B ⊆ A(G[Vx ]) incident to vertices in Vy .
We now construct a pattern. Let R = Ry (S \ B00 ) and G = Gy (S \ B00 ). Let P be the partition obtained from Px by
removing the vertex v from the set it is in or by removing the set {v} if it exists in Px . It is easy to verify that S \ B00
satisfies (R, G, P) at y: Definition 5(i) and (ii) are trivially satisfied by choice of R and G; Definition 5(iii) holds by
construction of P from Px , since G[Uy ] \ (S \ B00 ) is a subgraph of G[Ux ] \ S.
It remains to show that Procedure 2 applies to the pattern (R, P, G) and the set B0 in order to generate the
pattern (Rx , Px , Px ). Since S satisfies (Rx , Gx , Px ) at x, all vertices incident to arcs in B0 (if any) are contained in
the same set Pi ∈ Px and, hence, all of them except v are contained in the set Pi \ {v} ∈ P. Therefore, Procedure 2
applies to B0 and the pattern (R, G, P), produces some new pattern (R0 , G0 , P0 ), and updates Tabx (R0 , G0 , P0 ) with
Taby (R, G, P) + ω(B) − ω(B0 ).
It remains to show that, for at least one of the generated patterns, R0 = Rx , G0 = Gx , and P0 = Px . If B0 , 0/ then
0
P = Px by construction of P0 from P in Procedure 2. If B0 = 0,
/ then Px is clearly among the partitions P0 generated
from P by Procedure 2. Moreover, we already proved in the proof of Lemma 6(i) that S = (S \ B00 ) ∪ (B \ B0 ) satisfies
the pattern generated by Procedure 2 at x. Hence, R0 = Rx and G0 = Gx .
4.2.3. Join nodes
We use the following procedure to compute the table Tabx of a join node x under the assumption that the tables Taby
for all child nodes y of x have been computed correctly.
26
Procedure 3 (Join node). Let x be a join node with children y and z, that is, Vx = Vy = Vz . For each pair of patterns (R, Gy , Py ) for y and (R, Gz , Pz ) for z such that Py and Pz partition the vertices of Vy = Vz = Vx in the same way,
we construct a new pattern (R, G, P) as follows.
Let G0 be the graph containing all vertices and arcs of Gy and Gz , and for each u, w ∈ V (G0 ) such that u can reach w
in G0 add the arc (u, w) to G0 . Note that by Definition 2(iii) of a tree decomposition, Gy and Gz have only the vertices
in Vx in common.
Let P0 be the partition of Vx that partitions Vx in the same way as Py and Pz . We extend P0 to a partition for the
vertices of G0 : for each u ∈ V (G0 ) \Vx , add u to a set Pi of P0 that contains a vertex v with (v, u) being an arc of G0 . Since
there are no arcs between different sets of P in Gy or Gz , there is exactly one such set Pi ∈ P0 .
If we created some set P ∈ P0 with more than one vertex of V (G0 ) \Vx , then continue with a different pair of patterns.
Otherwise, we update Tabx (R, G0 , P0 ) with Taby (R, Gy , Py ) + Tabz (R, Gz , Pz ) − ω(A(G[Vx ])) + ω(A(R)).
We show that Procedure 3 fills the table associated with a join node x correctly. First, we show that there is a partial
partitioning set for G[Ux ] satisfying a pattern (R, G, P) at x and having weight at most Tabx (R, G, P) as computed by
Procedure 3.
P ROOF ( OF L EMMA 6(i) FOR JOIN NODES ). Let x be a join node with child nodes y and z, that is, Vx = Vy = Vz . Let
Sy be a partial partitioning set for G[Uy ] satisfying the pattern (R, Gy , Py ) at y and let Sz be a partial partitioning
set for G[Uz ] satisfying the pattern (R, Gz , Pz ) at z. We show that S = Sy ∪ Sz is a partial partitioning set for G[Ux ]
that satisfies the pattern (R, G0 , P0 ) constructed by Procedure 3. Since ω(S) = ω(Sy ) + ω(Sz ) − ω(Sy ∩ Sz ), wherein
Sy ∩ Sz = A(G[Vx ]) \ A(R), Lemma 6(i) follows for join nodes.
We show that S is indeed a partial partitioning set for G[Ux ], that is, we verify Definition 3. We first verify
Definition 3(ii) and then use it to verify Definition 3(i). Let s ∈ Ux \Vx be a sink such that the connected component
containing s in G[Ux ]\S contains a vertex of Vx . Then, s ∈ Uy \Vx or s ∈ Uz \Vx . Without loss of generality, let s ∈ Uy \Vx .
From Definition 2(iii) of a tree decomposition, we see that Uz ∩Uy ⊆ Vx and, hence, G[Uy ] \ Sy = G[Ux ] \ S − (Uz \Vz ).
It follows that there is also a connected component of G[Uy ] \ Sy = G[Ux ] \ S − (Uz \Vz ) that contains s and a vertex
of Vx and, therefore, s is reachable from some vertex v ∈ Vx in G[Ux ] \ S − (Uz \Vz ) and, hence, in G[Ux ] \ S. It also
follows that (v, s) is an arc in Gy and, by construction in Procedure 3, of G0 .
To verify Definition 3(i), for the sake of a contradiction, assume that there is a connected component of G[Ux ] \ S
that contains two sinks s1 , s2 in Ux \ Vx . Note that, by Definition 2(iii) of a tree decomposition, there are no arcs
between Uy \ Vx and Uz \ Vx . Hence, this connected component contains a vertex v of Vx ; otherwise, it would be a
connected component with two sinks outside of Vx already in either G[Uy ] \ Sy or G[Uz ] \ Sz . Thus, as seen in the
previous paragraph, we have arcs (s01 , s1 ) and (s02 , s2 ) with s01 , s02 ∈ Vx . It follows by construction of P0 from G0 in
Procedure 3 that s1 and s01 are in a set Pi ∈ P0 and s2 and s02 are in a set Pj ∈ P0 . We show i = j, which contradicts the
construction of P0 , since then Pi = Pj contains two vertices s1 < Vx and s2 < Vx .
Since s1 and s2 are in the same connected component of G[Ux ] \ S, also s01 and s02 are, since they can reach s1
and s2 , respectively. Hence, there is an undirected path p between s1 and s2 in G[Ux ] \ S. It consists of consecutive path
segments p0 that only have their endpoints u, w in Vx (possibly, such a path segment only consists of one arc). It follows
that such a path segment p0 is entirely contained in G[Uy ] \ Sy or G[Uz ] \ Sy and, hence, its endpoints u and w are in the
same set of Py or Pz . Since u, w ∈ Vx , by construction of P0 in Procedure 3, u and w are in the same set of P0 . It follows
that s01 and s02 are in the same set of P0 , and so are s1 and s2 .
It follows that S is indeed a partial partitioning set for G[Ux ] \ S. It remains to verify that S satisfies the pattern (R, G0 , P0 ) (Definition 5). Herein, Definition 5(i), R = Rx (S), is trivial. We verify (ii), that is, Gx (S) = G0 . Herein,
V (Gx (S)) ⊆ V (G0 ) we already verified when verifying Definition 3(ii). Now, assume that there are two vertices u, w
in G0 such that u can reach w in G[Ux ] \ S. Since, then, u is not a sink, it is in Vx . The directed path from u to w consists
of consecutive subpaths, each being entirely contained in G[Uy ] \ Sy or G[Uz ] \ Sz and thus, causing an arc in Gy or Gz
and, therefore, in G0 . It follows that u can reach w in G0 , which therefore has an arc (u, w). In the opposite direction, for
every arc (u, w) in G0 that is already in Gy or Gz , there is an directed path in either G[Uy ] \ Sy or in G[Uz ] \ Sz from u
to w and, thus, u can reach w in G[Ux ] \ S. For an arc (u, w) in G0 that is neither present in Gy nor Gz , there is an directed
path in G0 from u to w consisting only of arcs that are already present in Gy or Gz . Since we have seen that for each such
arc there is a corresponding directed path in G[Ux ] \ S, we have that u can reach w in G[Ux ] \ S.
For Definition 5(iii), it has been shown above that if two vertices of G0 are in the same connected component
of G[Ux ] \ S, then they are in the same set in P0 .
27
We now show that the minimum weight of a partial partitioning set for G[Ux ] satisfying a pattern (R, G, P) at x is at
least Tabx (R, G, P) as computed by by Procedure 3.
P ROOF ( OF L EMMA 6(ii) FOR JOIN NODES ). Let x be a join node with the child nodes y and z, that is Vx = Vy = Vz .
Assume that S is a minimum-weight partial partitioning set for G[Ux ] satisfying the pattern (R, Gx , Px ) at x. It is sufficient
to construct patterns (R, Gy , Py ) and (R, Gz , Pz ) that are satisfied by Sy := S ∩ A(G[Uy ]) at y and by Sz := S ∩ A(G[Uz ])
at z, respectively, such that from these patterns Procedure 3 generates exactly the pattern (R, Gx , Px ) to update
Tabx (R, Gx , Px ) with
Taby (R, Gy , Py ) + Tabz (R, Gz , Pz ) − ω(A(G[Vx ])) + ω(A(R))
≤ ω(Sy ) + ω(Sz ) − ω(Sy ∩ Sz )
= ω(S).
We first show that Sy is a partial partitioning set for G[Uy ]. Symmetrically, it follows that Sz is a partial partitioning
set for G[Uz ]. We first verify Definition 3(i). Since by Definition 2(iii), there are no arcs between vertices in Uy \Vy
and Uz \ Vz in G[Ux ], it follows from G[Uy ] \ Sy = G[Ux ] \ S − (Uz \ Vz ) that no connected component of G[Uy ] \ Sy
contains two sinks not in Vy = Vx . It remains to verify Definition 3(ii). To this end, let u be a sink in Uy \ Vy in a
connected component of G[Uy ] \ Sy containing a vertex of Vy . Then, by Definition 3(ii), the connected component
of G[Ux ] \ S containing u contains a directed path from some vertex in Vx to u. The subpath of this directed path that
contains only one vertex of Vx is preserved in G[Uy ] \ Sy . Hence, u is reachable from some vertex of Vx = Vy .
We now construct the patterns. To this end, let Gy := Gy (Sy ) and Gz := Gz (Sz ). Moreover, we choose Py and Pz such
that they partition the set Vx = Vy = Vz in the same way as Px and such that the vertices of V (Gy ) \Vy (or V (Gz ) \Vz ) are
in the same set as the other vertices of their connected components in G[Uy ] \ Sy (or G[Uz ] \ Sz ).
We show that Sy satisfies (R, Gy , Py ) at y. Analogously, it then follows that Sz satisfies (R, Gz , Pz ) at z. We verify
Definition 5. Since R = Rx (S) = Ry (Sy ) = Rz (Sz ) and Gy = Gy (Sy ) hold by definition, it remains to verify Definition 5(iii). To this end, observe that G[Uy ] \ Sy = G[Ux ] \ S − (Uy \Vy ). Now, assume, for the sake of a contradiction,
that there are two vertices v, w of Gy in different sets of Py but in the same connected component of G[Uy ] \ Sy . It
follows that v and w are in the same connected component of G[Ux ] \ S. If v, w ∈ Vy , then, by construction of Py from Px ,
the vertices v and w are in different sets of Px , contradicting Sx satisfying (R, Gx , Px ). If exactly one of v, w is in Vy ,
then v and w being in different sets of Py contradicts the construction of Py . If both v, w < Vy , then v and w are two
bag-reachable sinks in G[Uy ], which contradicts v and w being in the same connected component of G[Uy ] \ Sy .
Hence, indeed Sy satisfies (R, Gy , Py ) at y and Sz satisfies (R, Gz , Pz ) at z. Moreover, since Py and Pz partition Vx in
the same way, Procedure 3 applies to the patterns (R, Gy , Py ) and (R, Gz , Pz ) and produces a pattern (R, G0 , P0 ). If no
set of P0 contains more than one vertex of V (G0 ) \Vx , it indeed updates Tabx (R, G0 , P0 ).
Hence it remains to show that G0 = Gx and P0 = Px , as no set of Px contains two vertices of V (Gx ) \Vx by Definition 4(iii). We already showed in the proof of Lemma 6(i) for join nodes that S satisfies the pattern (R, G0 , P0 ) generated
by Procedure 3. Hence, G0 = Gx . Finally, by construction of P0 in Procedure 3, the vertices of Vx are partitioned the
same way by P0 and Px . For a vertex v ∈ V (G0 ) \ Vx , there is a vertex u in Vx that can reach v in G0 and, therefore,
in G[Ux ] \ S. Hence, u and v must be in the same set of both Px and P0 by construction of P0 in Procedure 3.
4.2.4. Running time
Having shown the correctness of the Procedures 1–3, we can finally complete the proof of Theorem 6 by analyzing the
running time of the procedures.
P ROOF ( OF T HEOREM 6). Lemma 6 showed that the presented dynamic programming algorithm is correct, that is, it
solves DAG PARTITIONING given a tree decomposition of the input graph. It remains to analyze the running time.
To this end, recall that each bag in a tree decomposition of width t contains at most t + 1 vertices. This allows
t+1
us to give an upper bound on the number of possible patterns (R, G, P). There are at most 3( 2 ) directed acyclic
graphs R on at most t + 1 vertices: for each pair (v, w) of vertices: there is either no arc, or an arc from v to w, or
2t+2
an arc from w to v. Similarly, there are at most 3( 2 ) directed graphs G on at most 2t + 2 vertices. Moreover, there
are at most (2t + 2)2t+2 partitions P of at most 2t + 2 vertices into at most 2t + 2 sets. Hence, each table has at most
28
2t+2
t+1
2
2
3( 2 ) · 3( 2 ) · (2t + 2)2t+2 = 3O(t +t logt) = 2O(t ) entries and looking up entries in the tables can be implemented to
2
run in O(log 2t ) time, which is polynomial in t.
In each leaf node, we set the single table entry to 0 in constant time.
In each forget node, Procedure 1 iterates over the entries of the table of the child node and for each entry spends
2
time polynomial in t. Thus, it spends a total of 2O(t ) time in each forget node.
To analyze the running time Procedure 2 spends in an introduce node, observe that there are at most t arcs
in A(G[Vx ]) incident to the introduced vertex v. Hence, there are at most 2t subsets of them. For each of these subsets
2
and for each entry of the child node, Procedure 2 spends time polynomial in t. This makes a total of 2O(t ) time spent in
each introduce node.
Finally, in a join node, Procedure 3 considers every pair of patterns of its two child nodes and for each combination
2
2
spends time polynomial in t. Hence, the total time spent in a join node is (2O(t ) )2 = 2O(t ) .
2
Since the nice tree decomposition has O(tn) nodes, the algorithm runs in 2O(t ) · n time.
5. Other parameters yield stronger NP-hardness results
In Sections 3 and 4, we have seen that DAG PARTITIONING is solvable in linear time when fixing the weight of the
partitioning set sought or the treewidth of the input graph. The question whether fixed-parameter algorithms can be
obtained for parameters that are smaller than the solution weight or the treewidth naturally arises [29, 25, 14]. One
parameter of interest is the maximum vertex outdegree in the graph: a citation network of journal articles will, for
example, have a small outdegree since journal articles seldom contain more than 50 references. In this section, however,
we will show that, among others, this parameter being small will not help solving DAG PARTITIONING efficiently.
Alamdari and Mehrabian [1] already showed that DAG PARTITIONING remains NP-hard even if the input graph
has only two sinks. We complement this negative result by showing that the problem remains NP-hard even if the
diameter or the maximum vertex degree of the input graph are constant. In conclusion, parameters like the number of
sinks, the graph diameter or maximum degree cannot lead to fixed-parameter algorithms unless P = NP.
Theorem 7. DAG PARTITIONING is solvable in linear time on graphs of diameter one, but NP-complete on graphs of
diameter two even if all arcs have unit weight.
P ROOF. On graphs of diameter one, the problem is trivial: a directed acyclic graph with diameter one is an acyclic
tournament, that is, there is no pair of vertices not joined by an arc. As such, it already contains exactly one source and
one sink. Hence, we just verify in linear time whether the input graph is an acyclic tournament and answer “yes” or
“no” accordingly.
For graphs of diameter two, we show NP-hardness by means of a polynomial-time many-one reduction from DAG
PARTITIONING, which is NP-hard even when all arcs have weight one. Therefore, we agree on all arcs in this proof
having weight one.
Given an arbitrary instance (G, ω, k) of DAG PARTITIONING, we add a gadget to G to obtain in polynomial
time an instance (G0 , ω 0 , k0 ) such that G0 is a graph of diameter two and such that (G, ω, k) is a yes-instance if and
only if (G0 , ω 0 , k0 ) is a yes-instance. We obtain a graph G0 from G by adding an acyclic tournament T consisting of
k + n + 2 vertices and outgoing arcs from the source s of T to all vertices of V (G) in G0 . We set k0 := k + n. Since every
vertex in G0 is in distance one from s, the constructed graph G0 has diameter two.
If (G, ω, k) is a yes-instance, then let S be a partitioning set with k arcs for G. A partitioning set S0 with k0 arcs
for G0 is obtained by adding to S the n arcs from the source s of T to all vertices of V (G) in G0 . Thus, (G0 , ω 0 , k) is a
yes-instance.
If (G0 , ω 0 , k0 ) is a yes-instance, then let S0 be a partitioning set with k0 arcs for G0 . By Observation 2, every vertex
in V (G) reaches at least one sink in G0 \ S0 . This sink cannot be the sink of T , since no vertex in T is reachable
from V (G). Thus, S0 has to disconnect the sink of T from all vertices of V (G) in G0 , where all paths between the sink
of T and V (G) are via the source s of T . Since T has k + n + 2 vertices, S0 cannot disconnect s from the sink of T and
thus, has to remove from G0 the n arcs connecting s to the vertices of V (G). Then, the remaining k arcs in S0 have to
be a partitioning set for G0 without the tournament T , which is precisely the original graph G. Thus, (G, ω, k) is a
yes-instance.
29
t1
t2
v2
t3
e{t1 ,t2 }
Z
e{t1 ,v1 }
Y
v1
t1
e{v1 ,t2 }
e{t2 ,v2 }
v1
t2
e{v2 ,t3 }
v2
t3
X
s1
s2
s3
Figure 13: Reduction from a M ULTIWAY C UT instance with the terminals t1 ,t2 , and t3 to DAG PARTITIONING. The top shows an instance I1 of
M ULTIWAY C UT, where the dotted edges are a multiway cut of size k1 = 3. The bottom shows the corresponding instance I2 of DAG PARTITIONING,
where the dotted arcs are a corresponding partitioning set of size k2 = k1 + 2(n − 3) = 7 (n is the number of vertices in the graph of the M ULTIWAY
C UT instance). The constructed vertex sets X,Y , and Z are highlighted using a gray background.
Theorem 8. DAG PARTITIONING is solvable in linear time on graphs of maximum degree two, but NP-complete on
graphs of maximum degree three even if all arcs have unit weight.
P ROOF. Any graph of maximum degree two consists of undirected cycles or undirected paths. Thus, the underlying
graph has treewidth at most two. We have seen in Theorem 6 that DAG PARTITIONING is linear-time solvable when
the treewidth of the input graph is bounded by a constant.
We prove the NP-hardness on graphs of maximum degree three. To this end, we adapt the polynomial-time manyone reduction from M ULTIWAY C UT to DAG PARTITIONING presented by Leskovec et al. [26]. In their reduction, we
replace vertices of degree greater than three by equivalent structures of lower degree.
M ULTIWAY C UT
Input: An undirected graph G = (V, E), a weight function ω : E → N, a set T ⊆ V of terminals, and an integer k.
Question: Is there a subset S ⊆ E with ∑e∈S ω(e) ≤ k such that the removal of S from G disconnects each terminal
from all the others?
We first recall the reduction from M ULTIWAY C UT to DAG PARTITIONING. From a M ULTIWAY C UT instance I1 :=
(G1 , ω1 , T, k1 ), we construct in polynomial time a DAG PARTITIONING instance I2 := (G2 , ω2 , k2 ) such that I1 is a
yes-instance if and only if I2 is. From I2 , we then obtain an instance I3 with maximum degree three. Since M ULTIWAY
C UT remains NP-hard even for three terminals and unit weights [11], we may assume |T | = 3 and, similarly as in the
proof of Theorem 7, we agree on all arcs in this proof having weight one. We now construct the DAG PARTITIONING
instance I2 = (G2 , ω2 , k2 ) from I1 = (G1 , ω1 , T, k1 ) as follows. The construction is illustrated in Figure 13.
1. Add three vertices s1 , s2 , s3 to G2 , forming the vertex set X,
2. add each vertex of G1 to G2 , forming the vertex set Y ,
3. for each edge {u, v} of G1 , add a vertex e{u,v} to G2 , forming the vertex set Z,
4. for each terminal ti ∈ T , add the arc (ti , si ) to G2 ,
5. for each vertex v ∈ Y \ T , add the arcs (v, si ) for i = 1, 2, 3 to G2 , and
6. for each edge {u, v} of G1 , add the arcs (e{u,v} , u) and (e{u,v} , v) to G2 .
30
w1
w3
w2
w4
w1
w3
w2
w4
w02
w03
w04
v
v
v0
s1
s2
s3
s1
Tv
s2
s3
Figure 14: Reduction of the degree of a vertex v to three. On the left, the original neighborhood of v is shown. The right side shows v after
modification. The tree structure Tv constructed in the proof is highlighted using a gray background.
Set k2 = k1 + 2(n − 3), where n is the number of vertices of G1 . We claim that I1 is a yes-instance if and only if I2 is a
yes-instance.
First, suppose that there is a multiway cut S of size at most k1 for G1 . Then, we obtain a partitioning set of size
at most k2 for G2 as follows: if a vertex v belongs to the same connected component of G1 \ S as terminal ti , then
remove every arc (v, s j ) with j , i from G2 . Furthermore, for each edge {u, v} ∈ S, remove either the arc (e{u,v} , u) or
the arc (e{u,v} , v) from G2 . One can easily check that we end up with a valid partitioning set of size k2 = k + 2(n − 3)
for G2 : we delete at most k arcs from Z to Y and, for each of the n − 3 vertices in Y \ T , delete two arcs from Y to X.
There are no arcs from X to Z.
Conversely, suppose that we are given a minimal partitioning set S of size at most k2 for G2 . Note that it has
to remove at least two of the three outgoing arcs of each vertex v2 ∈ Y \ T but cannot remove all three of them:
contrary to Observation 1, this would create a new sink. Thus, S deletes 2(n − 3) arcs from Y to X and the remaining
k2 − 2(n − 3) = k1 arcs from Z to Y . Therefore, we can define the following multiway cut of size k1 for G1 : remove an
edge {u, v} from G1 if and only if one of the arcs (e{u,v} , u) and (e{u,v} , v) is removed from G2 by S. Again, one can
easily check that we end up with a valid multiway cut.
It remains to modify the instance I2 = (G2 , ω2 , k2 ) to get an instance I3 = (G3 , ω3 , k3 ) of maximum degree three.
To this end, first we show how to reduce the outdegree of each vertex of G2 to two. Thereafter, we show how to reduce
the indegree of each vertex of G2 to one by introducing gadget vertices, each having indegree two and outdegree one.
The construction is illustrated in Figure 14.
Note that all vertices of G2 with outdegree larger than two are in Y . In order to decrease the degree of these
vertices, we obtain a graph G03 from G2 by carrying out the following modifications (see Figure 14) to G2 : for each
vertex v ∈ Y , with N + (v) = {s1 , s2 , s3 }, remove (v, s1 ) and (v, s2 ) add a new vertex v0 , and insert the three arcs (v, v0 ),
(v0 , s1 ), and (v0 , s2 ).
We show that (G03 , ω30 , k2 ) is a yes-instance if and only if (G2 , ω2 , k2 ) is. To this end, for v ∈ Y , let Tv be the
induced subgraph G03 [{v, v0 , s1 , s2 , s3 }]. In G2 , a minimal partitioning set removes exactly two of the outgoing arcs of v,
since s1 , s2 , and s3 are sinks. It is enough to show that a minimal partitioning set S removes exactly two arcs in Tv from G03
in such a way that there remains exactly one directed path from v to exactly one of s1 , s2 , or s3 . This remaining directed
path then one-to-one corresponds to the arc that a partitioning set would leave in G2 between v and s1 , s2 , or s3 . Since
s1 , s2 , and s3 are sinks, S indeed has to remove at least two arcs from Tv : otherwise, two sinks will belong to the same
connected component. However, due to Observation 1, S cannot remove more than two arcs from Tv . Moreover, again
exploiting Observation 1, the two arcs removed by S leave a single directed path from v to exactly one of s1 , s2 , or s3 .
We have seen that (G03 , ω30 , k2 ) is equivalent to (G2 , ω2 , k2 ) and that all vertices of G03 have outdegree two. To shrink
31
the overall maximum degree to three, it remains to reduce the indegrees. Note that the vertices newly introduced
in the previous step already have indegree one. We obtain graph G3 of maximum degree three from G03 as follows.
For each vertex v with |N − (v)| = |{w1 , . . . , wd − (v) }| ≥ 2, do the following (see Figure 14): for j = 2, . . . , d − (v), remove the arc (w j , v) and add a vertex w0j together with the arc (w j , w0j ). Moreover, add the arcs (w1 , w02 ), (wd − (v) , v),
and (w0j , w0j+1 ) for each j ∈ {2, . . . , d − (v) − 1}. Now, every vertex of V (G03 ) in G3 has indegree one and outdegree two,
while the newly introduced vertices have indegree two and outdegree one. It follows that all vertices in G3 have degree
at most three.
It remains to show that (G3 , ω3 , k2 ) is a yes-instance if and only if (G03 , ω30 , k2 ) is. It then follows that (G3 , ω3 , k2 ) is
a yes-instance if and only if (G2 , ω2 , k2 ) is. To this end, note that by Observation 1, among the introduced arcs, only the
arcs w1 , w02 and (w j , w0j ) can be removed by a minimal partitioning set. From this, there is a one-to-one correspondence
between deleting the arc w1 , w02 or (w j , w0j ) in the graph G3 and deleting the arc (w j , v) in the graph G03 .
6. Outlook
We have presented two fixed-parameter algorithms for DAG PARTITIONING, one with respect to the weight k of
the partitioning set sought and one with respect to the parameter treewidth t.
We demonstrated the feasibility of applying the fixed-parameter algorithm for the parameter k (Algorithm 1) to
large input instances with optimal partitioning sets of small weight. However, we were unable to solve the instances in
the data set of Leskovec et al. [26], since the weight of optimal partitioning sets is too large. We found out that the
heuristic presented by Leskovec et al. [26] finds nearly optimal partitioning sets on the instances that our algorithm
works on best. However, we have also seen that one does not have to specially craft adversarial instances to make the
heuristic perform badly. Surprisingly, our algorithm for DAG PARTITIONING is much simpler and more efficient than
the algorithm for M ULTIWAY C UT by Xiao [34], although M ULTIWAY C UT is much easier to approximate than DAG
PARTITIONING [23, 1].
On the theoretical side, we improved a fixed-parameter algorithm by Alamdari and Mehrabian [1] such that the
running time now depends on the treewidth of the input graph rather than on its pathwidth. However, our algorithm, as
well as the algorithm of Alamdari and Mehrabian [1], are practically inapplicable.
Towards solving the instances of Leskovec et al. [26] exactly in reasonable time, a challenging task would be to
analyze the data used by Leskovec et al. [26] in order to find parameters that are small and make DAG PARTITIONING
fixed-parameter tractable, that is, to take a data-driven approach to parameterizing DAG PARTITIONING.
Acknowledgments
We are thankful to Rolf Niedermeier and to the anonymous referees of CIAC’13 and Discrete Applied Mathematics for
helpful comments.
René van Bevern acknowledges supported by the Russian Foundation for Basic Research (RFBR), project 16-3160007 mol a dk, while working at Novosibirsk State University, and by the German Research Foundation (DFG),
project DAPA (NI 369/12), while working at TU Berlin. Robert Bredereck acknowledges support by DFG project PAWS
(NI 369/10). Morgan Chopin was supported by the DAAD during a three-month visit to TU Berlin in summer 2012.
Falk Hüffner acknowledges support by DFG projects PABI (NI 369/7) and ALEPH (HU 2139/1). Ondřej Suchý
acknowledges support by DFG project AREG (NI 369/9) while at TU Berlin and by the Czech Science Foundation,
project 14-13017P.
References
[1] S. Alamdari and A. Mehrabian. On a DAG partitioning problem. In Proceedings of the 9th International Workshop on Algorithms and
Models for the Web Graph (WAW’12), volume 7323 of Lecture Notes in Computer Science, pages 17–28. Springer, 2012. doi: 10.1007/
978-3-642-30541-2 2.
[2] A.-L. Barabási and R. Albert. Emergence of scaling in random networks. Science, 286(5439):509–512, 1999. doi: 10.1126/science.286.5439.
509.
[3] R. van Bevern. Towards optimal and expressive kernelization for d-Hitting Set. Algorithmica, 70(1):129–147, 2014. doi: 10.1007/
s00453-013-9774-3.
32
[4] R. van Bevern, S. Hartung, F. Kammer, R. Niedermeier, and M. Weller. Linear-time computation of a linear problem kernel for Dominating Set
on planar graphs. In Proceedings of the 6th International Symposium on Parameterized and Exact Computation (IPEC’11), volume 7112 of
Lecture Notes in Computer Science, pages 194–206. Springer, 2012. doi: 10.1007/978-3-642-28050-4 16.
[5] R. van Bevern, R. Bredereck, M. Chopin, S. Hartung, F. Hüffner, A. Nichterlein, and O. Suchý. Parameterized complexity of DAG Partitioning.
In Proceedings of the 8th International Conference on Algorithms and Complexity (CIAC’13), number 7878 in Lecture Notes in Computer
Science, pages 49–60. Springer, 2013.
[6] H. L. Bodlaender. A linear-time algorithm for finding tree-decompositions of small treewidth. SIAM Journal on Computing, 25(6):1305–1317,
1996. doi: 10.1137/S0097539793251219.
[7] H. L. Bodlaender. Kernelization: New upper and lower bound techniques. In Proceedings of the 4th International Workshop on Parameterized
and Exact Computation (IWPEC’09), volume 5917 of Lecture Notes in Computer Science, pages 17–37. Springer, 2009. doi: 10.1007/
978-3-642-11269-0 2.
[8] H. L. Bodlaender, B. M. P. Jansen, and S. Kratsch. Kernelization lower bounds by cross-composition. SIAM Journal on Discrete Mathematics,
28(1):277–305, 2014. doi: 10.1137/120880240.
[9] T. H. Cormen, C. E. Leiserson, R. L. Rivest, and C. Stein. Introduction to Algorithms. MIT Press, 2nd edition, 2001.
[10] M. Cygan, F. V. Fomin, L. Kowalik, D. Lokshtanov, D. Marx, M. Pilipczuk, M. Pilipczuk, and S. Saurabh. Parameterized Algorithms. Springer,
2015. doi: 10.1007/978-3-319-21275-3.
[11] E. Dahlhaus, D. S. Johnson, C. H. Papadimitriou, P. D. Seymour, and M. Yannakakis. The complexity of multiterminal cuts. SIAM Journal on
Computing, 23(4):864–894, 1994. doi: 10.1137/S0097539792225297.
[12] R. G. Downey and M. R. Fellows. Fundamentals of Parameterized Complexity. Springer, 2013. doi: 10.1007/978-1-4471-5559-1.
[13] S. Fafianie and S. Kratsch. A shortcut to (sun)flowers: Kernels in logarithmic space or linear time. In Proceedings of the 40th International
Symposium on Mathematical Foundations of Computer Science (MFCS’15), volume 9235 of Lecture Notes in Computer Science, pages
299–310. Springer, 2015. doi: 10.1007/978-3-662-48054-0 25.
[14] M. R. Fellows, B. M. P. Jansen, and F. A. Rosamond. Towards fully multivariate algorithmics: Parameter ecology and the deconstruction of
computational complexity. European Journal of Combinatorics, 34(3):541–566, 2013. doi: 10.1016/j.ejc.2012.04.008.
[15] J. Flum and M. Grohe. Parameterized Complexity Theory. Springer, 2006. doi: 10.1007/3-540-29953-X.
[16] L. Fortnow and R. Santhanam. Infeasibility of instance compression and succinct PCPs for NP. Journal of Computer and System Sciences, 77
(1):91–106, 2011. doi: 10.1145/1374376.1374398.
[17] J. Guo and R. Niedermeier. Invitation to data reduction and problem kernelization. ACM SIGACT News, 38(1):31–45, 2007. doi: 10.1145/
1233481.1233493.
[18] T. Hagerup. Simpler linear-time kernelization for Planar Dominating Set. In Proceedings of the 6th International Symposium on Parameterized
and Exact Computation (IPEC’11), volume 7112 of Lecture Notes in Computer Science, pages 181–193. Springer, 2012. doi: 10.1007/
978-3-642-28050-4 15.
[19] R. Impagliazzo and R. Paturi. On the complexity of k-SAT. Journal of Computer and System Sciences, 62(2):367–375, 2001. doi:
10.1006/jcss.2000.1727.
[20] R. Impagliazzo, R. Paturi, and F. Zane. Which problems have strongly exponential complexity? Journal of Computer and System Sciences, 63
(4):512–530, 2001. doi: 10.1006/jcss.2001.1774.
[21] H. Jeong, Z. Néda, and A. L. Barabási. Measuring preferential attachment in evolving networks. Europhysics Letters, 61(4):567–572, 2003.
doi: 10.1209/epl/i2003-00166-9.
[22] F. Kammer. A linear-time kernelization for the rooted k-leaf outbranching problem. Discrete Applied Mathematics, 193:126–138, 2015. doi:
10.1016/j.dam.2015.04.028.
[23] D. R. Karger, P. Klein, C. Stein, M. Thorup, and N. E. Young. Rounding algorithms for a geometric embedding of Minimum Multiway Cut.
Mathematics of Operations Research, 29(3):436–461, 2004. doi: 10.1287/moor.1030.0086.
[24] T. Kloks. Treewidth. Computations and Approximations, volume 842 of Lecture Notes in Computer Science. Springer, 1994.
[25] C. Komusiewicz and R. Niedermeier. New races in parameterized algorithmics. In Proceedings of the 37th International Symposium on
Mathematical Foundations of Computer Science (MFCS’12), volume 7464 of Lecture Notes in Computer Science, pages 19–30. Springer, 2012.
doi: 10.1007/978-3-642-32589-2 2.
[26] J. Leskovec, L. Backstrom, and J. Kleinberg. Meme-tracking and the dynamics of the news cycle. In Proceedings of the 15th ACM SIGKDD
International Conference on Knowledge Discovery and Data Mining (KDD’09), pages 497–506. ACM, 2009. doi: 10.1145/1557019.1557077.
[27] S. Lohr. Study measures the chatter of the news cycle. New York Times, July 13th:B1, 2009. New York edition.
[28] R. Niedermeier. Invitation to Fixed-Parameter Algorithms. Oxford University Press, 2006. doi: 10.1093/acprof:oso/9780198566076.001.0001.
[29] R. Niedermeier. Reflections on multivariate algorithmics and problem parameterization. In Proceedings of the 27th International Symposium
on Theoretical Aspects of Computer Science (STACS’10), volume 5 of LIPIcs, pages 17–32. Schloss Dagstuhl–Leibniz-Zentrum für Informatik,
2010. doi: 10.4230/LIPIcs.STACS.2010.2495.
[30] R. Niedermeier and P. Rossmanith. A general method to speed up fixed-parameter-tractable algorithms. Information Processing Letters, 73
(3-4):125–129, 2000. doi: 10.1016/S0020-0190(00)00004-1.
[31] D. D. S. Price. A general theory of bibliometric and other cumulative advantage processes. Journal of the American Society for Information
Science, 27(5):292–306, 1976. doi: 10.1002/asi.4630270505.
[32] F. Protti, M. Dantas da Silva, and J. Szwarcfiter. Applying modular decomposition to parameterized cluster editing problems. Theory of
Computing Systems, 44(1):91–104, 2009. doi: 10.1007/s00224-007-9032-7.
[33] C. Suen, S. Huang, C. Eksombatchai, R. Sosic, and J. Leskovec. NIFTY: A system for large scale information flow tracking and clustering. In
Proceedings of the 22nd International Conference on World Wide Web (WWW’13), pages 1237–1248. IW3C2, 2013. doi: 10.1145/2488388.
2488496.
[34] M. Xiao. Simple and improved parameterized algorithms for multiterminal cuts. Theory of Computing Systems, 46(4):723–736, 2010. doi:
10.1007/s00224-009-9215-5.
33
| 8cs.DS
|
Dichotomy for Digraph Homomorphism
Problems
arXiv:1701.02409v4 [cs.CC] 29 Jul 2017
Tomás Feder∗
Jeff Kinne
†
Arash Rafiey
‡
Abstract
Update – An issue has been found in the correctness of our algorithm, and we are
working to resolve the issue. Until a resolution is found, we retract our main claim
that our approach gives a combinatorial solution to the CSP conjecture. We remain
hopeful that we can resolve the issues. We thank Ross Willard for carefully checking
the algorithm and pointing out the mistake in the version of this manuscript. We briefly
explain one issue at the beginning of the text, and leave the rest of the manuscript
intact for the moment . Ross Willard is posting a more involved description of a
counter-example to the algorithm in the present manuscript. We have an updated
manuscript that corrects some issues while still not arriving at a full solution; we will
keep this private as long as unresolved issues remain.
Previous abstract – We consider the problem of finding a homomorphism from an
input digraph G to a fixed digraph H. We show that if H admits a weak-near-unanimity
polymorphism φ then deciding whether G admits a homomorphism to H (HOM(H))
is polynomial time solvable. This confirms the conjecture of Maroti and McKenzie
[MM08], and consequently implies the validity of the celebrated dichotomy conjecture
due to Feder and Vardi [FV93]. We transform the problem into an instance of the list
homomorphism problem where initially all the lists are full (contain all the vertices of
H). Then we use the polymorphism φ as a guide to reduce the lists to singleton lists,
which yields a homomorphism if one exists.
1 Introduction
Update – discussion of counter-example to correctness of the algorithm in this manuscript
– In this brief update we aim the discussion at those who already are familiar with our
∗
268 Waverley Street, Palo Alto, CA 94301, United States, [email protected]
Indiana State University, IN, USA, [email protected]
‡
Indiana State University, IN, USA [email protected] and Simon Fraser University, BC, Canada,
[email protected]
†
1
approach from reading this or another manuscript. Our algorithm makes a decision based
on the output of a test Tx,a,b on a smaller instance of the homomorphism problem. Here a, b
are possible images for x ∈ V (G), of a homomorphism from G to H.
The present manuscript assumes that the test Tx,a,b outputs“yes” and based on the correctness of the test a is removed from further consideration for x. The test Tx,a,b uses the
properties of the weak-nuf polymorphism φ. However, it is conceivable that the test Tx,a,b
fails and this means we should not remove a from the list of possible images of x. We had
incorrectly claimed in the manuscript that the properties of φ and pre-tests in the algorithm
gaurtanee the test always passes. But we can simply construct such an example where the
test must fail in the algorithm as follows. Let H be a digraph with two weakly connected
components H1 , H2 . The weak nuf polymorphism φ could be of arity 3 and such that for every a ∈ H1 and every b ∈ H2 , φ(a, b, b) = φ(b, b, a) = φ(b, a, b) = c for some c ∈ H2 . Suppose
there exists a homomorphism from G (weakly connected) to H that maps x to a and hence
the entire graph G must be mapped to H1 . Moreover, suppose there is no homomorphism
from G to H2 . The algorithm does consider the test Tx,a,b eventually for such G and H.
According to the test Tx,a,b , we remove a from further consideration for x which leads us to
remove the possible homomorphism from G to H.
Note that one can assume H is weakly connected as follows. Suppose H1 , H2 are balanced
digraphs with ` levels (we can partitioned the vertics of Hi , i = 1, 2, into ` parts where all
the arcs of Hi go from a vertex in some part j to part j + 1). Then an extra vertex a0 can be
added and connected to all the vertices of H1 , H2 on the lowest level. This way we obtain
the weakly connected digraph H = H1 ∪ H2 ∪ {a0 } with ` + 1 levels.
We may assume G is also balanced and has ` levels. An extra vertex x0 can be added to
G with arcs to every vertex of G on the lowest level. Now G0 = G ∪ {x0 } is also a balanced
digraph with ` + 1 levels. Note that in any homomorphism from G0 to H, x0 must be mapped
to a0 and any other vertex of G0 must map to H − {x}.
We note that Ross Willard may post a concrete counter-example (and further discussion of
his example) for which H contains 197 vertices in H. The example is inspired from instances
of the CSP problem, so called Semi-lattice block Maltsev.
Previous introduction – For a digraph G, let V (G) denote the vertex set of G and let
A(G) denote the arcs (aka edges) of G. An arc (u, v) is often written as simply uv to shorten
expressions. Let |G| denote the number of vertices in G.
A homomorphism of a digraph G to a digraph H is a mapping g of the vertex set of G to
the vertex set of H so that for every arc uv of G the image g(u)g(v) is an arc of H. A natural
decision problem is whether for given graphs G and H there is a homomorphism of G to H.
If we view graphs as digraphs in which each edge is replaced by the two opposite directed
arcs, we may apply the definition to graphs as well. An easy reduction from the k-coloring
problem shows that this decision problem is N P -hard: a graph G admits a 3-coloring if
and only if there is a homomorphism from G to K3 , the complete graph on 3 vertices. As
a homomorphism is easily verified if the mapping is given, the homomorphism problem is
2
contained in N P and is thus N P -complete.
The following version of the problem has attracted much recent attention. For a fixed
digraph H the problem HOM (H) asks if a given input digraph G admits a homomorphism
to H. Note that while the above reduction shows HOM (K3 ) is NP-complete, HOM (H)
can be easy (in P ) for some graphs H: for instance if H contains a vertex with a selfloop, then every graph G admits a homomorphism to H. Less trivially, for H = K2 (or
more generally, for any bipartite graph H), there is a homomorphism from G to K2 if and
only if G is bipartite. A very natural goal is to identify precisely for which digraphs H
the problem HOM (H) is easy. In the case of graphs the classification has turned out to
be this: if H contains a vertex with a self-loop or is bipartite, then HOM (H) is in P ,
otherwise it is N P -complete [HN90] (see [B05, S10] for shorter proofs). This classification
result implies a dichotomy of possibilities for the problems HOM (H) when H is a graph,
each problem being N P -complete or in P . However, the dichotomy of HOM (H) remained
open for general digraphs H. It was observed by Feder and Vardi [FV93] that this problem
is equivalent to the dichotomy of a much larger class of problems in N P , in which H is
a fixed finite relational structure. These problems can be viewed as constraint satisfaction
problems with a fixed template H [FV93], written as CSP (H). A constraint satisfaction
problem CSP (H) consists of (a) a relational structure H that specifies a set V of variables
that each come from some domain D and (b) a set C of constraints giving restrictions on
the values allowed on the variables.
The question is whether all constraints can be simultaneously satisfied. 3SAT is a prototypical instance of CSP, where each variable takes values of true or false (a domain size of
two) and the clauses are the constraints. Digraph homomorphism problems can also easily
be converted into CSPs: the variables V are the vertices of G, each must be assigned a vertex
in H (meaning a domain size of |V (H)|), and the constraints encode that each arc of G must
be mapped to an arc in H.
Feder and Vardi argued in [FV93] that in a well defined sense the class of problems
CSP (H) would be the largest subclass of N P in which a dichotomy holds. A fundamental
result of Ladner [L75] asserts that if P 6= N P then there exist N P -intermediate problems
(problems neither in P nor N P -complete), which implies that there is no such dichotomy
theorem for the class of all N P problems. Non-trivial and natural subclasses which do have
dichotomy theorems are of great interest. Feder and Vardi made the following Dichotomy
Conjecture: every problem CSP (H) is N P -complete or is in P . This problem has animated
much research in theoretical computer science. For instance the conjecture has been verified
when H is a conservative relational structure [B11], or a digraph with all in-degrees and
all-out-degrees at least one [BKN09].
Numerous special cases of this conjecture have been verified [ABISV09, B06, BH90, BHM88,
CVK10, D00, F01, F06, FMS04, LZ03, S78].
It should be remarked that constraint satisfaction problems encompass many well known
computational problems, in scheduling, planning, database, artificial intelligence, and constitute an important area of applications, in addition to their interest in theoretical computer
3
science [CKS01, D92, V00, K92].
While the paper of Feder and Vardi [FV93] did identify some likely candidates for the
boundary between easy and hard CSP -s, it was the development of algebraic techniques
by Jeavons [J98] that lead to the first proposed classification [BJK05]. The algebraic approach depends on the observation that the complexity of CSP (H) only depends on certain
symmetries of H, the so-called polymorphisms of H. For a digraph H a polymorphism φ
of arity k on H is a homomorphism from H k to H. Here H k is a digraph with vertex set
{(a1 , a2 , . . . , ak )|a1 , a2 , . . . , ak ∈ V (H)} and arc set {(a1 , a2 , . . . , ak )(b1 , b2 , . . . , bk ) | ai bi ∈
A(H) for all 1 ≤ i ≤ k}. For a polymorphism φ, φ(a1 , a2 , . . . , ak )φ(b1 , b2 , . . . , bk ) is an arc of
H whenever (a1 , a2 , . . . , ak )(b1 , b2 , . . . , bk ) is an arc of H k .
Over time, one concrete classification has emerged as the likely candidate for the dichotomy. It is expressible in many equivalent ways, including the first one proposed in
[BJK05]. There were thus a number of equivalent conditions on H that were postulated to
describe which problems CSP (H) are in P . For each, it was shown that if the condition is
not satisfied then the problem CSP (H) is N P -complete (see also the survey [HN08]). One
such condition is the existence of a weak near-unanimity polymorphism (Maroti and McKenzie [MM08]). A polymorphism φ of H of arity k is a k near unanimity polymorphism (k-nu)
on H, if φ(a, a, . . . , a) = a for every a ∈ V (H), and φ(a, a, . . . , a, b) = φ(a, a, . . . , b, a) =
· · · = φ(b, a, . . . , a) = a for every a, b ∈ V (H). If we only have φ(a, a, . . . , a) = a for every
a ∈ V (H) and φ(a, a, . . . , a, b) = φ(a, a, . . . , b, a) = · · · = φ(b, a, . . . , a) [not necessarily a] for
every a, b ∈ V (H), then φ is a weak k-near unanimity function (weak k-nu).
Given the N P -completeness proofs that are known, the proof of the Dichotomy Conjecture
reduces to the claim that a relational structure H which admits a weak near-unanimity
polymorphism has a polynomial time algorithm for CSP (H). As mentioned earlier, Feder
and Vardi have shown that is suffices to prove this for HOM (H) when H is a digraph. This
is the main result of our paper.
Note that the real difficulty in the proof of the graph dichotomy theorem in [HN90] lies in
proving the N P -completeness. By contrast, in the digraph dichotomy theorem proved here
it is the polynomial-time algorithm that has proven more difficult.
While the main approach in attacking the conjecture has mostly been to use the highly
developed techniques from logic and algebra, and to obtain an algebraic proof, we go in the
opposite direction and develop a combinatorial algorithm.
Our main result is the following.
Theorem 1.1 Let H be a digraph that admits a weak near-unanimity function. Then
HOM (H) is in P . Deciding whether an input digraph G admits a homomorphism to H
can be done in time O(|G|3 |H|k+1 ) (k is the arity of the weak near-unanimity function).
Together with the N P -completeness result of [MM08], this settles the CSP Conjecture in
the affirmative. We note that H is fixed and also k ≤ 2|H| according to [JB17]. Therefore
we may consider O(|G|3 ) as running time in Theorem 1.1.
4
Our Methods, Very High Level View We start with a general digraph H and a weak k-nu
φ of H. We turn the problem HOM (H) into a related problem of seeking a homomorphism
with lists of allowed images. The list homomorphism problem for a fixed digraph H, denoted
LHOM (H), has as input a digraph G, and for each vertex x of G an associated list (set)
of vertices L(x) ⊆ V (H), and asks whether there is a homomorphism g of G to H such
that for each x ∈ V (G), the image g(x) is in L(x). Such a homomorphism is called a list
homomorphism of G to H with respect to the lists L. List homomorphism problems are
known to have nice dichotomies. For instance when H is a reflexive graph (each vertex has
a loop), the problem LHOM (H) is polynomial when H is an interval graph and is N P complete otherwise [FH98]. Similar list homomorphism dichotomies were proved for general
graphs [FHH03, FHH07], and more recently also for digraphs [HR11]. In fact, motivated
by the results in [FH98, FHH03], Bulatov [B11] proved that the list version of constraint
satisfaction problems has a dichotomy for general relational systems.
It is not difficult to see that there are digraphs H such that HOM (H) is polynomial
while LHOM (H) is N P -complete. For instance, the reflexive four-cycle H has loops and
so HOM (H) is trivial, while LHOM (H) is NP-complete since H is not an interval graph.
However, we transform the problem HOM (H) into a restricted version of LHOM (H) in
which the lists satisfy an additional property related to the weak k-nu φ.
One of the common ingredients in CSP algorithms is the use of consistency checks to
reduce the set of possible values for each variable (see, for example the algorithm outlined
in [HN04] for CSP (H) when H admits a near unanimity function). Our algorithm includes
such a consistency check as a first step. We begin by performing a pair consistency check of
the list of vertices in the input digraph G. For each pair (x, y) of V (G) × V (G) we consider
a list of possible pairs (a, b), a ∈ L(x) (the list in H associated with x ∈ G) and b ∈ L(y).
Note that if xy is an arc of G and ab is not an arc of H then we remove (a, b) from the list
of (x, y). Moreover, if (a, b) ∈ L(x, y) and there exists z such that there is no c for which
(a, c) ∈ L(x, z) and (c, b) ∈ L(z, y) then we remove (a, b) from the list of (x, y). We continue
this process until no list can be modified. If there are empty lists then clearly there is no
list homomorphism.
After performing pair consistency checks (and repeating the consistency checks throughout
the algorithm), the main structure of the algorithm is to perform pairwise elimination, which
focuses on two vertices a, b of H that occur together in some list L(x), x ∈ V (G), and finds
a way to eliminate a or b from L(x) without changing a feasible problem into an unfeasible
one. In other words if there was a list homomorphism with respect to the old lists L, there
will still be one with respect to the updated lists L. This process continues until either
a list becomes empty, certifying that there is no homomorphism with respect to L (and
hence no homomorphism at all), or until all lists become singletons, specifying a concrete
homomorphism of G to H. This technique, due to the last author, has been successfully
used in several other papers [HR11, HR12, EHLR14]. In this paper, the choice of which a
or b is eliminated, and how, is governed by the given weak near-unanimity polymorphism
φ. In fact, we define a family of mappings fx , x ∈ V (G) which are each polymorphisms
5
derived from φ and use these polymorphisms as a guide. The heart of the algorithm is a
delicate procedure for updating the lists L(x) and polymorphisms fx in such a way that (i)
feasibility is maintained, and (ii) the polymorphisms fx remain polymorphisms (which is key
to maintaining feasibility). We present two approaches where the first algorithm is recursive
and the second algorithm is not recursive. For the sake of the completeness we repeat the
definitions in both parts. The introduction is presented once. We think the first algorithm
is conceptually is easier to follow.
2 First Approach (recursive algorithm )
An oriented walk (path) is obtained from a walk (path) by orienting each of its edges. The
net-length of a walk W , is the number of forward arcs minus the number of backward arcs
following W from the beginning to the end. An oriented cycle is obtained from a cycle by
orienting each of its edges. We say two oriented walks X, Y are congruent if they follow the
same patterns of forward and backward arcs.
Given digraphs G and H, let G × H k be a digraph on the vertices {(y; a1 , a2 , . . . , ak )|y ∈
V (G), ai ∈ V (H), 1 ≤ i ≤ k} with the arcs (y; a1 , a2 , ..., ak )(y 0 ; b1 , b2 , ..., bk ) where yy 0 is an
arc of G and each ai bi , 1 ≤ i ≤ k, is an arc of H. By convention, we shall further restrict the
use of the symbol G×H k to the digraph induced on the vertices {(y; a1 , a2 , . . . , ak )|y ∈ V (G),
ai ∈ L(y), 1 ≤ i ≤ k} where L(y) is the set of vertices in H that are being considered as
images of a homomorphism from G to H.
Definition 2.1 (Homomorphism consistent with Lists) Let G and H be digraphs. For
each x ∈ V (G), let list of x, L(x), be a subset of H. Let k > 1 be a constant integer.
A function f : G × H k → H is a homomorphism consistent with L if the following hold.
• List property : for every x ∈ V (G) and every a1 , a2 , . . . , ak ∈ L(x), f (x; a1 , a2 , . . . , ak ) ∈
L(x)
• Adjacency property: for every x, y ∈ V (G) and every a1 , ..., ak ∈ L(x), b1 , ..., bk ∈ L(y),
if xy is an arc of G and ai bi is an arc of H for each 1 ≤ i ≤ k then f (x; a1 , ..., ak )f (y; b1 , ..., bk )
is an arc of H.
In addition if f has the following property then we say f has the weak nu property.
• for every x ∈ V (G), {a, b} ⊆ L(x), we have f (x; a, b, b, ..., b) = f (x; b, a, b, ..., b) = ... =
f (x; b, b, b, ...a).
We note that this definition is tailored to our purposes and in particular differs from the
standard definition of weak k-nu as follows.
(a) f is based on two digraphs G and H rather than just H (we think of this as starting
with a traditional weak k-nu on H and then allowing it to vary somewhat for each
x ∈ G),
6
(b) We do not require that f (y; d, d, d, ..., d) = d (this is not required in our algorithm, and
in fact is more convenient to leave out).
Notation For simplicity let (bk , a) = (b, b, . . . , b, a) be a k-tuple of all b’s but with an a in
the k th coordinate. Let (x; bk , a) be a (k + 1)-tuple of x, (k − 1) b’s and a in the (k + 1)th
coordinate.
2.1 Main Procedure
Algorithm 1 The main algorithm for solving the digraph homomorphism problem.
1: function DigraphHom(G, H, φ)
. G and H digraphs, φ a weak k-nu on H
2:
for all x ∈ G, let L(x) = V (H)
3:
for all x ∈ G and a1 , ..., ak ∈ V (H), let f (x; a1 , ..., ak ) = φ(a1 , ..., ak )
4:
PreProcessing(G, H, L)
5:
RemoveNotMinority(G, H, L, f )
6:
Note: now, for all x ∈ V (G) and a, b ∈ L(x) we know f (x; bk , a) = a
7:
RemoveMinority(G, H, L, f )
8:
if RemoveMinority produces a homomorphism then return true
The main algorithm starts with applying the arc consistency and pair consistency on the
lists L by calling Algorithm 2.
Algorithm 4 (RemoveNotMinority function) is the key subroutine of the main algorithm.
It starts with w = (x; bk , a) where f (w) = c 6= a and then it will modify f by setting
f (x; e1 , e2 , . . . , ek ) = f (w) for every k-tuple e1 , e2 , . . . , ek ∈ L(x) with f (x; e1 , e2 , . . . , ek ) = a.
Now in order to have a homomorphism from G × H k to H consistent with L, it recursively
solves an instance of the problem to be able to modify f as necessary.
After the main loop in Algorithm 1, we end up with a so-called Maltsev or minority instance of the problem – in which we have a homomorphism f consistent with L such that
for every y ∈ V (G) and every c, d ∈ L(y) we have f (y; ck , d) = d. We argue in the next
subsection that such instances can be solved by using the known algorithm of [BD06] (see
the remark at the end of Subsection 2.2). The Maltsev/minority instances can also be solved
in a manner similar to our arguments for RemoveNotMinority (see Section 5.7) .
In what follows we give an insight of why the weak nu property of H is necessary for our
algorithm. For contrary suppose w1 = (x; bk , a) with f (w1 ) = c and w2 = (x; a, b, b, . . . , b)
with f (w2 ) = d. If d = a then in RemoveNotMinority we try to remove a from L(x) if we
start with w1 while we do need to keep a in L(x) because we later need a in L(x) for the
Maltsev algorithm. It might be the case that d 6= a but some non-minority pairs becomes
minority pairs during the execution of Algorithm 4. In other words, for some w3 = (x; bk , e)
with f (w3 ) 6= e we may set f (w3 ) to e. So we need to have f (w1 ) = f (w2 ), the weak nu
7
Algorithm 2 Update lists of x, y based on edge constraints and pair constraint. Call by
reference, i.e. the update to L will be reflected in the calling function
1: function PreProcessing(G, H, L)
2:
Input: digraphs G, H, lists L(x) ⊆ V (H) for each x
. L lists are unary and binary
3:
. The update to L would be available for the function calling PreProcessing
4:
ArcConsistency(G, H, L) and PairConsistency(G, H, L)
function ArcConsistency(G, H, L)
update=True
while update do
if ∃xy(yx) ∈ A(G), a ∈ L(x) s.t. @b ∈ L(y) with ab(ba) ∈ A(H) then
remove a from L(x) and set update=True.
else update=False.
11:
if there is an empty list then print ”no homomorphism” and terminate
5:
6:
7:
8:
9:
10:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
function PairConsistency(G, H, L)
for all (x, y) ∈ V (G) × V (G) do set L(x, y) = {(a, b)|a ∈ L(x), b ∈ L(y)}
for all x ∈ V (G) do set L(x, x) = {(a, a)|a ∈ L(x)}.
for all xy ∈ A(G), a ∈ L(x), b ∈ L(y) do
if ab 6∈ A(H) then remove (a, b) from L(x, y).
update=True
while update do
if ∃x, y, z s.t. @c ∈ L(z) s.t. (a, c) ∈ L(x, z)&(c, b) ∈ L(z, y) then
remove (a, b) from L(x, y) and set update=True.
else update=False.
if there is an empty list then print ”no homomorphism” and terminate
8
property, to start in the main algorithm, Algorithm 1. We do mention at the end of the
proof of Lemma 3.4 that we can start the entire Algorithm 4 from vertex w2 and following
the first coordinates.
First we give the necessary definition for Algorithm 4.
GL Subdigraph of G×H k : Let GL be the digraph with vertices V (GL ) = {(y; ak1 , a2 ) | y ∈
V (G), a1 , a2 ∈ L(y)} and with arcs :
A(GL ) = {(y; ak1 , a2 )(y 0 ; bk1 , b2 ) | yy 0 ∈ A(G), a1 b1 , a2 b2 ∈ A(H)}∪{(y; ak1 , a2 )(y 0 ; bk1 , b2 ) | y 0 y ∈
A(G), b1 a1 , b2 a2 ∈ A(H)}.
Note that a directed path in GL is an oriented path in G × H k . Let B be a walk in H
starting at some vertex a ∈ V (H). We say a directed path W in GL follows B if B is the
projection of W on the (k + 1)th coordinate, i.e. B is induced by the (k + 1)th coordinates
of the vertices in W .
We say a directed path W in GL follows a walk X in G when X is the projection of W
on the first coordinate, i.e. X is induced by the first coordinates of the vertices in W .
Definition 2.2 (Reachable from w in GL : GL (w), GrL (w)) Let w = (x; bk , a) with f (w) 6=
a.
• Let GL (w) be the induced sub-digraph of GL with vertices w0 = (y; ak1 , a2 ) such that
(b, a1 ) ∈ L(x, y) and w0 is reachable from w in GL . 1
• Let GrL (w) be the induced sub-digraph of GL (w) with vertices w1 = (y; ak1 , a02 ) such that
– (b, a1 ) ∈ L(x, y) and (a, a02 ) ∈ L(x, y)
– w0 is reachable from w in GL .
2
Definition 2.3 (Pair digraph to update f : ) Let L be the lists of the vertices of G. For
y ∈ V (G) and a1 ∈ L(y), let Pl (y, a1 ) denote a subset of vertices in L(y). For a positive
integer t > 1, let Plt (y, c1 ) be a subset of L(y) s.t. Plt (y, c1 ) = Pl (y, Plt−1 (y, c1 )). Here
Pl (y, L0 ) = ∪a1 ∈L0 Pl (y, a1 ) where L0 ⊆ L(y) and Pl1 (y, a1 ) = Pl (y, a1 ).
Let GPl be the digraph of vertices (y, c1 , c2 ) with y ∈ V (G) and c1 , c2 ∈ L(y) such that
c2 ∈ Pli (y, c1 ) for some i ≥ 1 and Pl (y, c2 ) = {c2 }.
The arc set of GPl is A(GPl ) = {(y, c1 , c2 )(z, d1 , d2 ) | yz ∈ A(G), c1 d1 , c2 d2 ∈ A(H) OR zy ∈
A(G), d1 c1 , d2 c2 ∈ A(H)}.
Let GPl (x, a, d) be the induced sub-digraph of GPl induced by the vertices of GPl that are
reachable from (x, a, d).
1
2
Suppose w0 is reachable from w via a path P in Gl . If P contains (x; dk , a0 ) then d = b.
Suppose w0 is reachable from w via a path P in Gl . If P contains (x; bk , a0 ) then a0 = a.
9
Definition 2.4 (f -closure of a list : ) We say a list L0 (y) ⊆ L(y) is closed under f if for
every k-tuple a01 , a02 , . . . , a0k ∈ L1 (y) we have f (y; a01 , a02 , . . . , a0k ) ∈ L0 (y).
For L0 (y) ⊆ L(y), let L0f (y) be a minimal set that includes all the element of L0 (y) and it
is closed under f .
Definition 2.5 (restriction of f to a sub-list:) Let L0 (y) ⊆ L(y) for every y ∈ V (G).
Let f |L0 denote the restriction of f under L0 , i.e. for all y ∈ V (G), and for all a1 , a2 , . . . , ak ∈
L0 (y) we have fL0 (y; a1 , a2 , . . . , ak ) = f (y; a1 , a2 , . . . , ak ).
Let X : x1 , x2 , . . . , xn be an oriented walk in G. Let L(X) denote the vertices of H that
lie in the list of the vertices of X. Let X[xi , xj ], 1 ≤ i ≤ j ≤ n, denote the induced sub-path
of X from xi to xj .
An overview of Algorithm 4 : Algorithm 4 starts with a vertex x ∈ V (G) and two vertices
a, b ∈ L(x) such that f (w = (x; bk , a)) = c. The goal is to modify f by setting f (w1 ) to c for
every w1 = (x; e1 , e2 , . . . , ek ) where f (w1 ) = a initially.
The change to f in x imposes a change to f in the neighbors of x. Let w1 = (y; ak1 , a2 ) be
a neighbor of w and suppose ca2 6∈ A(H). We do need to change f (w0 = (y; a01 , a02 , . . . , a0k ))
from a2 to an out-neighbor of c in L(y) in order to keep f as a homomorphism from G × H k
to H and consistent with the lists L. There are several possibilities to change f (w0 ). One
could consider any out-neighbor of b, say a1 and consider vertex w2 = (y; ak1 , a2 ) and modify
f (w0 ) to f (w2 ) which is an out-neighbor of c (see Figure 1 top). The decision for changing
f in y would also imply a change to f in some vertex z ∈ V (G) which is a neighbor of
y. In order to manage these modifications we construct an instance G0 , H, L0 , f 0 so called
Small-Instance.
We first give a reason of the construction of such an instance. Let X be an oriented
cycle in G. The goal is to replace an oriented cycle B containing vertex a in L(X) with an
oriented cycle C in L(X) that does not contain a. We may view B as an image of X under
a homomorphism from G to H that maps x to a.
The Algorithm 4 starts with empty lists E, L0 . It considers the vertices of GrL (w) and for
each vertex (y; ak1 , a2 ) ∈ GrL (w) it adds f (y; ak1 , a2 ) into E(y). Then it sets L0 (y) = Ef (y),
i.e. the closure of E(y) under f . This means L0 (y) contains f (y; c1 , c2 , . . . , ck ) for every
c1 , c2 , . . . , ck ∈ L0 (y). Now in this new instance, for every y ∈ V (G0 ) it defines f 0 to be the
restriction of f to the vertices in L0 (y). This would imply that f 0 is a homomorphism from
G × H k to H and consistent with the lists L0 . Note that L0 (x) = c and hence in the resulting
instance at least one list becomes smaller.
The algorithm solves the instance G0 , L0 , f 0 , H recursively. Note that the stopping point
is when for every y ∈ G0 and every d, e ∈ L0 (y), f 0 (y; dk , e) = e. In order to keep track
of changes to f 0 , it maintains a list Pl (y, a2 ); list of possible images for a2 , which indicates
f (w0 = (y; a01 , a02 , . . . , a0k )) may need to be changed from a2 to c2 ∈ Pl (y, a2 ) (see Figure 1).
We make the Pl list global so they will be accessible in recursion tree.
10
However, in the recursive call, c2 may change to something else. In other words, d2 ∈
Pl (y, c2 ), and this would mean f (w0 ) should be changed from a2 to d2 . In this case d2 ∈
Pl (y, Pl (y, a2 )) or equivalently d2 ∈ Pl2 (y, a2 ). At the bottom of the recursion tree we would
have an element ct ∈ Plt (y, a2 ) such that Pl (y, ct ) = {ct }.
Now the algorithm modifies f according to the Pl lists. It constructs a triple digraph
GPl (x, a, c) that represents the vertices of G for which the f values need to be changed
because of the initial change in L(x). It starts at w and it changes f (w1 = (x, e1 , e2 , . . . , ek ))
from a to c for any such w1 . Suppose the algorithm modifies f (w0 ) for w0 = (y; a01 , a02 , . . . , a0k )
from a2 to an element ct ∈ Plt (y, a2 ). Then it modifies f for f (z; ...) from b2 to an element
in Plt (z, b2 ) which is an out-neighbor of c2 and (y; ak1 , a2 )(z; bk1 , b2 ) ∈ A(GrL (w) see Figure 2).
This suggests the algorithm does a depth first search in digraph GPl (x, a, c) to update f in
the Update-f function.
At this point for every k-tuple e1 , e2 , . . . , ek in L(x), f (x, e1 , e2 , . . . , ek ) 6= a. However,
the algorithm has not modified f on the entire digraph affected by the initial change to f
in L(x). For this reason it constructs a Big-Instance G0 , H, L0 , f 0 . In the Big-Instance it
considers the GL (w). It constructs the lists L0 based on the already updated f after solving
the Small-Instance. It starts with the empty lists E, L0 . For each vertex (y; ak1 , a2 ) ∈ GL (w)
it does the following :
(a) If (y; ak1 , a2 ) ∈ GrL (w) and f (y; a01 , a02 , . . . , a0k ) changed from a2 to c2 in the Update-f
(after the Small-Instance) then add c2 into E(y). In this case we have Pl (y, c2 ) = {c2 }.
It is also possible that a2 = c2 and in this case a2 is added into E(y).
(b) If (y; ak1 , a2 ) 6∈ GrL (w), i.e. (a, a2 ) 6∈ L(x, y) then add f (y; ak1 , a2 ) into E(y). Now in this
case the f (y; ak1 , a2 ) is added into Pl (y, a2 ).
At the end the algorithm sets L0 (y) = Ef (y). Item (b) ensures that we do not repeat
part of the work done in the Small-Instance. The necessary information after the recursive
call for Small-Instance are still available in Pl lists (see Figure 1 middle). Now it runs
the PreProcessing and in particular the ArcConsistency on G0 , H, L0 to remove unnecessary
elements in L0 . Recall that G0 is the induced sub-digraph of G obtained by the projection
of GL (w) on the first coordinates. Note that because of condition in (a) we only need
to have elements d ∈ L0 (z) where (z; bk1 , b2 ) ∈ GL (w) \ GrL (w), (y; ak1 , a2 ) ∈ GrL (w) where
(c, d) ∈ L0 (y, z) and c ∈ L0 (y). This is achieved by running the ArcConsistency on lists L0 .
Note that by definition, L0 (x) does not contain a because we know that no k-tuple in x has
its f value as a. Therefore the resulting instance is smaller than the original one.
In the Big-Instance we pass f |L0 as f 0 . At the end the algorithm updates f according
to Pl lists as described for the Small-Instance and it removes a from L(x). We argue that
the changes to f can be extended upon the changes occurred to f after solving the SmallInstance. Note that all the information that are needed in the Update-f function can be
obtained by looking at the current values of the Pl lists.
11
c1
c2
c3
a04
b04
d04
c
a4
b4
d4
f (x; b, b, a) = c
a
a2
b2
d2
f (y; a1 , a1 , a2 ) = a4
f (y; a01 , a01 , a02 ) = a04
a01
b01
b1
a1
b
d01
a01
f (y; a4 , a4 , a04 ) = c1
a4 , a04 ∈ Pl (y, a2 )
c1 ∈ Pl2 (y, a2 )
a4 , a04 , c1 ∈ L0 (y)
f (z; b4 , b4 , b04 ) = c2
b4 , b04 ∈ Pl (y, b2 )
c2 ∈ Pl2 (z, b2 )
b4 , b04 , c2 ∈ L0 (z)
f (z; b1 , b1 , b2 ) = b4
f (z; b01 , b01 , b02 ) = b04
d1
b01
d01
b
a1
b1
d1
x
y
z
u
f (u; d1 , d1 , d2 ) = d4
f (u; d01 , d01 , d02 ) = d04
f (u; d4 , d4 , d04 ) = c3
d4 , d04 ∈ Pl (u, d2 )
c3 ∈ Pl2 (u, d2 )
d4 , d04 , c3 ∈ L0 (u)
c1
c2
c3
c4
a04
b04
d04
e004
f (x; b, b, a0 ) = c0
f (x; b, b, a) = c
c
0
f (z; b1 , b1 , b2 ) = b4
f (z; b01 , b01 , b02 ) = b04
f (z; b4 , b4 , b04 ) = c2
e04
b4
a4
c
d4
e02
b2
a2
a
a01
b01
b1
a1
b
d2
a01
b01
d01
e01
d1
e1
d01
e01
a0
a
f (u; d1 , d1 , d2 ) = d4
f (u; d01 , d01 , d02 ) = d04
f (u; d4 , d4 , d04 ) = c3
b
f (v; e1 , e1 , e02 ) = e04
f (v; e01 , e01 , e02 ) = e004
f (v; e04 , e04 , e004 ) = c4
b
a1
b1
d1
e1
b
x
y
z
u
v
x
L0 (y) = c1
c3 ∈ L0 (u)
L0 (z) = c2
e04 , e004 , c4 ∈ L0 (v)
Figure 1: The top figure is an example of small-instance. The middle figure : example of biginstance.
e3
e2
e1
a03
a02
a01
b03
b02
b01
d03
d02
d01
c1
c2
c3
a04
b04
d04
f (x; e1 , e2 , e3 ) = a
f (x; e1 , e2 , e3 ) = c
change
c
a4
b4
f (y; a01 .a02 , a03 ) = a2
f (y; a01 .a02 , a03 ) = c1
f (z; b01 , b02 , b03 ) = b2
f (z; b01 , b02 , b03 ) = c2
f (u; d01 , d02 , d03 ) = d2
f (u; d01 , d02 , d03 ) = c3
d4
a
a2
b2
d2
x
y
z
u
Figure 2: Updating f based on the result of the recursion.
12
Algorithm 3 Initialize Pl lists and call Remove-NM
1: function RemoveNotMinority(G, H, L, f )
2:
Input: digraphs G, H, lists L and, weak nu homomorphism f : G × H k → H
3:
For every y ∈ V (G) and every a1 ∈ L(y) set Pl (y, a1 ) = ∅.
. Pl lists are global
4:
Remove-NM(G, H, L, f )
. call by reference for L, f , i.e, the changes to f, L will be reflected to
RemoveNotMinority function
Algorithm 4 updating f so that it remains a homomorphism of G × H k to H consistent
with L and for every x ∈ V (G), a0 , b0 ∈ L(x), f (x; a0k , b0 ) = b0
1: function Remove-NM(G, H, L, f )
2:
Input: digraphs G, H, lists L and, weak nu homomorphism f : G × H k → H
3:
while ∃x ∈ V (G), & a, b ∈ L(x), a 6= b s.t. f (w = (x; bk , a)) = c 6= a &
∃ w2 ∈ V (G × H k ) with f (w2 ) = a do
4:
Let (G0 , L0 , f 0 )= Small-instance(G, L, f, w)
5:
Remove-NM(G0 , H, L0 , f 0 )
6:
for all y ∈ G0 , d ∈ L0 (y) do set Pl (y, d) = {d}.
7:
Update-f(L, Pl , f, x, a)
. Call by reference for f
0
0
0
8:
Let (G , L , f )= Big-instance(G, L, f, w)
9:
Remove-NM(G0 , H, L0 , f 0 )
10:
for all y ∈ G0 , d ∈ L0 (y) do set Pl (y, d) = {d}.
11:
Update-f(L, Pl , f, x, a)
12:
Remove a from L(x)
13:
PreProcessing(G, H, L)
. Update lists L after removing a
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
function Small-Instance(G, L, f, w )
Set G0 to be the induced sub-digraph of G with vertices y s.t. (y; ak1 , a2 ) ∈ GrL (w)
Create new empty lists E, L0 .
for all y ∈ V (G0 ) do
Set E(y) = {f (w1 = (y; ak1 , a2 ))|w1 ∈ GrL (w1 )}
Set L0 (y) = Ef (y)
. see definitions 2.4
0
for all y ∈ V (G ), a1 ∈ L(y) do
Pl (y, a1 ) = {f (w1 = (y; dk , a1 ))|w1 ∈ GrL (w)}
Set f 0 = f |L0
. see definitions 2.5
0
0
0
Return (G , L , f )
13
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
function Big-Instance(G, L, f, w )
Set G0 to be the induced sub-digraph of G with vertices y s.t. (y; ak1 , a2 ) ∈ GL (w)
Create new empty lists E, L0 .
for all w1 = (y; ak1 , a2 ) ∈ GrL (w) do
add c2 into E(y), where f (w1 = (y; ...)) was initially a2 and now f (w1 ) = c2 .
. Happened in Update-f function. Pl (y, c2 ) = {c2 } and possible that c2 = a2
for all y ∈ V (G0 ) do
E(y) = E(y) ∪ {f (w1 )|w1 = (y; ak1 , a2 ) ∈ GL (w) \ GrL (w)}
L0 (y) = Ef (y)
for all y ∈ V (G0 ), a2 ∈ L(y) s.t. (a, a2 ) 6∈ L(x, y) do
Pl (y, a2 ) = {f (w1 )|w1 = (y; ak1 , a2 )) ∈ GL (w)}
. w1 ∈ GL (w) \ GrL (w)
Set f 0 = f |L0
PreProcessing(G0 , L0 ) and Return (G0 , L0 , f 0 )
function Update-f(L, Pl , f, x, a )
. In Small-Instance GPl (x, a, c)
Construct GPl (x, a, d)
Initialize and empty stack ST
ST.Push((x, a, d)) and set V isit[(x, a, d)] to true
while ST is not empty do
v = (y, a1 , c1 ) = ST.P op()
for all (a01 , a02 , . . . , a0k ) ∈ L(y) s.t f (y; a01 , a02 , . . . , a0k ) = a1 do
set f (y; a01 , a02 , . . . , a0k ) = c1
for all v 0 where vv 0 ∈ A(GPl (x, a, d)) do
if V isit[v 0 ] = f alse then
Set V isit[v 0 ] to true and
ST.P ush(v 0 )
14
Description of the Algorithm 4 line by line
3. We go through list of each vertex x and take a, b with f (x; bk , a) 6= a. Going to
modify f so we don’t need a in the list of x. Note that we should take a such that
∃ f (w1 = (x; e1 , e2 , . . . , ek )) = a.
4. We are going to modify f (w1 ) to c. Construct Small-Instance in which L0 (x) = {c}
only.
5. We call to solve the same problem for instance G0 , H, L0 , f 0 . Here f 0 is the restriction
of f over the k-tuples in L0 . f 0 is a homomorphism from G × H k to H and consistent
with the lists L0 .
6. At this point each L0 (y) is either singleton or for every d, d0 ∈ L0 (y) we have f 0 (y; d0k , d) =
d. That’s why we set Pl (y, d) = {d}.
7. We Update f based on the Pl lists. Pl has the information of what should be the new
value for f (y, a01 , a02 , . . . , a0k ) that was initially a2 . The Update to f is with respect to
the Small-Instance. At this point no k-tuple in L(x) has f value a.
8. In order to update f everywhere we construct the Big-Instance. Since we have run the
Remove-NM for Small-Instance, the Big-Instance is smaller than the original instance.
L0 (x) does not have a.
11. Update f based on the Pl lists. Now at this point f is a homomorphism from G × H k
to H and consistent with the lists L and at this point no k-tuple in L(x) has value a.
12. At this point we remove a from L(x).
13. Since we have removed a from L(x) the L lists should be updated.
2–3. [Small-Instance:] To construct G0 and create the L0 lists.
4–6. [Small-Instance:] To make sure L0 is closed under f we set L0 (y) = Ef (y) for every
y ∈ V (G0 ); the image of each k-tuple in L0 (y) should be in L0 (y). Set f 0 to be the same
as f in L0 lists.
7–8. [Small-Instance :] Fill out Pl lists for Small-Instance.
9. [Small-Instance :] Set f 0 to be the the restriction of f on L0 lists.
4–8. [Big-Instance :] First we look at every w1 = (y; ak1 , a2 ) ∈ GrL (w). If the f value of some
k-tuple in y changed from a2 to c2 (when updating f after running Small-Instance)
then we need to add c2 into L0 (y). The goal is not to recompute L0 and recompute Pl
lists in Small-Instance.
15
11–13. [Big-Instance :] We update Pl for those Pl (y, a2 ) that have not been set in lines 4–7.
In other words, for we go through w1 = (y; ak1 , a2 ) ∈ GL (w) \ GrL (w) and add f (w1 )
into Pl (y, a2 ).
2. [Update-F :] Construct the digraph GPl (x, a, d). In the Small-Instance Pl (x, a) is c. In
the Big-Instance instead of c (when call recursively) we may have Pl (x, c) = c1 and
finally we may end up having d ∈ Plt (x, a) where Pl (x, d) = {d}.
4. [Update-F :] We start a depth first search DFS (using stack ST) starting from (x, a, d).
5–7. [Update-F :] Pop an element v = (y, a1 , c1 ) from the top of ST and if f (w1 ) = a1 for
w1 = (y; a01 , a02 , . . . , a0k ), then we change f (w1 ) to c1 (the f value of each such w1 is
changed once).
8–9. [Update-F :] We look at each neighbor of v in GPl (x, c, d) and if we have not visited
that neighbor we push it on the top of the stack ST.
2.2 Minority Algorithm (RemoveMinority)
In this section we show that once the minority case has been reached in our main algorithm,
we can reduce to an already solved setting for homomorphism testing – namely that of the
Maltsev case. We note that this section is independent of the rest of the algorithm.
Note that at this point for every a, b ∈ L(x) we have f (x; bk , a) = a and in particular when
a = b we have f (x; a, a, . . . , a) = a (idempotent property). This is because when a is in L(x)
then it means the Remove-NM procedure did not consider a and in fact did not change the
value of f (x; ...) from a to something else. Note that for the argument below we just need
the idempotent property for those vertices that are in L(x), x ∈ V (G).
A ternary polymorphism h0 is called Maltsev if for all a 6= b, h0 (a, b, b) = h0 (b, b, a) = a.
Note that the value of h0 (b, a, b) is unspecified by this definition.
Let G and H be as input to Algorithm 1, and suppose line 6 of the algorithm has been
reached. We define a homomorphism h : G × H 3 → H consistent with the lists L by setting
h(x; a, b, c) = f (x; a, b, b, . . . , b, c) for a, b, c ∈ L(x). Note that because f has the minority
property for all x ∈ G, a, b ∈ L(x), h is a Maltsev homomorphism consistent with the lists
L.
Note that for the argument below we just need the idempotent property for those vertices
that are in L(x), x ∈ V (G).
Let G0 be the structure obtained from G by making each arc a different binary relation.
In other words, G0 has vertices V (G) and |E(G)| binary relations Re , e ∈ E(G), where
Re = {xy} if e is the arc e = xy.
Let H 0 be the structure where V (H 0 ) is the disjoint union of L(x), x ∈ V (G), and there
are also |E(G)| binary relations Se , e ∈ E(G), where Se is the set of all ordered pairs ab with
ab ∈ E(H), a ∈ L(x), b ∈ L(y), where e = xy. Note that |V (H 0 )| ≥ |V (G0 )| if each L(x) is
non-empty. This may seem unusual for the homomorphism setting, but is certainly allowed.
16
Now note that there is an L-homomorphism of G to H (i.e., a list homomorphism consistent
with lists L) if and only if there is a homomorphism of G0 to H 0 . Homomorphisms of such
structures are mappings f : V (G0 ) → V (H 0 ) such that xy ∈ Re implies f (x)f (y) ∈ Se for all
e ∈ E(G).
Finally, note that the structure H 0 has a Maltsev polymorphism h0 of the ordinary kind.
Indeed, let hx be our Maltsev polymorphisms defined on L(x) by setting hx (a, b, c) =
h(x; a, b, c). We let h0 (a, b, c) = h(x; a, b, c) if a, b, c are from the same L(x), and for a, b, c
not from the same L(x) define h0 (a, b, c) = a unless a = b, in which case define it as
h0 (a, b, c) = c. The definition ensures that h is Maltsev. To check it is a polymorphism, note
that aa0 , bb0 , cc0 ∈ Se is only possible if a, b, c ∈ L(x), a0 , b0 , c0 ∈ L(y), where e = xy. For
those, we have the polymorphism property by assumption.
Now we have a structure with a Maltsev polymorphism, so the Bulatov-Dalmau [BD06]
algorithm applies and solves the homomorphism problem. Note that Corollary 4.2 of the
Bulatov-Dalmau paper explicitly mentions that it is polynomial in both the sizes of G and
H.
Therefore we have the following theorem.
Theorem 2.6 Suppose h : G × H k → H is a minority homomorphism consistent with lists
L on G. Then the existence of an L-homomorphism of G to H can be decided in polynomial
time.
Remark : We have communicated with the authors of [BD06] and they confirmed that
indeed we can apply their algorithm as explained above. We note that it is also possible
to give a direct algorithm for the minority case that is similar to how we handle the “not
minority” case.
3 Proofs
3.1 PreProcessing and List Update
We first show that the standard properties of consistency checking remain true in our setting
– namely, that if the PreProcessing algorithms succeed then f remains a homomorphism
consistent with the lists L if it was before the PreProcessing.
Lemma 3.1 If f is a homomorphism of G × H k → H consistent with L then f is a homomorphism consistent with L after running the pre-processing.
Proof: We need to show that if a1 , a2 , . . . , ak are in L(y) after the pre-processing then
f (y; a1 , a2 , . . . , ak ) ∈ L(y) after the pre-processing. By definition vertex a is in L(y) after
the pre-processing because for every oriented path Y (of some length m ) in G from y to a
fixed vertex z ∈ V (G) there is a vertex a0 ∈ L(z) and there exists a walk B in H from a to
a0 and congruent with Y that lies in L(Y ).
17
Let a01 , a02 , a03 , . . . , a0k ∈ L(z). Let Ai , 1 ≤ i ≤ k be a walk from ai to a0i in L(Y ) and
0
congruent to Y . Let Ai = ai , ai1 , a2i , . . . , am
i , ai and let Y = y, y1 , y2 , . . . , ym , z.
Since f is a homomorphism consistent with L before the pre-processing, f (y; a1 , a2 , . . . , ak ),
0
0
0
m
m
f (y1 ; a11 , a12 , . . . , a1k ), . . . , f (yi ; ai1 , ai2 , . . . , aik ), . . . , f (ym ; am
1 , a2 , . . . , ak ), f (z; a1 , a2 , . . . , ak ) is
a walk congruent with Y . This would imply that there is a walk from f (y; a1 , a2 , . . . , ak ) to
f (z; a01 , a02 , . . . , a0k ) congruent with Y in L(Y ) and hence f (y; a1 , a2 , . . . , ak ) ∈ L(y).
By a similar argument as in the proof of Lemma 3.1 we have the following lemma.
Lemma 3.2 If f is a homomorphism of G×H k → H, consistent with L and a1 , a2 , . . . , ak ∈
L(x), b1 , b2 , . . . , bk ∈ L(y), and (ai , bi ) ∈ L(x, y), 1 ≤ i ≤ k, after pre-processing then
(f (x; a1 , a2 , . . . , ak ), f (y; b1 , b2 , . . . , bk )) ∈ L(x, y) after the pre-processing.
Lemma 3.3 The pair list L0 (x, y) is not empty.
Proof: Let w = (x; bk , a) with f (w) = c 6= a. Suppose (b, a1 ) ∈ L(x, y), and a2 ∈ L(y).
Now by definition (Small-Instance or Big-Instance) there exists a vertex w1 = (y; ak1 , a2 ) in
GL (w) and hence f (w1 ) ∈ L0 (y). This would mean (c, f (w1 )) ∈ L0 (x, y) because there exists
a path from c to f (w1 ) in L0 (Y ) where Y is a path from x to y in G0 .
3.2 RemoveNotMinority Correctness Proof
The main argument is proving that after RemoveNotMinority function, f still is a homomorphism consistent with the lists and has weak nu property (Lemma 3.4). Moreover, after
RemoveNotMinority there still exists a homomorphism from G to H if there was one before
RemoveNotMinority (Lemma 3.5).
Lemma 3.4 If f is a homomorphism of G × H k → H, consistent with L and with weak nu
property before RemoveNotMinority then the modified f remains a homomorphism consistent
with L and with weak nu property afterwards.
Lemma 3.5 If there is a homomorphism g : G → H then there is a homomorphism from
G to H after RemoveNotMinority.
3.2.1 Proof of Lemma 3.4
It is enough to show the following :
f is a homomorphism of G × H k → H consistent with the lists (with weak nu
property) after removing a from L(x) in the while loop in Algorithm 4.
We need to address items 1,2,3 below.
1. The weak nu property is preserved.
2. The Running time of Remove-NM function is O(|G|3 |H|k+1 ).
18
3. The adjacency property is preserved: for an arbitrary arc yz ∈ A(G) (zy ∈ A(G))
and for every a01 , a02 , . . . , a0k ∈ L(y) and b01 , b02 , . . . , b0k ∈ L(z) where a0i b0i ∈ A(H) (b0i a0i ∈
A(H)), 1 ≤ i ≤ k, we have f (y; a01 , a02 , . . . , a0k )f (z; b01 , b02 , . . . , b0k ) ∈ A(H)
(f (z; b01 , b02 , . . . , b0k )f (y; a01 , a02 , . . . , a0k ) ∈ A(H)).
Proof of 1 : Since in the Update-f function we change the value f (y; a01 , a02 , ..., a0k ) from
a2 to f (y; ak1 , a2 ) for every k-tuple a01 , a02 , . . . , a0k ∈ L(y), we change f (y; b1 , b1 , . . . , b1 , b2 ) =
f (y; b1 , b1 , . . . , b2 , b1 ) = · · · = f (y; b2 , b1 , . . . , b1 ) to the same value. Therefore f still has the
weak k-nu property.
Proof of 2 : In the Algorithm 4 we consider pairs (x, a) ∈ V (G × H) where a ∈ L(x) and
∃w2 ∈ V (G × H k ) with f (w2 ) = a.
In the Update-F function for every y ∈ V (G0 ) and every a2 ∈ L(y), such that (y; ak1 , a2 ) ∈
GL (w), the f value of w1 with f (w1 ) = a2 is going to change to some new value. Once
the f value of some k-tuple in L(y) changed from a2 to something else, there would be no
k-tuple in L(y) that its value is set to a2 in the further steps of the Algorithm. Moreover,
for every (a, e) ∈ L(x, y) the value of f for w1 , w2 with f (w1 ) = a and f (w2 ) = e would
change simultaneously. In other words, if (y; ak1 , a2 ) is in GL (w) and (y; bk1 , a2 ) ∈ GL (w1 ) for
some w1 = (z; dk , e) then there exists some w10 = (z; d0k , e) ∈ GL (w). This implies that in
the main while loop we do not consider any w100 = (z; d00k , e) because there exists no k-tuple
b01 , b02 , . . . , b0k in L(z) such that f (z; b01 , . . . , b0k ) = e. This means that the vertices of GL are
partitioned and when we solve instance G0 , H, L0 , f 0 recursively we deal with one of these
partitions.
Considering all these partitions together, in the worst case we end up processing all the
vertices in G − x together with their lists to be able to process a ∈ L(x). At each step
one element (y, aj ), aj ∈ L(y), y ∈ V (G) is processed (aj is removed) to be able to process
a ∈ L(x); remove a from L(x). The actual work is writing down some new lists (each write
down takes O(|A(G)||H k |)). The overall size of the new lists is one smaller than the previous
lists. So the overall running time is O(|G||A(G)||H|k+1 ). Note that |A(G)| ∈ O(|G|2 ). By
considering the running time of the PreProcessing we conclude that the overall running time
is O(|G|3 |H|k+1 ).
Proof of 3 : Let ak+1 = f (y; a01 , a02 , . . . , a0k ) and bk+1 = f (z; b01 , b02 , . . . , b0k ). We need to show
ak+1 bk+1 ∈ A(H) when (y; a01 , . . . , a0k )(z; b01 , b02 , . . . , b0k ) is an arc of G × H k . In order to show
that we first prove the following claims. In Claims 3.6,3.7,3.8 at each step of the recursive call
we deal with some vertex x of G and two vertices a, b ∈ L(x) such that f (x; bk , a) = c 6= a,
and f (x; c, c, . . . , c) = c considered in Algorithm 4 (line 3). Moreover, the assumption is
that f is a homomorphism consistent with the lists L and if f (y; a01 , a02 , . . . , a0k ) = a1 then
f (y; a1 , a1 , . . . , a1 ) = a1 .
Claim 3.6 Let X : y = y, y1 , y2 , . . . , yn , x be a walk from a vertex y to vertex x in G.
Let C1 : c0 , c1 , c2 , . . . , cn , c and C2 : c00 , c01 , . . . , c0n , c be two oriented walks congruent to X
19
in L(X). Then f (y; ck0 , c00 ), f (y1 ; ck1 , c01 ), . . . , f (yi ; cki , c0i ), . . . , f (yn ; ckn , c0n ), f (x; c, c, . . . , c) is a
walk in L(X) from c0 = f (y; ck0 , c1 ) to c and congruent to X.
Proof: Since f is a homomorphism consistent with the lists, for every 0 ≤ i ≤ n − 1 we
have f (yi ; cki , c0i )f (yi+1 ; cki+1 , c0i+1 ) ∈ A(H) when yi yi+1 ∈ A(G), ci ci+1 , c0i c0i+1 ∈ A(H), and
f (yi+1 ; cki+1 , c0i+1 )f (yi ; cki , c0i ) ∈ A(H) when yi+1 yi ∈ A(G), ci+1 ci , c0i+1 c0i ∈ A(H).
The following claim will be used to show that we can extend the update to f from the
Small-Instance to the Big-Instance.
Claim 3.7 Let W1 , W2 be two cycles from w = (x; bk , a) to w = (x; bk , a) in GL (w) with the
following conditions :
• W1 , W2 are congruent and both follow oriented cycle X : x, x1 , x2 , . . . , xn , x in G.
• the projection of W1 , W2 on the (k + 1)th coordinate yield the same closed walk B :
a, a1 , a2 , . . . , ai , . . . , an , a in L(X).
Let B 0 : a, a1 , a2 , . . . , ai , a0i+1 , a0i+2 , . . . , a0n , a0 be a walk in L(X) congruent with B such that
(a, a0i+1 ) 6∈ L(x, xi+1 ). Let ei = f (wi ) and e0i = f (wi0 ) where wi and wi0 are the ith vertices of
W1 , W2 respectively. Then there exist walks C1 , C2 from ei , e0i to c0 in L0 (X) (constructed in
Big-Instance) that are congruent to B 0 [ai , a0 ] and there exists a walk C from f (xi ; eki , e0i ) to
c0 = f (x; bk , a0 ) in L0 (X) and congruent to B 0 [ai , a0 ].
Proof: Let W3 be the quasi-path from w to w0 = (x; bk , a0 ) and congruent to W1 that follows
X, B 0 and such that the j th vertex of W3 is of form (y; dkj , a0j ), where (y; dkj , aj ) ∈ W1 and
a0j is the j-th vertex of B 0 . Let W4 be the quasi-path from w to w0 = (x; bk , a0 ) and con0
gruent to W2 that follows X, B 0 such that the j th vertex of W4 is of form (y; d0k
j , aj ), where
0
th
(y; d0k
vertex of B 0 . Let f (vi ) = ei and f (ui ) = e0i where vi is
j , aj ) ∈ W2 and aj is the j
the ith vertex of W3 and ui is the ith vertex of W4 . Now C1 : f (vi ), f (vi+1 ), . . . , f (vn ), c0
is a walk in L0 (X) and congruent to B 0 [ai , a0 ]. This is because we add f (vj ), i ≤ j ≤ n
(and c0 ) into L0 (X) in the construction of the Big-Instance, and L0 (X) is closed under f .
Similarly C2 : f (ui ), f (ui+1 ), . . . , f (un ), c0 is a walk in L0 (X) and congruent to B 0 [ai , a0 ]. Now
according to Claim 3.6 the walk C : f (xi ; eki , e0i ), f (xi+1 ; eki+1 , e0i+1 ), . . . , f (xn ; ekn , en ), c0 where
ej ∈ C1 and e0j ∈ C2 , i ≤ j ≤ n, is from f (xi ; eki , e0i ) to c0 and it is congruent to B 0 [ai , a0 ].
Note that C is in L0 (X[xi , xn ], x) ⊂ L(X).
Claim 3.8 Let X : x, x1 , x2 , . . . , xn , x be an oriented cycle in G containing vertex x. Suppose
f (x; bk , a0 ) = c0 . Let B : a, a1 , a2 , . . . , ai , . . . , an , a and B 0 : a, a1 , a2 , . . . , ai , a0i+1 , a0i+2 , . . . , a0n , a0
be two congruent walks in L(X) such that (a, a0i+1 ) 6∈ L(x, xi+1 ). Let t ≥ 1 be an integer such
that ci ∈ Plt (xi , ai ) (Pl constructed during the execution of Remove-NM). Then there exists
a walk in L0 (X) from ci to c0 congruent with B 0 [ai , a0 ].
20
c1
c2
c3
c4
c0
e01
e02
e2
e03
e3
e04
e4
e1
c
a1
a
d01
a2
d02
d1
b
d01
d2
d02
a03
a04
a0
a3
a4
a
d03
d04
d3
d4
d03
d04
b
b
d1
d2
d3
d4
b
x
y
z
u
v
x
f (y; d1 , d1 , a1 ) = e1
f (y; d01 , d01 , a1 ) = e01
f (u; d3 , d3 , a03 ) = e3
f (u; d03 , d03 , a03 ) = e03
f (u; e3 , e3 , e03 ) = c3
Figure 3: An illustration for Claim 3.7.
Proof: Let W1 , W2 be two cycles from w = (x; bk , a) to w = (x; bk , a) in GL (w) that both
follow B and such that f (wi ), f (wi0 ) ∈ Pl (xi , ai ) where wi is the ith vertex of W1 and wi0 is
the ith vertex of W2 . According to Claim 3.7, there exist walks C1 , C2 in L0 (X) from c1 , c2 to
c0 where c1 , c2 ∈ Pl (y, ai ) and there exists a walk C 0 in L0 (X) from c3 = f (y; ck1 , c2 ) to c0 and
congruent with B 0 [ai , a0 ]. Now according to the Remove-NM in the recursive call we place
c3 into L0 (y) and also into Pl (y, c2 ). By applying Claim 3.6 on c3 , c1 we conclude that there
exists a walk from f (y; ck1 , c3 ) to c0 and congruent to C1 if the recursive call considers vertex
(y; ck1 , c3 ). We note that since L0 (y) is closed under f , f (y; ck1 , c3 ) ∈ L0 (y).
Note that in this case Pl2 (y, ai ) would contain c3 . Similarly we would have a walk from
f (y; ck3 , c1 ) to c0 congruent to C1 if the recursive call considers vertex (y; ck3 , c1 ). By repeatedly
applying Claim 3.6 we conclude that there exists a walk in L0 (X) from ci to c0 and congruent
to B 0 [ai , a0 ]
Claim 3.9 Consider the current lists L at some stage of Remove-NM algorithm. For every
(y, a2 ) where (y; ak1 , a2 ) ∈ GL (w) there exists d ∈ L0 (y) ∩ Plt (y, a2 ) for some integer 1 ≤ t ≤
|L(y)| such that Pl (y, d) = {d}.
0
Proof: Let a01 ∈ Pl (y, a2 ). Suppose there exists a02 ∈ Pl (y, a2 ) such that c1 = f (y; a0k
1 , a2 ) 6=
0
0
0
a2 . Now in the recursive call Pl (y, a2 ) is updated (we write down some vertices of L into
Pl (y, a02 )) and in particular c1 ∈ Pl (y, a02 ) and by definition c1 ∈ Pl2 (y, a2 ). To see that, L0 is
a subset of L where every a02 ∈ L0 (y) L0 is on a walk C 0 from c inside L0 (Y ). Here Y is a
walk in G that some path W in GL (w) from w to (y; ak1 , a02 ) follows Y, C 0 . Now according to
construction of Small-Instance and Big-Instance with respect to lists L0 , c1 ∈ Pl (y, a02 ).
21
Since at line 6 and 10 of the Remove-NM we set Pl (y, d) = {d}, and because at each step
of the recursion one element from L0 (y) will disappear from the rage of f 0 = f |L0 , after some
t ≤ |L| steps of the nested recursive calls for all d, e ∈ Lt , f (t) (y; dk , e) = e. Here Lt are
the lists passed to the Remove-NM after t steps of recursions, and f t is the corresponding
homomorphism (from G × H k to H) to Lt . Note that all the information that are needed in
the Update-f function can be obtained by looking at the current values of the Pl lists.
From Small-Instance to Big-Instance We need to argue that in the Big-Instance the
modification to f would be consistent with the update to f after calling the Remove-NM
on Small-Instance. In the function Update-f(L,Pl ,f, x, a) after Small-Instance we follow an
oriented path inside L(X) where X is an oriented cycle in G containing x ∈ V (G). We
modify the f value for (y; a01 , a02 , . . . , a0k ) from a2 to some element d2 ∈ Plt (y, a2 ) where y ∈ X
and (y, a2 , d2 ) is reachable from (x, a, d) in GPl (x, a, d) (t and d2 exist according to Claim
3.9). Now consider a vertex b2 in L(z) where yz ∈ A(X) but (a, b2 ) 6∈ L(x, z). In the
Big-Instance we consider an oriented path B 0 : a2 , b2 , . . . , a0 such that
• B 0 is congruent to yX[z, x] (portion of X from y to x) in L(X)
• a directed path W from (y; ak1 , a2 ) to (x; bk , a0 ) follows B 0 .
By Claim 3.8, there exists a walk C 0 from d2 ∈ Plt (y, a2 ) to c0 in L0 (X) which is congruent to
B 0 . This would imply that if we fix vertex c0 in L0 (x) in the recursive call, then corresponding
to B 0 in L(X) we would have walk C 0 in L0 (X) and the f value along B 0 would change to
elements in C 0 . Note that we may change the f value for some k-tuples in x from c0 ∈ L(x).
To be more precise, we would have the elements in C 0 to be considered for further changes
to f . Therefore the updates in f in the Big-Instance can be build upon the updated f on
the Small-Instance.
Why the adjacency preserved : Suppose the value of f (y; a01 , a02 , . . . , a0k ) changed from
a2 to c2 ∈ Plt (y, a2 ). Let b2 = f (z; b01 , b02 , . . . , b0k ). We note that a2 b2 ∈ A(H) because f was
initially a homomorphism consistent with the lists. Since f (y; b01 , b02 , . . . , a0k ) changed from
a2 to c2 , we have (y, a2 , c2 ) ∈ GPl (x, a, d).
There exists a1 ∈ L(y) such that (b, a1 ) ∈ L(x, y) (according to Lemma 3.3 at each step
of the recursion) and hence (y; ak1 , a2 ) ∈ GL (w). Let c2 ∈ Plt (y, a2 ) (c2 , t exist by Claim
3.9 ) and let c02 ∈ Pl (y, a2 ). Now there exists a vertex b1 ∈ L(z) such that (b, b1 ) ∈ L(x, z)
and a1 b1 ∈ A(H). According to Remove-NM we add an out-neighbor of c02 into L0 (z) and
hence Pl (z, b2 ) has an out-neighbor of c02 say d02 . By following the recursive calls we would
conclude that there exists some vertex d2 ∈ Plt (z, a2 ) which is an out-neighbor of c2 and
Pl (y, d2 ) = {d2 }. Therefore according to the function Update-f (note that each (y, a2 ) is
visited once in Update-f) we change f (z; b01 , b02 , . . . , b0k ) from b2 to d2 .
22
Closing remark Once we change the value of f (x; a1 , a2 , . . . , ak ) from a to c then potentially we need to modify the value for f (y; b1 , b2 , . . . , bk ) from an out-neighbor of a, say
a0 in L(y) to an out-neighbor of c. As far as the modifying f is concern it would yield the
same result if we start from (x; b, . . . , a, b, . . . , b), a is in the ith coordinate.
3.2.2 Proof of Lemma 3.5
It is enough to show that the following:
If there is a homomorphism g : G → H with g(x) = a then there is a homomorphism from G to H after removing a from L(x) in Algorithm 4 (line 12).
The Update-f function performs a depth first search in GPl (x, a, c), c = f (w, bk , a). For
simplicity we may assume that c ∈ Plt (x, a) with Pl (x, c) = {c}. The argument is similar for
the case when (x, a, d) ∈ GPl . We note that if vertex (y, a2 , c2 ) is visited in Update-f function
first and then after a while a vertex (y; a2 , c02 ) is visited then no changes would apply into f .
This is because there is no k-tuple a01 , a02 , . . . , a0k in L(y) such that f (y; a01 , a02 , . . . , a0k ) = a2 .
Let G3 be the induced sub-digraph of GPl (x, a, c) with the arcs ê = (y, a1 , c1 )(z, b1 , c2 )
such that ê ∈ A(GPl (x, a, c)) and c1 b1 6∈ A(H) when a1 b1 ∈ A(H) and b1 c1 6∈ A(H) when
b1 c1 ∈ A(H). Note that G3 contains (x, a, c).
Let G0 be the induced sub-digraph of G with vertices y such that (y, g(y), c2 ) ∈ G3 is
visited for the first time by the depth first search in function Update-f.
Observation 3.10 Let X be an oriented cycle in G containing vertex x. By construction
of Small-Instance, corresponding to g(X) say B, there exists a closed cycle C containing c
and congruent to B. Here g(X) is the image of X under g.
For every vertex y ∈ G0 set h(y) = c2 where (y, g(y), c2 ) ∈ GPl (x, a, c). For every y ∈
V (G) \ V (G0 ) set h(y) = g(y).
We need to show that h is a homomorphism from G to H. Let zz 0 (z 0 z) be an arc of
G where z ∈ V (G0 ). We only need to consider the case z ∈ V (G0 ). By definition we have
(z, g(z), c2 ) ∈ G3 .
Since zz 0 is an arc of G and g is a homomorphism from G to H, g(z)g(z 0 ) ∈ A(H).
If h(z)g(z 0 ) ∈ A(H) then we are done because in this case we have z 0 6∈ G0 . Suppose
h(z)g(z 0 ) 6∈ A(H). Now in the Small-Instance there exist vertices w1 = (z; ak1 , g(z)) and
w2 = (z 0 ; bk1 , g(z 0 )) that are in GrL (w). This means there exists a vertex c02 in L0 (z 0 ) that is an
out-neighbor of c2 and hence by definition (z 0 , g(z 0 ), c02 ) ∈ G3 . This would imply z 0 ∈ V (G0 )
and hence h(z 0 ) = c02 , implying that h(z)h(z 0 ) ∈ A(H).
Note that in the special case when z 0 = x we have g(z 0 ) = a and by the Observation above
we have h(z 0 ) = c. Therefore h is a homomorphism from G to H with h(x) = c.
23
3.3 Proof of Theorem 1.1
By Lemma 3.5 we preserve the existence of a homomorphism from G to H after Algorithm
4. By Lemma 3.4 f is still a homomorphism from G × H k → H consistent with the lists L
of G after Algorithm 4. Now we can apply Theorem 5.4. We observe that the running time
of PreProcessing function is O(|G|3 |H|2 ).
According to the proof of Lemma 3.4 (2) the running time of Algorithm 4 is O(|G|3 |H|k+1 ).
The running time of Algorithm 2.2 (O(|G||A(G)||H|k+1 ). Therefore the running time of the
Algorithm 1 is O(|G|3 |H|k+1 ).
4 Additional Algorithms and Proofs (New Minority
Algorithm)
This section collects additional statements and proofs mainly related to new Minority algorithm. We develops a direct algorithm to handle the minority case in our main algorithm.
Of course, it would be easier to appeal to Bulatov-Dalmau Maltsev result as explained in
Subsection 2.2.
A direct proof for Theorem 5.4 Let M 3 be the digraph of triple (y, c, d), x ∈ G, c, d ∈
L(y). There is an arc from (y, c, d) to (y 0 , c0 , d0 ) where yy 0 ∈ A(G), cc0 , dd0 ∈ A(H) and
cd0 6∈ A(H) or y 0 y ∈ A(G) and c0 c, d0 d ∈ A(H), d0 c 6∈ A(H). Note that M 3 is a graph but we
view M 3 as digraph and when we talk about a path in M 3 we mean an oriented path that
reflect the direction of the edge xy when (y, c, d)(y 0 , c0 , d0 ) is an arc of M 3 . Let Gx (a, b) be
the strong component of M 3 containing (x, a, b) such that for each vertex (y, c, d) ∈ Gx (a, b),
(a, c) ∈ L(x, y) and (b, d) ∈ L(x, y). We say a strong component Gx (a, b) of M 3 is invertible
if both (x, a, b) and (x, b, a) are in Gx (a, b).
Again the idea is similar to the one handling the RemoveNotMinority case. At each step we
consider a vertex x of G and two vertices a, b ∈ L(x) and try to eliminate one of the a, b from
L(x). To decide whether remove a or b we construct an instance of the problem say (G0 , H, L0 )
and solve this instance recursively. Based on the existence of a L0 -list homomorphism from
G0 to H we decide to remove a or b. Depending on Gx (a, b) being invertible or not two
different instances for the sub-problem are constructed. At the end we have singleton lists
and if there is a homomorphism from G to H with the singleton lists then success otherwise
we report there is no homomorphism from G to H. We denote the underline graph of digraph
G by U N (G). If Gx (a, b) is invertible we construct a new instance G0 , H, L0 and to construct
the lists L0 we use the Maltsev property.
Lemma 4.1 Let X be an oriented path in G and let B, C, D be three walks in L(X) all
congruent to X where B is from a to c and C is from b to c and D is from b to D. Then
there exists a walk E from a to d in L(X) that is congruent with X.
24
Algorithm 5 RemoveMinority – Using Matlsev Operations
1: function RemoveMinority(G, H, L)
2:
Input: x, a, b such that f (x; bk |i←a ) = a
3:
Define Maltsev consistent homomorphism h : G × H 3 → H where
h(x; a, b, c) = f (x; a, b, b, . . . , b, c) for a, b, c ∈ L(x).
4:
while ∃ x ∈ V (G) with |L(x)| ≥ 2 do
5:
Matlsev(G, H, L )
6:
if ∃ a list homomorphism from G to H then return True
7:
else return False.
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
24:
25:
26:
function Maltsev(G, H, L)
Let a, b ∈ L(x) be two distinct vertices. Construct Gx (a, b).
if Gx (a, b) is not invertible then
Set L0 (x) = a and L0 (y) = ∅ for every x 6= y.
Set G00 = ∅, and let G0 be an induced sub-digraph of G constructed as below.
for all y ∈ G s.t. ∃ a path from (x, a, b) to (y, c, d) in Gx (a, b) do
add y to G0 and c to L0 (y),
let yy 0 an arc of G0 when (y, c, d)(y 0 , c0 , d0 ) ∈ Gx (a, b).
if Gx (a, b) is invertible then
Set L0 (x) = a and L0 (y) = ∅ for every x 6= y.
Let G00 be the set of vertices y ∈ V (G) where (y, c, d, e) lies on an oriented path
P from (x, a, b, b) to (x, b, b, a) in U N (G × H 3 ) s.t. no intermediate vertex in P
is (x, a0 , b0 , c0 ) for a0 , b0 , c0 ∈ L(x)
for all arc (y, c, d, e)(y 0 , c0 , d0 , e0 ) ∈ P do
add h(y, c, d, e) to L0 (y), h(y 0 , c0 , d0 , e0 ) to L0 (y 0 ), and add arc yy 0 (y 0 y) to G00
Let G0 be the induced sub-digraph of G with vertices
y ∈ G \ G00 and arcs yy 0 where (y, c, d)(y 0 , c0 , d0 ) ∈ Gx (a, b) and add c into L0 (y)
when (y, c, d) ∈ Gx (a, b)
if Matlsev(G0 ∪ G00 , H, L0 ) then
remove b from L(x)
else remove a from L(x).
25
Proof: By following B, C, D on the vertices in X and applying the definition of polymorphism h, we conclude that E exists.
Lemma 5.10 implies the following corollary.
Corollary 4.2 If (a, c) ∈ L(x, y) and (b, c), (b, d) ∈ L(x, y) then (a, d) ∈ L(x, y).
Lemma 4.3 The Algorithm 9 runs in O(|A(G)||H|k+1 ). Moreover, if there is a homomorphism g from G to H with g(x) ∈ {a, b} then there is a homomorphism from G to H after
removing a or b from L(x) according to Maltsev-(G, H, L, h).
Proof: We spend O(|G||H|k+1 ) to define h. For every pair a, b ∈ L(x) the algorithm construct an instance of the problem which takes O(|A(G)||A(H)|3 ). The algorithm is recursive
and the depth of the recursion in |G||H|. Now by similar analysis as in Algorithm 4 we
conclude that the running time of function Maltsev(G, H, L) is O(|G||A(G)||H|4 ) (replace
|A(H)| by |H|2 ). Therefore the running time of the Algorithm 2.2 is O(|A(G)||H|k+1 ).
First consider the case that Gx (a, b) is not invertible. The Algorithm 9 is recursive
but at each recursive call the size of the input decreases by at least one. This is because we
do not add b into L0 (x) and hence we have an instance (G0 , L0 , H) of the problem in which at
least one vertex has a smaller size list. We also note that once we make a decision to remove
a vertex from a list the decision is not changed. Therefore the overall procedure in this
case is polynomial (assuming in each call Gx (a, b) is not invertible). According to Maltsev(G, H, L, h) if G0 does not admit a homomorphism to H then there is no homomorphism
from G to H that maps x to a since G0 is a sub-digraph of G.
No suppose b is removed according to Maltsev-(G, H, L, h) and there exists a homomorphism g from G to H with g(x) = b. Let ψ be the homomorphism from G0 to H. Note that
ψ(x) = a.
Define G1 be a sub-digraph of G0 consists of the vertices y such that (y, ψ(y), g(y)) is
reachable from (x, a, b) in Gx (a, b).
Now for every vertex y ∈ G1 set φ(x) = ψ(x) and for every y ∈ G \ G1 set φ(y) = g(y).
Now let zz 0 be an arc of G. If none of the z, z 0 ∈ G1 then clearly since g is a homomorphism, φ(z)φ(z 0 ) ∈ A(H). If both z, z 0 ∈ G1 then again since ψ is a homomorphism, we have
φ(z)φ(z 0 ) ∈ A(H). Suppose z ∈ G1 and z 0 6∈ G1 . Since z ∈ G1 there exists a path in Gx (a, b)
from (x, a, b) to (z, ψ(z), g(z)). Now if ψ(z)g(z 0 ) is an arc then we are done. Otherwise
(z 0 , c, g(z 0 )) ∈ Gx (a, b) where c ∈ L(z 0 ) and ψ(z)c ∈ A(H) (note that since ψ(z) is in L(z) it
must have an out-neighbor in L(z 0 )). This would mean z 0 ∈ G1 , a contradiction. Note that
when z 0 = x and z ∈ G1 then because (ψ(z), a) ∈ L(z, x) we have ψ(z)a ∈ A(H).
Second consider the case that Gx (a, b) is invertible. Again as we argued in the
previous case the algorithm is recursive but at each recursive call the size of the input
decreases and once a decision made (removing a or b ) it won’t change.
26
Observe that if there exists a homomorphism g that maps x to a then for every closed
walk X from x to x, the image of g(X) is a closed walk from a to a in H. Since there is
a walk BB from b to b in L(X) congruent and since (a, b) ∈ L(x, x), there is a walk BA
in L(X) congruent to X. Now by Lemma 5.10 there is a walk AB from a to b in L(X)
and congruent with X. These would imply that there exists a path P in U (G × H 3 ) from
(x, b, b, a) to (x, a, b, b). Now according to the definition of G00 for every walk BB in L(X)
from b to b congruent with X we keep a walk AA in L0 (X) congruent with X and is obtained
by adding h(y, c1 , c2 , c3 ) into L0 (y) where (y, c1 , c2 , c3 ) is in P .
This would mean that in the list of L0 (X), corresponding to ψ(X) (ψ is a homomorphism
from G to H where ψ(x) = a if there exists one) we have a path from a to a in L0 (X).
Therefore if there exists no homomorphism from G0 ∪ G00 → H that maps x to a then there
is no homomorphism from G to H that maps x to a.
Now suppose b is removed according to Maltsev-(G, H, L, h) and there exists a homomorphism g from G to H with g(x) = b. Let ψ1 be the homomorphism from G0 to H and ψ2 be
a homomorphism from G00 to H. Note that ψ1 (x) = ψ2 (x) = a. First we show that there is
no arc e (forward or backward) from a vertex y in G0 to a vertex in G00 . If this is the case
then there is a walk Q from x to y and there is a walk Q0 from y to x. Now since yz is an
arc then Qee−1 Q0 is a closed walk from x to x and hence we should have added z into G00 .
Note that G00 consists of all the induced oriented cycles including x and also all the paths
reaching out of vertices of these cycles except x.
Define G1 be a sub-digraph of G0 consists of vertices y such that (y, ψ1 (y), g(y)) is reachable
from (x, a, b) in Gx (a, b).
Now for every vertex y ∈ V (G1 ) set φ1 (y) = ψ(y) and for every y ∈ V (G) \ V (G1 )
set φ1 (y) = g(y). Define G2 be a sub-digraph of G00 consisting of vertices y such that
(y, ψ2 (y), g(y)) is reachable from (x, a, b) in Gx (a, b).
Now for every vertex y ∈ V (G2 ) set φ2 (y) = ψ2 (y) and for every y ∈ V (G) \ V (G2 )
set φ2 (y) = g(y). And finally let φ(y) = φ1 (y) when y ∈ V (G0 ) and φ(y) = φ2 (y) when
y ∈ V (G00 ). Now let zz 0 be an arc of G. We need to verify that φ(z)φ(z 0 ) is an arc of H.
If none of the z, z 0 ∈ V (G1 ∪ G2 ) then clearly since g is a homomorphism, φ(z)φ(z 0 ) ∈
A(H). If both z, z 0 ∈ V (Gi ) (i = 1, 2) then again since ψi is a homomorphism, we have
φ(z)φ(z 0 ) ∈ A(H). Suppose z ∈ V (G2 ) and z 0 6∈ V (G2 ). Since z ∈ V (G2 ), there exists a
closed walk ZZ 0 that contains z where Z is a path from x to z and Z 0 is a walk from z
to x. Note that ψ2 (ZZ 0 ) is a walk from a to a in L(ZZ 0 ) and g(ZZ 0 ) is a walk from b to
b in L(ZZ 0 ) and since (b, a) ∈ L(x × x) there exists a walk BA in L(ZZ 0 ) congruent with
ZZ 0 . These would imply that corresponding to ZZ 0 there exists a path P from (x, b, b, a) to
(x, a, b, b) containing vertex (z, ψ2 (z), d, g(z)). Now since g(z) ∈ L(z) there exists a vertex
c ∈ L(z 0 ) such that g(z)c ∈ A(H) and there exists e0 ∈ L(z 0 ) such that dd0 ∈ A(H).
Now we add the two arcs (z, ψ2 (z), d, g(z))(z 0 , c, d0 , g(z 0 )), (z 0 , c, d0 , g(z 0 ))(z, ψ2 (z), d, g(z))
into P and hence we obtain a path P 0 from (x, b, b, a) to (x, a, b, b) that goes through vertex
(z 0 , c, d0 , g(z 0 )) and hence z 0 ∈ G00 . Now since φ2 (z)g(z 0 ) 6∈ A(H), by definition z 0 ∈ V (G2 ), a
contradiction.
27
As we argue before one can show that φ1 is also a homomorphism from G to H. Therefore
φ is a homomorphism from G to H with φ(x) = a.
5 Second Approach (none recursive algorithm)
5.1 An introduction to the Algorithm
In this section we introduce the main parts of our algorithm. We encourage the reader to
take time to read and internalize as much of this section as possible, while consulting the
figures that are referenced.
1. We associate to each vertex x ∈ V (G) a list L(x), with each L(x) initially V (H). We
also consider pair lists L(x, y), where (a, b) ∈ L(x, y) means that it may be possible for
x and y to simultaneously map to a and b (respectively) for the same homomorphism.
2. We let the polymorphism on H be specialized for each vertex in G. We define a
homomorphism f : G × H k → H, i.e. f (x; a1 , a2 , . . . , ak ) for x ∈ V (G) and a1 , ..., ak ∈
V (H). Initially f (x; a1 , ..., ak ) = φ(a1 , ..., ak ) where φ is the weak k-nu polymorphism
given to us. We call f a weak k-nu homomorphism from G to H (see Def 5.1).
3. PreProcessing. We perform standard consistency checks that are done in CSP algorithms to prune the lists L(x) for every x, and pair lists L(x, y) for each x and y.
4. We refine f and work towards building a homomorphism g from G to H. The main
loop in the algorithm picks a vertex x ∈ V (G) and tries to remove one of the vertices
a ∈ V (H) from consideration for being g(x). At any given time, we have lists L(x) ⊆ H
for each x ∈ V (G) that have the vertices in H that are being considered for g(x).
5. For a, b ∈ L(x), if f (x; b, b, ..., b, a) = a we say this is the minority case. The minority
case is similar to a case of the homomorphism problem that is already solved, namely
the setting where the underlying polymorphism is Maltsev (see subsection 5.4). We
thus have our main loop choose x, b, a where f (x; b, b, ..., b, a) 6= a and attempt to
remove a from consideration. If f (x; b, b, ..., b, a) = a then we leave it alone. This is
the first place the existence of weak k-nu would be essential.
6. The main procedure, then, is to take an x ∈ V (G), special a, b ∈ L(x) such that
f (x; b, b, ..., b, a) 6= a and remove a from consideration – that is for any e1 , e2 , ..., ek ∈
L(x) if f (x; e1 , e2 , ..., ek ) = a then we would set f (x; e1 , e2 , . . . , ek ) = c.
This change to set f (x; e1 , ..., ek ) = c could break the homomorphism property of f (for
k = 3 if (y; a01 , a02 , a03 ) exists with xy ∈ A(G), ei a0i ∈ A(H) and cf (y; a01 , a02 , a03 ) ∈
/ A(H)).
If so, we update the values of f on these neighbors of x. See Figure 4.
28
e3
e2
e1
a03
a02
a01
b03
b02
b01
change
f (...)
c
a4
b4
a
a2
b2
f (x; e1 , e2 , e3 ) = a
f (y; a01 , a02 , a03 ) = a2
f (z; b01 , b02 , b03 ) = b2
f (x; e1 , e2 , e3 ) = c
f (y; a01 , a02 , a03 ) = a4
f (z; b01 , b02 , b03 ) = b4
H
H
b
a1
b1
b
a1
b1
x
y
z
f (x; b, b, a) = c
f (y; a1 , a1 , a2 ) = a4
f (z; b1 , b1 , b2 ) = b4
H
G
Figure 4: The solid lines are arcs, dotted lines are missing arcs, and no line means either could
be true. In the first column, initially f (x; e1 , e2 , e3 ) = a but we set f (x; e1 , e2 , e3 ) = c
where c = f (x; b, b, a). In the second column, initially f (y; a01 , a02 , a03 ) = a3 but we change
it to f (y; a01 , a02 , a03 ) = a4 where a4 = f (y; a1 , a1 , a2 ). The change in the first column
is so we can remove a from L(x); the change in the second column is to preserve the
homomorphism property (this change is only needed if there is a missing arc from c to
a neighbor a2 of a).
7. DFS of updates to f . We begin by changing f values from a to something else (c) on x,
and this forces some changes for f on y’s that are neighbors of x, which in turn forces
some changes for f on neighbors of y. This results in a DFS through G × H k . We can
imagine a subgraph Gf of G × H k where the vertices are the columns in Figure 4 (for
k = 3) and there is a connection between (y; a1 , a1 , a2 ) and (z; b1 , b1 , b2 ) if a change to
f occurs on the DFS.
8. Cycles in DFS and weak k-nu property. The DFS of updates to f should be careful of
difficulties such as cycles.
An almost-cycle is a cycle in the DFS of updates from some vertex (x; b, b, .., b, a) to
itself where the only change in the vertex is the coordinate of the a (see Figure 6).
DFS moves along such cycles and update f .
Since f is a weak k-nu, update to f starting at (x; b, b, .., b, a) would be consistent with
the changes starting at (x; b, b, ...., b, a, b, ..., b), a at ith coordinate.
9. At each step in the DFS, one of the coordinates (e.g., a2 in the second column of
Figure 4) is forced due to a missing arc (the dotted lines in the figures). The remaining
coordinates (e.g., a1 in the second column of Figure 4) are more free, and these are
chosen to ensure that the DFS of updates to f always moves in the direction of a
29
shortest almost-cycle (Figure 6). Different choices for the ”more free” coordinates
result in different paths through Gf because they use different values for updating f .
10. The goal is that the DFS can complete, and we have fixed any places the homomorphism property for f was broken. If we can fix all the problems, then f would still be
a homomorphism after the procedure is done. Then we remove a from L(x). Further,
for the proof of correctness we need to make sure that if we had a correct homomorphism g in mind, then after removing a from L(x) we could change g to remain a
homomorphism with respect to the lists L and set g(x) to something other than a.
Outline We have given most of the main ideas that are part of the algorithm and proofs.
We give definitions to setup the algorithm in Section 5.2, precisely define the algorithm in
Section 5.3, and give the correctness proof in Section 5.5. We encourage the reader to begin
by reading Section 5.3 and consult back to Section 5.2 as needed. We also encourage the
reader to look forward to Section 5.5 to keep in mind the main structure of the proofs.
5.2 Definitions
This section contains definitions that are used in the algorithms in Section 5.3 and the proofs
in Section 5.5.
An oriented walk (path) is obtained from a walk (path) by orienting each of its edges. The
net-length of a walk W , is the number of forward arcs minus the number of backward arcs
following W from the beginning to the end. An oriented cycle is obtained from a cycle by
orienting each of its edges. We say two oriented walks X, Y are congruent if they follow the
same patterns of forward and backward arcs.
Given digraphs G and H, let G × H k be a digraph on the vertices {(y; a1 , a2 , . . . , ak )|y ∈
V (G), ai ∈ V (H), 1 ≤ i ≤ k} with the arcs (y; a1 , a2 , ..., ak )(y 0 ; b1 , b2 , ..., bk ) where yy 0 is an
arc of G and each ai bi , 1 ≤ i ≤ k, is an arc of H. By convention, we shall further restrict the
use of the symbol G×H k to the digraph induced on the vertices {(y; a1 , a2 , . . . , ak )|y ∈ V (G),
ai ∈ L(y), 1 ≤ i ≤ k} where L(y) is the set of vertices in H that are being considered as
images of a homomorphism from G to H.
Definition 5.1 (Homomorphism consistent with Lists) Let G and H be digraphs. For
each x ∈ V (G), let L(x) ⊆ V (H). Let k > 1 be a constant integer.
A function f : G × H k → H is a homomorphism consistent with L if the following hold.
• List property : for every x ∈ V (G) and every a1 , a2 , . . . , ak ∈ L(x), f (x; a1 , a2 , . . . , ak ) ∈
L(x)
• Adjacency property: for every x, y ∈ V (G) and every a1 , ..., ak ∈ L(x), b1 , ..., bk ∈ L(y),
if xy is an arc of G and ai bi is an arc of H for each 1 ≤ i ≤ k then f (x; a1 , ..., ak )f (y; b1 , ..., bk )
is an arc of H.
30
In addition if f has the following property then we say f has the weak nu property.
• for every x ∈ V (G), {a, b} ⊆ L(x), we have f (x; a, b, b, ..., b) = f (x; b, a, b, ..., b) = ... =
f (x; b, b, b, ...a).
We note that this definition is tailored to our purposes and in particular differs from the
standard definition of weak k-nu as follows.
(a) f is based on two digraphs G and H rather than just H (we think of this as starting
with a traditional weak k-nu on H and then allowing it to vary somewhat for each
x ∈ G),
(b) We do not require that f (y; d, d, d, ..., d) = d (this is not required in our algorithm, and
in fact is more convenient to leave out).
Notation For simplicity let (bk , a) = (b, b, . . . , b, a) be a k-tuple of all b’s but with an a in
the k th coordinate. Let (x; bk , a) be a (k + 1)-tuple of x, (k − 1) b’s and a in the (k + 1)th
coordinate.
Digraph of Updates to f
Let Gf be a digraph where V (Gf ) = V (G × H k ) and with arcs :
A(Gf ) = {(y; ak1 , a2 )(y 0 ; bk1 , b2 ) | yy 0 ∈ A(G), a1 b1 , a2 b2 ∈ A(H), f (y; ak1 , a2 )b2 6∈ A(H)} ∪
{(y; ak1 , a2 )(y 0 ; bk1 , b2 ) | y 0 y ∈ A(G), b1 a1 , b2 a2 ∈ A(H), b2 f (y; ak1 , a2 ) 6∈ A(H)}.
When (y; ak1 , a2 )(y 0 ; bk1 , b2 ) is an arc of Gf then we say (y; ak1 , a2 ) avoids (y 0 ; bk1 , b2 ) in coordinate k, or simply say the avoidance appears in the k th coordinate. In Figure 4, (z; b1 , b1 , b2 )
is an out-neighbor of (y; b1 , b1 , b2 ) in Gf with avoidance in the third coordinate.
Note that a directed path in Gf would be an oriented path in G × H k . Let B be a walk in
H starting at some vertex a ∈ V (H). We say a directed path W in Gf (with avoidance in
coordinate k) follows B if B is a walk in H induced by (k + 1)th coordinates of the vertices
in W (see Figure 5).
We say directed path W in Gf follows walk X in G when X is the walk induced by the
first coordinates of the vertices in W .
For a fixed vertex x ∈ V (G) and fixed vertex b ∈ V (H), let Gx,b be a digraph of vertices
a0 ∈ L(x) where f (x; bk , a0 ) 6= a0 . The arcs of Gx,b are of form a0 a00 where there exists a
directed path in Gf from w0 = (x; bk , a0 ) to w00 = (x; bk , a00 ).
Definition 5.2 Let w = (x; bk , a) and w1 = (x; bk , a0 ) with f (w) 6= a and f (w1 ) 6= a0 . We
say directed path W from w to w1 in Gf is quasi-cycle when vertex (x; dk , d1 ) is in W then
d = b. If a0 = a then we say W is an almost-cycle.
31
c
a4
b4
c4
a
a2
b2
c2
b
a1
b1
c1
b
a1
b1
c1
x
y
z
u
f (x; b, b, a) = c
f (y; a1 , a1 , a2 ) = a4
f (z; b1 , b1 , b2 ) = b4
f (u; c1 , c1 , c2 ) = c4
A path from (x; b, b, a) to (u; c1 , c1 , c2 )
Figure 5: An illustration of the graph Gf , with each column representing a vertex in Gf . Solid
lines are arcs, dotted lines are missing arc, and no line means either could be true.
(x; b, b, a) is connected to (y; a1 , a1 , a2 ) with avoidance at coordinate 3, which in turn is
connected to (z; b1 , b1 , b2 ) with avoidance at coordinates 3, etc. The figure depicts a path
P in Gf – where the avoidance began on the 3rd coordinate, proceeding by the walk
a, a2 , b2 , c2 in H.
f (x; b, b, a) = c
f (y; b1 , b1 , b2 ) = b4
W :
f (z; c1 , c1 , c2 ) = c4
f (w; d1 , d1 , d2 ) = d4
f (v; i1 , i1 , i2 ) = i4
f (u; `1 , `1 , `2 ) = `4
c
b4
c4
d4
i4
`4
c
a
b2
c2
d2
i2
`2
a
b
b1
c1
d1
i1
`1
b
b
b1
c1
d1
i1
`1
b
x
y
z
w
u
x
(x; b, b, a), (y; b1 , b1 , b2 ), (z; c1 , c1 , c2 ), (w; d1 , d2 , d3 ),
v
(v; i1 , i1 , i2 ),
(u; `1 , `1 , `2 ), (x; b, b, a)
Figure 6: An almost-cycle in Gf . Each column is a vertex in Gf , and dotted lines are missing
arcs.
32
Consider the digraph Gx,b . Let S = Gx,b (a) be a strong component of Gx,b containing a.
We say S is a sink component if there is no arc from an element in S to any other vertex
in Gx,b outside S. We also say S is trivial if it has only one element.
Let w = (x; bk , a) where Gx,b (a) is a sink component in Gx,b . Let Grf (w) (r for restricted)
be the induced sub-digraph of Gf (w) containing vertices w0 = (y; ak1 , a2 ) reachable from w
such that (b, a1 ) ∈ L(x, y).
5.3 Main Procedure
In this subsection we present the main algorithm. The main algorithm is Algorithm 6.
Subroutines that are used by the main algorithm are Algorithms 7, 8, and a known result
discussed in Section 5.4. Algorithm 7 (PreProcessing) simply updates the lists L of vertices
in G based on local edge constraints, and also updates the pair lists of vertices in G (pair
consistency); standard textbook CSP algorithms for the homomorphism problem would repeatedly invoke the PreProcessing routine, and then make a decision (often greedy, or trying
all possible choices that remain).
Algorithm 6 The main algorithm for solving the digraph homomorphism problem.
1: function DigraphHom(G, H, φ)
. G and H digraphs, φ a weak k-nu on H
2:
for all x ∈ G, let L(x) = V (H)
3:
for all x ∈ G and a1 , ..., ak ∈ V (H), let f (x; a1 , ..., ak ) = φ(a1 , ..., ak )
4:
PreProcessing(G, H, L)
5:
while ∃x ∈ V (G), a, b ∈ L(x), a 6= b s.t. f (x; bk , a) 6= a & Gx,b (a) is a sink do
6:
RemoveNotMinority((x; bk , a))
7:
Remove a from L(x)
8:
PreProcessing(G, H, L)
9:
Note: now, for all x ∈ V (G) and a, b ∈ L(x) we know f (x; bk , a) = a
10:
RemoveMinority(G, H, L, f )
11:
if RemoveMinority produces a homomorphism then return true
Algorithm 8 (RemoveNotMinority function) is the key subroutine of the main algorithm.
It starts with w = (x; bk , a) where f (w) = c 6= a and then it starts modifying f by setting
f (x; e1 , e2 , . . . , ek ) = f (w) for every k-tuple e1 , e2 , . . . , ek ∈ L(x) with f (x; e1 , e2 , . . . , ek ) = a.
Now in order to have a homomorphism from G × H k to H consistent with L it performs a
depth first search (in Grf (w) ) to modify f as necessary. After the execution of Algorithm 8
we remove a from L(x) in Algorithm 6. We construct Gf once and then construct Gx,b and
then consider a sink component of Gx,b , to identify the vertex w = (x; bk , a).
After the main loop in Algorithm 6, we end up with a so-called Maltsev or minority instance of the problem – in which we have a homomorphism f consistent with L such that
for every y ∈ V (G) and every c, d ∈ L(y) we have f (y; ck , d) = d. We argue in the next
subsection that such instances can be solved by using the known algorithm of [BD06] (see
33
the remark at the end of Subsection 5.4). The Maltsev/minority instances can also be solved
in a manner similar to our arguments for RemoveNotMinority (see Section 5.7) .
Algorithm 7 Update lists of x, y based on edge constraints and pair constraint
1: function PreProcessing(G, H, L)
2:
Input: digraphs G, H, lists L(x) ⊆ V (H) for each x
3:
ArcConsistency(G, H, L);
4:
PairConsistency(G, H, L)
function ArcConsistency(G, H, L)
update=True
while update do
if ∃xy(yx) ∈ A(G), a ∈ L(x) s.t. @b ∈ L(y) with ab(ba) ∈ A(H) then
remove a from L(x) and set update=True.
else update=False.
11:
if there is an empty list then return no homomprphism
5:
6:
7:
8:
9:
10:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
function PairConsistency(G, H, L)
for every (x, y) ∈ V (G) × V (G) set L(x, y) = {(a, b)|a ∈ L(x), y ∈ L(y)}.
for all xy ∈ A(G), a ∈ L(x), b ∈ L(y) do
if ab 6∈ A(H) then remove (a, b) from L(x, y).
update=True
while update do
if ∃x, y, z s.t. @c ∈ L(z) s.t. (a, c) ∈ L(x, z)&(c, b) ∈ L(z, y) then
remove (a, b) from L(x, y) and set update=True.
else update=False.
if there is an empty list then return no homomorphism
An overview of Algorithm 8 : Algorithm 8 starts with vertex x ∈ V (G) and two
vertices a, b ∈ L(x) such that Gx,b (a) is a sink component. The Algorithm 8 has a depth
first search function so called NM-DFS that modifies f . The advantage of taking such a, b
is that we only need to modify f in Grf (w) where w = (x; bk , a) and this allows us to direct
the NM-DFS along the almost-cycles in Grf (w).
The crucial observation is that we construct Grf (w) once with the initial values. Then we
change f (x; e1 , e2 , . . . , ek ) from a to f (x; bk , a). Now updating f is translated into taking
different directed paths in Grf (w) starting at vertex w. All the step taken by the NM-DFS
along the paths in Grf (w) are guided based on the current values of f .
Suppose at the current step of the NM-DFS we are at vertex w0 = (y; ak1 , a2 ). Suppose we
have changed the image of some (y; a01 , a02 , . . . , a0k ) from a2 to f (w0 ). Note that if this is not
34
the case then there is no need to continue from w0 and it would mean retreat from w0 .
Now we look at each neighbor of y, say z. Then in the list of z we look for an out-neighbor
(in-neighbor) of a2 , say b2 , for which the f value of w1 = (z; b01 , b02 , . . . , b0k ) should be changed
to an out-neighbor of f (w0 ) because the f value of some vertex (y; a01 , . . . , a0k ) was changed
from a2 to f (w0 ). There are two possibilities :
In the first case the f value of w1 is b2 and it should be changed from b2 to an out-neighbor
of f (w0 ) because f (w0 )b2 6∈ A(H) (b2 f (w0 ) 6∈ A(H)).
In the second case, the f value of w1 has already been changed to some vertex d ∈ L(z) at
some earlier step and now this value should be changed to an out-neighbor of f (w0 ) because
f (w0 )b2 , f (w0 )d 6∈ A(H) (b2 f (w0 ), df (w0 ) 6∈ A(H)).
Therefore we find a neighbor of w0 , say w00 = (z; bk1 , b2 ) in Grf (w) to change f (w1 ) to f (w00 )
and take arc w0 w00 ∈ Gfr (w) in the NM-DFS.
Since w0 ∈ Grf (w), we have (b, a1 ) ∈ L(x, y) and hence there exists a vertex b1 ∈ L(z) such
that a1 b1 ∈ A(H), and (b, b1 ) ∈ L(x, z) (because of the PreProcessing).
Definition 5.3 Let w0 = (y; ak1 , a2 ) be a vertex in Grf (w). We say an arc w0 w00 ∈ Grf (w) is
nice if it lies on a shortest directed path (cycle) Q (starting at w) in Grf (w) and if there is a
choice we may assume Q is a shortest almost-cycle, i.e. Q ends at (x; bk , a).
To simplify the proof of the correctness we may direct the NM-DFS to take an w0 w00 when
f (w0 )f (w00 ) is an arc of H. If no such arc w0 w00 exists, then we place w0 w00 into queue Q
(pause) and go back to take different branches. Later NM-DFS comes back and will resume
from w0 w00 . We will explain how such situation occurs. Once NM-DFS retreat at some point
it considers vertex w1 where w1 = (y; dk1 , d2 ) and then take arc w1 w2 where w2 = (z; ek1 , e2 )
and f (w0 ) = d2 and f (w00 ) = e2 . If there is no branch to take NM-DFS removes an arc w0 w00
from Q and it resumes from there.
For simplicity one can assume that DFS does not put anything into queue Q.
In Algorithm 8 the vertices (x; bk |i←a ) and (x; bk |`←a ) of Gf are considered as the same
vertex.
The NM-DFS is used for two purposes.
1. To modify f such that at the end f is a homomorphism consistent with lists L and for
any k-tuple e1 , e2 , . . . , ek , f (x; e1 , e2 , . . . , ek ) 6= a.
2. To show that if there exists a homomorphism g : G → H consistent with lists L with
g(x) = a then there exists a homomorphism h : G → H consistent with lists L and
with h(x) 6= a. We use NM-DFS as a guide to define h.
We will show that in order to repair f we only need to do the NM-DFS inside Grf (w). Let
Gaf (w) (’a’ for almost-cycle) be the sub-digraph of Grf (w) containing vertices w0 such that
w0 lies on an almost cycle containing w. We observe that in order to update f in Gaf (w),
35
e3
e2
e1
a03
a02
a01
b03
b02
b01
d03
d02
d01
d
f (z; b1 , b1 , b2 ) = b4
c
a4
b4
d4
change
a
a2
b2
d2
b
a1
b1
d1
b
a1
b1
d1
x
y
z
u
f (y; a01 , a02 , a03 ) = a2
f (z; b01 , b02 , b03 ) = b2
f (y; a01 , a02 , a03 ) = a4
f (z; b01 , b02 , b03 ) = b4
f (z; b01 , b02 , b03 ) = d
Figure 7: Line 10 from Algorithm 8 where f (z; b01 , b02 , b03 ) changed from b2 to d at an earlier step
in the NM-DFS, and so now is changed to b4 along with other points that map to b2 .
we need to go through the cycles in Gaf (w) and hence we direct NM-DFS to go through the
shortest almost-cycles one by one.
Let Gof (w) (’o’ is for other) be the sub-digraph of Grf (w) containing those vertices that
do not lie on an almost-cycle. Therefore in order to update f in Gof (w) we need to direct
NM-DFS to go through the directed paths in Gof (w) and hence we direct NM-DFS to go
through the shortest paths one by one.
In what follows we argue how things could go wrong if H does not have the weak nu
property.
For contrary suppose w1 = (x; bk , a) with f (w1 ) = c and w2 = (x; a, b, b, . . . , b) with
f (w2 ) = d. If d = a then according to the NM-DFS we try to remove a from L(x) if we start
with w1 while we do need to keep a in L(x) because we later use the Maltsev algorithm. It
might not be the case that d 6= a but as one can see (shown in the proof of Lemma 3.3 (4))
some non-minority pairs becomes minority pairs. In other words, during NM-DFS for some
w3 = (x; bk , e) with f (w3 ) 6= e we may set f (w3 ) to e.
So we need to have f (w1 ) = f (w2 ), the weak nu property, to start. We do mention at
the end of the proof of Lemma 5.7 that it does not matter to start the NM-DFS from w1 or w2 .
In Conclusion : NM-DFS modifies f and the reason we use almost-cycles is because of
(2) above and the need for weak nu property is because modifying f starting at any of the
(x; bk , a), (x; b, b, . . . , b, a, b, . . . , b) (ith coordinate) would keep the property of f .
36
Algorithm 8 Remove a from L(x), updating f so it remains a homomorphism of G × H k
to H consistent with L.
1: function RemoveNotMinority(w = (x; bk , a))
2:
Input: x, a, b such that f (w) = c 6= a and Gx,b (a) is a sink
3:
if ∃ w1 = (x; a01 , a02 , . . . , a0k ) with f (w1 ) = a then
4:
for all w1 = (x; a01 , a02 , . . . , a0k ) with f (w1 ) = a do set f (w1 ) = c
5:
NM-DFS(w = (x; bk , a))
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
function NM-DFS(w0 = (y; ak1 , a2 ))
if w0 already visited then return
do
Let w00 = (z; bk1 , b2 ) 6= (x; bk , a) s.t w0 w00 is a nice arc of Grf (w)
. a2 b2 ∈ A(H), f (w0 )b2 6∈ A(H) & f (y; ...) changed from a2 to f (w0 )
if f (w0 )f (w00 ) 6∈ A(H) then add w0 w00 into queue Q and return
Let d ∈ L(z) s.t. f (z; ...) was changed from b2 to d.
. d = b2 if no change
0
if f (w )d 6∈ A(H) then
for all w2 = (z; ...) if f (w2 ) was changed from b2 to d or f (w2 ) = b2 do
set f (w2 ) = f (w00 )
NM-DFS(w00 )
while ∃ a neighbor of w0 that has non been considered
. It considers the neighbors of w0 and for each b2 only one NM-DFS(w00 ) call
while Q is not empty do remove w0 w00 from Q
if w0 w00 is an arc of Grf (w) then NM-DFS(w00 )
37
5.4 Minority Algorithm (RemoveMinority)
In this section we show that once the minority case has been reached in our main algorithm,
we can reduce to an already solved setting for homomorphism testing – namely that of the
Maltsev case. We note that this section is independent of the rest of the algorithm.
Note that at this point for every a, b ∈ L(x) we have f (x; bk |i←a ) = a and in particular
when a = b we have f (x; a, a, . . . , a) = a (idempotent property). This is because when a is
in L(x) then it means the RemoveNotMinority procedure did not consider a and in fact did
not change the value of f (x; ...) from a to something else. Note that for the argument below
we just need the idempotent property for those vertices that are in L(x), x ∈ V (G).
A ternary polymorphism h0 is called Maltsev if for all a 6= b, h0 (a, b, b) = h0 (b, b, a) = a.
Note that the value of h0 (b, a, b) is unspecified by this definition.
Let G and H be as input to Algorithm 6, and suppose line 9 of the algorithm has been
reached. We define a homomorphism h : G × H 3 → H consistent with the lists L by setting
h(x; a, b, c) = f (x; a, b, b, . . . , b, c) for a, b, c ∈ L(x). Note that because f has the minority
property for all x ∈ G, a, b ∈ L(x), h is a Maltsev homomorphism consistent with the lists
L.
Note that for every a, b ∈ L(x) we have f (x; bk |k←a ) = a and in particular if a = b
we have f (x; a, a, . . . , a) = a (idempotent property). Since a is in L(x), this means the
RemoveNotMinority procedure did not considered a and in fact did not change the value
of f (x; ...) from a to something else. Note that for the argument below we just need the
idempotent property for those vertices that are in L(x), x ∈ V (G).
Let G0 be the structure obtained from G by making each arc a different binary relation.
In other words, G0 has vertices V (G) and |E(G)| binary relations Re , e ∈ E(G), where
Re = {xy} if e is the arc e = xy.
Let H 0 be the structure where V (H 0 ) is the disjoint union of L(x), x ∈ V (G), and there
are also |E(G)| binary relations Se , e ∈ E(G), where Se is the set of all ordered pairs ab with
ab ∈ E(H), a ∈ L(x), b ∈ L(y), where e = xy. Note that |V (H 0 )| ≥ |V (G0 )| if each L(x) is
non-empty. This may seem unusual for the homomorphism setting, but is certainly allowed.
Now note that there is an L-homomorphism of G to H (i.e., a list homomorphism consistent
with lists L) if and only if there is a homomorphism of G0 to H 0 . Homomorphisms of such
structures are mappings f : V (G0 ) → V (H 0 ) such that xy ∈ Re implies f (x)f (y) ∈ Se for all
e ∈ E(G).
Finally, note that the structure H 0 has a Maltsev polymorphism h0 of the ordinary kind.
Indeed, let hx be our Maltsev polymorphisms defined on L(x) by setting hx (a, b, c) =
h(x; a, b, c). We let h0 (a, b, c) = h(x; a, b, c) if a, b, c are from the same L(x), and for a, b, c
not from the same L(x) define h0 (a, b, c) = a unless a = b, in which case define it as
h0 (a, b, c) = c. The definition ensures that h is Maltsev. To check it is a polymorphism, note
that aa0 , bb0 , cc0 ∈ Se is only possible if a, b, c ∈ L(x), a0 , b0 , c0 ∈ L(y), where e = xy. For
those, we have the polymorphism property by assumption.
Now we have a structure with a Maltsev polymorphism, so the Bulatov-Dalmau [BD06]
38
algorithm applies and solves the homomorphism problem. Note that Corollary 4.2 of the
Bulatov-Dalmau paper explicitly mentions that it is polynomial in both the sizes of G and
H.
Therefore we have the following theorem.
Theorem 5.4 Suppose h : G × H k → H is a minority homomorphism consistent with lists
L on G. Then the existence of an L-homomorphism of G to H can be decided in polynomial
time.
Remark : We have communicated with the authors of [BD06] and they confirmed that
indeed we can apply their algorithm as explained above. We note that it is also possible
to give a direct algorithm for the minority case that is similar to how we handle the “not
minority” case.
5.5 Proofs
5.5.1 PreProcessing and List Update
We first show that the standard properties of consistency checking remain true in our setting
– namely, that if the PreProcessing algorithms succeed then f remains a homomorphism
consistent with the lists L if it was before the pre-processing.
Lemma 5.5 If f is a homomorphism of G × H k → H consistent with L then f is a homomorphism consistent with L after running the pre-processing.
Proof: We need to show that if a1 , a2 , . . . , ak are in L(y) after the pre-processing then
f (y; a1 , a2 , . . . , ak ) ∈ L(y) after the pre-processing. By definition vertex a is in L(y) after
the pre-processing because for every oriented path Y (of some length m ) in G from y to a
fixed vertex z ∈ V (G) there is a vertex a0 ∈ L(z) and there exists a walk B in H from a to
a0 and congruent with Y that lies in L(Y ). L(Y ) denote the vertices that are in the list of
the vertices of Y .
Let a01 , a02 , a03 , . . . , a0k ∈ L(z). Let Ai , 1 ≤ i ≤ k be a walk from ai to a0i in L(Y ) and
0
congruent to Y . Let Ai = ai , ai1 , a2i , . . . , am
i , ai and let Y = y, y1 , y2 , . . . , ym , z.
Since f is a homomorphism consistent with L before the pre-processing, f (y; a1 , a2 , . . . , ak ),
m
m
0
0
0
f (y1 ; a11 , a12 , . . . , a1k ), . . . , f (yi ; ai1 , ai2 , . . . , aik ), . . . , f (ym ; am
1 , a2 , . . . , ak ), f (z; a1 , a2 , . . . , ak ) is
a walk congruent with Y . This would imply that there is a walk from f (y; a1 , a2 , . . . , ak ) to
f (z; a01 , a02 , . . . , a0k ) congruent with Y in L(Y ) and hence f (y; a1 , a2 , . . . , ak ) ∈ L(y).
By a similar argument as in the proof of Lemma 5.5 we have the following lemma.
Lemma 5.6 If f is a homomorphism of G×H k → H, consistent with L and a1 , a2 , . . . , ak ∈
L(x), b1 , b2 , . . . , bk ∈ L(y), and (ai , bi ) ∈ L(x, y), 1 ≤ i ≤ k, after pre-processing then
(f (x; a1 , a2 , . . . , ak ), f (y; b1 , b2 , . . . , bk )) ∈ L(x, y) after the pre-processing.
39
5.5.2 RemoveNotMinority Correctness Proof
The main argument is proving that after each step of the RemoveNotMinority function
f still is a homomorphism consistent with the lists and has weak nu property (Lemma
5.7). Moreover, after removing a vertex from the list of a vertex of G there still exists a
homomorphism from G to H if there was one before removing that vertex from the list
(Lemma 5.8).
Lemma 5.7 If f is a homomorphism of G × H k → H, consistent with L with weak nu
property before RemoveNotMinority(w = (x; bk , a)) then the modified f remains a homomrphism consistent with L with weak nu property afterwards. Moreover for every k-tuple
a1 , a2 , . . . , ak ∈ L(x), f (x; a1 , . . . , ak ) 6= a.
Lemma 5.8 If there is a homomorphism g : G → H with g(x) = a then there is a homomorphism from G to H after removing a from L(x) according to RemoveNotMinority((x; bk , a)).
5.5.3 Proof of Lemma 5.7
In order to show that f still is a homomorphism consistent with L with weak nu property,
we need to address items 1,2,3,4 below.
1. The weak nu property is preserved.
2. The RemoveNotMinority function stops in O(|A(G)||H|k+1 ).
3. The adjacency property is preserved: for an arbitrary arc yz ∈ A(G) (zy ∈ A(G))
and for every a01 , a02 , . . . , a0k ∈ L(y) and b01 , b02 , . . . , b0k ∈ L(z) where a0i b0i ∈ A(H) (b0i a0i ∈
A(H)), 1 ≤ i ≤ k, we have f (y; a01 , a02 , . . . , a0k )f (z; b01 , b02 , . . . , b0k ) ∈ A(H)
(f (z; b01 , b02 , . . . , b0k )f (y; a01 , a02 , . . . , a0k ) ∈ A(H)).
4. We argue that we only need to fix the problems caused by changing f (x; ...) from a
to c in Grf (x; bk , a). We show that if f (x; bk , a) = c 6= a and f (x; e1 , e2 , . . . , ek ) = a
then f (x; e1 , e2 , . . . , ek ) is changed to c (in the simple case) or to some other fixed
element (which will be used in the next lemma) not equal to a after the execution of
RemoveNotMinority((x; bk , a)).
Proof of 1 : Since we change the value f (y; a01 , a02 , ..., a0k ) from a2 to f (y; ak1 , a2 ) for every
k-tuple a01 , a02 , . . . , a0k ∈ L(y), we change f (y; b1 , b1 , . . . , b1 , b2 ) = f (y; b1 , b1 , . . . , b2 , b1 ) = · · · =
f (y; b2 , b1 , . . . , b1 ) to the same value. Therefore f still has the weak k-nu property.
Proof of 2 : Observe that during the NM-DFS the value of f (y; a01 , a02 , . . . , a0k ), y 6= x may
change several times. This is because after changing f (y; a01 , a02 , . . . , a0k ) from a2 to some ak+1
there might be the case that there is a path A0 from a to ak+1 in H and some directed path
Q in Grf (w) to some vertex (y; ck1 , ak+1 ).
40
However, the value of f (θ), θ = (x; e1 , e2 , . . . , ek ) is changed once or the value of f (θ) may
change a number of times while going through a circuit (explained in Case 2 in the proof of
(4)) but each time to a new value.
Note that the NM-DFS algorithm may end up traversing all the vertices in Gf but each
vertex at most once. There are at most |G||H k | vertices in Gf . The running time of the
NM-DFS is O(|A(G)|H|2k+1 ). Note that the size of Gf is at most |A(G)|H|k+1 and at each
step of the NM-DFS we may change a value of some k-tuple in a list of a vertex of G and
there are |H|k of them.
Proof of 3 : Let ak+1 = f (y; a01 , a02 , . . . , a0k ) and bk+1 = f (z; b01 , b02 , . . . , b0k ). If ak+1 bk+1 is an
arc of H then we are done. Otherwise without loss of generality we consider the following
two cases.
(a) The value of f (y; a01 , a02 , . . . , a0k ) changed from a2 to ak+1 for the last time and the value
of f (z; b01 , b02 , . . . , b0k ) was not changed.
(b) The value of f (y; a01 , a02 , . . . , a0k ) changed from a2 to ak+1 for the last time and the value
of f (z; b01 , b02 , . . . , b0k ) was changed after changing the value of f (y; a01 , a02 , . . . , a0k ).
Note that ak+1 is the last value of f (y; a01 , a02 , . . . , a0k ) and bk+1 is the last value of f (x; b01 , b02 , . . . , b0k ).
Observe that the original value of f (y; a01 , a02 , . . . , a0k ) may not be a2 and it could have been
changed several times before finally being set to ak+1 . For the sake of simplicity, we may
assume that a2 is the original value of f (y; a01 , a02 , . . . , a0k ). Otherwise we may consider an arc
(y; a001 , a002 , . . . , a00k )(z; b001 , b002 , . . . , b00k ) with f (y; a001 , a002 , . . . , a00k ) = a2 instead of arc
(y; a01 , a02 , . . . , a0k )(z; b01 , b02 , . . . , b0k ) and then we could look back at the earlier step of the NMDFS and apply the same argument as below.
According to the RemoveNotMinority function, there is a directed path W from w =
(x; bk , a) to w0 = (y; ak1 , a2 ) in Grf (w). Here w0 is the vertex visited by NM-DFS function.
Let b2 be the original value of f (z; b01 , b02 , . . . , b0k ). We note that a2 b2 is an arc of H because
initially f is a homomorphism consistent with L.
Proof of (a) : Here bk+1 = b2 . Now since ak+1 b2 6∈ A(H), there exists w00 = (y; bk1 , b2 )
such that w0 w00 ∈ A(Grf (w)). The NM-DFS considers all such arcs w0 w00 ∈ Gfr (w) and by
Claim 5.9 we may assume that w0 w00 is a nice arc (f (w00 ) is an out-neighbor of ak+1 ). Therefore NM-DFS should have changed the value of f (z; b01 , b02 , . . . b0k ) from b2 to f (w00 ) which is
a contradiction.
Proof of (b): Case 1. Suppose ak+1 b2 6∈ A(H).
Observe that there exists w00 = (z; bk1 , b2 ) such that w0 w00 ∈ A(Grf (w)). Recall that the
assumption is : f (z; b01 , b02 , . . . , b0k ) changed after f (y; a01 , a02 , . . . , a0k ). There are two scenarios
:
1. w00 is visited right after w0 .
41
a03
a02
a01
−
−
−
b03
b02
b01
d03
d02
d01
a5
a4 = d
b4
e4
c
a2
f (x; b, b, a) = c
f (y; a1 , a1 , a2 ) = a4
f (z; c1 , c1 , b2 ) = b4
f (u; e1 , e1 , e2 ) = e4
f (y; d1 , d1 , a2 ) = a5
change
b2
e2
a
a1
b
d1
a1
b
x
d1
y
b1
c1
f (y; a01 , a02 , a03 ) = a2
f (z; b01 , b02 , b03 ) = b2
f (u; d01 , d02 , d03 ) = e2
f (y; a01 , a02 , a03 ) = a4
f (z; b01 , b02 , b03 ) = b4
f (u; d01 , d02 , d03 ) = e4
f (y, a01 , a02 , a03 ) = a4
f (y, a01 , a02 , a03 ) = a5
e1
b1
c1
e1
z
u
z
0
Figure 8: Keep in mind for the proof of 3 (b) in the proof of Lemma 5.7
2. We reach w00 by following a directed path R in Grf (w) starting at w0 .
Suppose (1) occurs. This would mean we should have changed the value of f (z; b01 , b02 , . . . , b0k )
as we discussed in the proof of (a). Note that again here we appeal to Claim 5.9 for the nice
arc w0 w00 . We also note that in this case the value of f (z; b01 , b02 , . . . , b0k ) is going to change
for the first time in the NM-DFS.
Suppose (2) occurs. This means the value of f (z; b01 , b02 , . . . , b0k ) did change (from b2 ) following
R. Let Z : y, y1 , . . . , yr , z be the oriented path in G that R follows.
By adding arc yz (zy) we get an oriented cycle in G. Since Z(yz) is a cycle it may be that
the last vertex of R is w00 = (z; bk1 , b2 ) and hence by following the modification of f along the
path R we set f (z; b01 , b02 , . . . , b0k ) = bk+1 . Since f is a homomorphism for the vertices in R (by
Claim 5.9), f (w0 )f (w00 ) is an arc and hence ak+1 bk+1 must be an arc of H, a contradiction.
In other words, the modification is consistent.
Now let us assume that the last vertex of R is not w00 .
Let α = (z; ck1 , b2 ) be the last vertex of R (α 6= w00 ). Now in this case according to line 10
of the NM-DFS (see y as z and z as y in the algorithm, see also Figure 8) we have d = ak+1
and dbk+1 is not an arc of H.
First suppose a2 bk+1 6∈ A(H) (in Figure 8, a2 b4 6∈ A(H)) then according to NM-DFS
we should have considered the vertex α0 = (y; dk1 , a2 ) where αα0 is a nice arc and hence
we should have changed the value of f (y; a01 , a02 , . . . , a0k ) from ak+1 to an in-neighbor of
bk+1 , a contradiction (by Claim 5.9 we consider the nice arc αα0 and f (α)f (α0 ) is an arc
of H). Second suppose a2 bk+1 is an arc of H (in Figure 8, a2 b4 ∈ A(H)). In this case
there exists a nice arc w0 w2 in Grf where w0 = (y; ak1 , a2 ) and w2 = (z; bk1 , bk+1 ) (in Figure
42
8, (y; a1 , a1 , a2 )(z; b1 , b1 , b4 ) ∈ A(Grf (w)). Therefore we should have changed the value of
f (z; b01 , b02 , . . . , b0k ) from bk+1 , a contradiction that was the last value (here again by Claim
5.9 f (w0 )f (w2 ) is an arc of H).
Remark : Note that w00 could be one of the visited vertices but the important thing is
we change the value of some k-tuple in z to f (w00 ) (if necessary) and once the NM-DFS is
called with vertex w00 it returns. What is important to notice is that in the NM-DFS we
consider the arcs of Grf (w) (we visit each arc) and makes the update to f accordingly. Once
the NM-DFS traces back and come back to w00 it starts taking a different branch and it
continues updating f from w00 (the visited) accordingly.
Case 2. ak+1 b2 ∈ A(H). In this case the NM-DFS does not proceed to some vertex w00 = (z; bk1 , b2 ) where a2 b2 ∈ A(H). However, NM-DFS proceeds to some vertex
w1 = (z; dk1 , , b2 ) with f (w1 ) = bk+1 because f (z; b01 , b02 , ..., b0k ) has changed to bk+1 . Now
since ak+1 b2 ∈ A(H) and ak+1 bk+1 6∈ A(H), NM-DFS should have continued considering
some vertex w2 = (y; ek1 , ak+2 ) which is a neighbor of w1 in Grf (w) and should have changed
f (y; a01 , a02 , ..., a0k ) from ak+1 to an in-neighbor of bk+1 . But this is a contradiction because
ak+1 was the last value of f (y; a01 , a02 , ..., a0k ).
Claim 5.9 Suppose the NM-DFS is at vertex w0 = (y; ak1 , a2 ). Let b2 be an out-neighbor
(in-neighbor) of a2 . Then at some stage of the NM-DFS there exists a vertex w00 = (z; bk1 , b2 )
such that w0 w00 is a nice arc and f (w0 )f (w00 ) ∈ A(H).
Proof: Set w00 = (z; bk1 , b2 ) such that w0 w00 ∈ Grf (w). Initially f (w0 )f (w00 ) (f (w00 )f (w0 ) ∈
A(H)) is an arc of H because f is a homomorphism consistent with the lists. For contradiction suppose f (w0 )f (w00 ) is not an arc of H. This means either f (w0 ) did change and f (w00 )
did not change immediately or f (w00 ) did change and f (w0 ) did not change.
Case 1. f (w0 ) did change first. This means there is a directed path Q1 from w to some
α = (y; ck1 , c2 ) in Grf (w) and the NM-DFS went along Q1 and did change f (w0 ) from c2 to
f (α). But according to the rules of NM-DFS we don’t take arc w0 w00 and we will wait till we
take arc λ = (y; ck1 , c2 )(z; dk1 , d2 ) such that f (w00 ) = d2 is an is an out-neighbor (in-neighbor)
of c2 . Note that f (w0 ) is changed to f (α) according to the NM-DFS. If f (α)d2 6∈ A(H) then
the arc λ should have been taken by NM-DFS.
Case 2. f (w00 ) did change and f (w0 ) did not change immediately. This means there is a
path Q1 from w to β = (z; dk1 , d2 ) where f (w00 ) changed from d2 to f (β). Now the NM-DFS
moved along some path Q2 starting at β and reach w0 . At this point if f (w0 )d2 6∈ A(H) and
f (w0 )f (β) 6∈ A(H) then according to line 13 of the Algorithm 8 the NM-DFS should change
f (w00 ) from f (β) to an out-neighbor of f (w0 ). So this case can not happen.
43
c
d1
d2
a
c
a
c
d1
d2
a
b
b
b
b
b
b
b
b
b
b
x
x
x
x
x
Figure 9: Circuit of length 2
Proof of 4 : The NM-DFS starts from vertex w where w = (x; bk , a). The changes to f are
forced in the list of vertices y where w0 = (y; ak1 , a2 ) ∈ Gf , i.e. f (y; . . . ) should be changed
from a2 to something else. This means there exists a path from w to w0 . Since we only need
to change the image of f in y from a2 , we are free to choose a1 ∈ L(y) and hence we may
assume that NM-DFS is performed in Grf (w). We also note that since a is a sink component
of Gx,b (a), any shortest almost-cycle lies entirely in Grf (w).
Now in the remaining we analyze the behavior of NM-DFS depending on the final value of
f (x; e1 , e2 , . . . , ek ) that is initially a. The final value is used in Lemma 5.8. Let C = Gx,b (a)
be sink strong component of Gx,b containing a.
We say vertex a ∈ C is a source if there is no path from a to c in C where f (x; bk , a) = c.
We say the sink component C is invertible if it does not have a source.
Suppose C does not have a source. Now according to the definition of invertible, there is
a path in C starting at a that goes through the vertices c, d1 , d2 , . . . , dt−1 , dt in this order.
Thus there exist corresponding vertices α = (x; bk , a), β = (x; bk , c), λj = (x; bk , dj ), 1 ≤ j ≤ t
and there exists a quasi-cycle Q in Grf from α to λt that goes through β, λ1 , λ2 , . . . , λt in
this order and f (α) = c, f (β) = d1 , f (λ1 ) = d2 , . . . , f (λt−1 ) = dt , f (λt ) = a. If t = 1
then d1 = a and λt+1 = α. We say Q is a circuit in Grf (α) (see Figure 9). Let Q(t) be an
almost-cycle obtained from concatenating Q and a directed path Q0 from λt to α0 = (x; bk , a)
in Grf (α). Note that path Q0 exists because C is a strong component of Gx,b . Note that here
for simplicity we may assume that the last vertex of the circuit is λt with f (λt ) = a. It could
be the case the di plays the role of a.
Observe that if there exists an arc from a to a00 in C then this means we do need to modify
f (x; β10 , β20 , . . . , βk0 ) from a00 to f (w1 ) where w1 = (x; bk , a00 ).
Now we consider two cases depending on whether C is invertible or not.
44
Case 1. C is not invertible.
Note that since a is a source and C is a sink component, there is no arc from a to some
c where f (w) = c, and hence there is no path in Grf (α) from α to (x; bk , c). This means we
do not end up changing a value of f (x; α10 , α20 , . . . , αk0 ) after changing it from a to c and then
changing it from c to something else.
Case 2. C is invertible. We need to show that if there is an arc from a to c in C then
function NM-DFS would handle this situation and it fixes the value of f (x; b01 , b02 , . . . , b0k )
from a to some element not a. The main purpose of this case would be finding that fixed
element.
Since C is invertible, there exists a directed path P in C from a to dt which goes
through the vertices a, c, d1 , d2 , . . . , dt−1 (at this order). Consequently this means there
exist vertices : α, β = (x; bk , c), λi = (x; bk , di ), 1 ≤ i ≤ t and there exists a path Q :
α, . . . , β, . . . , λ1 , . . . , λ2 , . . . , λt in Grf (α) where f (α) = c, f (β) = d1 , f (λ1 ) = d2 , . . . ,
f (λt−1 ) = dt , f (λt ) = a.
Now according to the definition of a circuit there exists an almost-cycle Q(t). in Grf (α).
According to the function NM-DFS at some point the NM-DFS starts with α and it moves
along an almost-cycle Q(t) and changes the value of f (λt ) from a to c. As it moves along
Q(t), once it reaches β the NM-DFS algorithm changes the value of f (λt ) from c to d1 and
also changes the value of f (α) from c to d1 , once NM-DFS reaches λ1 on Q(t), it changes the
value of f (β) from d1 to d2 and the value of f (α) from d1 to d2 . It continues, until it reaches
λt−1 , and at this point the value of all f (λi ), 1 ≤ i ≤ t is set to dt as well as the value of
f (α), f (β). Now there does not exists a path from α in Grf (α) to some vertex α0 = (x; bk , dt )
with f (α0 ) = dt . This is because of the following observations :
Observation: There is no directed path R in Grf (ψ), ψ = (x; ck , d) with f (ψ) 6= d to some
vertex φ = (x; c0k , d0 ) with f (φ) = d0 . To see that suppose such a path R exists. Consider the
last arc of R say e = (y; dk1 , d2 )(x; c0k , d0 ) where we may assume yx ∈ A(G). Since e is an arc
of Grf (ψ) we have f (y; dk1 , d2 )d0 6∈ A(H) and f (y; dk1 , d2 )f (x; c0k , d0 ) ∈ A(H), a contradiction.
Observation: Since C is a source we do not need to consider circuits in other components
of Gx,b .
Now α becomes a source, since f (α) = dt . NM-DFS has changed f (x; a1 , a2 , . . . , ak ) from
a to dt and it does not change it anymore. NM-DFS starts the usual procedure as if there
is no circuit Q. Moreover since C is a sink component there is no path from C to another
invertible component.
Note that the NM-DFS goes through the vertices of λ1 , λ2 , . . . , λt−1 and it continues the
usual NM-DFS from λt−1 . In the next step it returns back and it starts from λt−2 and so on.
45
Closing remark We run the function NM-DFS(x; bk , a). Once we change the value of
f (x; a1 , a2 , . . . , ak ) from a to c (or dt when C is invertible) then potentially we need to modify
the value for f (y; b1 , b2 , . . . , bk ) from an out-neighbor of a, say a0 in L(y) to an out-neighbor
of c. As far as the modifying f is concern it would yield the same result if we start from
(x; a, b, . . . , b).
5.5.4 Proof of Lemma 5.8
Suppose g is a homomorphism from G to H with g(x) = a. We may assume that there exists
a (x; e1 , e2 , . . . , ak ) with f (x; e1 , e2 , . . . , ak ) = a. Otherwise in some previous step we have
shown that there exists a homomorphism that does not map x to a if there exists one that
maps x to a. Let w = (x; bk , a) and let c = f (w).
Let G0f be the induced sub-digraph of Grf (w) with vertices w0 = (y; ak1 , a2 ) such that :
• f (w1 ) = a2 , for w1 = (y; a01 , a02 , . . . , a0k ) where f (w1 ) was initially g(y),
• a2 is the value before the last time f (w1 ) is going be to change to f (y; ak1 , a2 ).
. the value of f (w1 ) is going to be changed to f (y; ak1 , a2 ) according to NM-DFS for
the last time.
Let G0 be the induced sub-digraph of G with the vertices y ∈ V (G) where (y; ak1 , a2 ) ∈ G0f .
Observation : Consider a path W in G0f from (x; bk , a) to (x; bk , a). Now let X be a
walk in G from x to x corresponding to W , i.e. W follows X. We note that by replacing a2
with f (y; ak1 , a2 ) for every y ∈ X where (y; ak1 , a2 ) is on path W , we get a mapping from X
to H and under this mapping the beginning of X and end of X are mapped to same vertex
c. Moreover this mapping is a homomorphism from X to H because f is a homomorphism
consistent with L.
Let Y be a path in G0 from x to y. Corresponding to Y there is a path in G0f from
(x; bk , a) to some (y; ak1 , a2 ). Now define h(y) = f (y; ak1 , a2 ). For every y ∈ V (G) \ V (G0 )
set h(y) = g ∗ (y) (here g ∗ (y) is the last value of f (y; a01 , a02 , . . . , a0k ) where initially was g(y)).
Note that h(x) = c.
We need to show that h is a homomorphism from G to H. Let zz 0 (z 0 z) be an arc of G
where z ∈ V (G0 ). We only need to consider the case z ∈ V (G0 ). By definition there exists a
path W from w to some w0 = (z; bk1 , b2 ) ∈ G0f with f (w0 ) = bk+1 . Note that if there does not
exists such path W , then by definition z 6∈ V (G0 ).
Since zz 0 is an arc of G and g is a homomorphism from G to H, g(z)g(z 0 ) ∈ A(H).
Moreover, according to the NM-DFS b2 d2 ∈ A(H) (here d2 is the value in which some vertex
w10 = (z 0 ; d01 , d02 , . . . , d0k ) is going to change from g(z 0 ) after visiting the arc zz 0 by the NMDFS). Note that d2 is in L(z 0 ). This is because either d2 = g(z 0 ) and hence it passes the
46
PreProcessing test or d2 = f (z 0 ; ...) and hence by Lemma 5.5 d2 ∈ L(z 0 ). Now if bk+1 d2 is an
arc of H then we are done. Otherwise according to the NM-DFS there exist vertex d1 ∈ L(z 0 )
such that b1 d1 , b2 d2 ∈ A(H), and w0 w00 , w00 = (z 0 ; dk1 , d2 ), is an arc of G0f . Since this is the
last time NM-DFS change the value of some f (w10 ) from g(z), we have w00 ∈ G0f and hence
z 0 ∈ V (G0 ). This would imply that h(z 0 ) = f (w00 ). Since f is a homomorphism consistent
with L, we have f (w0 )f (w00 ) ∈ A(H) and because we set h(z) = f (w0 ) and h(z 0 ) = f (w00 ) we
have h(z)h(z 0 ) ∈ A(H). Note that in the special case when z 0 = x we have g(z 0 ) = a and
by the Observation above we have h(z 0 ) = c. Therefore h is a homomorphism from G to H
with h(x) = c.
The argument is the same when the strong component Gx,b (a) does not have a source.
5.6 Proof of Theorem 1.1 (second approach)
By Lemma 5.8 we preserve the existence of a homomorphism from G to H after Algorithm
8. By Lemma 5.7 f is still a homomorphsim from G × H k → H consistent with the lists L
of G after Algorithm 8. Now we can apply Theorem 5.4.
According to the proof of Lemma 5.7 (2) the running time of Algorithm 4 is O(|A(G)||H|2k+1 ).
The running time of Algorithm 2.2 is O(|G||A(G)||H|k+1 ). In the Algorithm 6 we call
Algorithm 8 at most |G||H| times. We observe that the running time of PreProcessing
O(|G|3 |H|2 ). However, one single run but at each step of while loop in the Algorithm 6
we only remove one element of H therefore the overall time for PreProcessing would be
O(|G|3 |H|2 ). Therefore the running time of Algorithm 6 is O(|G|3 |H|2k+2 ).
5.7 Additional Algorithms and Proofs (New Minority Algorithm)
This section collects additional statements and proofs mainly related to new Minority algorithm. We develops a direct algorithm to handle the minority case in our main algorithm.
Of course, it would be easier to appeal to Bulatov-Dalmau Maltsev result as explained in
Subsection 5.4.
A direct proof for Theorem 5.4 Let M 3 be the digraph of triple (y, c, d), x ∈ G, c, d ∈
L(y). There is an arc from (y, c, d) to (y 0 , c0 , d0 ) where yy 0 ∈ A(G), cc0 , dd0 ∈ A(H) and
cd0 6∈ A(H) or y 0 y ∈ A(G) and c0 c, d0 d ∈ A(H), d0 c 6∈ A(H). Note that M 3 is a graph but we
view M 3 as digraph and when we talk about a path in M 3 we mean an oriented path that
reflect the direction of the edge xy when (y, c, d)(y 0 , c0 , d0 ) is an arc of M 3 . Let Gx (a, b) be
the strong component of M 3 containing (x, a, b) such that for each vertex (y, c, d) ∈ Gx (a, b),
(a, c) ∈ L(x, y) and (b, d) ∈ L(x, y). We say a strong component Gx (a, b) of M 3 is invertible
if both (x, a, b) and (x, b, a) are in Gx (a, b).
Again the idea is similar to the one handling the RemoveNotMinority case. At each step we
consider a vertex x of G and two vertices a, b ∈ L(x) and try to eliminate one of the a, b from
L(x). To decide whether remove a or b we construct an instance of the problem say (G0 , H, L0 )
47
and solve this instance recursively. Based on the existence of a L0 -list homomorphism from
G0 to H we decide to remove a or b. Depending on Gx (a, b) being invertible or not two
different instances for the sub-problem are constructed. At the end we have singleton lists
and if there is a homomorphism from G to H with the singleton lists then success otherwise
we report there is no homomorphism from G to H. We denote the underline graph of digraph
G by U N (G). If Gx (a, b) is invertible we construct a new instance G0 , H, L0 and to construct
the lists L0 we use the Maltsev property.
Lemma 5.10 Let X be an oriented path in G and let B, C, D be three walks in L(X) all
congruent to X where B is from a to c and C is from b to c and D is from b to D. Then
there exists a walk E from a to d in L(X) that is congruent with X.
Proof: By following B, C, D on the vertices in X and applying the definition of polymorphism h, we conclude that E exists.
Lemma 5.10 implies the following corollary.
Corollary 5.11 If (a, c) ∈ L(x, y) and (b, c), (b, d) ∈ L(x, y) then (a, d) ∈ L(x, y).
Lemma 5.12 The Algorithm 9 runs in O(|E(G)||H|k+1 ). Moreover, if there is a homomorphism g from G to H with g(x) ∈ {a, b} then there is a homomorphism from G to H after
removing a or b from L(x) according to Maltsev-(G, H, L, h).
Proof: We spend |G||H|k+1 to define h. For every pair a, b ∈ L(x) the algorithm construct an instance of the problem which takes O(|A(G)||A(H)|3 ). The algorithm is recursive
and the depth of the recursion in |G||H|. We conclude that the running time of function
Maltsev(G, H, L) is O(|G||A(G)||H|4 ) (replaces |A(H)| by |H|2 ) (see the analysis of Algorithm 4). Therefore the running time of the Algorithm 2.2 is O(|A(G)||H|k+1 ).
First consider the case that Gx (a, b) is not invertible. The Algorithm 9 is recursive
but at each recursive call the size of the input decreases by at least one. This is because we
do not add b into L0 (x) and hence we have an instance (G0 , L0 , H) of the problem in which at
least one vertex has a smaller size list. We also note that once we make a decision to remove
a vertex from a list the decision is not changed. Therefore the overall procedure in this
case is polynomial (assuming in each call Gx (a, b) is not invertible). According to Maltsev(G, H, L, h) if G0 does not admit a homomorphism to H then there is no homomorphism
from G to H that maps x to a since G0 is a sub-digraph of G.
No suppose b is removed according to Maltsev-(G, H, L, h) and there exists a homomorphism g from G to H with g(x) = b. Let ψ be the homomorphism from G0 to H. Note that
ψ(x) = a.
Define G1 be a sub-digraph of G0 consists of the vertices y such that (y, ψ(y), g(y)) is
reachable from (x, a, b) in Gx (a, b).
48
Algorithm 9 RemoveMinority – Using Matlsev Operations
1: function RemoveMinority(G, H, L)
2:
Input: x, a, b such that f (x; bk |i←a ) = a
3:
Define Maltsev consistent homomorphism h : G × H 3 → H where
h(x; a, b, c) = f (x; a, b, b, . . . , b, c) for a, b, c ∈ L(x).
4:
while ∃ x ∈ V (G) with |L(x)| ≥ 2 do
5:
Matlsev(G, H, L )
6:
if ∃ a list homomorphism from G to H then return True
7:
else return False.
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
24:
25:
26:
27:
function Maltsev(G, H, L)
Let a, b ∈ L(x) be two distinct vertices. Construct Gx (a, b).
if Gx (a, b) is not invertible then
Set L0 (x) = a and L0 (y) = ∅ for every x 6= y.
Set G00 = ∅, and let G0 be an induced sub-digraph of G constructed as below.
for all y ∈ G s.t. ∃ a path from (x, a, b) to (y, c, d) in Gx (a, b) do
add y to G0 and c to L0 (y),
let yy 0 an arc of G0 when (y, c, d)(y 0 , c0 , d0 ) ∈ Gx (a, b).
if Gx (a, b) is invertible then
Set L0 (x) = a and L0 (y) = ∅ for every x 6= y.
Let G00 be the set of vertices y ∈ V (G) where (y, c, d, e) lies on an oriented path
P from (x, a, b, b) to (x, b, b, a) in U N (G × H 3 ) s.t. no intermediate vertex in P
is (x, a0 , b0 , c0 ) for a0 , b0 , c0 ∈ L(x)
for all arc (y, c, d, e)(y 0 , c0 , d0 , e0 ) ∈ P do
add h(y, c, d, e) to L0 (y), h(y 0 , c0 , d0 , e0 ) to L0 (y 0 ), and add arc yy 0 (y 0 y) to G00
Let G0 be the induced sub-digraph of G with vertices
y ∈ G \ G00 and arcs yy 0 where (y, c, d)(y 0 , c0 , d0 ) ∈ Gx (a, b) and add c into L0 (y)
when (y, c, d) ∈ Gx (a, b)
if Matlsev(G0 ∪ G00 , H, L0 ) then
remove b from L(x)
else remove a from L(x).
49
Now for every vertex y ∈ G1 set φ(x) = ψ(x) and for every y ∈ G \ G1 set φ(y) = g(y).
Now let zz 0 be an arc of G. If none of the z, z 0 ∈ G1 then clearly since g is a homomorphism, φ(z)φ(z 0 ) ∈ A(H). If both z, z 0 ∈ G1 then again since ψ is a homomorphism, we have
φ(z)φ(z 0 ) ∈ A(H). Suppose z ∈ G1 and z 0 6∈ G1 . Since z ∈ G1 there exists a path in Gx (a, b)
from (x, a, b) to (z, ψ(z), g(z)). Now if ψ(z)g(z 0 ) is an arc then we are done. Otherwise
(z 0 , c, g(z 0 )) ∈ Gx (a, b) where c ∈ L(z 0 ) and ψ(z)c ∈ A(H) (note that since ψ(z) is in L(z) it
must have an out-neighbor in L(z 0 )). This would mean z 0 ∈ G1 , a contradiction. Note that
when z 0 = x and z ∈ G1 then because (ψ(z), a) ∈ L(z, x) we have ψ(z)a ∈ A(H).
Second consider the case that Gx (a, b) is invertible. Again as we argued in the
previous case the algorithm is recursive but at each recursive call the size of the input
decreases and once a decision made (removing a or b ) it won’t change.
Observe that if there exists a homomorphism g that maps x to a then for every closed
walk X from x to x, the image of g(X) is a closed walk from a to a in H. Since there is
a walk BB from b to b in L(X) congruent and since (a, b) ∈ L(x, x), there is a walk BA
in L(X) congruent to X. Now by Lemma 5.10 there is a walk AB from a to b in L(X)
and congruent with X. These would imply that there exists a path P in U (G × H 3 ) from
(x, b, b, a) to (x, a, b, b). Now according to the definition of G00 for every walk BB in L(X)
from b to b congruent with X we keep a walk AA in L0 (X) congruent with X and is obtained
by adding h(y, c1 , c2 , c3 ) into L0 (y) where (y, c1 , c2 , c3 ) is in P .
This would mean that in the list of L0 (X), corresponding to ψ(X) (ψ is a homomorphism
from G to H where ψ(x) = a if there exists one) we have a path from a to a in L0 (X).
Therefore if there exists no homomorphism from G0 ∪ G00 → H that maps x to a then there
is no homomorphism from G to H that maps x to a.
Now suppose b is removed according to Maltsev-(G, H, L, h) and there exists a homomorphism g from G to H with g(x) = b. Let ψ1 be the homomorphism from G0 to H and ψ2 be
a homomorphism from G00 to H. Note that ψ1 (x) = ψ2 (x) = a. First we show that there is
no arc e (forward or backward) from a vertex y in G0 to a vertex in G00 . If this is the case
then there is a walk Q from x to y and there is a walk Q0 from y to x. Now since yz is an
arc then Qee−1 Q0 is a closed walk from x to x and hence we should have added z into G00 .
Note that G00 consists of all the induced oriented cycles including x and also all the paths
reaching out of vertices of these cycles except x.
Define G1 be a sub-digraph of G0 consists of vertices y such that (y, ψ1 (y), g(y)) is reachable
from (x, a, b) in Gx (a, b).
Now for every vertex y ∈ V (G1 ) set φ1 (y) = ψ(y) and for every y ∈ V (G) \ V (G1 )
set φ1 (y) = g(y). Define G2 be a sub-digraph of G00 consisting of vertices y such that
(y, ψ2 (y), g(y)) is reachable from (x, a, b) in Gx (a, b).
Now for every vertex y ∈ V (G2 ) set φ2 (y) = ψ2 (y) and for every y ∈ V (G) \ V (G2 )
set φ2 (y) = g(y). And finally let φ(y) = φ1 (y) when y ∈ V (G0 ) and φ(y) = φ2 (y) when
y ∈ V (G00 ). Now let zz 0 be an arc of G. We need to verify that φ(z)φ(z 0 ) is an arc of H.
If none of the z, z 0 ∈ V (G1 ∪ G2 ) then clearly since g is a homomorphism, φ(z)φ(z 0 ) ∈
50
A(H). If both z, z 0 ∈ V (Gi ) (i = 1, 2) then again since ψi is a homomorphism, we have
φ(z)φ(z 0 ) ∈ A(H). Suppose z ∈ V (G2 ) and z 0 6∈ V (G2 ). Since z ∈ V (G2 ), there exists a
closed walk ZZ 0 that contains z where Z is a path from x to z and Z 0 is a walk from z
to x. Note that ψ2 (ZZ 0 ) is a walk from a to a in L(ZZ 0 ) and g(ZZ 0 ) is a walk from b to
b in L(ZZ 0 ) and since (b, a) ∈ L(x × x) there exists a walk BA in L(ZZ 0 ) congruent with
ZZ 0 . These would imply that corresponding to ZZ 0 there exists a path P from (x, b, b, a) to
(x, a, b, b) containing vertex (z, ψ2 (z), d, g(z)). Now since g(z) ∈ L(z) there exists a vertex
c ∈ L(z 0 ) such that g(z)c ∈ A(H) and there exists e0 ∈ L(z 0 ) such that dd0 ∈ A(H).
Now we add the two arcs (z, ψ2 (z), d, g(z))(z 0 , c, d0 , g(z 0 )), (z 0 , c, d0 , g(z 0 ))(z, ψ2 (z), d, g(z))
into P and hence we obtain a path P 0 from (x, b, b, a) to (x, a, b, b) that goes through vertex
(z 0 , c, d0 , g(z 0 )) and hence z 0 ∈ G00 . Now since φ2 (z)g(z 0 ) 6∈ A(H), by definition z 0 ∈ V (G2 ), a
contradiction.
As we argued before one can show that φ1 is also a homomorphism from G to H. Therefore
φ is a homomorphism from G to H with φ(x) = a.
Acknowledgements : We would like to thank Ross Willard and Pavol Hell for so many
helpful discussions and their useful comment as well as their enormous support. We would
like to thank Vı́ctor Dalmau for so many useful questions and helpful comments. We would
also like to thank Geoffrey Exoo, László Egri, for their comments.
References
[ABISV09]
E. Allender, M. Bauland, N. Immerman, H. Schnoor, and H. Vollmer. The complexity of satisfiability problems: refining schaefer’s theorem. Journal of Computer and
System Sciences, 75(4): 245–254 (2009).
[BHM88]
J. Bang-Jensen, P. Hell, G. MacGillivray. The complexity of colouring by semicomplete
digraphs. SIAM J. Discrete Math., 1 : 281–298 (1988).
[BH90]
J. Bang-Jensen, P. Hell. The effect of two cyles on the complexity of colourings by
directed graphs. Discrete Appl. Math., 26 : 1–23 (1990).
[BKN09]
L. Barto, Marcin Kozik, and Todd Niven. The CSP Dichotomy Holds for Digraphs
with No Sources and No Sinks (A Positive Answer to a Conjecture of Bang-Jensen
and Hell). SIAM J. Comput, 38(5) : 1782–1802 (2009).
[B02b]
A.A. Bulatov. A dichotomy constraint on a three-element set. In Proceedings of STOC
649–658 (2002).
[B05]
A.Bulatov. H-Coloring dichotomy revisited. Theoret. Comp. Sci., 349 (1) : 31-39
(2005).
[B06]
A. Bulatov. A dichotomy theorem for constraints on a three-element set. Journal of
the ACM, 53(1): 66–120 (2006).
51
[B11]
A.Bulatov. Complexity of conservative constraint satisfaction problems. Journal of
ACM Trans. Comput. Logic, 12(4) : 24–66 (2011).
[BD06]
A. Bulatov and V. Dalmau. A Simple Algorithm for Mal’tsev Constraints. SIAM J.
Comput., 36(1): 16–27 (2006).
[BJK05]
A.A. Bulatov, P. Jeavons, and A. Krokhin. Classifying the complexity of constraints
using finite algebras. SIAM journal on computing, 34(3): 720-742 (2005).
[JB17]
J. Bulin. Private communication.
[CCL13]
J.Y. Cai, X. Chen, P. Lu. Graph Homomorphisms with Complex Values: A Dichotomy
Theorem. SIAM J. Comput., 42(3): 924–1029 (2013).
[CVK10]
C. Carvalho, V.Dalmau, and A.A. Krokhin. CSP duality and trees of bounded pathwidth. Theor. Comput. Sci., 411(34–36): 3188–3208 (2010).
[CKS01]
N. Creignou, S. Khanna, and M. Sudan. Complexity Classifications of Boolean Constraint Satisfaction Problems. SIAM Monographs on Discrete Math. and Applications,
vol. 7 (2001).
[CL14]
P. Csikvári and Z. Lin. Graph homomorphisms between trees. Elec. J. Combin., 21 :
4–9 (2014).
[D92]
R. Dechter. Containt networks. Encyclopedia of Artificial Intelligence 276–285 (1992).
[D00]
V. Dalmau. A new tractable class of constraint satisfaction problems. In Proceedings
6th International Symposium on Artificial Intelligence and Mathematics, 2000.
[DF03]
V. Dalmau, D. Ford. Generalized satisfiability with k occurrences per variable: A
study through delta-matroid parity. In Proceedings of MFCS 2003, Lecture Notes in
Computer Science, 2747 : 358–367 (2003).
[EHLR14]
L.Egri, P.Hell, B.Larose, and A.Rafiey. Space complexity of List H-coloring : a dichotomy. In Proceedings of SODA, (2014).
[F01]
T. Feder. Homomorphisms to oriented cycles and k-partite satisfiability. SIAM J.
Discrete Math., 14 : 471–480 (2001).
[F06]
T. Feder A dichotomy theorem on fixed points of several nonexpansive mappings.
SIAM J. Discrete Math., 20 : 291–301 (2006).
[FHH99]
T.Feder, P.Hell, J.Huang. Bi-arc graphs and the complexity of list homomorphisms.
J. Graph Theory, 42 : 61–80 (1999).
[FH98]
T. Feder, P. Hell. List homomorphisms to reflexive graphs. J. Comb. Theory Ser., B
72 : 236–250 (1998).
[FHH99]
T. Feder, P. Hell, J. Huang. List homomorphisms and circular arc graphs. Combinatorica, 19 : 487–505 (1999).
52
[FHH03]
T. Feder, P. Hell, J. Huang. Bi-arc graphs and the complexity of list homomorphisms.
J. Graph Theory, 42 : 61–80 (2003).
[FHH07]
T. Feder, P. Hell, J. Huang. List homomorphisms of graphs with bounded degrees.
Discrete Math., 307 : 386–392 (2007).
[FMS04]
T. Feder, F. Madelaine, I.A. Stewart. Dichotomies for classes of homomorphism problems involving unary functions. Theoret. Comput. Sci., 314 : 1–43 (2004).
[FV93]
T.Feder and M.Vardi. Monotone monadic SNP and constraint satisfaction. In Proceedings of STOC, 612–622 (1993).
[FV98]
T. Feder and M.Y. Vardi. The computational structure of monotone monadic SNP
constraint satisfaction: A study through Datalog and group theory. SIAM Journal on
Computing, 28(1): 57–104 (1998).
[HN90]
P. Hell and J. Nešetřil. On the complexity of H-colouring. J. Combin. Theory B, 48 :
92–110 (1990).
[HN04]
P. Hell, J. Nešetřil. Graphs and Homomorphisms, Oxford University Press, 2004.
[HN08]
P.Hell, J.Nešetřil. Colouring, Constraint Satisfaction, and Complexity. Computer Science Review , 2(3): 143–163 (2008).
[HR11]
P.Hell and A.Rafiey. The Dichotomy of List Homomorphisms for Digraphs. In Proceedings of SODA 1703–1713 (2011).
[HR12]
P. Hell and A. Rafiey. The Dichotomy of Minimum Cost Homomorphism Problems
for Digraphs. SIAM J. Discrete Math., 26(4): 1597–1608 (2012).
[KSDR96]
L.G. Kroon, A. Sen, H. Deng, A. Roy. The optimal cost chromatic partition problem for trees and interval graphs. In Graph-Theoretic Concepts in Computer Science
(Cadenabbia, 1996), Lecture Notes in Computer Science, 1197 : 279–292 (1997).
[K92]
V. Kumar. Algorithms for constraint-satisfaction problems. AI Magazine, 13 :32–44
(1992).
[J98]
P. Jeavons. On the Algebraic Structure of Combinatorial Problems. Theor. Comput.
Sci., 200(1-2): 185-204 (1998).
[L75]
R. Ladner. On the Structure of Polynomial Time Reducibility. Journal of the ACM
(JACM), 22(1): 155–171 (1975).
[LZ03]
B. Larose, L. Zádori. The complexity of the extendibility problem for finite posets.
SIAM J. Discrete Math., 17 : 114–121 (2003).
[MM08]
M. Maroti, and R. McKenzie. Existence theorems for weakly symmetric operations.
Algebra Universalis, 59 : 463–489 (2008).
53
[S78]
T.J. Schaefer. The complexity of satisfiability problems. In Proceedings of STOC, 216–
226 (1978).
[S10]
M.Sigge. A new proof of the H-coloring dichotomy. SIAM J. Discrete Math., 23 (4) :
2204–2210 (2010).
[V00]
M.Y. Vardi. Constraint satisfaction and database theory: a tutorial. Proceedings of
the 19th Symposium on Principles of Database Systems (PODS), 76–85 (2000).
54
| 8cs.DS
|
International Journal of Computer Trends and Technology (IJCTT) – volume 14 number 2 – Aug 2014
Selection of Most Appropriate Backpropagation
Training Algorithm in Data Pattern Recognition
Hindayati Mustafidah #1, Sri Hartati #2, Retantyo Wardoyo #3, Agus Harjoko #4
1
Engineering Informatics - Muhammadiyah University of Purwokerto
Purwokerto, Indonesia
234
Computer Science – FMIPA - Gadjah Mada University
Yogyakarta, Indonesia
Abstract—There are several training algorithms for back
propagation method in neural network. Not all of these
algorithms have the same accuracy level demonstrated through
the percentage level of suitability in recognizing patterns in the
data. In this research tested 12 training algorithms specifically in
recognize data patterns of test validity. The basic network
parameters used are the maximum allowable epoch = 1000,
target error = 10-3, and learning rate = 0.05. Of the twelve
training algorithms each performed 20 times looping. The test
results obtained that the percentage rate of the great match is
trainlm algorithm with alpha 5% have adequate levels of
suitability of 87.5% at the level of significance of 0.000. This
means the most appropriate training algorithm in recognizing
the the data pattern of test validity is the trainlm algorithm.
Keywords—validity, appropriate, training algorithms, data
pattern recognition.
I. INTRODUCTION
One of the computer technology development in the
creation of the current approach in resolving an issue is soft
computing [1]. Soft Computing is part of the intelligent
system that is a model approach to computing by imitating
human reason and has the ability to reason and learn in an
environment that is filled with uncertainty and inaccuracy.
Soft Computing consists of several main components
constituent that is fuzzy systems, neural network, evolutionary
algorithms, and probabilistic reasoning. Performance of soft
computing in solving problems in various areas of life has
proven its benefits. One of the methods used in this study is an
Artificial Neural Network (ANN), which is the ideal solution
for problems that can’t be easily formulated using an
algorithm [2]. In the neural network there are several learning
algorithms. Based on how it works, learning algorithm is
divided into two parts with adaptation-based and trainingbased learning. The adaptation-based learning algorithm
consists of two algorithms namely learngd and learngdm,
while the training-based learning algorithm consisting of
twelve algorithms, they are traingd, traingdm, traingdx, trainrp,
traingda, traincgf, traincgp, traincgb, trainscg, trainbfg,
trainoss, and trainlm [3]. Training-based learning algorithm is
often referred to as a training algorithm.
The use of training algorithm in neural network has been
done in solving, for example, predicts student learning
achievement at Informatics Engineering in Muhammadiyah
ISSN: 2231-2803
University of Purwokerto that is based on the grades in the
subjects to be tested in the national exam as in High School
[4]. The training algorithm used is traingd and produces error
rate of 0.0664179 of target error 0.05. Seen that the resulting
error rates in excess of a expected target error. It shows that it
needs to review and study the learning algorithm that held the
most optimal in problem solving. Some studies have also been
carried out to determine the most optimal algorithm that gives
the smallest error i.e. comparing the optimum level of training
algorithms traingd, traingdm, learngd, and learngdm in case of
student learning achievement predictions at Informatics
Engineering Muhammadiyah University of Purwokerto [5].
Based on the results of the statistical tests of the four
algorithms with the alpha level 5%, obtained the value
significance of 0,632. It means that the four algorithms have
no differences in optimum level. In other words, the four
training algorithms have the same proper level. Based on this
case, it needs to be done further testing on other training
algorithms to obtain reference the most appropriate one.
On this research was performed testing against the twelve
training algorithms to obtain the most appropriate algorithm in
data pattern recognizing. Appropriateness is shown by the
percentage of matches in recognizing patterns of data i.e. data
patterns of training and test data pattern where the test data
used is similar to the training data. A case study used is of
data validity of test questions based on the percentage
category of test item in Bloom revised taxonomy developed
by Anderson-Krathwohl. This Data is used in the this research
by reason of the training algorithms obtained will be used in
the determination of the level of the validity of the question
test with a different case so that will be obtained the right
predictions of reserved test questions validity.
Validity is a measure that shows the levels of instrument
validity. A test is said to have validity if the result corresponds
to criterion, in the sense of having equality between the test
results and criteria [6]. A test called valid if the test were able
to precisely measure what is measured. If the data that is
retrieved from a valid instrument, it can be said that the
instrument is valid as it can give a description of the data
correctly in accordance with the fact or circumstance indeed.
In doing the analysis of validity about commonly used
correlation product moment formula [7]. A test is said to have
the validity of the construction when the details matter that
http://www.ijcttjournal.org
Page 92
International Journal of Computer Trends and Technology (IJCTT) – volume 14 number 2 – Aug 2014
build these tests measure every aspect of thinking as
mentioned in the learning objectives. Construction in this
sense is psychological invention itemize the contents of
several aspects such as: memory/knowledge, comprehension,
application, analysis, synthesis, and evaluation better known
as Bloom's taxonomy [8]. The six aspects by [9] developed as
six cognitive processes dimensions, namely: remembering
(C1), understanding (C2), applying (C3), analyzing (C4),
evaluating (C5), and creating (C6) that known a revision of
Bloom's Taxonomy by Anderson-Krathwohl. The validity of
the construction can be known with detailing and pair each
test item with every aspect of the learning objectives.
II. RESEARCH METHODS
The data used in this research are primary and secondary
data. Primary data are student’s scores and test questions at
National Exam in 2009 of the province of Yogyakarta and
Bengkulu. The primary data retrieved from the PUSPENDIK
(Centre for Educational Assessment) in Jakarta, while the
secondary data were retrieved from the data analysis of the
research about test analysis at 2003, 2004, 2005 in Banyumas
Regency and its surroundings ([10], [11], [12], [13], [14], [15],
[16], [17], [18], [19]). The research data were presented in
Table I.
TABLE I
RESEARCH DATA
C1
20
0
0
18.75
12.5
12.9
16.67
14.29
17.14
50
50
48
48
60
28.89
28.89
28.89
40
40
...
C2
35
15
15
12.5
6.25
9.68
56.67
62.86
57.14
30
30
42
42
36.67
37.78
37.78
37.78
38
38
0
0
Percentage
C3
C4
0
45
29
6
29
6
59.38 9.38
71.88 9.38
67.74 9.68
23.33 3.33
20
2.86
17.14 8.57
10
10
10
10
6
4
6
4
3.33
0
26.67 6.67
26.67 6.67
26.67 6.67
8
14
8
14
45
52.5
C5
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
C6
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
2.5
Data preprocessing
setting network
parameters
train the training
data
Simulating of testing
data = training data
Network output of
testing result
Data postprocessing
Camparing network
output simulated data and
training target data
Validity
0.351
0.308
0.376
0.232
0.205
0.186
0.752
0.296
0.345
0.361
0.232
0.071
0.175
0.218
0.184
0.221
0.155
0.162
0.201
0.458
This research uses mixed method, namely qualitative and
quantitative research. The qualitative research carried out by
developing a program code for processing data using the
twelve algorithms (Fig. 1), whereas quantitative research
conducted by statistical testing against the accuracy of the
twelve algorithms in recognizing data patterns shown by the
percentage of matches between the results of network
ISSN: 2231-2803
simulation data and the input data. These algorithms are
traingd, traingdm, traingdx, trainrp, traingda, traincgf, traincgp,
traincgb, trainscg, trainbfg, trainoss, and trainlm. The output
data of any algorithm is done by looping 20 times.
Fig. 1 The process of training and testing data
Statistical testing is done using SPSS software to
determine the most appropriate algorithm using one-way
analysis of variance (One-way ANOVA) [20]. The hypotheses
used are:
H0: the variants of the twelve training algorithms are the same
H1: the variants of the twelve training algorithms are not the
same.
The rejection of H0 is performed if the value significance
obtained is less than α (in this case α = 0.05 or 5%).
III. RESULTS AND DISCUSSION
A. The Structure of the Network
The network structure is shown in Fig. 2.
B. The Network Parameters
Parameters of the network includes the learning rate (LR) =
0.05, the maximum amount of epoch = 1000, and the level of
error = 0.001 (10-3).
http://www.ijcttjournal.org
Page 93
International Journal of Computer Trends and Technology (IJCTT) – volume 14 number 2 – Aug 2014
testing against both algorithms using independent samples ttest (Table VI).
TABLE III
POSTHOC TEST OF DUNCAN MULTIPLE COMPARISONS ANALYSIS
FOR 12 ALGORITHMS USING SPSS
Matched reached in percentages
Training
Subset for alpha = 0.05
N
Algorithms
1
2
3
4
20 58.125
Duncana traingda
traingd
20
64.125
traingdm
20
65.750
20
66.375
traingdx
20
80.500
trainrp
20
84.000
trainoss
traincgp
20
84.121
20
84.375
trainbfg
20
85.000
traincgf
20
85.375
trainscg
20
86.125
traincgb
trainlm
20
1.000 0.110
1.000
0.166
Sig.
Means for groups in homogeneous subsets are displayed.
a. Uses Harmonic Mean Sample Size = 20.000.
Analysis
Fig. 2 Neural network structure with 3 layers
C. Testing
ANOVA test results obtained significance of 0.000, it
means that the value is less than α, so H0 is rejected. In other
word, the twelve algorithms are not homogeneous (have not
the same variant) (Table II).
TABLE II
ANOVA TEST RESULTS OF THE 12 TRAINING ALGORITHMS
USING SPSS
Matched reached in percentages
Sum of
df
Squares
11
Between Groups 25020.286
4007.812
228
Within Groups
29028.099
239
Total
Mean
Square
F
Sig.
2274.571
129.398
0.000
17.578
To learn more the algorithms which have differences level
of appropriateness, performed using a Duncan POSTHOC
method with results as shown in the Table III.
Based on the Duncan method, it shows that traincgf,
trainscg, traincgb, and trainlm algorithms are on the lowest
subset with significance 0.086. This means that the four
training algorithms have the same appropriateness level at α =
5%. To get the most appropriate algorithm of four the training
algorithms, further testing needs to be done and the results as
presented in Table IV.
From the second test result, significance of 0.003 is
obtained which means that all four training algorithms have
inequality on the appropriateness level. Then the testing is
done using Duncan method that is indicated as in Table V.
5
85.000
85.375
86.125
87.500
0.086
TABLE IV
ANOVA TEST RESULTS OF 4 TRAINING ALGORITHMS
Matched reached in percentages
Sum of
df
Squares
Between Groups
73.125
3
Within Groups
371.875
76
Total
445.000
79
Mean
Square
F
Sig.
24.375
4.982
0.003
4.893
TABLE V
POSTHOC TEST OF DUNCAN MULTIPLE COMPARISONS ANALYSIS
FOR 4 ALGORITHMS
Duncan
Subset for alpha = 0.05
1
2
Algorithm traincgf,
trainscg, traincgb, trainlm
N
traincgf
20
85.000
trainscg
20
85.375
traincgb
20
86.125
trainlm
20
Sig.
86.125
87.500
0.133
0.053
Means for groups in homogeneous subsets are displayed.
Based on Table VI seen that with 95% confidence interval
and significance level of α = 0.000 (< 5%), the traincgb
algorithm is in the negative area which means that it has
smaller appropriateness level than trainlm.
It is seen that in this test was obtained similarities at two
algorithms, namely traincgb and trainlm on the significance of
0.053. Because it has not yet obtained the most appropriate
algorithm, testing was continued with an average difference of
ISSN: 2231-2803
http://www.ijcttjournal.org
Page 94
International Journal of Computer Trends and Technology (IJCTT) – volume 14 number 2 – Aug 2014
TABLE VI
THE TEST RESULTS OF THE INDEPENDENT SAMPLE T-TEST
Independent Samples Test
Levene's Test for
Equality of Variances
Matched of
traincgb &
trainlm
Equal variances
assumed
t-test for Equality of Means
Sig.
t
df
Sig. (2tailed)
Mean
Difference
Std. Error
Difference
73.962
0.000
-3.240
38
0.002
-1.375000
0.424380
-2.234113
-0.515887
-3.240
19.000
0.004
-1.375000
0.424380
-2.263238
-0.486762
Equal variances
not assumed
IV. CONCLUSIONS
The 12 training algorithms in neural network have been
tested, namely traingd, traingdm, traingdx, trainrp, traingda,
traincgf, traincgp, traincgb, trainscg, trainbfg, trainoss, and
trainlm in the appropriateness of recognizing the data pattern.
The network parameters include maximum epoch = 1000, the
learning rate = 0.05, and target error = 0.001. The test results
obtained that at the confidence interval of 95% the trainlm
algorithm is the algorithm that is most appropriate in
recognizing the data pattern with an average level of
appropriateness of 87.5%. It is advisable to test the training
algorithms in recognizing data patterns in the other case of
categories of test quality namely reliability, test difference
power, and difficulty level reserved.
[10]
[11]
[12]
[13]
[14]
REFERENCES
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
S. Kusumadewi, and S. Hartati, Neuro-Fuzzy : Fuzzy systems and
Neural Network Integration, Yogyakarta: Graha Ilmu, 2006.
M.T. Jones, Artificial Intelligence: A Systems Approach, New Delhi:
Infinity Science Press LLC, 2008.
S. Kusumadewi, Building Artificial Neural Network Using MATLAB
& EXCEL LINK, Yogyakarta: Graha Ilmu, 2004.
Harjono, and D. Aryanto, Application of Artificial Neural Networks to
Predict Students Study Achievement, SAINTEK ISSN 1411-2558, Vol.
5 No. 2. 2009.
H. Mustafidah, D.K. Hakim, and S. Sugiyanto, Optimization Level of
Training Algorithms in Neural Network (Case Studies of Student
Learning Achievement Predictions), JUITA, ISSN: 2086 – 9398, Vol.
II No. 3, pp. 159–166, May 2013.
S. Arikunto, The basics of Education Evaluation, Revised Edition,
Jakarta: Bumi Aksara, 2006.
L.L. Oriondo and E.M. Dallo-Antonio, Evaluating Educational
Outcomes (Tests, Measurement and Evaluation), Manila: Rex Book
Store, 1998.
M. Wijaya, Taxonomy Bloom in Digital Era Education, [Online],
Available: http://high-aims.blogspot.com/2009/02/taxsonomy-bloomin-digital-era.html, accessed on 12th January 2010, 2009.
L.W. Anderson, and D.R. Krathwohl, Framework Foundation for
Teaching, and Learning Assessment (a Revision of Bloom's Taxonomy
of Education) (translated by: Agung Prihantono), Yogyakarta: Pustaka
Pelajar, 2010.
ISSN: 2231-2803
95% Confidence Interval of
the Difference
Lower
Upper
F
[15]
[16]
[17]
[18]
[19]
[20]
C. Bennyarsih, Correlation between Remidial Learning and
Achievement of Learning Objectives in PKnPS Subject in Term the
Struggle against the Occupiers and the National Movement in Grade V
SD Negeri Segugus Pamuji UPK Karanglewas Semester II in
Academic Year of 2005/2006. Thesis, Purwokerto: FKIP
Muhammadiyah University of Purwokerto (UMP), 2006.
Y. Dwiningsih, The Curriulum 2004 Evaluation System Analysis of
Some Private High Schools on the Subjects of Biology Class X
Academic Year 2004/2005 in Banyumas Regency, Thesis, Purwokerto:
FKIP Muhammadiyah University of Purwokerto (UMP), 2006.
L. Handayani, Evaluation of Mathematics General Test Questions in
High School Class I Quarter 3 in Banyumas Regency, Thesis,
Purwokerto: FKIP Muhammadiyah University of Purwokerto (UMP),
2003.
D.F. Hani, Analysis of Test Items at Even Semester Class X in Biology
Curriculum 2004 Academic Year 2004/2005 in SMAN Sumpiuh
Banyumas Regency, Thesis, Purwokerto: FKIP Muhammadiyah
University of Purwokerto (UMP), 2006.
F. Nurhasanah, Evaluation of Mathematic Test Question at First Class
in SLTP Quarter 3 in Banyumas Regency, Thesis, Purwokerto: FKIP
Muhammadiyah University of Purwokerto (UMP), 2003.
D. Purwanti, An Attempt to Evaluate the Implementation of the
Curriculum 1994 Through Analysis of Test Items in Even Semester at
Final Exam in Study Program Biology Class I Academic Year
2003/2004 at SMA Veteran and SMA Jendral Sudirman Purwokerto,
Thesis, Purwokerto: FKIP Muhammadiyah University of Purwokerto
(UMP), 2005.
D.A. Rahayu, Evaluation of Mathematics Test Questions the Second
Class Quarter 3 in High School in Banyumas Regency, Thesis,
Purwokerto: FKIP Muhammadiyah University of Purwokerto (UMP),
2003.
R.D.P. Sari, Evaluation of the Quality of the Final Exam Questions in
Even Semester the Subjects of Biology Class X Academic Year
2004/2005 in SMAN 3 Purwokerto, SMAN 4 Purwokerto, and SMAN
Sokaraja, Thesis, Purwokerto: FKIP Muhammadiyah University of
Purwokerto (UMP), 2006.
E. Susiani, The Relationship between Students Discipline and Learning
Achievements at SMP Negeri 3 Margasari Tegal Regency in the
Second Class for Social Science Subjects, Thesis, Purwokerto: FKIP
Muhammadiyah University of Purwokerto (UMP), 2005.
N. Wardiningsih, Evaluation of Mathematic Final Test Question the
Even Semester Class II SMPN 1 Karangmoncol Purbalingga Regency,
Thesis, Purwokerto: FKIP Muhammadiyah University of Purwokerto
(UMP), 2004.
T. Taniredja and H. Mustafidah, Quantitative Research (an
Introduction), Bandung: Alfabeta, 2011.
http://www.ijcttjournal.org
Page 95
| 9cs.NE
|
arXiv:1703.03962v3 [math.AC] 4 Sep 2017
PRÜFER CONDITIONS UNDER THE AMALGAMATED
CONSTRUCTION
Y. AZIMI, P. SAHANDI AND N. SHIRMOHAMMADI
Abstract. In this paper we improve the recent results on the transfer of
Prüfer, Gaussian and arithmetical conditions on amalgamated constructions.
As an application we provide an answer to a question posed by Chhiti, Jarrar,
Kabbaj and Mahdou as well as we construct various examples.
1. Introduction
The notion of Prüfer domain appeared for the first time in [20]. Then Krull further studied and named such domains as Prüfer domains [15]. Prüfer domains play
a central role in Multiplicative ideal theory and several different characterizations
of these domains exist [11]. Many of these characterizations have been extended
to the case of rings with zero-divisors. Among them are the Prüfer, Gaussian and
arithmetical rings. It is commonly accepted to define Prüfer rings as the rings in
which every non-zero finitely generated regular ideal is projective [3, 12]. A commutative ring R is called Gaussian if cR (f g) = cR (f )cR (g) for every two polynomials
f, g ∈ R[X], where cR (f ) is the content ideal of f [22]. Finally, a commutative ring
is called an arithmetical ring if every finitely generated ideal is locally principal [14].
It is well known that an arithmetical ring is Gaussian and a Gaussian ring is Prüfer
and that all these coincide in the context of domains. The aim of this paper is
to investigate when the amalgamated algebra is a Prüfer, Gaussian or arithmetical
ring.
D’Anna, Finocchiaro, and Fontana in [6] and [7] have announced the following
ring construction. Let R and S be two commutative rings with unity, let J be
an ideal of S and let f : R → S be a ring homomorphism. They introduced the
following subring
R ⊲⊳f J := {(r, f (r) + j) | r ∈ R and j ∈ J}
of R × S, called the amalgamation of R with S along J with respect to f . This
construction generalizes the amalgamated duplication of a ring along an ideal (introduced and studied in [9]). Moreover, several classical constructions such as the
Nagata’s idealization, the R + XS[X] and the R + XSJXK constructions can be
studied as particular cases of this new construction (see [6, Example 2.5 and Remark
2.8]).
Assuming that J and f −1 (J) are regular ideals, Finocchiaro in [10, Theorem
3.1] obtained that R ⊲⊳f J can be Prüfer ring only in the trivial case, namely when
R and S are Prüfer rings and J = S. He also established some results about the
2010 Mathematics Subject Classification. Primary 13F05, 13A15.
Key words and phrases. Amalgamated algebra, arithmetical ring, Gaussian ring, Prüfer ring,
trivial extension.
1
2
Y. AZIMI, P. SAHANDI AND N. SHIRMOHAMMADI
transfer of some Prüfer-like conditions between R ⊲⊳f J and R. Meanwhile, among
other things, the authors in [4, Theorem 2.2] find necessary and sufficient conditions
under which Prüfer-like properties transfer between the amalgamated duplication
R ⊲⊳ I = R ⊲⊳f I of a local ring R and the ring R itself, where I is an ideal of R,
S = R and f is the identity map on R. In particular, they proved that R ⊲⊳ I is a
Prüfer ring if and only if R is a Prüfer ring and I = rI for every regular element
r of the unique maximal ideal of R. They then asked that is their characterization
valid in the global case? i.e., when R is Prüfer (not necessarily local) or locally
Prüfer [4, Question 2.11].
The outline of the paper is as follows. In Section 2, we determine the zerodivisors of the amalgamated algebra R ⊲⊳f J under certain conditions. In Section
3, among other things, we attempt to generalize the results mentioned above to the
case of the amalgamated algebra R ⊲⊳f J and, consequently, to answer [4, Question
2.11]. In Sections 4 and 5, we investigate the transfer of Gaussian and arithmetical
conditions of the amalgamated algebra R ⊲⊳f J. By various examples we examine
our results.
2. Preliminaries
To determine whether a ring is Prüfer (resp. total ring of quotients) or not, it
is of great importance to know zero-divisors and regular elements of the ring. This
section is devoted to determine the zero-divisor and regular elements of R ⊲⊳f J.
Let first us fix some notation which we shall use throughout the paper: R and S are
two commutative rings with unity, J a proper ideal of the ring S, and f : R → S
is a ring homomorphism. For a commutative ring A, Max(A) denotes the set of
maximal ideals of A, Jac(A) denotes the Jacobson radical of A, Reg(A) denotes the
set of regular elements of A, Z(M ) denotes the set of zero-divisors of an A-module
M and, for an ideal I of A, V(I) denotes the set of all prime ideals of A containing
I. An ideal of ring is called a regular ideal if it contains a regular element.
A full description for the set of zero-divisors of the duplication ring R ⊲⊳ I has
been provided in [17, Proposition 2.2]. In the following lemma we generalize it to
amalgamated algebra. Before we announce this generalization, we should perhaps
point out that {(r, f (r) + j) | j ′ (f (r) + j) = 0, for some j ′ ∈ J \ {0}} ⊆ Z(R ⊲⊳f J)
in general.
Lemma 2.1. There is the inclusion Z(R ⊲⊳f J) ⊆ {(r, f (r) + j) | r ∈ Z(R)} ∪
{(r, f (r) + j) | j ′ (f (r) + j) = 0, for some j ′ ∈ J \ {0}}, with equality if at least one
of the following conditions hold:
(1) f (Z(R)) ⊆ J and f −1 (J) 6= 0;
(2) f (Z(R))J = 0 and f −1 (J) 6= 0;
(3) J ⊆ f (R);
(4) J be a torsion R-module.
Proof. Assume that (r, f (r) + j) ∈ Z(R ⊲⊳f J). Then (r, f (r) + j)(s, f (s) + j ′ ) = 0
for some non-zero element (s, f (s) + j ′ ) ∈ R ⊲⊳f J. Hence rs = 0 and jf (s) +
j ′ (f (r) + j) = 0. If s = 0, then j ′ 6= 0 and j ′ (f (r) + j) = 0. Otherwise, r ∈ Z(R).
This proves the inclusion.
We already have the inclusion {(r, f (r) + j) | j ′ (f (r) + j) = 0, for some j ′ ∈
J \ {0}} ⊆ Z(R ⊲⊳f J). To complete the proof, it is enough for us to show that
{(r, f (r) + j) | r ∈ Z(R) and j ∈ J} ⊆ Z(R ⊲⊳f J) under the validity of any of
PRÜFER CONDITIONS UNDER THE AMALGAMATED CONSTRUCTION
3
conditions (1)–(4). Let (r, f (r) + j) ∈ R ⊲⊳f J where r ∈ Z(R) and j ∈ J. Then
there is a non-zero element s ∈ R such that rs = 0.
(1) If r 6= 0, we have (r, f (r) + j)(s, f (s) + (−f (s))) = 0. So assume that r = 0
and choose 0 6= a ∈ f −1 (J). Then (0, i)(a, 0) = 0.
(2) If r 6= 0, we have (r, f (r) + j)(s, f (s)) = 0. So assume that r = 0 and choose
0 6= a ∈ f −1 (J). Then (0, i)(a, 0) = 0.
(3) If f (s)J = 0, then (r, f (r) + j)(s, f (s)) = 0. Assume now that f (s)J 6= 0.
Then there is 0 6= k ∈ J such that f (s)k 6= 0. By assumption k = f (t) for some
t ∈ R. It is clear that st 6= 0 and (r, f (r) + i)(st, f (st) − f (s)k) = 0.
(4) Let t ∈ Reg(R) such that f (t)j = 0. It is easy to see that ts 6= 0 and
(r, f (r) + i)(ts, f (ts)) = 0.
We say that the amalgamated ring R ⊲⊳f J has the condition ⋆ if the equality
Z(R ⊲⊳f J) = {(r, f (r)+j) | r ∈ Z(R)}∪{(r, f (r)+j) | j ′ (f (r)+j) = 0, ∃j ′ ∈ J \{0}}
holds. Lemma 2.1 now says that the amalgamated ring R ⊲⊳f J involving one of
the mentioned conditions has the condition ⋆. In particular when f is surjective,
by part (3) of Lemma 2.1, R ⊲⊳f J has the condition ⋆. Therefore the amalgamated
duplication of a ring along an ideal has the condition ⋆. Although the trivial
extension of a ring by a module does not satisfy any conditions mentioned in Lemma
2.1, the next remark illustrates that the trivial extension of a ring by a module also
has this condition. So we have a bunch of examples satisfying the condition ⋆.
Let M be an R-module. Nagata (1955) gave a ring extension of R called the
trivial extension of R by M (or the idealization of M in R), denoted here by R ⋉ M
[19, page 2]. It should be noted that the module M becomes an ideal in R ⋉ M
and (0 ⋉ M )2 = 0. As in [6, Remark 2.8], if S := R ⋉ M , J := 0 ⋉ M , and
ι : R → S be the natural embedding, then R ⊲⊳ι J ∼
= R ⋉ M which maps the
element (r, ι(r) + (0, m)) to the element (r, m).
Remark 2.2. Let M be an R-module. By [1, Proposition 1.1], one has
Z(R ⋉ M ) = {(r, m) | r ∈ Z(R) ∪ Z(M ) and m ∈ M }.
This equality together with the isomorphism mentioned above shows that the trivial
extension R ⋉ M has the condition ⋆.
3. Transfer of Prüfer condition
In this section we investigate the transfer of Prüfer condition on the amalgamated
algebra R ⊲⊳f J. The following concept and lemmas are crucial in this investigation.
Let p be a prime ideal of R. Then (R, p) is said to has the regular total order
property if, for each pair of ideals a and b of R, at least one of which is regular,
the ideals aRp and bRp are comparable. Using this notion, Griffin [12, Theorem
13] obtained the following useful characterization of Prüfer rings which we use it
frequently.
Theorem 3.1. The ring R is Prüfer if and only if (R, m) has the regular total
order property for every m ∈ Max(R).
The following basic lemma will be used throughout this section.
Lemma 3.2. Let U be a multiplicatively closed subset of the ring R. If r ∈ Reg(R),
then r/1 ∈ Reg(RU ). The converse holds whenever U ∩ Z(R) = φ. In particular, if
4
Y. AZIMI, P. SAHANDI AND N. SHIRMOHAMMADI
Z(R) ⊆ Jac(R), then, for a maximal ideal m, one has r/1 ∈ Reg(Rm ) if and only
if r ∈ Reg(R).
This lemma enables us to present the following version of the theorem above
which is very useful for us.
Lemma 3.3. Assume that Z(R) ⊆ Jac(R) and that m is a maximal ideal of the
ring R. Then (R, m) has the regular total order property if and only if the principal
ideals xRm and yRm are comparable for all x, y ∈ R with x is regular.
Proof. Assume that the principal ideals xRm and yRm are comparable for all x, y ∈
R with x is regular. Let a and b be ideals of R, and x be a regular element of
R in a. Suppose that y/1 ∈ bm \ am . It follows that xRm ⊆ yRm . To conclude
the inclusion am ⊆ bm , it is enough for us to show that am ⊆ yRm . Suppose on
the contrary that there exists an element a/1 ∈ am \ yRm . Since aRm * xRm , it
follows that xRm ⊆ aRm . It turns out that a/1 is regular. This in conjunction with
Lemma 3.2 implies that a is regular. Hence yRm and aRm are comparable which
is a contradiction.
In the sequel, we will use the following properties of amalgamated rings without
explicit comments.
Remark 3.4. The following statements hold.
(1) ([8, Corollaries 2.5 and 2.7]) For p ∈ Spec(R) and q ∈ Spec(S) \ V(J), set
p′f :=p ⊲⊳f J := {(p, f (p) + j)|p ∈ p, j ∈ J},
qf :={(r, f (r) + j)|r ∈ R, j ∈ J, f (r) + j ∈ q}.
Then, one has the following.
(a) The prime ideals of R ⊲⊳f J are of the type qf or p′f , for q varying in
Spec(S) \ V(J) and p in Spec(R).
(b) Max(R ⊲⊳f J) = {p′f |p ∈ Max(R)} ∪ {qf |q ∈ Max(S) \ V(J)}.
(2) ([8, Proposition 2.9]) The following formulas for localizations hold.
(a) For any q ∈ Spec(S) \ V(J), the localization (R ⊲⊳f J)qf is canonically isomorphic to Sq . This isomorphism maps the element (r, f (r) +
j)/(r′ , f (r′ ) + j ′ ) to (f (r) + j)/(f (r′ ) + j ′ ).
(b) For any p ∈ Spec(R) \ V(f −1 (J)), the localization (R ⊲⊳f J)p′f is
canonically isomorphic to Rp . This isomorphism maps the element
(r, f (r) + j)/(r′ , f (r′ ) + j ′ ) to r/r′ .
(c) For any p ∈ Spec(R) containing f −1 (J), consider the multiplicative
subset Tp := f (R \ p) + J of S and set STp := Tp−1 S and JTp :=
Tp−1 J. If fp : Rp → STp is the ring homomorphism induced by f ,
then the ring (R ⊲⊳f J)p′f is canonically isomorphic to Rp ⊲⊳fp JTp .
This isomorphism maps the element (r, f (r) + j)/(r′ , f (r′ ) + j ′ ) to
(r/r′ , (f (r) + j)/(f (r′ ) + j ′ )).
We are now ready to state and prove the main result of this section. In fact, it
provides a partial converse of [10, Proposition 4.2] as well as a generalization of [4,
Theorem 2.2].
Theorem 3.5. Assume that f (Reg(R)) ⊆ Reg(S).
PRÜFER CONDITIONS UNDER THE AMALGAMATED CONSTRUCTION
5
(1) If R ⊲⊳f J is a Prüfer ring, then R is a Prüfer ring and JTm = f (r)JTm for
every m ∈ Max(R) and every r ∈ Reg(R).
(2) Assume that R ⊲⊳f J has the condition ⋆ and that Z(R ⊲⊳f J) ⊆ Jac(R ⊲⊳f
J). If R is a Prüfer ring and JTm = f (r)JTm for every m ∈ Max(R) and
every r ∈ Reg(R), then R ⊲⊳f J is a Prüfer ring.
Proof. (1) Assume that R ⊲⊳f J is a Prüfer ring. By [10, Proposition 4.2], R
is a Prüfer ring. Let r ∈ Reg(R), m ∈ Max(R), and k/x ∈ JTm . It is easy to
see that (r, f (r)) ∈ Reg(R ⊲⊳f J). By Theorem 3.1, the ideal (r, f (r))(R ⊲⊳f
J)m′ f is comparable with any ideal of (R ⊲⊳f J)m′ f . Hence the principal ideal of
Rm ⊲⊳fm JTm generated by (r/1, fm (r/1)) is comparable
with the principal ideal
generated by (0, k/x). If (r/1, fm (r/1)) Rm ⊲⊳fm JTm ⊆ (0, k/x) Rm ⊲⊳fm JTm ,
then (r/1, fm (r/1)) = (0, k/x)(s/z, fm(s/z) + j/y) for some (s/z, fm(s/z) + j/y) ∈
Rm ⊲⊳fm JTm which implies that r/1 = 0. This is a contradiction since r/1 ∈
fm
Reg(Rm ) by Lemma 3.2. So we
may assume that the inclusion (0, k/x) Rm ⊲⊳ JTmf ⊆
fm
(r/1, fm (r/1)) Rm ⊲⊳ JTm holds. Thus there exists (s/z, fm (s/z)+j/y) ∈ Rm ⊲⊳ m
JTm such that (0, k/x) = (r/1, fm (r/1))(s/z, fm (s/z) + j/y). This means that
s/z = 0 and k/x = fm (r/1)j/y = f (r)j/y ∈ f (r)JTm , as desired.
(2) By Theorem 3.1, it is sufficient to prove that (R ⊲⊳f J, M) has the regular
total order property for every maximal ideal M of R ⊲⊳f J. On the other hand,
it follows from the hypothesis that J ⊆ Jac(S) and Z(R) ⊆ Jac(R). Indeed if
J * Jac(S), there exists a q ∈ Max(S) \ V(J). Then {(0, j) | j ∈ J} ⊆ Z(R ⊲⊳f
J) ⊆ Jac(R ⊲⊳f J) ⊆ qf . This yields J ⊆ q a contradiction. For the second
inclusion assume a ∈ Z(A) and m ∈ Max(R). Then (a, f (a)) ∈ Z(R ⊲⊳f J) ⊆
Jac(R ⊲⊳f J) ⊆ m′f , which implies that a ∈ m. Therefore we have two cases to
consider:
Case 1. Assume that M = m′f for some m ∈ Max(R) such that f −1 (J) *
m. Let (r, f (r) + j) ∈ Reg(R ⊲⊳f J). Then (r, f (r) + j)/(1, 1) ∈ Reg((R ⊲⊳f
J)m′f ). Using the isomorphism (R ⊲⊳f J)m′f ∼
= Rm , one has r/1 ∈ Reg(Rm ) which
implies that r ∈ Reg(R) by Lemma 3.2. Since (R, m) has the regular total order
property, the ideal rRm is comparable with every ideal of Rm . Consequently, the
ideal (r, f (r) + j)(R ⊲⊳f J)m′f is comparable with every ideal of (R ⊲⊳f J)m′f .
Case 2. Assume that M = m′f for some m ∈ Max(R) such that f −1 (J) ⊆ m.
Let (r, f (r) + j) ∈ Reg(R ⊲⊳f J) and (r′ , f (r′ ) + j ′ ) ∈ R ⊲⊳f J. Since R ⊲⊳f J
has the condition ⋆, we have r ∈ Reg(R). Thus the principal ideals rRm and
r′ Rm are comparable because (R, m) has the regular total order property. We may
and do assume r′ Rm ⊆ rRm . So we have r′ /1 = (r/1)(t/u) for some t/u ∈ Rm .
We claim that there exists an element k/y ∈ JTm such that j ′ /1 − fm (t/u)j/1 =
(k/y)(fm (r/1) + j/1). If this is the case, then one has the equality
(r′ /1, fm(r′ /1) + j ′ /1) = (t/u, fm(t/u) + k/y)(r/1, fm (r/1) + j/1)
in the ring Rm ⊲⊳fm JTm which shows the following inclusion of principal ideals
(r′ /1, fm(r′ /1) + j ′ /1)(Rm ⊲⊳fm JTm ) ⊆ (r/1, fm (r/1) + j/1)(Rm ⊲⊳fm JTm ).
This inclusion yields the inclusion (r′ , f (r′ ) + j ′ )(R ⊲⊳f J)m′f ⊆ (r, f (r) + j)(R ⊲⊳f
J)m′f . It remains to prove the claim. To this end, one notices that our assumption
gives an element l/v ∈ JTm such that j/1 = fm (r/1)(l/v). Hence fm (r/1) + j/1 =
fm (r/1)((v + l)/v) and observe that the element (v + l)/v is a unit element of
STm .
6
Y. AZIMI, P. SAHANDI AND N. SHIRMOHAMMADI
We have the following corollaries.
Corollary 3.6. Assume that (R, m) is a local ring, J ⊆ Jac(S) and f (Reg(R)) ⊆
Reg(S).
(1) If R ⊲⊳f J is a Prüfer ring, then R is a Prüfer ring and J = f (r)J for
every r ∈ Reg(R).
(2) Assume further that R ⊲⊳f J has the condition ⋆. If R is a Prüfer ring and
J = f (r)J for every r ∈ Reg(R), then R ⊲⊳f J is a Prüfer ring.
Proof. It follows from the inclusion J ⊆ Jac(S) that the elements of Tm = f (R \
m) + J are units in S for every m ∈ Max(R). Hence, for every r ∈ R and every
m ∈ Max(R), JTm = f (r)JTm holds if and only if J = f (r)J holds. One also notices
from Remark 3.4(1)(b) that (R, m) is a local ring and J ⊆ Jac(S) if and only if
R ⊲⊳f J is local. The result is therefore immediate from Theorem 3.5.
We now deal with the amalgamated duplication. Assume that R = S and f is the
identity map on R. Let I be an ideal of R, 0 6= r ∈ R and m ∈ Max(R). One notices
that if m ⊇ I, then Tm = R \ m + I = R \ m, otherwise, ITm = 0 by [10, Remark
2.4]. Hence ITm = rITm holds if and only if Im = rIm . It also easily obtain that
Z(R ⊲⊳ I) ⊆ Jac(R ⊲⊳ I) if and only if Z(R) ⊆ Jac(R) and I ⊆ Jac(R).Therefore
we can derive the following corollary of Theorem 3.5. This corollary generalizes [4,
Theorem 2.2] to non-local case as well as provides an answer to [4, Question 2.11].
Corollary 3.7. Let I be an ideal of the ring R.
(1) If R ⊲⊳ I is a Prüfer ring, then R is a Prüfer ring and Im = rIm for every
m ∈ Max(R) and every r ∈ Reg(R).
(2) Assume further that Z(R) ⊆ Jac(R) and I ⊆ Jac(R). If R is a Prüfer ring
and Im = rIm for every m ∈ Max(R) and every r ∈ Reg(R), then R ⊲⊳ I is
a Prüfer ring.
Corollary 3.8. (See [4, Theorem 2.2]) Let (R, m) be a local ring and I a proper
ideal of R. Then R ⊲⊳ I is a Prüfer ring if and only if R is a Prüfer ring and
I = rI for every r ∈ Reg(R).
We shall now describe the behaviour of Prüfer condition on trivial extension.
Corollary 3.9. Let M be an R-module such that Z(M ) ⊆ Z(R).
(1) If R ⋉ M is a Prüfer ring, then R is a Prüfer ring and Mm = rMm for
every m ∈ Max(R) and every r ∈ Reg(R).
(2) Assume further that Z(R) ⊆ Jac(R). If R is a Prüfer ring and Mm = rMm
for every m ∈ Max(R) and every r ∈ Reg(R), then R ⋉ M is a Prüfer ring.
Proof. It follows from the inclusion Z(M ) ⊆ Z(R) that ι(Reg(R)) ⊆ Reg(R ⋉ M ).
On the other hand, for r ∈ R and m ∈ Max(R), we have Tm = ι(R \ m) + J =
R ⋉ M \ m ⋉ M where J = 0 ⋉ M . Hence JTm = ι(r)JTm holds if and only
if Mm = rMm holds. Finally, it is easy to see that our assumption Z(M ) ⊆ Z(R)
together with Remark 2.2 implies Z(R⋉M ) ⊆ Jac(R⋉M ). Consequently, Theorem
3.5 completes the proof.
It is worth pointing out that in the corollary above the assumption Z(M ) ⊆ Z(R)
in crucial. For example let (R, m) be a local integral domain which is not a valuation
domain (e.g. R := k + Y k(X)JY K, where k is a field and X, Y are indeterminates
PRÜFER CONDITIONS UNDER THE AMALGAMATED CONSTRUCTION
7
over k). The trivial extension R ⋉ (R/m) is a total ring of quotients, hence, Prüfer,
while R is not (see [16, Proposition 3.1(a)]).
Corollary 3.10. Assume that R is an integral domain or is a local ring and that
M is a torsion-free R-module. Then R ⋉ M is a Prüfer ring if and only if R is a
Prüfer ring and Mm = rMm for every m ∈ Max(R) and r ∈ Reg(R).
Applying the corollary above we obtain an easy proof for the following result of
Bakkari, Kabbaj, and Mahdou.
Corollary 3.11. (cf. [2, Theorem 2.1(1)]) Assume that A ⊆ B is an extension of
domains and let K be the quotient field of A. Then A ⋉ B is a Prüfer ring if and
only if A is a Prüfer domain and K ⊆ B.
Proof. It is well-known that B = ∩m∈Max(A) Bm . Hence, for every 0 6= a ∈ A, one
has Bm = aBm for every m ∈ Max(A) if and only if B = aB. This is the case if
and only if K ⊆ B. The conclusion is now clear by Corollary 3.9.
Total rings of quotients are trivial examples of Prüfer rings. In [10, Proposition
4.6], Finocchiaro gave conditions, under which, total ring of quotients property
transfers from R to R ⊲⊳f J. The next proposition provides a slight generalization
as well as is a partial converse of it.
Proposition 3.12. The following statements hold:
(1) Assume that J ⊆ Jac(S) and that R ⊲⊳f J has the condition ⋆. If R is a
total ring of quotients, then so is R ⊲⊳f J.
(2) Assume that f (Reg(R)) ⊆ Reg(S). If R ⊲⊳f J is a total ring of quotients,
then so is R.
Proof. The proof of (1) is similar to that of [10, Proposition 4.6], but, for the
reader’s convenience, we derive it under the general assumption that R ⊲⊳f J has
the condition ⋆. Assume that R is a total ring of quotients and let (r, f (r) + j) be a
non-unit element of R ⊲⊳f J. Hence (r, f (r)+j) ∈ m′f for some m ∈ Max(R); so that
r ∈ m. This in conjunction with the assumption that R is a total ring of quotients
yields r ∈ Z(R). It now follows from Lemma 2.1 that (r, f (r) + j) ∈ Z(R ⊲⊳f J).
Therefore R ⊲⊳f J is a total ring of quotients. To prove (2), let r ∈ R. If (r, f (r))
is a unit of R ⊲⊳f J, then obviously r is a unit of R. Otherwise (r, f (r)) is a zerodivisor of R ⊲⊳f J. Our desired result will be established if r is a zero-divisor of
R. If not, by Lemma 2.1, there exists j ∈ J such that jf (r) = 0, which contradicts
since f (r) is regular.
4. Transfer of Gaussian condition
Our goal in this section is to determine the behaviour of Gaussian condition
under amalgamated construction. Let X be an indeterminate over R. The content
ideal cR (f ) of a polynomial f ∈ R[X] is defined to be the ideal of R generated by
the coefficients of f . A ring R is called a Gaussian ring if cR (f g) = cR (f )cR (g) for
every two polynomials f, g ∈ R[x]. It is shown in [22, Theorem 2.2] that a local
ring R is Gaussian if and only if, for every two elements a, b ∈ R, the following two
properties hold: (i) (a, b)2 = (a2 ) or (b2 ); (ii) if (a, b)2 = (a2 ) and ab = 0, then
b2 = 0. Using this characterization, the authors in [4, Theorem 3.2(2) and Corollary
3.8(2)] established that the amalgamated duplication ring R ⊲⊳ I is Gaussian if and
2
only if R is Gaussian, Im
= 0 and rIm = r2 Im for every m ∈ Max(R) ∩ V(I) and
8
Y. AZIMI, P. SAHANDI AND N. SHIRMOHAMMADI
every r ∈ m. Then, again, by means of this characterization, for a local ring (R, m)
and assuming J 2 = 0, it is shown that R ⊲⊳f J is Gaussian if and only if so is R
and f (r)J = f (r)2 J for every r ∈ m [18, Theorem 2.1(2)]. In the following, we give
an improvement to this result provided that J ⊆ f (R).
Theorem 4.1. Assume that (R, m) is a local ring and J ⊆ f (R) ∩ Jac(S) (e.g. f
is surjective). Then R ⊲⊳f J is Gaussian if and only if R is Gaussian, J 2 = 0 and
f (r)J = f (r)2 J for every r ∈ m.
Proof. The backward direction is [18, Theorem 2.1(2)]. For the forward direction
assume that R ⊲⊳f J is Gaussian. Then R is a Gaussian ring, since the Gaussian
property is stable under factor rings. Next, let i, j ∈ J. Assume that j = f (a)
for some a ∈ R. By assumption we have ((a, j), (0, j))2 = ((0, j)2 ) or ((a, j)2 )
[22, Theorem 2.2]. If the first case happens, then a2 = 0; so that j 2 = 0. The
second case yields j 2 (1 − j ′ ) = 0 for some j ′ ∈ J. Thus j 2 = 0. Similarly i2 = 0.
Hence ij = 0. To this end one notices that R ⊲⊳f J is Gaussian. Let now 0 6=
r ∈ m and j ∈ J. The last equality is the direct consequence of the statement
((r, f (r)), (0, j))2 = ((r, f (r))2 ) or ((0, j)2 ) together with J 2 = 0.
In the course of this paper, for an R-module M , SuppR (M ) will denote the set
of all prime ideals p of R such that Mp 6= 0.
In a sense, the next corollary is a generalization of [4, Corollary 3.8(3)].
Corollary 4.2. Assume that J ⊆ f (R) ∩ Jac(S). Then R ⊲⊳f J is Gaussian if
2
and only if R is Gaussian, Jm
= 0 and f (r)Jm = f (r)2 Jm for every m ∈ Max(R) ∩
−1
V(f (J)) and r ∈ m.
Proof. One can see Jm = JTm for every m ∈ V(f −1 (J)) since Tm = f (R \ m) + J =
f (R \ m). It also follows from the assumption J ⊆ f (R) that JTm ⊆ fm (Rm ) for
every m ∈ SuppR (J) ∩ V(f −1 (J)). Hence Theorem 4.1 completes the proof.
The next example illustrates the assumption J ⊆ f (R) in the forward direction
of the theorem above is essential.
Example 4.3. Let Z be the ring of integers and p = pZ be the prime ideal generated
by a prime number p. Let X be an indeterminate over the field of rational numbers
Q, and let R := Zp , S := QJXK the formal power series ring over Q, J := XQJXK
the unique maximal ideal of S, f : R → S be the inclusion homomorphism. It
is clear that J * f (R). Notice that R ⊲⊳f J is isomorphic to the composite ring
extension Zp + XQJXK, by [6, Example 2.5]. It follows from [13, Theorem 1.3] that
R ⊲⊳f J is a Prüfer domain, hence, a Gaussian ring [11, Corollary 28.5], while
J 2 6= 0.
As an application of our results we construct some examples of Prüfer rings
which are not Gaussian.
Example 4.4. Let k be a field and X an indeterminate over k. Let R := k[X]/(X 8)
which is a total quotient ring, S := k[X]/(X 4), J := (X 2 )/(X 4 ) and f be the
canonical surjection. Then R ⊲⊳f J has the condition ⋆ since f is surjective. It
follows from Proposition 3.12(1) that R ⊲⊳f J is a total quotient ring, hence Prüfer.
However R ⊲⊳f J is not a Gaussian ring by Theorem 4.1, since XJ 6= X 2 J.
The next example handles the non-local case.
PRÜFER CONDITIONS UNDER THE AMALGAMATED CONSTRUCTION
9
Example 4.5. Let R := Z/48Z, S := Z/24Z, J := 6Z/24Z and f be the canonical
surjection. Then, as Example 4.4, R ⊲⊳f J has the condition ⋆ and it follows from
Proposition 3.12(1) that R ⊲⊳f J is a Prüfer ring. But, it is not a Gaussian ring
2
by Corollary 4.2, since the natural image of 12 = 62 − 24 in Jm
is not zero, where
m = 2Z/48Z.
5. Transfer of arithmetical condition
This section deals with arithmetical condition on amalgamated construction.
The ring R is said to be an arithmetical ring if every finitely generated ideal of
R is locally principal [14]. Thus R is an arithmetical ring if and only if R is a
locally chain ring. Recall that an R-module M is said to be uniserial if its set of
submodules is totally ordered by inclusion and R is a chain ring if it is uniserial as
R-module.
The special case of the following result has been appeared in [5, Proposition 1.1].
Theorem 5.1. Assume that J is a non-zero ideal of S. If R ⊲⊳f J is a chain ring,
then R is a valuation domain and J = (f (a) + j)J for every 0 6= a ∈ R and j ∈ J.
The converse holds provided that J is a uniserial R-module.
Proof. Assume that R ⊲⊳f J is a chain ring. Then, as a factor ring, R is a chain ring.
Let 0 6= a ∈ R and l ∈ J. Clearly (a, f (a)) ∈
/ (0, l)(R ⊲⊳f J). Then (0, l) = (r, f (r)+
f
i)(a, f (a)) for some (r, f (r) + i) ∈ R ⊲⊳ J; so that ra = 0 and l = f (a)i ∈ f (a)J.
Hence J = f (a)J and this shows that R is a valuation domain. Now if further
j ∈ J, there is (r, f (r) + k) ∈ R ⊲⊳f J such that (0, l) = (a, f (a) + j)(r, f (r) + k).
Hence ra = 0 which implies that r = 0. Thus l = (f (a) + j)k ∈ (f (a) + j)J.
Therefore J = (f (a) + j)J.
For the converse assume that J is a uniserial R-module. Let (a, f (a)+j), (b, f (b)+
i) ∈ R ⊲⊳f J be arbitrary. Assume that b 6= 0 and a = bc. Then there exists k ∈ J
such that j − f (c)i = (f (a) + i)k. Hence (a, f (a) + j) = (b, f (b) + i)(c, f (c) + k).
In the case that b = 0, there is r ∈ R such that j = ri(= f (r)i) since J is uniserial.
Hence (0, j) = (r, f (r))(0, i).
We notice that one can not necessarily deduce J is a uniserial R-module from the
assumption that R ⊲⊳f J is a chain ring. In fact, bearing in mind the notations of
Example 4.3, it is clear that the Zp -submodules of J generated by the two elements
X and X 2 are not comparable.
Corollary 5.2. The following statements hold:
(1) Assume that J is a non-zero ideal of S such that J 2 = 0. Then R ⊲⊳f J is a
chain ring if and only if R is a valuation domain, J is a uniserial R-module
and J = f (a)J for every 0 6= a ∈ R.
(2) Assume that J ⊆ f (R). Then R ⊲⊳f J is a chain ring if and only if R is a
chain ring and J = 0.
Proof. (1) It is obvious that J = (f (a) + j)J for every 0 6= a ∈ R and j ∈ J if and
only if J = f (a)J for every 0 6= a ∈ R. Then, in view of Theorem 5.1, it remains
for us to deduce J is uniserial when R ⊲⊳f J is a chain ring. Indeed, for i, j ∈ J,
one of the elements (0, i) and (0, j) divides the other. So we may assume (0, j) =
(r, f (r) + k)(0, i) for some (r, f (r) + k) ∈ R ⊲⊳f J. Thus j = (f (r) + k)i = f (r)i.
(2) One implication is clear. For the other, suppose, on the contrary, that J 6= 0.
Then one can choose a non-zero element j ∈ J and, hence, a non-zero element a ∈ R
10
Y. AZIMI, P. SAHANDI AND N. SHIRMOHAMMADI
such that j = f (a). This together with Theorem 5.1 shows that J = (f (a)−j)J = 0,
which is a contradiction.
Corollary 5.3. (see [4, Theorem 3.2(1)]) The ring R ⊲⊳ I is a chain ring if and
only if R is a chain ring and I = 0.
We now want to generalize Corollary 5.2 to arithmetical rings. To this end, we
need the concept of distributive lattice of submodules. Let M be a module over
the ring R. We say that M has a distributive lattice of submodules if M satisfies
one of the following two equivalent conditions:
(1) (N + L) ∩ K = (N ∩ K) + (L ∩ K) for every submodules N, L, K of M ;
(2) (N ∩ L) + K = (N + K) ∩ (L + K) for every submodules N, L, K of M .
It is known that M has a distributive lattice of submodules if and only if Mm is a
uniserial module for each maximal ideal m of R [5, Proposition 1.2].
The special case of part 2 of the following result was obtained in [4, Corollary
3.8(1)].
Corollary 5.4. The following statements hold:
(1) Assume that J is a non-zero ideal of S with J 2 = 0. Then R ⊲⊳f J is
an arithmetical ring if and only if R is arithmetical, Rp is a domain for
every p ∈ SuppR (J), J is locally divisible and has a distributive lattice of
submodules.
(2) Assume that J ⊆ f (R). Then R ⊲⊳f J is an arithmetical ring if and only if
R is arithmetical, Jm = 0 for every m ∈ Max(R) ∩ V(f −1 (J)) and Sq is a
chain ring for every q ∈ Max(S) \ V(J).
Proof. (1) One can employ [21, Lemma 2.8] together with Corollary 5.2 and the
above-mentioned result [5, Proposition 1.2] to deduce the assertion.
(2) Assume that R ⊲⊳f J is an arithmetical ring. First, note that, as a factor
ring, R is an arithmetical ring. Then we show that Jm = 0 for every m ∈ Max(R) ∩
V(f −1 (J)). Let m ∈ Max(R) ∩ V(f −1 (J)). By assumptions, (R ⊲⊳f J)m′f ∼
=
Rm ⊲⊳fm JTm is a chain ring and one has the inclusion JTm ⊆ fm (Rm ). Hence
JTm = 0 by Corollary 5.2. On the other hand, using the inclusions J ⊆ f (R) and
f −1 (J) ⊆ m, it is easily seen that Tm = f (R \ m) + J = f (R \ m). Therefore Jm = 0.
Finally, the isomorphism (R ⊲⊳f J)qf ∼
= Sq shows that Sq is a chain ring for every
q ∈ Max(S) \ V(J). The converse direction is easy to obtain.
The conclusion of the corollary above fails if the assumption J 2 = 0 or J ⊆ f (R)
is dropped. For example, let X be an indeterminate over the field of rational
numbers Q, and let R := Z, S := QJXK the formal power series ring over Q,
J := XQJXK, f : R → S be the inclusion homomorphism. It is clear that J * f (R)
and that J 2 6= 0. Notice that R ⊲⊳f J ∼
= Z + XQJXK [6, Example 2.5]. It follows
from [13, Theorem 1.3] that R ⊲⊳f J is a Prüfer domain, hence, an arithmetical
ring. However J is not locally divisible.
In concluding we give an example of a non-arithmetical Gaussian ring.
Example 5.5. Let k be a field and X an indeterminate over k. Set R := k[X],
S := k[X]/(X 2), J := (X)/(X 2 ) and let f be the canonical surjection. Then
R ⊲⊳f J is Gaussian by Corollary 4.2. However R ⊲⊳f J is not an arithmetical ring
by Corollary 5.4, since J 6= 0 = XJ.
PRÜFER CONDITIONS UNDER THE AMALGAMATED CONSTRUCTION
11
References
1. M. Axtell and J. Stickles, Zero-divisor graphs of idealizations, J. Pure Appl. Algebra, 204,
(2006), 235–243.
2. C. Bakkari, S. Kabbaj, and N. Mahdoua, Trivial extensions defined by Prüfer conditions J.
Pure App. Algebra, 214, (2010), 53–60.
3. H. S. Butts and W. Smith, Prüfer rings, Math. Z., 95, (1967), 196–211.
4. M. Chhiti, M. Jarrar, S. Kabbaj and N. Mahdou, Prüfer conditions in an amalgamated
duplication of a ring along an ideal, Comm. Algebra, 43, (2015), 249–261.
5. F. Couchot, Gaussian trivial ring ectensions and fqp-rings, Comm. Algebra, 43, (2015), 2863–
2874.
6. M. D’Anna, C. A. Finocchiaro, and M. Fontana, Amalgamated algebras along an ideal, in:
Commutative Algebra and Applications, Proceedings of the Fifth International Fez Conference
on Commutative Algebra and Applications, Fez, Morocco, 2008, W. de Gruyter Publisher,
Berlin, 2009, pp. 155–172.
7. M. D’Anna, C. A. Finocchiaro, and M. Fontana, Properties of chains of prime ideals in an
amalgamated algebra along an ideal, J. Pure Appl. Algebra, 214, (2010), 1633–1641.
8. M. D’Anna, C. A. Finocchiaro, and M. Fontana, New algebraic properties of an amalgamated
algebra along an ideal, Comm. Algebra, 44, (2016), 1836–1851.
9. M. D’Anna and M. Fontana, An amalgamated duplication of a ring along an ideal: the basic
properties, J. Algebra Appl., 6, No.3, (2007), 443–459.
10. C. A. Finocchiaro, Prüfer-like conditions on an amalgamated algebra along an ideal, Houston
J. Math., 40, (2014), 63–79.
11. R. Gilmer, Multiplicative Ideal Theory, New York, Dekker, 1972.
12. M. Griffin, Prüfer rings with zero divisors, J. Reine Angew. Math., 239/240, (1969), 55–67.
13. E. Houston and J. Taylor, Arithmetic properties in pullbacks, J. Algebra, 310, (2007), 235–
260.
14. C. U. Jensen, Arithmetical rings, Acta Math. Hungr., 17, (1966), 115–123.
15. W. Krull, Beitrage zur arithmetik kommutativer integritatsbereiche, Math. Z., 41, (1936),
545–577.
16. T. Lucas, Some results on Prüfer rings, Pasific J. Math., 124, (1986), 333–343.
17. H. Maimani and S. Yassemi, Zerodivisor graphs of amalgamated duplication of a ring along
an ideal, J. Pure Appl. Algebra, 212, (2008), 168–174.
18. M. A. S. Moutui and N. Mahdou, Amalgamated algebras along an ideal defined by Gaussian
condition, Journal of Taibah University for Science, 9, (2015), 373–379.
19. M. Nagata, Local Rings, Interscience, New York, 1962.
20. H. Prüfer, Untersuchungen uber teilbarkeitseigenschaften in korpern, J. Reine Angew. Math.,
168 (1932), 1-36.
21. P. Sahandi, N. Shirmohammadi, and S. Sohrabi, Cohen-Macaulay and Gorenstein properties
under the amalgamated construction, Comm. Algebra, 44, (2016), 1096–1109.
22. H. Tsang, Gauss’s Lemma Ph.D. thesis, University of Chicago, Chicago, 1965.
Department of Pure Mathematics, Faculty of Mathematical Sciences, University of
Tabriz, Tabriz, Iran.
E-mail address: [email protected]
Department of Pure Mathematics, Faculty of Mathematical Sciences, University of
Tabriz, Tabriz, Iran.
E-mail address: [email protected]
Department of Pure Mathematics, Faculty of Mathematical Sciences, University of
Tabriz, Tabriz, Iran.
E-mail address: [email protected]
| 0math.AC
|
1
Multi-objective Optimal Sizing and Energy
Management of Hybrid Energy Storage System for
Electric Vehicles
arXiv:1801.07183v1 [cs.SY] 19 Jan 2018
Huilong Yu, Member, IEEE, Federico Cheli, Francesco Castelli-Dezza, Member, IEEE, Dongpu
Cao, Member, IEEE, and Fei-Yue Wang, Fellow, IEEE
Abstract—Hybrid energy storage system (HESS) composed of
lithium-ion battery and supercapacitors has been recognized as
one of the most promising solutions to face against the high
cost, low power density and short cycle life of the battery-only
energy storage system, which is the major headache hindering the
further penetration of electric vehicles. In this work, the HESS
sizing and energy management problem of an electric race car
is investigated as a case study to improve the driving mileage
and battery cycle life performance. Compared with the existing
research, the distinctive features of this work are: (1) A dynamic
model and a degradation model of the battery are employed to
describe the dynamic behavior and to predict the cycle life of the
battery more precisely; (2) Considering the fact that the design
and control problems are coupled in most cases, in order to
achieve a global optimal design solution and an implementable
real-time energy management system, a Bi-level multi-objective
sizing and control framework based on non-dominated sorting
genetic algorithm-II and fuzzy logic control (FLC) is proposed
to size the HESS and to optimize the membership functions of a
FLC based EMS at the same time; (3) In order to improve the
optimization efficiency, a vectorized fuzzy inference system which
allows large scale of fuzzy logic controllers operating in parallel is
devised. At last, the Pareto optimal solutions of different HESSs
are obtained and compared to show the achieved enhancements
of the proposed Bi-level optimal sizing and energy management
framework.
Index Terms—Hybrid energy storage system, multi-objective
optimization, vectorized fuzzy interface, real-time energy management, electric vehicles.
I. I NTRODUCTION
I
N recent years, to face the challenge of air pollution, fossil
oil crisis, and greenhouse gas emissions, Electric vehicles (EVs) have gained unprecedented amount of attentions
from the governments, academia and industries all over the
world. After intensively developing over the last decades, the
worldwide promotion and application of EVs have reached
a considerable scale. However, the dynamic performance,
H. Yu was with the Department of Mechanical Engineering, Politecnico
di Milano, 20156, Milano, Italy. He is now with the Mechanical and
Mechatronics Engineering Department, University of Waterloo, N2L 3G1,
Waterloo, Canada (e-mail: [email protected]).
F. Cheli and F. Castelli Dezza are with the Department of Mechanical Engineering, Politecnico di Milano, 20156, Milano, Italy (e-mail: [email protected];[email protected]).
D. Cao is with the Mechanical and Mechatronics Engineering Department, University of Waterloo, N2L 3G1, Waterloo, Canada (e-mail:
dongp [email protected]).
FY. Wang is with the State Key Lab for Management and Control of
Complex Systems Institute of Automation, Chinese Academy of Sciences,
100190, Beijing, China (e-mail: [email protected]).
cost, durability of an EV are still significantly influenced
by the design, integration, and control of its energy storage
system (ESS) [1]. It is generally known that the batteryonly ESS with high cost and short cycle life has become
one of the biggest obstacles hindering further penetrations
of the EVs. Lithium-ion batteries with high energy density
and relatively good power density dominate the most recent
group of EVs in development, however, its degradation can
be accelerated when there is high peak discharging/charging
power demand during the acceleration/ deceleration process
[2]–[4]. Alternatively, supercapacitors (SCs) can tolerate much
more charging/discharging cycles and exhibit superior ability
to cope with high peak power, due to their specific energy
storage mechanism, but the low energy density hampers their
large scale application on EVs [5], [6]. A hybrid energy storage
system (HESS) composed by lithium-ion batteries and SCs
which can bridge the gap between them is considered as
one of the most promising solutions to solve the forgoing
problems entrenched in battery-only/SC-only energy [7]–[10].
The configuration of a HESS vary with different connections
of the battery, supercapacitor and DC/DC converter. The
employed HESS in this work is the most studied configuration
that using a bidirectional DC/DC converter to interface the
supercapacitor to the DC link of the battery in parallel,
where the voltage of supercapacitor can be used in a wide
range [8]. Existing research has demonstrated that HESS can
dramatically improve braking energy recuperation efficiency,
eliminate the need for battery over-sizing, and reduce the
weight and cost of the entire system [11]. However, the
HESS introduces complicated sizing, energy management and
integration problems [12].
The sizing problem of HESS aims to find the appropriate number of supercapacitor banks and battery cells that
minimizing the cost, mass and efficiency of the HESS or
maximize the battery cycle life. A multi-objective optimization
problem was formulated to minimize the overall HESS size
and maximize the battery cycle life, the formulated problem
was solved with the sample-based global search oriented
DIviding RECTangles (DIRECT) algorithm [13]. Ref. [14]
proposed a multi-objective sizing approach based on nondominated sorting genetic algorithm II (NSGA-II) for a semiactive HESS and obtained the Pareto frontier with battery
capacity loss and total cost of supercapacitors as objectives.
Convex optimization was introduced to solve the formulated
sizing and energy management problems of different kinds of
2
HESSs with weighted cost as objective [1], [15].
For the energy management strategy (EMS) of HESS, both
rule based approach and the optimization based approach are
widely investigated [16]. A time efficient utility functionbased control of a battery semi-active HESS was proposed
and carried out in [17] by formulating a weighted multiobjective optimization problem, then the formulated problem
is solved using the Karush-Kuhn-Tucker (KKT) conditions.
Ref. [18] proposed an energy management strategy for a HESS
based on fuzzy logic supervisory wavelet-transform frequency
decoupling approach, which aims to maintain the state of the
energy (SOE) of the supercapacitor at an optimal value, to
increase the power density of the ESS and to prolong the
battery lifetime. An explicit model predictive control (EMPC)
system for a HESS was proposed and validated in [19] to
make the HESS operating within specified constraints while
allocating high and low frequency current changes to the
supercapacitor and battery respectively. Ref. [20] developed
a real-time predictive power management control strategy
based on neural network and particle swarm optimization
algorithm to minimize the total cost including battery degradation and system energy. A variable charging/discharging
threshold method and an adaptive intelligence technique based
on historical data was proposed in [21] to improve the power
management efficiency and smooth the load of a HESS.
Two real-time energy management strategies based on KKT
conditions and neural network were investigated and validated
with experimental work in [22] to improve the battery state of
health performance of a HESS effectively.
The continuous previous efforts have improved the overall
performance of HESS considerably. However, most of the
aforementioned literature investigated the sizing and energy
management problems separately which can not obtain the
global optimal performance of the HESS since the design and
control problems of it are actually coupled in practice [23].
We can only obtain the sub-optimal solution when we try to
optimize one and fix the other. Some of the existing approach
in literature are off-line optimization methods which are quite
useful as the reference in designing real-time EMS but not
appropriate for real implementation, and most of the real-time
implementable EMSs are not able to achieve optimal performances. Besides, few existing work has considered the fact
that the total available amount of energy could be extracted
from the battery cells will vary when uses different discharging
C-rates in practical cases. This dynamic situation should be
considered during the optimization research especially when
the HESS is operating at limit conditions. Thus, it is necessary
to explore a framework that can achieve the global optimal
performance by incorporating both the sizing and real-time
feedback control problems of the HESS taking into account
the battery dynamics.
In order to force against the drawbacks of the state-of-theart methods, this work will explore the sizing and real-time
energy management problem of HESS as a coupled problem.
In particular, the HESS of an electric race car is investigated
as a case study. Although win the race is the only ultimate
goal on a circuit, we should try to minimize the cost for
a racing team and the environmental impact caused by the
waste battery as much as possible during a race or for the
offline training, which can match the spirit of the electric
racing better. Our goal of this work is to introduce the HESS
with proper sizing parameters and optimized EMS to improve
the cycle life of the battery without scarifying the mileage
of the electric race car too much. A multi-objective optimal
sizing and energy management framework supported by the
employed dynamic battery model, evaluation model and the
proposed vectorized fuzzy inference system is proposed in
this work. With this framework, one can obtain the Pareto
optimal solutions of the formulated multi-objective optimal
sizing and energy management problem and achieve both the
optimal sizing parameters and the static parameters of the EMS
at the same time for each Pareto optimal solution which are
ready for further real-time implementation.
The rest of this work is organized as follows: Section
2 describes the proposed Bi-level design and energy management framework, then presents the formulation of the
sizing and energy management problem. Section 3 elaborates
the modelling of the battery and supercapacitor. Section 4
details the devised FLC based on vectorized fuzzy inference
engine. In section 5, the simulation parameters and settings
are presented in detail. Section 6 and Section 7 illustrate the
obtained results and the concluding remarks respectively.
II. B I -L EVEL O PTIMAL D ESIGN AND C ONTROL
F RAMEWORK
The proposed Bi-level optimal design and control framework is presented as Figure 1. The power demand of the
driving profile Pdem , the battery state of charge xSOC and
supercapacitor state of energy xSOE are the inputs of the
FLC based EMS, while the outputs are requested power
from the battery Preqbat and from the supercapacitor Preqsc .
The outputs of the EMS are the inputs of the battery and
supercapacitor modeled in Section III, while the evaluation
indexes can be calculated with the outputs of the battery and
supercapacitor model.
Preqbat
xSOC
Battery
model
xSOC
Vb
Ibat
xSOE
EMS
xSOE
Pdem
Preqsc
SC
model
xmf
N sc
Evaluation
model
Optimal ?
Yes
End
Vsc
Isc
No
Multi-objective
optimization
Fig. 1. Framework of the Bi-Level design and control
The working scheme of the Bi-level optimal design and
control is illustrated as follows. Firstly, the multi-objective
algorithm will generate the sizing parameter matrix and the
corresponding static tuning parameter matrix of the energy
management system, in this work, the mentioned matrices
are respectively the number of supercapacitor banks and the
parameters of the membership functions (MFs) in different
3
pages of the optimization parameters. Secondly, the FLC based
EMS constructed with the new membership functions will
control the generated new HESS to output the demand power
from the battery and supercapacitor respectively. Then, the
maximum number of laps can be obtained when both the
battery and supercapacitor arrives at the minimum state of
charge values set in the constraints, while the capacity loss
of the battery is evaluated with the average current of the
battery during the whole scenario. There are quite a lot of
existing literature to model the capacity loss of the lithiumion battery. The capacity loss model is mostly validated by
discharging the battery with constant current C rate, and we
havent find any work that can predict the battery capacity
loss dynamically with validated experimental work. Thus, we
choose to estimate the capacity loss of the battery with average
load as many previous work did. When the Pareto-frontier of
the two evaluation indexes is obtained, the above iteration will
terminate, otherwise, it will continue.
The objective of this work is to find the optimal sizing
parameter Nsc and the parameter vector xmf defining the
membership functions which are respectively the key parameters of HESS design and the real-time FLC based EMS. The
optimized EMS will output the requested control command
series u(t) = [Preqbat , Preqsc ] to maximize the number of
traveled laps Jlaps and battery cycle life Jlifebat on a given
race circuit:
max J = [Jlaps (x(t), u(t), p), Jlifebat (x(t), u(t), p)]
(1)
subject to:
the first order dynamic constraints
ẋ(t) = f [x(t), u(t), t, p],
(2)
the boundaries of the state, control and design variables
xmin ≤ x(t) ≤ xmax
umin ≤ u(t) ≤ umax
(3)
pmin ≤ p ≤ pmax ,
the algebraic path constraints
g min ≤ g[x(t), u(t), t, p] ≤ g max ,
III. M ODELLING OF THE HESS
In this section, the dynamic characteristics of the implemented Lithium-ion battery are analyzed first, and a dynamic
battery model is employed after comparison. Then, the details
of the employed battery cycle life model are presented. A
simplified supercapacitor model is illustrated at the end of this
section.
A. Dynamic Battery model
In order to obtain the optimal sizing parameters and energy
management strategy for the HESS considering the characteristics of the battery in practical conditions, it is necessary to
implement a proper dynamic battery model that can describe
the battery dynamic behavior precisely. The most existing
battery models for the simulation of battery behavior basically
include the experimental, electrochemical and electrical ones
[24]–[26]. However, the experimental and electrochemical
model are not appropriate to simulate the battery dynamics
correctly. In comparison with the experimental and electrochemical battery model, specially devised electric circuit based
model are capable to represent the battery dynamics for the
purpose of state of charge (SOC) estimation. In this work, a
modified Shepherd model is employed to depict the dynamic
characteristics of the battery during charging and discharging
process [27]. The dynamic battery model are presented as
Equation (6) and Equation (7) with the assumption that the
internal resistance is constant and the thermal behavior of the
battery is neglected.
Discharge:
Vbatt = E0 −K
Qmax
Qmax
it−K
i−Ri+Ae(−B·it)
Qmax − it
Qmax − it
(6)
Charge:
(4)
and the boundary conditions
bmin ≤ b[x(t0 ), t0 , x(tf ), tf , p] ≤ bmax ,
variables x, control variables u, design parameters p and
the boundary constraints b will be presented in the following
paragraphs.
(5)
where ẋ is the first order derivative of the state variables, f is
the dynamic model, x, u, p are respectively the state, control
and design vector with their lower and upper bounds: xmin ,
umin , pmin and xmax , umax , pmax . While g and b are the
path and boundary equations respectively with their lower and
upper bounds g min , bmin and g max , bmax . The dimensions
of the input and output variables in Equations (2), (4) and (5)
are separately given as:
f : Rnx × Rnu × R × Rnp → Rnx
g : Rnx × Rnu × R × Rnp → Rng
b : Rnx × R × Rnx × R × Rnp → Rnb
In this work, the algebraic path constraint g is eliminated
by introducing a simple relaxation in Equation (16), the state
Qmax
Qmax
it−K
i−Ri+Ae(−B·it)
Qmax − it
it − 0.1Qmax
(7)
where Vbatt is the battery voltage (V ), E0 is the voltage
constant (V ), K is the polarization constant or polarization
resistance, Qmax is the total capacity, i is the battery current,
R is the internal resistance. The battery discharge (i > 0) or
charge (i < 0 ) it is denoted as
Z
it = idt.
(8)
Vbatt = E0 −K
The voltage amplitude A (V ), time constant inverse B
(Ah−1 ) of the exponential zone, the polarization resistance
K (Ω) and the voltage constant E0 (V) in Equations (6) and
(7) are calculated as follows [28]:
A = VF ull − Vexp
B=
3
Qexp
(9)
(10)
4
K=
(Vf ull − Vnom + A(exp(−BQnom ) − 1))(Q − Qnom )
Qnom
(11)
E0 = VF ull + K + R · i − A.
(12)
The typical discharge process of the lithium-ion battery is
divided into three stages as demonstrated in Figure 2. In the
initial stage, the battery voltage drops exponentially from the
fully charged voltage Vf ull , the voltage and extracted charge
at the end of this stage are Vexp and Qexp , respectively. The
voltage continues to decline slowly before reaching the nominal voltage Vnom and the corresponding discharge increases
to Qnom in the second stage. In the final stage, the battery
voltage drops abruptly to the minimum value and the discharge
reaches the maximum value Qmax .
in problems that are time-consuming and difficult to solve.
There should be more efforts put in improving the optimization
efficiency.
B. Battery Cycle Life Model
In recent years, substantial efforts have been made by both
the researchers and industries to develop models that can
predict the degradation of the lithium-ion batteries accurately
[29]–[31]. A revised semi-empirical model based on Arrhenius
equation is widely researched with large scale experiments,
and this model is mostly applied in optimization and control
problems related with batteries [29]. As presented from Equation (17) to Equation (20), the capacity loss of this model is
expressed as a function of the discharge current rate Crate ,
temperature T and ampere-hour throughout Ah .
−Ea
)(Ah )z
(17)
RT
where Qloss represents the battery capacity loss, A the preexponential factor, Ea the activation energy from Arrhenius
law (J), R is the gas constant of 8.314 , T is the absolute
temperature (K), Ah is the Ah-throughput, which represents
the amount of charge delivered by the battery during cycling.
The pre-exponential factor A in Equation (17) is proved
to be sensitive to the discharge current rate Crate with large
scale experiments in [32] , and it is fitted with the format as
Equation (18) in [33].
Qloss = A exp(
ln A = a · exp(−b · Crate ) + c
The activation energy can be fitted as a linear function of
discharge current rate Ref. [32],
Typical discharge curve
Fig. 2. Typical discharge curve of lithium-ion battery.
The state of charge of the battery xsoc and its derivative ẋsoc
is denoted as Equation (13) and Equation (14) respectively.
Z tf
1
xsoc = 100(1 −
idt)
(13)
3600Qmax 0
ẋsoc
1
=−
i
36Qmax
(14)
The charging/discharging i is denoted as
Preqbat
N V η , Preqbat ≥ 0
bat bat AD
(15)
i=
Preqbat ηAD
,
Preqbat < 0
Nbat Vbat
where ηAD is the efficiency of the DC/AC converter taking
into account of the motor efficiency as a constant value. In
this work, the number of battery cells Nbat are determined by
the available total mass of the HESS mHESS and the number of
the supercapacitor banks Nsc , as shown in Equation (16). The
total mass of the HESS is fixed in this work considering the
fact that the mass of a race car is strictly limited in general.
Nbat = b(mHESS − Nsc mbank )/mcell c
(18)
(16)
Although the modified Shepherd model can represent the
dynamic characteristics of the battery, it has not been employed so much in the existing literature related with the HESS
design and control due to its complicity, which usually results
Ea = d + e · Crate
(19)
where a, b, c, d, e are the correction parameters of the battery
cycle life model.
The Ah-throughput can be expressed as
Ztf
Ah =
i
dt
3600
(20)
0
where i is the discharge current, tf is the end time of the
current profile.
C. Supercapacitor Model
In this work, the capacity fading of the supercapacitor
is neglected considering the fact that it has much longer
cycle life than lithium-ion batteries. The supercapacitor model
is simplified to a series connection of a resistance and a
supercapacitor bank [6]. Also, the efficiency of the DC/DC
converter between the supercapacitor and the bus is assumed
to be a constant value of 0.95. The recursive supercapacitor
model is deduced as
p
Vct − Vct2 − 4Rsct Preqsc /(ηAD ηdc )
Preqsc ≥ 0
−
2Csct Rsct
V̇ct =
p
V − Vct2 − 4Rsct Preqsc ηAD ηdc
− ct
Preqsc < 0
2Csct Rsct
(21)
5
xSOE =
Vct 2
2
Vctmax
(22)
ẋSOE =
2Vct V̇ct
2
Vctmax
(23)
and big respectively. The basic idea of the fuzzy rule is to
utilize the the supercapacitor as a buffer to reduce the high
peak power impact on battery and absorb more regenerative
braking power.
Vct = Vc Nsc
(24)
B. Membership functions
Csct =Cbank /Nsc
(25)
Rst = Nsc Rs
(26)
where Vct is the total open circuit voltage of the supercapacitor
pack assuming that all banks have a uniform behavior, tk+1
is the time at step k + 1, Rst is the total equivalent series
resistance, Preqsc is the demand power from the supercapacitor, ηdc is the efficiency of the DC/DC converter, Csct is
the total capacity, SOE is the state of energy, Vctmax is the
initial open circuit voltage, Vc is the open circuit voltage of
one supercapacitor, Nsc is the total number of the banks, Rs
is the series resistance of one supercapacitor.
The actual total output power of the supercapacitor is
represented as
p
Vct − Vct2 − 4Rst Preqsc /ηAD ηdc
.
(27)
Psc = Vct ·
2Rst
IV. FLC BASED ON V ECTORIZED F UZZY I NFERENCE
E NGINE
In this work, the EMS is developed based on fuzzy logic
control (FLC), which has the features of real-time, adaptive
and intelligent. It allows different operators to merge nonlinearities and uncertainties in the best way and incorporate
heuristic control in the form of if-then rules. The developed
FLC in this section are composed of the if-then fuzzy rules,
fuzzification, fuzzy inference engine and defuzzification modules. To speed up the optimization and take the advantage
of the powerful matrix processing capability of MATLAB, a
vectorized fuzzy inference system (VFIS) presented in Figure
3 is developed for the first time according to the state-ofthe-art literature. The developed VFIS is capable to handle
Np × Ninp dimensional inputs with Np pages of membership
functions each time. This means that Np fuzzy controllers
(can be hundreds of thousands depends on the performance
of the utilized CPU) can work at the same time with the
same page number of inputs and outputs. The following
paragraph will present the detail of fuzzy rules, membership
functions, vectorized fuzzification, fuzzy inference engine, and
defuzzification operations of the developed vectorized FLC.
The concept of membership functions was introduced by
Zadeh in the first paper on fuzzy sets [34]. A membership
function is a curve or a function that defines how each point
of the input variables is mapped to a membership value
between 0 and 1. It is quite challenging to design the optimal
MFs for each HESS manually according to the engineering
experiences. Besides, considering that the performance of the
FLCs are sensitive to their MFs, different MFs of the FLCs
with the same fuzzy rules should be devised for different sizes
of HESS. Based on these considerations, the parameters of the
MFs are selected as parts of the parameters to be optimized
in this work.
The trapezoidal-shaped membership function is selected for
the fuzzy inference engine based on the considerations that it
has high flexibility, e.g., as shown in Figure 5, when a = b,
the trapezoid MF will change its shape to triangle [12].
C. Vectorized Fuzzification
During the fuzzification stage, the input variables are identified to the fuzzy sets (membership functions) they belong
to and the respective degree of membership to each relevance
will be assigned. For a FIS with trapezoidal shaped MFs and a
number of Ninp inputs, the fuzzy sets of each can be described
with a matrix S k = [ak , bk , ck , dk ] ∈ RNp ×Nti,k ×4 . Np is
the total page number of the inputs; Nti,k is the number of
fuzzy linguistic sets of state input k, k ∈ {1, 2, ..., Ninp } and
a, b, c, d are the variables that define one trapezoid in Figure
5. The input matrix is denoted as xk ∈ RNp , for Xk , its
membership matrix µk ∈ RNp ×Nti,k can be denoted as:
Xk − ak
µk (ak ≤ Xk < bk ) = b − a
k
k
µk (bk ≤ Xk ≤ ck ) = I
(28)
d
−
X
k
µk (ck < Xk ≤ dk ) = k
d k − ck
where ak , bk , ck , dk , I belong to RNp ×Nti,k , and Xk is denoted as:
Xk = [xk , xk , ..., xk ] ∈ RNp ×Nti,k
|
{z
}
(29)
Nti,k
A. Fuzzy rules
Fuzzy rules are a set of if-then linguistic rules used to
formulate the conditional relationships that comprise a fuzzy
logic controller, for instance, a fuzzy rule can be: if SOC is
Small and SOE is Big and Preq is Positive big then Psc is
Positive big. It is reasonable to devise the same if-then rules
for the control of different sizes of HESSs since the control
objectives of all the HESSs are the same in this work. The
developed fuzzy rules are demonstrated as Figure 4, where the
labels N, P, S, M, B means negative, positive, small, medium
The membership array U for input X can be constructed as:
U = {µ1 , ..., µk , ..., µNinp } ∈ RNp ×Nti,k ×Ninp
(30)
D. Vectorized Fuzzy Inference Engine
Fuzzy inference is the way of mapping an input space to an
output space using fuzzy logic. A FIS tries to formalize the
reasoning process of human language by means of fuzzy logic
(the built fuzzy If-Then rules). The process of fuzzy inference
involves all of the MFs, If-Then rules, linguistic variables of
6
Page 1
Page 2
Membership function optimization algorithm
c(Nto,Np)
d(Nto,Np)
Page Ns
b(1,Np,1) c(1,Np,1) b(2,Np,1) c(2,Np,1) b(Nti,k,Np,1) c(Nti,k,Np,1)
d(Nti,k,Np,2)
a(Nti,k,Np,2)
……
……
x
d(1,Np,Ninp) a(2,Np,Ninp) d(2,Np,Ninp) a(Nti,k,Np,Ninp)
Np×Ninp
(Np,Ninp)
Inputs
UoNto×Np (1, Np)
……
a(1,Np,Ninp)
……
b(1,Np)
b(1,Np,Ninp)c(1,Np,Ninp)b(2,Np,Ninp)c(2,Np,Ninp) b(Nti,k,Np,Ninp) c(Nti,k,Np,Ninp)
xNp×Ninp(Np,Ninp)
d(1,Np) a(2,Np) d(2,Np)
Output fuzzy
set of page Np
d(1,Np,1) a(2,Np,2) d(2,Np,2)
xNp×Ninp(Np,2)
d(Nti,k,Np,Ninp)
Input fuzzy sets of page Np
Fuzzification
Fuzzy Rules
xNp×Ninp(Np)
a(1,Np)
a(1,Np,1)
c(2,Np)
c(Nti,k,Np,2)
……
b(2,Np)
b((1,Np,1) c(1,Np,1) b(2,Np,2) c(2,Np,2) b(Nti,k,Np,2)
xNp×Ninp(Np,2)
b(Nto,Np)
d(Nti,k,Np,1)
c(1,Np)
a(Nti,k,Np,1)
xNp×Ninp(Np,1)
UoNto×Np (2, Np)
UoNto×Np (Nto, Np )
a(1,Np,1) d(1,Np,1) a(2,Np,1) d(2,Np,1)
Fuzzy Inference Engine
xNp×Ninp(Np,1)
a(Nto,Np)
……
Defuzzification
Outputs
Fig. 3. Framework of the vectorized FLC
a=b
c d
a=b=c
d
PB
PM
PS
NS
NM
NB
B
B
Change a,b,c,d
M
M
S
S
a
b
c
a b
d
c=d
Trapezoidal MF
NB
NM
NS
PS
PM
PB
a b=c d
Fig. 4. Fuzzy rules for the EMS of the HESS
the inputs and outputs. Mamdani’s fuzzy inference method is
the most commonly seen fuzzy methodology. The search-able
fuzzy inference engine is able to map only one page of the
inputs to one page of the outputs. This section will give an
elaborate description of the developed powerful VFIS which
allows a large number of FLCs operating in parallel based on
Mamdani’s fuzzy inference method.
The linguistic variables are programmed with their integer
indexes from the smallest to the biggest in turn in this work.
For instance, the fuzzy sets {NB, NM, NS, PS, PM, PB} of
the third input in Figure 4 are correspondingly mapped to
{1, 2, ..., Nti,k }, here Nti,k = 6, k = 3. The fuzzy rule matrix
< ∈ RNr ×(Ninp +No ) is constructed with the mapped integer
a b=c=d
Fig. 5. Flexible trapezoidal-shaped membership function
indexes, Nr is the number of fuzzy rules and No is the number
of outputs. For instance:
Rule : xSOC
<(Nr ) :
1
xSOE
3
Preq
6
Psc
6
(31)
where <(Nr ) denotes the fuzzy rule Nr , it means the rule
like: if SOC is Small and SOE is Big and Preq is Positive big
7
then Psc is Positive big. The working scheme of the VFIS is
illustrated as follows:
1) Repeatedly copy the membership matrix µk into Nr blocks,
and we can obtain:
µtemp
= [µk ; µk ; ...; µk ], µtemp
∈ RNp ×Nti,k ×Nr
k
k
|
{z
}
Lin
2
2
..
.
···
...
..
.
Nti,k
Nti,k
..
.
1
2
···
Nti,k
,···
1
1
..
.
2
2
..
.
···
...
..
.
Nti,k
Nti,k
..
.
1
2
···
Nti,k
{z
Np
N
r
}
(33)
3) Repeatedly copy the kth column of the rule matrix < ∈
RNr ×(Ninp +No ) into Np × Nti,k block arrangement <temp ∈
RNp ×Nti,k ×Nr , k ∈ {1, 2, . . . , Ninp }:
<k , <k , . . . , <k
<k , <k , . . . , <k
<temp = .
Np
(34)
.
.
..,
..
..,
...,
<k , <k , . . . , <k
|
{z
}
Nti,k
4) Get the effective membership matrix µeff,k for input k ∈
{1, 2, . . . , Ninp }:
µeff,k = µk (Lin == <temp ), µeff,k ∈ RNp ×Nti,k ×Nr
Ueff,o = Uo,temp (Lo == <otemp ), Ueff,o ∈ RNp ×Nr ×Nto
(41)
11) Merge the membership matrix of the output fuzzy sets in
all the fuzzy rules
Uo,final =
E. Vectorized Defuzzification
The purpose of defuzzification process is to produce a
quantifiable result in crisp logic based on the given fuzzy sets
and corresponding membership degrees. The defuzzification
process based on center of gravity method is demonstrated as
Figure 6.
UoNto×Np (1, Np)
a(1,Np)
b(1,Np)
k=1
2
2
..
.
···
...
..
.
Nto
Nto
..
.
1
2
···
Nto
,···
{z
Np
1
1
..
.
2
2
..
.
···
...
..
.
Nto
Nto
..
.
1
2
···
Nto
Nr
}
(38)
8) Repeatedly copy the column of output fuzzy sets in the
rule matrix < ∈ RNr ×(Ninp +No ) into a Np × Nto block
arrangement <otemp ∈ RNp ×Nr ×Nto , Nto is the number of
fuzzy linguistic sets of output :
<No , <No , . . . , <No
<No , <No , . . . , <No
<otemp = .
Np
(39)
.
.
..,
..
..,
...,
<No , <No , . . . , <No
|
{z
}
Nto
c(1,Np)
b(2,Np)
c(2,Np)
d(1,Np)
c(1,Np)
a(2,Np)
b(Nto,Np)
b(2,Np)
d(2,Np)
a(Nto,Np)
c(2,Np)
b(Nto,Np)
Np)
c(Nto,Np)
d(Nto,Np)
c(Nto,Np)
……
a(1,Np)
b(1,Np)
7) Create index matrix Lo ∈ RNp ×Nr ×Nto for output the fuzzy
sets:
1
1
..
.
UoNto×Np (2,
……
6) Get the mapped membership matrix Uo for the output fuzzy
sets:
Ninp
Uo = ∩ Uin (k), Uo ∈ RNp ×Nr
(37)
b(1,Np)
……
Ninp
(42)
By the above calculation, the membership of each trapezoid
of the output fuzzy set is obtained as Uo,final , and the next step
is the defuzzification.
Uin = { ∪ µeff,k (j), ∪ µeff,k (j), ..., ∪ µeff,k (j)} (36)
j=1
j=1
j=1
|
{z
}
Lo =
|
Uo,eff (i), Uo,final ∈ RNp ×Nto
i=1
Nti,k
Nti,k
Nr
[
(35)
5) Combine and get the final membership matrix Uin ∈
RNp ×Nr ×Ninp for all the input X:
Nti,k
Nto
10) Get the effective membership matrix Ueff,o of all the output
fuzzy sets:
2) Create index matrix Lin ∈ RNp ×Nti,k ×Nr for input k:
1
1
..
.
Uo,temp = {Uo , Uo , ..., Uo }, Uo,temp ∈ RNp ×Nr ×Nto (40)
|
{z
}
(32)
Nr
=
|
9) Repeatedly copy the membership matrix of the output fuzzy
sets Uo ∈ RNp ×Nr into Nto blocks Uo,temp :
d(1,Np)
c(1,Np)
a(2,Np)
b(2,Np)
d(2,Np)
a(Nto,Np)
c(2,Np)
b(Nto,Np)
d(Nto,Np)
c(Nto,Np)
……
UoNto×Np (Nto, Np )
a(1,Np)
d(1,Np)
a(2,Np)
d(2,Np)
a(Nto,Np)
d(Nto,Np)
Output fuzzy set of page Np
∑
Fig. 6. Defuzzification process based on center of gravity method
The procedure of the vectorized defuzzification is elaborated
as followings:
1) Discrete the output fuzzy sets into Ndis parts xo ∈ RNdis
from its minimum value xo,min to the maximum one xo,max ,
xo = [xo,min : (xo,max − xo,min )/(Ndis − 1) : xo,max ]
(43)
8
where the value of Ndis affects the accuracy of the crisp
output, for instance, the increasing of Ndis will improve the
precision but will increase the computational burden.
2) Repeatedly copy xo ∈ RNdis and output fuzzy set
S o = [ao , bo , co , do ] ∈ RNto ×Np ×4 , we can obtain xo,temp ∈
RNto ×Np ×Ndis and S o,temp ∈ RNto ×Np ×Ndis respectively:
xo , xo , . . . , xo
xo , xo , . . . , xo
xo,temp = .
Np
(44)
..
.
..,
., . . . , ..
xo , xo , . . . , xo
|
{z
}
vector is p = {Nsc , xmf }. As demonstrated in Figure 7, there
are 28 parameters of the devised membership functions plus
one design parameter of the HESS in one page of parameters
to be optimized. The design vector p is constrained by defining
pmin and pmax .
1
S
M
B
0.5
0
0
100
1
S
Nto
M
B
0.5
S o,temp = [S o , S o , ..., S o ]
|
{z
}
(45)
0
Ndis
0
3) Calculate the membership matrix of xo,temp
based on the output fuzzy set S o,temp
=
[ao,temp , bo,temp , co,temp , do,temp ] ∈ RNto ×Np ×Ndis ×4 :
xo,temp − ao,temp
µo (ao,temp ≤ xo,temp < bo,temp ) =
bo,temp − ao,temp
µo,temp (bo,temp ≤ xo,temp ≤ co,temp ) = I
d
− xo,temp
µo (co,temp < xo,temp ≤ do,temp ) = o,temp
do,temp − co,temp
(46)
4) Repeatedly copy the membership matrix Uo,f inal in to Ndis
blocks Uo,temp ∈ RNto ×Np ×Ndis :
Uo,temp = [Uo,f inal , Uo,f inal , ..., Uo,f inal ]
|
{z
}
(47)
Ndis
5) Find the effective membership matrix:
\
Ueff,o = Uo,temp µo , Ueff,o ∈ RNto ×Np ×Ndis
(48)
6) Merge the membership matrix obtained in last step:
Nr
[
Uo,x =
Ueff,o (i), Uo,x ∈ RNp ×Ndis
(49)
i=1
7) Calculate the crisp output matrix for all the input matrices:
NP
dis
y=
xo (i) ◦ Uo,x (xo (i))
i=1
NP
dis
, y ∈ RNp
(50)
Uo,x (xo (i))
i=1
In order to design the fuzzy rules and membership functions
conveniently, the devised vectorized FLC modules illustrated
above are developed in MATLAB with standard and user
friendly interfaces.
V. S IMULATION PARAMETERS AND S ETTINGS
The state variables includes the battery state of charge
xSOC and state of energy of the supercapacitor xSOE , x =
[xSOC , xSOE ]. The control variable output by the FLC in
this work is the requested power from the supercapacitor
u = Preqsc , the demand power from the battery can be
calculated by Preqbat = Pdem − Preqsc . The design parameter
100
1
0.5
0
-200
200
1
0.5
0
-200
Fig. 7. The design parameters of the MFs for the EMS
The operating profile of an electric race car is of great
difference with the one of conventional electric vehicle running
on a city road. Thus, standard driving cycles are not suitable
for the research on electric race car. The real driving cycle of
a race car in Nurburgring circuit is chosen as the test scenario.
The demand driving/braking power is calculated by Equation
(51). The corresponding velocity profile, acceleration profile
and demand power are demonstrated as Figure 8.
1
Pdem = ( ρCd Av 2 + f mv g + mv a)v
2
(51)
The detail simulation parameters of the race car, 53 Ah high
energy lithium-ion battery, 2.85V/3400F high performance
supercapacitor and the converters are illustrated in Table I.
In the FLC based EMS, the SOE and current of the
supercapacitor are constrained between 0.1 and 0.99, -2000
A and 2000A respectively. While the SOC of the lithiumion battery is constrained between 0.2 and 0.9, the current is
regulated by adjusting the requested power from the battery.
When the lithium-ion battery is exhausted, the simulation of
one iteration will be terminated and the objective functions will
be correspondingly evaluated. The temperature is for sure very
important in any kind of vehicle equipped with batteries since
it can affect the performance of the batteries directly. However,
it is very difficult to model the heat generation, dispassion and
the thermal control system of the energy storage system on an
9
300
200
100
0
10
0
20
40
60
80
100
120
0
20
40
60
80
100
120
0
20
40
60
80
100
120
0
-10
-20
200
0
-200
-400
Fig. 8. Power demand in Nurburgring circuit
TABLE I
PARAMETER VALUES OF THE SIMULATION
Parameters
Vehicle mass (kg)
Aerodynamics coefficient (h2 N/km2 )
Rolling resistance coefficient
Mass of the battery cell (kg)
Voltage constant of the battery cell (V )
Total capacity of the battery cell(Ah)
Polarization resistance of the battery cell (Ω)
Internal resistance of the battery cell (Ω)
Voltage amplitude of the battery cell (V )
Time constant inverse of the battery cell (Ah−1 )
Fitting parameter of pre-exponential factor
Fitting parameter of pre-exponential factor
Fitting parameter of pre-exponential factor
Fitting parameter of activation energy
Fitting parameter of activation energy
Mass of the supercapacitor bank (kg)
Supercapacitor bank capacity(F )
Supercapacitor equivalent series resistance (Ω)
DC/DC converter efficiency
DC/AC converter efficiency
Symbol
mv
ρCd A
f
mcell
E0
Qmax
K
R
A
B
a
b
c
d
e
mbank
Cbank
Rs
ηdc
ηAD
From the sizing point of view, using different number of
supercapacitors means different compromises between high
power density and high energy density. As it is demonstrated
in Figure 9, utilizing more supercapacitors can assist to
reduce the average current of the Lithium-ion battery which is
beneficial for longer cycle life of the battery, but cut down the
energy density of the HESS which results in shorter driving
mileage. When less supercapacitors are used, the results will
be opposite. It is also observed from Figure 9 that HESS with
the same design solutions (makers filled with the same color)
may achieve different values of both objective functions, which
means that for the same HESS with uniform fuzzy rules,
the parameters of the membership functions will determine
whether we can achieve the Pareto optimal solutions. Thanks
to the proposed Bi-level optimal sizing and control framework,
the corresponding sizing parameter Nsc of each HESS and the
membership function parameters xmf of the related EMS are
coupled and obtained at the same time for all the solutions
including those on the Pareto frontier.
Value
570
0.075
0.016
1.15
3.43
55
8.85×10−5
1.33×10−3
0.761
0.040
1.345
0.2563
9.179
46868
-470.3
0.52
3400
2.2×10−4
0.95
0.96
Fig. 9. Multi-objective sizing and control solutions when mHESS = 320kg
electric vehicle precisely. Actually, it is reasonable to assume
that the temperature is controlled at a constant value (23 o C)
by adjusting the thermal control system [1], [35].
In this work, a controlled elitist NSGA which is a variant of
NSGA-II [36] is implemented to solve the multi-objective optimization problem. Instead of only choosing the top-ranking
nondominated fronts, the controlled elitist GA also favors
individuals that can assist to improve the diversity of the
population even if their fitness values are relatively lower.
VI. R ESULTS
In this section, the results of the multi-objective optimal
sizing and control of the HESS are presented and analyzed in
detail. Figure 9 presented the achieved results when the total
mass of the HESS is limited at 320 kg. The population size
is set as 500 in the NSGA-II optimization algorithm, and the
optimization is terminated after about 13 hours in a ThinkPad
T470P laptop with Intel(R) Core(TM) i5-7300HQ CPU @2.50
GHz CPU and 16GB RAM, the number of total iterations is
1839.
Moreover, this work has investigated the optimal sizing and
control results of HESSs with different total mass. From Figure
10, we can drawn the following basic conclusions: 1) HESSs
with smaller total mass will cover fewer number of available
laps, but the available cycle life of the battery are longer due
to their shorter operating mileage; 2) We can achieve a pretty
decent compromised solution that can enhance both objective
functions with only about 40 supercapacitor banks and the
optimized membership functions.
In order to analyze the reason of the exhibited advantages of
the proposed Bi-level optimal sizing and control framework,
one solution from the Pareto frontier in Figure 9 (Nsc = 32) is
compared with the solution with same sizing parameter but the
initial devised membership functions. Figure 11 demonstrates
the initial and optimized membership functions with the dotted
lines and solid lines respectively.
The achieved available number of laps of the initial and
optimized solutions are very similar which are respectively
17.88 and 17.98. This is mostly due to the fact that the two
cases are implemented with the same HESS and the available
10
140
5800
120
6000
100
6200
80
6400
60
6600
40
supercapacitors very fast at a few seconds after starting the
operation and the average SOE is under 20% during the
simulation which is not capable to provide long-time high peak
power to protect the battery. While EMS with the optimized
membership functions tends to maintain the SOE of the
supercapacitors above 50%, which helps to play the role of
shaving the peak and filling the valley very well during the
whole driving profile. For instance, the curves in the dotted
box in Figure 12 (c) demonstrate that the requested power from
the battery is less after optimizing the MFs since the SOE is
maintained at a relatively high level due to the optimized EMS.
6800
20
19
18
17
16
15
14
13
12
11
10
200
150
100
Fig. 10. Pareto optimal solutions for HESS with different mass mHESS
50
1
0
M
S
B
0
20
40
60
80
100
120
140
160
180
200
0
20
40
60
80
100
120
140
160
180
200
0
20
40
60
80
100
120
140
160
180
200
0.5
100
50
0
0
10
20
30
40
50
60
70
80
90
100
0
1
-50
S
M
B
-100
0.5
-150
0
0
10
20
30
40
50
60
70
80
90
100
1
100
80
NB
NM
NS
PS
PM
PB
60
0.5
40
0
-200
20
-150
-100
-50
0
50
100
150
200
0
1
NB
NM NS
PS
PM
PB
0.5
Fig. 12. Pareto optimal solutions for HESS with different sizes
0
-200
-150
-100
-50
0
50
100
150
200
VII. C ONCLUSIONS
Fig. 11. The initial and optimized membership functions when Nsc = 32
mileage is mainly determined by the sizing parameters rather
than the control parameters. However, the available cycle life
of the battery are different which are respectively 6082 and
6463. This means that HESS with the optimized membership
functions improved the battery cycle life by 6.3%. Figure
12 presents the interested variables between 0-200s, as it
is illustrated in Figure 12 (a) and Figure 12 (b), the EMS
with initial devised membership functions tends to request
more high peak power from the battery and less from the
supercapacitors which will accelerate the degradation of the
battery. This phenomenon can be explained with the curve
of SOE in Figure 12 (c). We can see that EMS with the
initial devised membership functions tends to exhaust the
More supercapacitors do not always guarantee a better
overall performance especially when the total mass of the
HESS is limited. However, we are able to obtain a pretty
good balanced performance with less supercapatitors and the
optimized EMS by the proposed optimization framework. The
proposed Bi-level optimal sizing and control framework in this
work makes it possible to obtain the global optimal solutions
since it enables the optimization algorithm to search both the
design and control parameters simultaneously. The user could
choose the favored sizing solution from the obtained Pareto
frontier packaged with the optimal membership functions
based on a preferred compromise between the two objectives.
The obtained global optimal sizing parameters and optimal
parameters of the real-time controller on the Pareto frontier
can be put into real-time implementations. In addition to the
Bi-level optimal sizing and control framework, the devised
11
vectorized fuzzy logic controller with standard interfaces can
be used in other kinds of real time feedback control problems,
in particular, it can assist to dramatically improve the computational efficiency when need to optimize the parameters of
fuzzy logic controller.
R EFERENCES AND F OOTNOTES
R EFERENCES
[1] X. Hu, L. Johannesson, N. Murgovski, and B. Egardt, “Longevityconscious dimensioning and power management of the hybrid energy
storage system in a fuel cell hybrid electric bus,” Applied Energy, vol.
137, pp. 913–924, 2015.
[2] G. Suri and S. Onori, “A control-oriented cycle-life model for hybrid
electric vehicle lithium-ion batteries,” Energy, vol. 96, pp. 644–653,
2016.
[3] F. Sun, X. Hu, Y. Zou, and S. Li, “Adaptive unscented Kalman filtering
for state of charge estimation of a lithium-ion battery for electric
vehicles,” Energy, vol. 36, no. 5, pp. 3531–3540, 2011.
[4] Y. Zou, X. Hu, H. Ma, and S. E. Li, “Combined state of charge and
state of health estimation over lithium-ion battery cell cycle lifespan for
electric vehicles,” Journal of Power Sources, vol. 273, no. Supplement
C, pp. 793 – 803, 2015.
[5] A. Burke, “Ultracapacitors: why, how, and where is the technology,”
Journal of power sources, vol. 91, no. 1, pp. 37–50, 2000.
[6] L. Zhang, X. Hu, Z. Wang, F. Sun, and D. G. Dorrell, “A review of supercapacitor modeling, estimation, and applications: A control/management
perspective,” Renewable and Sustainable Energy Reviews, vol. 81, pp.
1868 – 1878, 2018.
[7] S. M. Lukic, S. G. Wirasingha, F. Rodriguez, J. Cao, and A. Emadi,
“Power management of an ultracapacitor/battery hybrid energy storage
system in an hev,” in 2006 IEEE Vehicle Power and Propulsion Conference, Sept 2006, pp. 1–6.
[8] J. Cao and A. Emadi, “A new battery/ultracapacitor hybrid energy
storage system for electric, hybrid, and plug-in hybrid electric vehicles,”
IEEE Transactions on power electronics, vol. 27, no. 1, pp. 122–132,
2012.
[9] T. Ma, H. Yang, and L. Lu, “Development of hybrid batterysupercapacitor energy storage for remote area renewable energy systems,” Applied
Energy, vol. 153, pp. 56–62, 2015.
[10] O. Ahmed and J. Bleijs, “An overview of dcdc converter topologies
for fuel cell-ultracapacitor hybrid distribution system,” Renewable and
Sustainable Energy Reviews, vol. 42, no. Supplement C, pp. 609 – 626,
2015.
[11] C. G. Hochgraf, J. K. Basco, T. P. Bohn, and I. Bloom, “Effect of
ultracapacitor-modified PHEV protocol on performance degradation in
lithium-ion cells,” Journal of Power Sources, vol. 246, pp. 965–969,
2014.
[12] H. Yu, D. Tarsitano, X. Hu, and F. Cheli, “Real time energy management
strategy for a fast charging electric urban bus powered by hybrid energy
storage system,” Energy, vol. 112, pp. 322–331, 2016.
[13] J. Shen, S. Dusmez, and A. Khaligh, “Optimization of sizing and battery
cycle life in battery/ultracapacitor hybrid energy storage systems for
electric vehicle applications,” IEEE Transactions on Industrial Informatics, vol. 10, no. 4, pp. 2112–2121, 2014.
[14] Z. Song, J. Li, X. Han, L. Xu, L. Lu, M. Ouyang, and H. Hofmann,
“Multi-objective optimization of a semi-active battery/supercapacitor
energy storage system for electric vehicles,” Applied Energy, vol. 135,
no. 0, pp. 212–224, 2014.
[15] X. Hu, N. Murgovski, L. M. Johannesson, and B. Egardt, “Comparison
of three electrochemical energy buffers applied to a hybrid bus powertrain with simultaneous optimal sizing and energy management,” IEEE
Transactions on Intelligent Transportation Systems, vol. 15, no. 3, pp.
1193–1205, 2014.
[16] S. F. Tie and C. W. Tan, “A review of energy sources and energy
management system in electric vehicles,” Renewable and Sustainable
Energy Reviews, vol. 20, pp. 82 – 102, 2013.
[17] H. Yin, C. Zhao, M. Li, and C. Ma, “Utility Function-Based RealTime Control of A Battery Ultracapacitor Hybrid Energy System,” IEEE
Transactions on Industrial Informatics, vol. 11, no. 1, pp. 220–231,
2015.
[18] S. Dusmez and A. Khaligh, “A supervisory power splitting approach
for a new ultracapacitor battery vehicle deploying two propulsion
machines,” IEEE Transactions on Industrial Informatics, vol. 10, no. 3,
pp. 1960–1971, 2014.
[19] B. Hredzak, V. G. Agelidis, and G. Demetriades, “Application of
explicit model predictive control to a hybrid battery-ultracapacitor power
source,” Journal of Power Sources, vol. 277, pp. 84–94, 2015.
[20] Q. Zhang, W. Deng, and G. Li, “Stochastic Control of Predictive Power
Management for Battery/Supercapacitor Hybrid Energy Storage Systems
of Electric Vehicles,” IEEE Transactions on Industrial Informatics, pp.
1–1, 2017.
[21] K. Jia, Y. Chen, T. Bi, Y. Lin, D. Thomas, and M. Sumner, “HistoricalData-Based Energy Management in a Microgrid With a Hybrid Energy
Storage System,” IEEE Transactions on Industrial Informatics, vol. 13,
no. 5, pp. 2597–2605, 2017.
[22] J. Shen and A. Khaligh, “Design and real-time controller implementation for a battery-ultracapacitor hybrid energy storage system,” IEEE
Transactions on Industrial Informatics, vol. 12, no. 5, pp. 1910–1918,
2016.
[23] H. Yu, F. Castelli-Dezza, and F. Cheli, “Optimal powertrain design and
control of a 2-iwd electric race car,” in 2017 International Conference
of Electrical and Electronic Technologies for Automotive, June 2017,
pp. 1–7.
[24] Y. Cao, R. C. Kroeze, and P. T. Krein, “Multi-timescale parametric
electrical battery model for use in dynamic electric vehicle simulations,”
IEEE Transactions on Transportation Electrification, vol. 2, no. 4, pp.
432–442, 2016.
[25] S. J. Moura, F. B. Argomedo, R. Klein, A. Mirtabatabaei, and M. Krstic,
“Battery state estimation for a single particle model with electrolyte
dynamics,” IEEE Transactions on Control Systems Technology, vol. 25,
no. 2, pp. 453–468, 2017.
[26] H. Chaoui and C. C. Ibe-Ekeocha, “State of charge and state of health
estimation for lithium batteries using recurrent neural networks,” IEEE
Transactions on Vehicular Technology, vol. 66, no. 10, pp. 8773–8783,
2017.
[27] O. Tremblay, “Experimental validation of a battery dynamic model for
ev applications experimental validation of a battery dynamic model for
ev applications,” in 24th International Battery, Hybrid and Fuel Cell
Electric Vehicle Symposium and Exhibition 2009, EVS 24, vol. 2, 2009,
pp. 930–939.
[28] O. Tremblay, L.-a. Dessaint, and a. I. Dekkiche, “A Generic Battery
Model for the Dynamic Simulation of Hybrid Electric Vehicles,” 2007
IEEE Vehicle Power and Propulsion Conference, no. V, pp. 284–289,
2007.
[29] J. Wang, P. Liu, J. Hicks-Garner, E. Sherman, S. Soukiazian, M. Verbrugge, H. Tataria, J. Musser, and P. Finamore, “Cycle-life model for
graphite-LiFePO4 cells,” Journal of Power Sources, vol. 196, no. 8, pp.
3942–3948, 2011.
[30] A. Cordoba-Arenas, S. Onori, Y. Guezennec, and G. Rizzoni, “Capacity
and power fade cycle-life model for plug-in hybrid electric vehicle
lithium-ion battery cells containing blended spinel and layered-oxide
positive electrodes,” Journal of Power Sources, vol. 278, pp. 473–483,
2015.
[31] R. Deshpande, M. Verbrugge, Y.-T. Cheng, J. Wang, and P. Liu, “Battery
Cycle Life Prediction with Coupled Chemical Degradation and Fatigue
Mechanics,” Journal of the Electrochemical Society, vol. 159, no. 10,
pp. A1730–A1738, aug 2012.
[32] R. Wang, Y. Chen, D. Feng, X. Huang, and J. Wang, “Development
and performance characterization of an electric ground vehicle with
independently actuated in-wheel motors,” Journal of Power Sources, vol.
196, no. 8, pp. 3962–3971, 2011.
[33] J. Shen, A. Hasanzadeh, and A. Khaligh, “Optimal power split and sizing
of hybrid energy storage system for electric vehicles,” in 2014 IEEE
Transportation Electrification Conference and Expo (ITEC), 2014, pp.
1–6.
[34] L. Zadeh, “Fuzzy sets,” Information and Control, vol. 8, no. 3, pp. 338
– 353, 1965.
[35] S. Ebbesen, P. Elbert, and L. Guzzella, “Battery state-of-health perceptive energy management for hybrid electric vehicles,” IEEE Transactions
on Vehicular Technology, vol. 61, no. 7, pp. 2893–2900, Sept 2012.
[36] K. Deb, “Multi-objective optimization using evolutionary algorithms:
an introduction,” Multi-objective evolutionary optimisation for product
design and manufacturing, pp. 1–24, 2011.
| 3cs.SY
|
1
A systematic process for evaluating structured
perfect Bayesian equilibria in dynamic games with
asymmetric information
arXiv:1508.06269v3 [math.OC] 18 Mar 2018
Deepanshu Vasal, Abhinav Sinha and Achilleas Anastasopoulos
Abstract
We consider both finite-horizon and infinite-horizon versions of a dynamic game with N selfish players who
observe their types privately and take actions that are publicly observed. Players’ types evolve as conditionally
independent Markov processes, conditioned on their current actions. Their actions and types jointly determine
their instantaneous rewards. In dynamic games with asymmetric information a widely used concept of equilibrium
is perfect Bayesian equilibrium (PBE), which consists of a strategy and belief pair that simultaneously satisfy
sequential rationality and belief consistency. In general, there does not exist a universal algorithm that decouples
the interdependence of strategies and beliefs over time in calculating PBE. In this paper, for the finite-horizon game
with independent types we develop a two-step backward-forward recursive algorithm that sequentially decomposes
the problem (w.r.t. time) to obtain a subset of PBEs, which we refer to as structured Bayesian perfect equilibria
(SPBE). In such equilibria, a player’s strategy depends on its history only through a common public belief and
its current private type. The backward recursive part of this algorithm defines an equilibrium generating function.
Each period in the backward recursion involves solving a fixed-point equation on the space of probability simplexes
for every possible belief on types. Using this function, equilibrium strategies and beliefs are generated through
a forward recursion. We then extend this methodology to the infinite-horizon model, where we propose a timeinvariant single-shot fixed-point equation, which in conjunction with a forward recursive step, generates the SPBE.
Sufficient conditions for the existence of SPBE are provided. With our proposed method, we find equilibria that
exhibit signaling behavior. This is illustrated with the help of a concrete public goods example.
Index Terms
Dynamic games, asymmetric information, perfect Bayesian equilibrium, sequential decomposition, dynamic
programming, signaling.
I. I NTRODUCTION
Several practical applications involve dynamic interaction of strategic decision-makers with private and
public observations. Such applications include repeated online advertisement auctions, wireless resource
sharing, and energy markets. In repeated online advertisement auctions, advertisers place bids for locations
on a website to sell a product. These bids are calculated based on the value of that product, which is
privately observed by the advertiser and past actions of other advertisers, which are observed publicly.
The authors are with the Department of Electrical Engineering and Computer Science, University of Michigan, Ann Arbor, MI, 48105
USA e-mail: {dvasal, absi, anastas}@umich.edu.
This work is supported in part by NSF grants CIF-1111061 and ECCS-1608361.
This paper has originally appeared on arXiv.org on August 26, 2015 as working paper 1508.06269 and revised on September 14, 2016 as
1609.04221.
2
Each advertiser’s goal is to maximize its reward, which for an auction depends on the actions taken by
others. In wireless resource sharing, players are allocated channels that interfere with each other. Each
player privately observes its channel gain and takes an action, which can be the choice of modulation or
coding scheme and also the transmission power. The reward it receives depends on the rate the player
gets, which is a function of each player’s channel gain and other players’ actions (through the signalto-interference ratio). Finally, in an energy market, different suppliers bid their estimated power outputs
to an independent system operator (ISO) that formulates the market mechanism to determine the prices
assessed to the different suppliers. Each supplier wants to maximize its return, which depends on its cost
of production of energy, which is its private information, and the market-determined prices which depend
on all the bids.
Dynamical systems with strategic players are modeled as dynamic stochastic games, introduced by
Shapley in [1]. Discrete-time dynamic games with Markovian structure have been studied extensively
to model many practical applications, in engineering as well as economics literature [2, 3]. In dynamic
games with perfect and symmetric information, subgame perfect equilibrium (SPE) is an appropriate
equilibrium concept and there exists a backward recursive algorithm to find all the SPEs of these games
(refer to [4, 5, 6] for a more elaborate discussion). Maskin and Tirole in [7] introduced the concept of
Markov perfect equilibrium (MPE) for dynamic games with symmetric information, where equilibrium
strategies are dependent on some payoff relevant Markovian state of the system, rather than on the
entire history. This is a refinement of the SPE. Some prominent examples of the application of MPE
include [8, 9, 10]. Ericson and Pakes in [8] model industry dynamics for firms’ entry, exit and investment
participation, through a dynamic game with symmetric information, compute its MPE, and prove ergodicity
of the equilibrium process. Bergemann and Välimäki in [9] study a learning process in a dynamic oligopoly
with strategic sellers and a single buyer, allowing for price competition among sellers. They study MPE
of the game and its convergence behavior. Acemoğlu and Robinson in [10] develop a theory of political
transitions in a country by modeling it as a repeated game between the elites and the poor, and study its
MPE.
In dynamic games with asymmetric information, and more generally in multi-player, dynamic decision
problems (cooperative or non-cooperative) with asymmetric information, there is a signaling phenomenon
that can occur, where a player’s action reveals part of its private information to other players, which in
turn affects their future payoff (see [11] for a survey of signaling models).1 In one of the first works
demonstrating signaling, a two-stage dynamic game was considered by Spence [15], where a worker
signals her abilities to a potential employer using the level of education as a signal. Since then, this
phenomenon has been shown in many settings, e.g., warranty as a signal for better quality of a product,
in [16], larger deductible or partial insurance as a signal for better health of a person, in [17, 18], and in
evolutionary game theory, extra large antlers by a deer to signal fitness to a potential mate, in [19].
1
There are however instances where even though actions reveal private information, at equilibrium the signaling effect is non-existent [12,
13] and [14, sec III.A]. Thus, MPE is an appropriate equilibrium concept for such games. In [12], authors extend the model of [8] where
firms’ set-up costs and scrap values are random and their private information. However, these are assumed to be i.i.d. across time and thus
the knowledge of this private information in any period does not affect the future reward. In[13],[14, sec III.A], authors discuss games
with one-step delayed information pattern, where all players get access to players’ private information with delay one. In this case as well,
signaling does not occur.
3
For dynamic games with asymmetric information, where players’ observations belong to different
information sets, in order to calculate expected future rewards players need to form a belief on the
observations of other players (where players need not have consistent beliefs). As a result, SPE or MPE, 2
are not appropriate equilibrium concepts for such settings. There are several notions of equilibrium for
such games, such as perfect Bayesian equilibrium (PBE), sequential equilibrium, and trembling hand
equilibrium [4, 5]. Each of these equilibrium notions consist of an assessment, i.e., a strategy and a
belief profile for the entire time horizon. The equilibrium strategies are optimal given the equilibrium
beliefs and the equilibrium beliefs are derived from the equilibrium strategy profile using Bayes’ rule
(whenever possible), with some equilibrium concepts requiring further refinements. Thus there is a cyclical
requirement of beliefs being consistent with strategies, which are in turn optimal given the beliefs,
and finding such equilibria can be thought of as being equivalent to solving a fixed point equation in
the space of strategy and belief profiles over the entire time horizon. Furthermore, these strategies and
beliefs are functions of histories and thus their domain grows exponentially in time, which makes the
problem computationally intractable. To date, there is no universal algorithm that provides simplification
by decomposing the aforementioned fixed-point equation for calculating PBEs.
Some practically motivated work in this category is the work in [20, 21, 22, 23]. Authors in [20, 21, 22]
study the problem of social learning with sequentially-acting selfish players who act exactly once in the
game and make a decision to adopt or reject a trend based on their estimate of the system state. Players
observe a private signal about the system state and publicly observe actions of past players. The authors
analyze PBE of the dynamic game and study the convergent behavior of the system under an equilibrium,
where they show occurrence of herding. Devanur et al. in [23] study PBE of a repeated sales game where
a single buyer has a valuation of a good, which is its private information, and a seller offers to sell a
fresh copy of that good in every period through a posted price.
A. Contributions
In this paper, we present a sequential decomposition methodology for calculating a subset of all PBEs for
finite and infinite horizon dynamic games with asymmetric information. Our model, consists of strategic
players having types that evolve as conditionally independent Markov controlled processes. Players observe
their types privately and actions taken by all players are observed publicly. Instantaneous reward for each
player depends on everyone’s types and actions. The proposed methodology provides a decomposition of
the interdependence between beliefs and strategies in PBE and enables a systematic evaluation of a subset
of PBE, namely structured perfect Bayesian equilibria (SPBE). Furthermore, we show that all SPBE can
be computed using this methodology. Here SPBE are defined as equilibria with players strategies based
on their current private type and a set of beliefs on each player’s current private type, which is common
to all the players and whose domain is time-invariant. The beliefs on players’ types are such that they
can be updated individually for each player and sequentially w.r.t. time. The model allows for signaling
amongst players as beliefs depend on strategies.
2
SPE and MPE are used for games where beliefs in the game are strategy-independent and consistent among players. Equivalently, these
beliefs are derived from basic parameters of the problems and are not part of the definition of the equilibrium concept.
4
Our motivation for considering SPBE stems from ideas in decentralized team problems and specifically
the works of Ho [24] and Nayyar et al. [25]. We utilize the agent-by-agent approach in [24] to motivate a
Markovian structure where players’ strategies depend only on their current types. In addition, we utilize
the common information based approach introduced in [25] to summarize the common information into
a common belief on players’ private types. Even though these ideas motivate the special structure of
our equilibrium strategies, they can not be applied in games to evaluate SPBE because they have been
developed for dynamic teams and are incompatible with equilibrium notions. Our main contribution is a
new construction based on which SPBE can be systematically evaluated.
Specifically, for the finite horizon model, we provide a two-step algorithm involving a backward
recursion followed by a forward recursion. The algorithm works as follows. In the backward recursion, for
every time period, the algorithm finds an equilibrium generating function defined for all possible common
beliefs at that time. This involves solving an one-step fixed point equation on the space of probability
simplexes. Then, the equilibrium strategies and beliefs are obtained through a forward recursion using the
equilibrium generating function obtained in the backward step and the Bayes update rule. The SPBE that
are developed in this paper are analogous to MPEs (for games with symmetric information) in the sense
that players choose their actions based on beliefs that depend on common information, and private types,
both of which have Markovian dynamics.
For the infinite horizon model, instead of the backwards recursion step, the algorithm solves a single-shot
time invariant fixed-point equation involving both an equilibrium generating function and an equilibrium
reward-to-go function. We show that using our method, existence of SPBE in the asymmetric information
dynamic game is guaranteed if the aforementioned fixed-point equation admits a solution. We provide
sufficient conditions under which this is true. We demonstrate our methodology of finding SPBE through
a multi-stage public goods game, whereby we observe the aforementioned signaling effect at equilibrium.
B. Relevant Literature
Related literature on this topic include [14, 26] and [27]. Nayyar et al. in [14, 26] consider dynamic
games with asymmetric information. There is an underlying controlled Markov process and players
jointly observe part of the process and whilst making additional private observations. It is shown that
the considered game with asymmetric information, under certain assumptions, can be transformed to
another game with symmetric information. A backward recursive algorithm is provided to find MPE of
the transformed game. For this strong equivalence to hold, authors in [14, 26] make a critical assumption
in their model: based on the common information, a player’s posterior beliefs about the system state and
about other players’ information are independent of the past strategies used by the players. This leads to
all strategies being non-signaling. Our model is different from this since we assume that the underlying
state of the system has independent components, each constituting a player’s private type. However, we
do not make any assumption regarding update of beliefs and allow the belief state to depend on players’
past strategies, which in turn allows the possibility of signaling in the game.
Ouyang et al. in [27] consider a dynamic oligopoly game with strategic sellers and buyers. Each seller
privately observes the valuation of their good, which is assumed to have independent Markovian dynamics,
5
thus resulting in a dynamic game of asymmetric information. The common belief is strategy dependent
and the authors consider equilibria based on this common information belief. It is shown that if all other
players play actions based on the common belief and their private information using equilibrium strategies,
and if all players use equilibrium belief update function, then player i faces a Markov decision process
(MDP) with respect to its action with state being the common belief and its private type. Thus calculating
equilibrium boils down to solving a fixed-point equation on belief update functions and strategies of all
players. Existence of such equilibrium is shown for a degenerate case where players act myopically at
equilibrium and the equilibrium itself is non-signaling.
Other than the common information based approach, Li et al. [28] consider a finite horizon zero-sum
dynamic game, where at each time only one player out of the two knows the state of the system. The
value of the game is calculated by formulating an appropriate linear program. Cole et al. [29] consider
an infinite horizon discounted reward dynamic game where actions are only privately observable. They
provide a fixed-point equation for calculating a subset of sequential equilibrium, which is referred to as
Markov private equilibrium (MPrE). In MPrE strategies depend on history only through the latest private
observation.
C. Notation
We use uppercase letters for random variables and lowercase for their realizations. For any variable, subscripts represent time indices and superscripts represent player indices. We use notation −i
to represent all players other than player i i.e. −i = {1, 2, . . . i − 1, i + 1, . . . , N }. We use notation
At:t0 to represent the vector (At , At+1 , . . . At0 ) when t0 ≥ t or an empty vector if t0 < t. We use
i+1
. . . , AN
At−i to mean (A1t , A2t , . . . , Ai−1
t , At
t ) . We remove superscripts or subscripts if we want to
represent the whole vector, for example At represents (A1t , . . . , AN
t ). In a similar vein, for any collection
i
i
of sets (X )i∈N , we denote ×i∈N X by X . We denote the indicator function of a set A by IA (·).
For any finite set S, ∆(S) represents the space of probability measures on S and |S| represents its
cardinality. We denote by Pg (or Eg ) the probability measure generated by (or expectation with respect
to) strategy profile g. We denote the set of real numbers by R. For a probabilistic strategy profile of
players (βti )i∈N where the probability of action ait conditioned on a1:t−1 , xi1:t is given by βti (ait |a1:t−1 , xi1:t ),
Q
j j
j
−i
we use the notation βt−i (a−i
t |a1:t−1 , x1:t ) to represent
j6=i βt (at |a1:t−1 , x1:t ). All equalities/inequalities
involving random variables are to be interpreted in the a.s. sense. For mappings with range function sets
f : A → (B → C) we use square brackets f [a] ∈ B → C to denote the image of a ∈ A through f and
parentheses f [a](b) ∈ C to denote the image of b ∈ B through f [a]. A controlled Markov process with
state Xt , action At , and horizon T is denoted by (Xt , At )t∈T .
The paper is organized as follows. In Section II, we present the model for games with finite and infinite
horizon. Section III serves as motivation for focusing on SPBE. In Section IV, for finite-horizon games, we
present a two-step backward-forward recursive algorithm to construct a strategy profile and a sequence
of beliefs, and show that it is a PBE of the dynamic game considered. In Section V, we extend that
methodology to infinite-horizon games. Section VII discusses concrete example of a public goods game
with two players and results are presented for both, finite and infinite horizon versions of the example.
All proofs are provided in appendices.
6
II. M ODEL AND P RELIMINARIES
4
We consider a discrete-time dynamical system with N strategic players in the set N = {1, 2, . . . N }. We
4
consider two cases: finite horizon T = {1, 2, . . . T } with perfect recall and infinite horizon with perfect
4
recall. The system state is Xt = (Xt1 , Xt2 , . . . XtN ), where Xti ∈ X i is the type of player i at time t, which
is perfectly observed and is its private information. Players’ types evolve as conditionally independent,
controlled Markov processes such that
P(x1 ) =
N
Y
Qi1 (xi1 )
(1a)
i=1
P(xt |x1:t−1 , a1:t−1 ) = P(xt |xt−1 , at−1 )
=
N
Y
Qit (xit |xit−1 , at−1 ),
(1b)
(1c)
i=1
where Qit are known kernels. Player i at time t takes action ait ∈ Ai on observing the actions a1:t−1 =
(ak )k=1,...,t−1 where ak = ajk j∈N , which is common information among players, and the types xi1:t ,
which it observes privately. The sets Ai , X i are assumed to be finite. Let g i = (gti )t∈T be a probabilistic
strategy of player i where gti : At−1 × (X i )t → ∆(Ai ) such that player i plays action Ait according to
4
Ait ∼ gti (·|a1:t−1 , xi1:t ). Let g = (g i )i∈N be a strategy profile of all players. At the end of interval t, player
i receives an instantaneous reward Rti (xt , at ). To preserve the information structure of the problem, we
assume that players do not observe their rewards until the end of game.3 The reward functions and state
transition kernels are common knowledge among the players. For the finite-horizon problem, the objective
of player i is to maximize its total expected reward
( T
)
X
4
Rti (Xt , At ) .
(2)
J i,g = Eg
t=1
For the infinite-horizon case, the transition kernels Qit are considered to not depend on time t. We also
substitute Rti (Xt , At ) = δ t Ri (Xt , At ) take limT →∞ in the above equation, where δ ∈ [0, 1) is the common
discount factor and Ri is the time invariant stage reward function for player i. With all players being
strategic, this problem is modeled as a dynamic game, DT for finite horizon and D∞ for infinite horizon,
with asymmetric information and simultaneous moves.
A. Preliminaries
Any history of this game at which players take action is of the form ht = (a1:t−1 , x1:t ). Let Ht be
4
the set of such histories, HT = ∪Tt=0 Ht be the set of all possible such histories in finite horizon and
4
i
i
H∞ = ∪∞
t=0 Ht for infinite horizon. At any time t player i observes ht = (a1:t−1 , x1:t ) and all players
together have hct = a1:t−1 as common history. Let Hti be the set of observed histories of player i at time t
and Htc be the set of common histories at time t. An appropriate concept of equilibrium for such games is
PBE [5], which consists of a pair (β ∗ , µ∗ ) of strategy profile β ∗ = (βt∗,i )t∈T ,i∈N where βt∗,i : Hti → ∆(Ai )
Alternatively, we could have assumed instantaneous reward of a player to depend only on its own type, i.e. be of the form Rti (xit , at ),
and have allowed rewards to be observed by the players during the game as this would not alter the information structure of the game
3
7
and a belief profile µ∗ = (i µ∗t )t∈T ,i∈N where i µ∗t : Hti → ∆(Ht ) that satisfy sequential rationality so that
∀i ∈ N , t ∈ T , hit ∈ Hti , β i
∗,i
i
Wti,β ,T (hit ) ≥ Wti,β ,T (hit )
(3)
where the reward-to-go is defined as
i,β i ,T
Wt
β i β ∗,−i , i µ∗t [hit ]
(hit ) , E
( T
X
)
Rni (Xn , An ) hit
,
(4)
n=t
and the beliefs satisfy some consistency conditions as described in [5, p. 331]. Similarly, for the game
D∞ PBE (β ∗ , µ∗ ) requires: ∀i ∈ N , t ≥ 1, hit ∈ Hti , β i
∗,i
i
Wti,β (hit ) ≥ Wti,β (hit )
(5)
where the reward-to-go is
i,β i
Wt
i ∗,−i , i µ∗ [hi ]
t t
(hit ) , Eβ β
(∞
X
)
Rni (Xn , An ) hit
.
(6)
n=t
In general, a belief for player i at time t, i µ∗t is defined on history ht = (a1:t−1 , x1:t ) given its private history
hit = (a1:t−1 , xi1:t ). Here player i’s private history hit = (a1:t−1 , xi1:t ) consists of a public part hct = a1:t−1
and a private part xi1:t . At any time t, the relevant uncertainty player i has is about other players’ types
t
j
x−i
1:t ∈ ×n=1 (×j6=i X ) and their future actions. In our setting, due to independence of types, and given the
common history hct , player i’s type history xi1:t does not provide any additional information about x−i
1:t ,
as will be shown later. For this reason we consider beliefs that are functions of each player’s history hit
only through the common history hct . Hence, for each player i, its belief for each history hct = a1:t−1 is
derived from a common belief µ∗t [a1:t−1 ]. Furthermore, as will be shown later, this belief factorizes into a
Q
∗
∗
product of marginals j∈N µ∗,j
t [a1:t−1 ]. Thus we can sufficiently use the system of beliefs, µ = (µt )t∈T ,
∗,i
−i
c
i
where µ∗t = (µ∗,i
t )i∈N , and µt : Ht → ∆(X ), with the understanding that player i’s belief on xt is
Q
∗,j
j
µt∗,−i [a1:t−1 ](x−i
t ) =
j6=i µt [a1:t−1 ](xt ). Under the above structure, all consistency conditions that are
required for PBEs [5, p. 331] are automatically satisfied.
III. M OTIVATION FOR STRUCTURED EQUILIBRIA
In this section, we present structural results for the considered dynamical process that serve as a
motivation for finding SPBE of the underlying game DT . Specifically, we define a belief state based on
common information history and show that any reward profile that can be obtained through a general
strategy profile can also be obtained through strategies that depend on this belief state and players’
current types, which are their private information. These structural results are inspired by the analysis
of decentralized team problems, which serve as guiding principles to design our equilibrium strategies.
While these structural results provide intuition and the required notation, they are not directly used in the
proofs for finding SPBE later in Section IV.
At any time t, player i has information (a1:t−1 , xi1:t ) where a1:t−1 is the common information among
players, and xi1:t is the private information of player i. Since (a1:t−1 , xi1:t ) increases with time, any strategy
of the form Ait ∼ gti (·|a1:t−1 , xi1:t ) becomes unwieldy. Thus it is desirable to have an information state in
8
a time-invariant space that succinctly summarizes (a1:t−1 , xi1:t ), and that can be sequentially updated. We
first show in Lemma 1 that given the common information a1:t−1 and its current type xit , player i can
discard its type history xi1:t−1 and play a strategy of the form Ait ∼ sit (·|a1:t−1 , xit ). Then in Lemma 2,
we show that a1:t−1 can be summarized through a belief πt , defined as follows. For any strategy profile
4
g, belief πt on Xt , πt ∈ ∆(X ), is defined as πt (xt ) = Pg (Xt = xt |a1:t−1 ), ∀xt ∈ X . We also define the
4
marginals πti (xit ) = Pg (Xti = xit |a1:t−1 ), ∀xit ∈ X i .
For player i, we use the notation g to denote a general policy of the form Ait ∼ gti (·|a1:t−1 , xi1:t ), notation
s, where sit : At−1 × X i → ∆(Ai ), to denote a policy of the form Ait ∼ sit (·|a1:t−1 , xit ), and notation m,
where mit : ∆(×i∈N X i ) × X i → ∆(Ai ), to denote a policy of the form Ait ∼ mit (·|πt , xit ). It should be
noted that since πt is a function of random variables a1:t−1 , m policy is a special type of s policy, which
in turn is a special type of g policy.
Using the agent-by-agent approach [24], we show in Lemma 1 that any expected reward profile of the
players that can be achieved by any general strategy profile g can also be achieved by a strategy profile
s.
Lemma 1: Given a fixed strategy g −i of all players other than player i and for any strategy g i of player
i, there exists a strategy si of player i such that ∀t ∈ T , xt ∈ X , at ∈ A,
i −i
i −i
Ps g (xt , at ) = Pg g (xt , at )
i −i
(7)
i −i
which implies J i,s g = J i,g g .
Proof: Please see Appendix A.
Since any si policy is also a g i type policy, the above lemma can be iterated over all players which implies
that for any g policy profile there exists an s policy profile that achieves the same reward profile i.e.,
(J i,s )i∈N = (J i,g )i∈N .
Policies of types s still have increasing domain due to increasing common information a1:t−1 . In order to
summarize this information, we take an equivalent view of the system dynamics through a common agent,
as taken in [30]. The common agent approach is a general approach that has been used extensively in
dynamic team problems [31, 32, 33, 34]. Using this approach, the problem can be equivalently described
as follows: player i at time t observes a1:t−1 and takes action γti , where γti : X i → ∆(Ai ) is a partial
(stochastic) function from its private information xit to ait , of the form Ait ∼ γti (·|xit ). These actions are
generated through some policy ψ i = (ψti )t∈T , ψti : At−1 → {X i → ∆(Ai )}, that operates on the common
information a1:t−1 such that γti = ψti [a1:t−1 ]. Then any policy of the form Ait ∼ sit (·|a1:t−1 , xit ) is equivalent
to Ait ∼ ψti [a1:t−1 ](·|xit ).
We call a player i’s policy through common agent to be of type ψ i if its actions γti are taken as
γti = ψti [a1:t−1 ]. We call a player i’s policy through common agent to be of type θi where θti : ∆(X ) →
{X i → ∆(Ai )}, if its actions γti are taken as γti = θti [πt ]. A policy of type θi is also a policy of type ψ i .
There is a one-to-one correspondence between policies of type si and of type ψ i and between policies of
type mi and of type θi . In summary, the notation for the various functional form of strategies is
Ait ∼ sit (·|a1:t−1 , xit )
Ait ∼ ψti [a1:t−1 ](·|xit ),
(8a)
Ait ∼ mit (·|πt , xit )
Ait ∼ θti [πt ](·|xit ).
(8b)
9
In the following lemma, we show that the space of profiles of type s is outcome-equivalent to the space
of profiles of type m.
Lemma 2: For any given strategy profile s of all players, there exists a strategy profile m such that
Pm (xt , at ) = Ps (xt , at )
∀t ∈ T , xt ∈ X , at ∈ A,
(9)
which implies (J i,m )i∈N = (J i,s )i∈N .
Proof: Please see Appendix B.
The above two lemmas show that any reward profile that can be generated through a policy profile of
type g can also be generated through a policy profile of type m. This is precisely the motivation for using
SPBE which are equilibria based on policies of type m. It should be noted that the construction of si
depends only on g i (as shown in (37)), while the construction of mi depends on the whole policy profile g
and not just on g i , since the construction of θi depends on ψ in (49). Thus any unilateral deviation of player
i in g policy profile does not necessarily translate to unilateral deviation of player i in the corresponding
m policy profile. Therefore g being an equilibrium of the game (in some appropriate notion) does not
necessitate the corresponding m also being an equilibrium. Thus the set of equilibria of type g contains
those of type m but not vice-versa (in general); characterizing the relationship between the two sets of
equilibria is an interesting open problem.
We end this section by noting that although finding general PBEs of type g of the games DT or D∞
would be a desirable goal, since the space of strategies is growing exponentially with time, it would be
computationally intractable. However, Lemmas 1 and 2 suggest that strategies of type m form a rich
class that achieves every possible reward profile. Since these strategies are functions of beliefs πt that
lie in a time-invariant space and are easily updatable, equilibria of this type are potential candidates for
computation through backward recursion. Our goal is to devise an algorithm to find structured equilibria
of type m of the dynamic games DT or D∞ .
Definition 1 (SPBE): A structured perfect Bayesian equilibrium is a PBE of the considered dynamic
game where at any time t, for any agent i, its equilibrium strategy βt∗,i is of type m (as in (8b)).
IV. A M ETHODOLOGY FOR SPBE
COMPUTATION IN FINITE HORIZON
In this section we consider the finite horizon dynamic game DT . In the previous section, (specifically
in Claim 1, included in the proof of Lemma 2 in Appendix B), it is shown that due to the independence
of types and their evolution as independent controlled Markov processes, for any strategy of the players,
Q
i i
the joint common belief can be factorized as a product of its marginals i.e., πt (xt ) = N
i=1 πt (xt ), ∀xt .
Since in this paper, we only deal with such joint beliefs, to accentuate this independence structure, we
define π t ∈ ×i∈N ∆(X i ) as vector of marginal beliefs where π t := (πti )i∈N . In the rest of the paper, we
will use π t instead of πt whenever appropriate, where of course, πt can be constructed from π t . Similarly,
we define the vector of belief updates as F (π, γ, a) := (F i (π i , γ i , a))i∈N where (using Bayes rule)
P πi (xi )γ i (ai |xi )Qi (xi |xi ,a)
P i i i i i
t
t t+1 t
xit Pt
if
i
i
i
i
i
x̃it π (x̃t )γ (a |x̃t ) > 0
π (x̃t )γ (a |x̃t )
x̃it
(10)
F i (π i , γ i , a)(xit+1 ) =
P
P
π i (xi )Qi (xi |xi , a) if
π i (x̃i )γ i (ai |x̃i ) = 0.
xit
t
t
t+1
t
x̃it
t
t
10
The update function F i defined above depends on time t through the kernel Qit (for the finite horizon
model). For notational simplicity we suppress this dependence on t. We also change the notation of
policies of type m and θ as follows, so they depend on π t instead of πt
mit : ×i∈N ∆(X i ) × X i → ∆(Ai )
θti : ×i∈N ∆(X i ) → X i → ∆(Ai ) .
(11a)
(11b)
In the following we present a backward-forward algorithm that evaluates SPBE. As will be shown in
Theorem 2, this is a “canonical” methodology, in the sense that all SPBE can be generated this way.
A. Backward Recursion
In this section, we define an equilibrium generating function θ = (θti )i∈N ,t∈T , where θti : ×i∈N ∆(X i ) →
{X i → ∆(Ai )}. In addition, we define a sequence of reward-to-go functions of player i at time t,
(Vti )i∈N ,t∈{1,2,...T +1} , where Vti : ×i∈N ∆(X i )×X i → R. These quantities are generated through a backward
recursive way, as follows.
1. Initialize ∀π T +1 ∈ ×i∈N ∆(X i ), xiT +1 ∈ X i ,
4
VTi +1 (π T +1 , xiT +1 ) = 0.
(12)
Q
2. For t = T, T − 1, . . . 1, ∀π t ∈ ×i∈N ∆(X i ), πt = i∈N πti , let θt [π t ] be generated as follows. Set
γ̃t = θt [π t ], where γ̃t is the solution, if it exists,4 of the following fixed-point equation, ∀i ∈ N , xit ∈ X i ,
o
n
i
i
i
i
γti (·|xit )γ̃t−i , πt
i
i
γ̃t (·|xt ) ∈ arg max
(13)
E
Rt (Xt , At ) + Vt+1 (F (π t , γ̃t , At ), Xt+1 ) xt ,
i
i
γt (·|xt )
i
where expectation in (13) is with respect to random variables (Xt−i , At , Xt+1
) through the measure
i i i −i −i −i
i
i
i
πt−i (x−i
t )γt (at |xt )γ̃t (at |xt )Qt+1 (xt+1 |xt , at ) and F is defined above.
Furthermore, using the quantity γ̃t found above, define
o
n
i
i 4 γ̃ti (·|xit )γ̃t−i , πt
i
i
i
i
Vt (π t , xt ) =E
(14)
Rt (Xt , At ) + Vt+1 (F (π t , γ̃t , At ), Xt+1 ) xt .
It should be noted that in (13), γ̃ti is not the outcome of a maximization operation as is the case in a best
response equation of a Bayesian Nash equilibrium. Rather (13) is a different fixed point equation. This is
because the maximizer γ̃ti appears in both, the left-hand-side and the right-hand-side of the equation (in
the belief update F (π t , γ̃t , At ) = (F i (πti , γ̃ti , At ))i∈N ). This distinct construction is pivotal in the proof of
Theorem 1, as will be further elaborated in the Discussion section.
B. Forward Recursion
As discussed above, a pair of strategy and belief profile (β ∗ , µ∗ ) is a PBE if it satisfies (4). Based on
θ defined above in (12)–(14), we now construct a set of strategies β ∗ and beliefs µ∗ for the game DT in
a forward recursive way, as follows.5 As before, we will use the notation µ∗t [a1:t−1 ] := (µ∗,i
t [a1:t−1 ])i∈N ,
4
The problem of existence in this step will be discussed in Section VI.
As discussed in the preliminaries subsection on Section II, the equilibrium beliefs in SPBE, µ∗t are functions of each player’s history hit
only through the common history hct and are the same for all players.
5
11
where µ∗,i
[a1:t−1 ] is a belief on xit , and µ∗t [a1:t−1 ] can be constructed from µ∗t [a1:t−1 ] as µ∗t [a1:t−1 ](xt ) =
QN ∗,it
i
c
i=1 µt [a1:t−1 ](xt ), ∀a1:t−1 ∈ Ht .
1. Initialize at time t = 1,
µ∗1 [φ](x1 )
:=
N
Y
Qi1 (xi1 ).
(15)
i=1
c
2. For t = 1, 2 . . . T, ∀i ∈ N , a1:t ∈ Ht+1
, xi1:t ∈ (X i )t
βt∗,i (ait |a1:t−1 , xi1:t ) = βt∗,i (ait |a1:t−1 , xit )
:= θti [µ∗t [a1:t−1 ]](ait |xit )
(16)
and
∗,i
i
i ∗
µ∗,i
t+1 [a1:t ] := F (µt [a1:t−1 ], θt [µt [a1:t−1 ]], at )
(17)
where F i is defined in (10).
We now state our main result.
Theorem 1: A strategy and belief profile (β ∗ , µ∗ ), constructed through the backward-forward recursion
algorithm is a PBE of the game, i.e., ∀i ∈ N , t ∈ T , a1:t−1 ∈ Htc , xi1:t ∈ (X i )t , β i ,
∗,i ∗,−i
βt:T
βt:T , µ∗t [a1:t−1 ]
E
( T
X
)
Rni (Xn , An ) a1:t−1 , xi1:t
n=t
i β ∗,−i , µ∗ [a
βt:T
t 1:t−1 ]
t:T
≥E
( T
X
)
Rni (Xn , An ) a1:t−1 , xi1:t
.
n=t
(18)
Proof: Please see Appendix C.
We emphasize that even though the backward-forward algorithm presented above finds a class of equilibrium strategies that are structured, the unilateral deviations of players in (18) are considered in the space
of general strategies, i.e., the algorithm does not make any bounded rationality assumptions.
The following result shows that the backward-forward construction described above is “canonical”, in
the sense that all SPBE can be found through this methodology. Clearly, an SPBE can be defined as a
PBE (β ∗ , µ∗ ) of the game that is generated through forward recursion in (15)–(17), using an equilibrium
generating function φ, where φ = (φit )i∈N ,t∈T , φit : ×i∈N ∆(X i ) → {X i → ∆(Ai )}, common belief update
function F and prior distributions Q1 . As a consequence, βt∗,i only depends on current type xit of player
i, and on the common information a1:t−1 through the set of marginals µ∗t [a1:t−1 ], and µ∗,i
t depends only
on common information history a1:t−1 .
Theorem 2 (Converse): Let (β ∗ , µ∗ ) be an SPBE. Then there exists an equilibrium generating function
φ that satisfies (13) in backward recursion ∀ πt = µ∗t [a1:t−1 ], ∀ a1:t−1 , such that (β ∗ , µ∗ ) is defined through
forward recursion using φ.6
Proof: Please see Appendix E.
6
Note that for π t 6= µ∗t [a1:t−1 ] for any a1:t−1 , φ can be arbitrarily defined without affecting the definition of (β ∗ , µ∗ ).
12
C. Discussion
Several remarks are in order with regard to the above methodology and the result.
Remark 1: The second sub-case in (10) dictates how beliefs are updated for histories with zero
probability. The particular expression used is only one of many possible updates than can be used here.
Dynamics that govern the evolution of public beliefs at histories with zero probability of occurrence affect
equilibrium strategies. Thus, the construction proposed for calculating PBEs in this paper will produce a
different set of equilibria if one changes the second sub-case above. The most well-known example of
another such update is the intuitive criterion proposed in [35] for Nash equilibria, later generalized to
sequential equilibria in [36]. The intuitive criterion assigns zero probability to states that can be excluded
based on data available to all players (in our case action profile history a1:t−1 ). Another example of belief
update is universal divinity, proposed in [37].
Remark 2: To highlight the significance of the unique structure of (13), one can think as follows. When
all players other than player i play structured strategies, i.e., strategies of the form Ajt ∼ mjt (·|π t , xjt ) =
θtj [π t ](·|xjt ), one may want to characterize the optimization problem from the viewpoint of the i-th player
in order to characterize its best response. In particular one may want to show that although player i can
play general strategies of the form Ait ∼ gti (·|xi1:t , a1:t−1 ), it is sufficient to best respond with structured
strategies of the form Ait ∼ mit (·|π t , xit ) = θti [π t ](·|xit ) as well. To show that, one may entertain the thought
that player i faces an MDP with state (Xti , Πt ), and action Ait at time t. If that were true, then player i’s
optimal action could be characterized (using standard MDP results) by a dynamic-programming equation
similar to (13), of the form
o
n
i
i
i
−i
i
i
γti (·|xit )γ̃t−i , πt
i
i
γ̃ti (·|xit ) ∈ arg max
,
γ
(·|x
),
γ̃
(·|·),
γ̃
,
A
),
X
)
x
(19)
E
(π
R
(X
,
A
)
+
V
(F
t
t
t
t t
t
t
t
t+1
t ,
t
t+1
i
i
γt (·|xt )
where, unlike (13), in the belief update equation the partial strategy γti (·|xit ) is also optimized over.
However, as it turns out, user i does not face such an MDP problem! The reason is that the update
equation π t+1 = F (π t , γt , at ) also depends on γti which is the partial strategy of player i and this has
not been fixed in the above setting. If however the update equation is first fixed (so it is updated as
π t+1 = F (π t , γ̃ti , γ̃t−i , at ) = F (π t , θt [πt ], at ), i.e., using the equilibrium strategies even for player i) then
indeed the problem faced by user i is the MDP defined above. It is now clear why (13) has the flavor
of a fixed-point equation: the update of beliefs needs to be fixed beforehand with the equilibrium action
γ̃ti even for user i, and only then user i’s best response can depend only on the MDP state (Xti , Πt ) thus
being a structured strategy as well. This implies that his optimal action γ̃ti appears both on the left and
right hand side of this equation giving rise to (13).
Remark 3: In this paper, we find a class of PBEs of the game, while there may exist other equilibria
that are not “structured”, and can not be found by directly using the proposed methodology. The rationale
for using structured equilibria over others is the same as that for using MPE over SPE for a symmetric
information game; a focussing argument for using simpler strategies being one of them.
V. A M ETHODOLOGY FOR SPBE COMPUTATION IN I NFINITE H ORIZON
In this section we consider the infinite horizon discounted reward dynamic game D∞ . We state the
fixed-point equation that defines the value function and strategy mapping for the infinite horizon problem.
13
This is analogous to the backwards recursion ((13) and (14)) that define the value function and θ mapping
for the finite horizon problem.
j
i
i
i
i
Define the set of functions V i : ×N
j=1 ∆(X ) × X → R and strategies γ̃ : X → ∆(A ) (which are
generated formally as γ̃ i = θi [π] for given π) via the following fixed-point equation: ∀ i ∈ N , xi ∈ X i ,
o
n
i
i
γ i (·|xi ),γ̃ −i ,π −i
i
i
0i
i
γ̃ (· | x ) ∈ argmax E
R (X, A) + δV F (π, γ̃, A), X | π, x ,
(20a)
γ i (·|xi )∈∆(Ai )
n
o
i
i
−i −i
i
V i (π, xi ) = Eγ̃ (·|x ),γ̃ ,π
Ri (X, A) + δV i F (π, γ̃, A), X 0 | π, xi .
(20b)
Note that the above is a joint fixed-point equation in (V, γ̃), unlike the backwards recursive algorithm
earlier which required solving a fixed-point equation only in γ̃. Here the unknown quantity is distributed
as (X −i , Ai , A−i , X 0 i ) ∼ π −i (x−i )γ i (ai | xi )γ̃ −i (a−i | x−i )Qi (x0 i | xi , a), and F i (·) is defined in (10).
Define the belief µ∗ inductively similar to the forward recursion from Section IV-B. By construction
the belief defined above satisfies the consistency condition needed for a PBE. Denote the strategy arising
out of γ̃ by β ∗ i.e.,
i
h
i,∗ i
∗
i
i
(21)
βt (at | x1:t , a1:t−1 ) = θ µt [a1:t−1 ] (ait | xit ).
Note that although the mapping θi is stationary, the strategy βti,∗ derived from it is not so. Below we state
the central result of this section, that the strategy-belief pair (β ∗ , µ∗ ) constructed from the solution of the
fixed-point equation (20) and the forward recursion indeed constitutes a PBE.
Theorem 3: Assuming that the fixed-point equation (20) admits an absolutely bounded solution V i (for
all i ∈ N ), the strategy-belief pair (β ∗ , µ∗ ) defined in (21) is a PBE of the infinite horizon discounted
reward dynamic game i.e., ∀ i ∈ N , β i , t ≥ 1, hit ∈ Hti ,
(∞
)
(∞
)
X
X
i,∗
−i,∗
∗
c
i
−i,∗
∗
c
Eβ ,β ,µt [ht ]
δ n−t Ri (Xn , An ) | hit ≥ Eβ ,β ,µt [ht ]
δ n−t Ri (Xn , An ) | hit .
(22)
n=t
n=t
Proof: Please see Appendix F.
Our approach to proving Theorem 3 is as follows. We begin by noting that the standard contraction
mapping arguments used in infinite horizon discounted reward MDPs/POMDPs viewed as a limit of
finite horizon problems, do not apply here, since the policy equation (20a) is not a maximization, but a
different fixed-point equation. So we attempt to “fit” the infinite horizon problem into the framework of
finite-horizon model developed in the previous section. We do that by first introducing a terminal reward
that depends on common beliefs, in the backward-forward recursion construction of Section IV for finite
horizon games. We consider a finite horizon, T > 1, dynamic game with rewards same as in the infinite
horizon version and time invariant transition kernels Qi . For each player i, there is a terminal reward
Gi (πT +1 , xiT +1 ) that depends on the terminal type of player i and the terminal belief. It is
assumed that
i,T
Gi (·) is absolutely bounded. We define the value functions Vt : ×j∈N ∆(X j ) × X i → R
and
i∈N ,t∈T
strategies γ̃ti,T
backwards inductively in the same way as in Section IV-A except Step 1, where
i∈N ,t∈T
i
∗
∗
instead of (12) we set VTi,T
+1 ≡ G . This consequently results in a strategy/belief pair (β , µ ), based on
the forward recursion in Section IV-B. Now, due to the above construction, the value function Vti,T from
above and V i from (20) are related (please see Lemma 9 in Appendix G). This result combined with
continuity arguments as T → ∞ complete the proof of Theorem 3.
14
VI. A N E XISTENCE R ESULT FOR THE F IXED -P OINT E QUATION
In this section, we discuss the problem of existence of signaling equilibria.7 While it is known that for
any finite dynamic game with asymmetric information and perfect recall, there always exists a PBE [4,
Prop. 249.1], existence of SPBE is not guaranteed. It is clear from our algorithm that existence of SPBE
boils down to existence of a solution to the fixed-point equation (13) in finite horizon and (20) in infinite
i
for all i ∈ N from
horizon. Specifically, for the finite horizon, at each time t given the functions Vt+1
the previous round (in the backwards recursion) equation (13) must have a solution γ̃ti for all i ∈ N .
Generally, existence of equilibria is shown through Kakutani’s fixed point theorem, as is done in proving
existence of a mixed strategy Nash equilibrium of a finite game [4, 38]. This is done by showing existence
of fixed point of the best-response correspondences of the game. Among other conditions, it requires the
“closed graph” property of the correspondences, which is usually implied by the continuity property of
the utility functions involved. For (13) establishing existence is not straightforward due to: (a) potential
discontinuity of the πt update function F when the denominator in the Bayesian update is 0 and (b)
i
potential discontinuity of the value functions, Vt+1
. In the following we provide sufficient conditions that
can be checked at each time t to establish the existence of a solution.
We consider a generic fixed-point equation similar to the one encountered in Section IV and Section V
and state conditions under which they are guaranteed to have a solution. To concentrate on the essential
aspects of the problem we consider a simple case with N = 2, type sets X i = {xH , xL } and action sets
Ai = {0, 1}. Furthermore, types are static and instantaneous rewards Ri (x, a) do not depend on x−i .
i
1
2
Given public belief π = (π 1 , π2 ) ∈ ×2i=1
∆(A ), value functions V , V , one wishes to solve the
following system of equations for γ̃ i (· | xi )
.
xi ∈{xH ,xL },i∈{1,2}
i
i
γ̃ (· | x ) ∈
γ i (·|xi ),γ̃ −i
argmax E
i i
i
1
1
1
1
2
2
2
2
i
i
(23)
R (x , A)+V
F (π , γ̃ , A ), F (π , γ̃ , A ) , x | x , π
γ i (·|xi )∈∆(Ai )
where the expectation is evaluated using the probability distribution on (A1 , A2 ),
h
i
i i
i
j
H
j j
H
j
L j j
L
γ (a | x ) π (x )γ̃ (a | x ) + π (x )γ̃ (a | x ) .
(24)
The probabilistic policy γ̃ can be represented by the 4-tuple p = p̃1L , p̃2L , p̃1H , p̃2H where p̃iH =
γ i (ai = 1 | xH ) and p̃iL = γ i (ai = 1 | xL ), i = 1, 2.
The fixed-point equation of interest reduces to
1H
p̃ ∈ argmax a π 2 p̃2H +(1−π 2 )p̃2L V 1 (F1 (π 1 , p̃1 ), F1 (π 2 , p̃2 ), xH )−V 1 (F0 (π 1 , p̃1 ), F1 (π 2 , p̃2 ), xH )
a∈[0,1]
+ 1 − π 2 p̃2H − (1 − π 2 )p̃2L V 1 (F1 (π 1 , p̃1 ), F0 (π 2 , p̃2 ), xH ) − V 1 (F0 (π 1 , p̃1 ), F0 (π 2 , p̃2 ), xH )
+ π 2 p̃2H + (1 − π 2 )p̃2L R1 (xH , 1, 1) − R1 (xH , 0, 1)
2 2H
2 2L
1 H
1 H
+ 1 − π p̃ − (1 − π )p̃
R (x , 1, 0) − R (x , 0, 0)
(25)
In the special case of uncontrolled types where player i’s instantaneous reward does not depend on its private type xit , the fixed point
equation always has a type-independent, myopic solution γ̃ti (·), since it degenerates to a best-response-like equation similar to the one for
computing Nash equilibrium. This result is shown in [27].
7
15
and three other similar equations for p̃1L , p̃2H , p̃2L .
πpH
πpH + πpL
π(1 − pH )
F0 (π, (pH , pL )) ,
π(1 − pH ) + π(1 − pL )
F1 (π, (pH , pL )) ,
(26a)
(26b)
and in both definitions, if the denominator is 0 then the RHS is taken as π.
A. Points of Discontinuities and the Closed graph result
Equation (25) and the other three similar equations are essentially of the form (for a given π)
x ∈ argmax af1 (x, y, w, z),
a∈[0,1]
w ∈ argmax cf3 (x, y, w, z),
c∈[0,1]
y ∈ argmax bf2 (x, y, w, z)
b∈[0,1]
z ∈ argmax df4 (x, y, w, z)
(27)
d∈[0,1]
with x, y, z, w as p̃1H , p̃1L , p̃2H , p̃2L , respectively.
Define Di ⊆ [0, 1]4 as the set of discontinuity points of fi and D , ∪4i=1 Di .
For any point x0 ∈ D, define S(x0 ) as the subset of indexes i ∈ {1, 2, 3, 4} for which fi (x) is
discontinuous at x0 .
Assumption (E1): At any point x0 ∈ D, ∀ i ∈ S(x0 ) one of the following is satisfied:
1) fi (x0 ) = 0, or
2) ∃ > 0 such that ∀ x ∈ B (x0 ) (inside an -ball of x0 ) the sign of fi (x) is same as the sign of fi (x0 ).
In the following we provide a sufficient condition for existence.
Theorem 4: Under Assumption (E1), there exists a solution to the fixed-point equation (27).
Proof: Please see Appendix H.
The above set of results provide us with an analytical tool for establishing existence of a solution to
the concerned fixed-point equation.
While the above analytical result is useful in understanding a theoretical basis for existence, it doesn’t
cover all instances. For instance, fixed-point equation arising out of (13) for t = 1 from Section VII-A,
does not satisfy assumption (E1). In the following we provide a more computationally orientated approach
to establishing existence and/or solving the generic fixed-point equation (27).
We motivate this case-by-case approach with the help of an example. Suppose we hypothesize that the
solution to (27) is such that x = 0, w = 0 and y, z ∈ (0, 1). Then (27) effectively reduces to checking if
there exists y ∗ , z ∗ ∈ (0, 1) such that
y ∗ ∈ argmax b f2 (0, y ∗ , 0, z ∗ ) z ∗ ∈ argmax d f4 (0, y ∗ , 0, z ∗ )
b
(28a)
d
f1 (0, y ∗ , 0, z ∗ ) ≤ 0
f3 (0, y ∗ , 0, z ∗ ) ≤ 0.
(28b)
Thus the 4-variable system reduces to solving a 2-variable system and 2 conditions to verify. For instance, if
f2 (0, y, 0, z), f4 (0, y, 0, z) as functions of y, z satisfy the conditions of Theorem 4 then the sub-system (28a)
has a solution. If one of these solution is also consistent with (28b) then this sub-case indeed provides a
solution to (27).
16
Generalizing the simplification provided in the above example, we divide solutions into 34 = 81 cases8
based on whether each of x, y, w, z are in {0}, (0, 1), {1}. There are (1) 16 corner cases where none are
in the strict interior (0, 1); (2) 32 cases where exactly one is in the strict interior (0, 1); (3) 24 cases where
2 variables are in the strict interior (0, 1); (4) 8 cases where 3 variables are in the strict interior (0, 1);
(5) 1 case where all 4 variables are in the strict interior (0, 1).
Similar to the calculations above, for each of the 81 cases one can write a sub-system to which the
problem (27) effectively reduces to. Clearly, if any one of the 81 sub-systems has a solution then the
problem (27) has a solution. Furthermore, searching for a solution reduces to an appropriate sub-problem
depending on the case.
The approach then is to enumerate each of these 81 cases (as stated above) and check them in order.
However this case-by-case division provides a computational simplification - not all cases require solving
the entire fixed-point equation. Whenever a variable, say y, is not in the strict interior (0, 1) then the
corresponding equation (27) need not be solved, since one only needs to verify the sign at a specific
point. Hence, all sub-cases of (1) reduce to simply checking the value of functions fi at corner points no need for solving a fixed-point equation. All sub-cases of (2) reduce to solving a 1-variable fixed-point
equation and three corresponding conditions to verify, etc.
VII. A C ONCRETE E XAMPLE OF M ULTI - STAGE INVESTMENT IN P UBLIC G OODS
Here we discuss both, a two-stage (finite) and an infinite-horizon version of a public goods example to
illustrate the methodology described above for the construction of SPBE.
A. A two stage public goods game
We consider a discrete version of Example 8.3 from [5, ch.8], which is an instance of a repeated public
goods game. There are two players who play a two-period game. In each period t, they simultaneously
decide whether to contribute to the period t public good, which is a binary decision ait ∈ {0, 1} for
players i = 1, 2. Before the start of period 2, both players know the action profile from period 1. In each
period, each player gets a reward of 1 if at least one player contributed and 0 if none contributed. Player
i’s cost of contributing is xi which is its private information. Both players believe that xi ’s are drawn
independently and identically with probability distribution Q with support {xL , xH }; 0 < xL < 1 < xH ,
and PQ (X i = xH ) = q ∈ (0, 1).
In our model this corresponds to N = 2, T = 2 and reward for player i in period t is Rti (x, at ) =
i
δ t Ri (x, at ), with Ri (x, at ) = (1 − xi )1(ait = 1) + a−i
t 1(at = 0). We set δ = 1 in this two-stage case. We
use the backward recursive algorithm from Section IV to find an SPBE of this game. Here the partial
iH
functions γti can equivalently be defined through the scalars piL
∈ [0, 1], for t = 1, 2 and i = 1, 2,
t , pt
where
γti (1|xL ) = piL
t ,
γti (1|xH ) = piH
t ,
γti (0|xL ) = 1 − piL
t ,
(29a)
γti (0|xH ) = 1 − piH
t ,
(29b)
iH
Henceforth, piL
is used interchangeably with γti .
t , pt
8
Generally, the number of cases is 3
PN
i=1
Mi
where N is the number of agents and Mi is the number of types for player i.
17
For t = 2 and for any fixed π 2 = (π21 , π22 ), where π2i = π2i (xH ) ∈ [0, 1] represents a probability measure
2L 1H
2H
on the event {X i = xH }. Let γ̃2 = (p̃1L
2 , p̃2 , p̃2 , p̃2 ) = θ2 [π 2 ] be defined through the fixed point
equation (13). Since 1 − xH < 0, p̃iH
2 = 0 is the solution. Thus the fixed-point equation can be reduced
to, ∀i ∈ {1, 2},
iL
−i −iL
L
+ piL
p̃iL
2 (1 − x ).
2 ∈ arg max(1 − p2 )(1 − π2 )p̃2
piL
2
This implies (31) below, the solutions to which are
0
if
1
if
p̃iL
2 =
arbitrary if
(30)
shown in Figure 1 in the space of (π21 , π22 ).
xL > 1 − (1 − π2−i )p̃−iL
2 ,
xL < 1 − (1 − π2−i )p̃−iL
2 ,
(31)
xL = 1 − (1 − π2−i )p̃−iL
2 .
2H
2L 1H
Fig. 1: Solutions of fixed point equation in (31). Solutions are shown as quadruplets (p̃1L
2 , p̃2 , p̃2 , p̃2 )
with intervals used whenever the solution is not uniquely defined.
Thus for any π 2 , there can exist multiple equilibria and correspondingly multiple θ2 [π 2 ] can be defined.
For any particular θ2 , at t = 1, the fixed point equation arising out of (13) defines θ1 [Q2 ], where Q2 = Q×Q
denotes the profile of initial belief.
Using one such θ2 defined below, we find an SPBE of the game for q = 0.1, xL = 0.2, xH = 1.2. We
use θ2 [π 2 ] as one possible set of solutions of (31), described below,
L 1−xL
( 1−x
π21 ∈ [0, xL ), π22 ∈ [0, xL )
1 , 1−π 2 , 0, 0)
1−π
2
2
(1, 0, 0, 0)
π21 ∈ [0, xL ], π22 ∈ [xL , 1]
2H
1L 2L 1H
θ2 [π2 ] = (p̃2 , p̃2 , p̃2 , p̃2 ) =
(32)
(0, 1, 0, 0)
π21 ∈ [xL , 1], π22 ∈ [0, xL ]
(1, 1, 0, 0)
π21 ∈ (xL , 1], π22 ∈ (xL , 1].
Then, through iteration on the fixed point equation and using the aforementioned θ2 [π 2 ], we numerically
2L 1H
2H
find (and analytically verify) that θ1 [Q2 ] = (p̃1L
1 , p̃1 , p̃1 , p̃1 ) = (0, 1, 0, 0) is a fixed point. Thus
β11 (A11 = 1|X 1 = xL ) = 0
β12 (A21 = 1|X 2 = xL ) = 1
β11 (A11 = 1|X 1 = xH ) = 0
β12 (A21 = 1|X 2 = xH ) = 0
18
with beliefs µ∗2 [00] = (q, 1), µ∗2 [01] = (q, 0), µ∗2 [10] = (q, 1), µ∗2 [11] = (q, 0) and (β2i (·|a1 , ·))i∈{1,2} =
θ2 [µ∗2 [a1 ]] is an SPBE of the game. In this equilibrium, player 2 at time t = 1, contributes according to
her type whereas player 1 never contributes, thus player 2 reveals her private information through her action
whereas player 1 does not. Since θ2 is symmetric, there also exists an (antisymmetric) equilibrium where
at time t = 1, players’ strategies reverse i.e. player 2 never contributes and player 1 contributes according
1−xL
1−xL
to her type. We can also obtain a symmetric equilibrium where θ1 [Q2 ] = ( (1−q)(1+x
L ) , (1−q)(1+xL ) , 0, 0) as
q
, resulting in beliefs µ∗2 [00] = (p, p), µ∗2 [01] = (p, 0), µ∗2 [10] = (0, p), µ∗2 [11] =
a fixed point when xL > 2−q
(0, 0) where p =
q(1+xL )
.
q(1+xL )+(1−xL )
B. Infinite horizon version
For the infinite horizon version we consider three values δ = 0, 0.5, 0.95 and solve the corresponding
fixed point equation (arising out of (20)) numerically to calculate the mapping θ. The fixed-point equation
is solved numerically by discretizing the π−space [0, 1]2 and all solutions that we find are symmetric
w.r.t. players i.e., p̃1L for π = (π 1 , π 2 ) is the same as p̃2L for π 0 = (π 2 , π 1 ) and similarly for p̃1H , p̃2H .
For δ = 0, the game is instantaneous and actually corresponds to the second round t = 2 play in
the finite horizon two-stage version above. Thus whenever player 1’s type is xH , it is instantaneously
profitable not to contribute. This gives p̃1H = 0, for all π. Thus we only plot p̃1L ; in Fig. 2 (this can be
inferred from the discussion and Fig. 1 above). Intuitively, with type xL the only values of π for which
player 1 would not wish to contribute is if he anticipates player 2’s type to be xL with high probability
and rely on player 2 to contribute. This is why for lower values of π 2 (i.e., player 2’s type likely to be
xL ) we see p̃1L = 0 in Fig. 2.
Fig. 2: p̃1L vs. (π 1 , π 2 ) at δ = 0; (p̃1L , p̃1H ) = θ1 [π 1 , π 2 ].
Now consider p̃1L plotted in Fig. 2 and 3. As δ increases, future rewards attain more priority and
signaling comes into play. So while taking an action, players not only look for their instantaneous reward
but also how their action affects the future public belief π about their private type. It is evident in the
figures that as δ increases, at high π 1 , up to larger values of π 2 player 1 chooses not to contribute when
his type is xL . This way he intends to send a “wrong” signal to player 2 i.e., that his type is xH and
subsequently force player 2 to invest. This way player 1 can free-ride on player 2’s investment.
Now consider p̃1H plotted in Fig. 3. For δ = 0 we know that not contributing is profitable, however as δ
increases from 0, players are mindful of future rewards and thus are willing to contribute at certain beliefs.
Specifically, coordination via signaling is evident here. Although it is instantaneously not profitable to
19
Fig. 3: p̃1L , p̃1H vs. (π 1 , π 2 ) at δ = 0.5 (upper left and right). p̃1L , p̃1H vs. (π 1 , π 2 ) at δ = 0.95 (lower left
and right).
contribute if player 1’s type is xH , by contributing at higher values of π 2 (i.e., player 2’s type is likely
xH ) and low π 1 , player 1 coordinates with player 2 to achieve net profit greater than 0 (reward when
no one contributes). This is possible since the loss when contributing is −0.2 whereas the profit from
free-riding on player 2’s contribution is 1.
Under the equilibrium strategy, beliefs Πt form a Markov chain. One can trace this Markov chain
to study the signaling effect at equilibrium. On numerically simulating this Markov chain for the above
example (at δ = 0.95) we observe that for almost all initial beliefs, within a few rounds players completely
learn each other’s private type truthfully (or at least with very high probability). In other words, players
manage to reveal their private type via their actions at equilibrium and to such an extent that it negates
any possibly incorrect initial belief about their type.
As a measure of cooperative coordination at equilibrium one can perform the following calculation.
Compare the value function V 1 (·, x) of agent 1 arising out of the fixed-point equation, for δ = 0.95 and
x ∈ {xH , xL } (normalize it by multiplying with 1 − δ so that it represents per-round value) with the best
possible attainable single-round reward under a symmetric mixed strategy with a) full coordination and
b) no coordination. Note that the two cases need not be equilibrium themselves, which is why this will
result in a bound on the efficiency of the evaluated equilibria.
In case a), assuming both agents have the same type x, full coordination can lead to the best possible
reward of 1+1−x
= 1 − x2 i.e., agent 1 contributes with probability 0.5 and agent 2 contributes with
2
probability 0.5 but in a coordinated manner so that it doesn’t overlap with agent 1 contributing.
In case b) when agents do not coordinate and invest with probability p each, then the expected singleround reward is p(1 − x) + p(1 − p). The maximum possible value of this expression is (1 − x2 )2 .
For x = xL = 0.2, the range of values of V 1 (π1 , π2 , xL ) over (π1 , π2 ) ∈ [0, 1]2 is [0.865, 0.894]. Whereas
full coordination produces 0.9 and no coordination 0.81. It is thus evident that agents at equilibrium end
20
up achieving reward close to the best possible and gain significantly compared to the strategy of no
coordination.
Similarly for x = xH = 1.2 the range is [0.3, 0.395]. Whereas full coordination produces 0.4 and no
coordination 0.16. The gain via coordination is evident here too.
VIII. C ONCLUSION
In this paper we presented a methodology for evaluating SPBE for games with asymmetric information
and independent private types evolving as controlled Markov processes. The main contribution is a time
decomposition akin to dynamic programming. This decomposition allows one to find SPBE that exhibit
signaling behavior with linear complexity in the time horizon. Using this methodology, dynamic LQG
games with asymmetric information are studied in [39] where it is shown that under certain conditions,
there exists an SPBE of the game with strategies being linear in players’ private types. In [40], authors
extend the finite-horizon model in this paper such that players do not observe their own types, rather make
independent noisy observations of their types. An analogous backward-forward algorithm is presented for
that model. It is worth noting that although structured strategies are useful in making the equilibrium
finding process tractable, no claim can be made about whether the resulting equilibrium outcomes are
better or worse than those corresponding to general strategies. We believe this is an interesting future
research direction. Another interesting future direction is dynamic mechanism design for asymmetric
information systems.
ACKNOWLEDGMENT
The authors wish to acknowledge Vijay Subramanian for his contribution to the paper. Achilleas
Anastasopoulos wishes to acknowledge Ashutosh Nayyar for the fruitful discussion and criticism of an
early draft of this work presented during the ITA 2012 conference.
A PPENDIX A
P ROOF OF L EMMA 1
We prove this Lemma in the following steps.
(a) In Claim 1, we prove that for any policy profile g and ∀t ∈ T , xi1:t for i ∈ N are conditionally
independent given the common information a1:t .
(b) In Claim 2, using Claim 1, we prove that for every fixed strategy g −i of the players −i, ((A1:t−1 , Xti ), Ait )t∈T
is a controlled Markov process for player i.
i 4 g i
i
(c) For a given policy g, we define a policy si of player i from g as sit (ait |a1:t−1
, xt ) = P (at |a1:t−1
, xt ).
(d) In Claim 3, we prove that the dynamics of this controlled Markov process (A1:t−1 , Xti ), Ait
under
i −i
t∈T
(si g −i ) are same as under g i.e. Ps g (xit , xit+1 , a1:t ) = Pg (xit , xit+1 , a1:t ).
(e) In Claim 4, we prove that w.r.t. random variables (xt , at ), xit is sufficient for player i’s private information
−i
history xi1:t i.e. Pg (xt , at |a1:t−1 , xi1:t , ait ) = Pg (xt , at |a1:t−1 , xit , ait ).
i −i
(f) From (c), (d) and (e) we then prove the result of the lemma that Ps g (xt , at ) = Pg (xt , at ).
21
Claim 1: For any policy profile g and ∀t,
Pg (x1:t |a1:t−1 ) =
N
Y
i
Pg (xi1:t |a1:t−1 )
(34)
i=1
Proof:
Pg (x1:t , a1:t−1 )
g
x1:t P (x1:t , a1:t−1 )
Qt
QN
i
i
i
i i
i
i
i
i
i
i
n=2 Qn (xn |xn−1 , an−1 )gn (an |a1:n−1 , x1:n )
i=1 Q1 (x1 )g1 (a1 |x1 )
= P QN
Qt
i
i
i
i i
i
i i
i
i
i
i=1 Q (x1 )g1 (a1 |x1 )
n=2 Qn (xn |xn−1 , an−1 )gn (an |a1:n−1 , x1:n )
x1:t
QN
Qt
i
i
i
i i
i
i
i
i
i
i
Q
(x
)g
(a
|x
)
Q
(x
|x
,
a
)g
(a
|a
,
x
)
n−1
1:n−1
n n
1:n
n=2 n n n−1
P 1 1 1 1 1
= Q i=1
Qt
N
i
i
i
i
i i
i
i
i
i
i
|x
(a
)g
)
Q
(x
|x
,
a
)g
(a
|a
,
x
)
Q
(x
i
1 1 1 1
1:n
i=1
n=2 n n n−1 n−1 n n 1:n−1
x1:t
Q
N
Y
Qi (xi )g i (ai |xi ) t Qin (xin |xin−1 , an−1 )gni (ain |a1:n−1 , xi1:n )
=
Qt
P 1 i1 i1 i1 i1 i n=2
i
i
i
i
i
i
Q
(x
)g
(a
|x
)
i
1
1
1
1
n=2 Qn (xn |xn−1 , an−1 )gn (an |a1:n−1 , x1:n )
x
i=1
Pg (x1:t |a1:t−1 ) = P
(35a)
(35b)
(35c)
(35d)
1:t
=
N
Y
i
Pg (xi1:t |a1:t−1 )
(35e)
i=1
Claim 2: For a fixed g −i , {(A1:t−1 , Xti ), Ait }t is a controlled Markov process with state (A1:t−1 , Xti ) and
control action Ait .
Proof:
Pg (ã1:t , xit+1 |a1:t−1 , xi1:t , ai1:t )
X
i
i
=
Pg (ã1:t , xit+1 , x−i
1:t |a1:t−1 , x1:t , at )
(36a)
x−i
1:t
=
X
i
−i
i
i
i
Pg (ã−i
t , xt+1 , x1:t |a1:t−1 , x1:t , at )I(a1:t−1 ,ait ) (ã1:t−1 , ãt )
(36b)
x−i
1:t
!
=
X
P
x−i
1:t
=P
g −i
g −i
(x−i
1:t |a1:t−1 )
Y
i
gtj (ãjt |a1:t−1 , xj1:t ) Qit (xit+1 |xit , ait , ã−i
t )I(a1:t−1 ,ait ) (ã1:t−1 , ãt )
(36c)
j6=i
(ã1:t , xit+1 |a1:t−1 , xit , ait ),
(36d)
i
where (36c) follows from Claim 1 since x−i
1:t is conditionally independent of x1:t given a1:t−1 and the
corresponding probability is only a function of g −i .
For any given policy profile g, we construct a policy si in the following way,
4
sit (ait |a1:t−1 , xit ) = Pg (ait |a1:t−1 , xit )
P
g i
i
xi1:t−1 P (at , x1:t |a1:t−1 )
=P P
i
i
g i
ãit
x̃i1:t−1 P (ãt , x̃1:t−1 xt |a1:t−1 )
P
gi
i
i i
i
xi1:t−1 P (x1:t |a1:t−1 )gt (at |a1:t−1 , x1:t )
=P P
Pgi (x̃i1:t−1 xit |a1:t−1 )gti (ãit |a1:t−1 , x̃i1:t−1 xit )
ãit
x̃i
1:t−1
(37a)
(37b)
(37c)
22
i
= Pg (ait |a1:t−1 , xit ),
(37d)
where dependence of (37c) on only g i is due to Claim 1.
Claim 3: The dynamics of the Markov process {(A1:t−1 , Xti ), Ait }t under (si g −i ) are the same as under
g i.e.,
i −i
Ps g (xit , xit+1 , a1:t ) = Pg (xit , xit+1 , a1:t )
∀t
(38)
Proof: We prove this by induction. Clearly,
i −i
Pg (xi1 ) = Ps g (xi1 ) = Qi1 (xi1 ).
(39)
i −i
Now suppose (38) is true for t − 1 which also implies that the marginals Pg (xit , a1:t−1 ) = Ps g (xit , a1:t−1 ).
Then
Pg (xit , a1:t−1 , xit+1 , at ) = Pg (xit , a1:t−1 )Pg (ait |a1:t−1 , xit )Pg (xit+1 , a1:t |xit , a1:t−1 , ait )
=P
si g −i
(xit , a1:t−1 )sit (ait |a1:t−1 , xit )P
=P
si g −i
(xit , a1:t−1 , xit+1 , at ),
g −i
(xit+1 , a1:t |xit , a1:t−1 , ait )
(40a)
(40b)
(40c)
where (40b) is true from induction hypothesis, definition of si in (37d) and since {(a1:t−1 , xit ), ait }t is a
controlled Markov process as proved in Claim 2 and its update kernel does not depend on policy g i .This
completes the induction step.
Claim 4: For any policy g,
−i
Pg (x̃t , ãt |a1:t−1 , xi1:t , ait ) = Pg (x̃t , ãt |a1:t−1 , xit , ait ).
(41)
−i
i
Pg (x̃t , ãt |a1:t−1 , xi1:t , ait ) = Ixit ,ait (x̃it , ãit )Pg (x̃−i
t , ãt |a1:t−1 , x1:t ).
(42)
Proof:
Now
−i
i
Pg (x̃−i
t , ãt |a1:t−1 , x1:t ) =
X
−i
i
Pg (x̃−i
1:t , ãt |a1:t−1 , x1:t )
(43a)
x̃−i
1:t−1
!
=
X
i
Pg (x̃−i
1:t |a1:t−1 , x1:t )
x̃−i
1:t−1
Y
gtj (ãjt |a1:t−1 , x̃j1:t )
(43b)
j6=i
!
=
X
−i
Pg (x̃−i
1:t |a1:t−1 )
x̃−i
1:t
Y
gtj (ãjt |a1:t−1 , x̃j1:t )
(43c)
j6=i
−i
−i
= Pg (x̃−i
t , ãt |a1:t−1 )
(43d)
where (43c) follows from Claim 1.
Hence
−i
−i
Pg (x̃t , ãt |a1:t−1 , xi1:t , ait ) = Ixit ,ait (x̃it , ãit )Pg (x̃−i
t , ãt |a1:t−1 )
(44a)
23
−i
= Pg (x̃t , ãt |a1:t−1 , xit , ait )
(44b)
Finally,
Pg (x̃t , ãt ) =
X
Pg (x̃t , ãt |a1:t−1 , xi1:t , ait )Pg (a1:t−1 , xi1:t , ait )
(45a)
a1:t−1 xi1:t ait
=
−i
X
Pg (x̃t , ãt |a1:t−1 , xit , ait )Pg (a1:t−1 , xi1:t , ait )
(45b)
a1:t−1 xi1:t ,ait
=
X
−i
Pg (x̃t , ãt |a1:t−1 , xit , ait )Pg (a1:t−1 , xit , ait )
(45c)
a1:t−1 xit ,ait
=
X
−i
i −i
Pg (x̃t , ãt |a1:t−1 , xit , ait )Ps g (a1:t−1 , xit , ait )
(45d)
a1:t−1 xit ,ait
i −i
= Ps g (x̃t , ãt ).
(45e)
where (45b) follows from (41) in Claim 4 and (45d) from (38) in Claim 3.
A PPENDIX B
P ROOF OF L EMMA 2
For this proof we will assume the common agents strategies to be probabilistic as opposed to being
deterministic, as was the case in Section III. This means actions of the common agent, γti ’s are generated
probabilistically from ψ i as Γit ∼ ψti (·|a1:t−1 ), as opposed to being deterministically generated as γti =
ψti [a1:t−1 ], as before. These two are equivalent ways of generating actions ait from a1:t−1 and xit . We avoid
using the probabilistic strategies of common agent throughout the main text for ease of exposition, and
because it conceptually does not affect the results.
Proof: We prove this lemma in the following steps. We view this problem from the perspective
of a common agent. Let ψ be the coordinator’s policy corresponding to policy profile g. Let πti (xit ) =
i
Pψ (xit |a1:t−1 ).
Q
i
i i
(a) In Claim 5, we show that πt can be factorized as πt (xt ) = N
i=1 πt (xt ) where each πt can be updated
i
through an update function πt+1
= F i (πti , γti , at ) and F i is independent of common agent’s policy ψ.
(b) In Claim 6, we prove that (Πt , Γt )t∈T is a controlled Markov process.
4
(c) We construct a policy profile θ from g such that θt (dγt |πt ) = Pψ (dγt |πt ).
(d) In Claim 7, we prove that dynamics of this Markov process (Πt , Γt )t∈T under θ is same as under ψ i.e.
Pθ (dπt , dγt , dπt+1 ) = Pψ (dπt , dγt , dπt+1 ).
(e) In Claim 8, we prove that with respect to random variables (Xt , At ), πt can summarize common
information a1:t−1 i.e. Pψ (xt , at |a1:t−1 , γt ) = P(xt , at |πt , γt ).
(f) From (c), (d) and (e) we then prove the result of the lemma that Pψ (xt , at ) = Pθ (xt , at ) which is
equivalent to Pg (xt , at ) = Pm (xt , at ), where m is the policy profile of players corresponding to θ.
Q
i i
i
Claim 5: πt can be factorized as πt (xt ) = N
i=1 πt (xt ) where each πt can be updated through an
i
update function πt+1
= F i (πti , γti , at ) and F i is independent of common agent’s policy ψ. We also say
π t+1 = F (π t , γt , at ).
24
Proof: We prove this by induction. Since π1 (x1 ) =
Q
i
suppose πt = N
i=1 πt . Then,
QN
i=1
Qit (xi1 ), the base case is verified. Now
πt+1 (xt+1 ) = Pψ (xt+1 |a1:t , γ1:t+1 )
(46a)
= Pψ (xt+1 |a1:t , γ1:t )
P ψ
P (xt , at , xt+1 |a1:t−1 , γ1:t )
= P xt
ψ
x̃t+1 x̃t P (x̃t , x̃t+1 , at |a1:t−1 , γ1:t )
P
QN i i i i i
i
xt πt (xt )
i=1 γt (at |xt )Qt (xt+1 |xt , at )
=P
QN i i i i i
i
i=1 γt (at |x̃t )Qt (x̃t+1 |x̃t , at )
x̃t x̃t+1 πt (x̃t )
P i i i i i i i
N
i
Y
xit πt (xt )γt (at |xt )Qt (xt+1 |xt , at )
P i i i i i
=
x̃it πt (x̃t )γt (at |x̃t )
i=1
=
N
Y
i
πt+1
(xit+1 ),
(46b)
(46c)
(46d)
(46e)
(46f)
i=1
where (46e) follows from induction hypothesis. It is assumed in (46c)-(46e) that the denominator is not
0. If denominator corresponding to any γti is zero, we define
X
i
πt+1
(xit+1 ) =
πti (xit )Qit (xit+1 |xit , at ),
(47)
xit
i
= F i (πti , γti , at ) and π t+1 = F (π t , γt , a1 ) where F i and F are
where πt+1 still satisfies (46f). Thus πt+1
appropriately defined from above.
Claim 6: (Πt , Γt )t∈T is a controlled Markov process with state Πt and control action Γt
Proof:
X
Pψ (dπt+1 , at , xt |π1:t , γ1:t )
(48a)
Pψ (dπt+1 |π1:t , γ1:t ) =
at ,xt
=
X
Pψ (xt |π1:t , γ1:t )
at ,xt
=
X
at ,xt
(N
Y
)
γti (ait |xit ) IF (πt ,γt ,at ) (πt+1 )
(48b)
i=1
πt (xt )
(N
Y
)
γti (ait |xit ) IF (πt ,γt ,at ) (πt+1 )
(48c)
i=1
= P(dπt+1 |πt , γt ).
(48d)
For any given policy profile ψ, we construct policy profile θ in the following way.
4
θt (dγt |πt ) = Pψ (dγt |πt ).
(49)
Claim 7:
Pψ (dπt , dγt , dπt+1 ) = Pθ (dπt , dγt , dπt+1 )
∀t ∈ T .
(50)
25
Proof: We prove this by induction. For t = 1,
Pψ (dπ1 ) = Pθ (dπ1 ) = IQ (π1 ).
(51)
Now suppose P ψ (dπt ) = P θ (dπt ) is true for t, then
Pψ (dπt , dγt , dπt+1 ) = Pψ (dπt )P ψ (dγt |πt )Pψ (dπt+1 |πt γt )
(52a)
= Pθ (dπt )θt (dγt |πt )P (dπt+1 |πt , γt )
(52b)
= Pθ (dπt , dγt , dπt+1 ).
(52c)
where (52b) is true from induction hypothesis, definition of θ in (49) and since (Πt , Γt )t∈T is a controlled
Markov process as proved in Claim 6 and thus its update kernel does not depend on policy ψ. This
completes the induction step.
Claim 8: For any policy ψ,
Pψ (xt , at |a1:t−1 , γt ) = P(xt , at |πt , γt ).
(53)
Proof:
Pψ (xt , at |a1:t−1 , γt ) = Pψ (xt |a1:t−1 , γt )
Y
γti (ait |xit )
(54a)
i∈N
= πt (xt )
Y
γti (ait |xit )
(54b)
i∈N
= P(xt , at |πt , γt ).
(54c)
Pψ (xt , at |a1:t−1 , γt )Pψ (a1:t−1 , γt )
(55a)
P(xt , at |πt , γt )Pψ (a1:t−1 , γt )
(55b)
Finally,
Pψ (xt , at ) =
X
a1:t−1 ,γt
=
X
a1:t−1 γt
=
X
P(xt , at |πt , γt )Pψ (πt , γt )
(55c)
P(xt , at |πt , γt )Pθ (πt , γt )
(55d)
πt ,γt
=
X
πt ,γt
= Pθ (xt , at ).
where (55b) follows from (53), (55c) is due to change of measure and (55d) follows from (50).
(55e)
26
A PPENDIX C
P ROOF OF T HEOREM 1
Proof: We prove (18) using induction and the results in Lemma 3, 4 and 5 proved in Appendix D.
For base case at t = T , ∀i ∈ N , (a1:T −1 , xi1:T ) ∈ HTi , β i
n
o
∗,i ∗,−i
∗
(56a)
EβT βT , µT [a1:T −1 ] RTi (XT , AT ) a1:T −1 , xi1:T = VTi (µ∗T [a1:T −1 ], xiT )
o
n
i ∗,−i
∗
≥ EβT βT , µT [a1:T −1 ] RTi (XT , AT ) a1:T −1 , xi1:T , (56b)
where (56a) follows from Lemma 5 and (56b) follows from Lemma 3 in Appendix D.
c
Let the induction hypothesis be that for t + 1, ∀i ∈ N , a1:t ∈ Ht+1
, xi1:t+1 ∈ (X i )t+1 , β i ,
( T
)
X
∗,i
∗,−i
∗
Eβt+1:T βt+1:T , µt+1 [a1:t ]
Rni (Xn , An ) a1:t , xi1:t+1
(57a)
n=t+1
∗,−i
i
βt+1:T
βt+1:T
, µ∗t+1 [a1:t ]
≥E
(
T
X
)
Rni (Xn , An ) a1:t , xi1:t+1
.
(57b)
n=t+1
Then ∀i ∈ N , (a1:t−1 , xi1:t ) ∈ Hti , β i , we have
( T
)
X
∗,i ∗,−i
∗
Eβt:T βt:T , µt [a1:t−1 ]
Rni (Xn , An ) a1:t−1 , xi1:t
=
n=t
i
∗
i
Vt (µt [a1:t−1 ], xt )
(58a)
o
n
i
i
≥E
Rti (Xt , At ) + Vt+1
(µ∗t+1 [a1:t−1 At ], Xt+1
) a1:t−1 , xi1:t
(
( T
X
∗,i
∗,−i
βti βt∗,−i , µ∗t [a1:t−1 ]
i
βt+1:T
βt+1:T
, µ∗t+1 [a1:t−1 ,At ]
=E
Rt (Xt , At ) + E
Rni (Xn , An )
βti βt∗,−i , µ∗t [a1:t−1 ]
(58b)
n=t+1
o
a1:t−1 , At , xi1:t+1 a1:t−1 , xi1:t
i ∗,−i
∗
≥ Eβt βt , µt [a1:t−1 ] Rti (Xt , At )+
( T
)
)
X
∗,−i
i
∗
i
Eβt+1:T βt+1:T µt+1 [a1:t−1 ,At ]
Rni (Xn , An ) a1:t−1 , At , xi1:t , Xt+1
a1:t−1 , xi1:t
βti βt∗,−i , µ∗t [a1:t−1 ]
=E
i β ∗,−i µ∗ [a
βt:T
t 1:t−1 ]
t:T
E
(
(58c)
(58d)
n=t+1
n
i
Rt (Xt , At )+
T
X
)
i
Rni (Xn , An ) a1:t−1 , At , xi1:t , Xt+1
a1:t−1 , xi1:t
o
(58e)
n=t+1
i β ∗,−i
βt:T
t:T
=E
µ∗t [a1:t−1 ]
( T
X
)
Rni (Xn , An ) a1:t−1 , xi1:t
,
(58f)
n=t
where (58a) follows from Lemma 5, (58b) follows from Lemma 3, (58c) follows from Lemma 5, (58d)
follows from induction hypothesis in (57b) and (58e) follows from Lemma 4. Moreover, construction of
θ in (13), and consequently definition of β ∗ in (16) are pivotal for (58e) to follow from (58d).
We note that µ∗ satisfies the consistency condition of [5, p. 331] from the fact that (a) for all t and for
every common history a1:t−1 , all players use the same belief µ∗t [a1:t−1 ] on xt and (b) the belief µ∗t can be
Q
∗,i
∗,i
c
i
factorized as µ∗t [a1:t−1 ] = N
i=1 µt [a1:t−1 ] ∀a1:t−1 ∈ Ht where µt is updated through Bayes’ rule F as
in Lemma 5 in Appendix B.
27
A PPENDIX D
I NTERMEDIATE LEMMAS USED IN PROOF OF T HEOREM 1
Lemma 3: ∀t ∈ T , i ∈ N , (a1:t−1 , xi1:t ) ∈ Hti , βti
Vti (µ∗t [a1:t−1 ], xit ) ≥
βti βt∗,−i , µ∗t [a1:t−1 ]
E
o
n
i
i
i
i
∗
∗
Rt (Xt , At ) + Vt+1 (F (µt [a1:t−1 ], βt (·|a1:t−1 , ·), At ), Xt+1 ) a1:t−1 , x1:t . (59)
Proof: We prove this Lemma by contradiction.
Suppose the claim is not true for t. This implies ∃i, βbti , b
a1:t−1 , x
bi1:t such that
o
n
i
βbti βt∗,−i , µ∗t [b
a1:t−1 ]
i
i
i
∗
∗
E
Rt (Xt , At ) + Vt+1 (F (µt [b
a1:t−1 , x
b1:t
a1:t−1 , ·), At ), Xt+1 ) b
a1:t−1 ], βt (·|b
> Vti (µ∗t [b
a1:t−1 ], x
bit ). (60)
We will show that this leads to a contradiction.
Construct
(
bit
a1:t−1 , x
bi1:t ) xit = x
βbti (ait |b
γ
bti (ait |xit ) =
arbitrary
otherwise.
(61)
Then for b
a1:t−1 , x
bi1:t , we have
Vti (µ∗t [b
a1:t−1 ], x
bit )
o
n
∗
∗
i
i
i i −i
i
= max
E
a1:t−1 ], βt (·|b
a1:t−1 , ·), At ), Xt+1 ) x
bt , (62a)
Rt (b
xt xt , at ) + Vt+1 (F (µt [b
γti (·|b
xit )
o
n
∗,−i
i
i
∗
i
i
a1:t−1 ], βt∗ (·|b
a1:t−1 , ·), At ), Xt+1
)x
bit
(62b)
≥ Eγbt (·|bxt )βt , µt [ba1:t−1 ] Rti (Xt , At ) + Vt+1
(F (µ∗t [b
n
o
X
i
∗
=
Rti (b
xit x−i
a1:t−1 ], βt∗ (·|b
a1:t−1 , ·), at ), xit+1 )
t , at ) + Vt+1 (F (µt [b
γti (·|b
xit )βt∗,−i , µ∗t [b
a1:t−1 ]
x−i
t ,at ,xt+1
=
i i
× µ∗,−i
[b
a1:t−1 ](x−i
γti (ait |b
xit )βt∗,−i (a−i
a1:t−1 , x−i
xit , at )
t
t )b
t |b
t )Qt (xt+1 |b
o
n
X
∗
∗
i
i i −i
i
a1:t−1 ], βt (·|b
a1:t−1 , ·), at ), xt+1 )
Rt (b
xt xt , at ) + Vt+1 (F (µt [b
(62c)
x−i
t ,at ,xt+1
i i
bi i a1:t−1 , x
a1:t−1 , x−i
xit , at )
bi1:t )βt∗,−i (a−i
× µ∗,−i
[b
a1:t−1 ](x−i
t
t |b
t )Qt (xt+1 |b
t )βt (at |b
o
n
bi ∗,−i ∗
i
∗
∗
i
i
= Eβt βt ,µt [ba1:t−1 ] Rti (b
xit x−i
,
a
)
+
V
(F
)
b
a
,
x
b
(µ
[b
a
],
β
(·|b
a
,
·),
A
),
X
t
1:t−1
t
1:t−1
t+1
1:t
t
t
t+1
t 1:t−1
(62d)
> Vti (µ∗t [b
a1:t−1 ], x
bit ),
(62f)
(62e)
where (62a) follows from definition of Vti in (14), (62d) follows from definition of γ
bti and (62f) follows
from (60). However this leads to a contradiction.
i
Lemma 4: ∀i ∈ N , t ∈ T , (a1:t , xi1:t+1 ) ∈ Ht+1
and βti
( T
)
X
i β ∗,−i , µ∗ [a
βt:T
]
i
i
t 1:t−1
t:T
E
Rn (Xn , An ) a1:t , x1:t+1
(63a)
n=t+1
=E
∗,−i
i
βt+1:T
βt+1:T
, µ∗t+1 [a1:t ]
(
T
X
n=t+1
)
Rni (Xn , An ) a1:t , xi1:t+1
.
(63b)
28
Thus the above quantities do not depend on βti .
∗,−i
Proof: Essentially this claim stands on the fact that µt+1
[a1:t ] can be updated from µt∗,−i [a1:t−1 ], βt∗,−i
Q
∗,j
∗,j
and at , as µ∗,−i
t+1 [a1:t ] =
j6=i F (µt [a1:t−1 ], βt , at ) as in Claim 5. Since the above expectations involve
−i
random variables Xt+1
, At+1:T , Xt+2:T , we consider the probability
∗,−i
i
Pβt:T βt:T
, µ∗t [a1:t−1 ]
i
(x−i
t+1 , at+1:T , xt+2:T |a1:t , x1:t+1 ) =
Nr
Dr
(64)
where
Nr =
X
i
∗,−i
∗,−i
, µ∗t [a1:t−1 ]
Pβt:T βt:T
, µ∗t [a1:t−1 ]
i
(x−i
t , at , xt+1 , at+1:T , xt+2:T |a1:t−1 , x1:t )
(65a)
x−i
t
=
X
i
Pβt:T βt:T
i
(x−i
t a1:t−1 , x1:t )
x−i
t
∗,−i
i
−i
βt:T βt:T
βti (ait |a1:t−1 , xi1:t )βt∗,−i (a−i
t |a1:t−1 , xt )Q(xt+1 |xt , at )P
=
X
, µ∗t [a1:t−1 ]
(at+1:T , xt+2:T |a1:t , xi1:t−1 , xt:t+1 )
(65b)
i i
i
µt∗,−i [a1:t−1 ](x−i
t )βt (at |a1:t−1 , x1:t )
x−i
t
i
∗,−i
∗
−i
i i
i
−i −i
−i
βt+1:T βt+1:T , µt+1 [a1:t ]
βt∗,−i (a−i
(at+1:T , xt+2:T |a1:t , xi1:t , xt+1 ),
t |a1:t−1 , xt )Q (xt+1 |xt , at )Q (xt+1 |xt , at )P
(65c)
where (65c) follows from the conditional independence of types given common information, as shown
in Claim 1, and the fact that probability on (at+1:T , x2+t:T ) given a1:t , xi1:t−1 , xt:t+1 , µ∗t [a1:t−1 ] depends on
∗,−i
i
βt+1:T
. Similarly, the denominator in (64) is given by
a1:t , xi1:t , xt+1 , µ∗t+1 [a1:t ] through βt+1:T
X i ∗,−i ∗
i
i
Pβt:T βt:T , µt [a1:t−1 ] (x̃−i
Dr =
t , at , xt+1 a1:t−1 , x1:t )
x̃−i
t
X
i
∗,−i
Pβt:T βt:T
, µ∗t
∗,−i −i
i
i i
i
i i
i
(x̃−i
(at |a1:t−1 , x̃−i
t |a1:t−1 , x1:t )βt (at |a1:t−1 , x1:t )βt
t )Q (xt+1 |xt , at )
(65d)
x̃−i
t
=
X
∗,−i −i
i i
i
i i
i
µ∗,−i
[a1:t−1 ](x̃−i
(at |a1:t−1 , x̃−i
t
t )βt (at |a1:t−1 , x1:t )βt
t )Q (xt+1 |xt , at ).
(65e)
x̃−i
t
By canceling the terms βti (·) and Qi (·) in the numerator and the denominator, (64) is given by
P
∗,−i −i
−i
−i
−i
µ∗,−i
[a1:t−1 ](x−i
(at |a1:t−1 , x−i
t
t )βt
t )Qt+1 (xt+1 |xt , at )
x−i
t
P
∗,−i −i
µ∗,−i
[a1:t−1 ](x̃−i
(at |a1:t−1 , x̃−i
t
t )βt
t )
x̃−i
t
i
∗,−i
∗
× Pβt+1:T βt+1:T , µt+1 [a1:t ] (at+1:T , xt+2:T |a1:t , xi1:t , xt+1 )
−i
= µ∗,−i
t+1 [a1:t ](xt+1 )P
=P
∗,−i
i
βt+1:T
βt+1:T
, µ∗t+1 [a1:t ]
∗,−i
i
βt+1:T
βt+1:T
, µ∗t+1 [a1:t ]
(at+1:T , xt+2:T |a1:t , xi1:t , xt+1 )
i
(x−i
t+1 , at+1:T , xt+2:T |a1:t , x1:t+1 ),
(65f)
(65g)
(65h)
∗,−i
where (65g) follows from using the definition of µt+1
[a1:t ](x−i
t ) in the forward recursive step in (17) and
the definition of the belief update in (46).
Lemma 5: ∀i ∈ N , t ∈ T , (a1:t−1 , xi1:t ) ∈ Hti ,
)
( T
X
∗,i ∗,−i ∗
Rni (Xn , An ) a1:t−1 , xi1:t .
(66)
Vti (µ∗t [a1:t−1 ], xit ) = Eβt:T βt:T ,µt [a1:t−1 ]
n=t
29
Proof: We prove the lemma by induction. For t = T ,
n
o
∗,i ∗,−i
∗
EβT βT , µT [a1:T −1 ] RTi (XT , AT ) a1:T −1 , xi1:T
X
∗,−i −i
∗,i i
i
=
(aT |a1:T −1 , x−i
RTi (xT , aT )µ∗T [a1:T −1 ](x−i
T )
T )βT (aT |a1:T −1 , xT )βT
(67a)
x−i
T aT
= VTi (µ∗T [a1:T −1 ], xiT ),
(67b)
where (67b) follows from the definition of Vti in (14) and the definition of βT∗ in the forward recursion
in (16).
i
Suppose the claim is true for t + 1, i.e., ∀i ∈ N , t ∈ T , (a1:t , xi1:t+1 ) ∈ Ht+1
)
( T
X
∗,i
∗,−i
∗
i
(68)
Rni (Xn , An ) a1:t , xi1:t+1 .
(µ∗t+1 [a1:t ], xit+1 ) = Eβt+1:T βt+1:T , µt+1 [a1:t ]
Vt+1
n=t+1
Then ∀i ∈ N , t ∈ T , (a1:t−1 , xi1:t ) ∈ Hti , we have
( T
)
X
∗,i ∗,−i
∗
Eβt:T βt:T , µt [a1:t−1 ]
Rni (Xn , An ) a1:t−1 , xi1:t
n=t
∗,i ∗,−i
βt:T
βt:T , µ∗t [a1:t−1 ]
=E
( T
X
n
∗,i ∗,−i
∗
Rti (Xt , At ) + Eβt:T βt:T , µt [a1:t−1 ]
)
)
i
Rni (Xn , An ) a1:t−1 , At , xi1:t , Xt+1
a1:t−1 , xi1:t
(69a)
n=t+1
∗,i ∗,−i
∗
= Eβt:T βt:T , µt [a1:t−1 ] Rti (Xt , At )+
( T
)
)
X
i
Rni (Xn , An ) a1:t−1 , At , xi1:t , Xt+1
a1:t−1 , xi1:t
(69b)
n=t+1
n
o
i
i
∗
i
i
=E
Rt (Xt , At ) + Vt+1 (µt+1 [a1:t−1 At ], Xt+1 ) a1:t−1 , x1:t
n
o
βt∗,i βt∗,−i , µ∗t [a1:t−1 ]
i
∗
i
i
i
=E
Rt (Xt , At ) + Vt+1 (µt+1 [a1:t−1 At ], Xt+1 ) a1:t−1 , x1:t
(69d)
= Vti (µ∗t [a1:t−1 ], xit ),
(69e)
∗,i ∗,−i
βt:T
βt:T , µ∗t [a1:t−1 ]
(69c)
where (69b) follows from Lemma 4 in Appendix D, (69c) follows from the induction hypothesis in (68),
i
(69d) follows because the random variables involved in expectation, Xt−i , At , Xt+1
do not depend on
∗,i
∗,−i
βt+1:T βt+1:T and (69e) follows from the definition of βt∗ in the forward recursion in (16), the definition
of µ∗t+1 in (17) and the definition of Vti in (14).
A PPENDIX E
P ROOF OF T HEOREM 2
Proof: We prove this by contradiction. Suppose for any equilibrium generating function φ that
generates (β ∗ , µ∗ ) through forward recursion, there exists t ∈ T , i ∈ N , a1:t−1 ∈ Htc , such that for
π t = µ∗t [a1:t−1 ], (13) is not satisfied for φ i.e. for γ̃ti = φi [π t ] = βt∗,i (·|µ∗t [a1:t−1 ], xit ),
n
o
i
γti (·|xi )γ̃t−i , πt
i
i
i
i
γ̃t 6∈ arg max
E
Rt (Xt , At ) + Vt+1 (F (π t , γ̃t , At ), Xt+1 ) xt .
(70)
i
γt
30
Let t be the first instance in the backward recursion when this happens. This implies ∃ γ
bti such that
n
o
i
i −i
i
i
(F (π t , γ̃t , At ), Xt+1
) xit
Eγbt (·|x )γ̃t , πt Rti (Xt , At ) + Vt+1
o
n
i
i −i
i
i
(F (π t , γ̃t , At ), Xt+1
(71)
) xit
> Eγ̃t (·|x )γ̃t , πt Rti (Xt , At ) + Vt+1
This implies for βbt (·|µ∗t [a1:t−1 ], ·) = γ
bti ,
)
( T
X
∗,i ∗,−i
∗
Rni (Xn , An ) a1:t−1 , xi1:t
Eβt:T βt:T , µt [a1:t−1 ]
n=t
βt∗,i βt∗,−i , µ∗t [a1:t−1 ]
=E
( T
X
n
∗,i ∗,−i
∗
Rti (Xt , At ) + Eβt:T βt:T , µt [a1:t−1 ]
)
)
Rni (Xn , An ) a1:t−1 , At , xi1:t+1
a1:t−1 , xi1:t
(72a)
n=t+1
βt∗,i βt∗,−i , µ∗t [a1:t−1 ]
=E
( T
X
n
∗,i
∗,−i
∗
Rti (Xt , At ) + Eβt+1:T βt+1:T , µt+1 [a1:t−1 ,At ]
)
)
Rni (Xn , An ) a1:t−1 , At , xi1:t+1
a1:t−1 , xi1:t
(72b)
n=t+1
n
o
i
i
Rti (Xt , At ) + Vt+1
(F (π t , γ̃t , At ), Xt+1
) xit
o
n
∗
i −i
bi
i
i
) xit
< Eβt (·|µt [a1:t−1 ],xt )γ̃t , πt Rti (Xt , At ) + Vt+1
(F (π t , γ̃t , At ), Xt+1
n
∗,i
∗,−i
∗
βbti βt∗,−i , µ∗t [a1:t−1 ]
=E
Rti (Xt , At ) + Eβt+1:T βt+1:T µt+1 [a1:t−1 ,At ]
( T
)
o
X
i
Rni (Xn , An ) a1:t−1 , At , xi1:t , Xt+1
a1:t−1 , xi1:t
i
i
−i
= Eγ̃t (·|xt )γ̃t
, πt
(72c)
(72d)
(72e)
n=t+1
∗,i
∗,−i
βbti ,βt+1:T
βt:T
, µ∗t [a1:t−1 ]
=E
( T
X
)
Rni (Xn , An ) a1:t−1 , xi1:t
,
(72f)
n=t
where (72b) follows from Lemma 4, (72c) follows from the definitions of γ̃ti and µ∗t+1 [a1:t−1 , At ] and
Lemma 5, (72d) follows from (71) and the definition of βbti , (72e) follows from Lemma 3, (72f) follows
from Lemma 4. However, this leads to a contradiction since (β ∗ , µ∗ ) is a PBE of the game.
A PPENDIX F
P ROOF OF T HEOREM 3
We divide the proof into two parts: first we show that the value function V i is at least as big as any
reward-to-go function; secondly we show that under the strategy βi∗ , reward-to-go is V i .
Part 1: For any i ∈ N , β i define the following reward-to-go functions
)
(∞
X
i
i
−i,∗
∗
c
(73a)
Wti,β (hit ) = Eβ ,β ,µt [ht ]
δ n−t Ri (Xn , An ) | hit
( T
)
n=t
X
i,β i ,T
i
β i ,β −i,∗ ,µ∗t [hct ]
n−t i
δ R (Xn , An ) + δ T +1−t V i (ΠT +1 , XTi +1 ) | hit .
Wt
(ht ) = E
(73b)
n=t
i
Since X i , Ai are finite sets the reward Ri is absolutely bounded, the reward-to-go Wti,β (hit ) is finite ∀
i, t, β i , hit .
31
For any i ∈ N , hit ∈ Hti ,
i,β i
i,β i ,T
i,β i ,T
i,β i
∗ c
i
i
i
∗ c
i
i
i
i
i
(ht ) + Wt
(ht ) − Wt (ht ) (74)
V µt [ht ], xt − Wt (ht ) = V µt [ht ], xt − Wt
Combining results from Lemmas 8 and 9 in Appendix G, the term in the first bracket in RHS of (74) is
non-negative. Using (73), the term in the second bracket is
∞
X
β i ,β −i,∗ ,µ∗ [hc ]
n−(T +1) i
i
T +1−t
i
i
t
t
−
δ
R (Xn , An ) + V (ΠT +1 , XT +1 ) | ht .
E
δ
(75)
n=T +1
The summation in the expression above is bounded by a convergent geometric series. Also, V i is bounded.
Hence the above quantity can be made arbitrarily small by choosing T appropriately large. Since the LHS
of (74) does not depend on T , this results in
i
(76)
V i µ∗t [hct ], xit ≥ Wti,β (hit ).
Part 2: Since the strategy β ∗ generated in (21) is such that βti,∗ depends on hit only through µ∗t [hct ]
i,∗
and xit , the reward-to-go Wti,β , at strategy β ∗ , can be written (with abuse of notation) as
)
(∞
X
i,∗
i,∗
∗
∗
c
Wti,β (hit ) = Wti,β (µ∗t [hct ], xit ) = Eβ ,µt [ht ]
(77)
δ n−t Ri (Xn , An ) | µ∗t [hct ], xit .
n=t
For any hit ∈ Hti ,
i,∗
Wti,β (µ∗t [hct ], xit )
V
i
(µ∗t [hct ], xit )
β ∗ ,µ∗t [hct ]
= E
β ∗ ,µ∗t [hct ]
= E
o
n
i,β i,∗
∗ c
∗ c
i
∗ c
i
i
R (Xt , At ) + δWt+1 F (µt [ht ], θ[µt [ht ]], At+1 ), Xt+1 | µt [ht ], xt
(78a)
∗ c
∗ c
i
∗ c
i
i
i
R (Xt , At ) + δV F (µt [ht ], θ[µt [ht ]], At+1 ), Xt+1 | µt [ht ], xt . (78b)
Repeated application of the above for the first n time periods gives
t+n−1
X
i,β i,∗
i,β i,∗
∗ c
i
β ∗ ,µ∗t [hct ]
m−t i
n
i
∗ c
i
δ
R (Xt , At ) + δ Wt+n Πt+n , Xt+n | µt [ht ], xt
(79a)
Wt
(µt [ht ], xt ) = E
m=t
t+n−1
X
i
∗ c
i
β ∗ ,µ∗t [hct ]
∗ c
i
i
m−t i
n i
V (µt [ht ], xt ) = E
δ
R (Xt , At ) + δ V Πt+n , Xt+n | µt [ht ], xt . (79b)
m=t
Here Πt+n is the n−step belief update under strategy and belief prescribed by β ∗ , µ∗ .
Taking differences results in
i,∗
Wti,β (µ∗t [hct ], xit ) − V i (µ∗t [hct ], xit )
= δ n Eβ
∗ ,µ∗ [hc ]
t t
n
i,∗
i,β
Wt+n
o
i
i
Πt+n , Xt+n
− V i Πt+n , Xt+n
| µ∗t [hct ], xit . (80)
Taking absolute value of both sides then using Jensen’s inequality for f (x) = |x| and finally taking
supremum over hit reduces to
i,∗
sup Wti,β (µ∗t [hct ], xit ) − V i (µ∗t [hct ], xit )
hit
≤ δ n sup Eβ
hit
∗ ,µ∗ [hc ]
t t
n
o
i,β i,∗
i
Wt+n
(Πt+n , Xt+n
) − V i (µ∗t [hct ], xit ) | µ∗t [hct ], xit . (81)
Now using the fact that Wt+n , V i are bounded and that we can choose n arbitrarily large, we get
i,∗
suphit |Wti,β (µ∗t [hct ], xit ) − V i (µ∗t [hct ], xit )| = 0.
32
A PPENDIX G
I NTERMEDIATE L EMMA USED IN P ROOF OF T HEOREM 3
In this section, we present four lemmas. Lemma 6 and 7 are intermediate technical results needed in
the proof of Lemma 8. Then the results in Lemma 8 and 9 are used in Appendix F for the proof of
Theorem 3. The proofs for Lemma 6 and 7 below aren’t stated as they are analogous (the only difference
being a non-zero terminal reward in the finite horizon model) to the proofs of Lemma 3 and 4, from
Appendix D, used in the proof of Theorem 1.
i
Define the reward-to-go Wti,β ,T for any agent i and strategy β i as
i
Wti,β ,T (hit )
β i ,β −i,∗ ,µ∗t [hct ]
=E
T
hX
i
δ n−t Ri (Xn , An ) + δ T +1−t Gi (ΠT +1 , XTi +1 ) | hit .
(82)
n=t
Here agent i’s strategy is β whereas all other agents use strategy β −i,∗ defined above. Since X i , Ai are
assumed to be finite and Gi absolutely bounded, the reward-to-go is finite ∀ i, t, β i , hit .
In the following, any quantity with a T in the superscript refers the finite horizon model with terminal
reward Gi . For further discussion, please refer to the comments after the statement of Theorem 3.
Lemma 6: For any t ∈ T , i ∈ N , hit and β i ,
h
i
i −i,∗ ∗ c
i,T
i
F (µ∗t [hct ], βt∗ (·|µ∗t [hct ], ·), At ), Xt+1
Vti,T (µ∗t [hct ], xit ) ≥ Eβ ,β ,µt [ht ] Ri (Xt , At ) + δVt+1
| hit . (83)
i
Lemma 7:
i
−i,∗
∗
c
Eβt+1:T ,βt+1:T ,µt+1 [ht ,at ]
T
h X
δ n−(t+1) Ri (Xn , An ) + δ T +1−t Gi (ΠT +1 , XTi +1 ) | hit , at , xit+1
i
n=t+1
i ,β −i,∗ ,µ∗ [hc ]
βt:T
t t
t:T
=E
T
h X
δ
n−(t+1)
i
R (Xn , An ) + δ
T +1−t
i
G
(ΠT +1 , XTi +1 )
|
hit , at , xit+1
i
. (84)
n=t+1
The result below shows that the value function from the backwards recursive algorithm is higher than any
reward-to-go.
Lemma 8: For any t ∈ T , i ∈ N , hit and β i ,
i
Vti,T (µ∗t [hct ], xit ) ≥ Wti,β ,T (hit ).
(85)
Proof: We use backward induction for this. At time T , using the maximization property from (13)
(modified with terminal reward Gi ),
VTi,T (µ∗T [hcT ], xiT )
i,T
(86a)
h
c
(·|xiT ),γ̃T−i,T ,µ∗T [ht ]
i
Ri (XT , AT ) + δGi F (µ∗T [hcT ], γ̃TT , AT ), XTi +1 | µ∗T [hcT ], xiT
h
i
i,T
−i,T ∗
i
c
≥ EγT (·|xT ),γ̃T ,µT [ht ] Ri (XT , AT ) + δGi F (µ∗T [hcT ], γ̃TT , AT ), XTi +1 | µ∗T [hcT ], xiT
, Eγ̃T
i
= WTi,β ,T (hiT )
(86b)
(86c)
(86d)
Here the second inequality follows from (13) and (14) and the final equality is by definition in (82).
Assume that the result holds for all n ∈ {t + 1, . . . , T }, then at time t we have
Vti,T (µ∗t [hct ], xit )
(87a)
33
i
−i,∗
≥ Eβt ,βt
,µ∗t [hct ]
βti ,βt−i,∗ ,µ∗t [hct ]
≥E
h
i
i,T
i
| hit
Ri (Xt , At ) + δVt+1
F (µ∗t [hct ], βt∗ (·|µ∗t [hct ], ·), At ), Xt+1
T
h
h X
−i,∗
i
i
βt+1:T
,βt+1:T
,µ∗t+1 [hct ,At ]
R (Xt , At ) + δE
δ n−(t+1) Ri (Xn , An )
(87b)
(87c)
n=t+1
+δ
T −t
i
G
(ΠT +1 , XTi +1 )
i ,β −i,∗ ,µ∗ [hc ]
βt:T
t t
t:T
=E
T
hX
δ
|
i
hit , At , Xt+1
n−t
i
i
|
hit
i
R (Xn , An ) + δ
T +1−t
i
G
(ΠT +1 , XTi +1 )
|
hit
i
(87d)
n=t
i
= Wti,β ,T (hit )
(87e)
Here the first inequality follows from Lemma 6, the second inequality from the induction hypothesis, the
third equality follows from Lemma 7 and the final equality by definition (82).
The following result highlights the similarities between the fixed-point equation in infinite horizon and
the backwards recursion in the finite horizon.
Lemma 9: Consider the finite horizon game with Gi ≡ V i . Then Vti,T = V i , ∀ i ∈ N , t ∈ {1, . . . , T }
satisfies the backwards recursive construction stated above (adapted from (13) and (14)).
Proof: Use backward induction for this. Consider the finite horizon algorithm at time t = T , noting
i
i
that VTi,T
+1 ≡ G ≡ V ,
h
i
−i,T
−i
i
i
γ̃Ti,T (· | xiT ) ∈ argmax EγT (·|xT ),γ̃T ,πT Ri (XT , AT ) + δV i F (π T , γ̃TT , AT ), XTi +1 | π T , xiT (88a)
γTi (·|xiT )∈∆(Ai )
i,T
VTi,T (π T , xiT ) = Eγ̃T
(·|xiT ),γ̃T−i,T ,π −i
T
h
i
Ri (XT , AT ) + δV i F (π T , γ̃TT , AT ), XTi +1 | π T , xiT .
(88b)
Comparing the above set of equations with (20), we can see that the pair (V, γ̃) arising out of (20)
satisfies the above. Now assume that Vni,T ≡ V i for all n ∈ {t + 1, . . . , T }. At time t, in the finite horizon
i,T
construction from (13), (14), substituting V i in place of Vt+1
from the induction hypothesis, we get the
i,T
i
same set of equations as (88). Thus Vt ≡ V satisfies it.
A PPENDIX H
P ROOF OF T HEOREM 4
Denote the vector correspondence defined by the RHS of (27) by
φ1 (x)
argmaxa af1 (x)
.
..
.. =
φ(x) =
.
φ4 (x)
(89)
argmaxd df4 (x)
where x = (x, y, w, z). For any x ∈ [0, 1]4 , φ(x) is non-empty and closed, since the argmax solution
always exists and is one of {0}, {1}, [0, 1]. If in addition φ also has a closed graph then by Kakutani
Fixed Point Theorem there exists a solution to (27).
Consider any sequence (xn , an , bn , cn , dn ) → (x0 , a0 , b0 , c0 , d0 ) such that ∀ n ≥ 1,
an ∈ argmax af1 (xn ),
a∈[0,1]
cn ∈ argmax cf3 (xn ),
c∈[0,1]
bn ∈ argmax bf2 (xn ),
(90a)
b∈[0,1]
dn ∈ argmax df4 (xn ).
d∈[0,1]
(90b)
34
We need to verify that (90) also holds for the limit (x0 , a0 , b0 , c0 , d0 ). If x0 ∈
/ D then due to continuity, (90)
indeed holds at the limit. For x0 ∈ D, for any i ∈ S(x0 ) if fi (x0 ) = 0 then in the relation to be verified,
the requirement is either of a0 , b0 , c0 , d0 ∈ [0, 1], which is always true. For x0 ∈ D1 ∩ D2{ ∩ D3{ ∩ D4{ , if
f1 (x0 ) > 0 then for any sequence xn → x0 , for large n the points in the sequence are within B (x0 )
and thus f1 (xn ) > 0 for large n. This means that the relation from (90) holds at the limit (noting that
f2 , f3 , f4 are continuous at x0 in this case).
Similarly if f1 (x0 ) < 0 and for any x0 ∈ D1{ ∩ D2 ∩ D3{ ∩ D4{ .For x0 ∈ D1 ∩ D2 ∩ D3{ ∩ D4{ if f1 (x0 ) > 0
and f2 (x0 ) < 0 then there exists an > 0 such that ∀ x ∈ B (x0 ) we have f1 (x) > 0 and f2 (x) < 0.
From this it follows that the relation (90) holds at the limit. Similar argument works for any other sign
combination of f1 , f2 , f3 , f4 .
The two arguments above cover all cases.
R EFERENCES
[1] L. S. Shapley, “Stochastic games,” Proceedings of the national academy of sciences, vol. 39, no. 10,
pp. 1095–1100, 1953.
[2] T. Başar and G. Olsder, Dynamic Noncooperative Game Theory, 2nd Edition. Society for Industrial
and Applied Mathematics, 1998.
[3] J. Filar and K. Vrieze, Competitive Markov decision processes. Springer Science & Business Media,
2012.
[4] M. J. Osborne and A. Rubinstein, A Course in Game Theory, ser. MIT Press Books. The MIT
Press, 1994, vol. 1.
[5] D. Fudenberg and J. Tirole, Game Theory. Cambridge, MA: MIT Press, 1991.
[6] G. J. Mailath and L. Samuelson, Repeated games and reputations: long-run relationships. Oxford
university press, 2006.
[7] E. Maskin and J. Tirole, “Markov perfect equilibrium: I. observable actions,” Journal of Economic
Theory, vol. 100, no. 2, pp. 191–219, 2001.
[8] R. Ericson and A. Pakes, “Markov-perfect industry dynamics: A framework for empirical work,”
The Review of Economic Studies, vol. 62, no. 1, pp. 53–82, 1995.
[9] D. Bergemann and J. Välimäki, “Learning and strategic pricing,” Econometrica: Journal of the
Econometric Society, pp. 1125–1149, 1996.
[10] D. Acemoğlu and J. A. Robinson, “A theory of political transitions,” American Economic Review,
pp. 938–963, 2001.
[11] D. M. Kreps and J. Sobel, “Chapter 25 signalling,” ser. Handbook of Game Theory with Economic
Applications. Elsevier, 1994, vol. 2, pp. 849 – 867.
[12] U. Doraszelski and A. Pakes, “A framework for applied dynamic analysis in IO,” Handbook of
industrial organization, vol. 3, pp. 1887–1966, 2007.
[13] E. Altman, V. Kambley, and A. Silva, “Stochastic games with one step delay sharing information
pattern with application to power control,” in Game Theory for Networks, 2009. GameNets’ 09.
International Conference on. IEEE, 2009, pp. 124–129.
35
[14] A. Nayyar, A. Gupta, C. Langbort, and T. Başar, “Common information based Markov perfect
equilibria for stochastic games with asymmetric information: Finite games,” IEEE Trans. Automatic
Control, vol. 59, no. 3, pp. 555–570, March 2014.
[15] M. Spence, “Job market signaling,” The quarterly journal of Economics, pp. 355–374, 1973.
[16] S. J. Grossman, “The informational role of warranties and private disclosure about product quality,”
The Journal of Law & Economics, vol. 24, no. 3, pp. 461–483, 1981.
[17] C. Wilson, “A model of insurance markets with incomplete information,” Journal of Economic theory,
vol. 16, no. 2, pp. 167–207, 1977.
[18] M. Rothschild and J. Stiglitz, “Equilibrium in competitive insurance markets: An essay on the
economics of imperfect information,” in Foundations of Insurance Economics. Springer, 1976,
pp. 355–375.
[19] A. Zahavi, “Mate selection–a selection for a handicap,” Journal of theoretical Biology, vol. 53, no. 1,
pp. 205–214, 1975.
[20] A. V. Banerjee, “A simple model of herd behavior,” The Quarterly Journal of Economics, pp. 797–
817, 1992.
[21] S. Bikhchandani, D. Hirshleifer, and I. Welch, “A theory of fads, fashion, custom, and cultural
change as informational cascades,” Journal of Political Economy, vol. 100, no. 5, pp. pp. 992–1026,
1992. [Online]. Available: http://www.jstor.org/stable/2138632
[22] L. Smith and P. Sörensen, “Pathological outcomes of observational learning,” Econometrica, vol. 68,
no. 2, pp. 371–398, 2000. [Online]. Available: http://dx.doi.org/10.1111/1468-0262.00113
[23] N. R. Devanur, Y. Peres, and B. Sivan, “Perfect Bayesian equilibria in repeated sales,” in Proceedings
of the Twenty-Sixth Annual ACM-SIAM Symposium on Discrete Algorithms. SIAM, 2015, pp. 983–
1002.
[24] Y.-C. Ho, “Team decision theory and information structures,” Proceedings of the IEEE, vol. 68,
no. 6, pp. 644–654, 1980.
[25] A. Nayyar, A. Mahajan, and D. Teneketzis, “Optimal control strategies in delayed sharing information
structures,” IEEE Trans. Automatic Control, vol. 56, no. 7, pp. 1606–1620, July 2011.
[26] A. Gupta, A. Nayyar, C. Langbort, and T. Başar, “Common information based Markov perfect
equilibria for linear-gaussian games with asymmetric information,” SIAM Journal on Control and
Optimization, vol. 52, no. 5, pp. 3228–3260, 2014.
[27] Y. Ouyang, H. Tavafoghi, and D. Teneketzis, “Dynamic oligopoly games with private Markovian
dynamics,” in Proc. 54th IEEE Conf. Decision and Control (CDC), 2015.
[28] L. Li and J. Shamma, “Lp formulation of asymmetric zero-sum stochastic games,” in 53rd IEEE
Conference on Decision and Control, Dec 2014, pp. 1930–1935.
[29] H. L. Cole and N. Kocherlakota, “Dynamic games with hidden actions and hidden states,” Journal
of Economic Theory, vol. 98, no. 1, pp. 114–126, 2001.
[30] A. Nayyar, A. Mahajan, and D. Teneketzis, “Decentralized stochastic control with partial history
sharing: A common information approach,” IEEE Transactions on Automatic Control, vol. 58, no. 7,
pp. 1644–1658, July 2013.
[31] A. Mahajan, “Optimal decentralized control of coupled subsystems with control sharing,” Automatic
36
[32]
[33]
[34]
[35]
[36]
[37]
[38]
[39]
[40]
Control, IEEE Transactions on, vol. 58, no. 9, pp. 2377–2382, 2013.
A. Mahajan and D. Teneketzis, “On the design of globally optimal communication strategies for
real-time communcation systems with noisy feedback,” IEEE J. Select. Areas Commun., no. 4, pp.
580–595, May 2008.
A. Nayyar and D. Teneketzis, “On globally optimal real-time encoding and decoding strategies in
multi-terminal communication systems,” in Proc. IEEE Conf. on Decision and Control, Cancun,
Mexico, Dec. 2008, pp. 1620–1627.
D. Vasal and A. Anastasopoulos, “Stochastic control of relay channels with cooperative and strategic
users,” IEEE Transactions on Communications, vol. 62, no. 10, pp. 3434–3446, Oct 2014.
I.-K. Cho and D. M. Kreps, “Signaling games and stable equilibria,” The Quarterly Journal of
Economics, vol. 102, no. 2, pp. 179–221, 1987.
I.-K. Cho, “A refinement of sequential equilibria,” Econometrica, vol. 55, no. 6, pp. 1367–1389,
1987.
J. S. Banks and J. Sobel, “Equilibrium selection in signaling games,” Econometrica, vol. 55, no. 3,
pp. 647–661, 1987.
J. Nash, “Non-cooperative games,” Annals of mathematics, pp. 286–295, 1951.
D. Vasal and A. Anastasopoulos, “Signaling equilibria for dynamic LQG games with asymmetric
information,” in Proc. IEEE Conf. on Decision and Control, Dec. 2016, pp. 6901–6908.
——, “Decentralized Bayesian learning in dynamic games,” in Proc. Allerton Conf. Commun.,
Control, Comp., Sept. 2016.
| 3cs.SY
|
Phase Transitions, Optimal Errors and Optimality of
Message-Passing in Generalized Linear Models
Jean Barbier†,♦,⊗ , Florent Krzakala? , Nicolas Macris† , Léo Miolane∗,⊗ and Lenka Zdeborová∪
arXiv:1708.03395v2 [cs.IT] 3 Nov 2017
Abstract
We consider generalized linear models (GLMs) where an unknown n-dimensional signal vector is observed through the application of a random matrix and a non-linear (possibly probabilistic) componentwise
output function. We consider the models in the high-dimensional limit, where the observation consists
of m points, and m/n → α where α stays finite in the limit m, n → ∞. This situation is ubiquitous in
applications ranging from supervised machine learning to signal processing. A substantial amount of theoretical work analyzed the model-case when the observation matrix has i.i.d. elements and the components
of the ground-truth signal are taken independently from some known distribution. While statistical physics provided number of explicit conjectures for special cases of this model, results existing for non-linear
output functions were so far non-rigorous. At the same time GLMs with non-linear output functions are
used as a basic building block of powerful multilayer feedforward neural networks. Therefore rigorously
establishing the formulas conjectured for the mutual information is a key open problem that we solve in
this paper. We also provide an explicit asymptotic formula for the optimal generalization error, and confirm the prediction of phase transitions in GLMs. Analyzing the resulting formulas for several non-linear
output functions, including the rectified linear unit or modulus functions, we obtain quantitative descriptions of information-theoretic limitations of high-dimensional inference. Our proof technique relies on
a new version of the interpolation method with an adaptive interpolation path and is of independent interest. Furthermore we show that a polynomial-time algorithm referred to as generalized approximate
message-passing reaches the optimal generalization error for a large set of parameters.
Contents
1 Introduction
2
2 Setting and main results
2.1 Generalized linear estimation: Problem statement . . . . . . . . . . . .
2.2 Two scalar inference channels . . . . . . . . . . . . . . . . . . . . . . .
2.3 Replica-symmetric formula and mutual information . . . . . . . . . . .
2.4 The case of discrete channels . . . . . . . . . . . . . . . . . . . . . . . .
2.5 Optimality of the generalized approximate message-passing algorithm
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
4
4
6
7
8
10
† Laboratoire de Théorie des Communications, Faculté Informatique et Communications, Ecole Polytechnique Fédérale de Lausanne,
Suisse.
♦ Probability and Applications Group, School of Mathematical Sciences, Queen Mary University of London, United-Kingdom.
? Laboratoire de Physique Statistique, CNRS & Université Pierre et Marie Curie & École Normale Supérieure & PSL Université, Paris,
France.
∗ Département d’Informatique de l’ENS, École Normale Supérieure & CNRS & PSL Research University & Inria, Paris, France.
∪ Institut de Physique Théorique, CNRS & CEA & Université Paris-Saclay, Saclay, France.
⊗ Corresponding authors: [email protected], [email protected]
1
3 Optimal errors in supervised learning
3.1 Teacher-student scenario . . . . . . . . . . . . . . . . . . . .
3.2 Generalization error . . . . . . . . . . . . . . . . . . . . . . .
3.3 Computing the generalization error . . . . . . . . . . . . . .
3.4 Computing the optimal errors with (small) side information
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
11
11
12
12
13
4 Application to concrete situations
15
4.1 Generic observations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
4.2 Phase diagrams of perfect learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
4.3 Examples of optimal generalization error . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
5 Proof of the replica formula by the adaptive interpolation method
5.1 Interpolating estimation problem . . . . . . . . . . . . . . . . . . .
5.2 Free entropy variation along the interpolation path . . . . . . . . .
5.3 Overlap concentration . . . . . . . . . . . . . . . . . . . . . . . . .
5.4 Canceling the remainder . . . . . . . . . . . . . . . . . . . . . . . .
5.5 Lower and upper matching bounds . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
23
23
25
26
27
28
Appendix A Some technicalities
A.1 The Nishimori identity . . . . . . . . .
A.2 Some properties of the scalar channels
A.3 Proof of Theorem 2 . . . . . . . . . . .
A.4 Proof of Proposition 1 . . . . . . . . . .
A.5 Boundedness of an overlap fluctuation
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
29
29
29
32
34
35
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Appendix B Proofs of the limits of optimal errors
36
B.1 Computation of the denoising error: Proof of Corollary 3 . . . . . . . . . . . . . . . . . . . . . 36
B.2 Optimal generalization error with side information: Proof of Proposition 2 . . . . . . . . . . . 37
Appendix C Proof of Proposition 4
37
C.1 Proof of (114) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
C.2 Proof that An vanishes as n → ∞ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
Appendix D Approximation
41
D.1 Relaxing the hypotheses on P0 and Φ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
D.2 Relaxing the hypotheses on ϕ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
Appendix E A sup-inf formula
45
Appendix F Concentration of free entropy and overlaps
48
F.1 Concentration of the free entropy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
F.2 Concentration of the overlap . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
1 Introduction
Modern statistical analysis and signal processing require solving very high-dimensional estimation problems
with a very large number of parameters. This paradigm arises in problems as diverse as learning with neural
networks [54] and regression problems [22] or compressed sensing in signal processing [24, 30]. Developing
algorithms up to these tasks, and understanding their limitations, has become a major challenge in computer
science, machine learning and statistics.
2
In many examples, it has been empirically observed that both the inference and learning tasks appear
to have intrinsic limitations when the available data becomes too scarce or too noisy. In some cases these
apparent thresholds are related to information-theoretic phenomena: There is simply not enough information
in the dataset. This is the case, as famously discussed by Shannon in his seminal paper on communication
theory [74], for the task of reconstructing a noisy signal when the noise is beyond the so-called Shannon
capacity of the communication channel. In many situations, including communications, there also seem to
exist jumps in the computational hardness, beyond which the most sophisticated known algorithms take
exponential time to solve the task: In this case the problem has become too complicated to be solved efficiently.
A substantial amount of work suggests that one can understand and locate these fundamental barriers in a
range of statistical models by thinking of them as phase transitions in the sense of physics. In fact, over
the last three decades or so, a large body of interdisciplinary work in the statistical physics community has
been applied to such problems [32,34,35,59,60,65,73,81,84] considering random instances, generated by given
statistical models, and then locating these phase transitions. Such models are in fact being widely used in fields
as diverse as (without any pretension of exhaustivity) statistical learning [3,17,28,33,38], compressed sensing
and signal processing [9, 11, 21, 29, 31, 55, 56, 62, 70, 84], communication theory [6, 12, 15, 41, 77], community
detection in networks [2, 23, 27, 51], combinatorial optimization problems [7, 50, 59] or to model the behavior
of neurons or neural nets [4, 5, 73, 81].
Many of these works, especially in the context of compressed sensing and machine learning, relied however on non-rigorous methods, using instead powerful heuristics like the replica and cavity approaches [60].
In the present contribution we leverage on these pioneering works and provide instead rigorous asymptotic
predictions in the case of generalized linear estimation models [64]. This includes a range of statistical models
of interest in many scientific fields such as random linear estimation in statistics, the teacher-student singlelayer perceptron problem, probit classification or quantized compressed sensing. For all these estimation and
learning problems our results completely vindicate the physics results —closing in some cases conjectures
open since almost three decades [16, 36, 42]— and considerably extend them. We also establish the value of
the generalization error after an optimal learning, which gives a bound on how accurately any algorithm is
able to predict outcome values for previously unseen data.
We compare the optimal results with the algorithmic ones provided by message-passing algorithms [31,70,
84] and observe that, while optimal performances are often thought to be intractable, in the present setting
they can actually be obtained using a polynomial-time scheme for a large set of parameters. There exists,
however, an interesting region of parameters where all so far known algorithms fail to provide a satisfactory
answer to the estimation and learning problems, while it is nevertheless information-theoretically possible
to do so. In this case there is a significant gap between what currently known polynomial algorithms can do
and what should be expected from the information-theoretic point of view.
Finally, our proof technique has an interest on its own. It exploits a powerful new technique called the
adaptive1 interpolation method [13] that is applicable to many other open problems in statistical estimation.
We informally summarise our main contributions here:
• We consider generalized linear estimation models where, given an unknown signal vector X∗ , one is
given the measurement vector Y = ϕ √1n ΦX∗ , with Φ a known random matrix with i.i.d. entries of
zero mean, and where ϕ (that can be probabilistic) acts componentwise.
• Our first main result is the rigorous determination of the expression of the (conditional) entropy H(Y|Φ)
of the observation variable, a quantity often called “the averaged free energy” in the statistical physics
literature, in the “high-dimensional regime” where the signal dimensionality n as well as the number of
In [13] the method involves a discrete time stochastic process and is called “stochastic interpolation method.” For dense graphical problems one can easily take a continuum limit in which the stochastic process does not appear explicitly and the method is
better called “adaptive interpolation method.”
1
3
observations m are jointly growing such that m/n → α > 0. As we shall see, many statistical quantities of interest, such as the mutual information between the observation and the unknown signal or
the Bayes-optimal generalization error, can be computed from this expression. We provide the proof of
this expression —which is our main technical contribution— in the last section. Only the simple linear
case was known rigorously so far [10, 14, 71], whereas including the non-linearity is an absolutely key
element towards building more powerful multilayer neural networks.
In fact, our results cover a number of cases discussed in the literature, often anticipated by statistical
physics techniques, and allow to rigorously prove many predictions obtained by the heuristic replica
method. The expression for the entropy was first famously conjectured in 1989 in [36] for the particular
case of ϕ(x) = sgn(x), a work that is at the basis of many significant developments. The generic
formula was also recently conjectured on the same heuristic basis [84]. Our proof yields a spectacular
confirmation of the cavity and replica methods.
• We demonstrate (with the technical assumption of the presence of small side-information, which is
equivalent to the presence of a small vanishing magnetic field in physics) the formula for the minimum
mean-square-error (MMSE) for the reconstruction of the unknown signal in the generalized linear estimation model.
• We also compute (under the same technivcal assumption) the Bayes-optimal generalization error in the
context of a supervised learning task of a rule used by the model (the so-called teacher-student setting). This estimator is optimal in the sense that it minimizes the label (the value of the output of ϕ)
mean-square-error among all possible student estimators. Again, this formula was anticipated with the
heuristic cavity and replica methods, and was proposed, for the restricted case ϕ(x) = sgn(x) (the
so-called perceptron problem), in pioneering works in the statistical physics community [42, 66, 73].
• While these results are information-theoretic, we also consider the performance of a well-known algorithm to solve random instances of generalized linear estimation problems, called generalized approximate message-passing (GAMP [31, 70, 84]). This algorithm, which also originated from statistical
physics [5, 46, 58, 78], is expected to be particularly powerful on these random instances, as proven for
instance for compressed sensing [14, 31, 52, 79]. Indeed we show that, for a large set of problems and
a large region of parameters, GAMP yields an optimal generalization error. There exists, however, a
region of parameters where GAMP does not reach the optimal results. It remains a challenging open
problem whether a polynomial algorithm that outperforms GAMP in this region exists.
• Finally, we study in depth the situation for many given choices of the function ϕ, and identify sharp
phase transitions and novel phase diagrams. By locating these transitions we clarify the difficulties
and challenges in solving complex non-linear high-dimensional statistical problems in many modelsituations, and characterize these problems in terms of optimal information-theoretical reconstruction.
2 Setting and main results
2.1
Generalized linear estimation: Problem statement
We now give a generic description of the observation model. Note that we describe here an estimation (or
inference) problem. In the title of the paper, we refer to generalized linear models. This is a very general
setting that will allow us to consider supervised learning problems (see Sec. 3). Precise hypotheses are given
below in Sec. 2.3.
4
Let n, m ∈ N∗ . Let P0 be a probability distribution over R and let (Xi∗ )ni=1 i.i.d.
∼ P0 be the components
i.i.d.
∗
∗
of a signal vector X (this is also denoted X ∼ P0 ). We fix a function ϕ : R × RkA → R and consider
i.i.d.
kA
(Aµ )m
µ=1 ∼ PA , where PA is a probability distribution over R , kA ∈ N. We acquire m measurements
through
√
1
Yµ = ϕ √ [ΦX∗ ]µ , Aµ + ∆Zµ ,
n
(1)
1 ≤ µ ≤ m,
i.i.d.
where (Zµ )m
µ=1 ∼ N (0, 1) is an additive Gaussian noise, ∆ > 0, and Φ is a m × n measurement matrix with
i.i.d. Φµi ∼ N (0, 1) entries (this hypothesis will actually be relaxed, see Sec. 2.3). The estimation problem
is to recover X∗ from the knowledge of Y = (Yµ )m
of the random
µ=1 , ϕ, Φ, ∆, P0 and PA (the realization
Pn
∗
stream A itself, if present in the model, is unknown). We use the notation [ΦX ]µ = i=1 Φµi Xi∗ . When
ϕ(x, A) = ϕ(x) = x we have a random linear estimation problem, whereas if, say, ϕ(x) = sgn(x) we
have a noisy single layer perceptron. Sec. 3 discusses various examples related to non-linear estimation and
supervised learning.
It is also fruitful to think of the measurements as the outputs of a “channel”,
1
Yµ ∼ Pout · √ [ΦX∗ ]µ
n
(2)
where the transition density, with respect to (w.r.t.) Lebesgue’s measure, is
Z
1
1
2
Pout y x = √
dPA (a)e− 2∆ (y−ϕ(x,a)) .
2π∆
(3)
Our hypotheses ensure that this is a well defined density. In fact (3) is sometimes called a “random function
representation” of a transition kernel. Our analysis uses both representations (1) and (2).
Throughout this paper we often adopt the language of statistical mechanics. In particular the random
variables Y (and also Φ, X∗ , A, Z) are called quenched variables because once the measurements are acquired
they have a “fixed realization.” An expectation taken w.r.t. all quenched random variables appearing in an
expression will simply be denoted by E without subscript. Subscripts are only used when the expectation
carries over a subset of random variables appearing in an expression or when some confusion could arise.
A fundamental role is played by the joint posterior distribution of (the signal) x and of (the random stream)
a given the quenched measurements Y (recall that X∗ , A and Z are unknown). The prior over the signal is
Q
Q
denoted dP0 (x) = ni=1 dP0 (xi ), and similarly dPA (a) = m
µ=1 dPA (aµ ). According to the Bayes formula
this joint posterior is given by
m
dP (x, a|Y, Φ) =
Y
1
1
− 1
√
dP0 (x)dPA (a)
e 2∆
Z(Y, Φ)
2π∆
µ=1
2
Yµ −ϕ( √1n [Φx]µ ,aµ )
,
(4)
where the partition function (the normalization factor) is defined as
Z
Z(Y, Φ) :=
m
Y
1
− 1
√
dP0 (x)dPA (a)
e 2∆
2π∆
µ=1
5
2
Yµ −ϕ( √1n [Φx]µ ,aµ )
.
(5)
Marginalizing (4) w.r.t. a yields the posterior of x, namely
1
dP0 (x)e−H(x;Y,Φ) ,
Z(Y, Φ)
Z
Z(Y, Φ) = dP0 (x)e−H(x;Y,Φ) ,
dP (x|Y, Φ) =
(6)
(7)
where the Hamiltonian is defined as
H(x; Y, Φ) := −
m
X
1
ln Pout Yµ √ [Φx]µ .
n
µ=1
(8)
From the point of view of statistical mechanics (6) is a Gibbs distribution and the integration over dP0 (x) in
the partition function is best thought as a “sum over annealed or fluctuating degrees of freedom” (note that in
the representation (5), (aµ )m
µ=1 also play the role of annealed variables). Let us introduce a standard statistical
mechanics notation for the expectation w.r.t. the joint posterior (4), the so called Gibbs bracket h−i defined as
Z
hg(x, a)i := dP (x, a|Y, Φ)g(x, a)
(9)
for any continuous bounded function g. The main quantity of interest here is the associated averaged free
entropy (or minus the averaged free energy)
fn :=
1
E ln Z(Y, Φ) .
n
(10)
It is perhaps useful to stress that Z(Y, Φ) is nothing else than the density of Y conditioned on Φ so we have
the explicit representation (used later on)
Z
1
fn = EΦ dYZ(Y, Φ) ln Z(Y, Φ)
n
Z
Z
1
∗
(11)
= EΦ dYdP0 (X∗ )e−H(X ;Y,Φ) ln dP0 (x) e−H(x;Y,Φ) ,
n
Q
where dY = m
µ=1 dYµ . Thus fn is minus the conditional entropy −H(Y|Φ)/n of the measurements.
One of the main contributions of this paper is the derivation, thanks to the adaptive interpolation method,
of the thermodynamic limit limn→∞ fn in the “high-dimensional regime”, namely when n, m → ∞ while
m/n → α > 0 (α is sometimes referred to as the “measurement rate” in compressed sensing terminology).
2.2
Two scalar inference channels
An important role in our proof of the asymptotic expression of the free entropy is played by simple scalar
inference channels. As we will see, the free entropy is expressed in terms of the free entropy of these channels.
This “decoupling property” results from the mean-field approach in statistical physics, used through in the
replica method to perform a formal calculation of the free entropy of the model [59,60]. Let us now introduce
these two scalar denoising models.
The first one is an additive Gaussian channel. Let r ≥ 0, which plays the role of a signal-to-noise ratio
(snr). Suppose that X0 ∼ P0 and that we observe
Y0 =
√
r X 0 + Z0 ,
(12)
where Z0 ∼ N (0, 1) independently of X0 . Consider the inference problem consisting of retrieving X0 from
6
the observations Y0 . The associated posterior distribution is
√
2
dP0 (x)e r Y0 x−rx /2
√
.
dP (x|Y0 ) = R
dP0 (x)e r Y0 x−rx2 /2
(13)
In this expression all the x-independent terms have been simplified between the numerator and the normalization. The free entropy associated with this channel is just the expectation of the logarithm of the
normalization factor
Z
√
2
ψP0 (r) := E ln dP0 (x)e r Y0 x−rx /2 .
(14)
The basic properties of ψP0 are presented in Appendix A.2.1.
The second scalar channel that appears naturally in the problem is linked to the transition kernel Pout
through the following inference model. Suppose that V, W ∗ i.i.d.
∼ N (0, 1) where V is known while the inference
∗
problem is to recover the unknown W from the following observation
√
√
Ye0 ∼ Pout · | q V + ρ − q W ∗ ,
(15)
where ρ > 0, q ∈ [0, ρ] and Pout is the transition density defined in (3). The free entropy for this model, again
related to the normalization factor of the posterior dP (w|Ye0 , V ), is
Z
√
√
(16)
ΨPout (q; ρ) = ΨPout (q) := E ln DwPout Ye0 | q V + ρ − q w ,
2
where Dw = dw(2π)−1/2 e−w /2 is the standard Gaussian measure. We prove in Appendix A.2.2 that this
function is convex, differentiable and non-decreasing w.r.t. its first argument.
√
√
√
Recall that we consider here Pout to be of the form (3), so that Ye0 = ϕ( q V + ρ − q W ∗ , A) + ∆Z,
where A ∼ PA and Z ∼ N (0, 1), independently of any other random variable. In the sequel, the joint
posterior distribution of (W ∗ , A) given (V, Ye0 ) will be particularly useful, so we define the Gibbs bracket for
the scalar channel (15) by:
1
e− 2∆
Z
hg(w, a)isc =
DwdPA (a)
R
2
√
√
Ye0 −ϕ( q V + ρ−q w,a)
2
√
√
1
− 2∆
Ye0 −ϕ( q V + ρ−q w,a)
g(w, a) ,
(17)
DwdPA (a)e
for any continuous bounded function g.
2.3
Replica-symmetric formula and mutual information
Let us now introduce our first main result, a single-letter replica-symmetric formula for the asymptotic free
entropy of model (1), (2). The result holds under the following rather general hypotheses. We will consider
two cases: The case where the entries of Φ are i.i.d. standard Gaussian random variables and the case of
general distributions. This leads to two sets of hypotheses.
(h1) The prior distribution P0 admits a finite second moment and has at least two points in its support.
(h2) There exists γ > 0 such that the sequence (E[|ϕ( √1n [ΦX∗ ]1 , A1 )|2+γ ])n≥1 is bounded.
(h3) (Φµi ) i.i.d.
∼ N (0, 1).
In the non-Gaussian case, the hypothesis (h3) has to be replaced by the following hypotheses:
(h3.a) The prior distribution P0 admits a finite third moment.
7
(h3.b) The random variables (Φµi ) are independent with zero mean, unit variance and finite third moment
that is bounded with n.
(h3.c) For almost-all values of a ∈ RkA (w.r.t. PA ), the function x 7→ ϕ(x, a) is continuous almost everywhere.
Let us define the replica-symmetric potential (or just potential). Call ρ := E[(X ∗ )2 ] where X ∗ ∼ P0 . Then
the potential is
fRS (q, r; ρ) = fRS (q, r) := ψP0 (r) + αΨPout (q; ρ) −
rq
.
2
(18)
We define also fRS (ρ, +∞) = limr→∞ fRS (ρ, r). From now on denote ψP0 0 (r) and Ψ0Pout (q) = Ψ0Pout (q; ρ)
the derivatives of ψP0 (r) and ΨPout (q; ρ) w.r.t. their first argument. We need also to define the set of the
critical points of fRS :
(
)
q = 2ψP0 0 (r)
Γ := (q, r) ∈ [0, ρ] × (R+ ∪ {+∞})
,
(19)
r = 2αΨ0Pout (q; ρ)
where, with a slight abuse of notation, we define ψP0 0 (+∞) = limr→∞ ψP0 0 (r) and Ψ0Pout (ρ) = limq→ρ Ψ0Pout (q).
These limits are well defined by convexity of ψP0 and ΨPout . The elements of Γ are called “fixed points of the
state evolution”. The main theorem of this paper is
Theorem 1 (Replica-symmetric formula). Suppose that hypotheses (h1) and (h2) hold. Suppose that either (h3)
or (h3.a)-(h3.b)-(h3.c) hold. Then, for the generalized linear estimation model (1), (2) the thermodynamic limit of
the free entropy (10) verifies
f∞ := lim fn = sup inf fRS (q, r) = sup fRS (q, r) .
n→∞
q∈[0,ρ] r≥0
(20)
(q,r)∈Γ
Moreover, the “sup inf” and the supremum over Γ in (20) are achieved over the same couples.
An immediate corollary of Theorem 1 is the limiting expression of the mutual information between the
observations and the hidden variables.
Corollary 1 (Single-letter formula for the mutual information). The thermodynamic limit of the mutual information for model (1), (2) between the observations and the hidden variables verifies
in :=
1
α
I(X∗ , A; Y |Φ) −−−→ i∞ := −f∞ − (1 + ln(2π∆)) .
n→∞
n
2
(21)
Proof. This follows from a simple calculation:
1
P (Y, X∗ , A|Φ)
1
1
I(X∗ , A; Y|Φ) = E ln
= − E ln P (Y|Φ) + E ln P (Y|X∗ , A, Φ)
∗
n
P (Y|Φ)P (X , A|Φ)
n
n
m
X
1
m
= −fn −
(Yµ − ϕ([ΦX∗ ]µ , Aµ ))2 −
E
ln(2π∆)
2n∆
2n
µ=1
m
m
= −fn −
−
ln(2π∆) .
2n 2n
2.4
(22)
The case of discrete channels
A natural question to ask now is: Can we extend the previous results without the Gaussian noise in (1)? This
question seems to be non-trivial for various reasons. First, without the “regularizing” Gaussian noise, the
8
transition kernel Pout may not have a density w.r.t. Lebesgue’s measure: How should we define the “entropylike” quantities that we study? Second, even in the situation where the transition kernel admits a density,
it seems not totally obvious to justify the ∆ → 0 limit in Theorem 1. This is mostly due to the fact that
Shannon’s entropy is not continuous w.r.t. the weak convergence, see for instance [69]. The results of [69]
provide sufficient conditions that the densities should verify in order to have convergence of the entropies.
So it should be possible to take the ∆ → 0 limit in Theorem 1 at the cost of additional technical assumptions,
but we prefer for simplicity to only consider discrete transition kernels.
Therefore, in this section we assume that ϕ takes its values in a finite set. We suppose also that hypotheses (h1) and (h2) hold, and that either (h3) or (h3.a)-(h3.b)-(h3.c) hold.
We define
ξn :=
1
E ln
n
Z
dP0 (x)dPA (a)
m
1
Y
1
δ ϕ √ [ΦX∗ ]µ , Aµ − ϕ √ [Φx]µ , aµ ,
n
n
µ=1
(23)
and for q ∈ [0, ρ],
Z
Ψϕ (q) = E ln
√
√
√
√
dPA (a)Dw δ ϕ( q V + ρ − q W, A) − ϕ( q V + ρ − q w, a) ,
(24)
where V, W i.i.d.
∼ N (0, 1) and A ∼ PA are independent random variables.
Theorem 2 (Replica-symmetric formula for discrete channels). Suppose that hypotheses (h1) and (h2) hold.
Suppose that either (h3) or (h3.a)-(h3.b)-(h3.c) hold. Then, for the generalized linear estimation model (1) with
∆ = 0 and an output function ϕ taking values in a finite set, the thermodynamic limit of the free entropy (23)
verifies
n
rq o
.
(25)
ξn −−−→ sup inf ψP0 (r) + αΨϕ (q) −
n→∞ q∈[0,ρ] r≥0
2
The proof of Theorem 2 is presented in Appendix A.3.
As an application of Theorem 2 we can compute the free energy of the “planted perceptron” on the hypercube and the sphere. This perceptron model has already been studied in physics [36] and more recently
in statistics, where it is known as “one-bit compressed sensing” [21, 83]. The limit of the free entropy follows from an application of Theorem 2 with ϕ(x) = sgn(x) and P0 = 21 δ−1 + 12 δ1 (for the hypercube) or
P0 = N (0, 1) (for the sphere). For µ ∈ {1, . . . , m}, we define
n
o
Sµ = x ∈ Rn sgn(x · Φµ ) = sgn(X∗ · Φµ ) .
(26)
We will use the notation N (x) = P(Z ≤ x) for Z ∼ N (0, 1). Let Sn be the unit sphere in Rn and µn the
uniform measure on Sn .
Corollary 2 (Free entropy of the planted perceptron). Let Z, V ∼ N (0, 1).
m
\
1
E ln #
Sµ ∩ {−1, 1}n
n
µ=1
n
h √q V
√q V i r(q + 1) o
√
√
ln N √
−
−−−→ ln(2) + sup inf E ln cosh( rZ + r) + 2αE N
, (27)
n→∞
2
1−q
1−q
q∈[0,1) r≥0
m
\
n1
h √q V
√q V i q o
1
E ln µn
Sµ ∩ Sn −−−→ sup
ln(1 − q) + 2αE N √
ln N √
+
.
(28)
n→∞ q∈[0,1) 2
n
2
1
−
q
1
−
q
µ=1
9
2.5
Optimality of the generalized approximate message-passing algorithm
While the main results presented until now are information-theoretic, our next one concerns the performance of a popular algorithm to solve random instances of generalized linear problems, called generalized
approximate message-passing (GAMP). We shall not re-derive its properties here, and instead refer to the original papers for details. This approach has a long history, especially in statistical physics [5, 46, 58, 78], error
correcting codes [72], and graphical models [80]. For a modern derivation in the context of linear models,
see [31, 52, 79]. The case of generalized linear models was discussed by Rangan in [70], and has been used for
classification purpose in [85].
We first need to define two so-called threshold functions that are associated to the two scalar channels
(12) and (15). The first one is the posterior mean of the signal in channel (12) with signal-to-noise ratio r:
gP0 (y, r) = E[X0 |Y0 = y] .
(29)
The second one is the posterior mean of W ∗ in channel (15) with “noise level” η = ρ − q:
√
gPout (e
y , v, η) = E[W ∗ |Ye0 = ye, q V = v] .
(30)
These functions act componentwise when applied to vectors.
Given initial estimates (b
x0 , v0 ) for the means and variances of the elements of the signal vector X∗ ,
GAMP takes as input the observation vector Y and then iterates the following equations with initialization
gµ0 = 0 for all µ = 1, . . . , m (we denote by u the average over all the components of the vector u and Φ| is
the transpose of the matrix Φ): From t = 1 until convergence,
V t = vt
t = Φb
t−1 /√n − V t gt−1
ω
x
t
t
t
∀ µ = 1, . . . m
gµ = gPout (Yµ , ωµ , V )
t
2
t
t
(31)
λ
= α gPout (Y, ω , V )
√
t
t−1
t
|
t
b
R = x
+λ Φ g / n
t
t t
x
b
=
g
(R
∀ i = 1, . . . n
P0
i
i, λ )
vt = (λt )−1 ∂ g (R, λt )|
∀ i = 1, . . . n
R P0
R=Rit
i
One of the strongest asset of GAMP is that its performance can be tracked rigorously in the limit n, m → ∞
via a procedure known as state evolution (SE), see [18, 19] for the linear case, and [45, 70] for the generalized
one. In our notations, state evolution tracks the asymptotic value of the overlap between the true hidden
bt defined as q t = limn→∞ X∗ · x
bt /n (that is related to the asymptotic
value X∗ and its estimate by GAMP x
bt by E t = ρ − q t , where recall that ρ = E[(X ∗ )2 ] with
mean-square error E t between X∗ and its estimate x
∗
X ∼ P0 ) via:
(
q t+1
rt
= 2ψP0 0 (rt ) ,
= 2αΨ0Pout (q t ; ρ) .
(32)
From Theorem 1 we realize that the fixed points of these equations correspond to the critical points of the
asymptotic conditional entropy in (20). In fact, in the replica heuristic, the optimizer q ∗ of the potential is
conjectured to give the optimal value of the overlap, a fact that was proven for the linear channel [10, 14]. If
the minimizer of (20) is unique, or is attractive within the GAMP dynamics, then q t converges to q ∗ when
t → +∞.
Perhaps more surprisingly, one can use GAMP in the teacher-student scenario (that we precisely describe
√
in Sec. 3.1) in order to provide an estimation of a new output Cnew = ϕ(Φnew ·X∗ / n, Anew ), where Φnew is
10
bt is the GAMP estimate of the expectation
a new row of the matrix and Anew ∼ PA a new random source. As x
of X∗ , with estimated variance vt , the natural heuristic is to consider for the posterior probability distribution
√
√
bt / n and variance V t = E t = ρ − q t
of the random variable Φnew · X∗ / n a Gaussian with mean Φnew · x
(the fact that the variance and MSE are equal follows from the Nishimori identity of Proposition 9 applied
to GAMP, see e.g. [84]). This allows to estimate the posterior mean of the output, which leads the GAMP
prediction (recall the Pout definition (3)):
b GAMP,t =
C
Z
1
e
dy dw y Pout (y|w) p
2π(ρ − q t )
1
− 2(ρ−q
t)
√1 Φnew ·b
xt −w
n
2
.
(33)
We show in Sec. 3.3.2 that this leads to a GAMP generalization error given by
GAMP,t
b GAMP,t 2
:= lim E Cnew − C
Egen
n→∞
p
√
p
2
= EV,A ϕ( ρ V, A)2 − EV EW,A ϕ( q t V + ρ − q t W, A) ,
(34)
where V, W i.i.d.
∼ N (0, 1) and A ∼ PA .
We will see in Sec. 3.2 that this formula matches the one for the Bayes-optimal generalization error, see
(40), up to the fact that instead of q ∗ (a maximizer obtained from the replica formula (20)) appearing in the
optimal error formula, here it is q t which appears. Thus clearly, when q t converges to q ∗ (we shall see that this
is the case in many situations in the examples of Sec. 4) this yields a very interesting and non trivial result:
GAMP achieves the Bayes-optimal generalization error in a plethora of models (a task again often believed to
be intractable) and this for large sets of parameters.
3 Optimal errors in supervised learning
3.1
Teacher-student scenario
Consider the following teacher-student scenario (also called planted model). We voluntarily employ terms
coming from machine learning, instead of the signal processing terminology used in the previous sections.
First the teacher randomly generates a classifier X∗ ∈ Rn (the signal in the estimation problem) with
i.i.d.
n
X∗ i.i.d.
∼ P0 , an ensemble of m = αn patterns (row-vectors) Φµ ∈ R for µ = 1, . . . , m such that Φµ ∼
i.i.d.
N (0, In ), and a stream A = (Aµ )m
µ=1 ∼ PA . The teacher chooses a model ϕ(x, A). For deterministic
models A is simply absent. Finally the teacher associates to each pattern a label Cµ ∈ R selected by the
√
classifier, namely Cµ = ϕ(Φµ · X∗ / n, Aµ ) for µ = 1, . . . , m. Stacking the rows (Φµ )m
µ=1 in a m × n
m
matrix Φ and denoting the vector of labels C = (Cµ )µ=1 , then the labels, patterns, source of randomness and
classifier verify
1
C = ϕ √ ΦX∗ , A ,
n
(35)
where it is understood that ϕ acts in componentwise fashion on vectors. The student is given the distributions
P0 , PA and the function ϕ and his task is to learn the classifier X∗ from the pattern-label pairs (Φµ , Cµ )m
µ=1 .
∗
How does one use this estimation to perform prediction? The teacher, still using the same X , will create
a new sample (and thus a new line of the matrix, or pattern, Φnew and a new random Anew ). The teacher
√
will first compute the new value of the projection of X∗ , that we denote znew = Φnew · X∗ / n, and then
generate the new random label Cnew = ϕ(znew , Anew ). He is now ready to evaluate the student generalization
performance.
11
3.2
Generalization error
An important quantity is the generalization error (or prediction error) which measures the performance of the
b new , C, Φ) the estimator used by the student (which is thus a measurable function
student. If we denote C(Φ
of the observations), the generalization error is defined as
h
2 i
b = E Cnew − C(Φ
b new , C, Φ)
Egen (C)
.
(36)
The optimal generalization error is then defined as the minimum of Egen over all estimators:
h
D
E2 i
opt
b = E Cnew − ϕ √1 Φnew · x, a
:= min Egen (C)
Egen
.
n
b
C
(37)
where x is sampled according to the posterior distribution P (x = X∗ |C, Φ) and a ∼ PA . h−i denotes the
joint expectation w.r.t. (x, a).
3.3
Computing the generalization error
In order to express the limit of the generalization error, we introduce the following function:
√
2
√
√
E : q 7→ E ϕ( ρ V, A)2 − E EW,A ϕ( q V + ρ − q W, A)
(38)
where V, W i.i.d.
∼ N (0, 1), A ∼ PA are independent random variables and EW,A denotes the expectation w.r.t.
W and A only. We recall ρ = E[(X ∗ )2 ].
3.3.1
Optimal generalization error
opt
The computation of the limit of Egen
is not trivial and requires a good knowledge of the posterior distribution.
In Sec. 3.4.1 we will show that under a small perturbation of the model (given some side information), one can
compute it rigorously. Here we only show that if the overlap between the planted solution X∗ and a sample (a
replica) x from the posterior distribution converges to some number q, then the optimal generalization error
converges to E(q).
Proposition 1. Assume that the overlap n1 x · X∗ converges in probability to some number 0 ≤ q < ρ. Then
opt
Egen
−−−→ E(q) .
n→∞
(39)
The proof of Proposition 1 is presented in Appendix A.4. This result by itself does not tell us how to
compute q. But in view of Proposition 2 we expect that q = q ∗ the maximizer of (20).
Conjecture 1 (Optimal generalization error). For almost all ∆ > 0 we have
opt
Egen
−−−→ E(q ∗ ) ,
n→∞
(40)
where q ∗ is a maximizer of (20).
The optimal generalization error is independent of the random function representation as long as both ϕ
√
R
R
and PA are known. EX dY Y k Pout (Y |X) = EX,a dY Y k exp{−(Y − ϕ(X, a))2 /(2∆)}/ 2π∆ implies
that the error (40) can be re-expressed equivalently as a function of the two first moments of the transition
probability Pout , namely (again, q ∗ is a maximizer of (20))
Z
h hZ
i2 i
p
p
√
opt
−−−→ EV
dY Y 2 Pout (Y | ρ V ) − EV Ew
dY Y Pout (Y | q ∗ V + ρ − q ∗ w)
− ∆ . (41)
Egen
n→∞
12
3.3.2
Generalization error of the generalized approximate message-passing algorithm
We now turn our attention to the generalization performance of the GAMP algorithm of Sec. 2.5. The following claim is from [70], where it is stated as a claim because some steps of the proof are missing. The paper [45]
affirms in its abstract to prove the claim of [70], but without further details. For these reasons, we believe that
the claim holds, however we prefer to state it here as a claim (instead of a theorem).
Claim 1 ( [70]). We have almost surely for all t ∈ N,
lim
n→∞
1 t
1 t 2
b · X∗ = lim kb
x
x k = qt .
n→∞
n
n
(42)
Provided that Claim 1 holds, we can deduce the generalization error of GAMP.
b GAMP,t is given by (33). Then
Claim 2 (GAMP generalization error). Recall that C
GAMP,t
b GAMP,t 2 = E(q t ) .
:= lim E Cnew − C
Egen
n→∞
(43)
Claim 2 is proved following the same proof strategy as for Proposition 1, see Appendix A.4.
3.3.3
Optimal denoising error
Another interesting error measure to study is the following “denoising error”. Assume that the observations
are noisy, as in (1). The goal here is to denoise the observations Yµ and estimate the signal ϕ √1n [ΦX∗ ]µ , Aµ .
The minimum denoising error (in L2 sense) is actually a simple corollary from the replica-symmetric
formula of Theorem 1 and follows from a so-called “I-MMSE relation”, see [40].
Corollary 3 (Limit of the optimal denoising error). For almost every ∆ > 0, for any optimal couple (q ∗ , r∗ )
of (20) we have
D 1
E 2 i
1 h 1
E ϕ √ ΦX∗ , A − ϕ √ Φx, a
n→∞ m
n
n
p
p
√
= E ϕ( ρ V, A)2 − E ϕ( q ∗ V + ρ − q ∗ w, a)i2sc ,
lim
(44)
where the Gibbs brackets are defined by (9) and (17) and V ∼ N (0, 1), (w, a) ∼ h−isc .
The proof of Corollary 3 is presented in Appendix B.1.
3.4
3.4.1
Computing the optimal errors with (small) side information
Optimal generalization error with side information
As presented in Sec. 3.3.1, the computation of the optimal generalization error requires a good knowledge of
the posterior distribution, typically that the overlap concentrates around a deterministic value. Proving this
kind of result seems to be highly non-trivial. An easier strategy would be to proceed as in the previous section
by using “I-MMSE relations”: Differentiating the free entropy fn w.r.t. the model parameters allows to access
interesting quantities.
However, in order to obtain an expression for the generalization error, we need to assume that the new
“test label” Cnew is also observed by the student in the scenario of Sec. 3.1 but with a large amount of noise, so
that we can differentiate fn w.r.t. this noise. For this reason we introduce the following “train-test” observation
model.
The set of rows of Φ and labels are divided into two sets by the teacher: The training set S tr of size βm, β ∈
[0, 1], that will be used by the student in order to learn the classifier, and the test set S te of size (1 − β)m that
13
will be used by the teacher in order to evaluate the performance of the student. For the training
the
√ set both βm
tr
tr
patterns and associated noisy labels are given to the student, namely S = {(Yµ = Cµ + Zµ ∆ ; Φµ )}µ=1
where Zµ ∼ N (0, 1) (recall the labels are given by (35)). Here ∆tr is strictly positive (but typically small)
and known by the student who can thus optimally tune its “confidence” in the training labels during the
training/learning stage. For the test set, the previously unseen patterns
to classify are given to the student
√
te
te
te
but the labels are (almost) not, namely S = {(Yµ = Cµ + Zµ ∆ ; Φµ )}m
µ=βm+1 where ∆ is typically
te
very large. We are particularly interested in the case ∆ → ∞ when the student has no information about
the labels of the test set but for the aforementioned technical reasons we will have to work with very large,
but finite, ∆te .
Define ∆µ = ∆tr if µ ≤ βm, ∆µ = ∆te else. Then the inference of the classifier X∗ from the noisy
p
labels Yµ = Cµ + Zµ ∆µ is a slight extension of model (1): This inference problem is nothing else than
te = (Φ )m
a supervised learning of the classifier by the student. Define Cte = (Cµ )m
µ µ=βm+1
µ=βm+1 and Φ
as the m(1 − β)-dimensional vector of labels and the m(1 − β) × n matrix, respectively, both restricted to
te
the test set and similarly for Ate = (Aµ )m
µ=βm+1 . Then the optimal generalization error at finite ∆ in this
“train-test” observation model is
h
D 1
E 2 i
1
side
:=
Egen
E Cte − ϕ √ Φte x, ate
(1 − β)m
n
D 1
E 2 i
h 1
1
.
(45)
E ϕ √ Φte X∗ , Ate − ϕ √ Φte x, ate
=
(1 − β)m
n
n
Here the Gibbs bracket h−i is associated to the joint posterior (4) but with ∆ replaced by ∆µ in order to
take into account that the noise varies in the training and test sets. Moreover ate is the restriction of a to its
components belonging to the test set, namely ate = (aµ )m
µ=βm+1 .
By a straightforward extension of the interpolation method presented in Sec. 5 one can generalize Theorem 1 to take into account that the noise variance ∆µ differs in the training and test sets. This leads to an
asymptotic free entropy rigorously given by
n
rq o
ts
ts
tr (q; ρ) + α(1 − β)ΨP te (q; ρ) −
f∞
= sup inf fRS
(q, r) = sup inf ψP0 (r) + αβΨPout
.
(46)
out
2
q∈[0,ρ] r≥0
q∈[0,ρ] r≥0
tr and ∆ = ∆te respectively.
tr and ΨP te are obtained from (16) using Pout given by (3) with ∆ = ∆
Here ΨPout
out
ts (inside
Note that in the limit ∆te → ∞ (which means that the student has no access to the test set labels), fRS
the brackets in (46)) collapses on fRS given by (18) up to a trivial additive constant which corresponds to
0
te and a rescaling of the measurement rate α = αβ.
ΨPout
Differentiating the free entropy w.r.t. ∆te , we deduce the limit of this optimal generalization error with
side information:
Proposition 2 (Optimal generalization error with side information). For almost every ∆te > 0 we have
p
p
√
side
lim Egen
= E ϕ( ρ V, A)2 − E ϕ( q ∗ (∆te ) V + ρ − q ∗ (∆te ) w, a)i2sc,∆te ,
n→∞
(47)
where q ∗ (∆te ) is a maximizer of (46) and where h−isc,∆te is the Gibbs bracket associated with the posterior
distribution of the second scalar channel (17) with ∆ = ∆te .
The proof of Proposition 2 is presented in Appendix B.2. A natural continuation would be to show that
for almost every value of ∆tr , q ∗ (∆te ) converges, as ∆te → ∞, to a maximizer q ∗ of the replica formula (20)
from Theorem 1. Using theorems for limit-expectations permutation, it should not be difficult to show then
side
that lim lim Egen
= E(q ∗ ). Then, provided we can justify the following permutation of limits we would
∆te →∞ n→∞
14
obtain
lim E opt
n→∞ gen
lim E side
n→∞ ∆te →∞ gen
= lim
side
lim Egen
= E(q ∗ ) .
= lim
(48)
∆te →∞ n→∞
We conjecture this formula to be true, but we were however unable to provide the technical justifications
mentioned above.
3.4.2
Optimal reconstruction error with side information
Analogously to the previous section, one can add small side information to our statistical model (1) by supposing that one observes also
Yi0 = Xi∗ +
√
∆side Zi0 ,
(49)
1 ≤ i ≤ n,
side > 0 quantifies the noise level of the
where (Zi0 )ni=1 i.i.d.
∼ N (0, 1) independently of everything else. Here ∆
side information and will typically be very large. The proof of Theorem 1 can be directly adapted to obtain
that the new free energy converges to
n
side
side
f∞
= sup inf fRS
(q, r) = sup inf ψP0 r +
q∈[0,ρ] r≥0
q∈[0,ρ] r≥0
1
∆side
+ αΨPout (q; ρ) −
rq o
.
2
(50)
By then taking the derivative w.r.t. (∆side )−1 , which is again a I-MMSE relation [40], one deduces the minimum mean-square-error (MMSE) for the reconstruction of X∗ :
Proposition 3 (Reconstruction MMSE with side information). For almost all ∆side > 0,
MMSEn :=
1 ∗
2
E X − hxi
−−−→ ρ − q ∗ (∆side ),
n→∞
n
(51)
where q ∗ (∆side ) is a maximizer of (50).
4 Application to concrete situations
In this section, we show how our main results can be applied to several models of interest in fields ranging
from machine learning to signal processing, and unveil several interesting new phenomena in learning of
generalized linear models. For various specific cases of prior P0 and output Pout , we evaluate numerically the
free entropy potential (18), its stationary points Γ and identify which of them gives the information-theoretic
results, i.e. is the maximizer in (20). We also identify which of the stationary points corresponds to the result
obtained asymptotically by the GAMP algorithm, i.e. the fixed point of the state evolution (32). Finally we
compute the corresponding generalization error (43). We stress that in this section the results are based on
numerical investigation of the resulting formulas: We do not aim at rigor that would involve precise bounds
and more detailed analytical control for the corresponding integrals.
4.1
Generic observations
Using the functions gPout and gP0 introduced in Sec. 2.5 we can rewrite the fixed point equations (19) as
q = 2ψP0 0 (r) = E[gP0 (Y0 , r)2 ] ,
α
√
E[gPout (Ye0 , q V, ρ − q)2 ] ,
r = 2αΨ0Pout (q) =
ρ−q
15
(52)
(53)
where the expectation in (52) corresponds to the scalar channel (12) and the expectation in (53) corresponds
to the second scalar channel (15).
Non-informative fixed point and its stability: It is interesting to analyze under what conditions q ∗ = 0
is the optimizer of (20). Notice that q ∗ = 0 corresponds to the error on the recovery of the signal as large as
it would be if we had no observations at our disposition. Theorem 1 gives that any optimal couple (q ∗ , r∗ ) of
(20) should be a fixed point of the state evolution equations (52)-(53). A sufficient condition for (q, r) = (0, 0)
to be a fixed point of (52)-(53) is that
(a) The transition density Pout (y|z) is even in the argument z.
(b) The prior P0 has zero mean.
In order to see this, notice that if Pout (y|z) is even in z then from the definition (30) of the function gPout we
have gPout (y, 0, ρ) = 0 and consequently from (53) we have Ψ0Pout (0) = 0. For the second point, notice that
we have ψP0 0 (0) = 21 EP0 [X0 ]2 = 0.
We assume now that the transition density Pout (y|z) is even in the argument z and that the prior P0 has
zero mean. In order for q = 0 to be the global maximizer q ∗ of (20) or to be a relevant fixed point of the state
evolution (32) (relevant in the sense that GAMP might indeed converge to it in a practical setting) we need
q = 0 to be a numerically stable fixed point of the above equations (52)-(53). We therefore need to expand
(52)-(53) around q = 0, and doing so, we obtain that q = 0 is stable if
2
2αΨ00Pout (0) × 2ψP00 0 (0) = αE hw2 isc − hwi2sc − 1
< 1,
(54)
where the expectation corresponds to the scalar channel (15) with q = 0 and the Gibbs bracket h−isc is given
by (17). The expectation quantifies how the observation of Ye0 in the scalar channel (15) modifies the variance
of W ∗ (which is 1 without any observation). Rewriting this condition more explicitly into a form that is
convenient for numerical evaluation we get
Z
αρ
R
0 (y|√ρz) 2
DzzPout
dy R
< 1,
√
DzPout (y| ρz)
(55)
0 (y|ω) stands for the partial derivative
where recall that Dz is a standard Gaussian measure and where Pout
w.r.t. ω. We conjecture that the condition (54) (or equivalently (55) for differentiable models) delimits precisely
the region where polynomial-time algorithms do not perform better than “random guessing” (see the discussion below, where we will make this stability condition explicit for several examples of symmetric output
channels).
Exact recovery fixed point: Another particular fixed point of (52)-(53) that we observe is the one corresponding to exact recovery q ∗ = ρ. A sufficient and necessary condition for this to be a fixed point is that
limq→ρ Ψ0Pout (q) = +∞. Heuristically, this means that the integral of the Fisher information of the output
channel should diverge:
2
Z
−ω
0 (y|ω)2
e 2ρ Pout
dydω √
= +∞ ,
(56)
2πρ Pout (y|ω)
0 (y|ω) denotes the partial derivative w.r.t. ω. This typically means that the channel should be
where again Pout
noiseless. For the Gaussian channel with noise variance ∆, the above expression equals 1/∆. For the probit
√
√
channel where Pout (y|z) = erfc(−yz/ 2∆)/2 the above expression at small ∆ is proportional to 1/ ∆.
Numerical stability of the exact recovery fixed point was also investigated, but we did not obtain any
unified expression. The stability depends non-trivially on both the properties of the output channel, but also
16
on the properties of the prior. Below we give several examples where exact recovery either is or is not possible,
or where there is a phase transition between the two regimes.
4.2
Phase diagrams of perfect learning
In this section we consider deterministic (noiseless) output channels and ask: How many measurements are
needed in order to perfectly recover the signal?
Our crucial point is to compare with the well explored phase diagram of Bayesian (noiseless) compressed
sensing in the case of the linear channel [52, 53]. As the number of samples (measurements) varies we encounter four different regimes of parameters:
• The EASY phase: This is the region in the parameter space where GAMP achieves perfect reconstruction.
• The IMPOSSIBLE phase: Region where perfect reconstruction is information-theoretically impossible.
• The HARD phase: Region where the perfect reconstruction is information-theoretically possible, but
where GAMP is unable to achieve it. However, in this region GAMP leads to a better generalization
error than the one corresponding to the non-informative fixed point. It remains a challenging open
question whether polynomial-time algorithms can achieve perfect reconstruction in this regime.
• The HARD-RG phase: This “hard-random guess” phase corresponds to the region where perfect reconstruction is information-theoretically possible but where GAMP only achieves an error as bad as
randomly guessing, given by the trivial fixed point. In this phase as well, the existence of polynomialtime exact recovery algorithms is an open question. This phase does not exist for the linear channel.
Some of the codes used in this section can be consulted online on the github repository [1].
4.2.1
The linear channel
The case of exact recovery of a sparse signal after it passed trough a noiseless linear channel, i.e. ϕ(x) = x, is
studied in the literature in great details, especially in the context of compressed sensing [24, 31]. For a signal
with a fraction ρ of non-zero entries it is found that as soon as α > ρ, perfect reconstruction is theoretically
possible, although it may remain computationally difficult. The whole field of compressed sensing builds on
the realization that, using the `1 norm minimization technique, one can efficiently recover the signal for larger
α values than the so-called Donoho-Tanner transition [29, 31].
In the context of the present paper, when the empirical distribution of the signal is known, one can fairly
easily beat the `1 transition and reconstruct the signal up to lower values of α using the Bayesian GAMP
algorithm [31, 52, 53, 70]. In this case, three different phases are present [52, 53]: i) For α < ρ, perfect reconstruction is impossible; ii) for ρ < α < αs reconstruction is possible, but not with any known polynomialcomplexity algorithm; iii) for α > αs , the so-called spinodal transition computed with state evolution, GAMP
provides a polynomial-complexity algorithm able to reach perfect reconstruction. The line αs (ρ) depends on
the distribution of the signal. For a Gauss-Bernoulli signal with a fraction ρ of non-zero (Gaussian) values we
compare the GAMP performance to the optimal one in Fig. 1 (left and right).
4.2.2
The rectified linear unit (ReLU) channel
Let us start by discussing the case of a generalized linear model with the ReLU output channel, i.e. ϕ(x) =
√
max(0, x), with a signal coming from a Gauss-Bernoulli distribution P0 (x) = ρ exp(−x2 /2)dx/ 2π + (1 −
ρ)δ(x) with a fraction ρ of non-zero (Gaussian) values. We are motivated by the omnipresent use of the ReLU
activation function in deep learning, and explore its properties for GLMs that can be seen as a simple single
layer neural network.
17
Absolute value channel, Gauss-Bernoulli X*
Symmetric Door activation, width Κ, Rademacher X*
3
1.2
1
2.5
EASY
EASY
1.5
IMPOSSIBLE
0
1
HARD
IMPOSSIBLE
1
0.5
ρ
0.5
IMPOSSIBLE
Perfect recovery
Spinodal
instability at q=0
Spinodal CS
0
HARD,
RG
1.5
HARD
HARD, RG
0.2
α=m/n
0.6
0.4
EASY
2
HARD
α=m/n
α=m/n
0.8
Relu activation, Gauss-Bernoulli signal
2
0.5
instability at q=0
Perfect recovery
Spinodal
0
1
0
0.5
1
1.5
Κ
Perfect recovery
Spinodal
Naive spinodal
0
2
2.5
0
0.5
ρ
1
Figure 1: Phase diagrams for several non-linear output channels showing the region where a perfect recovery
is possible/hard/impossible. Left: The phase diagram for the signless output with ϕ(x) = |x| with a GaussBernoulli signal, as a function of α = m/n and the fraction of non-zero components ρ. We find that a perfect
recovery is IMPOSSIBLE for α < ρ. Perfect recovery becomes possible starting from α > ρ, as for the linear
channel. On the algorithmic side, GAMP is not able to perform better than a random guess as long as α < 0.5. We
denote this region HARD-RG, for “not better than random guess”. For larger values, the inference using GAMP
leads to better results than a purely random guess but cannot reach perfect recovery, so the problem remains
HARD. GAMP can perfectly identify the hidden signal only for values of α larger than the so-called spinodal
(or algorithmic threshold), when the problem becomes EASY. For comparison we also show the algorithmic
transition for the linear output (the spinodal CS dashed line). Middle: Phase diagram for the door function
problem, with ϕ(x) = sgn(|x| − K), for a Rademacher signal as a function of α and K. The same four regions of
parameters are observed and depicted. Right: Phase diagram for the ReLU output, with ϕ(x) = max(0, x), again
with the EASY, HARD and IMPOSSIBLE regions. Here it is always possible to generalize better than chance using
GAMP. The so-called naive spinodal shows for comparison the algorithmic performance when using information
only about the non-zero observations.
Our analysis shows that a perfect generalization (and thus a perfect reconstruction of the signal as well)
is possible whenever the number of samples per dimension (measurement rate) α > 2ρ, and impossible when
α < 2ρ. This is very intuitive, since half of the measurements (those non-zero) are giving as much information
as in the linear case, thus the factor 2.
How hard is it to actually solve the problem with an efficient algorithm? The answer is given by applying
the state evolution analysis to GAMP, which tells us that only for even larger values of α, beyond the spinodal
transition, does GAMP reach a perfect recovery. Notice, however, that this spinodal transition occurs at a
significantly lower measurement rate α than one would reach just keeping the non-zero measurements. This
shows that, actually, these zero measurements contain a useful information for the algorithm. The situation
is shown in the right side of Fig. 1: The zero measurements do not help information-theoretically but they,
however, do help algorithmically.
4.2.3
The signless channel
We now discuss the signless channel where only the absolute value of the linear mixture is observed, i.e.
ϕ(x) = |x|. This case can be seen as the real-valued analog of the famous phase retrieval problem. We again
consider the signal to come from a Gauss-Bernoulli distribution with a fraction ρ of non-zero (Gaussian)
values.
Sparse phase retrieval has been well explored in the literature in the regime where the number s of nonzeros is sub-leading in the dimension, s = O(n). This case is known to present a large algorithmic gap.
While analogously to compressed sensing exact recovery is information-theoretically possible for a number
of measurement O(s ln(n/s)), best known algorithms achieve it only with O(s2 / ln n) measurements [67],
see also [75] and references therein for a good discussion of other related literature. This is sometimes referred
18
to as the s2 barrier. We are not aware of a study where, as in our setting, the sparsity is s = ρn and the number
of measurements is αn with α = O(1). Our analysis in this regime hence sheds a new light on the hardness
of the problem of recovering a sparse signal from signless measurements.
Our analysis of the mutual information shows that a perfect reconstruction is information-theoretically
possible as soon as α > ρ: In other words, the problem is –information-theoretically– as easy, or as hard as
the compressed sensing one. This is maybe less surprising when one thinks of the following algorithm: Try
all 2m choices of the possible signs for the m outputs, and solve a compressed sensing problem for each of
them. Clearly, this should yields a perfect solution only in the case of the actual combination of signs.
Algorithmically, however, the problem is much harder than for the linear output channel. As shown in
the left side of Fig. 1, for small ρ one requires a much larger fraction α of measurements in order for GAMP
to recover the signal. For the linear channel the algorithmic transition αs (ρ) → 0 as ρ → 0, while for the
signless channel we get αs (ρ) → 1/2 as ρ → 0. In other words if one looses the signs one cannot perform
recovery in compressed sensing with less than n/2 measurements.
What we observe in this example for α < 1/2 is in the statistical physics literature on neural networks
known as retarded learning [43]. This appears in problems where the ϕ(x) function is symmetric, as seen at
the beginning of this section: There is always a critical point of the mutual information with an overlap value
q = 0. For this problem, this critical point is actually “stable” (meaning that it is actually a local minimum in
q) for all α < 1/2 independently of ρ. To see that, we have to go back to (54). For the absolute value channel,
√
the posterior distribution of W ∗ given Ye0 = ρ |W ∗ | in the second scalar channel (15) is a mean of two Dirac
masses on −W ∗ and W ∗ . Thus, the posterior variance is hw2 isc − hwi2sc = (W ∗ )2 . Consequently, (54) leads
to the stability condition α < 1/2.
This has the two following implications: i) In the IMPOSSIBLE phase, when α < 1/2 and ρ > α, the
minimum at q = 0 is actually the global one. In this case, the MMSE and the generalization error are the ones
given by using 0 as a guess for each element of X∗ ; in other words, there is no useful information that one
can exploit and no algorithmic approach can be better than a random guess. ii) In the POSSIBLE but HARD
phase when α < 1/2, GAMP initialized at random, i.e. close to the q = 0 fixed point, will remain there. This
suggests that in this region, even if a perfect reconstruction is possible, it will still be very hard to beat a
random guess with a tractable algorithm.
4.2.4
The symmetric door channel
The third output channel we study in detail is the symmetric door channel, where ϕ(x) = sgn(|x| − K). In
case of channels with discrete set of outputs exact recovery is only possible when the prior is also discrete. In
the present case we consider the signal to be Rademacher, where each element is chosen at random between
1 and −1, i.e. P0 (x) = (δ(x − 1) + δ(x + 1))/2. This channel was studied previously using the replica method
in the context of optimal data compression [44].
This output channel is in the class of symmetric channels for which overlap q = 0 is a fixed point. This
fixed point is stable for α < αc (K). Exact recovery is information-theoretically possible above αIT (K) and
tractable with the GAMP algorithms above the spinodal transition αs (K). The values of these three transition
lines are depicted in the center panel of Fig. 1.
We note that αIT ≥ 1 is a generic bound on exact recovery for every K, required by a simple counting
argument. While a-priori it is not clear whether this bound is saturated for some K, we observe that it is for
K = 0.67449 such that half of the observed measurements are negative and the rest positive. This is, however,
not efficiently achievable with GAMP. The saturation of the αIT ≥ 1 bound was remarked previously in the
context of the work [44] on optimal data compression. Our work predicts that this information-theoretic
result will not be achievable with known efficient algorithms.
19
Generalization error
Perceptron, Rademacher X*
Perceptron, Gauss-Bernoulli X*, ρ=0.2
Symmetric Door, Rademacher X*
1
1
0.8
0.8
0.6
0.6
0.6
0.4
0.4
0.4
0.2
0.2
SE
Optimal
AMP,N=1044
Logistic,N=10
0.2
0
0
0.5
1
Optimal and SE
AMP, N=104
Logistic,N=104
SE
Optimal
AMP,N=104
0.8
1
2-layer NN
0
0
0
1
α
1.5
2
5
10 15 20
0
0
0.5
1
1.5
2
2.5
α
3
3.5
4
4.5
5
0
0.5
1
1.5
2
2.5
α
Figure 2: Generalization error in three classification problems as a function of the number of data-samples per
dimension α. The red line is the Bayes-optimal generalization error, while the green one shows the (asymptotic)
performances of GAMP as predicted by the state evolution (SE), when different. For comparison, we also show
the result of GAMP (black dots) and, in blue, the performance of a standard out-of-the-box solver, both tested on a
single randomly generated instance. Left: Perceptron, with ϕ(x) = sgn(x) and a Rademacher (±1) signal. While
a perfect generalization is information-theoretically possible starting from α = 1.249(1), the state evolution
predicts that GAMP will allow such perfect prediction only from α = 1.493(1). The results of a logistic regression
with fine-tuned ridge penalty with the software scikit-learn [68] are shown for comparison. Middle: Perceptron
with Gauss-Bernoulli coefficients for the signal. No phase transition is observed in this case, but a smooth
decrease of the error with α. The results of a logistic regression with fine-tuned `1 sparsity-enhancing penalty
(again with [68]) are very close to optimal. Right: The symmetric door activation rule with parameter K =
0.67449 chosen in order to observe the same number of occurrence of the two classes. In this case there is a sharp
phase transition at α = 1 from a situation where it is impossible to learn the rule, so that the generalization is
not better than a random guess, to a situation where the optimal generalization error drops to zero. However,
GAMP identifies the rule perfectly only starting from αs = 1.566 (GAMP error stays 1 up to αstab = 1.36, see
the black dashed curve). Interestingly, this non-linear rule seems very hard to learn for other existing algorithms.
Using Keras [26], a neural network with two hidden layers was able to learn approximately the rule, but only for
much larger training set sizes (shown in inset).
4.3
Examples of optimal generalization error
Besides the formula for the mutual information, the main result of this paper is the formula (40) (or (47) when
∆te → +∞) for the optimal generalization error, and formula (43) for the generalization error achieved by
the GAMP algorithm. In this section we evaluate both these generalization errors for several cases of priors
and output functions. We study both regression problems, where the output is real-valued, and classification
problems, where the output is discrete.
While in realistic regression and classification problems the matrix Φ corresponds to the data, and is thus
not i.i.d. random, we view the practical interest of our theory as a benchmark for state-of-the art algorithms.
Our work provides an exact asymptotic analysis of optimal generalization error and sample complexity for a
range of simple rules where a teacher uses random data to generate labels. The challenge for state-of-the-art
multi-purpose algorithms is to try to match as closely as possible the performance that can be obtained with
GAMP that is fine-tuned to the specific form of the output and prior.
4.3.1
Threshold output: The perceptron
The example of non-linear output that is the most widely explored in the literature is the threshold output,
where the deterministic output (or “activation”) function is ϕ(x) = sgn(x). This output in the teacher-student
setting of the present paper is known as the perceptron problem [36], or equivalently, the one-bit compressed
sensing in signal processing [21]. Its solution has been discussed in details within the replica formalism (see
for instance [16, 46, 66, 83]) and we confirm all of these heuristic computations within our approach. Let
20
V ∼ N (0, 1). The formula (40) for the generalization error then reduces to (recall q ∗ is a maximizer of (20))
lim E opt
n→∞ gen
Z
=1−
2 Z V
DV √
π 0
q
q∗
2(ρ−q ∗ )
−t2 /2
dt e
2
h
= 1 − E erf V
s
2 i
q∗
.
2(ρ − q ∗ )
(57)
In Fig. 2 (left) we plot the optimal generalization error of the perceptron with a Rademacher signal, the
state evolution prediction of the generalization error of the GAMP algorithm, together with the error actually achieved by GAMP on one randomly generated instance of the problem. We also compare these to the
performance of a standard logistic regression. As expected from existing literature [36,73] we confirm that in
this case the information-theoretic transition appears at a number of samples per dimension αIT = 1.249(1),
while the algorithmic transition is at αs = 1.493(1). Logistic regression does not seem to be able to match
the performance on GAMP in this case.
In Fig. 2 (center) we plot the generalization error for a Gauss-Bernoulli signal with density ρ = 0.2. Cases
as this one were studied in detail in the context of one-bit compressed sensing [83] and GAMP was found to
match the optimal generalization performance with no phase transitions observed. In this case the logistic
regression is rather close to the performance of GAMP.
4.3.2
Symmetric Door
The next classification problem, i.e. discrete output rule, we study is the symmetric door function ϕ(x) =
sgn(|x| − K). In this case the generalization error (40) becomes
lim E opt
n→∞ gen
hn K − √q ∗ V
K + √q ∗ V
o2 i
= 1 − EV erf p
− erf − p
−1
.
2(ρ − q ∗ )
2(ρ − q ∗ )
(58)
In Fig. 2 (right) we plot the generalization error for K = 0, 67449 such that 1/2 of the outputs are 1 and
1/2 are −1. The symmetric door output is an example of function for which the optimal generalization error
for α < αIT = 1 (for that specific value of K, see phase diagram in center of Fig. 1)) is as bad as if we were
guessing randomly. The GAMP algorithm still achieves such a bad generalization until αstab = 1.36, and
achieves perfect generalization only for α > αs = 1.566.
Interestingly, labels created from this very simple symmetric door rule seem to be very challenging to
learn for general purpose algorithms. We tried to optimize parameters of a two-layers neural network and
only managed to get the performances shown in the inset of Fig. 2 (right). It is an interesting theoretical
challenge whether a deeper neural network can learn this simple rule from fewer samples.
4.3.3
Linear regression
The additive white Gaussian noise (AWGN) channel, or linear regression, is defined by ϕ(x, A) = x+σA with
A ∼ N (0, 1). This models the (noisy) linear regression problem, as well as noisy random linear estimation
and compressed sensing. In this case (40) leads to
lim E opt
n→∞ gen
= ρ − q∗ + σ2 .
(59)
This result agrees with the generalization error analyzed heuristically in [73] in the limit σ → 0. Fig. 3 (left)
depicts the generalization error for this example. Performance of GAMP in this case very close to the one of
LASSO.
21
Generalization error
AWGN σ= 0.032, Gauss-Bernoulli X*, ρ=0.2
ReLu, Gauss-Bernoulli X*, ρ=0.2
0.08
0.07
0.06
0.05
0.04
0.03
0.02
0.01
0
0.2
0.15
0.1
Optimal and SE
AMP,N=104
LASSO,N=104
0.05
0
0
0.2
0.4
0.6
Absolute value, Gauss-Bernoulli X*, ρ=0.5
0.2
SE
Optimal
AMP, N=104
Keras,N=104
0.15
SE
Optimal
AMP,N=104
0.6
0.1 0.4
0.2
0.05
PM
0
0 1 2 3 4
0
0
0.2
0.4
0.6
α
α
0.8
1
1.2
0
0.2
0.4
0.6
0.8
1
1.2
1.4
α
Figure 3: The generalization error for three regression problems is plotted as a function of the number of samples
per dimension α. The red line is again the Bayes-optimal generalization error, while the green one shows the
(asymptotic) performances of GAMP as predicted by the state evolution (SE), when different. Again, we also
show the result of GAMP on a particular instance (black dots) and, in blue, the performance of an out-of-the-box
solver. Left: White Gaussian noise output and a Gauss-Bernoulli signal. For this choice of noise, there is no sharp
transition (as opposed to what happens at smaller noises). The results of a LASSO with fine-tuned `1 sparsityenhancing penalty (with [68]) are very close to optimal. Middle: Here we analyze a ReLU output function
ϕ(x) = max(0, x), still with a Gauss-Bernoulli signal. Now there is an information-theoretic phase transition at
α = 2ρ = 0.4, but GAMP requires αs = 0.589 to reach perfect recovery. We show for comparison the result of
a maximum likelihood estimation performed with Keras [26] and observe a sizeable gap in performance. Right:
The signless output function ϕ(x) = |x|. The information-theoretic perfect recovery starts at α = ρ = 0.5,
but the problem is again harder algorithmically for GAMP that succeeds only above αs = 0.90. Again, the
problem appears to be hard for other solvers. In inset, we show the performance for the estimation problem
using PhaseMax [37], which is able to learn the rule only using about four times as many measurement than
needed information-theoretically.
4.3.4
Rectified linear unit (ReLU)
In Fig. 3 (center) we analyze the generalization error for the ReLU output function, ϕ(x) = max(0, x). This
channel models the behavior of a single neuron with the rectified linear unit activation [54] widely used in
multilayer neural networks. In this case (40) becomes after simple algebra and Gaussian integration by parts
(again V ∼ N (0, 1)),
lim E opt
n→∞ gen
h
ρ q∗
1 + EV V 2 erf V
= −
2
4
s
2 i (ρ − q ∗ )3/2 1
q∗
q∗
√
−
+
2(ρ − q ∗ )
ρ + q ∗ 2π ρπ
r
ρ + q∗
.
ρ − q∗
(60)
For sparse Gauss-Bernoulli signals in Fig. 3 (center) we observe again the information-theoretic transition
to perfect generalization to be distinct from the algorithmic one. At the same time our test with existing
algorithms were not able to closely match the performance of GAMP. This hence also remains an interesting
benchmark.
4.3.5
Signless channel
In Fig. 3 (right) we analyze the generalization error for the signless output function where ϕ(x) = |x|. This
models a situation similar to compressed sensing, except that the sign of the output has been lost. This is
a real-valued analog of the phase retrieval problem as discussed in Sec. 4.2.3. In this case the generalization
error (40) becomes
p
opt
lim Egen
= ρ − EV b(V q ∗ , ρ − q ∗ )2 ,
n→∞
(61)
where
r
b(x, y) =
2y − x2y2
x
x xn
x o
e
+ erfc − √
−
1 + erf − √
.
π
2
2
2y
2y
22
(62)
Our comparison with the performance of a state-of-the-art algorithm PhaseMax [37] suggests that also
for this simple benchmark there is room for improvement in term of matching the performance of GAMP.
4.3.6
Sigmoid, or logistic regression
Let us also consider an output function with auxiliary randomization. After having generated the classifier
X∗ , the teacher randomly associates the label +1 to the pattern Φµ with probability fλ (n−1/2 Φµ ·X∗ ), where
fλ (x) = (1 + exp(−λx))−1 ∈ [0, 1] is the sigmoid of parameter λ > 0, and the label −1 with probability
1 − fλ (n−1/2 Φµ · X∗ ). One of the (many) possible ways for the teacher to do so is by selecting ϕ(x, A) =
I(A ≤ fλ (x)) − I(A > fλ (x)), where I(E) is the indicator function of the event E. He then generates a
i.i.d.
stream of uniform random numbers A i.i.d.
∼ U[0,1] and obtains the labels through (35). Let V, w ∼ N (0, 1). In
this setting the error (40) becomes
opt
lim Egen
= 2 − 4 EV
n→∞
p
p
Ew fλ ( q ∗ V + ρ − q ∗ w)
2
.
(63)
This fomula reduces to (57) when λ → ∞ as it should.
5 Proof of the replica formula by the adaptive interpolation
method
We now prove Theorem 1. Our main tool will be an interpolation method recently introduced in [13] and
called “adaptive interpolation method”. Here we formulate the method as a direct evolution of the Guerra and
Toninelli interpolation method developed in the context of spin glasses [39]. In constrast with the discrete
and more pedestrian version of the adaptive interpolation method presented in [13] (and called in this case
“stochastic interpolation method”), here we employ a continuous approach which is more straightforward
(see [13] for the links between the discrete and continuous versions of the method) and that has also been
recently used in [8] for studying non-symmetric tensor estimation.
We will prove Theorem 1 under the following hypotheses:
(H1) The prior distribution P0 has a bounded support.
(H2) ϕ is a bounded C 2 function with bounded first and second derivatives w.r.t. its first argument.
(H3) (Φµi ) i.i.d.
∼ N (0, 1).
These assumptions will then be relaxed in Appendix D to the assumptions (h1), (h2) and (h3) or (h3.a)-(h3.b)(h3.c). Since the observations (1) are equivalent to the rescaled observations
1
Yeµ := ∆−1/2 Yµ = ∆−1/2 ϕ √ [ΦX∗ ]µ , Aµ + Zµ ,
n
1 ≤ µ ≤ m,
(64)
the variance ∆ of the Gaussian noise can be “incorporated” inside the function ϕ. Thus, it suffices to prove
Theorem 1 for ∆ = 1 and we suppose, for the rest of the proof, that we are in this equivalent case.
5.1
Interpolating estimation problem
We introduce an “interpolating estimation problem” that interpolates between the orginal problem (2) at t = 0,
t ∈ [0, 1] being the interpolation parameter, and the two scalar problems described in Sec. 2.2 at t = 1 which
are analytically tractable. For t ∈ (0, 1) the interpolating estimation problem is a mixture of the original
and scalar problems. This interpolation scheme is inspired from the interpolation paths used by Talagrand to
23
study the perceptron, see [76]. There are two major differences between the “non-planted perceptron” studied
by Talagrand, and the “planted perceptron” that we are investigating:
• In the planted case, the presence of a planted solution forces (under small perturbations) the correlations
to vanish for all values of the parameters, see [48,63]. In the non-planted case, proving such decorelation
is much more involved, and is proved only in a limited region of the parameter space, see [76].
• However, in the planted case, there can be arbitrarily many (see Remark 21 in [82]) solutions to the
state evolution equations (19), whereas in the region studied by [76], there is only one solution. For this
reason, our interpolation method needs to be more sophisticated in order to interpolate with the “right
fixed point”.
Let q : [0, 1] → [0, ρ] be a continuous “interpolation function” and r ≥ 0. Define
r
St,µ :=
1−t
[ΦX∗ ]µ +
n
sZ
sZ
t
t
q(v)dv Vµ +
0
(ρ − q(v))dv Wµ∗
(65)
0
where Vµ , Wµ∗ i.i.d.
∼ N (0, 1). Consider the following observation channels, with two types of observations
obtained through
(
Yt,µ ∼ Pout ( · | St,µ ) ,
1 ≤ µ ≤ m,
√
(66)
0
Yt,i
=
r t Xi∗ + Zi0 ,
1 ≤ i ≤ n,
m
where (Zi0 )ni=1 i.i.d.
∼ N (0, 1). We assume that V = (Vµ )µ=1 is known. Then the inference problem is to recover
∗
∗
m
∗
∗
n
both unknowns W = (Wµ )µ=1 and X = (Xi )i=1 from the knowledge of V, Φ and the “time-dependent”
0
0 n
observations Yt = (Yt,µ )m
µ=1 and Yt = (Yt,i )i=1 .
We now understand that rt appearing in the second set of measurements in (66), and the terms 1 − t,
Rt
Rt
q(v)dv
and 0 (ρ − q(v))dv appearing in the first set all play the role of signal-to-noise ratios in the in0
terpolating model, with t giving more and more “power” (or weight) to the scalar inference channels when
increasing. Here is the first crucial and novel ingredient of our interpolation scheme. In the classical interpolation method, these signal intensities would all take a trivial form (i.e. would be linear in t) but here, the
non-trivial (integral) dependency in t of the two latter intensities through the use of the interpolation function
q allows for much more flexibility when choosing the interpolation path. This will allow us to actually choose
the “optimal interpolation path” (this will become clear soon).
Define uy (x) := ln Pout (y|x) and, with a slight abuse of notations,
r
st,µ = st,µ (x, wµ ) :=
1−t
[Φx]µ +
n
sZ
sZ
t
0
t
(ρ − q(v))dv wµ .
q(v)dv Vµ +
(67)
0
We introduce the interpolating Hamiltonian
Ht (x, w; Y, Y0 , Φ, V) := −
m
X
n
ln Pout (Yµ |st,µ ) +
µ=1
=−
m
X
2
1X 0 √
Yi − t r xi
2
i=1
uYt,µ (st,µ ) +
µ=1
n
1X √
2
t r (Xi∗ − xi ) + Zi0
2
.
(68)
i=1
The dependence in Φ and V of the Hamiltonian is through the (st,µ )m
µ=1 . We also introduce the corresponding
Gibbs bracket h−it which is the expectation operator w.r.t. the t-dependent posterior distribution of (x, w)
24
given (Yt , Yt0 , Φ, V). It is defined as
1
hg(x, w)it :=
Zt (Yt , Yt0 , Φ, V)
Z
0
dP0 (x)Dw g(x, w) e−Ht (x,w;Yt ,Yt ,Φ,V) ,
(69)
Q
2 /2
−wµ
for every continuous bounded function g on Rn × Rm . In (69) Dw = (2π)−m/2 m
is the
µ=1 dwµ e
0
m-dimensional standard Gaussian distribution and Zt (Yt , Yt , Φ, V) is the appropriate normalization:
Z
0
0
Zt (Yt , Yt , Φ, V) := dP0 (x)Dw e−Ht (x,w;Yt ,Yt ,Φ,V) .
(70)
Finally the interpolating free entropy is
fn (t) :=
1
E ln Zt (Y, Y0 , Φ, V) .
n
(71)
One verifies easily that
(
fn (0) = fn − 21 ,
fn (1) = ψP0 (r) −
1+rρ
2
+
R1
m
n ΨPout ( 0 q(t)dt; ρ) ,
(72)
where fn is given by (11). We emphasize a crucial property of the interpolating model: It is such that at t = 0
we recover the original model and thus fn (0) = fn − 1/2 (the trivial constant comes from the purely noisy
measurements of the second channel in (66)), while at t = 1 we have the two scalar inference channels and
thus the associated terms ψP0 and ΨPout discussed in Sec. 2.2 appear in fn (1). These are precisely the terms
appearing in the potential (18).
5.2
Free entropy variation along the interpolation path
From the understanding of the previous section, it is at this stage very natual to evaluate the variation of free
entropy along the interpolation path, which allows to “compare” the original and purely scalar models thanks
to the identity
1
fn = fn (0) + = fn (1) −
2
Z
0
1
1
dfn (t)
dt + ,
dt
2
(73)
where the first equality follows from (72). As discussed above, part of the potential (18) appears in fn (1). If the
interpolation is properly done, the missing terms required to obtain the potential on the r.h.s. of (73) should
naturally appear. Then by choosing the optimal interpolation path thanks to the non-trivial snr dependencies
in t (i.e. by selecting the proper interpolating function q), we will be able to show the equality between the
replica formula and the true entropy fn .
We thus now compute the t-derivative of the free entropy along the interpolation path (see Appendix C
for the proof). Let u0y (x) be the derivative (w.r.t. x) of uy (x). Then we have the following.
Proposition 4 (Free entropy variation). For model (2), the derivative of the free entropy (71) verifies, for all
t ∈ (0, 1),
m
E
dfn (t)
1 D 1 X 0
rq(t) rρ
=− E
uYt,µ (St,µ )u0Yt,µ (st,µ ) − r Q − q(t)
+
−
+ On (1) ,
dt
2
n
2
2
t
µ=1
25
(74)
where On (1) is a quantity that goes to 0 in the n, m → ∞ limit, uniformly in t ∈ [0, 1] and the overlap is
n
Q :=
1X ∗
Xi xi .
n
(75)
i=1
5.3
Overlap concentration
The next lemma plays a key role in our proof. Essentially it states that the overlap concentrates around
its mean, a behavior called “replica symmetric” in statistical physics. Similar results have been obtained in
the context of the analysis of spin glasses [76]. Here we use a formulation taylored to Bayesian inference
problems as developed in the context of LDPC codes, random linear estimation [14] and Nishimori symmetric
spin glasses [47, 48, 57].
We introduce a “small” perturbation of the interpolating estimation problem by adding to the Hamiltonian
(68) a term
n 2
X
√
x
bi
i − xi Xi∗ − xi Z
2
(76)
i=1
where (Zbi )ni=1 i.i.d.
∼ N (0, 1). This√term can be interpreted as having a set of extra observations coming from
a Gaussian side-channel Ybi = Xi∗ + Zbi for i = 1, . . . , n and that preserves the Nishimori identity (see
Appendix A.1). The new Hamiltonian Ht, (x, w; Y, Y0 , Φ, V) defines a new Gibbs bracket h−in,t, and free
entropy fn, (t), and all the set up of Sec. 5.1 and Proposition 4 trivially extend. This perturbation induces
only a small change in the free entropy, namely of the order of :
Lemma 1 (Small free entropy variation under perturbation). Let S > 0 such that the support of P0 is included
in [−S, S]. For all > 0 and all t ∈ [0, 1],
|fn, (t) − fn (t)| ≤
S2
.
2
(77)
Proof. A simple computation gives
∂fn, (t)
1
S2
= EhQin,t, ≤
,
∂
2
2
(78)
which proves the lemma.
Moreover, this small perturbation forces the overlap to concentrate around its mean:
Lemma 2 (Overlap concentration). For any 0 < a < b < 1,
b
Z
lim
n→∞ a
Z
1
dt E Q − EhQin,t,
d
0
2
n,t,
= 0.
(79)
In Appendix F.2 we briefly sketch the main steps of the proof for the convenience of the reader and refer
to [13] for more details where the method has been streamlined.
∗
Lemma 2 implies that there exists a sequence (n )n≥1 ∈ (0, 1)N that converges to 0 such that
Z
lim
n→∞ 0
1
dt E Q − EhQin,t,n
2
n,t,n
= 0.
(80)
(n )n≥1 converges to 0, so Lemma 1 gives that fn,n (t) and fn (t) have the same limit (provided it exists). In
26
the rest of the section, in order to alleviate the notations, we abusively remove the perturbation subscript n
since it makes no difference for the computation of the limit of the free entropy.
5.4
Canceling the remainder
Note from (72) and (18) that the second and third terms appearing in (74) are precisely the missing ones
that are required in order to obtain the expression of the potential on the r.h.s. of (73). Thus in order to
prove Theorem 1 we would like to “cancel” the Gibbs bracket in (74), which is the so called remainder (once
integrated over t). This is made possible thanks to the adaptive interpolating parameter q. We would like to
choose q(t) = E hQit , which is approximately equal to Q because it concentrates, see Lemma 2. However,
Rt
E hQit depends on 0 q(v)dv (and on r too). The equation q(t) = E hQit is therefore a first order differential
Rt
equation over t 7→ 0 q(v)dv.
Proposition 5 (Existence of the optimal interpolation function). For all r ≥ 0 the differential equation
(81)
q(t) = E hQit
(r)
admits a unique solution qn (t) on [0, ρ] and the mapping
Z
1
r ≥ 0 7→
(82)
qn(r) (v)dv
0
is continuous.
Rt
Proof. One verify easily that E hQit is a bounded C 1 function of ( 0 q(v)dv, r). The proposition follows then
from an application of the parametric Cauchy-Lipschitz theorem.
Using this optimal choice for the interpolating function allows then to relate the potential and free entropy.
(r )
n
Proposition 6. Let (rn )n≥1 ∈ RN
be the solution of (81). Then
+ be a bounded sequence. For n ∈ N, let qn
Z
1
fn = fRS
qn(rn ) (v)dv, rn
+ On (1) .
(83)
0
(r )
Proof. qn n satisfies (81). Therefore by the Cauchy-Schwarz inequality
m
D 1 X
E
u0Yt,µ (St,µ )u0Yt,µ (st,µ ) − rn Q − qn(rn ) (t)
dt E
n
t
0
µ=1
m
Z 1
D 1 X
2 E 1/2 Z 1
2 1/2
≤
dt E
u0Yt,µ (St,µ )u0Yt,µ (st,µ ) − rn
dt E Q − qn(rn ) (t) t
= On (1) .
n
t
0
0
Z
1
µ=1
The last equality uses that the first factor is bounded (independently of t) because we supposed that Pout is
generated by (64) with assumptions (H1), (H2) and (H3) (see Appendix A.5 for proof details) and the second
factor goes to 0 when n, m → ∞ by (80), (81). Therefore from (74)
Z
0
1
rn
dfn (t)
dt =
dt
2
Z
1
qn(rn ) (t)dt −
0
rn ρ
+ On (1) .
2
(84)
When replacing this identity in (73) and combining this with (72) we reach
m
fn = ψP0 (rn ) + ΨPout
n
Z
1
qn(rn ) (t)dt; ρ
0
27
rn
−
2
Z
0
1
qn(rn ) (t)dt + On (1) .
(85)
Recalling that m/n → α as m, n → ∞ allows to recognize from (18) the claimed identity (83).
5.5
Lower and upper matching bounds
We now possess all the necessary tools to prove Theorem 1 in three steps.
(i) We prove that, under assumptions (H1), (H2) and (H3), limn→∞ fn = supr≥0 inf q∈[0,ρ] fRS (q, r).
(ii) Under hypothesis (H2), the function ΨPout is convex, Lipschitz and non-decreasing (Proposition 11). We
thus apply Corollary 5 of Appendix E to get supr≥0 inf q∈[0,ρ] fRS (q, r) = supq∈[0,ρ] inf r≥0 fRS (q, r). We
then deduce from (i) that limn→∞ fn = supq∈[0,ρ] inf r≥0 fRS (q, r) under (H1)-(H2)-(H3).
(iii) Finally, the approximation arguments given in Appendix D permit to relax (H1)-(H2) to the weaker
hypotheses (h1)-(h2) and allow to replace the Gaussian assumption (H3) on Φ by (h3.a)-(h3.b)-(h3.c).
This proves the first equality of Theorem 1. The last equality in (20) and the remaining part of Theorem 1
follow then from Lemma 13.
It remains thus to tackle (i). Let us start by the lower bound.
Proposition 7 (Lower bound). The free entropy (10) verifies
lim inf fn ≥ sup inf fRS (q, r) .
n→∞
r≥0 q∈[0,ρ]
(86)
Proof. By Proposition 6 we have that for any r ≥ 0,
fn ≥ fRS
1
Z
qn(r) (t)dt, r
+ On (1) ≥ inf fRS (q, r) + On (1)
q∈[0,ρ]
0
(87)
and thus
lim inf fn ≥ inf fRS (q, r) .
n→∞
q∈[0,ρ]
(88)
This is true for all r ≥ 0 thus we obtain Proposition 7.
We now turn our attention to the converse bound.
Proposition 8 (Upper bound). The free entropy (10) verifies
lim sup fn ≤ sup inf fRS (q, r) .
r≥0 q∈[0,ρ]
n→∞
(89)
Proof. Let K = 2αΨ0Pout (ρ; ρ). The mapping from equation (82) is continuous, consequently the application
[0, K] →
[0, K]
R 1 (r)
0
r
7→ 2αΨPout 0 qn (t)dt; ρ
(90)
is continuous (recall that Ψ0Pout denotes the derivative of ΨPout w.r.t. its first argument, and is shown to be
R 1 (r∗ )
continuous and bounded in Appendix A.2.2). It admits therefore a fixed point rn∗ = 2αΨ0Pout ( 0 qn n (t)dt; ρ).
Proposition 6 gives then
Z 1 ∗
fn = fRS
qn(rn ) (t)dt, rn∗ + On (1) .
(91)
0
28
We now remark that
fRS
Z
1
∗
qn(rn ) (t)dt, rn∗ = inf fRS (q, rn∗ ) .
q∈[0,ρ]
0
(92)
Indeed, the function grn∗ : q ∈ [0, ρ] 7→ fRS (q, rn∗ ) is convex (because of Proposition 11) and its q-derivative is
gr0 n∗ (q) = αΨ0Pout (q) −
rn∗
.
2
(93)
R 1 (r∗ )
R 1 (r∗ )
Since gr0 n∗ ( 0 qn n (t)dt) = 0 by definition of rn∗ , the minimum of grn∗ is necessarily achieved at 0 qn n (t)dt.
Combining (91) with (92) we reach
fn = inf fRS (q, rn∗ ) + On (1) ≤ sup inf fRS (q, r) + On (1)
q∈[0,ρ]
r≥0 q∈[0,ρ]
(94)
which allows to deduce Proposition 8.
From the arguments given at the beginning of the section, this ends the proof of Theorem 1.
Appendix A: Some technicalities
A.1
The Nishimori identity
Proposition 9 (Nishimori identity). Let (X, Y) ∈ Rn1 × Rn2 be a couple of random variables. Let k ≥ 1 and
let X(1) , . . . , X(k) be k i.i.d. samples (given Y) from the conditional distribution P (X = · |Y), independently
of every other random variables. Let us denote h−i the expectation operator w.r.t. P (X = · |Y) and E the
expectation w.r.t. (X, Y). Then, for all continuous bounded function g we have
Ehg(Y, X(1) , . . . , X(k) )i = Ehg(Y, X(1) , . . . , X(k−1) , X)i .
(95)
Proof. This is a simple consequence of Bayes formula. It is equivalent to sample the couple (X, Y) according
to its joint distribution or to sample first Y according to its marginal distribution and then to sample X
conditionally to Y from its conditional distribution P (X = · |Y). Thus the (k + 1)-tuple (Y, X(1) , . . . , X(k) )
is equal in law to (Y, X(1) , . . . , X(k−1) , X).
A.2
A.2.1
Some properties of the scalar channels
The additive Gaussian scalar channel
We recall here some properties of the free entropy of the first scalar channel (12).
Proposition 10. Let X0 ∼ P0 be a real random variable with finite second moment. Let r ≥ 0 and Y0 =
√
rX0 + Z0 , where Z0 ∼ N (0, 1) is independent from X0 . Then the function
Z
√
2
ψP0 : r 7→ E ln dP0 (x)e r Y0 x−rx /2
is convex, differentiable, non-decreasing and 12 E[X02 ]-Lipschitz on R+ . Moreover, ψP0 is strictly convex, if P0 is
not a Dirac measure.
Proof. Denote by h−ir the posterior distribution of X0 given Y0 in the channel (12). Using Gaussian integration by parts and the Nishimori property (Proposition 9), one verify easily that for all r ≥ 0
2ψP0 0 (r) = EhxX0 ir = Ehxi2r ∈ [0, E[X02 ]] ,
29
where x ∼ h−ir , independently of everything else. ψP0 0 is non-decreasing because it is linked to the minimal
mean square error mmse(r) of the channel (12) which is non-increasing with r:
h
2 i
mmse(r) = E X0 − E[X0 |Y0 ]
= E[X02 ] − 2ψP0 0 (r) .
It remains to show that ψP0 is strictly convex when P0 differs from a Dirac mass. When P0 admits a finite
moment of order 4, one can compute the second derivative and again, using Gaussian integration by parts
and the Nishimori identity one obtain:
2 i
1 h
ψP00 0 (r) = E hx2 ir − hxi2r
.
2
(N )
Without the 4th moment assumption, one can proceed by truncation. We consider now the distribution P0
(N )
of X0 = X0 I(−N ≤ X0 ≤ N ). h−ir,N will denote the corresponding posterior distribution. We apply
the previous computation to obtain
ψ 0 (N ) (t)
P0
−
ψ 0 (N ) (s)
P0
1
=
2
t
Z
h
hx2 ir,N − hxi2r,N
E
2 i
dr ,
(96)
s
for 0 < s < t. Now suppose that ψP0 0 is not strictly increasing. We can then find 0 < s < t such that
ψP0 0 (s) = ψP0 0 (t). The sequence of convex functions (ψP (N ) )N converges to ψP0 which is differentiable. A
0
standard analysis lemma gives then that the derivatives (ψ 0
(N )
P0
)N converges to ψP0 0 . Therefore, equation (96)
and Fatou’s Lemma lead to
Z t
Z t h
2
2 i
2
2
0≤
E lim inf hx ir,N − hxir,N
dr ≤ lim inf
E hx2 ir,N − hxi2r,N
dr = 0 .
s
N →∞
N →∞
s
It is not difficult to verify that hx2 ir,N − hxi2r,N −−−−→ hx2 ir − hxi2r almost surely. We conclude
N →∞
Z
t
h
E
hx2 ir − hxi2r
2 i
dr = 0 .
s
We can thus find r ∈ [s, t] such that hx2 ir − hxi2r = 0 almost surely. This means that h−ir is almost surely a
Dirac distribution. This is absurd because h−ir admits almost surely a strictly positive density w.r.t. P0 , and
P0 is assumed not to be a Dirac distribution. This concludes the proof.
A.2.2
The non-linear scalar channel
We prove here some properties of the free entropy of the second scalar channel (15).
Proposition 11. Let Pout be as in (3) where ϕ : R × RkA → R is a measurable function. Suppose that for
(Z, A) ∼ N (0, 1) ⊗ PA we have
√
E[ϕ( ρZ, A)2 ] < ∞ .
(97)
Let V, W ∗ i.i.d.
∼ N (0, 1) and Y be the output of the scalar channel (15). Then for all q ∈ [0, ρ],
Z
ΨPout (q) = E ln
w2
√
e− 2
√
dw √ Pout Y | q V + ρ − q w
2π
is well define and the function ΨPout is convex, non-decreasing on [0, ρ] and differentiable on [0, ρ). Moreover, if
ϕ is bounded and C 2 with respect to its first coordinate, with bounded first and second derivatives, then ΨPout is
Lipschtitz on [0, ρ].
30
Notice that (97) is for instance verified under hypotheses (h2).
Proof. We have for all x, y ∈ R
1
Pout (y|x) = √
2π∆
Z
1
1
2
exp −
(y − ϕ(x, a)) dPA (a) ≤ √
.
2∆
2π∆
Thus, using Jensen’s inequality
√
1
1
√
Ea,w (Y − ϕ( qV + ρ − qw))2
− ln(2π∆) −
2
2∆
w2
Z
√
1
e− 2
√
≤ ln dw √ Pout Y | q V + ρ − q w ≤ − ln(2π∆) ,
2
2π
(98)
where Ea,w denotes the expectation w.r.t. (w, a) ∼ N (0, 1) ⊗ PA . By assumptions, the left-hand side is integrable, so that ΨPout (q) is well defined for all q ∈ [0, ρ].
Let us assume now that ϕ is bounded and C 2 w.r.t. its first coordinate, with bounded first and second
derivatives. In that case, we will be able to use continuity and differentiation under the expectation, because
all the domination hypotheses will be easily verified. It is thus easy to check that ΨPout is continuous on [0, ρ].
We compute now the first derivative. Recall that h−isc , defined in (17), denotes the posterior distribution
of W ∗ given Y . We will use the notation uy (x) = ln Pout (y|x). For q ∈ (0, ρ) we have
E
√
√
1 D
√
√
Ψ0Pout (q) = E u0Y ( qV + ρ − qw)u0Y ( qV + ρ − qW ∗ )
2
sc
E2
√
1 D 0 √
= E uY ( qV + ρ − qw)
≥ 0,
2
sc
where w ∼ h−isc , independently of everything else. ΨPout is therefore non-decreasing. Using the boundedness assumption on ϕ, it is not difficult to check that Ψ0Pout is indeed bounded. ΨPout is thus Lipschtitz.
√
√
We will now compute Ψ00Pout . To lighten the notations, we write u0Y (w) for u0Y ( qV + ρ − qw). We
compute
D
E
h 1
D
E i
1
∂q E u0Y (w)u0Y (W ∗ )
=E
W ∗ u0Y (W ∗ ) u0Y (w)u0Y (W ∗ )
√ V − √
2 q
2 ρ−q
sc
sc
D 1
E
1
+2E
W ∗ u00Y (W ∗ )u0Y (w)
√ V − √
2 q
2 ρ−q
sc
D 1
E
1
+E
W ∗ u0Y (W ∗ )2 u0Y (w)
√ V − √
2 q
2 ρ−q
sc
D
E D 1
E
1
−E u0Y (W ∗ )u0Y (w)
w u0Y (w)
√ V − √
2 q
2 ρ−q
sc
sc
(A)
(B)
(C)
(D)
(99)
Notice that (A) = (C). We compute, using Gaussian integration by parts and the Nishimori identity (Pro-
31
position 9)
D
E i 1 h
D
E i
1 h
(A) = E u0Y (W ∗ ) u0Y (W ∗ )u00Y (w)
+ E u0Y (W ∗ ) u0Y (W ∗ )u0Y (w)2
2
2
sc
sc
D
E D
E i
1 h 0
− E uY (W ∗ ) u0Y (W ∗ )u0Y (w)
u0Y (w)
2
sc
D
E D
D
Esc
D
E
− E u00Y (W ∗ )u0Y (w)
+ E u00Y (W ∗ )u0Y (w)2
(B) = E u00Y (W ∗ )u00Y (w)
sc
sc
D 1
E
1
∗
0
∗ 0
(1) 0
(2)
√
V
−
W
u
(W
)u
(w
)u
(w
)
(D) = −E
√
Y
Y
Y
2 q
2 ρ−q
sc
sc
(100)
u0Y (w)
E
sc
(101)
= −Ehu0Y (W ∗ )u00Y (w(1) )u0Y (w(2) )isc − Ehu0Y (W ∗ )u0Y (w(1) )2 u0Y (w(2) )isc
+ Ehu0Y (W ∗ )u0Y (w(1) )u0Y (w(2) )isc hu0Y (w)isc
(102)
We now replace (100), (101) and (102) in (99):
D
E
D
E
D
E
− E u0Y (W ∗ )2 u0Y (w(1) )u0Y (w(2) )
+ E u0Y (W ∗ )2 u0Y (w)2
2Ψ00Pout (q) = E u0Y (W ∗ )2 u00Y (w)
E sc
E sc D
D
E sc D
00
∗ 0
(1) 0
(2)
00
∗ 0
2
00
∗ 00
− E uY (W )uY (w )uY (w )
+ E uY (W )uY (w)
+ E uY (W )uY (w)
−
sc
0
∗ 00
(1) 0
EhuY (W )uY (w )uY (w(2) )isc
Using the identity u00Y (x) + u0Y (x)2 =
Ψ00Pout (q)
00 (Y |x)
Pout
Pout (Y |x) ,
−
sc
0
∗ 0
EhuY (W )uY (w(1) )2 u0Y (w(2) )isc
+
sc
0
4
EhuY (w)isc .
this factorizes and gives
E 2 2 i
D
00 (Y |√q V + √ρ − qw) E
√
1 hD Pout
√
√
= E
≥ 0.
− u0Y ( q V + ρ − qw)
√
2
Pout (Y | q V + ρ − qw) sc
sc
(103)
ΨPout is thus convex on [0, ρ].
It remains now to prove that ΨPout is convex without the boundedness and C 2 assumptions on ϕ. Using
the same arguments than in Appendix D one see that we can approximate with arbitrary precision ΨPout with
a function ΨP̂out , where P̂out can be represented by a couple (ϕ̂, ∆) and ϕ̂ is bounded and C 2 w.r.t. its first
coordinate, with bounded first and second derivatives. This proves that ΨPout is convex. The fact that ΨPout
is differentiable on [0, ρ) follows from differentiation under the expectation sign. In order to see it, we define
√
√
X = q V + ρ − q W ∗ . Then, for all q ∈ [0, ρ):
Z
ΨPout (q) = E
1
−
e
dX p
2π(ρ − q)
√
(X− q V )2
2(ρ−q)
Z
Z
dY Pout (Y |X) ln
dx p
1
2π(ρ − q)
e
−
√
(x− q V )2
2(ρ−q)
Pout (Y |x) .
(104)
We are now in a good setting to differentiate under the expectation sign. The domination follows by the same
convexity argument than for (98).
A.3
Proof of Theorem 2
Define for ∆ > 0,
m
ln(2π∆)
fˆn (∆) = fn +
2n
Z
m
X
√
1
1
= E log dP0 (x)dPa (a) exp −
(ϕ([ΦX]µ , Aµ ) − ϕ([Φx]µ , aµ ) + ∆Zµ )2
n
2∆
µ=1
32
Lemma 3.
m
fˆn (∆) −−−→ ξn −
∆→0
2n
Proof. This follows from two applications of the dominated convergence Theorem. Define
Z
m
X
√
1
(ϕ([ΦX]µ , Aµ ) − ϕ([Φx]µ , aµ ) + ∆Zµ )2
h(X, A, Z, ∆) = log dP0 (x)dPa (a) exp −
2∆
µ=1
and
Z
g(X, A, Z) = log
dP0 (x)dPa (a)
m
Y
Z2
− 2µ
δ (ϕ([ΦX]µ , Aµ ) − ϕ([Φx]µ , aµ )) e
µ=1
By the dominated convergence Theorem (the domination is here obvious), for all Φ, X, A, Z
h(X, A, Z, ∆) −−−→ g(X, A, Z) .
∆→0
Now, for all Φ, X, A, Z, x, a we have
m
Y
δ(ϕ([ΦX]µ , Aµ ) − ϕ([Φx]µ , aµ )) e−
2
Zµ
2
m
1 X
√
(ϕ([ΦX]µ , Aµ )−ϕ([Φx]µ , aµ )+ ∆Zµ )2 ≤ 1
≤ exp −
2∆
µ=1
µ=1
Consequently, g(X, A, Z) ≤ h(X, A, Z, ∆) ≤ 0 . Thus h is lower and upper bounded by integrable random
variables (that do not depend on ∆). We can thus apply the dominated convergence Theorem to obtain the
Lemma.
(∆)
We will need to write explicitely the dependence of Pout in ∆ by writting: Pout . We also define
Ψ̂P (∆) (q) = ΨP (∆) (q) +
out
out
1
ln(2π∆)
2
Let (∆k )k≥0 be a sequence of positive numbers that converges to 0.
Lemma 4. The sequence Ψ̂ (∆k )
converges uniformly on [0, ρ] to Ψϕ − 21 .
Pout
k≥0
Proof. The pointwise convergence is proved using the same arguments as in the previous lemma. The uni
form convergence follows from Dini’s second Theorem, Ψ̂ (∆k ) k≥0 is indeed a sequence of non-decreasing
Pout
functions on [0, ρ] that converges pointwise to Ψϕ − 12 , which is continuous over [0, ρ] (as a pointwise limit
of convex functions, for instance).
The uniform convergence implies then the convergence of the sup − inf :
Corollary 4.
n
n
rq o
rq o α
sup inf ψP0 (r) + αΨP (∆) (q) −
−−−→ sup inf ψP0 (r) + αΨϕ (q) −
−
∆→0 q∈[0,ρ] r≥0
out
2
2
2
q∈[0,ρ] r≥0
We have now all the tools needed for the proof of Theorem 2.
Proof of Theorem 2. We extend by continuity (see Lemma 3) fˆn at ∆ = 0, to obtain a sequence of convex
33
functions on [0, 1]. Define for ∆ ∈ [0, 1]
o
n
supq∈[0,ρ] inf r≥0 ψP0 (r) + αΨ (∆) (q) − rq
Pout
n
o2
fˆ∞ (∆) =
rq
sup
− α2
q∈[0,ρ] inf r≥0 ψP0 (r) + αΨϕ (q) − 2
if ∆ > 0 ,
if ∆ = 0 .
By Theorem 1, fˆ∞ is the pointwise limit of the sequence of convex functions (fˆn )n≥1 on (0, 1], it is thus
continuous on (0, 1]. By Corollary 4 we obtain that fˆ∞ is continuous on [0, 1].
Moreover, fˆn is differentiable on (0, 1) and it is not difficult to check that its derivative is uniformly
bounded in n:
"
2 #
m
1
0
∀n ≥ 1, sup |fn (∆)| ≤
E ϕ √ [ΦX]1 , A1
n
n
∆∈(0,1)
which is bounded by assumption (h2). This implies that fˆn (0) → fˆ∞ (0), otherwise the derivative of fˆn would
explode near 0.
A.4
Proof of Proposition 1
In this section, we prove Proposition 1. In order to prove Proposition 1, we will need the following lemma.
For k ≥ 1 we denote by Sk+ the set of positive semi-definite k × k matrices and by Sk+∗ the set of positive
definite k × k matrices.
Lemma 5. Let (Cn )n≥1 be a sequence of random variables in Sk+ , that converges in law to C, such that C is
almost surely positive definite. Let f : Rk → R be a measurable function and Z ∼ N (0, Ik ) be independent
from (Cn ). Assume that the sequence (f (Cn Z))n≥1 is uniformly integrable. Then
E[f (Cn Z)] −−−→ E[f (CZ)]
n→∞
Proof. By the uniform integrability assumption, it suffices to show that (f (Cn Z))n converges to f (CZ) in
law. Let g : R → R be a continuous bounded function. Let EZ denote the expectation w.r.t. Z only. Then the
function
h : Sk+ → R
M 7→ EZ [g(f (MZ))]
is continuous over Sk+∗ . Indeed, for M ∈ Sk+∗ ,
1
h(M) =
k/2
(2π) |detM|k
Z
1 | −2
exp − x M x g(f (x))dx .
2
The discontinuity points of h are thus in Sk+ \ Sk+∗ , which has 0 measure for the law of C (because C is
a.s. positive definite). Consequently, by Portemanteau Theorem E[h(Cn )] → E[h(C)] which proves that
(f (Cn Z))n converges to f (CZ) in law.
We can now prove Proposition 1.
Proof. We define
1
Cn =
n
kX∗ k2 x · X∗
x · X∗ kxk2
!
34
P
−−−→
n→∞
ρ q
q ρ
!
∈ S2+∗
By the Nishimori property, one verify easily that
*
2
2 +
Φnew · X∗
Φnew · x
Φnew · x
1
Φnew · X∗
√
√
√
√
E ϕ
,A − ϕ
,a
,A − ϕ
,a
= E
ϕ
2
n
n
n
n
n
n
*
2 +
∗
Φnew · X
Φnew · x
1
√
√
,A − ϕ
,a
= E Ea,A ϕ
2
n
n
n
Now, we remark that
2
Φnew · X∗
Φnew · x
√
√
Ea,A ϕ
,A − ϕ
,a
n
n
is simply a measurable function that we denote f of
Φnew · X∗ Φnew · x
√
= C1/2
, √
n Z
n
n
in law, where Z ∼ N (0, I2 ) is independent from Cn . We can thus apply Lemma 5 (the uniform integrability
assumption follows from our hypothesis (h2)):
*
2 +
Φnew · X∗
Φnew · x
√
√
lim E Ea,A ϕ
,A − ϕ
,a
n→∞
n
n
n
h
2 i
√
√
√
√
1/2
= E[f (C Z)] = E ϕ( qZ1 + ρ − qZ2 , A) − ϕ( qZ1 + ρ − qZ20 , a)
for Z1 , Z2 , Z20 i.i.d.
∼ N (0, 1). By expanding the square of the right-hand side and applying Bayes rule, one
obtains exactly twice the right-hand side of (39). This concludes the proof.
A.5
Boundedness of an overlap fluctuation
In this appendix we show that the “overlap fluctuation”
m
m
D 1 X
2 E
D 1 X
2 E
0
0
2
E
uYt,µ (St,µ )uYt,µ (st,µ ) − rn
≤ 2rn + 2E
u0Yt,µ (St,µ )u0Yt,µ (st,µ )
n
n
t
t
µ=1
(105)
µ=1
is bounded uniformly in t under hypothesis (H2) on ϕ. From the representation (3) (recall we consider ∆ = 1
in Sec. 5)
uYt,µ (s) = ln Pout (Yt,µ |s)
Z
1
1
2
= ln dPA (aµ ) √ e− 2 (Yt,µ −ϕ(s,aµ ))
2π
(106)
and thus
u0Yt,µ (s)
R
=
1
2
dPA (aµ )(Yt,µ − ϕ(s, aµ ))ϕ0 (s, aµ )e− 2 (Yt,µ −ϕ(s,aµ ))
R
1
2
dPA (aµ )e− 2 (Yt,µ −ϕ(s,aµ ))
(107)
where ϕ0 is the derivative w.r.t. the first argument. From (1) at ∆ = 1 we get |Yt,µ | ≤ sup |ϕ| + |Zµ |, where
the supremum is taken over both arguments of ϕ, and thus immediately obtain for all s ∈ R
|u0Yt,µ (s)| ≤ (2 sup |ϕ| + |Zµ |) sup |ϕ0 | .
35
(108)
From (108) and (105) we see that it suffices to check that
2
m2
E (2 sup |ϕ| + |Zµ |)2 (sup |ϕ0 |)2
≤ C(ϕ)
2
n
where C(ϕ) is a constant depending only on ϕ. This is easily seen by expanding all squares and using that
m/n is bounded.
Appendix B: Proofs of the limits of optimal errors
B.1
Computation of the denoising error: Proof of Corollary 3
One can verify easily that in (defined by (21)) is a concave differentiable function of ∆−1 (see [40]). Thus its
limit i∞ is also a concave function of ∆−1 . Therefore, a standard analysis lemma gives that the derivative of
in w.r.t. ∆−1 converges to the derivative of i∞ at every point at which i∞ is differentiable (i.e. almost every
points, by concavity). Let us now compute these two derivatives. Using the I-MMSE Theorem from [40], one
obtain
m
1
D 1
E
1 X
∂in
∗
√
√
=
E
ϕ
[ΦX
]
,
A
−
ϕ
[Φx]
,
a
µ
µ
µ
µ
∂∆−1
2n
n
n
2
.
(109)
µ=1
Second, define
h(q, ∆) =
n
qr o
α
(1 + ln(2π∆)) + αΨPout (q) + inf ψP0 (r) −
.
r≥0
2
2
(110)
Then i∞ = − supq∈[0,ρ] h(q, ∆). The computation of ∂ΨPout (q)/∂∆−1 follows from the I-MMSE theorem
√
√
√
of [40]. Indeed, if we denote S = ϕ( qV + ρ − qW ∗ , A), notice that ΨPout (q) = −I(S; S + ∆Z | V ) −
√
√
√
1
2 log(2πe∆), because ΨPout (q) = −H(S + ∆Z | V ) and I(S + ∆Z; S | V ) = H(S + ∆Z | V ) − H(S +
√
∆Z | V, S) = −ΨPout (q) − 12 log(2πe∆). Therefore
√
√
∂ΨPout (q)
∆
∂
∆ 1
=
−
I(S
+
∆Z;
S
|
V
)
=
−
MMSE(S
+
∆Z | V, S)
∂∆−1
2
∂∆−1
2
2
p
p
∆ 1 √
=
−
E ϕ( ρ V, A)2 − E ϕ( q ∗ V + ρ − q ∗ w, a)i2sc .
2
2
Consequently,
p
p
∂h(q, ∆)
α √
2
∗V +
∗ w, a)i2
E
ϕ(
ρ
V,
A)
−
E
ϕ(
q
ρ
−
q
.
=
−
sc
∂∆−1
2
Now, Theorem 1 from [61] gives that at every ∆−1 at which i∞ is differentiable
p
p
∂i∞
∂
α √
2
∗V +
∗ w, a)i2
ρ
V,
A)
−
E
ϕ(
q
ρ
−
q
=
−
sup
h(q,
∆)
=
E
ϕ(
.
sc
∂∆−1
∂∆−1 q∈[0,ρ]
2
where q ∗ ∈ [0, ρ] is a point where the supremum above is achieved, and thus correspond to an optimal couple
in (20). As explained above, ∂in /∂∆−1 converges to ∂i∞ /∂∆−1 at every ∆−1 at which i∞ is differentiable,
which concludes the proof.
36
B.2
Optimal generalization error with side information: Proof of Proposition 2
Here the arguments are the same as for the proof of Corollary 3, but we repeat them briefly for the reader.
By a straightforward generalization of (22), fn is related to the mutual information I(X∗ , A; Y|Φ) between
the noisy labels and the classifier and stream A through
in :=
1
α
I(X∗ , A; Y|Φ) = −fn − − αβ ln(2π∆tr ) − α(1 − β) ln(2π∆te ) ,
n
2
(111)
Now in order to access the generalization error we employ the classical I-MMSE relation [40] for Gaussian
noise but restricted to the test set. It takes the following form in the present setting
(1 − β)m side
din
=
Egen .
te
−1
d(∆ )
2n
(112)
Again we use that in is a sequence of concave functions in (∆te )−1 and thus its limit (which exists by Theorem 1) is concave too. Again, these concavity properties allow to permute limits and differentiation. Therefore from (111) and (112) we find that for almost every ∆te > 0
lim E side
n→∞ gen
= ∆te −
2
d
sup inf f ts (q, r) .
α(1 − β) d(∆te )−1 q∈[0,ρ] r≥0 RS
(113)
It remains to evaluate this derivative using (46) to finally assess the asymptotic generalization error. This is
done as in the previous section using an envelope Theorem, see [61].
Appendix C: Proof of Proposition 4
0 (y|x) and P 00 (y|x) the first
Recall u0y (x) is the x-derivative of uy (x) = ln Pout (y|x). Moreover denote Pout
out
and second x-derivatives, respectively, of Pout (y|x). We will first prove that for all t ∈ (0, 1)
m
n
1 X
E
dfn (t)
1 D 1 X 0
rq(t) rρ An
=− E
uYµ (St,µ )u0Yµ (st,µ ) − r
Xi∗ xi − q(t)
+
−
−
,
dt
2
n
n
2
2
2
t
µ=1
(114)
i=1
where
m
n
1
i
h 1 X
00 (Y |S )
Pout
1 X
µ t,µ
√
((Xi∗ )2 − ρ)
ln Zt .
An := E √
n
n µ=1 Pout (Yµ |St,µ )
n
(115)
i=1
Once this is done, we will prove that An goes to 0 as n → ∞ uniformly in t ∈ [0, 1], in order to obtain
Proposition 4.
C.1
Proof of (114)
Recall definition (71) which becomes, when written as a function of the interpolating Hamiltonian (68),
Z
Z
1
0
0
∗
∗ −Ht (X∗ ,W∗ ;Y,Y 0 ,Φ,V)
fn (t) = EΦ,V dYdY dP0 (X )DW e
ln dP0 (x)Dw e−Ht (x,w;Y,Y ,Φ,V) .
n
(116)
37
We will need the Hamiltonian t-derivative Ht0 given by
Ht0 (X∗ , W∗ ; Y, Y0 , Φ, V)
=−
m
X
dSt,µ
µ=1
dt
u0Yµ (St,µ )
1
−
2
r
n
rX ∗ 0 √
Xi (Yi − t rXi∗ ) .
t
(117)
i=1
The derivative of the interpolating free entropy thus reads, for 0 < t < 1,
1
dfn (t)
1
= − E Ht0 (X∗ , W∗ ; Y, Y0 , Φ, V) ln Zt − E Ht0 (x, w; Y, Y0 , Φ, V) t
dt
{z
} |n
{z
}
|n
T1
(118)
T2
where recall the definition of Zt = Zt (Y, Y0 , Φ, V) given by (70).
Let us compute T1 . Let 1 ≤ µ ≤ m. Let us start with the following term
i
h dS
i 1 h
[ΦX∗ ]µ
ρ − q(t)
q(t)
t,µ 0
Vµ + qR
Wµ∗ u0Yµ (St,µ ) ln Zt .
E
+ qR
uYµ (St,µ ) ln Zt = E − p
t
t
dt
2
n(1 − t)
0 q(s)ds
0 (ρ − q(s))ds
(119)
Let us compute the first term of the right-hand side of the last identity. By Gaussian integration by parts w.r.t
Φµi we obtain
1
p
E [ΦX∗ ]µ u0Yµ (St,µ ) ln Zt
n(1 − t)
n
hZ
i
X
1
∗
∗
0
=p
E
dYdY0 e−Ht (X ,W ;Y,Y ,Φ,V) Φµi Xi∗ u0Yµ (St,µ ) ln Zt
n(1 − t) i=1
=
n
1 X ∗ 2 00
E (Xi ) uYµ (St,µ ) + u0Yµ (St,µ )2 ln Zt + E Xi∗ xi u0Yµ (St,µ )u0Yµ (st,µ ) t
n
i=1
n
n
i
D1 X
E
h1 X
P 00 (Yµ |St,µ )
(Xi∗ )2 out
ln Zt + E
Xi∗ xi u0Yµ (St,µ )u0Yµ (st,µ ) ,
=E
n
Pout (Yµ |St,µ )
n
t
i=1
(120)
i=1
where we used the identity
u00Yµ (x) + u0Yµ (x)2 =
00 (Y |x)
Pout
µ
.
Pout (Yµ |x)
(121)
We now compute the second term of the right hand side of (119). Using again Gaussian integrations by parts
but this time w.r.t Vµ , Wµ∗ i.i.d.
∼ N (0, 1) as well as the previous formula, we obtain similarly
h
i
q(t)
ρ − q(t)
E qR
Vµ + qR
Wµ∗ u0Yµ (Sµ,t ) ln Zt
t
t
0 q(s)ds
0 (ρ − q(s))ds
i
hZ
q(t)
ρ − q(t)
∗
∗
0
=E
dYdY0 e−Ht (X ,W ;Y,Y ,Φ,V) qR
Vµ + qR
Wµ∗ u0Yµ (Sµ,t ) ln Zt
t
t
0 q(s)ds
0 (ρ − q(s))ds
i
h P 00 (Y |S )
µ µ,t
ln Zt + E q(t)u0Yµ (Sµ,t )u0Yµ (sµ,t ) t .
(122)
= E ρ out
Pout (Yµ |Sµ,t )
38
Combining equations (119), (120) and (122) together, we have
h dS
i
t,µ 0
−E
uYµ (St,µ ) ln Zt
dt
n
n
h
i 1 D 1 X
E
00
1 P (Yµ |Sµ,t ) 1 X ∗ 2
= E out
(Xi ) − ρ ln Zt + E
Xi∗ xi − q(t) u0Yµ (St,µ )u0yµ (st,µ ) . (123)
2 Pout (Yµ |Sµ,t ) n
2
n
t
i=1
i=1
As seen from (118) it remains to compute, recalling that for 1 ≤ j ≤ n, Yj0 −
Gaussian integration by parts w.r.t Zj0 ∼ N (0, 1),
√
t rXj∗ = Zj0 and then again a
Z
h
i
∗ 0 √
∗ 0
0
∗
∗ 0
E Xj (Yj − t rXj ) ln Zt = E Xj Zj ln Zt = E Xj Zj ln dP0 (x)Dw e−Ht (x,w;Y,Y ,Φ,V)
Z
n
m
h
n X
√
2 oi
1X √
∗ 0
= E Xj Zj ln dP0 (x)Dw exp −
uYµ (st,µ ) +
trXi∗ + Zi0 − t r xi
2
µ=1
i=1
∗ √
∗
0
t r(Xj − xj ) + Zj t
= E Xj
√
(124)
= t r ρ − EhXj∗ xj it .
Thus, by taking the sum,
1
−
2
r
n
n
E
i rρ r D 1 X
r h1 X ∗ 0 √
E
Xi (Yi − t rXi∗ ) ln Zt =
− E
Xi∗ xi .
t n
2
2 n
t
i=1
(125)
i=1
Therefore, for all t ∈ (0, 1),
m
n
1
i rρ rq(t)
00 (Y |S )
1 h 1 X Pout
1 X
µ µ,t
∗ 2
√
√
T1 = E
((Xi ) − ρ)
ln Zt +
−
2
Pout (Yµ |Sµ,t )
n
2
2
n
n
µ=1
i=1
m
n
1 X
E
1 D 1 X 0
uYµ (St,µ )u0Yµ (st,µ ) − r
+ E
Xi∗ xi − q(t)
.
2
n
n
t
µ=1
(126)
i=1
To obtain (114), it remains to show that T2 = 0. This is a direct consequence of the Nishimori identity (see
Appendix A.1):
T2 =
1
E Ht0 (x, w; Y, Y0 , Φ)
n
t
=
1
E Ht (X∗ , W∗ ; Y, Y0 , Φ) = 0 .
n
(127)
For obtaining the Lemma, it remains to show that An goes to 0 uniformly in t ∈ [0, 1].
C.2
Proof that An vanishes as n → ∞
We now consider the final step, that is showing that An given by (115) vanishes in the n → ∞ limit uniformly
in t ∈ [0, 1] under conditions (H1)-(H2)-(H3). First we show that
m
n
h 1 X
i
00 (Y |S )
Pout
1 X
µ t,µ
√
fn (t)E √
((Xi∗ )2 − ρ) = 0 .
n µ=1 Pout (Yµ |St,µ )
n
(128)
i=1
Once this is done, we use the fact that n1 ln Zt concentrates around fn (t) to prove that An converges to 0
R 00
as n → ∞. We start by noticing the simple fact that for all s ∈ R, Pout
(y|s)dy = 0. Consequently, for
39
µ ∈ {1, . . . , m},
h P 00 (Y |S )
i Z
out µ t,µ
∗
00
E
X , St = dYµ Pout
(Yµ |St,µ ) = 0 .
Pout (Yµ |St,µ )
(129)
Thus, using the “tower property” of the conditionnal expectation:
n
m
n
m
ii
h X
X
h X
hX
00 (Y |S ) i
00 (Y |S )
Pout
Pout
µ t,µ
µ t,µ
X ∗ , St = 0
E
((Xi∗ )2 − ρ)
=E
((Xi∗ )2 − ρ) E
Pout (Yµ |St,µ )
Pout (Yµ |St,µ )
i=1
µ=1
µ=1
i=1
(130)
which implies (128). We now show that An goes to 0 uniformly in t ∈ [0, 1] as n → ∞. Using successively
(128) and the Cauchy-Schwarz inequality, we have
m
n
h 1 X
i
1
00 (Y |S )
Pout
1 X
µ t,µ
√
√
|An | = E
ln Zt − fn (t)
((Xi∗ )2 − ρ)
n
n µ=1 Pout (Yµ |St,µ )
n
i=1
m
n
h 1 X
2 i1/2 h 1
2 i1/2
00 (Y |S ) 2
Pout
1 X
µ t,µ
√
≤ E √
E
.
((Xi∗ )2 − ρ)
ln Zt − fn (t)
n
n µ=1 Pout (Yµ |St,µ )
n
(131)
i=1
Using again the “tower property” of conditional expectations
m
n
n
m
h X
2 i
h X
ii
2 h X
00 (Y |S ) 2 X
00 (Y |S ) 2
Pout
Pout
µ t,µ
µ t,µ
∗ 2
∗ 2
E
((Xi ) − ρ)
=E
X∗ , St .
((Xi ) − ρ) E
Pout (Yµ |St,µ )
Pout (Yµ |St,µ )
µ=1
i=1
µ=1
i=1
(132)
Now, using the fact that conditionally on St , the random variables
we have
00 (Y |S
Pout
µ t,µ )
Pout (Yµ |St,µ ) 1≤µ≤m
are i.i.d. and centered,
m
m
h X
i
h X
i
h P 00 (Y |S ) 2
i
00 (Y |S ) 2
00 (Y |S ) 2
Pout
Pout
µ t,µ
µ t,µ
out 1 t,1
X∗ , St = E
St = mE
St . (133)
E
Pout (Yµ |St,µ )
Pout (Yµ |St,µ )
Pout (Y1 |St,1 )
µ=1
µ=1
Under condition (H2), it is not difficult to show that there exists a constant C > 0 such that
h P 00 (Y |S ) 2
i
out 1 t,1
E
St ≤ C .
Pout (Y1 |St,1 )
(134)
Combining now (134), (133) and (132) we obtain that
m
n
n
h X
2 i
h X
2 i
00 (Y |S ) 2 X
Pout
µ t,µ
((Xi∗ )2 − ρ)
≤ mC E
((Xi∗ )2 − ρ)
= mnC Var (X1∗ )2 .
E
Pout (Yµ |St,µ )
µ=1
i=1
i=1
Going back to (131), therefore there exists a constant C 0 > 0 such that
2 i1/2
h 1
|An | ≤ C 0 E
ln Zt − fn (t)
.
n
(135)
By Theorem 4 we have E[(n−1 ln Zt − fn (t))2 ] → 0 as n → ∞ uniformly in t ∈ [0, 1]. Thus An goes to 0
uniformly in t ∈ [0, 1] as n → ∞. This ends the proof of Proposition 4.
40
Appendix D: Approximation
In this section, we suppose that Theorem 1 holds for channels of the form (1) under the hypotheses (H1), (H2)
and (H3).
We show in this section that this imply that Theorem 1 holds under the hypotheses (h1), (h2), and either
(h3) or (h3.a)-(h3.b)-(h3.c). Again, as explained at the beginning of Sec. 5, it suffices to consider the case where
∆ = 1. We start by relaxing the hypothesis (H1).
D.1
Relaxing the hypotheses on P0 and Φ
The goal of this section is to prove the following proposition:
Proposition 12. Suppose that (h1) holds and that ϕ is C ∞ with compact support. Suppose that (h3) or (h3.a)(h3.b)-(h3.c) hold. Then Theorem 1 holds.
Proposition 12 will follow from Lemma 6 and Lemma 7 below.
Lemma 6. Suppose that (h1) (H2) and (h3) hold. Then Theorem 1 holds.
Proof. The ideas are basically the same that in [55] (Sec. 6.2.2). We omit the details here for the sake of
brevity.
We now relax the Gaussian assumption on the “measurement matrix” Φ.
Lemma 7. Suppose that (h1) holds, and that ϕ is C ∞ with compact support. Suppose that (h3.a)-(h3.b) hold.
Then Theorem 1 holds.
Proof. The proof is based on the Lindeberg generalization theorem (Theorem 2 from [49]) which is a variant
of the generalized “Lindeberg principle” from [25].
Theorem 3 (Lindeberg generalization theorem). Let (Ui )1≤i≤n and (Vi )1≤i≤n be two collection of random
variables with independent components and f : Rn → R a C 3 function. Denote ai = |EUi − EVi | and bi =
|EUi2 − EVi2 |. Then
n
X
bi
ai E|∂i f (U1:i−1 , 0, Vi+1:n )| + E|∂i2 f (U1:i−1 , 0, Vi+1:n )|
|Ef (U) − Ef (V)| ≤
2
i=1
Z Ui
Z Vi
1
1
3
2
3
2
+ E
|∂ f (U1:i−1 , 0, Vi+1:n )|(Ui − s) ds + E |∂i f (U1:i−1 , 0, Vi+1:n )|(Vi − s) ds .
2 0 i
2 0
Let (Φ0µ,i ) i.i.d.
∼ N (0, 1) and let (Φµ,i ) be a family of independent random variables, with zero mean and
unit variance. Let fn0 be the free entropy (10) with design matrix Φ0 and fn be the free entropy (10) with
design matrix Φ.
We will apply Theorem 3 to the function
F : U ∈ Rm×n 7→
1
E ln
n
Z
− 12
dPA (a)dP0 (x) e
m
P
2
ϕ √1n [UX∗ ]µ ,Aµ −ϕ √1n [Ux]µ ,aµ +Zµ
µ=1
x,a
where the expectation E is taken w.r.t. X∗ , A and Z. We have
fn = EF (Φ)
and
41
fn0 = EF (Φ0 ) .
It is not difficult to verify that F is a C 3 function and that for all 1 ≤ µ ≤ m and 1 ≤ i ≤ n:
∂3F
3
∂Uµ,i
≤
∞
C
,
n5/2
for some constant C that only depends on ϕ and the first three moments of P0 . Thus, an application of
Theorem 3 gives
C
|fn − fn0 | ≤ √ .
n
By Proposition 6, we know that Theorem 1 holds for fn0 , thus it holds for fn .
D.2
Relaxing the hypotheses on ϕ
It remains to relax the hypotheses on ϕ. This section is dedicated to the proof of the following proposition,
which is of course exactly the statement of Theorem 1.
Proposition 13. Suppose that (h1) and (h2) hold. Suppose that (h3) or (h3.a)-(h3.b)-(h3.c) hold. Then, Theorem 1
holds for the output channel (1).
To prove Proposition 13 we will approximate the function ϕ with a function ϕ̂ which is C ∞ with compact
support. In the following, G is a standard Gaussian random variable, independent of everything else.
Proposition 14. Suppose that (h1) and (h2) hold. Suppose that (h3) or (h3.a)-(h3.b)-(h3.c) hold. Then, for all
> 0, there exist ϕ̂ ∈ C ∞ (R × RkA ) with compact support, such that
√
√
E (ϕ( ρG, A) − ϕ̂( ρG, A))2 ≤ ,
and for n large enough, we have
"
E
ϕ
1
√ [ΦX∗ ]1 , A1
n
− ϕ̂
1
√ [ΦX∗ ]1 , A1
n
2 #
≤ .
Proof. We will distinguish the cases where (h3) holds and where (h3.a)-(h3.b)-(h3.c) holds.
Case 1: The Gaussian case: (h3) holds.
Notice that [ΦX∗ ]1 = kX∗ kG in law. Thus, if |X1∗ | is a constant random variable, then ϕ is in L2 (R × RkA )
√
with the measure induced by ( ρG, A1 ). The result follows by density of the C ∞ functions with compact
support in L2 .
We consider now the case where |X1∗ | is not constant. We start with a useful lemma.
Lemma 8. Let f : R → R+ be a measurable function. Let G ∼ N (0, 1) and 0 < a ≤ b. Then
√
r
E[f ( aG)] ≤
√
b
E[f ( bG)].
a
√
√
In particular, if f ( bG) is integrable then f ( aG) is also integrable.
R
√
√
2
Proof. For x ∈ [a, b], notice that: xE[f ( xG)] = (2π)−1/2 e−g /(2x) f (g)dg which is clearly non-decreasing
in x.
42
|X1∗ | is not constant and ρ = E[(X1∗ )2 ], therefore there exists ρ < r < r0 such that P(r ≤ (X1∗ )2 ≤ r0 ) > 0.
Consequently, using Lemma 8,
√
√
P(r ≤ (X1∗ )2 ≤ r0 )E[ϕ( rG, A1 )2 ] = E[1r≤(X1∗ )2 ≤r0 ϕ( rG, A1 )2 ]
|X ∗ |
≤ E[1r≤(X1∗ )2 ≤r0 √1 ϕ(|X1∗ |G, A1 )2 ] ≤
r
r
r0
E[ϕ(|X1∗ |G, A1 )2 ] < ∞.
r
√
Therefore E[ϕ( rG, A1 )2 ] < ∞.
√
Let > 0. We have just proved that ϕ ∈ L2 (R × RkA ) with the measure induced by ( rG, A1 ). There
√
√
exists a C ∞ function with compact support ϕ̂ such that E (ϕ( rG, A) − ϕ̂( rG, A))2 ≤ . Thus by
Lemma 8
r
r
√
√
r √
r
√
2
2
E (ϕ( rZ, A) − ϕ̂( rZ, A)) ≤
.
E (ϕ( ρZ, A) − ϕ̂( ρZ, A)) ≤
ρ
ρ
2
1
1
√
√
. By the law of large numbers, n1 kXk2
It remains to bound E ϕ
kXkG, A1 − ϕ̂
kXkG, A1
n
n
converges to ρ in probability as n → ∞. Thus P( n1 kXk2 ∈
/ [ρ/2, r]) −−−→ 0. We now apply Hölder’s
n→∞
inequality:
"
2 #
1
1
E 1kXk2 /n∈[ρ/2,r]
ϕ √ kXkG, A1 − ϕ̂ √ kXkG, A1
/
n
n
2
"
2+γ # 2+γ
γ
1
1
≤ P(kXk2 /n ∈
/ [ρ/2, r]) 2+γ E ϕ √ kXkG, A1 − ϕ̂ √ kXkG, A1
n
n
γ
≤ C 2+γ ,
for some constant C > 0 and for n large enough. It remain to bound
"
E 1kXk2 /n∈[ρ/2,r] ϕ
1
√ kXkG, A1
n
!2 #
1
− ϕ̂ √ kXkG, A1
n
r
2
√
√
nr
≤ E 1kXk2 /n∈[ρ/2,r]
ϕ rG, A1 − ϕ̂ rG, A1
kXk2
r
2 i
√
√
2r h
≤
E ϕ rG, A1 − ϕ̂ rG, A1
ρ
r
2r
.
≤
ρ
Case 2: The non-Gaussian case: (h3.a)-(h3.b)-(h3.c) hold.
In that case, by the Central Limit Theorem (using the fact that the third moments of (Xi∗ Φ1,i ) are bounded
with n, because of hypotheses (h3.a) and (h3.b))
[ΦX∗ ]1
√
(d)
√
, A1 −−−→ ( ρG, A1 ) .
(136)
n→∞
n
This implies that
ϕ
[ΦX∗ ]1
√
, A1
n
√
(d)
−−−→ ϕ( ρG, A1 ) .
n→∞
(137)
Indeed, let g : R → R be a continuous bounded function. We have, by Portemanteau Theorem, that for almost
43
all a ∈ RkA (with respect to PA )
[ΦX∗ ]1
√
√
,a
−−−→ E [g ◦ ϕ( ρG, a)] .
E g◦ϕ
n→∞
n
because for almost all a, g◦ϕ(·, a) is continuous almost-everywhere and bounded. By dominated convergence,
it follows that
[ΦX∗ ]1
√
√
E g◦ϕ
−−−→ E [g ◦ ϕ( ρG, A1 )] .
, A1
n→∞
n
∗]
√ 1 , A1 )n is by assumption (h2) bounded in L2 , thus
That finishes the proof of (137). The sequence (ϕ [ΦX
n
√
by (137) we have that E[ϕ( ρG, A1 )2 ] < ∞.
√
Let > 0. We have just proved that ϕ ∈ L2 (R × RkA ) with the measure induced by ( ρG, A1 ). There
√
√
exists a C ∞ function with compact support ϕ̂ such that E (ϕ( ρG, A) − ϕ̂( ρG, A))2 ≤ .
One deduce from (136) and (137) that
2
1
1
√
√
(d)
∗
∗
−−−→ (ϕ( ρG, A) − ϕ̂( ρG, A))2 .
ϕ √ [ΦX ]1 , A1 − ϕ̂ √ [ΦX ]1 , A1
n→∞
n
n
Now, hypothesis (h2) gives that the sequence above is uniformly integrable. This gives that
2
1
1
√
√
E ϕ √ [ΦX∗ ]1 , A1 − ϕ̂ √ [ΦX∗ ]1 , A1
−−−→ E(ϕ( ρG, A) − ϕ̂( ρG, A))2 ≤ .
n→∞
n
n
Consequently, the left-hand side is smaller that 2 for n large enough. This concludes the proof.
In the remaining of this section, we prove Proposition 13. Let > 0. Let ϕ and ϕ̂ as in Proposition 14. Let
fn be the free entropy associated to ϕ and fˆn the free entropy corresponding to ϕ̂.
Lemma 9. There exists a constant C > 0 such that for n large enough
√
|fn − fˆn | ≤ C .
Proof. Consider the observation channel given by
√ 1
Yt,µ =
tϕ √n [ΦX∗ ]µ , Aµ + Zµ
√
Ŷt,µ =
1 − tϕ̂ √1n [ΦX∗ ]µ , Aµ + Ẑµ
for 1 ≤ µ ≤ m. Let fn (t) denote the interpolating free energy:
1
fn (t) = E ln
n
Z
− 12
dPA (a)dP0 (x) e
m
P
√
Yt,µ − tϕ
µ=1
√1 [Φx]µ ,aµ
n
2
2
√
+ Ŷt,µ − 1−tϕ̂ √1n [Φx]µ ,aµ
.
x,a
(x,a)
In order to shorten the notations, we will now write ϕµ
We compute, for 0 < t < 1:
fn0 (t)
i
m h (X∗ ,A) 2
(X∗ ,A) 2
E (ϕ̂1
) − (ϕ1
) +
=
2n
i
m h (X∗ ,A) 2
(X∗ ,A) 2
=
E (ϕ̂1
) − (ϕ1
) +
2n
(x,a)
and ϕ̂µ
for ϕ √1n [Φx]µ , aµ and ϕ̂ √1n [Φx]µ , aµ .
m
E
1 X D (x,a) (X∗ ,A)
∗ ,A)
− ϕ̂(x,a)
ϕ̂(X
E ϕµ ϕµ
µ
µ
2n
t
µ=1
D
E
m
(x,a) (X∗ ,A)
(X∗ ,A)
(x,a)
(x,a) (X∗ ,A)
E ϕ1 (ϕ1
− ϕ̂1
) + (ϕ1
− ϕ̂1 )ϕ̂1
.
2n
t
44
We start by controlling the first term:
h
i
1/2
(X∗ ,A) 2
(X∗ ,A) 2
(X∗ ,A)
(X∗ ,A) 2
(X∗ ,A)
(X∗ ,A) 2
≤ E ϕ̂1
E (ϕ̂1
) − (ϕ1
)
+ ϕ1
E ϕ̂1
− ϕ1
√
≤ C0 ,
by Proposition 14, for some constant C0 and n large enough. The two other terms can be bounded the same
way:
D
E
(x,a) (X∗ ,A)
ϕ1 (ϕ1
−
(X∗ ,A)
ϕ̂1
)
1/2
(X∗ ,A) 2
(X∗ ,A)
(X∗ ,A) 2
≤ E ϕ1
E ϕ̂1
− ϕ1
E
t
√
≤ C0 ,
by Proposition 14, for n large enough. Consequently, there exists a constant C > 0 such that for n large
√
enough and for all 0 < t < 1, |fn0 (t)| ≤ C . Notice that t 7→ fn (t) is continuous over [0, 1], fn (0) = fˆn
R1
√
and fn (1) = fn , hence |fn − fˆn | ≤ 0 |fn0 (t)|dt ≤ C .
Let Pout denote the transition kernel associated to ϕ and P̂out the one associated to ϕ̂. Analogously to the
previous Lemma, one can show:
Lemma 10. There exists a constant C 0 > 0 such that for all q ∈ [0, ρ]
√
|ΨPout (q) − ΨP̂out (q)| ≤ C 0 .
From there we obtain that
√
sup inf fRS (q, r) − sup inf fˆRS (q, r) ≤ C 0 .
q∈[0,ρ] r≥0
q∈[0,ρ] r≥0
(138)
Applying Theorem 1 for Pout , we obtain that for n large enough |fn − supq∈[0,ρ] inf r≥0 fRS (q, r)| ≤ . We
now combine this with (138) and Lemma 9 we obtain that for n large enough
√
√
fn − sup inf fRS (q, r) ≤ fˆn − sup inf fˆRS (q, r) + (C + C 0 ) ≤ (C + C 0 ) + ,
q∈[0,ρ] r≥0
r≥0 q∈[0,ρ]
which concludes the proof of Proposition 13.
Appendix E: A sup-inf formula
This appendix gathers some useful lemmas for the manipulations of “sup-inf” formulas like (20). We first need
to indroduce some notations about convex functions. Let f be a convex function on some interval I ⊂ R. For
t ∈ I we will denote respectlively by f 0 (t− ) and f 0 (t+ ) the left- and right-hand derivatives of f at t. We also
define the subgradient of f at t as ∂f (t) = [f 0 (t− ), f 0 (t+ )]. If t is the right (resp. left) border of I, then we
define ∂f (t) = [f 0 (t− ), +∞) (resp. ∂f (t) = (−∞, f 0 (t+ )]).
Lemma 11. Let f and g be two convex, Lipschitz, non-decreasing functions on R+ . For q1 , q2 ∈ R+ we define
ψ(q1 , q2 ) = f (q1 ) + g(q2 ) − q1 q2 . Then
sup inf ψ(q1 , q2 ) =
q1 ≥0 q2 ≥0
sup
ψ(q1 , q2 ) =
q1 ∈∂g(q2 )
q2 =f 0 (q1+ )
45
sup
q1 ∈∂g(q2 )
q2 ∈∂f (q1 )
ψ(q1 , q2 ) ,
(139)
and these extremas are achieved at some (possibly many) couples. Moreover, if g is strictly convex, then the above
extremas are achieved precisely on the same couples (q1 , q2 ) and f is differentiable at q1 .
Corollary 5. Let f be a convex, Lipschitz, non-decreasing function on R+ . Define ρ = supx≥0 f 0 (x+ ). Let
g : [0, ρ] → R be a convex, Lipschitz, non-decreasing function. For q1 ∈ R+ and q2 ∈ [0, ρ] we define ψ(q1 , q2 ) =
f (q1 ) + g(q2 ) − q1 q2 . Then
sup
inf ψ(q1 , q2 ) = sup
inf ψ(q1 , q2 ) .
q2 ∈[0,ρ] q1 ≥0
q1 ≥0 q2 ∈[0,ρ]
Proof. In order to apply Lemma 11 we need to extend g on R+ . We thus define for x ≥ 0
g̃(x) =
(
g(x)
if x ≤ ρ
g(ρ) + (x − ρ)g 0 (ρ+ ) if x ≥ ρ
g̃ is simply equal to g that we extend for x ≥ ρ using its tangent at ρ. Obviously g̃ is a convex, Lipschitz,
non-decreasing function on R+ . One can thus apply Lemma 11:
sup inf f (q1 ) + g̃(q2 ) − q1 q2 = sup inf f (q1 ) + g̃(q2 ) − q1 q2
q1 ≥0 q2 ≥0
q2 ≥0 q1 ≥0
and both “sup-inf” are achieved on [0, supx≥0 g̃ 0 (x+ )]×[0, supx≥0 f 0 (x+ )]. By definition, ρ = supx≥0 f 0 (x+ ),
thus
sup
inf f (q1 ) + g̃(q2 ) − q1 q2 = sup inf f (q1 ) + g̃(q2 ) − q1 q2
q1 ≥0 q2 ≥0
q1 ≥0 q2 ∈[0,ρ]
= sup inf f (q1 ) + g̃(q2 ) − q1 q2 = sup
q2 ≥0 q1 ≥0
inf f (q1 ) + g̃(q2 ) − q1 q2
q2 ∈[0,ρ] q1 ≥0
which concludes the proof because g̃(q2 ) = g(q2 ) for q2 ∈ [0, ρ].
To prove Lemma 11 we will need the following lemma on the Fenchel-Legendre transform.
Lemma 12. Let V ⊂ R be a non-empty interval and let g : V → R be a convex function. Define
g ∗ : x ∈ R 7→ sup xy − g(y) ∈ R ∪ {+∞} .
(140)
y∈V
Let Dg∗ = {x ∈ R | g ∗ (x) < ∞}. Then g ∗ is a convex function on the interval Dg∗ 6= ∅. Moreover, Dg∗ = {a}
if and only if g : x 7→ ax. For all x ∈ Dg∗ the set of maximizers of (140) is of the form [ax , bx ], where
ax , bx ∈ R ∪ {±∞}. Then, the left-hand and right-hand derivatives of g ∗ at x are respectlively ax and bx .
In particular, if g is strictly convex then g ∗ is differentiable around every point in the interior of Dg∗ .
Proof. g is convex, so there exists an affine function h : x 7→ ax + b such that ∀x ∈ V, g(x) ≥ h(x). Compute
g ∗ (a) = sup ay − g(y) ≤ sup ay − h(y) = b < ∞ .
y∈V
y∈V
Thus Dg∗ 6= ∅. It is easy to verify that Dg∗ is convex. g ∗ is then convex as a supremum of linear functions.
For x ∈ R we define the function ϕx : y 7→ xy − g(y). Let x ∈ Dg∗ . Because of convexity, g ∗ is left- and
right-hand differentiable at x. We are only going to show that the left-hand derivative of g ∗ at x is equal to
ax , the result for the right-hand derivative is proved analogously.
First of all, notice that if ax = −∞ then g ∗ (x0 ) = +∞ for any x0 < x, so the left-hand derivative of g ∗ is
not defined. We will thus concentrate on the cases where ax ∈ R or ax = +∞.
46
Let us consider first the case where ax is finite. By definition, g ∗ (x) = ϕx (ax ). Let now x0 < x. We have
ax (x − x0 ) = ϕx (ax ) − ϕx0 (ax ) ≥ g ∗ (x) − g ∗ (x0 ) ,
which implies that (g ∗ )0 (x− ) ≤ ax . Let now x0 < x in Dg∗ . Notice that, by convexity of g, bx0 ≤ ax . The
fact that ax achieves the maximum in (140) implies that x ∈ ∂g(ax ). We have then, by convexity g(bx0 ) ≥
g(ax ) + (bx0 − ax )x which implies
g ∗ (x) − g ∗ (x0 ) = ax x − g(ax ) − bx0 x0 + g(bx0 )
≥ ax − g(a) − bx0 x0 + g(a) + (bx0 − ax )x
≥ bx (x − x0 ) .
When x0 → x, it is not difficult to verify that bx0 → ax . We obtain therefore that (g ∗ )0 (x− ) ≤ ax . We conclude
(g ∗ )0 (x− ) = ax .
Suppose now that ax = +∞. Let x0 < x. Notice that ax0 is finite, because the converse would imply
that g ∗ (x) = +∞. Using the result we just proved, we have ax0 = (g ∗ )0 (x0− ) ≤ (g ∗ )0 (x− ), where the last
inequality is due to convexity. When x0 → x it is again not difficult to verify that ax0 → ax = +∞. We
conclude (g ∗ )0 (x− ) = +∞.
Proof of Lemma 11. By convexity of g, for all q1 , q2 ≥ 0 such that q1 ∈ ∂g(q2 ), we have ψ(q1 , q2 ) = inf q20 ≥0 ψ(q1 , q20 ).
Consequently:
sup inf ψ(q1 , q2 ) ≥
q1 ≥0 q2 ≥0
ψ(q1 , q2 ) ≥
sup
q1 ∈∂g(q2 )
q2 ∈∂f (q1 )
sup
ψ(q1 , q2 ) .
q1 ∈∂g(q2 )
q2 =f 0 (q1+ )
Let us now prove the converse bound. Let Lg = supy≥0 g 0 (y + ). Lg is finite because g is Lipschitz. Notice
that for 0 ≤ y < Lg , g ∗ (y) < ∞ while for y > Lg , g ∗ (y) = +∞. f is continuous on [0, Lg ] and g ∗ is convex
on [0, Lg ] (g ∗ (Lg ) may be equal to +∞). Therefore f − g ∗ achieves its suppremum at some q1∗ ∈ [0, Lg ].
If 0 < q1∗ < Lg then the optimality condition at q1∗ gives f 0 (q1∗− ) − g ∗0 (q1∗− ) ≥ 0 and f 0 (q1∗+ ) − g ∗0 (q1∗+ ) ≤ 0.
Thus
(141)
g ∗0 (q1∗− ) ≤ f 0 (q1∗− ) ≤ f 0 (q1∗+ ) ≤ g ∗0 (q1∗+ ) .
We know by Lemma 12 that [g ∗0 (q1∗− ), g ∗0 (q1∗+ )] is the set of maximizers of q2 7→ q1∗ q2 − g(q2 ). Consequently
q2∗ := f 0 (q1∗+ ) maximizes q2 7→ q1∗ q2 − g(q2 ). This gives q1∗ ∈ ∂g(q2∗ ). We conclude:
sup inf ψ(q1 , q2 ) = f (q1∗ ) + g(q2∗ ) − q1∗ q2∗ ≤
q1 ≥0 q2 ≥0
sup
f (q1 ) + g(q2 ) − q1 q2 .
q1 ∈∂g(q2 )
q2 =f 0 (q1+ )
If now q1∗ = 0, the optimality condition is now: f 0 (0+ ) − g ∗0 (0+ ) ≤ 0. f is non-decreasing, hence 0 ≤
f 0 (0+ ) ≤ g ∗0 (0+ ). g is non-decreasing, thus the set of maxizers of g ∗ (0) is, by Lemma 12 [0, g ∗0 (0+ )]. Therefore q2∗ := f (0+ ) maximizes q2 7→ q1∗ q2 − g(q2 ) and we conclude similarly as before.
If q1∗ = Lg , then the optimality condition is f 0 (q1∗− ) − g ∗0 (q1∗− ) ≥ 0. Therefore f 0 (q1+ ) ≥ f 0 (q1− ) ≥ g ∗0 (q1∗− ).
Again by Lemma 12, q2∗ := f (q1∗+ ) maximizes q2 7→ q1∗ q2 − g(q2 ), we conclude similarly as before.
The equalities (139) are now proved. Moreover, we have seen in the proof above that there exists a couple
(q1∗ , q2∗ ) that achieves all three extremas.
Suppose now that g is strictly convex. This means (by Lemma 12) that g ∗ is differentiable.
47
Let (q1 , q2 ) ∈ R2+ be a couple that achieves the supremum in one of the last two sup. This means in particular that q1 ∈ ∂g(q2 ). By convexity of g we have then ψ(q1 , q2 ) = inf q20 ≥0 ψ(q1 , q20 ). Since ψ(q1 , q2 ) =
supq10 ≥0 inf q20 ≥0 ψ(q10 , q20 ), (q1 , q2 ) achieves the sup-inf.
Let (q1 , q2 ) be a couple that achieves the sup-inf. It remains to show that q1 ∈ ∂g(q2 ) and q2 = f10 (q1+ ).
First of all, the fact that ψ(q1 , q2 ) = inf q20 ≥0 ψ(q1 , q20 ) implies (by convexity) that q1 ∈ ∂g(q2 ). Now, as in the
proof above we have to distiguish whenever q1 ∈ (0, Lg ), q1 = 0 or q1 = Lg . When q1 ∈ (0, Lg ), the strict
convexity of g implies the differentiability of g ∗ , therefore the inequality (141) gives that f is differentiable
at q1 and f 0 (q1 ) = g ∗0 (q1 ) = q2 . The case q1 = 0 goes the same way. It remains to see that q1 could not be
equal to Lg . Indeed, by strict convexity, the infimum of q20 7→ ψ(Lg , q20 ) is only achieved when q20 → +∞.
∗
Lemma 12 gives then that g ∗ (L−
g ) = +∞, which gives that the function q1 7→ f (q1 ) − g (q1 ) is decreasing
on [Lg − , Lg ], for some > 0. The supremum can therefore not be achieved at q1 = Lg .
Lemma 13. Let g be a strictly convex, differentiable, Lipschitz non-decreasing function on R+ . Define ρ =
supx≥0 g 0 (x). Let f be a convex, non-decreasing function on [0, ρ], differentiable on [0, ρ). For q1 , q2 ∈ R+ we
define ψ(q1 , q2 ) = f (q1 ) + g(q2 ) − q1 q2 . Then
sup
inf ψ(q1 , q2 ) =
q1 ∈[0,ρ] q2 ≥0
where
sup
(
Γ=
(142)
ψ(q1 , q2 ) ,
(q1 ,q2 )∈Γ
(q1 , q2 ) ∈ [0, ρ] × (R+ ∪ {+∞})
q1 = g 0 (q2 )
q2 = f 0 (q1 )
)
,
where all the function are extended by there limits at the points at which they may not be defined (for instance
g 0 (+∞) = lim g 0 (q)). Moreover, the above extremas are achieved precisely on the same couples.
q→∞
Proof. The proof goes almost exactly as for Lemma 11. The only difference with Lemma 11 and Corollary 5
above is that, without the Lipschitz assumption on f , the extremas in (142) can be achieved for (q1 , q2 ) =
(ρ, +∞) ∈ Γ, when f 0 (ρ− ) = +∞. It remains only to deal with this case, because the others can be handled
exactly as in the proof of Lemma 11.
First, suppose that q1 = ρ maximizes q1 7→ inf q2 ≥0 ψ(q1 , q2 ) over [0, ρ]. In that case, the infimum in q2 is
achieved for q2 → +∞, because by assumption q2 7→ g(q2 ) − ρq2 is strictly decreasing. Thus in that case we
still have
sup inf ψ(q1 , q2 ) = ψ(ρ, +∞) ≤ sup ψ(q1 , q2 ) .
q1 ∈[0,ρ] q2 ≥0
(q1 ,q2 )∈Γ
It remains now to show that ψ(ρ, +∞) ≤ supq1 inf q2 ψ(q1 , q2 ). Let q2 ≥ 0 and define q1 = g 0 (q2 ). By
convexity
ψ(q1 , q2 ) ≤ sup inf
ψ(q10 , q20 ) .
0
q10 ∈[0,ρ] q2 ≥0
Now notice that f (q1 ) + g(q2 ) − ρq2 ≤ ψ(q1 , q2 ) and that the left-hand side of this last inequality goes to
ψ(ρ, +∞) as q2 → ∞. This concludes the proof.
Appendix F: Concentration of free entropy and overlaps
F.1
Concentration of the free entropy
The goal of this Appendix is to prove that the free entropy of the interpolating model studied in Sec. 5.1
concentrates around its expectation. To simplify the notations we use C(ϕ, S, α), for a generic strictly positive
48
constant depending only on ϕ, S and α (S the supremum over signal values). It is also understood that n and
m are large enough and m/n → α.
Theorem 4. Under assumptions (H1), (H2) and (H3) we can find a constant C(ϕ, S, α) > 0 such that
"
2 #
C(ϕ, S, α)
1
1
≤
E
ln Zt − E
ln Zt
.
n
n
n
(143)
The remaining of this Appendix is dedicated to the proof of Theorem 4. We first recall some set-up and
notation for the convenience of the reader. The interpolating Hamiltonian (68)-(67) is (we indicate only the
annealed variables in its arguments)
−
m
X
µ=1
n
1X 0 √
(Yi − trxi )2
ln Pout (Yµ |st,µ (x, wµ )) +
2
(144)
i=1
where
r
st,µ (x, wµ ) =
1−t
[Φx]µ + k1 (t)Vµ + k2 (t)wµ ,
n
sZ
sZ
t
q(v)dv,
k1 (t) =
0
t
(ρ − q(v))dv
k2 (t) =
0
We find it convenient to use the random function representation (1) for the interpolating model, namely
q
1−t
Y = ϕ
∗ ] + k (t)V + k (t)W ∗ , A
[ΦX
+ Zµ ,
t,µ
µ
1
µ
2
µ
µ
n
√
Y 0 = rtX ∗ + Z 0
t,i
i
i
i.i.d.
In this representation the random variables (Aµ )1≤µ≤m i.i.d.
∼ PA are arbitrary, and (Zµ )1≤µ≤m ∼ N (0, 1). We
have
Z
1
2
1
Pout (Yt,µ |st,µ (x, w)) = dPA (aµ ) √ exp − Yt,µ − ϕ(st,µ (x, wµ ), aµ
2
2π
Z
1
1
2
(145)
= dPA (aµ ) √ exp − (Γt,µ (x, wµ , aµ ) + Zµ )
2
2π
where, using the random function representation,
Γt,µ (x, wµ , aµ )
(146)
r 1 − t
r 1 − t
=ϕ
[ΦX∗ ]µ + k1 (t)Vµ + k2 (t)Wµ∗ , Aµ − ϕ
[Φx]µ + k1 (t)Vµ + k2 (t)wµ , aµ .
n
n
From (144), (145), (146) we can express the free entropy of the interpolating model as
Z
1 n
m
1
ln Zt = ln
dP0 (x)dPA (a)Dw e−Ht (x,w,a) −
ln(2π)
n
n
2n
(147)
where Dw denote the standard m-dimensional Gaussian measure and where the Hamiltonian Ht is
Ht (x, w, a) =
m
n
µ=1
i=1
√
1X
1X √
Γt,µ (x, wµ , aµ )2 +
( rtXi∗ + Zi0 − trxi )2 .
2
2
(148)
The interpretation here is that x, w, a are annealed variables and Φ, V, A, Z, Z0 , X∗ , W∗ are quenched. The
inference problem is to recover X∗ , W∗ given all other quenched variables.
49
Our goal is to show that the free energy (147) concentrates around its expectation. We will first show concentration w.r.t. all Gaussian variables Φ, V, Z, Z0 , W∗ thanks to the classical Gaussian Poincaré inequality,
then the concentration w.r.t. A and finally the one w.r.t. X∗ using classical bounded differences arguments.
The order in which we prove the concentrations matters. We recall here these two variances bounds. The
reader can refer to [20] (Chapter 3) for detailled proofs of these statements.
Proposition 15 (Gaussian Poncaré inequality). Let U = (U1 , . . . , UN ) be a vector of N independent standard
normal random variables. Let g : RN → R be a continuously differentiable function. Then
(149)
Var(g(U)) ≤ E k∇g(U)k2 .
Proposition 16. Let U ⊂ R. Let g : U N → R a function that satisfies the bounded difference property, i.e.,
there exists some constants c1 , . . . , cN ≥ 0 such that
sup
u1 ,...uN ∈U N
u0i ∈U
|g(u1 , . . . , ui , . . . , uN ) − g(u1 , . . . , u0i , . . . , uN )| ≤ ci , for all 1 ≤ i ≤ N .
Let U = (U1 , . . . , UN ) be a vector of N independent random variables that takes values in U. Then
N
1X 2
Var(g(U)) ≤
ci .
4
(150)
i=1
Before we proceed it is useful to remark
m
n
µ=1
i=1
1
1
1 X 2
1 X 02 m
ln Zt = ln Ẑt −
Zµ −
ln(2π)
Zi −
n
n
2n
2n
2n
(151)
where
1
1
ln Ẑt = ln
n
n
Z
dP0 (x)dPA (a)Dw e−Ĥt (x,w,a)
(152)
and
m
Ĥt (x, w, a) =
o
1 Xn
Γt,µ (x, wµ , aµ )2 + 2Zµ Γt,µ (x, wµ , aµ )
2
µ=1
n
o
√
√
√
1 Xn √
+
( rtXi∗ − trxi )2 + 2Zi0 ( rtXi∗ − trxi ) .
2
(153)
i=1
In order
Theorem 4 it remains to show that there exists a constant C(ϕ, S, α) > 0 such that
to prove
C(ϕ,S,α)
1
. This concentration property together with (151) implies (143).
Var n ln Ẑt ≤
n
F.1.1
Concentration with respect to Gaussian random variables Zµ , Zi0 , Vµ , Wµ∗ , Φµi
In this section, we prove that n1 ln Ẑt is close to its expectation w.r.t. the Gaussian random variables Z, Z0 , V, W∗ , Φ.
Lemma 14. Let EG denotes the expectation w.r.t. Z, Z0 , V, W∗ , Φ only. There exists an constant C(ϕ, S, α) > 0
such that
"
2 #
1
1
C(ϕ, S, α)
E
ln Ẑt − EG
ln Ẑt
≤
.
(154)
n
n
n
50
Lemma 14 follows, by Pythagorean theorem, from Lemmas 15 and 16 below.
Lemma 15. Let EZ,Z0 denotes the expectation w.r.t. Z, Z0 only. There exists an constant C(ϕ, S, α) > 0 such
that
"
2 #
1
C(ϕ, S, α)
1
≤
E
ln Ẑt − EZ,Z0
ln Ẑt
.
(155)
n
n
n
Proof. We consider here g =
random variables. We have
1
n
ln Ẑt only as a function of Z and Z0 and work conditionnaly to all other
m
X
∂g
k∇gk =
∂Zµ
2
µ=1
2
n
X
∂g
+
∂Zi0
2
.
(156)
i=1
Each of these partial derivatives are of the form ∂g = n−1 h∂ Ĥt iĤt where the Gibbs bracket h−iĤt pertains
to the effective Hamiltonian (153). We find
∂g
= n−1 hΓt,µ iĤt ≤ 2n−1 sup |∂x ϕ|
∂Zµ
√
√
∂g
= n−1 h rtXi∗ − trxi iĤt ≤ 2n−1 S
0
∂Zi
and replacing in (156) we get k∇gk2 ≤ 2n−1 ( m
n sup |∂x ϕ| + S). Applying Proposition 15 we have
"
EZ,Z0
2 #
1
1
C(ϕ, S, α)
ln Ẑt − EZ,Z0
ln Ẑt
≤
n
n
n
(157)
almost surely. Taking the expectation in (157) gives the lemma.
Lemma 16. There exists an constant C(ϕ, S, α) > 0 such that
"
2 #
1
1
C(ϕ, S, α)
E EZ,Z0
ln Ẑt − EG
ln Ẑt
.
≤
n
n
n
(158)
Proof. We consider here g = n−1 EZ,Z0 [ln Ẑt ] as a function of V, W∗ , Φ and we work conditionnaly to the
other random variables.
hD
∂Γt,µ E i
∂g
= n−1 EZ,Z0 (Γt,µ + Zµ )
∂Vµ
∂Vµ Ht
r
h
i
2
−1
−1
)(2 sup |∂x ϕ|)
≤ n EZ,Z0 (2 sup |ϕ| + |Zµ |)(2 sup |∂x ϕ|) = n (2 sup |ϕ| +
π
The same inequality holds for
∂g
∂Wµ∗
. To compute the derivative w.r.t. Φµi we first remark
r
r 1 − t
n
∂Γt,µ
1−t ∗
=
Xi − xi ∂x ϕ
[ΦX∗ ]µ + k1 (t)Vµ + k2 (t)Wµ∗ , Aµ
∂Φµi
n
n
r
1−t
o
− ∂x ϕ
[Φx]µ + k1 (t)Vµ + k2 (t)wµ , aµ .
n
51
Therefore
hD
∂Γt,µ E i
∂g
= n−1 EZ,Z0 (Γt,µ + Zµ )
∂Φµi
∂Φµi Ht
r
h
i
2
−3/2
−3/2
≤n
EZ,Z0 (2 sup |ϕ| + |Zµ |)(4S sup |∂x ϕ|) = n
(2 sup |ϕ| +
)(4S sup |∂x ϕ|)
π
Putting these inequalities together we find
m
m n
X
∂g 2 X X
+
∂Vµ
µ=1
µ=1
µ=1 i=1
r
m
2
≤ 2 (2 sup |ϕ| +
)(2 sup |∂x ϕ|) +
n
π
k∇gk2 =
m
X
∂g
∂Vµ
2
+
∂g
∂Φµi
2
mn
(2 sup |ϕ| +
n3
r
2
)(4S sup |∂x ϕ|) .
π
We the lemma follows again from Proposition 15.
F.1.2
Bounded difference with respect to Aµ
The next step is an application of the variance bound of Lemma 16 to show that n1 EG [ln Ẑt ] concentrates w.r.t.
A (we still keep X∗ fixed for the moment).
Lemma 17. Let EA denotes the expectation w.r.t. A only. There exists an constant C(ϕ, S, α) > 0 such that
"
2 #
1
1
C(ϕ, S, α)
E EG
ln Ẑt − EA EG
ln Ẑt
.
(159)
≤
n
n
n
Proof. Let us consider n1 EG [ln Ẑt ] as a function g of A only. Let ν ∈ {1, . . . , m}. We must estimate variations
(ν)
(ν)
g(A) − g(A(ν) ) corresponding to two configurations A and A(ν) with Aµ = Aµ for µ 6= ν and Aν = Ãν .
(ν)
(ν)
We will use the notations Ĥt and Γt,µ to denote respectively the quantities Ĥt and Γt,µ where A is replaced
by A(ν) . By an application of Jensen’s inequality one finds
1
1
(ν)
(ν)
EG hĤt − Ĥt iĤ(ν) ≤ g(A) − g(A(ν) ) ≤ EG hĤt − Ĥt iĤt
t
n
n
(160)
where the Gibbs brackets pertain to the effective Hamiltonians (153). From (153) we obtain
(ν)
Ht
− Ht =
m
1
1 X (ν)2
(ν)
(ν)2
(ν)
Γt,µ − Γ2t,µ + 2Zµ (Γt,µ − Γt,µ ) =
Γt,ν − Γ2t,ν + 2Zν (Γt,ν − Γt,ν ) .
2
2
µ=1
Consequently
D
E
D
E
1
1
(ν)2
(ν)
(ν)2
(ν)
EG Γt,ν − Γ2t,ν + 2Zν (Γt,ν − Γt,ν ) (ν) ≤ g(A) − g(A(ν) ) ≤
EG Γt,ν − Γ2t,ν + 2Zν (Γt,ν − Γt,ν )
.
2n
2n
Ht
Ht
(161)
(ν)2
(ν)
Notice that Γt,ν − Γ2t,ν + 2Zν (Γt,ν − Γt,ν ) ≤ 4 sup |ϕ|2 + 4|Zν | sup |ϕ|. Thus we conclude by (161) that g
satisfies a bounded difference property:
r !
2
2
|g(A) − g(A(ν) )| ≤ sup |ϕ| sup |ϕ| + 2
.
(162)
n
π
Lemma 17 follows then by an application of Proposition 16.
52
F.1.3
Bounded difference with respect to Xi∗
∗
Let EΘ = EA EhG denotes
i the expectation w.r.t. all quenched variables except X . It remains to bound the
variance of EΘ n1 ln Ẑt (which only depends on X∗ ).
Lemma 18. There exists an constant C(ϕ, S, α) > 0 such that
"
2 #
1
1
C(ϕ, S, α)
E EΘ
ln Ẑt − E
ln Ẑt
≤
.
n
n
n
(163)
Proof. The lemma is proved using again a bounded difference argument. n1 EΘ [ln Ẑt ] is a function of X∗
that we denote by g. We will show that g satisfies a bounded difference property. Let j ∈ {1, . . . , n}. Let
∗(j)
X∗ , X∗(j) ∈ [−S, S]n be two input signals such that Xi = Xi∗ for i 6= j.
We are going to interpolate between g(X∗ ) and g(X∗(j) ). For s ∈ [0, 1] we define
ψ(s) = g(sX∗ + (1 − s)X∗(j) )
Obviously ψ(1) = g(X∗ ) and ψ(0) = g(X∗(j) ). Using Gaussian integration by parts, it is not difficult to
verify that for s ∈ [0, 1]
C(ϕ, S, α)
|ψ 0 (s)| ≤
.
n
This implies that |g(X∗ ) − g(X∗(j) )| ≤
F.1.4
C(ϕ,S,α)
n
and using Proposition 16 we obtain the lemma.
Proof of Theorem 4
ln Ẑt ≤
From Lemmas 14, 17 and 18 above, we obtain using Pythagorean theorem that Var
some constant C(ϕ, S, α) > 0. As mentioned before, this implies (thanks to (151)) Theorem 4.
F.2
1
n
C(ϕ,S,α)
n
for
Concentration of the overlap
In this appendix we give the main steps of the proof of Lemma 2. We denote by h−in,t, the Gibbs measure
associated to the perturbed Hamiltonian
n
X
√
x2
Ht (x, w; Y, Y ) +
i − xi Xi∗ − xi Ẑi
2
0
i=1
i.e., the sum of (68) and (76). It is crucial that the second term is a perturbation which preserves the Nishimori
identity of Appendix A.1. We note that the precise form of the first term does not matter and all subsequent
arguments are generic as long as it is a Hamiltonian whose Gibbs distribution satisfies this Nishimori identity.
The corresponding average free entropy is denoted fn, (t) and we call Fn, (t) the free entropy for a realization
of the quenched variables, that is Fn, (t) = n−1 ln Zt (Y, Y0 ).
Let
n
L :=
1 X x2i
xi zbi
− xi si − √ .
n
2
2
i=1
Up to the prefactor n−1 this quantity is the derivative of the perturbation term in (76). The fluctuations of the
53
overlap Q = n−1
Pn
∗
i=1 Xi xi
E (L − EhL in,t, )2
and those of L are related through the remarkable identity
n,t,
1
= Eh(Q − EhQin,t, )2
4
n,t,
1
+ E[hQ2 in,t, − hQi2n,t, ]
2
n
1 X
+ 2
E[hXi2 in,t, − hXi i2n,t, ].
4n
i=1
A detailed derivation can be found in Appendix IX of [13] and involves only some algebra using the Nishimori
identity and integrations by parts w.r.t. the Gaussian Ẑi in the perturbation term. Lemma 2 is then a direct
consequence of the following:
Proposition 17 (Concentration of L on E [hL i] ). Under assumptions (H1), (H2) and (H3) we have for any
0 < a < b < 1,
Z
b
d E (L − EhL in,t, )2
lim
n→+∞ a
n,t,
= 0.
(164)
The proof of this proposition is broken in two parts. Notice that
E (L − EhL in,t, )2
n,t,
= E (L − hL in,t, )2
n,t,
+ E (hL in,t, − EhL in,t, )2 .
(165)
Thus it suffices to prove the two following lemmas. The first lemma expresses concentration w.r.t. the posterior
distribution (or “thermal fluctuations”) and is an elementary consequence of concavity properties of the free
energy and the Nishimori identity.
Lemma 19 (Concentration of L on hL i ). Under assumptions (H1), (H2) and (H3), we have for any 0 < a <
b < 1,
Z
lim
n→+∞ a
b
d E (L − hL in,t, )2
n,t,
=0
(166)
The second lemma expresses the concentration of the average overlap w.r.t. the realizations of quenched
disorder variables and is a consequence of the concentration of the free energy (more precisely equation (143)
in Appendix F.1).
Lemma 20 (Concentration of hL i on EhL i ). Under assumptions (H1), (H2) and (H3), we have for any 0 <
a < b < 1,
Z
lim
n→+∞ a
b
d E (hL in,t, − E[hL in,t, ])2 = 0
(167)
The reader is referred to Sec. V of [13] for the proof of these two Lemmas. We point out that the analysis
gives a rate of decay O(n−1 ) for (166) which is optimal but a weaker decay rate for (167). However any decay
rate will suit for the present proof of the replica formula.
Acknowledgments
Jean Barbier acknowledges funding from the Swiss National Science Foundation (grant 200021-156672). Florent
Krzakala acknowledges funding from the ERC under the European Union 7th Framework Programme Grant
Agreement 307087-SPARCS. Lenka Zdeborová acknowledges funding from the European Research Council
(ERC) under the European Union’s Horizon 2020 research and innovation programme (grant agreement No
714608 - SMiLe). Part of this work was done while Léo Miolane was visiting EPFL.
54
References
[1] Github repository of numerical experiments for the current paper. https://github.com/sphinxteam/
GeneralizedLinearModel2017, 2017.
[2] E. Abbe. Community detection and stochastic block models: recent developments. ArXiv e-prints arxiv
1703.10146, March 2017.
[3] Madhu Advani and Surya Ganguli. An equivalence between high dimensional bayes optimal inference
and m-estimation. In Advances in Neural Information Processing Systems, pages 3378–3386, 2016.
[4] Carlo Baldassi, Christian Borgs, Jennifer T. Chayes, Alessandro Ingrosso, Carlo Lucibello, Luca Saglietti,
and Riccardo Zecchina. Unreasonable effectiveness of learning neural networks: From accessible states
and robust ensembles to basic algorithmic schemes. Proceedings of the National Academy of Sciences,
113(48):E7655–E7662, 2016.
[5] Carlo Baldassi, Alfredo Braunstein, Nicolas Brunel, and Riccardo Zecchina. Efficient supervised learning
in networks with binary synapses. Proceedings of the National Academy of Sciences, 104(26):11079–11084,
2007.
[6] J. Barbier, M. Dia, and N. Macris. Proof of threshold saturation for spatially coupled sparse superposition
codes. In 2016 IEEE International Symposium on Information Theory (ISIT), pages 1173–1177, July 2016.
[7] J. Barbier, F. Krzakala, L. Zdeborová, and P. Zhang. The hard-core model on random graphs revisited.
In Journal of Physics Conference Series, volume 473 of Journal of Physics Conference Series, page 012021,
December 2013.
[8] J. Barbier, N. Macris, and L. Miolane. The Layered Structure of Tensor Estimation and its Mutual Information. In 47th Annual Allerton Conference on Communication, Control, and Computing (Allerton),
September 2017.
[9] Jean Barbier. Statistical physics and approximate message-passing algorithms for sparse linear estimation
problems in signal processing and coding theory. PhD thesis, Université Paris Diderot, 2015.
[10] Jean Barbier, Mohamad Dia, Nicolas Macris, and Florent Krzakala. The mutual information in random
linear estimation. In 2016 54th Annual Allerton Conference on Communication, Control, and Computing
(Allerton), 2016.
[11] Jean Barbier, Mohamad Dia, Nicolas Macris, Florent Krzakala, Thibault Lesieur, and Lenka Zdeborová.
Mutual information for symmetric rank-one matrix estimation: A proof of the replica formula. In Advances in Neural Information Processing Systems 29, page 424–432. 2016.
[12] Jean Barbier and Florent Krzakala. Approximate message-passing decoder and capacity-achieving sparse
superposition codes.
[13] Jean Barbier and Nicolas Macris. The stochastic interpolation method: A simple scheme to prove replica
formulas in bayesian inference. Arxiv preprints arxiv 1705.02780, 2017.
[14] Jean Barbier, Nicolas Macris, Mohamad Dia, and Florent Krzakala. Mutual information and optimality
of approximate message-passing in random linear estimation. arXiv preprint arXiv:1701.05823, 2017.
[15] A. R. Barron and A. Joseph. Toward fast reliable communication at rates near capacity with gaussian
noise. In 2010 IEEE International Symposium on Information Theory, pages 315–319, June 2010.
[16] Eric B Baum and Yuh-Dauh Lyuu. The transition to perfect generalization in perceptrons. Neural computation, 3(3):386–401, 1991.
55
[17] M. Bayati and A. Montanari. The lasso risk for gaussian matrices. IEEE Transactions on Information
Theory, 58(4):1997–2017, April 2012.
[18] Mohsen Bayati, Marc Lelarge, and Andrea Montanari. Universality in polytope phase transitions and
message passing algorithms. The Annals of Applied Probability, 25(2):753–822, 2015.
[19] Mohsen Bayati and Andrea Montanari. The dynamics of message passing on dense graphs, with applications to compressed sensing. IEEE Transactions on Information Theory, 57(2):764–785, Feb 2011.
[20] Stéphane Boucheron, Gábor Lugosi, and Olivier Bousquet. Concentration inequalities. In Advanced
Lectures on Machine Learning, pages 208–240. Springer, 2004.
[21] Petros T Boufounos and Richard G Baraniuk. 1-bit compressive sensing. In Information Sciences and
Systems, 2008. CISS 2008. 42nd Annual Conference on, pages 16–21. IEEE, 2008.
[22] Peter Bühlmann and Sara Van De Geer. Statistics for high-dimensional data: methods, theory and applications. Springer Science & Business Media, 2011.
[23] F. Caltagirone, M. Lelarge, and L. Miolane. Recovering asymmetric communities in the stochastic block
model. ArXiv e-prints, October 2016.
[24] E. J. Candes and T. Tao. Near-optimal signal recovery from random projections: Universal encoding
strategies? IEEE Transactions on Information Theory, 52(12):5406–5425, Dec 2006.
[25] Sourav Chatterjee et al. A generalization of the lindeberg principle. The Annals of Probability, 34(6):2061–
2076, 2006.
[26] François Chollet. keras. https://github.com/fchollet/keras, 2015.
[27] A. Decelle, F. Krzakala, C. Moore, and L. Zdeborová. Inference and Phase Transitions in the Detection
of Modules in Sparse Networks. Physical Review Letters, 107(6):065701, August 2011.
[28] David Donoho and Andrea Montanari. High dimensional robust m-estimation: asymptotic variance via
approximate message passing. Probability Theory and Related Fields, 166(3):935–969, Dec 2016.
[29] David Donoho and Jared Tanner. Observed universality of phase transitions in high-dimensional geometry, with implications for modern data analysis and signal processing. Philosophical Transactions of
the Royal Society of London A: Mathematical, Physical and Engineering Sciences, 367(1906):4273–4293,
2009.
[30] David L. Donoho, Arian Maleki, and Andrea Montanari. Message-passing algorithms for compressed
sensing. Proc. Natl. Acad. Sci., 106(45):18914–18919, 2009.
[31] David L Donoho, Arian Maleki, and Andrea Montanari. Message-passing algorithms for compressed
sensing. Proceedings of the National Academy of Sciences, 106(45):18914–18919, Nov 2009.
[32] Viktor Dotsenko. An introduction to the theory of spin glasses and neural networks, volume 54. World
Scientific, 1995.
[33] Noureddine El Karoui, Derek Bean, Peter J. Bickel, Chinghway Lim, and Bin Yu. On robust regression
with high-dimensional predictors. Proceedings of the National Academy of Sciences, 110(36):14557–14562,
2013.
[34] Andreas Engel and Christian Van den Broeck. Statistical mechanics of learning. Cambridge University
Press, 2001.
[35] Konrad H Fischer and John A Hertz. Spin glasses, volume 1. Cambridge university press, 1993.
56
[36] Elizabeth Gardner and Bernard Derrida. Three unfinished works on the optimal storage capacity of
networks. Journal of Physics A: Mathematical and General, 22(12):1983, 1989.
[37] Tom Goldstein and Christoph Studer. Phasemax: Convex phase retrieval via basis pursuit. arXiv preprint
arXiv:1610.07531, 2016.
[38] Rémi Gribonval and Pierre Machart. Reconciling" priors" &" priors" without prejudice? In Advances in
Neural Information Processing Systems, pages 2193–2201, 2013.
[39] Francesco Guerra and Fabio Lucio Toninelli. The thermodynamic limit in mean field spin glass models.
Communications in Mathematical Physics, 230(1):71–79, 2002.
[40] Dongning Guo, S. Shamai, and S. Verdú. Mutual information and minimum mean-square error in gaussian channels. IEEE Transactions on Information Theory, 51(4):1261–1282, April 2005.
[41] Dongning Guo and Sergio Verdú. Randomly spread cdma: Asymptotics via statistical physics. IEEE
Transactions on Information Theory, 51(6):1983–2010, June 2005.
[42] Géza Györgyi. First-order transition to perfect generalization in a neural network with binary synapses.
Physical Review A, 41(12):7097, 1990.
[43] D Hansel, G Mato, and C Meunier. Memorization without generalization in a multilayered neural network. EPL (Europhysics Letters), 20(5):471, 1992.
[44] Tadaaki Hosaka, Yoshiyuki Kabashima, and Hidetoshi Nishimori. Statistical mechanics of lossy data
compression using a nonmonotonic perceptron. Physical Review E, 66(6):066126, 2002.
[45] Adel Javanmard and Andrea Montanari. State evolution for general approximate message passing algorithms, with applications to spatial coupling. Information and Inference: A Journal of the IMA, 2(2):115–
144, 2013.
[46] Y Kabashima. Inference from correlated patterns: a unified theory for perceptron learning and linear
vector channels. Journal of Physics: Conference Series, 95(1):012001, 2008.
[47] S. B. Korada and N. Macris. Tight bounds on the capacity of binary input random cdma systems. IEEE
Transactions on Information Theory, 56(11):5590–5613, Nov 2010.
[48] Satish Babu Korada and Nicolas Macris. Exact solution of the gauge symmetric p-spin glass model on a
complete graph. Journal of Statistical Physics, 136(2):205–230, 2009.
[49] Satish Babu Korada and Andrea Montanari. Applications of the Lindeberg principle in communications
and statistical learning. IEEE Transactions on Information Theory, 57(4):2440–2450, 2011.
[50] F. Krzakala, A. Montanari, F. Ricci-Tersenghi, G. Semerjian, and L. Zdeborová. Gibbs states and the set
of solutions of random constraint satisfaction problems. Proceedings of the National Academy of Science,
104:10318–10323, June 2007.
[51] F. Krzakala, C. Moore, E. Mossel, J. Neeman, A. Sly, L. Zdeborová, and P. Zhang. Spectral redemption in
clustering sparse networks. Proceedings of the National Academy of Science, 110:20935–20940, December
2013.
[52] Florent Krzakala, Marc Mézard, François Sausset, YF Sun, and Lenka Zdeborová. Statistical-physicsbased reconstruction in compressed sensing. Phys. Rev. X, 2:021005(18), May 2012.
[53] Florent Krzakala, Marc Mézard, Francois Sausset, Yifan Sun, and Lenka Zdeborová. Probabilistic reconstruction in compressed sensing: algorithms, phase diagrams, and threshold achieving matrices. Journal
of Statistical Mechanics: Theory and Experiment, 2012(08):P08009(57), 2012.
57
[54] Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. Nature, 521(7553):436–444, 2015.
[55] M. Lelarge and L. Miolane. Fundamental limits of symmetric low-rank matrix estimation. ArXiv e-prints,
November 2016.
[56] T. Lesieur, L. Miolane, M. Lelarge, F. Krzakala, and L. Zdeborová. Statistical and computational phase
transitions in spiked tensor estimation. ArXiv e-prints, January 2017.
[57] N. Macris. Griffith-Kelly-Sherman correlation inequalities: A useful tool in the theory of error correcting
codes. IEEE Transactions on Information Theory, 53(2):664–683, Feb 2007.
[58] M Mézard. The space of interactions in neural networks: Gardner’s computation with the cavity method.
Journal of Physics A: Mathematical and General, 22(12):2181–2190, 1989.
[59] Marc Mezard and Andrea Montanari. Information, physics, and computation. Oxford University Press,
2009.
[60] Marc Mézard, Giorgio Parisi, and Miguel-Angel Virasoro. Spin glass theory and beyond. 1987.
[61] Paul Milgrom and Ilya Segal. Envelope theorems for arbitrary choice sets. Econometrica, 70(2):583–601,
2002.
[62] L. Miolane. Fundamental limits of low-rank matrix estimation: the non-symmetric case. ArXiv e-prints,
February 2017.
[63] Andrea Montanari. Estimating random variables from random sparse observations. European Transactions on Telecommunications, 19(4):385–403, 2008.
[64] John Ashworth Nelder and R Jacob Baker. Generalized linear models. Wiley Online Library, 1972.
[65] Hidetoshi Nishimori. Statistical physics of spin glasses and information processing: an introduction, volume
111. Clarendon Press, 2001.
[66] Manfred Opper and David Haussler. Generalization performance of bayes optimal classification algorithm for learning a perceptron. Physical Review Letters, 66(20):2677, 1991.
[67] Samet Oymak, Amin Jalali, Maryam Fazel, Yonina C Eldar, and Babak Hassibi. Simultaneously structured
models with application to sparse and low-rank matrices. IEEE Transactions on Information Theory,
61(5):2886–2908, 2015.
[68] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer,
R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and E. Duchesnay.
Scikit-learn: Machine learning in Python. Journal of Machine Learning Research, 12:2825–2830, 2011.
[69] Francisco J Piera and Patricio Parada. On convergence properties of shannon entropy. Problems of
Information Transmission, 45(2):75–94, 2009.
[70] S. Rangan. Generalized approximate message passing for estimation with random linear mixing. In 2011
IEEE International Symposium on Information Theory Proceedings, pages 2168–2172, July 2011.
[71] G. Reeves and H. D. Pfister. The replica-symmetric prediction for compressed sensing with gaussian
matrices is exact. In 2016 IEEE International Symposium on Information Theory (ISIT), pages 665–669,
July 2016.
[72] Tom Richardson and Ruediger Urbanke. Modern coding theory. Cambridge university press, 2008.
[73] H. S. Seung, H. Sompolinsky, and N. Tishby. Statistical mechanics of learning from examples. Phys. Rev.
A, 45:6056–6091, Apr 1992.
58
[74] Claude E Shannon. A mathematical theory of communication, part i, part ii. Bell Syst. Tech. J., 27:623–656,
1948.
[75] Mahdi Soltanolkotabi. Structured signal recovery from quadratic measurements: Breaking sample complexity barriers via nonconvex optimization. arXiv preprint arXiv:1702.06175, 2017.
[76] Michel Talagrand. Mean field models for spin glasses: Volume I: Basic examples, volume 54. Springer
Science & Business Media, 2010.
[77] Toshiyuki Tanaka. A statistical-mechanics approach to large-system analysis of cdma multiuser detectors. IEEE Transactions on Information Theory, 48(11):2888–2910, Nov 2002.
[78] David J Thouless, Philip W Anderson, and Robert G Palmer. Solution of‘solvable model of a spin glass’.
Philosophical Magazine, 35(3):593–601, 1977.
[79] Jeremy P Vila and Philip Schniter. Expectation-maximization gaussian-mixture approximate message
passing. IEEE Transactions on Signal Processing, 61(19):4658–4672, 2013.
[80] Martin J Wainwright, Michael I Jordan, et al. Graphical models, exponential families, and variational
inference. Foundations and Trends® in Machine Learning, 1(1–2):1–305, 2008.
[81] Timothy L. H. Watkin, Albrecht Rau, and Michael Biehl. The statistical mechanics of learning a rule.
Rev. Mod. Phys., 65:499–556, Apr 1993.
[82] Yihong Wu and Sergio Verdú. Optimal phase transitions in compressed sensing. IEEE Transactions on
Information Theory, 58(10):6241–6263, Oct 2012.
[83] Yingying Xu, Yoshiyuki Kabashima, and Lenka Zdeborová. Bayesian signal reconstruction for 1-bit
compressed sensing. Journal of Statistical Mechanics: Theory and Experiment, 2014(11):P11015, 2014.
[84] Lenka Zdeborová and Florent Krzakala. Statistical physics of inference: thresholds and algorithms.
Advances in Physics, 65(5):453–552, 2016.
[85] Justin Ziniel, Philip Schniter, and Per Sederberg. Binary linear classification and feature selection via
generalized approximate message passing. In Information Sciences and Systems (CISS), 2014 48th Annual
Conference on, pages 1–6. IEEE, 2014.
59
| 7cs.IT
|
What Really is Deep Learning Doing?
∗
Chuyu Xiong
Independent Researcher, New York, USA
Email: [email protected]
arXiv:1711.03577v1 [cs.LG] 6 Nov 2017
Abstract
Deep learning has achieved a great success in many areas, from computer vision to natural
language processing, to game playing, and much more. Yet, what deep learning is really doing is
still an open question. There are a lot of works in this direction. For example, [6] tried to explain
deep learning by group renormalization, and [5] tried to explain deep learning from the view of
functional approximation. In order to address this very crucial question, here we see deep learning
from perspective of mechanical learning and learning machine (see [1], [2]). From this particular
angle, we can see deep learning much better and answer with confidence: What deep learning is
really doing? why it works well, how it works, and how much data is necessary for learning. We
also will discuss advantages and disadvantages of deep learning at the end of this work.
Keywords: Mechanical learning, learning machine, deep learning, X-form, universal
learning machine, internal representation space, data sufficiency
1
Introduction
In recent years, deep learning (a branch of machine learning) has achieved many successes in lot of
fields. However, a clear theoretical framework of deep learning is still missing. Consequently, there are
many fundamental questions about deep learning are still open. For example: What is deep learning
really doing? Is it really learning, or just a kind of fancy approximation to a function? Why it indeed
has so many success? Why deep learning needs big data? For a particular problem, how much data
is sufficient to drive deep learning to learn? Up to now, there is no satisfactory answer for these
fundatamental questions. We here are trying to address these questions from a new angle.
We introduced term ”Mechanical learning” in [1]. Mechanical learning is a computing system that is
based on a simple set of fixed rules (this is so called mechanical), and can modify itself according to
incoming data (this is so called learning). A learning machine is a system that realizes mechanical
learning.
In [2], we described learning machine in a lot of details. By doing so, we gained some useful knowledges
and insights of mechanical learning.
In this short article, by using those knowledges and insights, we are trying to view deep learning from
a new angle. First, we will briefly discuss learning machine, pattern, internal representation space,
X-form, data sufficiency, and learning strategies and methods. Then, we will use the view of learning
machine to see deep learning. We start from the simplest deep learning, i.e., 2-1 RBM, then go to
3-1 RBM, N -1 RBM, N -M RBM, stacks of RBMs. Then we discuss the learning dynamics of deep
learning. By this approach, we see clearly what deep learning is doing, why deep learning is working,
under what condition deep learning can learn well, how much data are needed, and what disadvantages
deep learning has.
2
Mechanical Learning and Learning Machine
We here very briefly sum up the discussions that we did in [1] and [2]. A learning machine M has 2
major aspects: it is an IPU, i.e. it is able to process information; and it is learning, i.e. its information
∗ Great
thanks for whole heart support of my wife.
1
2
What really is deep learning?
processing ability is changing according to data. Without learning (since it is a machine we design,
we can stop learning), M is very similar to a CPU. However, one major difference between learning
machine and CPU is: learning machine treat incoming data according to its pattern, not bit-wise.
Thus, in order to understand a learning machine, it’s absolutely necessary to understand pattern well.
There are 2 kinds of patterns: objective pattern and subjective pattern. Subjective pattern is crucial
for learning machine. In [2], we proved one theorem: For any objective pattern Po , we can find a
proper subjective pattern p that can express Po well and is build upon a least set of base patterns.
To describe subjective patterns, it is best to use X-form, which is one algebraic expression upon some
base patterns. X-form is one very important mathematical object. X-form could have sub-forms.
X-form and its sub-forms actually forms the fundamental fabric of a learning machine.
We also defined learning by teaching and learning without teaching. Then, further specify typical
mechanical learning. Learning by teaching requires we know learning machine and the pattern to
learn well. By these knowledges, we can design a teaching sequence to make learning machine learn
well. We proved that if a learning machine has certain capabilities for learning by teaching, it is
universal, i.e. able to learn anything.
However, most learning is not learning by teaching. In order to understand typical mechanical learning,
we introduced internal representation space. Structurally, a learning machine M has components:
input space, output space, internal representation space, and learning methods and learning strategies.
The most important part is internal representation space. We studies internal representation space
E in details, and revealed, in fact, it is equivalent to a collection of X-forms. This fact tells us
that learning is nothing but a dynamics on E, moving from one X-form to another. With clear
and reachable internal representation space, learning can be understood much better, and can be
done much more efficiently. For example, we can unify all 5 kinds of learning – logic reasoning,
connectionism, probabilistic approach, analogy, and evolution (see [3]) – together on E naturally.
For mechanical learning, we need to understand data sufficiency. This is very crucial concept. We
use X-form and its sub-forms to define data sufficient to support one X-form and sufficient to bound
one X-form. With sufficient data, we can see how learning strategy and learning method work. There
could be many learning strategies and learning methods. We show 3 learning strategies: 1. Embed
X-form into parameter space. 2. Squeeze X-form from inside to higher abstraction. 3. Squeeze Xform from inside and outside to higher abstraction. We prove that with certain capabilities, the last
2 strategies and methods will make universal learning machine. Of course, this is theoretical results,
since we have not designed a specific learning machine yet.
Here, we will show that deep learning is actually doing mechanical learning by the first strategy, i.e.
embed X-forms into parameter space. Such a fact will help us to understand deep learning much
better.
3
See Deep Learning from View of Learning Machine
According to our definition, if without human intervention, deep learning is mechanical learning. Of
course, this ”if” is a big if. Often, deep learning program is running with a lot of human intervention,
specially at the time of model set up. We will restrict our discussion to Hinton’s original model [4],
i.e., a stack of RBMs. Each level of RBM is clearly a N -M learning machine (N andM are dimensions
of input and output). Hinton’s deep learning model is by stacking RBM together. If without further
human intervention, it is a learning machine. This is the original model of deep learning. Other deep
learning program can be thought as variations based on this model. Though in the past few years deep
learning has leaped forward greatly, stacking RBM together still reflects the most typical properties
of deep learning.
Thus, we would expect many things in mechanical learning could be applied to deep learning. The
point is, we are viewing deep learning from a quite different angle – angle of mechanical learning. For
example, we can view Hinton’s original deep learning program [4] as one 258-4 learning machine, and
we ask what is the internal representation space of it? We expect such angle and questions would
reveal useful things for us. The central questions indeed are: what is the internal representation
space of deep learning? what is the learning dynamics? At first, seems it is quite hard since learning
is conducted on a huge parameter space (dimension could be hundreds of millions), and learning
methods are overwhelmingly a big body of math. However, when we apply the basic thoughts of
learning machine to deep learning, starting from simplest RBM, i.e. 2-1 RBM, we start to see much
3
Chuyu Xiong
more clearly.
2-1 RBM
2-1 RBM is the simplest. However, it is also very useful since we can examine all details, and such
details will give us a good guide on more general RBM.
2
2-1 RBM is one IPU. We know 2-1 IPU totally has 16 processing (22 ). But, we only consider those
processing: p(0, 0) = 0, so totally 8 processing, which we denote as Pj , j = 0, . . . , 7 (see [1]). For 2-1
RBM, any processing p can be written as: for input (i1 , i2 ), the output o is:
1 if x ≥ 0
2
o = p(i1 , i2 ) = Sign(ai1 + bi2 ), where (a, b) ∈ R , Sign(x) =
0 if x < 0
The parameters (a, b) determine what the processing really is. Parameter space R2 has infinite many
choices of parameters. But, there are only 6 processing, thus, for many different parameters, the
processing is actually same. We can see all processing in below table:
(0,0)
(1,0)
(0,1)
(1,1)
Region
X-form
P0
0
0
0
0
R4
0
P1
0
1
0
0
R3
b1
P2
0
0
1
0
R5
b2
P3
0
1
0
1
R2
b1 + b3
P4
0
0
1
1
R6
b2 + b3
P5
0
1
1
0
None
b1 + b2
P6
0
0
0
1
None
b3
P7
0
1
1
1
R1
b1 + b2 + b3
Tab. 1. Table for all processing of 2-1 RBM
Fig. 1. Parameter space of 2-1 RBM that is cut into 6 regions
In first row of table, Pi , i = 0, . . . , 7 are all processing of 2-1 IPU. Under first row, there is value table
for each processing. We point out some quite familiar processing: P7 is actually logical OR gate, P6
is logical AND gate, P5 is logical XOR gate. Note, P5 , P6 are processing for 2-1 IPU, but not in 2-1
RBM. It is well known, 2-1 RBM has no XOR and AND (i.e. no P5 , P6 ).
Rj , j = 1, . . . , 6 indicate regions in parameter space R2 , each region for one processing. There are
only 6 regions, since 2-1 RBM only has 6 processing. We briefly discuss how we get these regions. See
illustration in figure.
Suppose p is processing. Line a = 0 cuts R2 into 2 regions: a ≥ 0 and a < 0. If (a, b) is in first region,
then p(1, 0) = 1, in second, p(1, 0) = 0. Line b = 0 is perpendicular to a = 0, so, it cuts the previous
4
What really is deep learning?
2 regions into 4 regions: a ≥ 0, b ≥ 0 and a ≥ 0, b < 0 and a < 0, b ≥ 0 and a < 0, b < 0. Clearly, if
b ≥ 0, p(0, 1) = 1, if b < 0, p(0, 1) = 0. Line a + b = 0 could no longer cuts the previous 4 regions into
8 regions, it could only cut 2 regions (2nd, 4th quadrant) into 4 regions (R2 , R3 , R5 , R6 ). So, totally,
we have 6 regions, and each region is for one processing. This argument about regions is very simple,
yet very effective. We can extend this argument to N -1 RBM.
Each region is for one processing. So, region can be used to represent processing. That is 6th row
in the table shown. Yet, a much better expression is by X-form ([2]). We explain them here. Here
b0 = (0, 0), b1 = (1, 0), b2 = (0, 1), b3 = (1, 1) are base patterns. For 2-dim pattern space, there are
only these 4 base patterns. But, bi can also be used to represent one processing of 2-1 IPU, i.e. bi
is such a processing: when input is bi , output is 1, otherwise output is zero. X-forms are expressions
based on all N -1 base patterns, operations +, ·, ¬, composition, Cartesian product, and apply them
consecutively. Example, b1 + b3 , b1 · b2 , b1 + ¬b2 are X-forms. Any processing of 2-1 IPU can be
expressed by at least one X-form [2]. For example, if region is R3 , processing is P1 , X-form is b1 .
Another example, region is R1 , processing is P7 (this is OR gate), X-form is b1 + b2 + b3 . P5 is a
processing of 2-1 IPU (XOR gate), but not in 2-1 RBM, its X-form is b1 + b2 . The 7th row in the table
shows X-forms representing processing. We can say that each processing is represented by a region,
and by a X-form as well.
When 2-1 RBM is learning, clearly, parameter (a, b) is adapting. But, only when (a, b) cross region,
processing changes. Before crossing, change of parameters is just for preparation for crossing (perhaps
many parameter changes are just wasted). Learning is moving from one region to another. Or,
equivalently, learning is from one X-form to another. Such view is crucial. Now, we are clear, on
surface, learning on 2-1 RBM is a dynamics on parameter space R2 , but real learning dynamics is on
6 regions (or X-forms). Such indirectness causes a lot of problem.
3-1 RBM
Just increase input dimension 1, we have 3-1 RBM. To discuss it, we can gain some insights for general
RBM. For 3-1 RBM, still we can write: for any input (i1 , i3 , i3 ) ∈ B3 , output o ∈ B is:
1 if x ≥ 0
o = p(i1 , i2 , i3 ) = Sign(ai1 + bi2 + ci3 ), where (a, b, c) ∈ R3 , Sign(x) =
0 if x < 0
However, while we can easily write down all possible processing of 2-1 RBM, it would be hard to do
3
so for 3-1 RBM. For 3-1 IPU, we know that the number of all possible processing is 22 = 28 = 256.
Since only considering such processing p: p(0, 0, 0) = 0, the number becomes 256/2 = 128. We expect
3-1 RBM has less processing. But, how many possible processing of 3-1 RBM could have?
Following the guidance that 2-1 RBM gives to us, i.e. to consider the hyperplane generated by
nonlinearity that cuts parameter spaces, we examine parameter space (a, b, c) ∈ R3 , and following
planes: a = 0, b = 0, c = 0, a + b = 0, a + c = 0, b + c = 0, a + b + c = 0. These planes are
naturally obtained. For example, if we consider the input (1, 0, 0), it is easy to see plane a = 0 is
where cut the value of output: 1 or 0. So, the above planes are associated with following inputs:
(1, 0, 0), (0, 1, 0), (0, 0, 1), (1, 1, 0), (1, 0, 1), (0, 1, 1), (1, 1, 1)
We can clearly see that in one region that is cut out by above 7 planes, the output values are same.
Therefore, one region actually is representing one processing: in the region, processing is same. So,
question of how many possible processing becomes how many possible regions cut out by the 7 planes.
We do counting for the regions below.
First, a = 0 cuts parameter space into 2 pieces: R11 , R21 . Second, b = 0 perpendicular to a = 0,
so, it cuts each region R11 , R21 into 2 pieces, we then have 4 regions: R12 , R22 .R32 , R42 . Then, c = 0
perpendicular to a = 0 and b = 0, so, we have 8 regions: Rj3 , j = 1, . . . , 8. Then, consider a + b = 0.
This plane no longer could be perpendicular to all a = 0, b = 0, c = 0. We will not have 2 ∗ 8 = 16
regions. We only have 1.5 ∗ 8 = 12 regions. Following the same argument, we have: For a + c = 0,
1.5 ∗ 12 = 18 regions. For b + c = 0, 1.5 ∗ 18 = 27 regions. For a + b + c = 0, 1.5 ∗ 27 < 41 regions.
So, for 3-1 RBM, there are at most 41 possible processing, comparing to 128 possible processing of
full 3-1 IPU. However, there are possibility that the number of processing is even less than 41, since
among those regions, it is possible that 2 different regions give same processing. We do not consider
these details here.
Since regions can be represented by X-form, each processing 3-1 RBM can be represented by at least
one X-form. b1 = (1, 0, 0), b2 = (0, 1, 1), b3 = (0, 0, 1), . . . , b7 = (1, 1, 1) are X-form for all base patterns.
5
Chuyu Xiong
Examples, X-form b1 + b2 is in 3-1 RBM. But, b1 · b2 is not in 3-1 RBM. There are a lot of such X-form
that is not in 3-1 RBM.
Learning dynamics on 3-1 RBM is also in such way: on surface, it is dynamics on R3 , but real learning
dynamics is on 41 regions (or X-forms).
N -1 RBM
The argument for 3-1 RBM can be extended to N -1 RBM (See details in [2]). We consider hyperplanes
N
and regions cut off by these hyperplanes. The number of these regions is less than: 2N 1.52 −N −1 .
N
Compare to the number of all processing of N -1 IPU, which is 22 −1 , easy to see, N -1 RBM has much
less processing. This means that N -1 RBM could not express many processing.
For N -1 RBM, still we can write: for any input (i1 , . . . , iN ) ∈ BN , output o ∈ B is:
o = p(i1 , . . . , i3 ) = Sign(a1 i1 + a2 i2 + . . . + aN i3 ), where (a1 , . . . , aN ) ∈ RN , Sign(x) =
1 if x ≥ 0
0 if x < 0
a1 = 0, . . . , a1 + a2 = 0, . . . , a1 + a2 + a3 = 0 . . . , . . . , a1 + a2 + a3 + . . . = 0, . . .
(1)
N
N
There are 1 hyperplanes such as a1 = 0; 2 hyperplanes such as a1 + a2 = 0; . . . . We also have
this: First N hyperplanes will cut parameter space into 2N regions. Then, later each hyperplanes will
cut more regions by the rate of multiplying factor 1.5. Thus, we can see the number of regions are:
2N ∗ 1.5K2 ∗ 1.5K3 ∗ . . . ∗ 1.5KN
where K2 =
N
2
is the number of hyperplanes such as a + b = 0, etc.
And, we have the equation:
N
2
N
N
N
N
=
+
+
+ ... +
0
1
2
N
(2)
So,
K2 + K3 + . . . + KN =
N
N
N
N
N
+
+ ... +
= 2N −
−
= 2N − N − 1
2
3
N
1
0
(3)
Thus, the number of regions are
N
2N ∗ 1.52
−N −1
3 N
= 2N ∗ ( )2 −N −1
2
This is a very big number. Yet, compare to the total possible processing of full IPU, it is quite small.
See their quotient:
N
22
2N −N −1
3 2N −N −1 = 2 ∗ (4/3)
N
2 ∗ (2)
N
It tells that full IPU has f = 2 ∗ (4/3)2 −N −1 times more processing comparing to RBMs. This is
huge difference. Say, just for N = 10, f is more than 120 digits, i.e. the number of processing of full
IPU would has more 120 digits at the end than the number of RBMs.
Also, each region can be expressed by at least one X-form. For examples, b1 + bN , b1 + b3 + b5 , etc.
Learning dynamics on N -1 RBM is in such way: on surface, it is dynamics on RN , but real learning
N
dynamics is on those 2N 1.52 −N −1 regions (or X-forms).
N -M RBM
Suppose Ri , i = 1, . . . , M are M N -1 RBMs, we can form a N -M RBM, denote as R = (R1 , . . . .RM ),
whose processing are p = (p1 , p2 , . . . , pM ), where pi , i = 1, . . . , M are processing of Ri . So, R is
Cartesian product of Ri , i = 1, . . . , M .
Since all Ri are cut into regions, and in each region, processing is same, we can see R is also cut into
regions, and each region is a Cartesian product of regions of Ri : R = R1 × R2 × . . . × RM , where Ri
6
What really is deep learning?
N
is one region from i-th RBM Ri . Thus, the number of all possible regions of R is (2N 1.52 −N −1 )M =
N
N
2N M 1.5M (2 −N −1) . This is a much smaller number than 2M 2 , which is the number of total possible
processing for N -M IPU.
X-form for each region of R, are actually Cartesian product of X-form for those regions of Ri . Suppose
R = R1 × R2 × . . . × RM , and fi are X-forms for region Ri in Ri , i = 1, . . . , M , then X-form for R is
f = (f1 , . . . , fM ). For example, (b1 , b1 + b3 , . . . , b2 · b4 ) is one X-form of R.
Learning dynamics on N -M RBM is in such way: on surface, it is dynamics on parameter space RN M ,
N
but real learning dynamics is on those 2N M 1.5M (2 −N −1) regions (or X-forms).
Stacking RBMs
Consider a N -M RBM R1 , and a M -L RBM R2 , stacking them together, we get one N -L IPU R: A
processing p of R are composition of processing p1 , p2 of R1 , R2 : p(i1 , i2 , . . . , iN ) = p2 (p1 (i1 , i2 , . . . , iN )).
And we denote as: R = R1 ⊗ R.
The parameter space of R clearly is RN M × RM L . We know RN M is cut into some regions, in each
region processing is same. So does RM L . Thus, RN M +M L is cut into some regions, in each region
processing is same, and these regions are Cartesian product of regions in RN M and RM L . So, we
know number of total possible processing R equals total possible processing of R1 times R2 , i.e.
N
M
N
M
2N M 1.5M (2 −N −1) × 2M L 1.5L(2 −M −1) = 2N M +M L 1.5M (2 −N −1)+L(2 −M −1) .
N
We can easily see if M is large enough, the above number will become greater than 2L2 , which is
total possible processing of R. We can see, at least potentially, R has enough ability to become a N -L
full IPU. But, we will not consider here. In fact, it is very closely related to the so called Universal
Approximation Theorem. Indeed, stacking RBM together is powerful.
X-form can be expressed as composition as well. For example, consider 3 2-1 RBM R1 , R2 , and
R3 . Using R1 and R2 to form a 2-2 RBM, and using R3 to stack on it, we get a 2-1 IPU R:
R = R3 ⊗ (R1 , R2 ). If for this case, R1 has X-form b1 , and R2 has X-form b2 , and R3 has X-form
b1 + b2 + b3 , them, R has X-form (b1 + b2 + b3 )(b1 , b2 ). Easy to see this X-form is processing P5 (XOR
gate), which is not expressible by one 2-1 RBM. So, putting 3 2-1 RBMs together, more X-form can
be expressed.
4
Learning Dynamics of Deep Learning
With these understandings RBMs, which is the most essential building block of deep learning, we can
see how the model of deep learning is build up, and how learning dynamics is doing. Clearly, today’s
deep learning is much more than original Hinton’s model of stacking RBMs (see [4]). But, we would
first talking about this model.
The deep learning model is by stacking more RBMs (this is so called deep). Once several RBMs are
putting together, a deep learning model is formed. Suppose Rj , are Nj -Nj+1 RBM, j = 1, . . . , J, where
N1 , . . . , NJ , NJ+1 are sequence of integers. So, we can stack these RBM together to form one N1 -NJ+1
IPU, whose processing could be written in such way: p(i1 , i2 , . . . , iN1 ) = pJ (. . . p2 (p1 (i1 , i2 , . . . , iN1 ))
where each pj is processing of Rj . We denote this IPU as R = R1 ⊗ . . . ⊗ RJ . All parameters of R
form a huge Euclidean space RN1 N2 +...+NJ NJ+1 . We can denote this huge parameter space as R∗ .
Then, clearly, deep learning is conducted on R to reach a good processing by modifying the parameters
in R∗ . Of course, it requires skills to modify such a huge number of parameters. There are methods,
such as CD (convergent divergence), SGD (stochastic gradient descent), etc. are invented for such
purpose.
However, no matter what methods are used to modify the parameters, it is modifying parameters
to form the dynamics of learning. So, seems the phase space of learning dynamics is on the space
R∗ . But, this is just on surface. As we discussed in last section, the true dynamics is not conducted
on parameters, but on those regions. The learning dynamics is conducted on these regions cut by
N1
hyperplanes and Cartesian products. The number of regions are huge: 2N1 N2 1.5N2 (2 −N1 −1) × . . ..
More precisely, the situation is: as learning, a huge number of parameters are adapting, but only
when parameters cross region, the processing of R changes. Before crossing, processing remains same,
the changes of parameter at most can be thought as the preparation for crossing (perhaps many
such changes of parameters are just wasted). Thus, learning dynamics is to move from one region
7
Chuyu Xiong
to another. We also know each region is associated with one X-form. Thus, learning dynamics is to
move from one X-form to another X-form.
Fig. 2. Illustration on parameter space is cut into regions
Fig. 2 gives one illustration on parameter space is cut into regions. Of course, R∗ is very high
dimension Euclidean space, so regions could be shown on paper precisely, and Fig. 2 is just one
illustration. However, this illustration gives us one clear picture about deep learning structure.
The deep learning structure is formed by these factors: how many RBMs, dimension of each RBM, how
to stack, how to do Cartesian product. Once the structure is formed, if no further human intervention
(such as manually adjust numbers or subroutines in model), the structure will not change. Such
structure is formed by people at set up time. So, for fixed structure, we will have fixed region cut (as
illustrated in Fig. 2). Further, we will have a fixed set of X-forms, and learning is conducted on this
set of X-forms.
We can see one example. R1 , R2 , R3 are 3 2-1 RBMs. We put them like this: R = R3 ⊗ (R1 , R2 ).
We have 3 parameter space (a, b), (c, d), (e, f ). We have 6 regions for each parameter space. Put them
together, we have 6x6x6 = 216 regions. R is one 2-1 IPU. So, R at most has 8 processing. Thus,
among those 216 regions some different regions will have same processing. But, each region will have
one X-form. That is to say, for one processing, there could have several X-form associated with.
For example, consider this region: R13 × (R31 , R52 ). This gives processing P5 (XOR gate). Normally,
for this processing, we can use X-form b1 + b2 for it. But, for the region, naturally, the X-form is:
(b1 +b2 +b3 )⊗(b1 , b2 ). That is to say, this X-form will generate the same processing as b1 +b2 . Another
region: R23 × (R21 , R62 ) will give the same processing. And, the X-form is: (b1 + b3 ) ⊗ (b1 + b3 , b2 + b3 ).
For deep learning model build on stacking RBMs (original Hinton’s model, [4]), as we discussed in
last section, the situation is same: the parameter space R∗ is cut into regions (by hyperplanes, etc),
and each region is associated with one X-form, when parameter cross the boundary of the regions,
one X-form moves to another X-form, learning dynamics is conducted on this set of X-form.
This is the learning dynamics of deep learning, this is what really deep learning is doing.
For more complicated deep learning, such as convolution is used, connection pruning is done, nonlinearity is other than sign function (like ReLU), the situation will be more complicated. However,
if there is no human intervention, it will surely be mechanical learning. We still can prove that the
learning dynamics is the same: the parameter space R∗ is cut into regions, and each region is associated with one X-form, when parameter cross the boundary of the regions, one X-form moves to
another X-form, learning dynamics is conducted on this set of X-form.
To prove this for general deep learning mathematically, additional works are needed. We will do this
work in other place. But, we have no doubt this can be done.
Such a learning strategy is exactly what we described in [2]: ”Embedded X-forms into parameter
space”.
8
5
What really is deep learning?
Remark
Now, we know the fact: deep learning is using strategy of ”Embedded X-forms into parameter space”.
This fact is very essential and many consequences can be derived from it. Here we make some
comments.
True nature of deep learning:
On surface, deep learning seems build a model from data feed into eventually (by using neural network,
stacking RBMs, and more other tools). However, as we reveal in previous sections, it is not such a
case. Essentially, a deep learning model is doing this: at the time of model setting up, to cut the
huge parameter space R∗ into many regions, and each region is associated with one X-form that is
one logical statement, then driven by big amount of data, following certain learning dynamics, i.e. to
move from region to another region, which is equivalent to move from one X-form to another X-form,
and eventually to reach a satisfactory X-form, which is the learning result.
So, we say that deep learning is not to building up a model from data input, but it is to choose a
good X-form from a set of X-forms established at the time deep learning model is set up. This is the
true nature of deep learning.
Such a view is different than popular view about deep learning. However, it is true and help us to
understand deep learning better. For example, [6] might be right, there are some group renormalization
going on, but, it missed this issue: group renormalization happens at the setting stage not at the
learning stage. Another example, [5] gives a very good explanation about the power of multi-stagecomposition. However, it failed to realize that learning is not only to get a good processing, but to
find a best possible X-form for the good processing (since one particular processing could have many
associated X-form, and some of such X-form is bad, some of such X-form is good).
Fundamental limitation of deep learning:
The fundamental limitation of a deep learning model is from its nature: it acts on a pre-selected set
of X-forms X, which is formed at the time the model is set up.
So, mostly likely, a deep learning model could not be an universal learning machine [2]. If it is
universal, X must contains at least one X-forms for any possible processing. This is nearly impossible.
Actually, a deep learning model is set up by human, and is for one particular task. So, most likely, X
only contains X-forms specially for this task. And, this deep learning model is limited by X.
If the learning target is for a particular processing, and if in X there is at least one X-form associated
to this processing, a deep learning model could possible to get the target. Otherwise, a deep learning
model could not reach the processing, no matter how hard to try and how much data. In another
word, the model is a bad model. But, deep learning has no any method to tell if a model is good or
bad before trying it out. This is one huge limitation.
Yet, even X contains a X-form associated with the desired processing, we still do not know if the
X-form is a good one. As we know in [2], there are many X-forms associated with one processing,
some is bad, and some is good. For example, one X-form is more robust for certain conditions. If X
does not contain the robust X-form, no matter how hard to try and how much data, learning could
not get a robust solution. Again, deep learning has no any method to tell if X contains such X-form.
.
These limitations are fundamental and are derived from the fact: deep learning is chosen X-form from
a pre-selected set X, not dynamically building a X-form from input data.
Logical or Probabilistic:
Quite often, people think deep learning is doing probability learning. They think that a probability
model is essential for deep learning since a lot data feed in, specially, stochastic gradient descent is one
very essential learning method. However, we would like to point out: deep learning fundamentally is
viewing its learning target logically. Why? Each X-form in X is a logical statement, often a very long
logical statement (so called deep). So, the very essential thing is: when a deep learning model does its
information processing, it is doing according to a solid logical statement. Not doing the information
processing probabilistically.
Of course, the way to get the X-form might not be pure logical, it could involve a lot of probabilistic
views and actions, such as stochastic gradient descent. However, we would point out, even the way
to get the X-form, could be pure deterministic not probabilistic. It is possible to design a pure
deterministic learning dynamics, at least in theory.
Chuyu Xiong
9
Why works well:
Practice shows, deep learning works very well for many problems. Now, we can see the reason for
such success clearly: when a deep learning model for one problem is set up by experienced people,
the desired X-form is already build into the model. More precisely, if the desired X-form is X, when
deep learning model is set up, we have X ∈ X, where X is the pre-selected set of X-forms. If so, it is
possible to learn the X-form X successfully, so the processing associated with X. Thus, the success
of a deep learning model depends on its set up. Having a good set up, the model will work well.
Otherwise, no matter what data and what efforts, the model will not work well.
Of course, besides the set up of model, learning methods are crucial. It is not easy to choose the right
X from X at all. We would like to point out the methods currently used indeed have some advantages:
1. Methods act on Euclidean space, which is most easy to calculate, with many sophisticated
algorithms, library, packages, and hardwares available.
2. Methods are mostly doing linear algebraic calculus, which are easy to be parallelized. And, high
parallelization is the key of its success. However, this advantage is build on this fact: no dynamic
adaption. If dynamic adaption is used (such as recently introduced Capsule), this advantage
might be lost.
Data for deep learning:
As discussed above, the logical statement (X-form) is the core of deep learning. Without supporting
data, a deep learning model could not reach a sophisticated logical statement (X-form). We defined
data sufficiency in [2], which tells what kind of data can support one X-form (logical statement).
Of course, the data sufficiency we defined is only the first step to understand data. Since deep learning
approaching the desired X-form by some learning methods, it is easy to see that we need more data
than just sufficient to support one X-form. The relationships here could be quite complicated, which
would be the topic for further research.
However, we can tell that data sufficient to support and sufficient to bound the desired X-form is
the necessary condition for deep learning. In this sense, for so called big data for deep learning, we
understand the necessity and lower bound.
Disadvantages of deep learning:
Deep learning has some fundamental disadvantages from its root. We list some of them below:
1. It is acting on huge parameter space, but, actual learning dynamics is on a fixed set of regions
(which is equivalent to a set of X-forms). This indirectness makes every aspects of learning
harder, especially, it is near impossible to know what is exactly happening in learning dynamics.
2. Successful learning needs data sufficient to support and to bound. This is very costly.
3. The structure of learning is setup by human. Once setup, structure (how many layers, how big
a layer is, how layers fitting together, how to do convolution, etc) is not be able to change. This
means that learning is restricted on a fixed group of regions, equivalent a fixed group of X-forms.
If best X-form is not in this set, deep learning has no way to reach best X-form, no matter how
big data are and how hard to try. Consequently, it is not universal learning machine.
4. It is very costly to embed X-forms into a huge parameter space. Perhaps, among all computing
spend on learning, only a very small fraction is used on critical part, i.e. moving X-form to
another, and most are just wasted.
5. Since there is no clear and reachable internal representation (due to the embedding), it will be
very hard to do advanced learning, such as to unite all 5 learning methods together (see [3], [2]).
6. Since there is no clear internal representation space, it is hard to define initial X-form, which is
very essential to efficiency improving and several stages of learning.
Looking forward to universal learning machine:
Since deep learning model is not universal learning machine, naturally, we looking forward to universal
learning machine. We discussed this in [1] and [2]. There, we proved that with certain capabilities,
we can make universal learning machine. Also, we actually invented a concrete universal learning
machine which is in the patent application process. We think universal learning machine has many
advantages over deep learning model. There are many research works needed to be done for universal
learning machine.
10
What really is deep learning?
References
[1] Chuy Xiong. Discussion on Mechanical Learning and Learning Machine, arxiv.org, 2016.
http://arxiv.org/pdf/1602.00198.pdf
[2] Chuy Xiong. Descriptions of Objectives and Processes of Mechanical Learning, arxiv.org, 2017.
http://arxiv.org/abs/1706.00066.pdf
[3] Pedro Domingos. The Master Algorithm, Talks at Google.
https://plus.google.com/117039636053462680924/posts/RxnFUqbbFRc
[4] E. G. Hinton. Learning multiple layers of representation, Trends in Cognitive Sciences, Vol. 11,
pp 428-434.. http://www.cs.toronto.edu/ hinton/absps/tics.pdf
[5] Henry W. Lin, Max Tegmark, David Rolnick. Why does deep and cheap learning work so well?,
arxiv.org, 2016.
http://arxiv.org/pdf/1608.08225.pdf
[6] Pankaj Mehta, David J. Schwab, David Rolnick. An exact mapping between the Variational
Renormalization Group and Deep Learning, arxiv.org, 2014.
http://arxiv.org/pdf/1410.3831.pdf
| 9cs.NE
|
Bayesian Nonparametric Inference for M/G/1 Queueing
Systems
Moritz von Rohrscheidt∗1 and Cornelia Wichelhaus
†2
arXiv:1703.07072v1 [math.ST] 21 Mar 2017
1 Ruprecht-Karls
2 Technische
Universität Heidelberg
Universität Darmstadt
March 22, 2017
Abstract
In this work, nonparametric statistical inference is provided for the continuous-time
M/G/1 queueing model from a Bayesian point of view. The inference is based on observations of the inter-arrival and service times. Beside other characteristics of the system,
particular interest is in the waiting time distribution which is not accessible in closed
form. Thus, we use an indirect statistical approach by exploiting the Pollaczek-Khinchine
transform formula for the Laplace transform of the waiting time distribution. Due to this,
an estimator is defined and its frequentist validation in terms of posterior consistency and
posterior normality is studied. It will turn out that we can hereby make inference for the
observables separately and compose the results subsequently by suitable techniques.
Keywords: Bayesian Statistics, Nonparametric Inference, Queues in Continuous Time
1
Introduction
Much attention has been drawn to the analysis and the statistics of queueing systems. Beside approaches to statistical inference for queueing systems in the classical frequentist sense,
during the last three decades there has taken place mentionable research in what is frequently
called Bayesian queues. By that we mean inference for queueing systems from the Bayesian
point of view. Bayesian statistical approaches are often feasible and useful in the area of
Operations Research since one is able to express prior knowledge about the system which is
often present, see also e.g. the foreword by Dennis Lindley in De Finetti (1974). It is worth
to mention here some of the recent works on Bayesian queues. Armero and Bayarri (1994)
deals with Markovian queues in the Bayesian sense, i.e. queues with a homogeneous Poisson
process as input stream and exponential service times with varying number of service stations.
In this parametric model, the parametrically conjugate Gamma prior is put on the parameter
space which contains the arrival and service parameter with the aim to evolve a Bayesian test
for the stability of the queueing system.
∗
†
[email protected]
[email protected]
1
This work forms the base for further works on Bayesian inference for systems with Markovian
characters under several generalizations cf. e.g. Armero and Conesa (1998, 2004, 2006) and
Ausín et al. (2007, 2008).
Generalizations in many ways were performed as for example a non-Markovian setting for
the inter-arrival times, cf. Wiper (1998) and Ausín et al. (2007, 2008). Assuming the arrival
stream of customers to be Poisson is often well-suited and ensures desirable properties of the
stochastic process in the background of the system, meaning that there is a description in
terms of an embedded Markov chain. In contrast, the service times often have to be modeled
to be governed by a more general distribution. Moreover, more flexible classes of service time
distributions often preserve the aforementioned property. Thus, it is reasonable to generalize
the distribution of the service times which, in the classical parametric treatment, is taken to
be exponential. Models with generalizations of the service time distribution can be found in
Insua et al. (1998), where Erlang and hyperexponential distributions are employed to model
the service times more flexibly. Ausín et al. (2004) use the well known probabilistic result
that the class of phase-type distributions forms a dense set in the space of all probability
distributions [cf. Asmussen (2008, p.84)]. They model the service time distribution semiparametrically, assign prior distributions to its parameters for eventually inferring the system
by MCMC procedures.
However, one is often interested in a even more general approach for modeling the service time
distribution which leads to a non-parametric Bayesian approach to estimate the unknown service time distribution. The first work in a discrete-time framework concerning this was given
by Conti (1999) for a Geo/G/1 system which can be seen to be the discrete-time analog of the
continuous-time M /G/1. Geo/G/1 is frequently used to model communication systems where
information encoded in packages of fixed size is transmitted by a serving unit. The server is
assumed to be able to transmit one package during one time slot. Hence, the randomness of
the service times can rather be seen to be given in form of random batch sizes, i.e. marks of
the point process governing the arrival stream. Employing a Dirichlet process prior for the distribution of the magnitude of these marks, Conti obtains estimators for various characteristics
of the queue by a well known functional relationship [see also e.g. Grübel and Pitts (1992)]
of the inter-arrival and the service time distribution with the waiting time distribution. Since
the target of Conti’s work was to obtain estimates for the waiting times of customers, he took
this rather indirect approach mainly because assigning a prior distribution to the waiting time
distribution and updating it by data consisting of the marked arrival stream is infeasible. He
obtained large sample properties for the estimator of the probability generating function of
the waiting time distribution. These were uniform posterior consistency and a Bernstein-von
Mises type result. The former roughly says that the posterior distribution, i.e. the prior updated by the data, will center around the true value while the latter forms a Bayesian analog
to the central limit theorem and gives an idea how this centering looks like. For the Bernsteinvon Mises theorem results in Freedman (1963) were employed.
The aim of the present work is to take a similar way of Bayesian statistical inference for the
continuous-time M /G/1 system which is the more appropriate model in many situations where
time-continuity is more reasonable. For example one could think of customers arriving at a
cashier in a supermarket, cars arriving at a traffic jam, goods arriving at a storing center and
many others. Therefor, a continuous-time analog for the functional relationship of observables
and objects of interest is used which is known as the functional Pollaczek-Khinchine formula
in honor of Felix Pollaczek and Aleksandr Khinchine who derived the steady state behaviour
2
of the M /G/1 system in the 1930s. Thereby, the philosophy is analog to that of Conti (1999),
meaning that the main target will be the nonparametric estimation of the waiting time distribution on basis of observations of the arrival stream and the service times. However, a richer
class of prior distributions is used which allows to express prior knowledge more flexibly. We
feel that this is more reasonable for a system evolving in continuous time. The combination of
this larger class of prior distributions together with the assumption of continuous time leads
to more intricate proofs of results related to the discrete-time analog. As usual, the Bayesian
approach is appreciable if prior knowledge and only few data is available. However, since
typically large data samples are accessible, large sample properties as posterior consistency
and posterior normality are investigated.
The paper is organized the following way. In section 2 the underlying queueing model is
introduced and the functional relationship between observables and characteristics of interest
is briefly described. Section 3 is devoted to the assignment of prior distributions to the
random distributions governing the observables. Therefor, we review some facts of Bayesian
statistics and call to mind a large family of non-parametric priors for eventually obtaining
suitable estimators. Subsequently, frequentist validations of the suggested estimators is given
in section 4 and section 5. While section 4 deals with the concentration of the posterior law
of the random quantities in form of posterior consistency results, section 5 is devoted to the
question how these concentrations take place yielding Bernstein-von Mises type results.
3
2
The Queueing Model
The present section is devoted to the underlying queueing model. The model is chosen, in
Kendall’s notation [Kendall (1953)], to be the M /G/1/∞-FIFO system. This notation means
that (M) indistinguishable customers arrive consecutively according to a homogeneous Poisson
process. Here, the M stands for "Markovian" to depict the memorylessness property of the
exponential distribution which governs the independent inter-arrival times (An )n∈Z . Note that
most of the appreciable properties of the M /G/1 system are due to this assumption. Moreover,
customers requiring service are served according to a general service time distribution (G)
concentrated on R+ . The service times are assumed to form a sequence of i.i.d. random
variables (Sn )n∈Z . The consecutive services are accomplished by one (1) reliable service station
in a first-in-first-out (FIFO) manner. Customers who cannot be served immediately are stored
in an infinitely large waiting-room (∞) and thus form a queue which, provided the queue is
stable and in steady state, is distributed according to a stationary distribution. For the sake
of clarity of what follows, we introduce the following notation. We hereby assume that all
random variables are defined on a common underlying probability space (Ω, A, P).
• Qt ≡ queue length at time t ∈ R
• Nt ≡ number of customers in the entire system at time t ∈ R,
i.e. Nt = Qt + 1 in case Qt > 0 for t ∈ R
• Wn ≡ waiting time of customer n ∈ Z
• Dn ≡ sojourn time in the system of customer n ∈ Z
• An ≡ time between arrival of n-th and (n − 1)-st customer n ∈ Z
• Sn ≡ service time of customer n ∈ Z
Letting λ denote the arrival rate of the Poisson arrival stream, µ the mean of the amount of
time the server needs to complete the service of a customer and defining the so-called traffic
intensity ρ ∶= E[S]/E[A] = λµ, a well known assumption for the existence of a stationary
distribution for N is ρ < 1. If the queue is stable, by stationarity it is common to denote above
quantities without an index tacitly assuming that the system has already run for an infinitely
large time-horizon. If additional one assumes E[S 2 ] < ∞ it can be shown that
E[N ] = ρ +
ρ2 + λVar[S]
.
2(1 − ρ)
This formula is known as the Pollaczek-Khinchin mean formula and relates the mean of the
distribution of N to that of A, S and to the variance of S, Var[S]. However, this formula only
relates parameters of aforementioned distributions to each other instead of the entire distributions. However, in Bayesian statistics one is typically interested in the entire distribution of a
parameter one is uncertain in. A functional relationship is given by the Pollaczek-Khinchine
k
transform formula. To state it, define the following objects. Let n(z) = ∑∞
k=0 z P(N = k) be
∞
the probability generating function (p.g.f.) of the distribution of N and g(z) = ∫0 e−zs dG(s)
the Laplace-Stieltjes transform (LST) of the service time distribution. Then the following
functional relationship can be obtained, see e.g. Haigh (2004) or Nelson (2013).
n(z) = g(λ(1 − z))
(1 − z)(1 − ρ)
,
g(λ(1 − z)) − z
z ∈ [0, 1].
4
Manipulations of this formula give
q(z) =
(1 − z)(1 − ρ)
,
g(λ(1 − z)) − z
and
z ∈ [0, 1],
w(z) =
z(1 − ρ)
,
z − λ(1 − g(z))
z ∈ R+ ,
where q(z) denotes the p.g.f. of the queue-length distribution and w(z) the LST of the waitingtime distribution. The latter two quantities are of special statistical interest since they provide
the essential information about the development of the queue.
5
3
Prior Assignments and Estimators
All the transforms n(z), q(z) and w(z) introduced in section 2 depend on the arrival rate λ
and on the LST of the service time distribution g(z). These values are typically assumed to
be unknown and thus need to be inferred. For this inference, we choose a Bayesian approach
which is further introduced in the present section. Since it does not make sense to a subjectivist statistician to talk about "fixed but unknown" parameters [see De Finetti (1974)],
another philosophical concept is used. That is one interprets the "fixed but unknown" parameter as a random quantity itself. This approach leads to the concept of exchangeability which
provides a "meaningful, observable character" of the data. This approach is briefly reviewed.
Let (Ω, A, P) be an abstract probability space. Call an infinite sequence of random variables
(Xi )∞
i=1 with Xi ∶ Ω → R, i ∈ N, exchangeable if for any m ∈ N and any permutation π of m
elements it holds that
(E)
L [X1 , . . . , Xm ] = L [Xπ(1) , . . . , Xπ(m) ] ,
where L denotes the joint law of the respective random object.
Now, assume that we observe n inter-arrival times An1 = (A1 , . . . , An ) between the first
(n + 1) consecutive customers as well as the service times S1n = (S1 , . . . , Sn ) of the first n
customers. The data An1 and S1n are assumed to be the first n projections of two independent
∞
sequences of exchangeable random variables A∞
1 and S1 which, in turn, are assumed to be
independent of each other, i.e. L(Si ; i ∈ I∣Aj ; j ∈ J) = L(Si ; i ∈ I) for all finite subsets I, J ⊂ N.
We turn to de Finetti’s theorem for Polish spaces [Hewitt and Savage (1955)]. Let P(S)
denote the space of all probability measures on some Polish space S and consider in particular
P(R) equipped with the topology of weak convergence of measures. This leads to a measurable
space (P(R), BP(R) ) which is itself Polish [e.g. Kechris (1995)]. By the de Finetti theorem
for exchangeables, it holds for all n ∈ N and for all measurable subsets Ai ⊂ R, i = 1, . . . , n that
X1∞ = (X1 , X2 . . . ) is exchangeable if and only if there is a unique mixing measure ν ∈ P(P(R))
such that
n
P(Xi ∈ Ai ; i = 1, . . . , n) = ∫
P(R)
∏ P (Xi ∈ Ai )ν(dP ).
i=1
The right-hand side reflects the equivalent property of the sequence X1∞ being exchangeable
the following way. The data X1∞ are conditionally i.i.d. given some probability measure
P ∈ P(R), in symbols write X1∞ ∣P ∼ ⊗N P . The probability measure P itself is random and
distributed according to the mixing measure ν which is called the prior distribution from a
Bayesian statistical point of view. Moreover, notice that unconditionally the data X1∞ are
in general not independent. Indeed, by a result from Kingman (1978), it is easily seen that
in general data are positively correlated, a fact that makes Bayesian statistics a theory of
statistical prediction and thus convenient for other theories as e.g. machine learning.
For actual applications, however, the prior ν and the integration in de Finetti’s theorem
becomes infeasible in general. This is mainly due to the fact that one has merely a nonconstructive proof of the existence of ν. Another difficulty is that the set P(R) is quite large.
One has mainly two ways to choose to circumvent this problem in applications. The first
is to shrink the support of ν to a reasonable subset of P(R) by an additional "symmetry
6
constraint" on the sequence of random variables, thus, to put additional information on the
data, to boil down above integration to a finite dimensional parameter space. The second is
to model ν non-parametrically by genius probabilistic tools. We use the first approach here
to get a prior distribution for the random arrival rate λ ∶ Ω → R+ and the second to get a prior
distribution for the random service time distribution G ∶ Ω → P(R+ ).
3.1
Arrival Rate
In case of the inter-arrivals an additional symmetry assumption on the law of the infinite
exchangeable sequence A∞
1 will give rise to a mixing measure which is supported by all exponential distributions. This additional symmetry condition is stated as follows. Let n ∈ N
and for i = 1, ..., n, Bi ∈ BR+ , the Borel sigma-field on R+ . Furthermore, let ci ∈ R be real
constants such that ∑ni=1 ci = 0 and Ci = ci + Bi = {ci + r ∶ r ∈ Bi } ⊂ R+ , i = 1, . . . , n. The law of
the exchangeable sequence A∞
1 fulfills the symmetry condition
P(Ai ∈ Bi ; i = 1, ..., n) = P(Ai ∈ Ci ; i = 1, ..., n),
(S)
for all n, Bi , Ci as above if and only if it is a mixture of exponential distributions [Diaconis
and Ylvisaker (1985)], in symbols
n
P(Ai ∈ Bi ; i = 1, ..., n) = ∫ ∏ P (Ai ∈ Bi )ν(dP )
E
i=1
n
= ∫ ∏∫
R+ i=1
Bi
λe−λxi dxi ν̃(dλ),
where E denotes the space of all exponential distributions on R+ and ν̃ denotes the pushforward measure of ν along the natural parametrization ˜∶ E → R+ ; P ↦ λ of the exponential
distributions. Moreover, if in above situation additionally it holds E[A2 ∣A1 ] = αA1 + β for real
constants α, β > 0, then the mixing measure ν̃ can be shown to be a Gamma distribution.
Since the M /G/1 queueing model implies a mixture of exponential distributions for the
joint law of the inter-arrivals, from a Bayesian point of view, we assume A∞
1 to meet constraints
(E), (S) and that E[A2 ∣A1 ] = αA1 + β holds as well (note that by exchangeability, this extends
to all random inter-arrival times). The latter directly leads to a conjugate prior for the arrival
rate. To be more precise, if we assume λ to be a random variable distributed according to a
Gamma distribution with hyper parameters a, b > 0, then the posterior distribution given the
data X1n is a Gamma distribution as well with updated hyper-parameters (a + n, b + ∑ni=1 Ai ),
i.e. the family of gamma distributions is closed with respect to exponential sampling. In
summary, we assume the following sampling scheme
λ∣(a, b) ∼ Γ(a, b)
A∞
1 ∣λ ∼ ⊗ E(λ),
N
which leads, through Bayes theorem, to the posterior distribution, the Bayes estimate for
7
squared error loss and the posterior predictive distribution density, respectively, given by
n
λ∣(a, b), An1 ∼ Γ (a + n, b + ∑ Ai )
i=1
a+n
=
b + ∑ni=1 Ai
(a + n) (b + ∑ni=1 Ai )a+n
f (an+1 ∣An1 , (a, b)) =
.
(b + ∑ni=1 Ai + an+1 )a+n+1
EΓ [λ∣An1 , (a, b)]
The latter leads to the predictive value for the next observation
E[An+1 ∣An1 , (a, b)] =
n
1
b
.
∑ Ai +
a + n − 1 i=1
a+n−1
Note that the latter equation again reflects the learning process which does not exist in the frequentistic approach in such an explicit form and that for n = 1 it is given by E[A2 ∣A1 , (a, b)] =
1/aA1 + b/a.
3.2
Service Time Distribution
Since the M /G/1 model does not imply a parametric mixture for the service time random
variables as for the inter-arrivals, assigning a suitable prior is a more difficult task. Not being
able to shrink the support of the mixing measure ν in the de Finetti theorem to a finitedimensional set, we have to choose a prior that supports most of P(R+ ). The common way is
to parametrize P(R+ ) by a reasonable dense subset. This is taken to be the set of all discrete
∞
distributions on R+ , Pd (R+ ) = {P ∈ P(R+ ) ∶ P (⋅) = ∑∞
i=1 wi δxi (⋅); wi ∈ [0, 1], ∑i=1 wi = 1, xi ∈
R+ ; ∀i ∈ N}. It is well-known that the most famous non-parametric prior in Bayesian statistics,
namely the Dirichlet process prior obtained in Ferguson (1973), samples discrete probability
measures with probability one. Moreover, it is known that it has full weak support of all of
P(R+ ) if its base measure has support all of R+ . However, here we will use a slightly more
general family of prior distributions that enables us to model more flexibly prior beliefs of the
true data generating distribution G0 . This larger family will be a subclass of so called neutral
to the right prior processes, namely the beta-Stacy processes. Although these priors sample
discrete probabilities with probability one, too, an analogue of the result concerning the weak
support is known as well. The class of neutral to the right priors is now briefly introduced
and its most important properties will be stated.
3.2.1
Neutral to the right Priors
Let F(R+ ) denote the space of all cumulative distribution functions (c.d.f.) on R+ . Then, a
random distribution function F ∈ ΩF (R+ ) is said to be neutral to the right (NTR) if for each
k > 1 and 0<t1 < t2 < ... < tk the normalized increments
F (t1 ),
F (t2 ) − F (t1 )
F (tk ) − F (tk−1 )
, ...,
1 − F (t1 )
1 − F (tk−1 )
are independent assuming F (t) < 1, ∀t ∈ R+ . That is, for all i ∈ N, F̄ (ti )/F̄ (ti−1 ) is independent
of the sigma-field generated by F up to time ti−1 , σ({F (t) ∶ t < ti−1 ), where F̄ (⋅) ∶= 1 − F (⋅) is
the survival function associated to F . This essentially asserts that the proportion of mass that
8
F assigns to (ti , ∞) with respect to (ti−1 , ∞) does not depend on how F behaves left of ti−1 .
This property coined the name neutral to the right. Doksum (1974) has shown that F (⋅) ∈ ΩF
is NTR if and only if L(F (⋅)) = L(1 − exp[−A(⋅)]) for some independent increment process
A(⋅) which is almost surely non-decreasing, right continuous and such that limt→−∞ A(t) = 0
and limt→∞ A(t) = ∞. Such objects are called increasing additive processes, see e.g. Sato
(1999). For more details on the construction of NTR priors see e.g. Phadia (2015). Since
independent increment processes are well understood, the definition of NTR priors leads to
a rich class of non-parametric priors which are analytically tractable. Another nice feature
of NTR priors is that this family is conjugate with respect to (right-censored) exchangeable
data. A fact that makes NTR priors appreciable in statistical survival analysis. Notice, that
a Dirichlet process prior updated by right censored data is not longer Dirichlet.
The next proposition makes a statement about the weak support of a NTR prior. Recall
that the topological support of a measure is the smallest closed set with full measure. In the
following theorem F(R+ ) is identified with the space of all probability measures P(R+ ) which
is equipped with the weak topology.
Proposition 3.1. Dey et al. (2003)
Let F be a random distribution function which is governed by a NTR prior Π ∈ P(F), i.e.
F ∼ Π, and let A(⋅) = − log(1 − F (⋅)) be the corresponding positive increasing additive process
with Lévy measure L. Then, Π has full support if L has full support. The assertion remains
true if F(R+ ) is equipped with the sup-norm.
Proof. It has to be shown that all weak neighborhoods of any probability measure have positive Π-mass. Since continuous distributions are dense in P(R+ ) with respect to the weak
topology, it suffices to show the assertion for all weak neighborhoods of continuous distributions. Now, choose some random distribution function F0 and some > 0 and consider
U ∶= {F ∶ sup ∣F (t) − F0 (t)∣ < }. Then [e.g. Ghosh and Ramamoorthi (2003))] W ⊆ U for
0≤t<∞
some weak neighborhood W of F0 . Moreover, there is δ > 0, m ∈ N and 0 < t1 < t2 < ⋯ < tm
such that {F ∶ ∣F (ti , ti+1 ] − F0 (ti , ti+1 ]∣ < δ, i = 1, . . . , m} ⊆ W . Hence, it is enough to show
that sup-neighborhoods restricted to compact sets possess positive prior probability. By the
homeomorphism φ ∶ F (⋅) ↦ − log[1 − F (⋅)] the problem can be translated to the analogous one
for the corresponding Lévy process. That is to show that L(A) gives positive probability to
sets of the form C ∶= {A ∶ sup ∣A(t) − φ−1 F0 (t)∣ < γ} for fixed r ∈ Q. Now, take a partition
0≤t≤r
ρ = ∪ki=1 (ai , ai+1 ] of (0, r] such that sup φ−1 F0 (aj , aj+1 ] < γ and define
1≤j≤k
Bi ∶= (ai , ai+1 ] × (φ−1 F0 (ai , ai+1 ] − γ/k, φ−1 F0 (ai , ai+1 ] + γ/k),
B ∶= ∩ki=1 {A ∶ #{(t, A({t}) ∈ Bi } = 1} .
It follows that B ⊆ C and, since L was assumed to have full support,
k
φ−1 Π(C) ≥ φ−1 Π(B) = ∏ L(Bi )e−L(Bi ) > 0.
i=1
9
3.2.2
Beta Processes and Beta-Stacy Processes
For our purposes, we choose a NTR prior with corresponding process Y (⋅) being driven by a
certain class of Lévy measures. These were studied by Hjort (1990) and Walker and Muliere
(1997). Hjort studied beta-processes from a survival analysis viewpoint and therefor elicited
a non-parametric prior for the cumulative hazard function (c.h.f.) given for F ∈ F(R+ ) by
t
H(t) = ∫
0
dF (s)
, t > 0.
F̄ (s)
Walker and Muliere (1997) give the definition of the analog of the beta-process as a prior for
the c.d.f. directly as follows. F is said to be distributed according to a beta-Stacy process
+
with parameters (c(⋅), H(⋅)) ∈ RR
+ × F(R+ ) (for short F ∼ BS(c, H)) if for all t ≥ 0 the
corresponding process Y (⋅) fulfilling F (⋅) = 1 − exp[−Y (⋅)] has Lévy measure
t
dx
−xc(s)(1−H(s))
dLt (x) =
c(s)dHc (s),
∫ e
1 − e−x
0
for all x > 0, where Hc (t) = H(t) − ∑ H(tk ) is the continuous part of H with ti as the fixed
k∶tk <t
points of discontinuity of H. Since we are in a continuous-time framework, we will always
choose H to be continuous. However, discontinuities appear in the Lévy measure governing
the posterior law. Note that EBS [F (⋅)] = H(⋅) is the prior guess on the c.d.f. F and the
function c(⋅) acts like a tuning parameter affecting the magnitude of the increments and is
sometimes interpreted as the "flexible" belief in the prior guess. For the sake of clarity, note
that the Dirichlet process with finite measure α as parameter admits a similar representation
with Lévy-measure
t
dx
−xc(1−ᾱ(s))
dDt (x) =
cᾱ(ds),
∫ e
1 − e−x
0
α(⋅)
where c = α(R) and ᾱ(⋅) denotes the c.d.f. corresponding to the probability measure c . In
the case of Dirichlet processes it is well-known that the prior guess on the random probability
measure equals 1/cα(⋅) and c itself is often interpreted as the strength of belief in the prior
guess. Thus, a Dirichlet process is a beta-Stacy process whose parameters are determined by
α alone.
As already mentioned, the beta-Stacy process is a parametrically conjugate prior, meaning
that the posterior law of F given exchangeable possibly right-censored data is a beta-Stacy
process as well. To be more precise, let S1n be the first n projections of infinite exchangeable
data S1∞ . Let S1∞ be conditional iid given the c.d.f. F and let F ∼ BS(c, H).
n
n
i=1
i=1
Further, define M (t) = ∑ 1[t,∞) (Si ) and N (t) = ∑ 1[0,t] (Si ). Then it holds [Walker and
Muliere (1997, Theorem 4)] that F ∣S1n ∼ BS(c∗n , Hn∗ ), where c∗n (⋅) and Hn∗ (⋅) are given by
Hn∗ (t) = 1 − ∏ (1 −
s∈[0,t]
c∗n (t) =
c(s)dH(s) + dN (s)
),
c(s)H̄(s) + M (s)
c(t)H̄(t) + M (t) − N (t)
.
H¯∗ (t)
10
Thereby, Hn∗ (t) is defined by means of the product integral, see Gill and Johansen (1990).
Note, that the posterior process possesses fixed points of discontinuity at the observations and
that the posterior guess on F , i.e. the Bayes estimate with respect to squared error loss is
given by F̂n (⋅) = Hn∗ (⋅).
Having a look at the Pollazcek-Khinchine transform formulas in section 2, one may ask if
they are well defined for almost all λ and F drawn from their respective prior and posterior
distributions. While the posterior of the mean inter-arrival times is straightforward by definition, the mean and the second moment of the random c.d.f. describing the general service
times deserves more attention. The existence of functionals of c.d.f.s drawn according to a
beta-Stacy process was studied in Epifani et al. (2003). Relating A(⋅) and c(⋅) to the existence
of a certain functional, they obtained sufficient conditions for moments of order m to exist
[equation (10) in their article]. We will assume throughout that this condition holds at least
for the moment of second order. Moreover, note that they obtained an explicit formula for
the prior moments [equation (11)] as well as for the posterior moments [equation (13)].
In summary, we assume the following sample scheme for the service times in the M /G/1
system.
G∣(c, H) ∼ BS(c, H)
S1∞ ∣G ∼ ⊗ G.
N
This leads to the posterior distribution
G∣(c, H), S1n ∼ BS(c∗n , Hn∗ ).
Notice again that this updating continues to hold for right-censored observations. Hence,
the model can be enlarged in that one does not have to keep exact track of the customers
and is still able to make reasonable inference even if e.g. it is solely known that the service of
customers has exceeded a certain threshold.
3.3
Estimators for Queueing Characteristics
In this subsection, we study the Bayes estimators, i.e. the posterior means, of several characteristics of the M /G/1 system. Not all posterior laws of each single characteristic are obtainable
explicitly. Hence, for those characteristics whose posterior laws are not obtainable in closed
form, we define suitable estimators by replacing Bayes estimators for the corresponding values.
Assume for the general service time distribution G ∼ BS(c, H) and let
Ĝn = EBS [G∣S1n ] ,
µ̂n = EBS [∫
λ̂n =
∞
0
EΓ [λ∣An1 ]
tdG(t)∣S1n ] and
be the Bayes estimates with respect to squared error loss. Note that with Mn (s) = ∑nj=1 δXj [s, ∞),
as in Epifani et al. (2003), µ̂n can be given as
µ̂n = ∫
0
∞
t
exp [− ∫
0
t β(s) + M (s) − 1
α(ds)
n
] exp [− ∫
Nn (ds)] dt,
β(s) + Mn (s)
β(s) + Mn (s)
0
11
where if H is continuous [see Phadia (2015)]
• β(t) = c(t)[1 − H(t)]
t
• α(t) = ∫0 c(s)dH(s).
Observe that ρ̂n = EBS⊗Γ [λµ∣S1n , An1 ] = µ̂n λ̂n by above independence assumption. However, in
Bayesian statistics, one is interested in the entire posterior law rather than merely in a certain
functional. The posterior law of ρ is obtainable explicitly in the following form
PBS⊗Γ (ρ ≤ t∣An1 , S1n ) = PBS⊗Γ (µλ ≤ t∣An1 , S1n )
∞
=∫
0
=∫
=
0
∞
PBS⊗Γ (µ ≤ t/λ∣An1 , S1n , λ)PBS⊗Γ (λ∣An1 , S1n )dλ
PBS (µ ≤ t/λ∣S1n , λ)PΓ (λ∣An1 )dλ
∞
(b + ∑ni=1 Ai )a+n
n
a+n−1 −λ(b+∑n
i=1 Ai ) dλ.
e
∫ PBS (µ ≤ t/λ∣S1 , λ)λ
Γ(a + n)
0
Note that PBS (µ ≤ t/λ∣S1n , λ) can be stated more explicitly in form of a density by means of
Proposition 4 in Regazzini et al. (2003). This explicit form is omitted here due to technical
reasons and left to the interested reader. If one observes the queueing system and is, as is
usually the case, not aware of the condition ρ0 < 1, one can thus obtain the posterior probability that the system is stable PBS⊗Γ (ρ < 1∣An1 , S1n ).
We close this section by defining estimates for:
• mean service time: µ̂n
• traffic intensity: ρ̂n = λ̂n µ̂n
∞
• LST of service-time distribution: gn∗ (z) = ∫0 e−zt dĜn (t)
• LST of waiting-time distribution: wn∗ (z) =
• pgf of queue-size distribution: qn∗ (z) =
z(1−ρ̂n )
∗ (z))
z−λ̂n (1−gn
(1−ρ̂n )(1−z)
∗ (λ̂ (1−z))−z
gn
n
• system-size distribution: n∗n (z) = qn∗ (z)gn∗ (λ̂n (1 − z)).
Notice that for the random arrival- and service rate and the random distribution function
of the service times the natural Bayes estimator is used, i.e. the minimizer with respect to
squared error loss. They are analytically tractable and obtainable in closed form. For the
remaining queueing characteristics obvious plug-in estimators are used. The reason therefor
is that a closed form of the push-forward law under the mapping
(λ, G(⋅)) ↦ f (⋅),
where f ∈ {n, g, q}, of the prior ΠBS⊗Γ is not easy to obtain. Neither is known how to update such a pushed prior by exchangeable data (S, A)n1 which we have access to in a natural
Bayesian way.
However, the goodness of these estimates in a rigorous mathematical sense is established
in the next sections.
12
4
Consistency Results
Posterior consistency provides a tool for validation of Bayesian procedures. Roughly speaking, it is defined to be the property of the posterior law to center around the true parameter
when the number of observed data increases. As a consequence of posterior consistency two
different priors will asymptotically lead to the same prediction, a fact often called merging of
prior opinions. So, prior information is consecutively washed away by the new state of information provided by the data. Posterior consistency is the most desired property of Bayesian
procedures since it states that one can recover the true measure that generates the data. For
examples of inconsistent Bayes procedures see e.g. Diaconis and Freedman (1986) and Kim
and Lee (2001) and references therein.
We will have to deal with two different posterior consistency issues, a parametric one for
the posterior of the arrival rate and a non-parametric one for the service time distribution
and values depending on it. The first was considered by Doob (1949) in a rather general way
by the use of martingale theory. It clarifies that, under very weak constraints (the state- and
the parameter spaces are assumed to be Polish spaces and the likelihood is identifiable), there
is a subset of the parameter space with full prior mass such that the sequence of posterior
laws is consistent at any parameter of that subset that is taken as the true one. However,
one problem is that in high-dimensional parameter spaces the set with full prior mass might
become topologically small. Therefore, especially nonparametric Bayesian statistical problems
deserve a deeper study of posterior consistency since the set of possible likelihoods can not
longer be assumed to be parametrized finitely.
First of all, the definition of posterior consistency in a rather general framework is given as
it can be found e.g. in Schervish (1995). Notice that for any Polish state space S the space of
all probability measures, P(S), can be equipped with the topology induced by weak convergence which has neighborhood bases for P ∈ P(S) given by the collection of sets of the form
UP ; = {Q ∈ P(R) ∶ ∣∫ fi dP − ∫ fi dQ∣ < , fi ∈ Cb (R), for all i = 1, .., k}. This topology makes
P(S) itself a Polish space [see Kechris (1995)] and the topology can be metrized e.g. by the
Prohorov metric. The Borel σ-field, B(P), with respect to the weak topology serves as a
natural measure-theoretical structure to turn P(S) into a measurable space (P(S), B(P)).
B(P) turns out to be the smallest σ-field on P(S) that makes the mappings P ↦ P (A)
(B(P), B([0, 1]))-measurable for all A ∈ B(S). In the sequel Π, occasionally with an appropriate index, shall denote a prior distribution and Πn the corresponding posterior distribution
after having seen the first n projections of the exchangeable data. A distribution (or a value
that parametrizes a distribution) indexed by 0 will always denote the true data generating
measure.
Definition 1. Let Π ∈ P(P(R)) be a prior distribution, i.e. the distribution of some random
probability measure P ∈ P(R) and let X1∞ be a sequence of exchangeable data which is conditionally i.i.d. given P . Moreover, let P0 ∈ P(R) be the true data generating distribution and
(Πn )n≥0 = (Π(⋅∣X1n ))n≥0 be the sequence of posterior laws. Then call (Πn )n≥0 weakly consistent
at P0 if for all weak neighborhoods UP0 , of P0 it holds
n→∞
Πn (UP0 , ) Ð→ 1,
for P0∞ -almost all data sequences X1∞ .
13
Here, P0∞ denotes the true joint law governing the sequence X1∞ .
Needless to say that in the case of the arrival rate λ things become easier because, due
to the additional judgment concerning the symmetry of the distribution of the exchangeable
sequence of inter-arrival times, one can reduce the problem to that of parametric consistency
that is rather easy to handle by conjugacy.
Proposition 4.1. Let ΠΓ stand for the Gamma prior of the random arrival rate λ and
(ΠΓ;n (⋅))n≥1 = (ΠΓ (⋅∣An1 ))n≥1 be the sequence of posterior laws. Then, for all > 0 and for Pλ∞0
almost all sequences A∞
1 it holds
n→∞
ΠΓ;n (∣λ − λ0 ∣ < ) Ð→ 1.
Proof. Taking into account that the gamma prior is conjugate for exponentially distributed
data, the posterior density can be shown to be that of a Γ(a + n, b + ∑ni=1 Ai ) distribution. By
well known properties of the Gamma distribution one has EΓ;n [λ] = (a + n)(b + ∑ni=1 Ai )−1 and
VarΓ;n [λ] = (a + n) (b + ∑ni=1 Ai )−2 . Applying the continuous mapping theorem the assertion
follows from the strong law of large numbers and the Markov inequality.
Let us now turn to the nonparametric part concerning the random c.d.f. G. The following
two results are due to Dey et al. (2003) and give the consistency of the Bayes estimator and
the random c.d.f. G.
Lemma 4.2 (Dey et al. (2003)). Let G0 ∈ F(R+ ) be the true c.d.f. of the service times.
Furthermore suppose that G0 is continuous and that G0 (t) < 1 for all t ∈ R+ . Let G ∼ BS(c, H)
and ΠBS;n denote the posterior distribution of G. Then, for G∞
0 -almost all sequences of data
S1∞ and all t ∈ R+ it holds
n→∞
EBS;n [G(t)] Ð→ G0 (t).
Next, the posterior consistency of the random service time distribution G induced by a
beta-Stacy prior will be stated. Since the service time distribution function G(⋅) is seen to be
a random function rather than a random variable, we investigate deviations in the sup-norm.
Theorem 4.3 (Dey et al. (2003)). Let G0 ∈ F(R+ ) be the true c.d.f. of the service times.
Again, suppose that G0 is continuous and that G0 (t) < 1 for all t ∈ R+ . Let G ∼ BS(c, H) and
ΠBS;n denote the posterior distribution of G. Then, for all > 0 it holds
n→∞
ΠBS;n ( sup ∣G(t) − G0 (t)∣ < ) Ð→ 1
0≤t<∞
for
G∞
0 -almost
all sequences S1∞ .
Remark 1. We stress some peculiarity of a certain class of neutral to the right measures. For
a full formal treatment see Dey et al. (2003). As already mentioned in the previous section,
neighborhoods w.r.t. the sup-norm contain some weak neighborhoods. These, in turn, are
given as finite intersections of sets of the form {G ∶ ∣G(t) − G0 (t)∣ < γ}. Thus, it is enough to
show that the posterior mass of such sets converges to one. But since the Lévy measure that
corresponds to the beta-Stacy prior is of the form L(dt, ds) = a(t, s)dsK(dt) for suitable a and
K, the convergence of the expected value of the posterior law to the true c.d.f. already ensures
that the posterior variance vanishes with increasing data size.
14
As an immediate consequence one has the uniform consistency of the Bayes estimator.
Corollary 4.4. Let the conditions of Lemma 4.2 be fulfilled. Then the Bayes estimate
EBS;n [G(⋅)] of the service time distribution is uniformly consistent at the true continuous
service time distribution G0 (⋅), that is
n→∞
sup ∣EBS;n [G(t)] − G0 (t)∣ Ð→ 0,
0≤t<∞
∞
for G∞
0 -almost all sequences S1 .
Proof. By above theorem one has
RRR
RRR
R
RRR
sup ∣EBS;n [G(t)] − G0 (t)∣ = sup RRR∫ (G(t) − G0 (t)) ΠBS;n (dG)RRRRR
RRR
0≤t<∞
0≤t<∞ RR
RRF
R
≤
sup ∣G(t) − G0 (t)∣ ΠBS;n (dG)
∫
{G∶sup∣G(t)−G0 (t)∣≥}
0≤t<∞
+
sup ∣G(t) − G0 (t)∣ ΠBS;n (dG)
∫
{G∶sup∣G(t)−G0 (t)∣<}
0≤t<∞
n→∞
< ΠBS;n ( sup ∣G(t) − G0 (t)∣ ≥ ) + Ð→ .
0≤t<∞
Since > 0 can be chosen arbitrarily small, the proof is completed and the assertion follows.
The next lemma establishes the uniform consistency on R+ of the service time LST in
posterior law.
Proposition 4.5. Let g(z) = ∫ e−sz dG(s) denote the LST of the random service time distribution G possessing a beta-Stacy process prior and g0 (z) = ∫ e−sz dG0 (s) the LST of the
corresponding true data generating distribution. Then, if the constraints of Theorem 4.3 are
fulfilled, it holds
n→∞
ΠBS;n ( sup ∣g(z) − g0 (z)∣ < ) Ð→ 1,
0≤z<∞
∞
for all > 0 and for G∞
0 -almost all data sequences S1 .
Proof. The assertion of the proposition follows from Theorem 4.3 and the continuous mapping
theorem applied to the mapping G(⋅) ↦ g(⋅) which is continuous w.r.t. the sup-norm. Indeed,
take a δ > 0 and let UG0 ,δ be a uniform δ-neighborhood of the true service time distribution
and let G ∈ UG0 ,δ with corresponding LST g. Then, by integration by parts of the Riemann-
15
Stieltjes integral it holds that
sup ∣g(z) − g0 (z)∣ = sup ∣∫
0≤z<∞
0≤z<∞
∞
0
e−sz [G − G0 ](ds)∣
= sup ∣[e−sz [G(s) − G0 (s)]∣s=∞ − [e−sz [G(s) − G0 (s)]∣s=0 − ∫
0
0≤z<∞
∞
= sup ∣∫
0≤z<∞
≤ sup ∫
0≤z<∞
0
∞
0
< δ sup ∫
[G(s) − G0 (s)]de−zs ∣ = sup ∣∫
0≤z<∞
∞
0
∞
[G(s) − G0 (s)]de−zs ∣
[G(s) − G0 (s)]ze−zs ds∣
ze−zs ∣G(s) − G0 (s)∣ ds
∞
ze−zs ds = δ,
0
0≤z<∞
which completes the proof and shows the claim.
Next, the consistency of the estimator gn∗ (z) of the LST of the service time distribution
is studied. On basis of the previous result, the following lemma establishes the uniform
consistency on R+ of the estimator gn∗ (z).
Lemma 4.6. Let gn∗ (⋅), g0 (⋅) as above. Then, under the constraints of Theorem 4.3, it holds
sup ∣gn∗ (z) − g0 (z)∣ Ð→ 0,
n→∞
0≤z<∞
∞
for G∞
0 -almost all sequences of data S1 .
Proof. Let > 0 be arbitrary and let Ĝn (⋅) ∶= EBS;n [G(⋅)]. Then by properties of the RiemannStieltjes integral, one has
sup ∣gn∗ (z) − g0 (z)∣
0≤z<∞
∞
= sup ∣∫
0≤z<∞
0
= sup ∣ [e
e−zs dĜn (s) − ∫
∞
0
−zs
Ĝn (s)]∣
s=∞
0≤z<∞
− [e
−zs
e−zs dG0 (s)∣
s=0
−zs
− [e
= sup ∣∫
0≤z<∞
0
= sup ∣∫
0≤z<∞
∞
0
∞
Ĝn (s)d(e
−sz
∞
Ĝn (s)]∣
∞
)−∫
0
−∫
0
Ĝn (s)de−zs
G0 (s)]∣s=∞ + [e−zs G0 (s)]∣s=0 + ∫
0
−sz
G0 (s)d(e
∞
G0 (s)de−zs ∣
)∣
[Ĝn (s) − G0 (s)]ze−zs ds∣
< sup ∣Ĝn (s) − G0 (s)∣ .
0≤s<∞
Hence, the assertion follows from Corollary 4.4 and the proof is completed.
∞
Since the Bayesian estimate of the random mean of G, i.e. EBS;n [∫0 tdG(t)], is used
to define estimators of several queueing characteristics, its posterior consistency is examined
next. For the most prominent prior process, namely the Dirichlet process a rather general
result is known. This is reviewed briefly. So, let P ∼ Dα be a random probability measure
16
that is distributed according to a Dirichlet prior with finite measure α as parameter. Let
f ∶ S → R be measurable. Then [c.f. Feigin and Tweedie (1989)] ∫ ∣f ∣ dα < ∞ ⇒ ∫ ∣f ∣ dP < ∞
with Dα probability one and EDα [∫ f dP ] = ∫ f dEDα [P ] = ∫ f dα. This fact in combination
with an assumption that the state space is countable makes it easy to show that posterior
consistency of the random measure P induces posterior consistency of the random mean of P .
However, here neither the state space is assumed to be countable nor the random measure is
assumed to possess a Dirichlet prior. Thus the posterior consistency of the random mean of G
which is drawn according to a beta-Stacy process is studied in more depth and an affirmative
result is given.
Lemma 4.7. Suppose G(⋅) is drawn according to a beta-Stacy process with parameters (α, β),
where α is a measure on R+ and β(s) ≥ 1 such that
(1.) ∫R+ [β(s)]−1 α(ds) = ∞
(2.) ∫R+ exp [− ∫0 [β(s)]−1 α(ds)] dt < ∞.
t
Then the posterior expectation of the random mean of G converges G∞
0 -a.s. to the true one.
∞
it
holds
-almost
all
sequences
of
data
X
That is for G∞
1
0
EBS;n [∫
tdG(t)] ∶= EBS [∫
R+
R+
n→∞
tdG(t)∣X1n ] Ð→ ∫
tdG0 (t).
R+
Proof. First of all note that, due to Epifani et al. (2003), the first moment of the random
mean under the prior does exist and is given as
∞
EBS;n [∫
tdG(t)] = ∫ exp [− ∫
R+
0
t
0
⎡
⎤
⎢
α(ds)
β(Xi ) + Mn (Xi ) − 1 ⎥⎥
⎢
]×⎢ ∏
⎥ dt,
β(s) + Mn (s)
⎢{i∶Xi ≤t} β(Xi ) + Mn (Xi ) ⎥
⎣
⎦
where Mn (s) = ∑nj=1 δXj [s, ∞). Thus, letting Nn (s) = ∑ni=1 δXi [0, s), one obtains
EBS;n [∫
∞
tdG(t)]
R+
t
= ∫ exp [− ∫
0
0
t
β(s) + Mn (s)
α(ds)
] exp [− ∫ log (
) dNn (s)] dt.
β(s) + Mn (s)
β(s) + Mn (s) − 1
0
Since β(s) ≥ 1, by elementary properties of the logarithm, it follows
1
β(s) + Mn (s)
1
≤ log (
)≤
,
β(s) + Mn (s)
β(s) + Mn (s) − 1
β(s) + Mn (s) − 1
which in turn implies
∞
t
∫ exp [− ∫
0
0
tdG(t)]
≤ EBS;n [∫
∞
R+
≤ ∫ exp [− ∫
0
t
Nn (ds)
α(ds)
] × exp [− ∫
] dt
β(s) + Mn (s)
0 β(s) + Mn (s) − 1
t
0
α(ds) + Nn (ds)
] dt.
β(s) + Mn (s)
17
such that it remains to show that the bounding terms converge to the mean of the true c.d.f..
From a straight-forward application of the monotone convergence theorem and by continuity
of the exponential function, it follows
∞
lim EBS;n [∫
n→∞
R+
∞
t
= ∫ exp [− ∫
0
0
lim
0 n→∞
0
1/nNn (ds)
] dt
1/nβ(s) + 1/nMn (s)
G0 (ds)
] dt.
1 − G0 (s)
G0 (ds)
1−G0 (s)
Now, note that
t
tdG(t)] = ∫ exp [− ∫
= λ(s)ds = P(s < S ≤ s + ∆s∣S > s), where λ(s) = lim∆s→0
t
t G (ds)
P(s<S≤s+∆s)
∆s[1−G0 (s)]
0
is the cumulative
usually denotes the hazard function. Hence Λ(t) = ∫0 λ(s)ds = ∫0 1−G
0 (s)
hazard until t ≥ 0. On the other hand Λ(t) = − log(1 − G0 (t)). Thus,
t
exp [− ∫
0
t
G0 (ds)
] = exp [− ∫ λ(s)ds] = exp [−Λ(t)] = exp [log(1 − G0 (t))] = 1 − G0 (t),
1 − G0 (s)
0
which in turn completes the proof.
The next lemma establishes the contraction of the mass of the posterior law of the random
mean.
Lemma 4.8. If, in addition to the assumptions of the previous lemma, the parameter (α, β)
meets the condition
√
∫
exp [− ∫
R+
t
[β(s)]−1 α(ds)] dt < ∞
0
then the posterior variance of the random mean ∫ tG(dt) vanishes G∞
0 -a.s. as the size of the
∞
∞
data increases. That is, for G0 -almost all sequences of data X1 it holds
VBS;n [∫
R+
n→∞
tG(dt)] ∶= VBS [∫
R+
tG(dt)∣X1n ] Ð→ 0.
Proof. Due to Lemma 4.7 and the continuous mapping theorem the claim follows if the second
moment of the random mean under the posterior converges a.s. to the square of the mean of
the true distribution function G0 (⋅). From Epifani et al. (2003) the second moment under the
prior law exists and the second moment under the posterior law of the random mean is given
18
by
∞
EBS;n [(∫
2
tdG(t)) ]
0
∞ ∞
s
α(dx)
α(dx)
] exp [− ∫
]
β(x) + Mn (x) + 1
0 β(x) + Mn (x)
r
= 2 ∫ ∫ exp [− ∫
0
r
0
×
∞
r
= 2 ∫ exp [− ∫
0
0
⎛
β(Xj ) + Mn (Xj ) − 1 ⎞
β(Xi ) + Mn (Xi ) ⎞ ⎛
dsdr
∏
∏
⎝{i∶Xi ≤r} β(Xi ) + Mn (Xi ) + 1 ⎠ ⎝{j∶Xj ≤s} β(Xj ) + Mn (Xj ) ⎠
r
α(dx)
β(x) + Mn (x)
] exp [− ∫ log (
) dNn (x)]
β(x) + Mn (x) + 1
β(x) + Mn (x) + 1
0
∞
s
∫ exp [− ∫
r
0
s
α(dy)
β(y) + Mn (y)
] exp [− ∫ log (
) dNn (y)] dsdr
β(y) + Mn (y)
β(y) + Mn (y) − 1
0
Hence, by similar arguments as in the proof of the previous lemma and using Fubini’s
theorem, it follows that for G∞
0 −almost all sequences of data
∞
EBS;n [(∫
0
2
∞
n→∞
r
tdG(t)) ] Ð→ 2 ∫ exp [− ∫
0
0
∞
s dG (y)
dG0 (x)
0
] ∫ exp [− ∫
] dsdr
1 − G0 (x)
0 1 − G0 (y)
r
∞ ∞
r
= 2 ∫ ∫ exp [− ∫
0
0
r
s dG (y)
dG0 (x)
0
] exp [− ∫
] dsdr.
1 − G0 (x)
0 1 − G0 (y)
(#)
Furthermore, a straight-forward application of Fubini’s theorem yields
∞
∞ ∞
s
∫ ∫ g(r, s)drds = ∫ ∫ g(r, s)dsdr,
0
0
r
0
where g is an arbitrary integrable function g ∶ [0, ∞)2 → R+ . Now, let f (r, s) denote the
integrand of the last integral of equation (#). Since f is symmetric, i.e. f (r, s) = f (s, r), it
follows from above equality and further application of Fubini’s theorem
∞ ∞
∞
∞ ∞
s
2 ∫ ∫ f (r, s)dsdr = ∫ ∫ f (r, s)drds + ∫ ∫ f (r, s)dsdr
r
∞ r
0
0
0
∞ ∞
r
∞ ∞
0
= ∫ ∫ f (s, r)dsdr + ∫ ∫ f (r, s)dsdr = ∫ ∫ f (r, s)drds.
0
0
0
r
0
0
Therefore,
∞
EBS;n [(∫
0
2
n→∞
tdG(t)) ] Ð→ (∫
0
∞
2
tdG0 (t)) ,
which completes the proof of the assertion.
19
Theorem 4.9. Under the assumptions of Lemma 4.7 and Lemma 4.8 the mean of the random
c.d.f. G possesses the property of posterior consistency. That is for all > 0 and G∞
0 -almost
all sequences of data it holds
ΠBS (∣∫
n→∞
tG(dt) − ∫
R+
R+
tG0 (dt)∣ > ∣X1n ) Ð→ 0.
Proof. This is a direct consequence of the two previous lemmas and the Markov inequality.
Remark 2. A brief discussion of the assumptions of above lemmas and theorem, respectively,
is given. Assumption (1.) of Lemma 4.7 is an artifact of survival analysis and ensures that the
prior (α, β), that is to be chosen, actually leads to a cumulative hazard function. Assumption
(2.) of the first lemma and the additional assumption of the second lemma, respectively,
ensures the existence of the first, resp. second, moment of the posterior distribution of the
random mean. These conditions are given in Epifani et al. (2003) Proposition 4.
Moreover, this work extends results which concern the existence of certain random functionals w.r.t. to a random c.d.f. drawn according to a NTR prior. Those results can be
understood as a generalization of a work by Feigin and Tweedie (1989) which gives conditions
under which certain functionals of a random measure drawn according to a Dirichlet prior
exist in terms of the base measure (the prior parameter). At this place it should be emphasized
that they investigated this problem by creating a new approach to the Dirichlet process which is
often not mentioned in the literature. To be more precise, they show that the Dirichlet process
can be extracted as the invariant distribution of a measure-valued Markov chain and exploit
this theory to show sufficiency of the conditions of their existence theorem. Since Dirichlet
priors are as well included in the family of NTR priors the extension by Epifani et al. seems
natural.
The last assumption, i.e. that β(s) ≥ 1 ensures posterior consistency of the random mean
of the c.d.f. governed by the beta-Stacy prior and is used in the proofs of both lemmas. Hence,
a prior for the random c.d.f. G is to be chosen, in terms of its parameter, such that the
sequence of posteriors of the G-mean is consistent as long as this property is desired. It is
interesting to see how advanced information on the posterior consistency of the random mean
influences the prior knowledge of the random c.d.f. itself. To put it another way, information
concerning the posterior consistency of the random mean is indeed prior information on G.
However, the other way around is not true in general, i.e. posterior consistency of the random
c.d.f. does not generally imply posterior consistency of the random mean. That is because this
functional is not continuous. However, posterior consistency of G implies that of its truncated
mean. That means integration is restricted to a compact set and would lead to a weaker form
of consistency in the sense of compact convergence.
As an immediate consequence of the previous results one obtains the consistency of the
traffic intensity of the M /G/1 queue.
Corollary 4.10. Let µ ∶= ∫ sG(ds), µ0 ∶= ∫ sG0 (ds) be the random and true mean of the
service-time distribution and ρ0 = λ0 µ0 the true traffic intensity. Further, let ΠBS⊗Γ =
ΠBS ⊗ ΠΓ be the prior on (P(R+ ), R+ ) that is formed by taking the product-measure of ΠBS
and ΠΓ , ΠBS⊗Γ;n the posterior, respectively, and define λ̂n ∶= Eλ;n [λ] ∶= EΓ [λ∣An1 ], µ̂n ∶=
∞
∞
EBS;n [∫0 tdG(t)] ∶= EBS [∫0 tdG(t)∣S1n ] and ρ̂n ∶= EBS⊗Γ;n [ρ] ∶= EBS⊗Γ [ρ∣An1 , S1n ].
Then, under the assumptions of Proposition 4.1 and Theorem 4.9, one has
20
(i) for all > 0
n→∞
ΠBS⊗Γ (∣ρ − ρ0 ∣ ≥ ∣S1n , An1 ) Ð→ 0,
∞
∞
for Pλ∞0 ⊗ G∞
0 -almost all data sequences (A1 , S1 ),
(ii)
n→∞
∣ρ̂n − ρ0 ∣ Ð→ 0,
∞
∞
for Pλ∞0 ⊗ G∞
0 -almost all data sequences (A1 , S1 ).
Proof. (i) Recalling the posterior consistency of the arrival rate λ in Proposition 4.1 and the
service rate µ in Theorem 4.9 one has
ΠBS⊗Γ;n (∣ρ − ρ0 ∣ ≥ ) = ΠBS⊗Γ;n (∣µλ − µλ0 + µλ0 − λ0 µ0 ∣ ≥ )
≤ ΠBS⊗Γ;n (∣µλ − µλ0 ∣ + ∣µλ0 − λ0 µ0 ∣ ≥ )
≤ ΠBS⊗Γ;n (∣µλ − µλ0 ∣ ≥ /2) + ΠBS⊗Γ;n (∣µ − µ0 ∣ ≥
)
2λ0
≤ ΠBS⊗Γ;n ((µ0 + δ)∣λ − λ0 ∣ ≥ /2, ∣µ − µ0 ∣ < δ)
+ ΠBS;n (∣µ − µ0 ∣ ≥ δ) + ΠBS;n (∣µ − µ0 ∣ ≥
≤ ΠΓ;n (∣λ − λ0 ∣ ≥
2(µ0 + δ)
)
2λ0
) + ΠBS;n (∣µ − µ0 ∣ ≥ δ) + ΠBS;n (∣µ − µ0 ∣ ≥
),
2λ0
from which the assertion of (i) follows.
(ii) Straightforward one has a.s.
∣ρ̂n − ρ0 ∣ ≤ ∣µ̂n λ0 − µ̂n λ̂n ∣ + λ0 ∣µ̂n − µ0 ∣
≤ (µ0 + γ)∣λ0 − λ̂n ∣ + λ0 ∣µ̂n − µ0 ∣
by selecting γ > 0 such that a.s. ∣µ̂n − µ0 ∣ < γ, for all n sufficiently large. The assertion of (ii)
then follows from the proof of Proposition 4.1 and Lemma 4.7 completing the proof.
We are now in a position to state the main theorem of this section, i.e. the consistency of
the estimators for the waiting time LST, the queue length p.g.f. and the system size p.g.f.,
defined in section 3.3.
Theorem 4.11. Under the assumptions of Proposition 4.1, Proposition 4.5 and Theorem 4.9
one has
(i) for all > 0
n→∞
ΠBS⊗Γ ( sup ∣f (z) − f0 (z)∣ ≥ ∣S1n , An1 ) Ð→ 0.
0≤z<∞
(ii)
sup ∣fn∗ (z) − f0 (z)∣ Ð→ 0,
n→∞
0≤z<∞
for
Pλ∞0
∞
∞
⊗ G∞
0 -almost all data sequences (A1 , S1 ), where f (⋅) ∈ {n(⋅), q(⋅), w(⋅)}.
21
Proof. We show the result for f = w, that is in the case of the p.g.f. of the waiting-time
distribution. The other cases are treated similarly and we omit the details. We intent to
apply the continuous mapping theorem. Therefor note that the mapping (g(⋅), λ, µ) ↦ w(⋅)
is continuous with respect to the suitable topologies induced by the sup-norm. Hence, the
assertion of the theorem is implied by the results of the present section.
We shall stress that above theorem is important, especially for applications, since it states
that one can reduce the difficult task of making Bayesian inference for queueing characteristics
to that of making inference for the observables separately in order to compose it subsequently
to that of the objects of interest. Qualitatively this will lead to reasonable asymptotic inference. This is appreciable because it is a non-feasible task to place a prior on the distributions of
aforementioned characteristics in a way that it can be updated with data given by observations
that we have access to.
22
5
Posterior Normality
Asymptotic normality results, so called Bernstein-von Mises theorems, for the posterior law
serve as an additional validation of Bayesian procedures. Especially in situations where the
exact posterior law is not available or hard to compute they are useful from a rather applied
viewpoint to get an approximation of the posterior law. While the previous section has shown
that the posterior concentrates around the true value when data increases, normality results
give an idea how, asymptotically speaking, it does concentrate and how fluctuations around
this centering appear. Often it is true that the posterior law, if centered and rescaled appropriately, resembles a centered Gaussian distribution with a certain covariance structure. For
applications, special interest lies in this limiting covariance structure.
The earliest result dates back to de Laplace (1774) who approximates the posterior of
a beta distribution by a normal integral. Sergei Bernstein and Richard von Mises gave a
rather modern version of this approach by applying a general result about infinite products
of functions to the sequence of posterior densities, see e.g. Johnson (1967) for a review and
references. These approaches of expanding posterior densities coined the name. Nowadays,
the parametric case is well understood and results including centering with the MLE or the
Bayes estimate as well can be found in Schervish (1995) or Ghosh and Ramamoorthi (2003).
However, the results needed for an asymptotic behavior of the posterior of infinite-dimensional
parameters, i.e. in the nonparametric case, go deeper and deserve separate investigation. Posterior normality results involving Dirichlet process priors or NTR processes can be found in
Conti (1999) who employs a result by Freedman (1963) on asymptotic normality in the finitedimensional case.
We use a general result by Kim and Lee (2004) on the asymptotic normality of NTR
processes to obtain the asymptotic behavior of the posterior law of the waiting time LST.
This result is stated next. For the sake of clarity, it is stated for the non-censored case.
However, it can be enlarged to the situation where data is right-censored. For a positive real
number τ let (D[0, τ ], ∥⋅∥τ ) denote the space of all cadlag functions on [0, τ ] equipped with
the sup-norm and L[X] the law of a random object X.
Theorem 5.1 (Kim and Lee (2004)). Suppose that G(⋅) = 1 − exp[A(⋅)] is a random c.d.f.
drawn according to a NTR prior Π with corresponding increasing additive process A(⋅) whose
Lévy measure L is given by
t
L([0, t], B) = ∫ ∫
0
B
gs (x)
dxλ(s)ds,
x
1
where ∫0 gt (x)dx = 1 for all t ∈ R+ . Let the true c.d.f. G0 be continuous and such that
G0 (t) < 1 for all t ∈ R+ . Moreover let the following conditions be fulfilled for a τ > 0,
•
sup
(1 − x)gt (x) < ∞,
t∈[0,τ ],x∈[0,1]
• there is a function q(t) ∶ R+ → R such that 0 < inf q(t) < sup q(t) < ∞ and, for some
t∈[0,τ ]
23
t∈[0,τ ]
α > 1/2 and > 0,
sup
t∈[0,τ ],x∈[0,]
∣
gt (x) − q(t)
∣ < ∞,
xα
• λ(t) is bounded and positive on (0, ∞).
∞
Then it holds that for G∞
0 -almost all sequences of data S1
√
lim Π ( n[G(⋅) − EΠ;n [G(⋅)]]∣S1n ) = L[A(⋅)]
n→∞
weakly on (D[0, τ ], ∥⋅∥τ ), where A(⋅) denotes a centered Gaussian process given by A(t) =
(G0 (t) − 1)W (A0 (t)) with a Brownian motion W. The covariance structure then is h(u, v) ∶=
Cov[A(u), A(v)] = (1 − G0 (u))(1 − G0 (v))min(A0 (u), A0 (v)).
Proof. See Kim and Lee (2004) for the proof as well as for a discussion of the constraints of
the theorem.
Above theorem can well be stated as follows. The posterior distribution of the scaled
and centered random process looks more and more like a Gaussian process as the sample
size increases. The limiting process is centered, i.e. the expectation function is the constant
function taking only the value zero. Next result ensures that above assertions hold for betaStacy processes.
Corollary 5.2 (Kim and Lee (2004)). The assertion of Theorem 5.1 for the random c.d.f. G
holds, with G being governed by a beta-Stacy process prior with parameters (c(⋅), H(⋅)), where
H is a continuous distribution function with continuous density a with respect to the Lebesgue
measure such that 0 < inf t c(t)[1 − H(t)] < supt c(t)[1 − H(t)] < ∞, i.e. for G∞
0 -almost all
sequences of data S1∞ it holds
√
lim ΠBS;n ( n[G(⋅) − Ĝn (⋅)]) = L[A(⋅)]
n→∞
weakly on (D(1) [0, ∞), ∥⋅∥∞ ) with A(⋅) as in Theorem 5.1, where D(1) [0, ∞) denotes the space
of cadlag functions bounded by one on [0, ∞).
Proof. By Theorem 5 of Dey et al. (2003), a beta-Stacy process is a transformed beta process.
More precisely, a process Λ is a beta-Stacy process with parameters (c(⋅), H(⋅)) if and only
⋅ dH(s)
if it is a beta process with parameters (c(⋅)[1 − H(⋅)], ∫0 1−H(s) ). Further on, Kim and Lee
(2004) showed that a beta process fulfills the constraints of the theorem as long as
positive and continuous on R+ and c(t)[1 − H(t)] is as required.
a(s)
1−H(s)
is
Next, we turn to the asymptotic behavior of the LST g of the service time distribution
G and its random mean µ ∶= ∫ xdG(x). Since a plug-in estimator for g is employed, we will
use the functional delta method to provide the asymptotic normality result. Recall that in
the previous section a posterior consistency result was given for the random mean µ, i.e. it
was shown that the posterior law of µ centers a.s. around its true value µ0 ∶= ∫ xG0 (x). This
suggests the conjecture that a normality result might be present in this case as well. However,
even if Regazzini et al. (2003) provide results to approximate the density of the random mean
of random measures, we were not able to show the density in our case to be approximated
24
by a normal density if the sample size increases. The main reason for this is that it does not
seem to be straight-forward to obtain a suitable expansion of this approximated density.
Hence, in the following we make the rather technical assumption that there is prior knowledge of the kind that service times can not exceed a certain sufficiently large threshold M ∈ R+ .
From a practical point of view this is a rather gentle constraint. Let FM ∶= FM (R+ ) ∶= {F ∈
F(R+ ) ∶ F (t) = 1, ∀t ≥ M } be the space of all c.d.f.’s whose corresponding probability measure has support [0, M ]. Since it is well known that the prior guess on the c.d.f. G under a
beta-Stacy prior BS(c, H) is given as EBS;n [G] = H, in the following we take H ∈ FM such
that H is continuous on [0, M ]. Recall that G(⋅) = 1 − exp[−A(⋅)], where A is a non-negative
increasing additive process with Lévy measure
t
dx
−xc(s)[1−H(s)]
dNt (x) =
c(s)H(ds).
∫ e
1 − e−x
0
From the theory of increasing additive processes it is well known [see e.g. Sato (1999)] that
for all t > 0 A(t) = At is a random variable governed by a infinitely divisible distribution φt .
Let φ̂t (ξ) denote the characteristic function of φ, i.e.
φ̂t (ξ) = ∫ exp[iξs]φ(ds) = exp [− ∫
∞
0
(1 − eiξx )dNt (x)] .
Furthermore, since (At ) has independent increments one has φ̂t (ξ) = φ̂s (ξ)φ̂s,t (ξ) for all s < t,
where φ̂s,t (ξ) denotes the characteristic function of the difference At − As . Now, since H ∈ FM
it follows for t > M,
φ̂t (ξ) = exp [− ∫
= exp [− ∫
∞
0
∞
0
(1 − eiξx ) dNt (x)] = exp [− ∫
∞
0
t
1 − eiξx
−xc(s)[1−H(s)]
c(s)H(ds)dx]
∫ e
−x
1−e
0
iξx
M
1−e
e−xc(s)[1−H(s)] c(s)H(ds)dx]
∫
−x
1−e
0
⋅ exp [− ∫
0
∞
t
1 − eiξx
−xc(s)[1−H(s)]
c(s)H(ds)dx]
∫ e
−x
1−e
M
= φ̂M (ξ),
the process (At )t≥M is a.s. constant. That implies that the corresponding c.d.f. G is constant
from M onwards. In order to ensure that it is indeed a distribution function, we set G(t) ∶= 1,
(M )
for all t ≥ M . The support of the truncated prior ΠBS law will still be all of FM .
In order to achieve a posterior normality result for the random LST g, we show that the
∞
mapping Φ ∶ D(1) [0, M ] → C (1) [0, ∞); G ↦ ∫0 e−sz dG(s) is Hadamard differentiable, where
D(1) [0, M ] and C (1) [0, ∞) denote the space of cadlag and continuous functions, bounded by
one, respectively. Moreover, since we are solely interested in distribution functions on R+ ,
w.l.o.g. it is assumed that D(1) [0, M ] consists only of functions starting at zero. For a good
reference on Hadamard differentiability and applications including the functional delta method
see Kosorok (2008).
25
Lemma 5.3. The mapping
Φ ∶ (D(1) [0, M ], ∥.∥M ) → (C (1) [0, ∞), ∥.∥∞ )
M
G ↦ Φ[G](●) ∶= ∫
e−●s G(ds)
0
is Hadamard differentiable.
t↘0
Proof. Let t ↘ 0 and ht ∈ D(1) [0, M ], such that ht Ð→ h ∈ D(1) [0, M ] w.r.t. the sup-norm.
Then, by properties of the Riemann-Stieltjes integral, one has
∥
M
Φ[G + tht ](z) − Φ[G](z)
−∫
e−zs dh(s)∥
t
0
∞
M
[ht (s) − h(s)](−z)e−zs ds∣
M
( sup ∣ht (x) − h(x)∣) (−z)e−zs ds∣
= sup ∣∫
0≤z<∞
0
≤ sup ∣∫
0≤z<∞
0
0≤x<M
M
≤ ∥ht − h∥M sup ∫
0≤z<∞
t↘0
ze−zs ds
0
= ∥ht − h∥M Ð→ 0.
Define
Φ′G ∶ D(1) [0, M ] → C (1) [0, ∞)
M
h ↦ Φ′G [h](z) = ∫
e−zs dh(s).
0
Since the Riemann-Stieltjes integral is linear in the integrator, the mapping Φ′G is linear.
Moreover, if sup ∣F (s) − G(s)∣ < δ, it follows
0≤s<M
RRR
RRR M
M
M
RRR
RRR
−zs
−zs
RRR∫ e dF (s) − ∫ e dG(s)RRR ≤ ∥F − G∥M sup ∫ ze−zs ds < δ,
RRR
RRR
0≤z<∞
0
0
R
R0
′
thus the continuity of ΦG . Hence the mapping Φ is Hadamard differentiable with derivative
Φ′G .
Now, the lemma will be applied in combination with the functional delta method to obtain
the posterior normality of the service time LST centered suitably at its respective Bayes
(M )
estimator. Write ΠBS;n for the posterior law induced by the M -truncated beta-Stacy process.
Corollary 5.4. Let gn∗M (z) = ∫0 e−zs dEBS;n [G](s). Under the assumptions of Theorem 5.1
∞
and Corollary 5.2, it holds for G∞
0 -almost all sequences S1
(M ) √
lim ΠBS;n ( n [g(⋅) − gn∗M (⋅)]) = L[G(⋅)]
M
n→∞
on C (1) ([0, ∞), ∣∣.∣∣∞ ), where G(z) is a centered Gaussian process with covariance structure
γ(⋅, ⋅) given by
M
γ(u, v) = Cov [G(u), G(v)] = uv ∫
0
M
∫
e−(us+vt) h(u, v)dudv,
0
where h(⋅, ⋅) is defined in Theorem 5.1.
26
Proof. By the functional delta method applied to the mapping in the previous lemma one has
G(z) = Φ′G [(G0 (⋅) − 1)W (A0 (⋅))] (z) = ∫
M
0
e−zs d [(G0 (s) − 1)W (A0 (s)] .
Using a Riemann sum approximation for above integral, one concludes that the process G(⋅)
is a Gaussian process. Further, by well-known properties of the Riemann-Stieltjes integral,
one gets
M
G(z) = z ∫
0
(1 − G0 (s))W (A0 (s)) e−sz ds.
Using Fubini’s theorem it is immediately seen that E[G(z)] = 0 for any z ∈ R+ . Again using
Fubini’s theorem, the covariance structure of G(⋅) is obtained as
Cov [G(u), G(v)] = E [G(u)G(v)]
M
= uvE [∫
0
M
= uv ∫
0
M
∫
0
0
M
= uv ∫
0
M
∫
M
∫
(1 − G0 (s))W (A0 (s))e−us (1 − G0 (t))W (A0 (t))e−vt dsdt]
e−(us+vt) (1 − G0 (s))(1 − G0 (t))E [W (A0 (s))W (A0 (t))] dsdt
e−(us+vt) h(s, t)dsdt.
0
Next, we investigate the posterior normality of the mean of the random c.d.f. G. Since
M
no exact results seem obtainable, we use the plug-in estimator µ∗n ∶= ∫0 [1 − EBS;n [G(t)]]dt
M
which, in general, does not equal EBS;n [∫0 [1 − G(t)]dt]. This estimator in combination with
the M -truncated c.d.f.’s enables us to use the functional delta method for obtaining normality
results.
Lemma 5.5. Let M be an arbitrary positive real number. Then, the mapping
Ψ ∶ (FM , ∥⋅∥M ) → ([0, M ], ∣⋅∣)
M
G ↦ Ψ[G] ∶= ∫
[1 − G(s)]ds
0
is Hadamard-differentiable with derivative Ψ′ [h] = − ∫0 h(s)ds.
M
t↘0
Proof. Take t ↘ 0 and ht ∈ FM , such that ht Ð→ h ∈ FM . Then
∣
M
M
Ψ[G + tht ] − Ψ[G]
t↘0
+∫
h(s)ds∣ = ∣∫
h(s) − ht (s)ds∣ ≤ M ∥h − ht ∥M Ð→ 0.
t
0
0
Obviously, the derivative of Ψ is linear and continuous w.r.t. to the considered topologies.
Corollary 5.6. Under the assumptions of Theorem 5.1 and Corollary 5.2 it holds for G∞
0
almost all data S1∞ that
(M ) √
lim ΠBS;n ( n[µ − µ∗n ]) = L[H],
n→∞
M
M
where H is a centered Gaussian random variable with variance η ∶= Var[H] = ∫0 ∫0 h(s, t)dsdt.
27
Proof. By the previous lemma and the functional delta method the limiting variable is given by
M
Ψ′ [(G0 (⋅) − 1)W (A0 (⋅))] = ∫0 (1 − G0 (s))W (A0 (s))ds which is seen to be centered Gaussian
by a Riemann sum approximation in combination with Fubini’s theorem. Moreover, again by
Fubini’ theorem
Var[H] = E [H2 ] = E [∫
M
=∫
0
M
0
M
∫
0
(1 − G0 (s))W (A0 (s))(1 − G0 (t))W (A0 (t))dsdt]
M
∫
h(s, t)dsdt.
0
Next we consider the asymptotic normality of the arrival-rate when centering with its
Bayes estimate.
Proposition 5.7. Let λ̂n ∶= EΓ;n [λ]. Then, for Pλ∞0 -almost all sequences of data A∞
1 one has
√
lim ΠΓ ( n [λ − λ̂n ] ∣An1 ) = N ,
n→∞
where N is a centered Gaussian random variable with precision λ20 .
Proof. By Theorem 1.4.3. in Ghosh and Ramamoorthi (2003) the convergence of the posterior
√
distribution of n [λ − λ̂n ] to a centered normal distribution directly follows. Moreover, the
variance of this limiting Gaussian variable is given by the inverse Fisher information at the
true arrival rate. Checking the necessary conditions for interchanging integral and derivative
is left to the interested reader. The Fisher information is obtained as
I(λ0 ) = Eλ0 [(
2
∂
∂2
= λ−2
log (λe−λA )) ] = −Eλ0 [ 2 [log(λ) − λA]]
0 .
∂λ
∂λ
∣λ=λ
0
We are now in a position to formulate the posterior normality of the waiting-time LST.
However, the same techniques can be applied to show posterior normality of several other
queueing characteristics like e.g. for the queue length p.g.f. or the sojourn time LST. The
waiting-time distribution is of special interest since it gives a qualitative idea about the loss of
information that can occur in a M /G/1 system and thus helps to ensure a well-working system.
However, since the exact posterior law of the waiting time distribution is not obtainable in
closed form, asymptotic approximations are given. These results extend results of the previous
section where it was shown that the plug-in estimator is reasonable to make inference. Roughly,
we prove that the posterior law of the LST follows, asymptotically speaking, a Gaussian
(M )
quantity that is centered around the plug-in estimator. Let ΠBS⊗Γ = ΠBS ⊗ ΠΓ denote the
prior on the parameter space R+ × FM .
z(1−ρ)
z−λ(1−g(z))
z(1−λ̂n µ∗n )
∗M (z))
z−λ̂n (1−gn
Theorem 5.8. Let w(z) =
be the LST of the waiting time distribution as given in
section 2 and wn∗M (z) =
be its plug-in estimator. Then, under the assumptions
∞
∞
of Theorem 5.1 and Corollary 5.2, for G∞
0 ⊗ Pλ0 -almost all sequences of data (S, A)1 it holds
that
√
lim ΠBS⊗Γ ( n [w(⋅) − wn∗M (⋅)] ∣An1 , S1n ) = L[Z(⋅)]
n→∞
28
weakly on C([0, ∞), ∣∣.∣∣∞ ), where Z(⋅) is a centered Gaussian process with covariance structure
ζ(u, v) = Cov[Z(u), Z(v)] given by
ζ(u, v) =
µ20
w0 (u)w0 (v) 2
w0 (u)(1 − g0 (u)) w0 (v)(1 − g0 (v))
[λ
η
+
×
] + λ−2
0
0
2
2
(1 − ρ0 )
λ0 (1 − g0 (u)) − u λ0 (1 − g0 (v)) − v
λ0
w0 (u)w0 (v)λ20
γ(u, v)
[λ0 (1 − g0 (u)) − u] [λ0 (1 − g0 (v)) − v]
µ0 w0 (u)w0 (v) w0 (u)(1 − g0 (u)) w0 (v)(1 − g0 (v))
−
[
+
]
[λ0 (1 − ρ0 )]2
u
v
⎡
⎤
⎥
λ0 w0 (u)w0 (v) ⎢⎢
−us
−tv
⎥,
w
(u)
e
h(s,
t)d(s,
t)
+
w
(v)
e
h(s,
t)d(s,
t)
−
0
0
∫
∫
⎢
⎥
2
(1 − ρ0 )
⎢
⎥
⎣
⎦
[0,M ]2
[0,M ]2
+
where γ(⋅, ⋅) is given in Corollary 5.4.
Proof. Taking a similar route of proving as in the proof of Theorem 3 in Conti (1999), we
√
begin with a decomposition of n [w(z) − wn∗M (z)]. The decomposition yields
√
−w(z) √
n [w(z) − wn∗M (z)] =
n [λµ − λ̂n µ∗n ]
1−ρ
√
1
1
+ z(1 − λ̂n µ∗n ) n [
]
−
z − λ(1 − g(z)) z − λ̂n (1 − g(z))
√
1
1
)
+ z(1 − λ̂n µ∗n ) n (
−
z − λ̂n (1 − g(z)) z − λ̂n (1 − gn∗M (z))
=∶ Z1;n (z) + Z2;n (z) + Z3;n (z)
Now, the three terms of the sum are investigated separately and it will be shown that they
possess the same asymptotic distribution as objects whose asymptotic is easier to obtain.
These objects will be tagged by an additional ∗ superscript. First write
Z1;n (z) =
−w(z) √
n [µ∗n (λ − λ̂n ) + (µ − µ∗n )(λ − λ0 ) + λ0 (µ − µ∗n )]
1−ρ
and note that by the uniform posterior consistency results of the previous section and the
M
continuity of the mapping G ↦ ∫0 (1 − G(s))ds one has
lim ΠBS⊗Γ [Z1;n (z)∣An1 , S1n ] = lim ΠBS⊗Γ [
n→∞
=∶
n→∞
∗
lim ΠBS⊗Γ [Z1;n
(z)∣An1 , S1n ]
n→∞
−w0 (z) √
n(µ0 (λ − λ̂n ) + λ0 (µ − µ∗n ))∣An1 , S1n ]
1 − ρ0
=∶ L[Z1 (z)].
For Z2;n (z), note that the mapping λ ↦ [z − λ(1 − g(z))]−1 is analytic in a suitably chosen
neighborhood of λ0 . Its derivative is given by
λ↦
1 − g(z)
[z − λ(1 − g(z))]2
.
29
Thus, a Taylor expansion of that mapping yields
Z2;n (z) = (1 − λ̂n µ∗n )z
1 − g(z)
√
2
n (λ − λ̂n ) ,
[z − λ̄(1 − g(z))]
for a suitably chosen λ̄ ∈ [λ0 , λ̂n ]. Therefore, using consistency results and continuous mapping, one gets
lim ΠBS⊗Γ [Z2;n (z)∣S1n , An1 ] = lim ΠΓ [(1 − ρ0 )z
n→∞
n→∞
1 − g0 (z)
[z − λ0 (1 − g0 (z))]
√
2
n (λ − λ̂n ) ∣An1 ]
∗
∣An1 ] =∶ L[Z2 (z)].
=∶ lim ΠΓ [Z2;n
n→∞
−1
Another Taylor expansion for the mapping x ↦ [z − λ̂n (1 − x)]
as before yields
(M )
lim ΠBS⊗Γ [Z3;n (z)∣S1n , An1 ] = lim ΠBS [−
n→∞
n→∞
and analogous reasoning
√
w0 (z)λ0
n [g(z) − gn∗M (z)] ∣S1n ]
λ0 (1 − g0 (z)) − z
(M )
∗
∣S1n ] =∶ L[Z3 (z)].
=∶ lim ΠBS [Z3;n
n→∞
Now, the convergence of the posterior law of the waiting time LST follows from the previous
results of the present section. What remains is the calculation of the covariance structure.
This, in turn, is easily obtained by above decomposition and the assumed independence of the
arrivals and services or their prior laws, respectively.
3
3
3
Cov [∑ Zi (u), ∑ Zi (v)] = ∑ Cov [Zi (u), Zi (v)] + ∑ Cov [Zi (u), Zj (v)]
i=1
i=1
i=1
i≠j
3
= ∑ Cov [Zi (u), Zi (v)]
i=1
+ Cov [Z1 (u), Z2 (v)] + Cov [Z2 (u), Z1 (v)]
+ Cov [Z1 (u), Z3 (v)] + Cov [Z3 (u), Z1 (v)]
By the previous results of this section it follows
3
µ20
w0 (u)w0 (v) 2
[λ
η
+
]
0
(1 − ρ0 )2
λ20
w0 (u)(1 − g0 (u)) w0 (v)(1 − g0 (v))
+ λ−2
×
0
λ0 (1 − g0 (u)) − u λ0 (1 − g0 (v)) − v
w0 (u)w0 (v)λ20
+
γ(u, v).
[λ0 (1 − g0 (u)) − u] [λ0 (1 − g0 (v)) − v]
∑ Cov [Zi (u), Zi (v)] =
i=1
Furthermore, by the independence assumption of the prior laws of the inter-arrival rate and
the service time distribution, one has
Cov [Z1 (u), Z2 (v)] =
µ0
w0 (u)w0 (v)(g0 (v) − 1)
×
.
v − λ0 (1 − g0 (v))
λ20 (1 − ρ0 )
30
has
Furthermore, using the previous results of the present section and Fubini’s theorem, one
w0 (u)w0 (v)λ20
E[H × G(v)]
(1 − ρ0 )(λ0 (1 − g0 (v)) − v)
w0 (u)w0 (v)λ20
=
(1 − ρ0 )(λ0 (1 − g0 (v)) − v)
Cov [Z1 (u), Z3 (v)] =
M
× E [∫
0
M
(1 − G0 (s))W (A0 (s))ds × v ∫
0
(G0 (t) − 1)W (A0 (t))e−tv dt]
M
M
vw0 (u)w0 (v)λ20
=
e−tv h(s, t)dsdt
∫
∫
(1 − ρ0 )(λ0 (1 − g0 (v)) − v) 0
0
w0 (u)w0 (v) 2
e−tv h(s, t)d(s, t).
λ w0 (v) ∫
=−
(1 − ρ0 )2 0
[0,M ]2
Finally, compounding above covariance structures yields ζ(⋅, ⋅).
The covariance structure ζ(⋅, ⋅) depends on the unknown objects. However, one can use
the suggested estimators and plug them in place of the true ones. This might be helpful to
implement the problem and the provided consistency results ensure accuracy as long as the
sample size is large enough.
31
Funding: This work was supported by the Deutsche Forschungsgemeinschaft (German
Research Foundation) within the programme "Statistical Modeling of Complex Systems and
Processes—Advanced Nonparametric Approaches", grant GRK 1953.
Corresponding Author
Dr. Cornelia Wichelhaus
Technische Universität Darmstadt
Schlossgartenstraße 7
64289 Darmstadt
Germany
[email protected]
Moritz von Rohrscheidt
Ruprecht-Karls Universität Heidelberg
Berliner Straße 41-49
69120 Heidelberg
Germany
[email protected]
32
References
Armero, C. and M. Bayarri
1994. Prior Assessments for Prediction in Queues. The Statistician, Pp. 139–153.
Armero, C. and D. Conesa
1998. Inference and Prediction in Bulk Arrival Queues and Queues with Service in Stages.
Applied Stochastic Models and Data Analysis, 14(1):35–46.
Armero, C. and D. Conesa
2004. Statistical Performance of a Multiclass Bulk Production Queueing System. European
Journal of Operational Research, 158(3):649–661.
Armero, C. and D. Conesa
2006. Bayesian Hierarchical Models in Manufacturing Bulk Service Queues. Journal of
Statistical Planning and Inference, 136(2):335–354.
Asmussen, S.
2008. Applied Probability and Queues, volume 51. Springer Science & Business Media.
Ausín, M. C., R. E. Lillo, and M. P. Wiper
2004. Bayesian estimation for the M/G/1 queue using a phase-type approximation. Journal
of Statistical Planning and Inference, 118(1):83–101.
Ausín, M. C., R. E. Lillo, and M. P. Wiper
2007. Bayesian Control of the Number of Servers in a GI/M/c Queueing System. Journal
of Statistical Planning and Inference, 137(10):3043–3057.
Ausín, M. C., R. E. Lillo, and M. P. Wiper
2008. Bayesian Prediction of the Transient Behaviour and Busy Period in short-and longtailed GI/G/1 Queueing Systems. Computational Statistics & Data Analysis, 52(3):1615–
1635.
Conti, P. L.
1999. Large Sample Bayesian analysis for Geo/G/1 discrete-time Queueing Models. The
Annals of Statistics, Pp. 1785–1807.
De Finetti, B.
1974. Theory of Probability, volume 1. Wiley New York.
de Laplace, P.
1774. Mémoire sur les suites récurro-récurrentes et sur leurs usages dans la théorie des
hasards. Mém. Acad. Roy. Sci. Paris, 6:353–371.
Dey, J., R. Erickson, and R. Ramamoorthi
2003. Some Aspects of neutral-to-right Priors. International Statistical Review, 71(2):383–
401.
Diaconis, P. and D. Freedman
1986. On the Consistency of Bayes Estimates. The Annals of Statistics, Pp. 1–26.
33
Diaconis, P. and D. Ylvisaker
1985. Quantifying Prior Opinion. Bayesian Statistics, 2:133–156.
Doksum, K.
1974. Tailfree and Neutral Random Probabilities and their Posterior Distributions. The
Annals of Probability, Pp. 183–201.
Doob, J. L.
1949. Application of the Theory of Martingales. Le calcul des probabilites et ses applications,
Pp. 23–27.
Epifani, I., A. Lijoi, and I. Prünster
2003. Exponential Functionals and Means of neutral-to-the-right Priors.
90(4):791–808.
Biometrika,
Feigin, P. D. and R. L. Tweedie
1989.
Linear Functionals and Markov Chains associated with Dirichlet Processes.
105(03):579–585.
Ferguson, T. S.
1973. A Bayesian Analysis of some Nonparametric Problems. The Annals of Statistics,
Pp. 209–230.
Freedman, D. A.
1963. On the Asymptotic Behavior of Bayes’ Estimates in the Discrete Case. The Annals
of Mathematical Statistics, Pp. 1386–1403.
Ghosh, J. K. and R. Ramamoorthi
2003. Bayesian Nonparametrics. Springer Series in Statistics.
Gill, R. D. and S. Johansen
1990. A survey of Product-Integration with a View toward Application in Survival Analysis.
The Annals of Statistics, Pp. 1501–1555.
Grübel, R. and S. M. Pitts
1992. A Functional Approach to the Stationary Waiting Time and Idle Period Distributions
of the GI/G/1 Queue. The Annals of Probability, Pp. 1754–1778.
Haigh, J.
2004. Probability Models. Springer Science & Business Media.
Hewitt, E. and L. J. Savage
1955. Symmetric Measures on Cartesian Products. Transactions of the American Mathematical Society, 80(2):470–501.
Hjort, N. L.
1990. Nonparametric Bayes Estimators based on Beta Processes in Models for Life History
Data. The Annals of Statistics, Pp. 1259–1294.
Insua, D. R., M. Wiper, and F. Ruggeri
1998. Bayesian Analysis of M/Er/1 and M/H_k/1 Queues. Queueing Systems, 30(3-4):289–
308.
34
Johnson, R. A.
1967. An Asymptotic Expansion for Posterior Distributions. The Annals of Mathematical
Statistics, Pp. 1899–1906.
Kechris, A. S.
1995. Classical Descriptive Set Theory. Springer.
Kendall, D. G.
1953. Stochastic Processes occurring in the Theory of Queues and their Analysis by the
Method of the imbedded Markov Chain. The Annals of Mathematical Statistics, Pp. 338–
354.
Kim, Y. and J. Lee
2001. On Posterior Consistency of Survival Models. The Annals of Statistics, Pp. 666–686.
Kim, Y. and J. Lee
2004. A Bernstein-von Mises Theorem in the Nonparametric Right-Censoring Model. The
Annals of Statistics, Pp. 1492–1512.
Kingman, J. F.
1978. Uses of Exchangeability. The Annals of Probability, Pp. 183–197.
Kosorok, M.
2008. Introduction to Empirical Processes and Semiparametric Inference. Springer, New
York.
Nelson, R.
2013. Probability, Stochastic Processes, and Queueing Theory: the Mathematics of Computer
Performance Modeling. Springer Science & Business Media.
Phadia, E. G.
2015. Prior Processes and their Applications. Springer.
Regazzini, E., A. Lijoi, and I. Prünster
2003. Distributional Results for Means of Normalized Random Measures with Independent
Increments. The Annals of Statistics, Pp. 560–585.
Sato, K.
1999. Lévy Processes and Infinitely Divisible Distributions.
Schervish, M. J.
1995. Theory of Statistics. Springer Science & Business Media.
Walker, S. and P. Muliere
1997. Beta-Stacy Processes and a Generalization of the Pólya-Urn Scheme. The Annals of
Statistics, Pp. 1762–1780.
Wiper, M.
1998. Bayesian Analysis of Er/M/1 and Er/M/c Queues. Journal of Statistical Planning
and Inference, 69(1):65–79.
35
| 10math.ST
|
1
Summarization of User-Generated Sports Video
by Using Deep Action Recognition Features
arXiv:1709.08421v1 [cs.CV] 25 Sep 2017
Antonio Tejero-de-Pablos, Member, IEEE, Yuta Nakashima, Member, IEEE, Tomokazu Sato, Member, IEEE,
Naokazu Yokoya, Life Senior Member, IEEE, Marko Linna, and Esa Rahtu
Abstract—Automatically generating a summary of sports video
poses the challenge of detecting interesting moments, or highlights, of a game. Traditional sports video summarization methods leverage editing conventions of broadcast sports video that
facilitate the extraction of high-level semantics. However, usergenerated videos are not edited, and thus traditional methods
are not suitable to generate a summary. In order to solve this
problem, this work proposes a novel video summarization method
that uses players’ actions as a cue to determine the highlights
of the original video. A deep neural network-based approach
is used to extract two types of action-related features and to
classify video segments into interesting or uninteresting parts.
The proposed method can be applied to any sports in which
games consist of a succession of actions. Especially, this work
considers the case of Kendo (Japanese fencing) as an example of
a sport to evaluate the proposed method. The method is trained
using Kendo videos with ground truth labels that indicate the
video highlights. The labels are provided by annotators possessing
different experience with respect to Kendo to demonstrate how
the proposed method adapts to different needs. The performance
of the proposed method is compared with several combinations
of different features, and the results show that it outperforms
previous summarization methods.
Video summary
Index Terms—Sports video summarization; user-generated
video; action recognition; deep learning; 3D convolutional neural
networks; long short-term memory.
I. I NTRODUCTION
T
HE widespread availability of cameras has led to an
enormous and ever-growing collection of unedited and
unstructured videos generated by users around the world
[1]. A popular domain corresponds to sports videos taken at
public events and professional/amateur matches. These types
of user-generated sports videos (UGSVs) are often lengthy
with several uninteresting parts, and thus many of them are
stored and are never reviewed. A convenient way to review,
transfer, and share the video via channels, such as social
network services, includes generating summaries of a UGSV
that only shows the interesting parts or highlights.
Automatic video summarization is a challenging problem
that involves extracting semantics from video. Traditional usergenerated video summarization methods target general videos
in which contents are not limited to a specific domain because
of the difficulty in extracting semantics from an unstructured
A. Tejero-de-Pablos, Y. Nakashima, T. Sato and N. Yokoya were with
the Graduate School of Information Science, Nara Institute of Science and
Technology, Nara, 630-0101 Japan. e-mail: [email protected].
M. Linna is with the Center for Machine Vision Research, University
of Oulu, Oulu, 90014 Finland. E. Rahtu is with the Department of Signal
Processing, Tampere University of Technology, Tampere, 33101 Finland
Manuscript received July, 2017
Fig. 1. An overview of the proposed method to generate a summary of
user-generated sports video (UGSV) based on players’ actions. Two types of
features that represent players’ actions, namely body joint-based and holistic
actions, are used to extract highlights from the original video.
video. As opposed to extracting semantics, these methods
use low-level visual features and attempt to reduce visual redundancy using clustering-based approaches [2]. More recent
user-generated video summarization methods use deep neural
network-based features to extract higher-level semantics [3].
With respect to sports video and especially with respect
to professional sports in broadcast TV programs, there exist
a number of summarization methods that leverage editing
conventions to extract high-level semantics by exploiting a
knowledge of the specific sport [4], [5]. For example, broadcast sports video contains slow-motion replays [6], narration
and superimposed text [7], and specific camera work [8].
This type of video editing constitutes the basis for heuristic
rules that aid in the determination of highlights (or certain
interesting moments of a game such as a free kick in soccer
or a pitch in baseball). Additionally, broadcast video is often
edited by following the structure of the sport (i.e., “downs”
2
in American football), and this constitutes another cue for
summarization [9].
UGSV lies in in-between general user-generated video and
broadcast sports video. Given a specific sport, domain knowledge can be used to generate a UGSV summary. However,
UGSV does not typically follow any editing convention or
structure, and thus a different type of cues is required to grab
the semantics.
This paper describes a novel method for UGSV summarization. Our observation with respect to semantics extraction is
that a game in most sports consists of a succession of players’
actions, and thus the actions can be one of the most important
cues to determine if a certain part of video is interesting or
not. For example, a definitive smash in tennis is more likely to
be enjoyed by tennis viewers than a repetitive ball exchange.
Also, a feint in boxing might not be interesting by itself, but
viewers would surely enjoy it if it is followed by an uppercut
that knocks out the opponent. Based on this observation, the
proposed method uses players’ actions to model the highlights
of a sports game (Fig. 1).
Inspired by recent methods for action recognition in video,
the proposed method uses a two-stream architecture that
extracts two types of action features for action representation.
One type involves players’ body joint positions estimated in
2D or 3D (obtainable from depth maps). Body joint-based
features provide a precise representation of actions. The other
type involves holistic features that can be obtained with deep
convolutional neural networks (CNNs) designed to extract
spatio-temporal features from video. Holistic features help to
capture actions in their context. Subsequently, long short-term
memory (LSTM) is used to model the temporal dependencies
of the extracted features for highlight classification. In our
summaries, a highlight may contain one or more actions
performed by the players. Several types of body joint-based
and holistic features are comparatively evaluated for UGSV
summarization.
We consider the case of Kendo (Japanese fencing) as an
example of a sport to evaluate the proposed method. This work
is an extension of our previous work in [10].
The main contributions of this work are as follows:
•
•
•
•
A novel UGSV summarization method that determines
highlights in a video by using players’ action features
and a deep neural network.
A comparison of several action feature extraction methods, i.e., body joint features (RGB image-based 2D
joint positions and depth map-based 3D joint positions)
and holistic features (C3D [11] and CNN-ISA [12]) to
demonstrate their adequacy to model video highlights.
A new UGSV dataset with 246 min of Kendo videos in
which each second of a video has a label that indicates
whether or not it is a part of a highlight. The labels were
provided by annotators with and without experience in
Kendo.
Objective and subjective evaluations of the proposed
method. Users with and without experience in Kendo
were surveyed to investigate the adequacy of the proposed
method with respect to individual needs.
II. R ELATED WORK
This section introduces existing video summarization methods in terms of the types of video (i.e., broadcast sports video
and user-generated video). This section also reviews existing
work in action recognition, which constitutes a key technique
for modeling highlights in the proposed method.
A. Action recognition from video
Traditional approaches for action recognition use conventional classifiers and hand-crafted features that represent human movement in RGB video [13], [14], [15]. However,
the field of action recognition recently benefits from deep
neural networks for feature extraction and classification. An
example corresponds to three-dimensional convolutional neural networks (3D CNNs) that constitute an extension of CNNs
applied to images (2D CNNs). While 2D CNNs perform only
spatial operations in a single image, 3D CNNs also perform
temporal operations while preserving temporal dependencies
among the input video frames [11]. Le et al. [12] used a 3D
CNN with independent subspace analysis (CNN-ISA) and a
support vector machine (SVM) to recognize human actions
from video. Additionally, Tran et al. [11] designed a CNN
called C3D to extract video features that were subsequently
fed to an SVM for action recognition.
Another state-of-the-art CNN-based action recognition
method employed two types of streams, namely a spatial
appearance stream and a temporal motion stream [16], [17].
Videos are decomposed into spatial and temporal components,
i.e., into an RGB and optical flow representation of its frames,
and fed into two separate 3D CNNs. Each stream separately
provides a score for each possible action, and the scores from
two streams were later combined to obtain a final decision.
This architecture is supported by the two-stream hypothesis of
neuroscience in which the human visual system is composed of
two different streams in the brain, namely the dorsal stream
(spatial awareness and guidance of actions) and the ventral
stream (object recognition and form representation) [18].
In addition to RGB videos, other methods leverage depth
maps obtained from commodity depth sensors (e.g. Microsoft
Kinect) to estimate the human 3D pose for action recognition
[19], [20], [21]. The third dimension provides robustness to
occlusions and variations from the camera viewpoint.
B. Broadcast sports video summarization
Summarization of sports video focuses on extracting interesting moments (i.e., highlights) of a game. A major approach
leverages editing conventions such as those present in broadcast TV programs. Editing conventions are common to almost
all videos of a specific sport and allow automatic methods
to extract high-level semantics [22], [23]. Ekin et al. [7]
summarized broadcast soccer games by leveraging predefined
camera angles in edited video to detect soccer field elements
(e.g., goal posts). Similar work used slow-motion replays to
determine key events in a game [6] and predefined camera
motion patterns to find scenes in which players scored in
basketball/soccer games [8].
3
In addition to editing conventions, the structure of the sport
also provides high-level semantics for summarization. Certain
sports are structured in “plays” that are defined based on
the rules of the sport and are often easily recognized in
broadcast videos [4], [24], [25]. For example, Li et al. [9]
summarized American football games by leveraging their turnbased structure and recognizing “down” scenes from the video.
Other methods used metadata in sports videos [5], [26] since it
contains high-level descriptions (e.g., “hits” may be annotated
in the metadata with their timestamps for a baseball game). A
downside of these methods is that they cannot be applied to
sports video without any editing conventions, structures, and
metadata. Furthermore, they are based on heuristics, and thus
it is difficult to generalize them to different sports.
Existing work also proposed several methods that are not
based on heuristics. These methods leverage variations between scenes that are found in broadcast video (e.g., the closeup in a goal celebration in soccer). Chen et al. [27] detected
intensity variations in color frames to segment relevant events
to summarize broadcast videos of soccer, basketball, and tennis. Mendi et al. [28] detected the extrema in the optical flow
of a video to extract the frames with the highest action content
and construct a summary for broadcast rugby video. These
methods can be more generally applied to broadcast videos,
but they lack high-level semantics, and thus the extracted
scenes do not always correspond to the highlights of the game.
C. User-generated video summarization
Sports video includes a somewhat universal criterion on the
extent to which a “play” is interesting (e.g., a homerun in a
baseball game should be an interesting play for most viewers).
In contrast, user-generated video in general do not have a clear
and universal criterion to identify interesting moments. Additionally, neither editing conventions nor specific structures that
can be used to grab high-level semantics can be leveraged [29].
Hence, many video summarization methods for user-generated
video are designed to reduce the redundancy of a lengthy
original video as opposed to determining interesting moments.
Traditional methods uniformly sample frames [30] or cluster
them based on low-level features, such as color [2], to extract a
brief synopsis of a lengthy video. These methods do not extract
highlights of the video, and therefore researchers proposed
other types of summarization criteria such as important objects
[31], attention [32], [33], and interestingness [34].
Recent methods use deep neural networks to automatically
learn a criterion to model highlights. Yang et al. [35] extracted
features from ground-truth video summaries to train a model
for highlight detection. Otani et al. [36] use a set of both
original videos and their textual summaries that are generated
via majority voting by multiple annotators to train a model to
find video highlights. Video titles [37] and descriptions [3] can
also be used to learn a criterion to generate summaries. The
aforementioned methods employed networks with CNNs and
LSTMs, and this requires a large amount of data for training.
The generation of these types of large summarization datasets
for training their network is non-viable for most researchers,
and thus their models are built on pre-trained networks such
as VGG [38] and GoogLeNet [39].
III. UGSV SUMMARIZATION USING ACTION FEATURES
UGSV summarization inherits the intricacies of usergenerated video summarization. The extraction of high-level
semantics is not trivial in the absence of editing conventions.
However, given a specific sport, it is possible to leverage
domain knowledge to facilitate the extraction of high-level
semantics. The idea in the present work for semantics extraction involves utilizing players’ actions, as they are the
main constituents of a game. Our previous work [10] applied
an action recognition technique to sports videos to determine
combinations of actions that interest viewers by using a hidden
Markov model with Gaussian Mixture emissions. To the best
of our knowledge, this work was the first to use a UGSV
summarization based on players’ actions.
A major drawback of the previous work [10] involves the
usage of the outputs of a classic action recognizer as features
to determine the highlights of the UGSV. Moreover, in addition
to the UGSV summarization dataset, the method also requires
an action dataset of the sport to train the action recognizer.
Another drawback of [10] is that it only uses features from
3D joint positions (that are estimated by, e.g., [40]). They
provide rich information on players’ actions but miss other
potential cues for summarization contained in the appearance
of the scene. For example, appearance is useful when the joint
position estimation fails.
In this work, the method in [10] is extended by employing
a two-stream deep neural network [16], [17]. Our new method
considers two different types of inputs, namely RGB frames
of video and body joint positions, and each of them are transformed through two separate neural networks (i.e., streams).
These two streams are then fused to form a single action
representation to determine the highlights. Our method does
not require an action recognizer; and the proposed network
is trained from the lower layer to the top layers by using a
UGSV summarization dataset.
Given the proposed method, it is necessary for the target
sports to satisfy the following conditions: (1) a game consists
of a series of recognizable actions performed by each player
and (2) players are recorded from a close distance for joint
position estimation. However, it is expected that the idea of
using action recognition-related features for UGSV summarization is still valid for most types of sports.
A. Overview
In this work, UGSV summarization is formulated as a
problem of classifying a video segment in the original video
as interesting (and thus included in the summary) or uninteresting. A two-stream neural network is designed for this
problem, and it is trained in a supervised manner with ground
truth labels provided by multiple annotators.
Figure 1 shows an overview of the proposed method.
The method first divides the original input video into a set
S = {st } of video segments in which RGB frames can be
accompanied by their corresponding depth maps. A video
segment st is then fed into the two-stream network. The body
joint-based feature stream considers RGB frames (and depth
maps) in st to obtain body joint-based features xt , and the
4
...
Time
Fig. 2. In the video segmentation, a video segment st contains frames inbetween t − 1 and t + 2 sec. Each video segment overlaps with adjacent ones
for 2 sec.
holistic feature stream computes holistic features yt from the
RGB frames. The former stream captures the players’ motion
in detail by explicitly estimating their body joint positions. The
latter stream represents entire frames in the video segment, and
this is helpful to encode, for example, the relationship between
the players. The features X = {xt } and Y = {yt } are then
used for highlight classification by considering the temporal
dependencies among the video segments. The highlight summaries correspond to a concatenation of the segments that are
classified as interesting.
B. Video segmentation
Various methods have been proposed to segment a video
(e.g., its content [27]). In the proposed method, the original
input video of length T seconds (sec) is uniformly segmented
into multiple overlapping segments, so that subsequently a
second t of video can be represented by extracting action
features from a video segment st , i.e., S = {st |t = 1, . . . , T }.
Thus, T also corresponds to the number of video segments
in S, and st corresponds to the video segment that contains
frames from sec t − 1 to sec t + τ − 1. For a finer labeling
of highlights, short video segments are required. We choose
a τ = 3, for which adjacent video segments overlap by 2
sec as shown in Fig. 2. Each segment st may contain a
different number of frames, especially when the input video is
captured with an RGB-D camera (e.g., Microsoft Kinect) due
to automatic exposure control.
C. Body joint-based feature stream
In this stream (Fig. 3), a sequence of positions of the
players’ body joints (e.g., head, elbow, etc.) that represent the
movement of the players irrespective of their appearance is
used to obtain a detailed representation of players’ actions.
Specifically, two types of joint representations are employed
in this work, namely 3D positions from depth maps or 2D
positions from RGB frames.
With respect to the 3D body joint positions, the skeleton
tracker (e.g., [40]) is used as in the previous work [10], and it
estimates 3D positions from depth maps. The 3D positions are
usually represented in the camera coordinate system, and thus
they are view-dependent, thereby introducing extra variations.
Fig. 3. In the body joint-based feature stream, an LSTM is fed with the body
joint positions estimated from players on each frame uft to model temporal
dependencies and extract a feature vector ht . Additionally, these body joint
positions are also used to calculate an activity measure for all players at . The
body joint-based feature vector is their concatenation xt .
Therefore, the 3D positions from the camera coordinate system
are transformed to each player’s coordinate system in which
the origin corresponds to one of the body joints (e.g., torso).
In the absence of depth maps (which is likely in current
user-generated video), 2D body joint positions can still be
estimated from RGB frames. Recent methods in human pose
estimation leverage 2D CNNs to learn spatial relationships
among human body parts and estimate 2D joint positions
[41]. These types of 2D positions are not as robust relative
to view variations as 3D positions. However, they can be
extracted from RGB frames alone without using depth maps.
The given 2D body joint positions are also transformed to
positions relative to the player’s coordinate system to ensure
that they are translation invariant.
The use of an activity measure works positively while
extracting highlights [10]. In order to calculate the activity
measure of a certain player q in the video segment s (we
omit subscript t in this subsection for notation simplicity), the
volume (or plane for the 2D case) around the player is divided
into a certain number of regions, and the ratio rv of the number
of frames in the video segment in which the joint j falls into
region v is calculated. The activity measure aq is defined as
the entropy obtained based on rv . With respect to each joint
j in player q’s body, we compute the entropy as follows:
X
ej = −
rv log(rv ).
(1)
v
Then, we calculate the activity measure for player q as follows:
aq =
J
X
j=1
ej .
(2)
5
The activity measure for all players in a segment is calculated.
More details on the activity measure can be found in [10].
Let ufqj in R3 or R2 (a row vector) denote the 3D or 2D
relative position of joint j of player q in frame f of video
segment s. Subsequently, given the number of players Q and
of estimated body joints J, the concatenation of the body joints
of all players in frame f is defined as follows:
uf = (uf11 · · · ufqj · · · ufQJ ).
(3)
As shown in Fig. 3, vectors u1 to uF are passed through
an LSTM to model the temporal dependencies of the joint
positions of players’ bodies in s. After feeding the last vector
uF , the hidden state vector h of the LSTM is considered as a
representation of {uf }. The state of the LSTM is reset to all
zeros prior to feeding the next video segment. It is assumed
that the number of players Q does not change. However, some
players can be out of the field-of-view of the camera. In this
case, the corresponding elements in u and a are substituted
with zeros.
The proposed method represents a video segment s by
concatenating the LSTM output and the activity measure of
all players in one vector as follows:
Fig. 4. The recurrent neural network architecture for highlight classification
consists of a single LSTM layer and several fully-connected layers. The body
joint-based features xt and holistic features yt extracted from video segment
st are input to calculate the probability pt that the segment is interesting.
where a denotes the concatenation of (a1 · · · aQ ).
This stream represents a video segment s by using a holistic
feature vector y that corresponds to the output of one of the
aforementioned 3D CNNs.
D. Holistic feature stream
E. Highlight classification using LSTM
This stream encodes a video segment s in a spatio-temporal
representation. We rely on state-of-the-art 3D CNNs over RGB
frames. Training a 3D CNN from scratch requires thousands of
videos [42] that are not available for the proposed task. Recent
work on deep neural networks for computer vision [11], [17],
[43] shows that the activations of an upper layer of a CNN
are useful for other related tasks without requiring fine-tuning.
Thus, 3D CNN in which parameters are pre-trained with largescale datasets can be used instead to leverage a huge amount
of labeled training data [44]. The proposed method utilizes
a 3D CNN for action feature extraction pre-trained with a
publicly available action recognition dataset, such as Sports1M [42]. Unlike our previous work [10], which required to
classify players’ actions, it is not necessary to use a sportspecific action recognition dataset.
Two types of holistic representations of video segments
extracted using 3D CNNs are employed, namely CNN-ISA
[12] and C3D [11]. Specifically, CNN-ISA provides a representation robust to local translation (e.g., small variations in
players’ or camera motion) while it is selective to frequency,
rotation, and velocity of such motion. The details of CNNISA can be found in [12]. CNN-ISA achieved state-of-theart performance in well-known datasets for action recognition
such as YouTube [45], Hollywood2 [46], and UCF sports [47].
Additionally, C3D features provide a representation of objects,
scenes, and actions in a video. The network architecture and
other details can be found in [11]. C3D pre-trained with the
Sports-1M dataset achieved state-of-the-art performance on
action recognition over the UCF101 dataset [48].
Figure 4 shows the network architecture designed to extract
highlights of UGSV using the features xt and yt from video
segment st . The temporal dependencies among video segments
are modeled using an LSTM, and the network outputs the
probability pt that the video segment st is interesting. First,
the features are concatenated to form vector zt = (xt yt ).
Vector zt then goes through a fully-connected layer to reduce
its dimensionality.
It is assumed that interesting video segments are related to
each other in time, in the same way a skillful boxer first feints
a punch prior to hitting to generate an opening in the defense.
Existing work in video summarization uses LSTMs to extract
video highlights [35] since it allows the modeling of temporal
dependencies across longer time periods when compared to
other methods [49]. Following this concept, an LSTM layer
is introduced to the network for highlight classification. The
hidden state of the LSTM from each time step goes through
two fully-connected layers, and this results in a final softmax
activation of two units corresponding to “interesting” and
“uninteresting.”
The proposed method provides control over the length L
of the output summary. The softmax activation of the unit
corresponding to “interesting” is considered as the probability
pt ∈ [0, 1] that segment st is interesting, and skimming curve
formulation [50] is applied to the sequence of probabilities by
decreasing a threshold θ from 1 until a set of segments whose
total length is highest below L is determined (Fig. 5). The
segments in which the probability exceeds θ are concatenated
to generate the output summary in the temporal order. Hence,
x = (h a),
(4)
6
Fig. 5. A summary is generated by concatenating segments in which the
probability pt of being a highlight surpasses a certain threshold θ. The
threshold is chosen to fit the summary length.
the resulting summary may contain multiple consecutive interesting segments.
F. Network training
A pre-trained CNN is used in the holistic features stream
(i.e., CNN-ISA or C3D) while the LSTMs and fully-connected
layers are trained from scratch. Hence, during training, the
parameters in the holistic feature stream (i.e., CNN layers)
are fixed and those in the body joint-based feature stream (i.e.,
lstmJ ) and highlight classification (i.e., fc1, lstmH , fc2, and
fc3) are updated.
The UGSV dataset contains video and ground truth labels
lt ∈ {0, 1} for every second, where lt = 1 implies that the
period from t sec to t + 1 sec of the video is “interesting”
and lt = 0 otherwise. Label lt is assigned to its corresponding
video segment st .
With respect to training, cross-entropy loss ` is used:
X
`=
lt log pt .
(5)
IV. E XPERIMENTS
The proposed method is evaluated objectively and subjectively. With respect to the objective evaluation, the performance of the proposed method is compared while using
different representations of the players’ actions. Specifically,
only body joint features (3D or 2D), only holistic motion
features (CNN-ISA or C3D), and a combination of both
features are evaluated. Subsequently, the completeness of the
highlights of the generated summaries are examined. With
respect to the subjective evaluation, users with and without
experience in the sport are surveyed to study their opinions
with respect to the summaries.
A. Implementation details
For the evaluation, Kendo (Japanese fencing) was selected
as an example of a sport. Kendo is a martial art featuring two
players and a set of recognizable actions (e.g., attacking and
parrying). We used the UGSV Kendo dataset in [10], which
contains 90 min of self-recorded Kendo matches divided in 10
RGB-D videos taken with a Microsoft Kinect v2, and extended
it by adding 18 more self-recorded RGB-D Kendo videos.
The total length of the videos is 246 min with a framerate of
approximately 20 fps.
The body joint-based feature stream was configured for Q =
2 players since Kendo is a two-player sport. The tracker in [40]
was used to estimate J = 15 3D body joint positions from
depth maps: head, neck, torso, right shoulder, right elbow,
right wrist, left shoulder, left elbow, left wrist, right hip, right
knee, right ankle, left hip, left knee, and left ankle. In order to
estimate the 2D positions of the players’ joints from the RGB
frames, the CNN-based method proposed by Linna et al. [51]
was used. We pre-trained this joint estimation CNN with the
human pose dataset used by Linna et al. [51], and then we
fine-tuned it with our extended UGSV Kendo video dataset.
The network provides J = 13 joints (that is the same as the
3D case with the exception of neck and torso). Therefore, the
size of vector uft is Q×J ×3 = 90 in the case of 3D positions
and Q × J × 2 = 52 in the case of 2D. Given that the size
of lstmJ is the same as that of the input and that the size of
at is Q = 2, the feature vector xt for the stream is ∈ R92 for
3D and ∈ R54 for 2D.
With respect to the holistic feature stream, either the CNNISA [12] or C3D [11] networks were used. The UGSV
Kendo dataset is not sufficiently large to train the CNNs
from scratch, and thus networks pre-trained with an action
recognition dataset were used. The CNN-ISA was trained
in an unsupervised way with the Hollywood2 dataset that
consists of 2859 videos [46]. For this network, we followed
the configuration in [52]. We used a vector quantization
representation of the extracted features with a codebook size
of 400, thereby resulting in a feature vector yt ∈ R400 for
each segment st . The C3D was trained with the Sports-1M
dataset [42] that consisted of 1.1 million videos of sports
activities. The C3D features were extracted as indicated in [11]
by uniformly sub-sampling 16 frames out of approximately
60 frames in st (the number of frames in st may vary for
different segments due to the variable framerate of Microsoft
Kinect v2) and subsequently the activations from layer fc6
(i.e., yt ∈ R4096 ) were extracted.
The proposed method was implemented in Chainer [53].
The learning rate was calculated by the adaptive moment
estimation algorithm (Adam) [54] with α = 0.001. Sigmoid
activation was introduced after the fully-connected layers.
Table I summarizes the number of learnable parameters for
each layer, which varies based on the choice of features.
B. Results
For annotating the ground truth, 15 participants were invited
and divided into two groups, namely experienced (E, 5 people)
and inexperienced (NE, 10 people), based on their experience
in the target sport (i.e., Kendo). It was assumed that the
highlights preferred by the E and NE groups would exhibit significant variations, and an aim of the study included evaluating
the extent to which the proposed method adapts to the needs
of each group. For this, the participants annotated manually
the highlights of the 28 videos. The ground truth labels of the
videos were separately obtained for both E and NE groups.
With respect to each one-second period t of video, the ground
truth label is lt = 1 if at least 40% of the participants annotated
it as interesting (i.e., 2 people in group E and 4 people in
7
TABLE I
S IZE OF THE LEARNABLE ELEMENTS IN THE NETWORK WITH RESPECT TO THE FEATURES USED (input × output).
F EATURE VECTOR SIZES ARE DETAILED IN S ECTION IV-A)
lstmJ
fc1
lstmH
fc2
fc3
Body
3D joints
90 × 90
92 × 50
50 × 50
50 × 20
20 × 2
joint-based features only
2D joints
Action recognition
52 × 52
—
54 × 50
402 × 400
50 × 50
400 × 400
50 × 20
400 × 100
20×2
100 × 2
Holistic features only
CNN-ISA
C3D
—
—
400 × 400
4096 × 400
400 × 400
400 × 400
400 × 100
400 × 100
100 × 2
100 × 2
group NE). Otherwise, lt = 0. Due to group E’s technical
knowledge of Kendo, their highlights contain very specific
actions (e.g., decisive strikes and counterattacks). Conversely,
group NE selected strikes as well as more general actions (e.g.,
parries and feints), and thus their labeled highlights are almost
three times as long as group E’s (please refer to the durations
in Appendix A).
The network was separately trained with each group’s
ground truth labels in the leave-one-out (LOO) fashion, i.e.,
27 videos were used for training and a summary of the
remaining video was generated for evaluation purposes. The
CNN for 2D pose estimation was trained independently prior
to each experiment, it was fine-tuned with the 27 training
videos in order to estimate the joints of the video used for
evaluation. This process was repeated for each video and for
each group E and NE, to result in 28 experienced summaries
and 28 inexperienced summaries. The generated summaries
had the same length as their respective ground truth. Figure 5
illustrates a few examples of the frames of a video as well as
highlight frames extracted by the proposed method (framed in
orange).
1) Objective evaluation by segment f-score: The ability of
the proposed method to extract highlights was evaluated in
terms of the f-score. In the proposed method, a one-second
period of video is as follows:
• true positive (TP), if it is in the summary and lt = 1,
• false positive (FP), if it is in the summary but lt = 0,
• false negative (FN), if it is not in the summary but lt = 1,
• true negative (TN), if it is not in the summary and lt = 0.
The f-score is subsequently defined as follows:
2TP
.
(6)
2TP + FP + FN
Table II shows the f-scores for the summaries generated with
the labels of both E and NE groups. In addition to the features
described in Section IV-A, it includes the results of using
the features from our previous work in UGSV summarization
[10]. The features were obtained by feeding the 3D body joint
representation of players’ actions to the action recognition
method in [55] and considering the action classification results.
Additionally, the proposed architecture was also compared
with that of the method used in the previous work [10]
that uses a hidden Markov model with Gaussian mixture
emission (GMM-HMM) over the same action recognition
results mentioned above. Finally, the results of using k-means
clustering are included, since k-means is widely accepted as
a baseline for user-generated video summarization [56]. To
f-score =
Body joint-based and holistic features
3D joints + CNN-ISA
2D joints + CNN-ISA
90 × 90
52 × 52
492 × 400
454 × 400
400 × 400
400 × 400
400 × 100
400 × 100
100 × 2
100 × 2
TABLE II
F- SCORE COMPARISON OF DIFFERENT COMBINATIONS OF FEATURES AND
OTHER UGSV SUMMARIZATION METHODS .
Method
Body joint-based
features
Holistic features
Body joint-based
and holistic features
Other UGSV
summarization
3D joints
2D joints
Action recognition [10]
CNN-ISA
C3D
3D joints + CNN-ISA
2D joints + CNN-ISA
GMM-HMM [10]
k-means clustering
Group E
0.53
0.45
0.48
0.50
0.27
0.58
0.57
0.44
0.28
Group NE
0.83
0.77
0.76
0.79
0.60
0.85
0.81
0.79
0.61
implement the k-means clustering baseline, the video segments
S were clustered based on the concatenated features 3D joints
+ CNN-ISA, and the summary was created by concatenating
in time the cluster centroids. The number of clusters for each
video were configured such that the resulting summary length
is equal to that of the ground truth.
With respect to using a single feature (i.e. 3D joins, 2D
joints, CNN-ISA, C3D, or action recognition), 3D joints obtain
the best performance. Although C3D features perform well
in action recognition tasks [11], the results were significantly
worse than that of other features in the summarization task.
The dimensionality of the C3D features is significantly higher
when compared to that of others, and thus the dataset may
not be sufficient to train the network well. Fine-tuning C3D
using the Kendo dataset might improve its performance. In
contrast, CNN-ISA also uses RGB frames and obtains better
results when compared to those of C3D and even 2D joints.
This implies that it is also possible to obtain features from
RGB frames that allow the modeling of UGSV highlights.
The decrease in the performance of 2D joints with respect to
3D joints may indicate that view variations in the same pose
negatively affect the body joint-based features stream. The
action recognition feature had an intermediate performance.
A potential reason is that the action recognition feature is
based on a classic approach for classification, so useful cues
contained in the 3D body joint positions degenerated in this
process. From these results, the features that performed better
for highlight classification correspond to CNN-ISA holistic
features and 3D body joint-based features.
Several state-of-the-art action recognition methods enjoy
improvements in performance by combining handcrafted
spatio-temporal features (e.g., dense trajectories) and those
learned via CNNs [11], [17]. This is also true in the present
8
𝑐=70%
𝑐 =0% 𝑐 =100%
Summary
Ground truth
Fig. 6. Association of highlights with respect to the greedy algorithm. Each
highlight in the ground truth is uniquely associated to a highlight in the
generated summary (two summary highlights cannot share the same ground
truth highlight). The completeness of a summary highlight corresponds to the
percentage of overlap with the ground truth (0% if unassociated).
work where a combination of CNN-ISA with 3D joints
achieves the best performance. The combination of CNNISA with 2D joints also provides a considerable boost in
performance and especially for the experienced summaries.
This supports our hypothesis that a two-streams architecture
also provides better results for UGSV summarization.
Finally, as shown in the lowest part of Table II, the results
of the proposed method outperform the results of previous
work, as well as those of the clustering-based baseline. While
clustering allows a wider variety of scenes in the summary,
this is not a good strategy for UGSV summarization that
follows a different criterion based on interestingness. The
proposed method also outperforms the previous work [10],
which used the classification results of an action recognition
method to train a GMM-HMM for highlight modeling. We
can conclude that it is not necessary to explicitly recognize
players’ actions for UGSV summarization, which may actually
degrade performance when compared to that in the case of
directly using action recognition features.
2) Objective evaluation by highlight completeness: A highlight may consist of consecutive video segments. Hence,
although missing a segment may not significantly impact the
f-score, it affects the continuity of the video, and thereby the
comprehensibility and the user experience of the summary.
Given this, a criterion is defined to evaluate the completeness
c of an extracted highlights as the fraction of overlap between
the extracted highlight and its associated ground truth highlights (Fig. 6). The association between extracted and ground
truth highlights is not trivial, and it was performed by using
a greedy algorithm in which the total c of all highlights is
maximized. An extracted highlight is considered as a TP if its
completeness c exceeds a certain percentage C%, and based
on this, the precision and recall of the highlights are calculated
as follows:
precision =
TP
,
TP + FP
recall =
TP
.
TP + FN
(7)
In the experiment, the threshold θ was varied from 0 to 1
over the probability p to generate the recall-precision curve of
group E and NE.
Figure 7 shows the curves produced for C = 50%, 70%,
and 90%. We observe that reducing C to 50% significantly
increases the number of complete highlights. The presence of
incomplete highlights is attributed to the way highlights are
extracted. First, the high p segments are extracted, and then the
highlight is completed with low p segments as the threshold
θ decreases (Fig. 8). However, prior to the completion of a
Fig. 7. Recall-precision curves for different completeness values (up: labels
E, down: labels NE). The gap between the curves C = 50% and C = 70%
shows that a significant number of the highlights are missing for a maximum
of half the interesting segments.
highlight, high p segments from other highlights are extracted
and, in a few cases, the low p segments are never extracted.
Specifically, the parts before and after an interesting Kendo
technique normally correspond to low p segments since they
are not present in every ground truth highlight annotated by
the participants.
The reason for the increased number of incomplete segments (less TP) in the NE summaries is because the inexperienced group annotated a higher number of highlights.
3) Subjective evaluation: The same participants who annotated the original videos were asked to participate in a survey
to assess their opinion on the ground truth and the generated
summaries. The three videos with the highest, median and
lowest f-scores (averaged over groups E and N E) were
selected. With respect to each video, participants were shown
the ground truth and the summaries generated with the best
feature combination (i.e., 3D joints + CNN-ISA) using both
group E and NE labels. As a result, each participant watched
12 videos (3 f-scores × 4 video types).
The participants were asked to:
• (Q1) assign a score in a Likert scale from 1 (very
few highlights are interesting) to 5 (most highlights are
interesting) based on their satisfaction with the contents
of each of the 12 videos.
• (Q2) state their opinion on the videos and the criteria
followed while assigning a score.
Table III shows the results of Q1 grouped by video type and
video f-score. The scores are averaged for group E and NE
9
𝑝
𝑝
𝜃
Ground
truth
Time
Fig. 8. Original length: 10 min 40 sec. Summary length: 1 min. The highlights summary is generated by applying a threshold θ to the probability of
interestingness p. Video segments with higher p are extracted prior to segments with lower p, and thus in a few cases the beginning/end segments of the
highlights are missing when compared to the ground truth.
separately.
In the context of Q1, with respect to the video type,
both experienced and inexperienced participants assigned a
higher score to the ground truth videos than to the generated
summaries. This is because some of the summaries contain
uninteresting video segments and also the completeness of
the highlights is worse when compared to that of ground
truth videos. The potential reasons as to why the ground truth
videos did not obtain a perfect score are mainly attributed to
the following two factors: (1) The ground truth summaries
are created by combining labels from several participants via
majority voting, and thus the original labels of each participant are lost. (2) The ground truth also contains incomplete
highlights due to errors when the participants annotated the
videos. Additionally, experienced participants preferred the NE
ground truth to the E summaries plausibly because they do not
find incomplete highlights interesting since context is missing.
Conversely, inexperienced participants tend to appreciate the
highlights from the experienced participants more than their
own highlights. This is potentially because the highlights from
the experienced participants are briefer and contain certain
techniques (e.g. counterattacks) that make summaries more
interesting when compared to those of the inexperienced
participants. The results for Q1 in terms of the f-score type
demonstrate the high correlation to the f-score (i.e., a video
with a higher f-score tends to receive a higher subjective
score).
With respect to Q2, participants provided their opinion on
the summaries. A few experienced participants found the highlights as too short and this even included complete highlights
in the ground truth. This occurs because only the segments
labeled as highlights by at least 40% of the participants (i.e., 2
people in group E and 4 people in group NE) were included in
the ground truth, and thus some labeled segments were left out.
Inexperienced participants state the usefulness of the proposed
method to extract highlights based on interesting actions as
well as time saved by watching the highlights as opposed to the
whole video. In addition, for a few inexperienced participants
incomplete highlights make the summaries difficult to follow.
From this evaluation, we conclude that the labels from
experienced users contain a better selection of Kendo techniques. Due to the negative impact of incomplete highlights
on the summaries, it is necessary to consider extra temporal
consistency in pt . Therefore, although combining the labels of
several participants is convenient to generate the ground truth,
the process introduces incomplete highlights (Section IV-B2)
and alters personal preferences. Thus, instead of combining
labels from different participants, another possibility is to
create personalized summaries with a higher quality ground
truth or to include user profiles such as that proposed in [5].
V. C ONCLUSION
This paper has described a novel method for automatic
summarization of UGSV, especially demonstrating the results for Kendo (Japanese fencing) videos. Given the lack
of editing conventions that permit the use of heuristics, a
different cue, i.e., players’ actions, is used to acquire high-level
semantics from videos to generate a summary of highlights.
The presented two-stream method combines body joint-based
features and holistic features for highlights extraction. The
best combination among the evaluated features corresponds
to a combination of 3D body joint-based features and CNNISA features [12]). In contrast to the previous work [10],
the results indicate that it is not necessary to explicitly
recognize players’ actions in order to determine highlights.
Alternatively, deep neural networks are leveraged to extract a
feature representation of players’ actions and to model their
temporal dependency. Specifically, LSTM is useful to model
the temporal dependencies of the joint positions of players’
bodies in each video segment as well as the highlights in
the entire video. In order to generate appealing summaries,
players’ 3D body joint positions from depth maps offer the
best performance. However, in the absence of depth maps, 2D
body joint positions and holistic features extracted from RGB
images are also used for summarization.
The future work includes fine-tuning the proposed network to illustrate its potential performance and evaluating the
method in the context of a wider variety of sports (e.g., boxing,
fencing, and table tennis).
10
TABLE III
S UBJECTIVE EVALUATION RESULTS WITH RESPECT TO THE VIDEO TYPE AND F - SCORE .
E ACH CELL CONTAINS THE MEAN ± THE STANDARD DEVIATION OF SCORES .
Group E
Group NE
Ground truth E
3.2±0.99
3.57±0.72
Video type
Ground truth NE
Summary E
3.07±1.04
2.6±1.23
3.5±1.07
3.2±0.83
ACKNOWLEDGMENT
This work was supported in part by JSPS KAKENHI Grant
Number 16K16086.
A PPENDIX A
Table IV lists the duration of the videos in the dataset used
in the experiments and their respective ground truth highlights
as annotated by users.
TABLE IV
D URATION OF THE VIDEO DATASET AND GROUND TRUTHS .
ID
#1
#2
#3
#4
#5
#6
#7
#8
#9
#10
#11
#12
#13
#14
#15
#16
#17
#18
#19
#20
#21
#22
#23
#24
#25
#26
#27
#28
Total
Original video
10 min 48 sec
5 min 10 sec
5 min 18 sec
9 min 37 sec
9 min 59 sec
10 min 5 sec
10 min 3 sec
10 min 10 sec
5 min 17 sec
5 min 14 sec
4 min 58 sec
20 min 40 sec
10 min 15 sec
10 min 16 sec
10 min 37 sec
10 min 37 sec
5 min 14 sec
5 min 4 sec
10 min 57 sec
5 min 36 sec
5 min 36 sec
10 min 48 sec
9 min 44 sec
10 min 23 sec
10 min 7 sec
10 min 40 sec
4 min 59 sec
8 min 13 sec
4 hours 6 min 11 sec
Ground truth E
1 min 11 sec
49 sec
1 min 9 sec
1 min 37 sec
2 min 33 sec
1 min 28 sec
48 sec
45 sec
32 sec
22 sec
53 sec
1 min 24 sec
53 sec
58 sec
47 sec
34 sec
16 sec
32 sec
38 sec
27 sec
33 sec
58 sec
1 min 11 sec
54 sec
28 sec
49 sec
33 sec
47 sec
24 min 49 sec
Ground truth NE
2 min 21 sec
1 min 7 sec
1 min 58 sec
2 min 17 sec
2 min 42 sec
2 min 55 sec
1 min 45 sec
2 min 14 sec
1 min 14 sec
1 min 30 sec
1 min 50 sec
4 min 14 sec
2 min 50 sec
5 min 8 sec
2 min 44 sec
2 min 21 sec
1 min 44 sec
2 min 21 sec
2 min 11 sec
1 min 21 sec
1 min 35 sec
1 min 59 sec
2 min 48 sec
2 min 25 sec
1 min 57 sec
2 min 5 sec
2 min 13 sec
2 min 10 sec
1 hour 3 min 59 sec
R EFERENCES
[1] Y. G. Jiang, Q. Dai, T. Mei, Y. Rui, and S. F. Chang, “Super fast
event recognition in internet videos,” IEEE Transactions on Multimedia,
vol. 17, no. 8, pp. 1174–1186, 2015.
[2] R. W. Lienhart, “Dynamic video summarization of home video,” in Proc.
SPIE Electronic Imaging, 1999, pp. 378–389.
[3] M. Otani, Y. Nakashima, E. Rahtu, J. Heikkilä, and N. Yokoya, “Video
summarization using deep semantic features,” in Proc. Asian Conference
on Computer Vision, 2016, pp. 361–377.
[4] W. Li, Z. Zhang, and Z. Liu, “Action recognition based on a bag of
3D points,” in Proc. IEEE Conference on Computer Vision and Pattern
Recognition, 2010, pp. 9–14.
[5] N. Nitta, Y. Takahashi, and N. Babaguchi, “Automatic personalized
video abstraction for sports videos using metadata,” Multimedia Tools
and Applications, vol. 41, no. 1, pp. 1–25, 2009.
Summary NE
2.73±0.87
2.9±0.97
Highest
3.3±0.95
3.48±0.83
Video f-score
Median
Lowest
2.85±0.97
2.55±1.18
3.03±0.91
3.38±0.95
[6] H. Pan, P. Van-Beek, and M. I. Sezan, “Detection of slow-motion replay
segments in sports video for highlights generation,” in Proc. IEEE
International Conference on Acoustics, Speech, and Signal Processing,
vol. 3, 2001, pp. 1649–1652.
[7] A. Ekin, A. M. Tekalp, and R. Mehrotra, “Automatic soccer video
analysis and summarization,” IEEE Transactions on Image Processing,
vol. 12, no. 7, pp. 796–807, 2003.
[8] G. Xu, Y. F. Ma, H. J. Zhang, and S. Q. Yang, “An HMM-based
framework for video semantic analysis,” IEEE Transactions on Circuits
and Systems for Video Technology, vol. 15, no. 11, pp. 1422–1433, 2005.
[9] B. Li and M. I. Sezan, “Event detection and summarization in sports
video,” in Proc. IEEE Workshop on Content-based Access of Image and
Video Libraries, 2001, pp. 132–138.
[10] A. Tejero-de-Pablos, Y. Nakashima, T. Sato, and N. Yokoya, “Human
action recognition-based video summarization for RGB-D personal
sports video,” in Proc. IEEE International Conference on Multimedia
and Expo, 2016, pp. 1–6.
[11] D. Tran, L. Bourdev, R. Fergus, L. Torresani, and M. Paluri, “Learning
spatiotemporal features with 3d convolutional networks,” in 2015 IEEE
International Conference on Computer Vision, 2015, pp. 4489–4497.
[12] Q. V. Le, W. Y. Zou, S. Y. Yeung, and A. Y. Ng, “Learning hierarchical
invariant spatio-temporal features for action recognition with independent subspace analysis,” in Proc. IEEE Conference on Computer Vision
and Pattern Recognition, 2011, pp. 3361–3368.
[13] D. Y. Chen, S. W. Shih, and H. Y. M. Liao, “Human action recognition
using 2-D spatio-temporal templates,” in Proc. IEEE International
Conference on Multimedia and Expo, 2007, pp. 667–670.
[14] H. Meng, N. Pears, and C. Bailey, “A human action recognition system
for embedded computer vision application,” in Proc. IEEE Conference
on Computer Vision and Pattern Recognition, 2007, pp. 1–6.
[15] S. Calderara, R. Cucchiara, and A. Prati, “Action signature: A novel
holistic representation for action recognition,” in Proc. IEEE International Conference on Advanced Video and Signal Based Surveillance,
2008, pp. 121–128.
[16] K. Simonyan and A. Zisserman, “Two-stream convolutional networks for
action recognition in videos,” in Proc. Advances in Neural Information
Processing Systems, 2014, pp. 568–576.
[17] C. Feichtenhofer, A. Pinz, and R. Wildes, “Spatiotemporal residual networks for video action recognition,” in Proc. Conference and Workshop
on Neural Information Processing Systems, 2016, pp. 3468–3476.
[18] M. A. Goodale and A. D. Milner, “Separate visual pathways for
perception and action,” Trends in neurosciences, vol. 15, no. 1, pp. 20–
25, 1992.
[19] L. Xia, C. C. Chen, and J. K. Aggarwal, “View invariant human action
recognition using histograms of 3D joints,” in Proc. IEEE Conference
on Computer Vision and Pattern Recognition, 2012, pp. 20–27.
[20] M. Martı́nez-Zarzuela, F. J. Dı́az-Pernas, A. Tejero-de-Pablos,
D. González-Ortega, and M. Antón-Rodrı́guez, “Action recognition
system based on human body tracking with depth images,” Advances
in Computer Science: an International Journal, vol. 3, no. 1, pp. 115–
123, 2014.
[21] X. Cai, W. Zhou, L. Wu, J. Luo, and H. Li, “Effective active skeleton
representation for low latency human action recognition,” IEEE Transactions on Multimedia, vol. 18, no. 2, pp. 141–154, 2016.
[22] J. Choi, W. J. Jeon, and S. C. Lee, “Spatio-temporal pyramid matching
for sports videos,” in Proc. ACM International Conference on Multimedia Information Retrieval, 2008, pp. 291–297.
[23] M. Chen, S. C. Chen, M. L. Shyu, and K. Wickramaratna, “Semantic
event detection via multimodal data mining,” IEEE Signal Processing
Magazine, vol. 23, no. 2, pp. 38–46, 2006.
[24] D. Tjondronegoro, Y. P. Chen, and B. Pham, “Integrating highlights for
more complete sports video summarization,” IEEE MultiMedia, vol. 11,
no. 4, pp. 22–37, 2004.
11
[25] C. Liang, J. Kuo, W. Chu, and J. Wu, “Semantic units detection and
summarization of baseball videos,” in Proc. IEEE Midwest Symposium
on Circuits and Systems, vol. 1, 2004, pp. 297–300.
[26] A. Divakaran, K. A. Peker, R. Radhakrishnan, Z. Xiong, and R. Cabasson, “Video summarization using mpeg-7 motion activity and audio
descriptors,” in Video Mining. Boston, MA: Springer US, 2003, pp.
91–121.
[27] C. Y. Chen, J. C. Wang, J. F. Wang, and Y. H. Hu, “Motion entropy
feature and its applications to event-based segmentation of sports video,”
EURASIP Journal on Advances in Signal Processing, vol. 2008, pp. 1–8,
2008.
[28] E. Mendi, H. B. Clemente, and C. Bayrak, “Sports video summarization
based on motion analysis,” Computers & Electrical Engineering, vol. 39,
no. 3, pp. 790–796, 2013.
[29] X. S. Hua, L. Lu, and H. J. Zhang, “Optimization-based automated home
video editing system,” IEEE Transactions on Circuits and Systems for
Video Technology, vol. 14, no. 5, pp. 572–583, 2004.
[30] M. Mills, J. Cohen, and Y. Y. Wong, “A magnifier tool for video data,”
in Proc. ACM SIGCHI Conference on Human Factors in Computing
Systems, 1992, pp. 93–98.
[31] S. Lu, Z. Wang, T. Mei, G. Guan, and D. D. Feng, “A bag-of-importance
model with locality-constrained coding based feature learning for video
summarization,” IEEE Transactions on Multimedia, vol. 16, no. 6, pp.
1497–1509, 2014.
[32] Y. F. Ma, X. S. Hua, L. Lu, and H. Zhang, “A generic framework of
user attention model and its application in video summarization,” IEEE
Transactions on Multimedia, vol. 7, no. 5, pp. 907–919, 2005.
[33] G. Evangelopoulos, A. Zlatintsi, A. Potamianos, P. Maragos, K. Rapantzikos, G. Skoumas, and Y. Avrithis, “Multimodal saliency and fusion
for movie summarization based on aural, visual, and textual attention,”
IEEE Transactions on Multimedia, vol. 15, no. 7, pp. 1553–1568, 2013.
[34] W. T. Peng, W. T. Chu, C. H. Chang, C. N. Chou, W. J. Huang, W. Y.
Chang, and Y. P. Hung, “Editing by viewing: automatic home video
summarization by viewing behavior analysis,” IEEE Transactions on
Multimedia, vol. 13, no. 3, pp. 539–550, 2011.
[35] H. Yang, B. Wang, S. Lin, D. Wipf, M. Guo, and B. Guo, “Unsupervised
extraction of video highlights via robust recurrent auto-encoders,” in
Proc. IEEE International Conference on Computer Vision, 2015, pp.
4633–4641.
[36] M. Otani, Y. Nakashima, T. Sato, and N. Yokoya, “Video summarization
using textual descriptions for authoring video blogs,” Multimedia Tools
and Applications, vol. 76, no. 9, pp. 12 097–12 115, 2017.
[37] Y. Song, J. Vallmitjana, A. Stent, and A. Jaimes, “Tvsum: Summarizing
web vvideos using titles,” in Proc. IEEE Conference on Computer Vision
and Pattern Recognition, 2015, pp. 5179–5187.
[38] K. Simonyan and A. Zisserman, “Very deep convolutional networks for
large-scale image recognition,” in Proc. International Conference on
Learning Representations, 2015.
[39] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan,
V. Vanhoucke, and A. Rabinovich, “Going deeper with convolutions,”
in Proc. IEEE Conference on Computer Vision and Pattern Recognition,
2015, pp. 1–9.
[40] Z. Zhang, “Microsoft Kinect sensor and its effect,” IEEE MultiMedia,
vol. 19, no. 2, pp. 4–10, 2012.
[41] S. E. Wei, V. Ramakrishna, T. Kanade, and Y. Sheikh, “Convolutional
pose machines,” in Proc. IEEE Conference on Computer Vision and
Pattern Recognition, 2016, pp. 3714–3722.
[42] A. Karpathy, G. Toderici, S. Shetty, T. Leung, R. Sukthankar, and
L. Fei-Fei, “Large-scale video classification with convolutional neural
networks,” in Proc. IEEE Conference on Computer Vision and Pattern
Recognition, 2014, pp. 1725–1732.
[43] K. H. Zeng, T. H. Chen, J. C. Niebles, and M. Sun, “Title generation
for user generated videos,” in Proc. European Conference on Computer
Vision, 2016, pp. 609–625.
[44] Y. Jia, E. Shelhamer, J. Donahue, S. Karayev, J. Long, R. Girshick,
S. Guadarrama, and T. Darrell, “Caffe: Convolutional architecture for
fast feature embedding,” in Proc. ACM International Conference on
Multimedia, 2014, pp. 675–678.
[45] J. Liu, J. Luo, and M. Shah, “Recognizing realistic actions from videos
“in the wild”,” in Proc. IEEE Conference on Computer Vision and
Pattern Recognition, 2009, pp. 1996–2003.
[46] M. Marszalek, I. Laptev, and C. Schmid, “Actions in context,” in Proc.
IEEE Conference on Computer Vision and Pattern Recognition, 2009,
pp. 2929–2936.
[47] M. D. Rodriguez, J. Ahmed, and M. Shah, “Action MACH a spatiotemporal Maximum Average Correlation Height filter for action recog-
[48]
[49]
[50]
[51]
[52]
[53]
[54]
[55]
[56]
nition,” in Proc. IEEE Conference on Computer Vision and Pattern
Recognition, 2008, pp. 1–8.
K. Soomro, A. R. Zamir, and M. Shah, “UCF101: A dataset of 101
human actions classes from videos in the wild,” CRCV-TR-12-01, 2012.
J. Y.-H. Ng, M. Hausknecht, S. Vijayanarasimhan, O. Vinyals, R. Monga,
and G. Toderici, “Beyond short snippets: Deep networks for video
classification,” in Proc. IEEE Conference on Computer Vision and
Pattern Recognition, 2015, pp. 4694–4702.
B. T. Truong and S. Venkatesh, “Video abstraction: A systematic
review and classification,” ACM Transactions on Multimedia Computing,
Communications, and Applications, vol. 3, no. 1, pp. 3:1–3:37, 2007.
M. Linna, J. Kannala, and E. Rahtu, “Real-time human pose estimation
from video with convolutional neural networks,” CoRR, abs/1609.07420,
2016.
H. Wang, M. M. Ullah, A. Klaser, I. Laptev, and C. Schmid, “Evaluation
of local spatio-temporal features for action recognition,” in Proc. British
Machine Vision Conference, 2009, pp. 124–1.
S. Tokui, K. Oono, S. Hido, and J. Clayton, “Chainer: A next-generation
open source framework for deep learning,” in Proc. Conference and
Workshop on Neural Information Processing Systems, 2015, pp. 1–6.
D. P. Kingma and J. L. Ba, “Adam: a method for stochastic optimization,” in Proc. International Conference on Learning Representations,
2015, pp. 1–13.
A. Tejero-de-Pablos, Y. Nakashima, N. Yokoya, F. J. Dı́az-Pernas, and
M. Martı́nez-Zarzuela, “Flexible human action recognition in depth
video sequences using masked joint trajectories,” EURASIP Journal on
Image and Video Processing, vol. 2016, no. 1, pp. 1–12, 2016.
Y. Cong, J. Yuan, and J. Luo, “Towards scalable summarization of
consumer videos via sparse dictionary selection,” IEEE Transactions
on Multimedia, vol. 14, no. 1, pp. 66–75, 2012.
Antonio Tejero-de-Pablos received a B.E. in telematics engineering and a M.E.in telecommunications
engineering from University of Valladolid, Spain, in
2009 and 2012 respectively. He received a Ph.D. in
information sciences from Nara Institute of Science
and Technology (NAIST), Japan, in 2017. He is
currently a researcher at the University of Tokyo. His
research interests include action recognition, video
summarization, user interfaces and general-purpose
computing on GPU. He is a member of the IEEE.
Yuta Nakashima received a B.E. and a M.E. in
communication engineering and a Ph.D. in engineering from Osaka University, Osaka, Japan in
2006, 2008, and 2012, respectively. He was an
assistant professor at Nara Institute of Science and
Technology from 2012 to 2016 and is currently
an associate professor at the Institute for Datability
Science, Osaka University. He was a visiting scholar
at UNCC in 2012 and at CMU from 2015-2016.
His research interests include computer vision and
machine learning and their applications. His main
research focus includes video content analysis using machine learning approaches. He is a member of IEEE, ACM, IEICE, and IPSJ.
Tomokazu Sato received a B.E. in computer and
system science from Osaka Prefecture University in
1999. He received a M.E. and a Ph.D. in information
sciences from Nara Institute of Science and Technology (NAIST) in 2001 and 2003, respectively. He
was an assistant professor at NAIST during 20032011. He served as a visiting researcher at Czech
Technical University in Prague from 2010-2011. He
is an associate professor at NAIST since 2011. His
research interests include computer vision and mixed
reality.
12
Naokazu Yokoya received a B.E., a M.E., and a
Ph.D. in information and computer sciences from
Osaka University in 1974, 1976, and 1979, respectively. He joined the Electrotechnical Laboratory
(ETL) at the Ministry of International Trade and Industry in 1979. He was a visiting professor at McGill
University in 1986-1987 and was a professor at
Nara Institute of Science and Technology (NAIST)
during 1992-2017. He is the president at NAIST
since April 2017. His research interests include
image processing, computer vision, and mixed and
augmented reality. He is a life senior member of IEEE.
Marko Linna received a M.Sc. in computer science
and engineering from the University of Oulu, Finland, in 2016. He is a research assistant at the Center
for Machine Vision and Signal Analysis (CMVS)
research center at the University of Oulu. His main
research interests include computer vision.
Esa Rahtu received a MSc in 2004 and a Doctor
of Technology in electrical engineering in 2007
from the University of Oulu, Finland. His thesis
was selected as the best doctoral thesis by the
Pattern Recognition Society of Finland. He was also
awarded a highly competitive three-year postdoctoral
scholarship from the Academy of Finland. Currently,
he is an assistant professor at the Tampere University
of Technology in Finland. His main research interests include computer vision and machine learning
with a focus on features extraction, object classification and categorization, face recognition, image retrieval, and registration
and segmentation of deformable scenes.
| 1cs.CV
|
NON SIMPLICITÉ DU GROUPE DE CREMONA SUR TOUT CORPS
ANNE LONJOU
arXiv:1503.03731v3 [math.AG] 14 Sep 2016
Abstract. Using a theorem of F. Dahmani, V. Guirardel and D. Osin we prove that the
Cremona group in 2 dimension is not simple, over any field. More precisely, we show that
some elements of this group satisfy a weakened WPD property which is equivalent in our
particular context to the M. Bestvina and K. Fujiwara’s one. (An English version of this paper
is available on the webpage of the author.)
Introduction
Une question importante lors de l’étude d’un groupe est de savoir s’il est simple ou non, et
dans ce dernier cas de construire des sous-groupes distingués. Ces questions se sont posées dès
la fin du 19ème siècle pour le groupe Bir(P2k ), le groupe des transformations birationnelles du
plan projectif sur un corps k. Cependant, il a fallu attendre 2013 pour que S. Cantat et S. Lamy
[CL13] y répondent dans le cas où k est un corps algébriquement clos.
L’espace de Picard-Manin associé à P2k est la limite inductive des groupes de Picard des surfaces
obtenues en éclatant toutes suites finies de points de P2k , infiniment proches ou non (voir [Man86],
[Can11] et le paragraphe §2.1). Il est muni d’une forme d’intersection de signature (1, ∞). En
considérant une nappe d’hyperboloïde, nous pouvons lui associer un espace hyperbolique de
dimension infinie, noté H∞
k . La stratégie de S. Cantat et S. Lamy est de faire agir, par isométries,
le groupe de Cremona sur cet espace hyperbolique. Ils obtiennent ainsi de nombreux sous-groupes
distingués. Leur article est constitué de deux parties.
Dans la première partie, ils définissent la notion d’élément tendu (« tight element » dans leur
terminologie) que nous énonçons dans le cas particulier du groupe de Cremona, de la manière
suivante : un élément g de Bir(P2k ) est tendu si l’isométrie correspondant à g est hyperbolique et
vérifie deux conditions. La première est que son axe soit rigide, c’est-à-dire si l’axe d’un conjugué
de g est proche de l’axe de g sur une distance suffisamment grande alors les deux axes sont
confondus. La seconde condition est que le stabilisateur de l’axe de g soit le normalisateur du
groupe <g>. Plus précisément, l’élément hyperbolique g doit vérifier :
• Pour tout ε > 0, il existe C ≥ 0 tel que si un élément f dans Bir(P2k ) satisfait l’inégalité
Diam(Tubε (Axe(g))∩Tubε (f Axe(g))) ≥ C alors f Axe(g) = Axe(g), où Tubε signifie le voisinage
tubulaire de rayon ε.
• Pour tout élément f dans Bir(P2k ), si f Axe(g) = Axe(g) alors f gf −1 = g ±1 .
Ce critère leur permet d’établir une variante de la propriété de la petite simplification :
Théorème 1. [CL13] Soit k un corps algébriquement clos. Si g ∈ Bir(P2k ) est tendu alors il
existe un entier n non nul tel que pour tout élément h non trivial appartenant au sous-groupe
normal engendré par g n , le degré de h vérifie deg(h) ≥ deg(g n ). En particulier, le sous-groupe
normal g n de Bir(P2k ) est propre.
2010 Mathematics Subject Classification. 14E07, 20F65.
Key words and phrases. groupe de Cremona, propriété WPD, espace hyperbolique.
1
NON SIMPLICITÉ DU GROUPE DE CREMONA SUR TOUT CORPS
2
La seconde partie consiste à montrer qu’il existe des éléments tendus dans Bir(P2k ). Nous y
reviendrons à la fin de cette introduction.
L’objectif du présent article est d’obtenir une preuve de la non simplicité du groupe de Cremona qui fonctionne pour tout corps k :
Théorème 2. Pour tout corps k, le groupe de Cremona Bir(P2k ) n’est pas simple.
Pour démontrer ce théorème, nous n’allons pas utiliser le fait qu’un élément soit tendu mais
plutôt le fait qu’il satisfasse la propriété WPD (« weakly properly discontinuous »), propriété
proposée par M. Bestvina et K. Fujiwara [BF02] en 2002 dans le contexte du mapping class
group. Un élément g satisfait la propriété WPD, si pour tout ε ≥ 0, il existe un point x et un
entier n strictement positif tels qu’il n’existe qu’un nombre fini d’éléments du groupe G déplaçant x et g n (x) d’au plus ε. Les éléments hyperboliques que nous étudions possèdent un axe. Par
conséquent, nous emploierons de manière équivalente la terminologie introduite par R. Coulon
dans son exposé au séminaire Bourbaki [Cou14] : le groupe G agit discrètement le long de l’axe
de g. L’avantage de la formulation de R. Coulon est d’expliciter le rôle du groupe G. À noter que
D. Osin [Osi16] a englobé cette notion ainsi que d’autres sous le terme d’action acylindrique. Il
unifie ainsi plusieurs travaux concernant des groupes différents.
Récemment, F. Dahmani, V. Guirardel et D. Osin [DGO14] généralisent également la théorie
de la petite simplification pour des groupes agissant par isométries sur des espaces δ-hyperboliques.
Rappelons qu’un espace métrique géodésique X est δ-hyperbolique si pour chaque triangle de
X, tout côté est contenu dans le δ-voisinage de la réunion de ses deux autres côtés. Une de
leur motivation est d’étudier le mapping class group d’une surface de Riemann hyperbolique.
Ce groupe agit sur le complexe des courbes qui est un espace δ-hyperbolique non localement
compact (tout comme l’espace H∞
k mentionné plus haut). Dans ce contexte, ils construisent des
sous-groupes distingués propres qui sont de plus libres et purement pseudo-Anosov, c’est-à-dire
que tous leurs éléments non triviaux sont pseudo-Anosov. Ils répondent ainsi à deux questions
restées longtemps ouvertes.
Le lien entre la théorie de la petite simplification et la propriété WPD se fait au travers de
deux énoncés [Gui14, Theorem 1.3 et Corollary 2.9]. Le premier nous dit que dans le groupe
normal engendré par une famille vérifiant la propriété de petite simplification, les éléments ont
une grande longueur de translation, et le second nous dit que lorsqu’un élément g satisfait la
propriété WPD, les conjugués de <g n > forment une famille satisfaisant la propriété de petite
simplification. En réunissant ces deux énoncés (voir également [DGO14, Theorem 5.3, Proposition
6.34]), nous obtenons le théorème :
Théorème 3. [DGO14] Soit C un nombre réel positif. Soient G un groupe agissant par isométries
sur un espace X δ-hyperbolique et g un élément hyperbolique de G. Si G agit discrètement le long
de l’axe de g alors il existe n ∈ N tel que pour tout élément h non trivial et appartenant au sousgroupe normal engendré par g n , L(h) > C où L est la longueur de translation. En particulier,
pour n assez grand, le sous-groupe normal g n de G est propre. De plus, ce sous-groupe est
libre.
En utilisant ce théorème, la preuve du théorème 2 se résume à exhiber des éléments satisfaisant
la propriété WPD. Nous trouvons de tels éléments dans Aut(A2k ) que nous identifions à un sousgroupe de Bir(P2k ) via l’application qui envoie (x, y) ∈ A2k sur [x : y : 1] ∈ P2k . Notre résultat
principal est alors :
Théorème 4. Soient n ≥ 2 et k un corps de caractéristique ne divisant pas n. Considérons
l’action du groupe Bir(P2k ) sur H∞
où k̄ est la clôture algébrique de k. Le groupe Bir(P2k ) agit
k̄
NON SIMPLICITÉ DU GROUPE DE CREMONA SUR TOUT CORPS
3
discrètement le long de l’axe de l’application :
hn :
A2k −→
(x, y) 7−→
A2k
.
(y, y n − x)
Remarquons que si k est un corps algébriquement clos de caractéristique p > 0, pour tout
entier k ≥ 1, le sous-groupe normal engendré par hkp est le groupe Bir(P2k ). En effet, hp normalise
les translations :
(xp − y, x) ◦ (x + a, y + b) ◦ (y, y p − x) = (x + ap − b, y + a).
Nous obtenons que hkp contient une translation non triviale. Le théorème de Noether permet
alors de montrer l’égalité hkp = Bir(P2k ), voir par exemple [CD13, Proposition 5.12] ou [Giz94,
p.42]. Plus généralement, sur un corps infini de caractéristique divisant n, l’application hn ne
satisfait pas la propriété WPD, ce qui justifie l’énoncé du théorème 4.
En fait, comme conséquence des résultats de F. Dahmani, V. Guirardel et D. Osin, nous
obtenons des propriétés plus précises que la seule non-simplicité :
Théorème 5. Soit k un corps. Le groupe Bir(P2k ) contient des sous-groupes distingués libres, et
est SQ-universel.
Rappelons qu’un groupe G est SQ-universel si tout groupe dénombrable se plonge dans un
quotient de G. Plus de détails se trouvent par exemple dans [Gui14, Theorem 2.14].
Nous terminons cette introduction en comparant notre preuve du théorème 2 à des résultats
récents de [SB15] d’une part, et à la stratégie de [CL13] d’autre part.
Dans son article [SB15, Corollary 7.11], N.I. Shepherd-Barron prouve que tout élément hyperbolique du groupe de Cremona sur un corps k fini engendre un sous-groupe normal propre.
En particulier, Bir(P2k ) n’est pas simple pour tout corps k fini. Dans ce même article (Theorem
7.6), il donne un critère en terme de la longueur de translation de g, pour qu’une transformation hyperbolique g soit tendue, et donc que le sous-groupe distingué engendré par l’une de ses
puissances soit propre.
Théorème 6. [SB15] Supposons que la caractéristique de k soit nulle ou que k soit de caractéristique p > 0 et algébrique sur Fp . Si la longueur de translation d’un élément hyperbolique g de
Bir(P2k ) n’est pas le logarithme d’un entier quadratique, ni de la forme log pn dans le cas où la
caractéristique est strictement positive, alors une certaine puissance de g est tendue.
Une hypothèse sans doute excessive est faite sur le corps dans le cas de la caractéristique
positive (k doit être isomorphe à un sous-corps de F¯p ), dans le but d’éviter le problème mentionné
ci-dessus des transformations normalisant le sous-groupe des translations. Cependant, même
si nous pouvions enlever cette hypothèse et obtenir ainsi une démonstration alternative de la
non simplicité de Bir(P2 ) sur tout corps, une telle preuve reste peu élémentaire car elle repose
lourdement sur les articles [CL13] et [BC16].
Dans ce dernier article, J. Blanc et S. Cantat s’intéressent aux degrés dynamiques de toutes
les applications birationnelles de surface projective. Ils prouvent notamment qu’il n’y a pas de
degré dynamique dans l’intervalle ]1, λL [ où λL est le nombre de Lehmer (« gap property »).
Ils obtiennent comme corollaire que pour tout élément hyperbolique g du groupe de Cremona,
l’indice de <g> dans son centralisateur est fini. Comme remarqué par R. Coulon [Cou14], ceci
implique que si un élément est tendu alors Bir(P2k ) agit discrètement le long de l’axe de cet
élément.
Dans [CL13] une relation entre l’entier n du théorème 1 et la longueur de translation de g est
donnée. Une autre conséquence de la « gap property » est que l’entier n peut être choisi de façon
NON SIMPLICITÉ DU GROUPE DE CREMONA SUR TOUT CORPS
4
10795
uniforme : n ≥ max{ 139347
λL , λL + 374} convient.
Dans leur article [CL13], S. Cantat et S. Lamy exhibent de deux manières différentes des
éléments tendus, selon si le corps est C ou un corps algébriquement clos quelconque.
Dans le cas où k est un corps algébriquement clos, ils éclatent des points particuliers de P2k
pour obtenir une surface S ayant des automorphismes de grands degrés dynamiques. Par exemple,
éclater les 10 points doubles d’une sextique rationnelle donne une telle surface, dite de Coble.
Si le corps n’est pas algébriquement clos, les automorphismes sont à coefficients dans la clôture
algébrique mais pas dans le corps k initial, leur preuve ne s’étend donc pas dans ce cas là.
Dans le cas où le corps est C, ils considèrent un élément « général » du groupe de Cremona
de la forme g = a ◦ J où a ∈ PGL3 (C) et J est une transformation de Jonquières, et ils montrent
qu’il est tendu. Le qualificatif « général » signifie que n’importe quel élément a ∈ PGL3 (C)
convient, quitte à enlever un nombre dénombrable de fermés de Zariski propres dans PGL3 (C).
Ainsi si le corps de base est dénombrable, il se peut qu’il n’y ait pas d’application générale. C’est
pourtant cette méthode que nous généralisons. En effet, dans leur preuve, il est nécessaire que
g satisfasse deux conditions. La première impose aux points base de g et de g −1 d’être dans P2k .
La seconde impose aux points base des itérés de g d’être disjoints de ceux des itérés de g −1 . Les
applications hn que nous considérons se décomposent comme la composée d’une involution de
Jonquières et de l’involution linéaire qui échange les coordonnées, hn = (y, x) ◦ (y n − x, y). La
2
première condition n’est pas vérifiée car hn et h−1
n n’ont qu’un seul point base dans Pk mais nous
verrons que ce n’est pas un problème. Concernant la seconde condition, composer par l’élément
(y, x) ∈ PGL3 (k) permet de séparer les points base des itérés de hn de ceux des itérés de h−1
n .
Ainsi cet élément particulier joue le même rôle que l’élément général a ci-dessus. La preuve présentée dans cet article ne repose pas sur [BC16, CL13] à l’exception du lemme 21 à la section
2.3 qui est une adaptation directe de [CL13, Proposition 5.7].
L’article est organisé comme suit. Dans la première section, nous rappelons la construction
formelle d’un espace hyperbolique de dimension infinie puis nous introduisons la notion de « tube
géodésique », qui nous permet notamment de court-circuiter [CL13, Lemma 3.1]. Le résultat
principal de cette section permet d’affaiblir, dans un espace hyperbolique de dimension infinie,
les hypothèses assurant à un groupe d’agir discrètement le long de l’axe d’un de ses éléments.
Dans la seconde section, nous rappelons la définition de l’espace de Picard-Manin ainsi que la
construction de l’espace hyperbolique H∞
associé. Nous utilisons l’action du groupe Bir(P2k ) sur
k̄
∞
Hk̄ afin d’exhiber des éléments satisfaisant la propriété WPD. Ceci nous permet de prouver le
théorème 4 et par conséquent de montrer que le groupe de Cremona n’est pas simple pour tout
corps.
Remerciements
Je remercie vivement Stéphane Lamy, mon directeur de thèse, qui m’a proposé ce sujet. Je le
remercie également pour ses nombreuses relectures. Merci au rapporteur pour ses commentaires
qui ont permis de rendre certains passages plus clairs.
1. Propriété WPD dans un espace hyperbolique de dimension infinie
L’objectif de cette section est de montrer que lorsqu’un groupe agit sur un espace hyperbolique de dimension infinie, nous pouvons affaiblir les hypothèses qui assurent qu’un groupe agit
discrètement le long de l’axe d’un de ses éléments.
1.1. Espace hyperbolique de dimension infinie. Nous rappelons ici la construction des
espaces hyperboliques de dimension infinie (voir par exemple [Can11, Section 6.3]). Soit (H, h·, ·i)
NON SIMPLICITÉ DU GROUPE DE CREMONA SUR TOUT CORPS
5
un espace de Hilbert réel de dimension infinie. Soient u ∈ H de norme 1 et u⊥ son complémentaire
orthogonal. Tout élément v ∈ H s’écrit de manière unique v = vu u + vu⊥ où vu ∈ R et vu⊥
appartient à u⊥ . Nous définissons une forme bilinéaire symétrique B sur H de signature (1, ∞)
de la manière suivante : B(x, y) = xu yu − hxu⊥ , yu⊥ i. Notons H∞ la nappe d’hyperboloïde définie
par :
H∞ := {x ∈ H | B(x, x) = 1 et B(u, x) > 0}.
L’espace H∞ muni de la distance d définie par cosh d(x, y) := B(x, y) est un espace métrique
complet et de dimension infinie. Remarquons que si l’intersection de H∞ avec un sous-espace
vectoriel de dimension n+1 de H est non vide alors c’est une copie de l’espace hyperbolique usuel
Hn . En particulier, il existe une unique géodésique reliant deux points de H∞ , obtenue comme
l’intersection de H∞ avec le plan vectoriel contenant ces deux points. Dans la pratique, nous nous
ramenons souvent à H2 en considérant un plan hyperbolique de H∞ . De ce fait, tout triangle de
H∞ est isométrique à un triangle de H√2 . Ceci implique que H∞ est CAT(−1) et δ-hyperbolique
pour la même constante δ = log(1 + 2) ≈ 0.881373587 que H2 (voir [Cal07, Example 1.23]).
Dans la suite lorsque nous parlerons d’un espace hyperbolique H∞ , il sera toujours de cette
forme.
Introduisons à présent quelques définitions et notations. Soit f une isométrie de H∞ , sa longueur de translation est définie par L(f ) = inf∞ d(x, f (x)). Si la longueur de translation de
x∈H
f est strictement positive, f est dite hyperbolique. Dans ce cas, elle possède un axe invariant
déterminé par les points réalisant l’infimum :
Axe(f ) := {x ∈ H∞ | d(x, f (x)) = L(f )}.
De plus, f s’étend de manière unique au bord ∂H∞ . Pour plus de détails sur la notion de bord,
nous renvoyons à [BH99, p.260]. L’isométrie hyperbolique f possède exactement deux points fixes
sur ∂H∞ ; l’un répulsif noté b− , et l’autre attractif noté b+ . Ce sont les bouts de l’axe de f que
l’on oriente de b− vers b+ . Nous définissons ainsi une relation d’ordre sur les points de Axe(f ). Le
point x est dit plus petit que y, noté x < y si x ∈ ]b− , y] (et par symétrie y ∈ [x, b+ [). Lorsqu’un
point x est sur l’axe de f nous notons respectivement x − ε et x + ε les deux points sur Axe(f )
situés à distance ε de x tels que x − ε < x + ε. Remarquons que pour tout point x ∈ H∞ , la suite
(f ±n (x))n∈N converge vers b± .
1.2. Propriété WPD. Soit G un groupe qui agit par isométrie sur un espace métrique (X, d).
Pour toute partie A de X et pour toute constante ε ≥ 0, nous notons
Fixε A := {g ∈ G | d(a, ga) ≤ ε ∀a ∈ A},
le fixateur à ε près de A par G. Dans la suite nous ne préciserons plus que les actions considérées
sont isométriques. Le lemme suivant est bien connu.
Lemme 7. Soient G un groupe agissant sur un espace métrique X et g un élément de G. Les
deux propriétés suivantes sont équivalentes :
(1) Il existe y ∈ X tel que pour tout ε ≥ 0, il existe N ∈ N tel que Fixε {y, g N (y)} soit fini.
(2) Pour tout x ∈ X, pour tout ε ≥ 0, il existe N ∈ N tel que Fixε {x, g N (x)} soit fini.
Démonstration. Soit y vérifiant le point (1). Soient x un point quelconque de X et ε ≥ 0. Posons
ε0 = 2 d(x, y) + ε et N l’entier du point (1) associé à ε0 . Par l’inégalité triangulaire, nous avons
l’inclusion :
Fixε {x, g N (x)} ⊂ Fixε0 {y, g N (y)}.
Ainsi, l’ensemble Fixε {x, g N (x)} est fini.
NON SIMPLICITÉ DU GROUPE DE CREMONA SUR TOUT CORPS
6
Nous disons dans la situation du lemme que g satisfait la propriété WPD (« weak proper discontinuity »), ou encore lorsque la notion d’axe est bien définie que le groupe G agit discrètement
le long de l’axe de g (voir [BF02] et [Cou14]).
Dans les prochains paragraphes, nous introduisons la notion de « tube géodésique ». Ceci nous
permet d’affaiblir les hypothèses garantissant la propriété WPD dans le contexte d’un groupe
agissant sur H∞ .
1.3. Quadrilatères hyperboliques. Nous rappelons ici un lemme donnant des relations trigonométriques pour des quadrilatères hyperboliques ayant trois angles droits.
Lemme 8. Soit ADCB un quadrilatère de H2 dont les angles en B, C et D sont droits, alors
la relation suivante est vérifiée :
tanh d(A, B) = tanh d(D, C) cosh d(C, B).
A
e
D
b
d
π
2
−γ
γ
C
a
c
B
Figure 1. Quadrilatère ADCB.
Démonstration. Dans le triangle rectangle ABC, notons a, b, c les longueurs opposées aux sommets A, B, C, et γ l’angle en C, et dans le triangle ACD, d et e, les longueurs opposées aux
sommets A et C (voir Figure 1). Les triangles rectangles ABC et ACD satisfont les relations
classiques (voir par exemple [Bea95, Theorem 7.11.2]) :
π
tanh d
tanh c
tanh a
, sin γ = cos( − γ) =
et tan γ =
.
tanh b
2
tanh b
sinh a
Cela nous donne l’égalité cherchée :
cos γ =
tanh c = tanh d cosh a.
1.4. Tubes géodésiques. Fixons un espace (H∞ , d) comme défini précédemment. Soit Γ une
géodésique de H∞ . Nous introduisons à présent la notion de tube géodésique autour de Γ qui sera
centrale dans la preuve de la proposition principale de cette section (Proposition 11). Le mot
« géodésique » est employé ici pour signifier que lorsque l’on considère l’intersection de ce tube
avec un plan hyperbolique P contenant Γ, le quadrilatère plein obtenu est bordé de segments
géodésiques.
Soient x et x0 deux points de H∞ et Γ la géodésique passant par ces deux points. Soit prΓ la
projection qui envoie un point y ∈ H∞ sur le point de Γ le plus proche de y. Nous notons
∞
Γ⊥
| prΓ y = x}
x := {y ∈ H
l’hyperplan orthogonal à Γ en x. Nous définissons le tube Tεx,x0 comme l’enveloppe convexe de
0
⊥
l’union des deux convexes fermés B̄(x, ε) ∩ Γ⊥
x et B̄(x , ε) ∩ Γx0 (voir Figure 2, à noter que les
tubes sont pleins). Le rayon de ce tube en z ∈ [x, x0 ], noté rεx,x0 (z), est
ε
rεx,x0 (z) := sup{d(z, u) | u ∈ Γ⊥
z ∩ Tx,x0 }.
NON SIMPLICITÉ DU GROUPE DE CREMONA SUR TOUT CORPS
ε
rεx,x0 (z)
n
z
x
B̄(x, ε) ∩
Γ⊥
x
7
Γ
x0
Figure 2. Tube Tεx,x0 .
1
Tεx,x
0
2
Tεy,y
0
Γ
ε2
1
Figure 3. Tube Tεx,x
0 traversant le tube Ty,y 0 .
1
Soient x ≤ y < y 0 ≤ x0 quatre points d’une même géodésique Γ. Nous dirons que le tube Tεx,x
0
ε2
ε1
ε1
0
traverse Ty,y0 si nous avons les inégalités : rx,x0 (y) ≤ ε2 et rx,x0 (y ) ≤ ε2 (voir Figure 3).
ε2
1
Remarque 9. Cette définition implique que pour tout z ∈ [y, y 0 ], rεx,x
0 (z) ≤ ry,y 0 (z). En effet,
ε1
soit u ∈ Γ⊥
z ∩ Tx,x0 , alors u appartient à un segment géodésique ayant ses extrémités dans
ε
ε
ε1
ε2
ε2
ε2
⊥
⊥
1
2
Γ⊥
y ∩ Tx,x0 ⊂ Ty,y 0 et Γy 0 ∩ Tx,x0 ⊂ Ty,y 0 . Par convexité du tube Ty,y 0 , u appartient à Γz ∩ Ty,y 0 .
L’espace H∞ étant CAT(−1) et donc à fortiori CAT(0), la distance est convexe. Cela revient
à dire (voir [BH99, p.176]) que pour tous segments géodésiques [a, b] et [c, d] et pour tous points
z ∈ [a, b] et z 0 ∈ [c, d] ayant les mêmes coordonnées barycentriques, c’est-à-dire d(a, z) = t d(a, b)
et d(c, z 0 ) = t d(c, d) avec t ∈ [0, 1], nous avons d(z, z 0 ) ≤ (1 − t) d(a, c) + t d(b, d). Nous utilisons
la convexité de la distance de la façon suivante.
Lemme 10. Soit G un groupe agissant sur H∞ . Considérons quatre points x, x0 , z, z 0 sur une
géodésique Γ vérifiant x ≤ z < z 0 ≤ x0 . Pour tout ε ≥ 0 nous avons :
(1) Fixε {x, x0 } ⊂ Fixε {z, z 0 }.
(2) le tube Tεx,x0 traverse le tube Tεz,z0 .
d(x,z)
0
0
Démonstration. Posons t := d(x,x
0 ) . Soit g ∈ Fix ε {x, x }. Comme g agit par isométrie, z ∈ [x, x ]
et g · z ∈ [g · x, g · x0 ] ont les mêmes coordonnées barycentriques. Nous avons alors par convexité
de la distance :
d(z, g · z) ≤ t d(x0 , g · x0 ) + (1 − t) d(x, g · x) ≤ ε.
En faisant de même pour z 0 , le point (1) est démontré.
Par le même argument et en se plaçant dans un plan contenant le segment [x, x0 ], nous avons :
rεx,x0 (z) ≤ t rεx,x0 (x0 ) + (1 − t) rεx,x0 (x) ≤ ε.
En faisant de même pour z 0 , le point (2) est prouvé.
1.5. Affaiblissement des hypothèses de la propriété WPD. Nous prouvons dans cette
section le résultat suivant, qui nous permettra de vérifier plus facilement qu’une isométrie satisfait
la propriété WPD.
NON SIMPLICITÉ DU GROUPE DE CREMONA SUR TOUT CORPS
8
Proposition 11. Soient G un groupe agissant sur H∞ et h un élément hyperbolique de G. Les
deux propriétés suivantes sont équivalentes :
(1) Il existe y ∈ Axe(h), η > 0 et n, k ∈ N tels que Fixη {h−k (y), hn (y)} soit fini.
(2) Il existe w ∈ Axe(h) tel que pour tout ε ≥ 0, il existe N ∈ N tel que Fixε {w, hN (w)} soit
fini.
Le point (2) correspond à la définition donnée dans le lemme 7.(1), et le point (1) est la
version avec les hypothèses affaiblies qui nous sera utile dans la section 2. La preuve découle
immédiatement du lemme 13.
Lemme 12. Soient G un groupe agissant sur H∞ et Γ une géodésique de H∞ . Pour toutes
constantes fixées ε ≥ 0, η > 0 et pour tous points y, y 0 , z, z 0 ∈ Γ avec y ≤ z < z 0 ≤ y 0 , le tube
Tεy,y0 traverse le tube Tηz,z0 dès que les distances d(y, z) et d(y 0 , z 0 ) sont suffisamment grandes.
x01
x1
z1
w1
ε
ε
Γ
x
z
w
z0
x0
Figure 4. Trace du tube Tεx,x0 dans le plan P.
Démonstration. Si η ≥ ε, il suffit de prendre y ≤ z et y 0 ≥ z 0 (voir lemme 10.(2)).
Si maintenant η < ε, nous voulons trouver deux points x et x0 sur Γ avec x ≤ z et x0 ≥ z 0
de sorte que rεx,x0 (z) = η et rεx,x0 (z 0 ) = η. Ainsi en utilisant le lemme 10.(2), tous les points
y, y 0 vérifiant d(y, z) ≥ d(x, z) et d(y 0 , z 0 ) ≥ d(x0 , z 0 ) satisferont le lemme. Soit w le milieu de
[z, z 0 ]. Considérons deux points x, x0 ∈ Γ avec x ≤ z et x0 ≥ z 0 et tels que w soit le milieu du
segment [x, x0 ]. Par symétrie, nous pouvons nous placer dans un plan P contenant la géodésique
Γ. Dans ce plan, la trace du tube Tεx,x0 est un quadrilatère plein. Notons x1 et x01 deux sommets
de ce quadrilatère situés dans un même demi-plan délimité par Γ (voir Figure 4). Nous avons
alors d(x1 , x) = d(x01 , x0 ) = ε. Soit w1 le milieu du segment géodésique [x1 , x01 ]. Par symétrie, la
géodésique passant par les points w et w1 est la géodésique orthogonale aux géodésiques [x1 , x01 ]
et Γ. Appelons z1 le point sur le segment géodésique [x1 , w1 ] vérifiant prΓ z1 = z. Le lemme 8
nous donne :
(
tanh d(w, w1 ) cosh d(w, z) = tanh d(z1 , z)
tanh d(w, w1 ) cosh d(w, x) = tanh d(x1 , x) = tanh ε.
En combinant les égalités, nous obtenons :
tanh ε cosh d(w, z)
= tanh d(z, z1 ).
cosh d(w, x)
Prenons x tel que
tanh ε cosh d(w, z)
.
tanh η
Alors d(z, z1 ) = η, et le point w étant le milieu des segments [z, z 0 ] et [x, x0 ], nous obtenons par
symétrie que d(z 0 , z10 ) = η, comme attendu.
d(w, x) = argcosh
NON SIMPLICITÉ DU GROUPE DE CREMONA SUR TOUT CORPS
9
Lemme 13. Soient G un groupe agissant sur H∞ et h une isométrie hyperbolique de G. Pour
toutes constantes fixées ε ≥ 0, η > 0, pour tous points z, z 0 , w ∈ Axe(h) avec z < z 0 , il existe
N, M ∈ N tels que si l’ensemble Fixη {z, z 0 } est fini alors Fixε {h−N (w), hM (w)} l’est aussi.
Tεh−N (w)+ε,hM (w)−ε
η/3
Tz−ε,z0 +ε
Axe(h)
z
w
−N
h
z0
z−ε
(w) + ε
z0 + ε
hM (w) − ε
η/3
Figure 5. Tube Tεh−N (w)+ε,hM (w)−ε traversant le tube Tz−ε,z0 +ε .
Démonstration. Soient ε ≥ 0, η > 0, z,z 0 et w trois points de Axe(h). D’après le lemme 12, il
existe deux entiers N et M suffisamment grands de sorte que le tube Tεh−N (w)+ε,hM (w)−ε traverse
η/3
Tz−ε,z0 +ε (voir Figure 5).
Supposons que l’ensemble Fixε {h−N (w), hM (w)} soit infini. Considérons une suite (fn )n∈N
d’éléments deux à deux distincts de cet ensemble, et notons pn et p0n les projections respectives
de fn (h−N (w)) et fn (hM (w)) sur l’axe de h. Ainsi, pour tout n, nous avons :
[fn (h−N (w)), fn (hM (w))] ⊂ Tεpn ,p0n .
En appliquant le lemme 10.(2) nous obtenons que le tube Tεpn ,p0n traverse le tube Tεh−N (w)+ε,hM (w)−ε
η/3
qui, par construction, traverse le tube Tz−ε,z0 +ε . De plus, par le lemme 10.(1), nous obtenons
l’inclusion :
Fixε {h−N (w), hM (w)} ⊂ Fixε {z, z 0 }.
Par conséquent, pour tout n, fn (z) appartient à la boule fermée B̄(z, ε). Notons zn la projection
de fn (z) sur Axe(h). En particulier, la suite (zn )n∈N appartient au compact [z − ε, z 0 + ε]. Ceci
signifie que d’une part, quitte à prendre une sous-suite, nous pouvons supposer que la suite
(zn )n∈N converge et que pour tout k ≥ 0, d(z0 , zk ) ≤ η3 , et d’autre part que d(zn , fn (z)) ≤ η3 .
Par conséquent, nous avons pour tout n ≥ 0 :
d(f0 (z), fn (z)) ≤ d(f0 (z), z0 ) + d(z0 , zn ) + d(zn , fn (z)) ≤ η.
Quitte à prendre une sous-suite de (fn )n∈N et en considérant maintenant le point z 0 , nous obtenons, par le même argument, l’inégalité précédente en remplaçant z par z 0 . Ainsi, la suite
(f0−1 fn )n≥0 est contenue dans Fixη {z, z 0 } qui est donc infini comme attendu.
2. Application au groupe de Cremona
Soient k un corps et k̄ sa clôture algébrique. Dans cette section, nous rappelons la construction
de l’espace de Picard-Manin sur k̄, et celle de l’espace hyperbolique de dimension infinie associé.
Ainsi, nous pourrons montrer que le groupe Bir(P2k ) agit discrètement le long de l’axe de certains
de ses éléments en utilisant la version affaiblie établie dans la section 1.5.
NON SIMPLICITÉ DU GROUPE DE CREMONA SUR TOUT CORPS
10
2.1. Action du groupe de Cremona sur l’espace de Picard-Manin. Plaçons nous sur k̄.
Rappelons rapidement la construction de l’espace de Picard-Manin et comment le groupe Bir(P2k̄ )
agit dessus. Plus de précisions se trouvent dans [CL13] et [Can11].
Soit S une surface projective lisse. Le groupe de Néron-Severi associé à S, noté N1 (S), est le
groupe des diviseurs à coefficients réels sur S à équivalence numérique près. Il est muni d’une
forme bilinéaire symétrique, la forme d’intersection. Pour tout diviseur D sur S nous notons
{D}S sa classe de Néron-Severi ou {D} s’il n’y a pas d’ambiguïté sur la surface. Si π : S 0 −→ S
est un morphisme birationnel entre deux surfaces, alors le tiré en arrière
π ∗ : N1 (S) ,→ N1 (S 0 )
qui à la classe d’un diviseur associe la classe de sa transformée totale, est un morphisme injectif
qui préserve la forme d’intersection. De plus, N1 (S 0 ) est isomorphe à
(1)
π ∗ (N1 (S)) ⊕ (
R{Ep∗ }),
⊕
p∈Bs(π −1 )
où Bs(π −1 ) est l’ensemble des points base de π −1 (infiniment proches ou pas) et Ep∗ est la
transformée totale, vue dans S 0 , du diviseur exceptionnel Ep obtenu en éclatant le point p. Cette
somme est orthogonale relativement à la forme d’intersection.
Considérons la limite inductive des groupes de Néron-Severi des surfaces S 0 obtenues comme
éclatements de S (nous dirons que de telles surfaces S 0 dominent S) :
Z C (S) = lim N1 (S 0 ),
−→
S 0 →S
où l’indice C fait référence aux b-diviseurs de Cartier (pour plus de précisions voir [Fav10]).
Remarquons que pour toute surface S 0 dominant S, le groupe N1 (S 0 ) est plongé dans Z C (S).
En fait, si nous considérons un diviseur D sur S, à chaque surface S 0 dominant S, nous pouvons
lui faire correspondre une classe de Néron-Severi {D}S 0 dans N1 (S 0 ). Ces éléments sont tous
identifiés dans Z C (S) et correspondent à une classe d de Z C (S) notée en lettre minuscule.
Exemple 14. Considérons un point q appartenant au diviseur exceptionnel Ep , issu de l’éclatement
d’une surface S au point p. Notons Sp la surface obtenue en éclatant le point p et Sp,q celle en
éclatant successivement les points p et q. La classe ep correspond à {Ep }Sp dans N1 (Sp ) et à
{Ẽp + Eq }Sp,q dans N1 (Sp,q ) où Ẽp est la transformée stricte de Ep dans Sp,q .
Définissons la forme d’intersection sur Z C (S). Pour cela, considérons c et d deux éléments de
Z C (S). Il existe une surface S1 dominant S telles que les classes c et d correspondent respectivement à {C}S1 et {D}S1 dans N1 (S1 ). La forme d’intersection est donnée par : c·d = {C}S1 ·{D}S1 .
Montrons qu’elle ne dépend pas du choix de la surface S1 . En effet, considérons une autre surface
S2 où les éléments c et d se réalisent comme des éléments de N1 (S2 ). En résolvant l’application
birationnelle allant de S1 vers S2 nous obtenons une surface S3 dominant S1 et S2 c’est-à-dire
qu’il existe deux morphismes birationnels π : S3 → S1 et σ : S3 → S2 faisant commuter le
diagramme :
S3
π
σ
S1
S2 .
S
NON SIMPLICITÉ DU GROUPE DE CREMONA SUR TOUT CORPS
11
La forme d’intersection étant stable par tiré en arrière nous avons :
{C}S1 · {D}S1 = {C}S3 · {D}S3 = {C}S2 · {D}S2 .
Par la suite, nous nous intéressons à l’espace de Hilbert défini par
X
X
Z(S) = {{D0 }S +
λp ep | λp ∈ R,
λ2p < ∞ et {D0 }S ∈ N1 (S)},
p
que nous appelons l’espace de Picard-Manin (voir [CL13] et [Can11]). C’est le complété L2 de
Z C (S). Les classes ep (nous gardons les notations introduites dans l’exemple 14) où p est un
point de S ou d’une surface dominant S, sont d’auto-intersection −1, orthogonales deux à deux
et orthogonales à N1 (S). La forme d’intersection est donc de signature (1, ∞) et préserve la
décomposition orthogonale :
Z(S) = N1 (S) ⊕ (⊕Rep ).
(2)
p
Exemple 15. Soient p et q les points définis dans l’exemple 14. Considérons les classes de PicardManin ep et eq . Plaçons nous sur la surface Sp,q puisqu’elle domine la surface Sp contenant le
diviseur Ep et qu’elle contient le diviseur Eq . Dans N1 (Sp,q ) les classes ep et eq correspondent
respectivement à {Ẽp + Eq } et à {Eq }. Nous avons donc :
ep · ep = Ẽp2 + Eq2 + 2Ẽp · Eq = −2 − 1 + 2 = −1
ep · eq = Ẽp · Eq + Eq2 = 1 − 1 = 0.
Si π : S 0 → S est un morphisme birationnel, nous avons mentionné que l’application induite
par tiré en arrière sur les groupes de Néron-Severi est une injection. Précisément :
π∗ :
N1 (S)
{D0 }S
,→
7→
{D̃0 }S 0 +
1
NP
(S 0 )
mq (D0 )eq ,
q∈Bs(π −1 )
où D̃0 est la transformée stricte dans S 0 de D0 et mq (D0 ) la multiplicité de D0 au point q.
Regardons ce que nous obtenons au niveau des espaces de Picard-Manin. D’après (2) puis (1),
nous avons la décomposition
0
1
Z(S ) = N (S) ⊕ (
⊕
Req ) ⊕ (
⊕
Rer ).
q∈Bs(π −1 )
−1 )
r ∈Bs(π
/
Nous définissons un isomorphisme π# de Z(S 0 ) vers Z(S) qui consiste à considérer comme exceptionnelles au-dessus de S les classes eq qui étaient dans l’espace de Néron-Severi de S 0 . Ainsi
−1
au niveau des
l’injection π ∗ au niveau des espaces de Néron-Severi devient un isomorphisme π#
espaces de Picard-Manin. Plus précisément, nous avons
−1
π#
: Z(S) = N1 (S) ⊕ (⊕Rep ) −→
Z(S 0 ) = N1 (S 0 ) ⊕ (
p
{D0 }S +
P
λ p ep
p
7→
⊕
Rer )
−1 )
r ∈Bs(π
/
!
{D̃0 }S 0 +
P
(mq (D0 ) + λq )eq
q∈Bs(π −1 )
+
P
λ r er .
−1 )
r ∈Bs(π
/
À présent, considérons l’espace
H∞ (S) = {c ∈ Z(S) | c · c = 1 et c · d0 > 0},
où d0 ∈ N1 (S) est une classe ample. Muni de la distance définie par d(c, c0 ) = argcosh(c · c0 ) pour
tous c, c0 ∈ H∞ (S), c’est un espace hyperbolique de dimension infinie comme ceux introduits
dans la section 1. Nous nous intéressons dans cet article à H∞ (P2k̄ ) que nous notons H∞
. Tout
k̄
P
∞
élément de Hk̄ est de la forme λ` ` + p λp ep où ` correspond à la classe d’une droite dans P2k̄ ,
P
λ` > 0 et λ2` − p λ2p = 1. Considérons une résolution de f ∈ Bir(P2k̄ ) :
NON SIMPLICITÉ DU GROUPE DE CREMONA SUR TOUT CORPS
12
S
π
P2k̄
σ
f
P2k̄ .
Le groupe Bir(P2k̄ ) agit sur H∞
via l’application (f, c) 7→ f# (c) où f# est définie par
k̄
f# = σ# ◦ (π# )−1 .
Remarquons que (f# )−1 = (f −1 )# .
Remarque 16. Si f est un isomorphisme d’un voisinage U de x sur un voisinage V de f (x) alors
nous avons f# (ex ) = ef (x) .
Explicitons sur un exemple l’action de f sur `, la classe d’une droite de P2k̄ .
Exemple 17. Soit f : P2s 99K P2b une application quadratique et notons q1 , q2 et q3 les points
base (infiniment proches ou non) de f −1 . Nous mettons un indice s comme source et un indice b
comme but pour plus de clarté. La transformée par f d’une droite générale L de P2s (ne passant
pas par les points base de f ) est une conique C de P2b qui passe par les points base de f −1 .
Considérons une résolution π : S → P2s et σ : S → P2b de f . Nous voulons exprimer {π ∗ (L)}s
dans la base de N1 (S) issue de σ ∗ , c’est-à-dire dans la base : σ ∗ ({L}b ), {Eq∗1 }b , {Eq∗2 }b , {Eq∗3 }b .
La droite L ne passant pas par les points base de f nous avons π ∗ (L) = L̃ qui correspond à la
transformée stricte de C par σ : C̃ = σ ∗ (C) − Eq∗1 − Eq∗2 − Eq∗3 . Or, σ ∗ ({C}b ) = 2σ ∗ ({L}b ). Ainsi,
nous avons écrit {π ∗ (L)}s dans la base voulue, {π ∗ (L)}s = {2σ ∗ (L) − Eq∗1 − Eq∗2 − Eq∗3 }b . Si nous
réécrivons cette égalité dans les espaces de Picard-Manin nous avons :
f# (`) = 2` − eq1 − eq2 − eq3 .
De manière générale, pour tout f ∈ Bir(P2k̄ ),
f# (`) · ` = deg(f ).
. Par la suite, nous allons nous intéresser aux ap2.2. Action des applications hn sur H∞
k̄
n
plications hn : (x, y) 7→ (y, y − x) où n ≥ 2. Elles appartiennent à Bir(P2k ) qui est inclus dans
Bir(P2k̄ ). Nous nous plaçons donc sur k̄. Fixons l’entier n ≥ 2.
Considérons les applications birationnelles a : (x, y) 7→ (y, x) et jn : (x, y) 7→ (y n − x, y). Nous
avons :
−1
hn = a ◦ jn et h−1
◦ hn ◦ a.
n =a
L’application jn étant une application de Jonquières elle a un point base p0 de multiplicité n − 1
et 2n−2 points base de multiplicité 1 (voir [KSC04, Definition 2.28]). En homogénéisant, le point
p0 a pour coordonnées [1 : 0 : 0]. C’est le seul point base de l’application jn qui appartient à P2k̄ .
Les autres points base de jn forment une tour de points infiniment proches au-dessus de p0 (voir
[Lam02, Lemme 9]). Pour 1 ≤ k ≤ 2n − 2, notons pk le point base de jn appartenant au diviseur
exceptionnel obtenu en éclatant le point base pk−1 . Ainsi le point pk est infiniment proche du point
pk−1 . Plus précisément, la figure 6 représente les transformées strictes des diviseurs exceptionnels
obtenus en résolvant hn munis de leur auto-intersection respective. La configuration obtenue,
illustrée sur la figure, découle du fait que les n − 1 points base p1 , . . . , pn−1 se trouvent sur les
transformées strictes respectives de Ep0 . À l’exception des points p0 et p1 , remarquons que les
points pk dépendent de n. Cependant, pour ne pas alourdir davantage l’écriture nous ne mettons
pas d’indice n. L’application a étant un automorphisme de P2k̄ , hn possède les mêmes points base
que jn . En conjuguant par a, nous obtenons par symétrie les points base de h−1
n que nous notons
{qk }0≤k≤2n−2 . Remarquons que q0 = [0 : 1 : 0].
NON SIMPLICITÉ DU GROUPE DE CREMONA SUR TOUT CORPS
13
Ep0
−n
Epn−1
−2
Epn−2
E pn
−2
−2
..
.
Ep1
Ep2n−3
−2
−2
..
.
Ep2n−2
−1
Figure 6. Arbre des transformées strictes des diviseurs exceptionnels obtenus
en résolvant les points d’indétermination de hn .
L’isométrie hn# est hyperbolique et de longueur de translation log(n) (voir [CL13, Theorem
4.4, Remark 4.5]). Nous allons décrire son axe. Pour alléger l’écriture nous notons e+
n (respectivement e−
)
les
sommes
avec
multiplicités
des
classes
des
diviseurs
exceptionnels
issus
de la
n
résolution de hn (respectivement de hn−1 ) :
( +
en = (n − 1)ep0 + ep1 + · · · + ep2n−2
.
e−
n = (n − 1)eq0 + eq1 + · · · + eq2n−2
L’action sur ` de hn# et de ses itérés est donnée par :
2
2
−
−
hn# (`) = n` − e−
n , hn# (`) = n ` − nen − hn# (en ), etc.
La suite ( n1k hkn# (`))k∈N converge dans l’espace de Picard-Manin vers un élément b+
n d’autointersection 0 qui s’identifie à un point du bord ∂H∞
.
Cet
élément
correspond
à
un
bout de
k̄
−
Axe(hn# ). De même, la suite ( n1k h−k
(`))
converge
vers
un
élément
b
.
Ces
deux
classes
se
k∈N
n#
n
réécrivent :
∞
∞
+
X
X
hin# (e−
h−i
n)
n# (en )
+
−
bn = ` −
et bn = ` −
.
i+1
i+1
n
n
i=0
i=0
L’action de hn sur les classes ep est donnée par la remarque suivante car hn est un automorphisme polynomial.
Remarque 18. Soit f un automorphisme du plan affine A2 de degré au moins 2, que nous étendons
en une application birationnelle de P2 . L’ordre des éclatements dans la résolution minimale de
f −1 est uniquement déterminé. En effet à chaque étape de la résolution l’unique point base propre
est l’image de la droite à l’infini. Ainsi il y a dans ce contexte une notion de « dernier diviseur »
produit lors de la résolution minimale de f −1 , et si p est un point de la droite à l’infini qui n’est
pas le point base propre de f , alors il existe un point q sur ce dernier diviseur produit tel que la
transformation birationnelle induite par f donne un isomorphisme local entre des voisinages de
p et q. En particulier, par la remarque 16, nous avons f# (ep ) = eq .
Lemme 19. Considérons l’ensemble des éléments des 4n − 2 suites (hin# (eqk ))i∈N , (h−i
n# (epk ))i∈N
avec 0 ≤ k ≤ 2n − 2. Ces éléments sont tous deux à deux orthogonaux.
Démonstration. Construisons les points (hn# (eqk )), pour 0 ≤ k ≤ 2n − 2. Les autres points
se construisent de la même façon. L’application hn est régulière en q0 et hn (q0 ) = q0 est un
point base de h−1
n . D’après la remarque 18, il existe un point sur le dernier diviseur exceptionnel
NON SIMPLICITÉ DU GROUPE DE CREMONA SUR TOUT CORPS
14
Eq2n−2 issu de l’éclatement de tous les points {qk }0≤k≤2n−2 , que nous notons q2n−1 , tel que
hn# (eq0 ) = eq2n−1 . Si nous notons π0 l’éclatement du point q0 , alors l’application hn ◦ π0 est
régulière en q1 et hn ◦ π0 (q1 ) = q0 qui est un point base de (hn ◦ π0 )−1 . Toujours d’après la
remarque 18, il existe un point sur le diviseur exceptionnel obtenu en éclatant le point q2n−1 ,
que nous notons q2n , et vérifiant hn# (eq1 ) = eq2n . En réitérant le procédé, nous obtenons qu’il
existe un point, noté q2n−1+k , sur le diviseur exceptionnel issu de l’éclatement du point q2n−2+k
tel que hn# (eqk ) = eq2n−1+k .
À présent, nous pouvons montrer, par récurrence sur i, que hin# (eqk ) = eqi(2n−1)+k pour tout
i ≥ 1 et tout k tel que 0 ≤ k ≤ 2n − 2.
En faisant de même, nous obtenons également que h−i
n# (epk ) = epi(2n−1)+k pour tout i ≥ 1 et
tout k tel que 0 ≤ k ≤ 2n − 2. Ainsi tous les éléments de toutes les suites sont deux à deux
orthogonaux.
Considérons à présent pour chaque n ≥ 2, le point wn qui est le projeté de ` sur Axe(hn# ).
−
L’axe de hn# étant uniquement déterminé par b+
n et bn , wn est une combinaison linéaire de ces
−
2
+ 2
− 2
+
−
deux classes ; wn = αb+
n + βbn . Nous avons 1 = wn = 2αβ car (bn ) = 0 = (bn ) et bn · bn = 1.
De plus, wn · ` = α + β doit être minimal car wn est le projeté de `. Nous obtenons finalement :
∞
−i +
X
√
hin# (e−
1
n ) + hn# (en )
.
(3)
wn = 2` − √ rn où rn =
i+1
n
2
i=0
Remarque 20. Écrivons tous les termes de rn :
h (eq )
eq
ep
eq
eq
ep
rn = (n − 1) 0 + 1 + · · · + 2n−2 + (n − 1) 0 + · · · + 2n−2 + (n − 1) n# 2 0 + · · · .
n
n
n
n
n
n
Le lemme 19 implique que tous les termes de rn sont orthogonaux deux à deux. Ainsi, la classe
de tout diviseur exceptionnel ei a un nombre d’intersection non nul avec au plus un seul des
termes de rn .
Remarquons également, même si nous n’utiliserons pas ce fait, que les points pi (respectivement qi ), construits dans le lemme 19, sont les points base des itérés de hn (respectivement de
2
h−1
n ). Le diagramme suivant donne une idée de preuve dans le cas de hn .
S̃
π̃
σ̃
S
π
P2k̄
S
σ
hn
π
P2k̄
σ
hn
P2k̄
2.3. Preuve du résultat principal. Soient k un corps et n ≥ 2 un entier tels que la caractéristique de k ne divise pas n. Considérons les applications
hn :
A2k −→
(x, y) 7−→
A2k
(y, y n − x)
appartenant au groupe Bir(P2k ). Nous montrons dans ce paragraphe le théorème 4, c’est-à-dire que
le groupe Bir(P2k ) agit discrètement le long de l’axe de hn# . Plus précisément, nous allons montrer
que le groupe Bir(P2k̄ ) agit discrètement le long de l’axe de hn# , ce qui impliquera immédiatement
la même propriété pour le groupe Bir(P2k ). Nous avons besoin du lemme suivant tiré de [CL13,
Proposition 5.7]. Remarquons que nous avons effectué deux modifications dans notre version de
l’énoncé. D’une part nous voulons vérifier le lemme pour les transformations hn , qui ne sont pas
NON SIMPLICITÉ DU GROUPE DE CREMONA SUR TOUT CORPS
f# (`)
b−
n
15
H∞
k̄
εn f# (wn ) +
bn
wn
)
Axe(hn#
`
Figure 7. Action de f sur les classes ` et wn .
des applications générales. D’autre part le rapporteur, que nous remercions, nous a signalé une
erreur de calcul d’intersection dans la preuve, qui en fait se trouvait déjà dans [CL13]. C’est ce
qui explique que dans notre énoncé la constante epsilon dépende de n, alors qu’elle était uniforme
dans l’article original. Ceci est suffisant pour notre but, signalons cependant qu’il serait possible
de préserver l’uniformité de la constante au prix d’une preuve un peu plus longue.
Avant d’énoncer le lemme, rappelons que wn est le projeté sur l’axe de hn# de la classe de `
et est donné par l’égalité (3).
Lemme 21. Il existe εn > 0 tel que tout élément de Fixεn {wn } ⊂ Bir(P2k̄ ) soit un automorphisme
de P2k̄ .
i
√
√ i
1
2 + n√
−
argcosh
2 , cette constante vérifie
Démonstration. Choisissons εn ∈ 0, argcosh
2
alors :
√
√
1
3
argcosh( 2) + εn < argcosh
2 + √ < argcosh √
n 2
2 .
(∗)
√
2 argcosh( 2) + εn < argcosh 4
La deuxième inégalité vient du fait que :
√
5
√
√
√
1
2 argcosh( 2)+εn < argcosh( 2)+argcosh
2+ √ < argcosh( 2)+argcosh √ < argcosh 4,
n 2
2 2
√
5
avec argcosh( 2) + argcosh 2√2 ' 0, 881 + 1, 171 = 2, 052 et argcosh 4 ' 2, 063.
Soit f ∈ Fixεn {wn }. Nous voulons montrer que f est de degré 1. D’après l’inégalité triangulaire
et le fait que d(f# (wn ), wn ) ≤ εn , nous obtenons (voir Figure 7) :
d(f# (`), `) ≤ d(f# (`), f# (wn )) + d(f# (wn ), wn ) + d(wn , `)
≤ 2 d(wn , `) + εn .
Rappelons que cosh d(f# (`), `) = f# (`) · ` = deg f , donc deg f ≤ cosh(2 d(wn , `) + εn ). Re√
marquons que d(wn , `) = argcosh(wn · `) = argcosh 2 (en utilisant (3)). Finalement, d’après le
choix (∗) de εn , nous majorons le degré de f par 4 :
√
deg f ≤ cosh(2 argcosh( 2) + εn ) < 4.
Il nous reste à montrer que f ne peut pas être de degré 2 ou 3. Dans chacun
√ des deux cas
nous raisonnons par l’absurde et arrivons à l’inégalité d(f# (`), wn ) > argcosh( 2) + εn . Or, par
l’inégalité triangulaire, nous obtenons :
d(f# (`), wn ) ≤ d(f# (`), f# (wn )) + d(f# (wn ), wn )
√
≤ d(wn , `) + εn = argcosh( 2) + εn ,
ce qui nous mène à une contradiction.
NON SIMPLICITÉ DU GROUPE DE CREMONA SUR TOUT CORPS
16
Supposons que deg(f ) = 3, alors f −1 possède un point base de multiplicité 2 et quatre points
base de multiplicité 1, d’après les relations de Noether (voir [KSC04, p.39]). Ainsi, l’action de f#
sur la classe ` de Picard-Manin est définie par f# (`) = 3` − 2e0 − e1 − e2 − e3 − e4 avec ei les
classes des diviseurs exceptionnels au-dessus de P2k̄ . D’après la remarque 20, nous obtenons :
1
(−2(n − 1) − (n − 1) − 1 − 1 − 1) = −3,
n
ce qui nous donne −f# (`) · rn ≥ −3. Par hypothèse nous avons f# (`) · ` = 3, ce qui nous permet,
en utilisant (3), d’établir l’inégalité :
√
√
3
1
3
f# (`) · wn = 2f# (`) · ` − √ f# (`) · rn ≥ 3 2 − √ = √ .
2
2
2
En prenant le cosinus hyperbolique réciproque et en utilisant (∗), nous arrivons à la contradiction
attendue :
√
3
d(f# (`), wn ) ≥ argcosh √ > argcosh( 2) + εn .
2
Supposons maintenant que deg f = 2, alors d’après les relations de Noether, f −1 possède trois
points base de multiplicité 1. De même que dans le cas précédent, nous considérons l’action de
f# sur ` ; f# (`) = 2` − e0 − e1 − e2 avec ei les classes des diviseurs exceptionnels au-dessus de
P2k̄ . D’après la remarque 20, nous avons :
(2e0 + e1 + e2 + e3 + e4 ) · rn ≥
−f# (`) · rn = (e0 + e1 + e2 ) · rn ≥
1
1
(−(n − 1) − (n − 1) − 1) = −2 + .
n
n
De plus, f# (`) · ` = 2, ce qui implique :
√
√
√
1
1
2
1
f# (`) · wn = 2f# (`) · ` − √ f# (`) · rn ≥ 2 2 − √ + √ = 2 + √ .
2
2 n 2
n 2
De même que précédemment, nous obtenons la contradiction :
√
√
1
d(f# (`) · wn ) ≥ argcosh
2 + √ > argcosh( 2) + εn .
n 2
Démonstration du théorème 4. Le groupe Bir(P2k ) étant inclus dans le groupe Bir(P2k̄ ), il agit sur
l’espace H∞
. Nous allons montrer que Bir(P2k̄ ) agit discrètement le long de l’axe de hn# , c’estk̄
2
à-dire qu’il existe ε > 0 tel que l’ensemble Fixε {h2n# (wn ), h−2
n# (wn )} ⊂ Bir(Pk̄ ) soit fini. Ainsi,
2
−2
2
2
Fixε {hn# (wn ), hn# (wn )}∩Bir(Pk ) sera fini ce qui signifie que le groupe Bir(Pk ) agit discrètement
le long de l’axe de hn# et le théorème sera démontré.
Choisissons un εn > 0 vérifiant (∗), comme dans le lemme 21. Soit f une application birationnelle de P2k̄ appartenant à Fixεn {h2n# (wn ), h−2
n# (wn )}. Par le lemme 10.(1), nous avons l’inclusion :
−1
Fixεn {h2n# (wn ), h−2
n# (wn )} ⊂ Fixεn {hn# (wn ), hn# (wn )} ⊂ Fixεn {wn }.
De ce fait, f appartient à Fixεn {wn } et en utilisant le lemme 21, nous obtenons que f est de
degré 1 :
f : [x : y : z] 7→ [ax + ky + bz : lx + cy + dz : hx + my + jz].
Le but est maintenant de trouver des contraintes sur les coefficients de f pour montrer qu’il
n’y a qu’un nombre fini de choix possibles pour une telle application. L’application f appartient
−1
−1
−1
aussi à Fixεn {hn# (wn ), h−1
n# (wn )} d’où d(f# hn# (wn ), hn# (wn )) ≤ εn . L’application hn# étant
une isométrie nous avons :
d(hn# f# h−1
n# (wn ), wn ) ≤ εn .
−1
−1
De plus, hn# f# hn# = (hn f hn )# donc d’après le lemme 21 l’application hn f h−1
n est un automorphisme. En considérant la courbe C contractée sur p0 par h−1
(la
droite
à
l’infini)
et le fait
n
que f est un automorphisme, nous avons f (p0 ) = p0 , ce qui implique que l = h = 0. En utilisant
NON SIMPLICITÉ DU GROUPE DE CREMONA SUR TOUT CORPS
17
le même argument et le fait que d(f# hn# (wn ), hn# (wn )) ≤ εn nous obtenons f (q0 ) = q0 qui
implique k = m = 0. À présent, nous pouvons supposer j = 1. Ainsi, f est un automorphisme
affine :
f : (x, y) 7→ (ax + b, cy + d).
−1
Calculons hn f h−1
et
h
f
h
et
regardons
à quelles conditions ce sont des automorphismes de
n
n
n
degré 1 de A2k̄ .
n
n
hn f h−1
n = (y, y − x) ◦ (ax + b, cy + d) ◦ (x − y, x)
= (cx + d, (cx + d)n + ay − axn − b)
= (cx + d, xn (cn − a) + ncn−1 dxn−1 + · · · + ncdn−1 x + ay + dn − b).
De même, nous avons :
n n
n−1 n−1
h−1
by
+ · · · + nabn−1 y + cx + bn − d, ay + b).
n f hn = (y (a − c) + na
−1
2
Comme les applications hn f h−1
n et hn f hn sont des automorphismes affines de Ak̄ , les coefficients
a et c de f vérifient les relations :
cn = a et c = an
Cela signifie que a et c appartiennent à l’ensemble Un2 −1 des racines (n2 − 1)-ème de l’unité car
f est un automorphisme et donc a et c sont non nuls.
Supposons dans un premier temps que n ≥ 3. En considérant les coefficients de xn−1 et y n−1
nous avons : ncn−1 d = nan−1 b = 0. Comme la caractéristique du corps ne divise pas n et que a
et c sont non nuls, nous obtenons
d = b = 0.
Pour n ≥ 3, f est donc de la forme (x, y) 7→ (ax, cy) avec a, c ∈ Un2 −1 et an = c, ainsi l’ensemble
Fixεn {h2n# (wn ), h−2
n# (wn )} est fini.
Si maintenant n = 2, nous ne pouvons pas conclure directement. Par hypothèse, f appartient
à Fixε2 {h22# (w2 ), h−2
2# (w2 )} cela signifie que
−2
2
2
d(f# h−2
2# (w2 ), h2# (w2 )) ≤ ε2 et d(f# h2# (w2 ), h2# (w2 )) ≤ ε2 .
2
Comme h−2
2# et h2# sont des isométries nous avons en fait :
−2
2
d(h22# f# h−2
2# (w2 ), w2 ) ≤ ε2 et d(h2# f# h2# (w2 ), w2 ) ≤ ε2 .
2
D’après le lemme 21, les applications h22 f h−2
et h−2
2
2 f h2 sont de degré 1. De plus le point p0
(respectivement le point q0 ) est le seul point base de h22 (respectivement de h−2
2 ) appartenant
−1
−1
2
à Pk̄ . Par le même argument que précédemment, h2 f h2 préserve p0 et h2 f h2 préserve q0 .
Nous obtenons que 2cd = 0 et 2ab = 0. Comme nous sommes en caractéristique différente
de 2 et que f est un automorphisme nous devons avoir a et c non nuls donc b = d = 0.
Finalement, lorsque n = 2, f est de la forme (x, y) 7→ (ax, cy) avec a, c ∈ U3 et a2 = c.
L’ensemble Fixε2 {h22# (w2 ), h−2
2# (w2 )} est fini comme attendu.
Références
[BC16]
[Bea95]
[BF02]
[BH99]
J. Blanc & S. Cantat – « Dynamical degrees of birational transformations of projective surfaces », J.
Amer. Math. Soc. 29 (2016), p. 415–471.
A. F. Beardon – The geometry of discrete groups, Graduate Texts in Mathematics, vol. 91, SpringerVerlag, New York, 1995.
M. Bestvina & K. Fujiwara – « Bounded cohomology of subgroups of mapping class groups », Geometry and Topology 6 (2002), p. 69–89 (electronic).
M. R. Bridson & A. Haefliger – Metric spaces of non-positive curvature, Grundlehren der Mathematischen Wissenschaften [Fundamental Principles of Mathematical Sciences], vol. 319, Springer-Verlag,
Berlin, 1999.
NON SIMPLICITÉ DU GROUPE DE CREMONA SUR TOUT CORPS
[Cal07]
[Can11]
[CD13]
[CL13]
[Cou14]
[DGO14]
[Fav10]
[Giz94]
[Gui14]
[KSC04]
[Lam02]
[Man86]
[Osi16]
[SB15]
18
D. Calegari – Foliations and the geometry of 3-manifolds, Oxford Mathematical Monographs, Oxford
University Press, Oxford, 2007.
S. Cantat – « Sur les groupes de transformations birationnelles des surfaces », Annals of Mathematics.
Second Series 174 (2011), p. 299–340.
D. Cerveau & J. Déserti – Transformations birationnelles de petit degré, Cours Spécialisés [Specialized
Courses], vol. 19, Société Mathématique de France, Paris, 2013.
S. Cantat & S. Lamy – « Normal subgroups in the Cremona group », Acta Math. 210 (2013), p. 31–94.
R. Coulon – « Théorie de la petite simplification : une approche géométrique », Séminaire Bourbaki
1089 (2014).
F. Dahmani, V. Guirardel & D. Osin – « Hyperbolically embedded subgroups and rotating families
in groups acting on hyperbolic spaces », À paraître aux Memoirs of the A.M.S., arxiv :1111.7048v5
(2014).
C. Favre – « Le groupe de Cremona et ses sous-groupes de type fini », Astérisque (2010), no. 332,
p. Exp. No. 998, vii, 11–43, Séminaire Bourbaki. Volume 2008/2009. Exposés 997–1011.
M. H. Gizatullin – « The decomposition, inertia and ramification groups in birational geometry », in
Algebraic geometry and its applications (Yaroslavl’, 1992), Aspects Math., E25, Vieweg, Braunschweig,
1994, p. 39–45.
V. Guirardel – « Geometric small cancellation », in Geometric Group Theory (M. Bestvina, M. Sageev
& K. Vogtmann, éds.), vol. 21, IAS/Park City Mathematics Institute, 2014.
J. Kollár, K. E. Smith & A. Corti – Rational and nearly rational varieties, Cambridge Studies in
Advanced Mathematics, vol. 92, Cambridge University Press, Cambridge, 2004.
S. Lamy – « Une preuve géométrique du théorème de Jung », Enseign. Math. (2) 48 (2002), no. 3-4,
p. 291–315.
Y. I. Manin – Cubic forms, North-Holland Mathematical Library, vol. 4, North-Holland Publishing
Co., Amsterdam, 1986.
D. Osin – « Acylindrically hyperbolic groups », Trans. Amer. Math. Soc. 368 (2016), p. 851–888.
N. Shepherd-Barron – « Some effectivity questions for plane Cremona transformations »,
arxiv :1311.6608v4 (2015).
Institut de Mathématiques de Toulouse, Université Paul Sabatier, 118 route de Narbonne, 31062
Toulouse Cedex 9, France
E-mail address: [email protected]
| 4math.GR
|
arXiv:1101.2245v3 [cs.DS] 4 Oct 2015
Invertible Bloom Lookup Tables
Michael T. Goodrich
Dept. of Computer Science
University of California, Irvine
http://www.ics.uci.edu/˜goodrich/
Michael Mitzenmacher
Dept. of Computer Science
Harvard University
http://www.eecs.harvard.edu/˜michaelm/
Abstract
We present a version of the Bloom filter data structure that supports not only
the insertion, deletion, and lookup of key-value pairs, but also allows a complete
listing of the pairs it contains with high probability, as long the number of keyvalue pairs is below a designed threshold. Our structure allows the number of keyvalue pairs to greatly exceed this threshold during normal operation. Exceeding the
threshold simply temporarily prevents content listing and reduces the probability
of a successful lookup. If entries are later deleted to return the structure below
the threshold, everything again functions appropriately. We also show that simple
variations of our structure are robust to certain standard errors, such as the deletion
of a key without a corresponding insertion or the insertion of two distinct values
for a key. The properties of our structure make it suitable for several applications,
including database and networking applications that we highlight.
1
Introduction
The Bloom filter data structure [5] is a well-known way of probabilistically supporting
dynamic set membership queries that has been used in a multitude of applications (e.g.,
see [8]). The key feature of a standard Bloom filter is the way it trades off query
accuracy for space efficiency, by using a binary array T (initially all zeroes) and k
random hash functions, h1 , . . . , hk , to represent a set S by assigning T [hi (x)] = 1 for
each x ∈ S. To check if x ∈ S one can check that T [hi (x)] = 1 for 1 ≤ i ≤ k,
with some chance of a false positive. This representation of S does not allow one to
list out the contents of S given only T . This aspect of Bloom filters is sometimes
viewed as a feature, in settings where some degree of privacy protection is desired
(e.g., see [2, 3, 18, 27, 30]). Still, in many domains one would benefit from a similar
set representation that would also allow listing out the set’s contents [17].
1
In this paper, we are interested not in simply representing a set, but instead in methods for probabilistically representing a lookup table (that is, an associative memory) of
key-value pairs, where the keys and values can be represented as fixed-length integers.
Unlike previous approaches (e.g., see [6, 9]), we specifically desire a data structure that
supports the listing out of all of its key-value pairs. We refer to such a structure as an
invertible Bloom lookup table (IBLT).
1.1
Related Work
Our work can be seen as an extension of the invertible Bloom filter data structure of
Eppstein and Goodrich [17], modified to store key-value pairs instead of only keys.
Our analysis, however, supersedes the analysis of the previous paper in several respects, in terms of efficiency and tightness of the analysis (as well as correcting some
small deficiencies). In particular, our analysis demonstrates the natural connection between these data structures and cores of random hypergraphs, similar to the connection
found previously for cuckoo hashing and erasure-correcting codes (e.g., see [16, 23]).
This provides both a significant constant factor reduction in the required space for the
data structure, as well as an important reduction in the error probability (to inverse
polynomial, from constant in [17]). In addition, our IBLT supports some usage cases
and applications (discussed later in this section) that are not supported by a standard
invertible Bloom filter.
Given its volume, reviewing all previous work on Bloom filters is, not surprisingly,
beyond the scope of this paper (e.g., see [7, 8, 33] for some excellent surveys). Nevertheless, two closely related works include Bloomier filters [9] and Approximate Concurrent State Machines (ACSMs) [6], which are structures to store and track key-value
pairs. Song et al. [35] store elements in a Bloom-like hash table in a fashion that is
somewhat similar to that of a cuckoo hash table [15, 26, 29]. Cohen and Matias [12] describe an extended Bloom filter that allows for multiplicity queries on multisets. While
an IBLT is not intended for changing values, it can be used in such settings where a
key-value pair can be explicitly deleted and a new value for the key re-inserted. Again,
an IBLT has additional features, including listing, graceful handling of data exceeding the listing threshold, and counting multiplicities, which make it useful for several
applications where these other structures are insufficient.
Another similar structure is the recently developed counter braid architecture [22],
which keeps an updatable count field for a set of flows in a compressed form, with the
compressed form arising by a careful use of hashing and allowing reconstruction of
the count for each flow. Unlike an IBLT, however, the flow list must be kept explicitly
to read out the flow counts, and such lists do not allow for direct lookups of individual values. There are several other differences from our work due to their focus on
counters, but perhaps most notable is their decoding algorithm, which utilizes belief
propagation. Additional work in the area of approximate counting of a similar flavor
but with very different goals from the IBLT includes the well-known CM-sketch [13]
and recent work by Price [32].
2
1.2
Our Results
We present a deceptively simple variation of the Bloom filter data structure that is
designed for key-value pairs and further avoids the limitation of previous structures
(such as [6, 9]) that do not allow the listing of contents. As mentioned above, we call
our structure an invertible Bloom lookup table, or IBLT for short. Our IBLT supports
insertions, deletions, and lookups in O(k) time, where k is the number of random
hash functions used (which will typically be a constant in practice). Just as Bloom
filters have false positives, our lookup operation works only with constant probability,
although this probability can be made quite close to 1.
Our data structure also allows for a complete listing of the key-value pairs the
structure contains, with high probability, whenever the current number, n, of such pairs
lies below a certain threshold capacity, t, a parameter that is part of the structure’s
design. This listing takes O(t) time. In addition, because the content-listing operation
succeeds with high probability, one can also use it as a backup in the case that a standard
lookup fails—namely, if a lookup fails, perform a listing of key-value pairs until one
can retrieve a value for the desired key.
Our IBLT construction is also space-efficient, requiring space1 at most linear in t,
the threshold number of keys, even if the number, n, of stored key-value pairs grows
well beyond t (for example, to polynomial in t) at points in time. One could of course
instead keep an actual list of key-value pairs with linear space, but this would require
space linear in n, i.e., the maximum number of keys, not the target number, t, of keys.
Keeping a list also necessarily requires more computationally expensive lookup operations than our approach supports.
We further show that with some additional checksums we can tolerate various natural errors in the system. For example, we can cope with key-value pairs being deleted
without first being inserted, or keys being inserted with the same value multiple times,
or keys mistakenly being inserted with multiple values simultaneously. Interestingly,
together with its contents-listing ability, this error tolerance leads to a number of applications of the IBLT, which we discuss next.
1.3
Applications and Usage Cases
There are a number of possible applications and usage cases for invertible Bloom
lookup tables, some of which we explore here.
Database Reconciliation Suppose Alice and Bob hold distinct, but similar, copies,
DA and DB , of an indexed database, D, and they would like to reconcile the differences between DA and DB . For example, Alice could hold a current version of D
and Bob could hold a backup, or Alice and Bob could represent two different copies
of someone’s calendar database (say, respectively on a desktop computer and a smartphone) that now need to be symmetrically synchronized. Such usage cases are natural
in database systems, particularly for systems that take the approach advocated in an
1 As in the standard RAM model, we assume in this paper that keys and values respectively fit in a
single word of memory (which, in practice, could actually be any fixed number of memory words), and we
characterize the space used by our data structure in terms of the number of memory words it uses.
3
interesting recent CACM article by Stonebraker [36] that argues in favor of sacrificing consistency for the sake of availability and partition-tolerance, and then regaining
consistency by performing database reconciliation computations whenever needed. Incidentally, in a separate work, Eppstein et al.2 are currently empirically exploring a
similar application of the invertible Bloom table technology for the reconciliation of
two distributed mirrors of a filesystem (say, in a peer-to-peer network).
To achieve such a reconciliation with low overhead, Alice constructs an IBLT, B,
for DA , using indices as keys and checksums of her records as values. She then sends
the IBLT B to Bob, who then deletes index-checksum pairs from B corresponding to all
of his entries in DB . The remaining key-value pairs corresponding to insertions without deletions identify records that Alice has that Bob doesn’t have, and the remaining
key-value pairs corresponding to deletions without insertions identify records that Bob
has that Alice doesn’t have. In addition, as we show, Bob can also use B to identify
records that they both possess but with different checksums. In this way, Alice needs
only to send a message B of size O(t), where t here is an upper bound on the number of differences between DA and DB , for Bob to determine the identities of their
differences (and a symmetric property holds for a similar message from Bob to Alice).
Tracking Network Acknowledgments As another example application, consider a
router R that would like to track TCP sessions passing through R. In this case, each
session corresponds to a key, and may have an associated value, such as the source
or the destination, that needs to be tracked. When such flows are initiated in TCP,
particular control messages are passed that can be easily detected, allowing the router
to add the flow to the structure. Similarly, when a flow terminates, control messages
ending the flow are sent. The IBLT supports fast insertions and deletions and can be
used to list out the current flows in the system at various times, as long as the number
of flows is less than some preset threshold, t. Note this work can be offloaded simply
by sending a copy of the IBLT to an offline agent if desired. Furthermore, the IBLT can
return the value associated with a flow when queried, with constant probability close
to 1. Finally, if at various points, the number of flows spikes above its standard level to
well above t, the IBLT will still be able to list out the flows and perform lookups with
the appropriate probabilities once the total load returns to t or below. Again, this is a
key feature of the IBLT; all keys and values can be reconstructed with high probability
whenever the number of keys is below a design threshold, but if the number of keys
temporarily exceeds this design threshold and later returns to below this threshold, then
the functionality will return at that later time.
In this networking setting, sometimes flows do not terminate properly, leaving them
in the data structure when they should disappear. Similarly, initialization messages
may not be properly handled, leading to a deletion without a corresponding insertion.
We show that the IBLT can be modified to handle such errors with minimal loss in
performance. Specifically, we can handle keys that are deleted without being inserted,
or keys that erroneously obtain multiple values. Even with such errors, we provide
conditions for which all valid flows can still all be listed with high probability. Our
experimental results also highlight robustness to these types of errors. (Eventually, of
2 Personal
communication.
4
course, such problematic keys should be removed from the data structure. We do not
concern ourselves with removal policies here; see [6] for some possibilities based on
timing structures.)
Oblivious Selection from a Table As a final motivating application, consider a scenario where Alice has outsourced her data storage needs, including the contents of an
important indexed table, T , of size n, to a cloud storage server, Bob, because Alice has
very limited storage capacity (e.g., Alice may only have a smartphone). Moreover, because her data is sensitive and she knows Bob is honest-but-curious regarding her data,
she encrypts each record of T using a secret key, and random nonces, so that Bob cannot determine the contents of any record from its encryption alone. Such encryptions
are not sufficient, however, to fully protect the privacy of Alice’s data, as recent attacks
show that the way Alice accesses her data can reveal its contents (e.g., see [10]). Alice
needs a way of hiding any patterns in the way she accesses her data.
Suppose now that Alice would like to do a simple SELECT query on T and she
is confident that the result will have a size at most t, which is much less than n but
still more than she can store locally. Thus, she cannot use techniques from private
information retrieval [11, 37], as that would either require storing results back with
Bob in a way that could reveal selected indices or using yet another server besides
Bob. She could use techniques from recent oblivious RAM simulations [19, 20, 31] to
obfuscate her access patterns, but doing so would require O(n log2 n) I/Os. Therefore,
using existing techniques would be inefficient.
By using an IBLT, on the other hand, she can perfrom her SELECT query much more
efficiently. The advantage comes from the fact that an insertion in an IBLT accesses a
random set of cells (that is, memory locations) whose addresses depend (via random
hash functions) only on the key of the item being inserted. Alice thus uses all the
indices for T as keys, one for each record, and accesses memory as though inserting
each record into an IBLT of size O(t). In fact, Alice only inserts those records that
satisfy her SELECT query. However, since Alice encrypts each write using a secret
key and random nonces, Bob cannot tell when Alice’s write operations are actually
changing the records stored in the IBLT, and when a write operation is simply rewriting
the same contents of a cell over again re-encrypted with a different nonce. In this way
Alice can obliviously create an IBLT of size O(t) that contains the result of her query
and is stored by Bob. Then, using existing methods for oblivious RAM simulation [20],
she can subsequently obliviously extract the elements from her IBLT using O(t log2 t)
I/Os. With this approach Bob learns nothing about her data from her access pattern. In
addition, the total number of I/Os for her to perform her query is O(n + t log2 t), which
is linear (and optimal) for any t that is O(n/ log2 n). We are not currently aware of any
other way that Alice can achieve such a result using a structure other than an IBLT.
5
2
A Simple Version of the Invertible Bloom Lookup Table
In this section, we describe and analyze a simple version of the IBLT. In the sections
that follow we describe how to augment and extend this simple structure to achieve
various additional performance goals.
The IBLT data structure, B, is a randomized data structure storing a set of key-value
pairs. It is designed with respect to a threshold number of keys, t; when we say the
structure is successful for an operation with high probability it is under the assumption
that the actual number of keys in the structure at that time, which we henceforth denote
by n, is less than or equal to t. Note that n can exceed t during the course of normal
operation, however.
As mentioned earlier, we assume throughout that, as in the standard RAM model,
keys and values respectively fit in a single word of memory (which, in practice, could
actually be any fixed number of memory words) and that each such word can alternatively be viewed as an integer, character string, floating-point number, etc. Thus,
without loss of generality, we view keys and values as positive integers.
In many cases we take sums of keys and/or values; we must also consider whether
word-value overflow when trying to store these sums in a memory word. (That is, the
sum is larger than what fits in a data word.) As we explain in more detail at appropriate
points below, such considerations have minimal effects. In most situations, with suitably sized memory words, overflow may never be a consideration. Alternatively, if we
work in a system that supports graceful overflows, so that (x + y) − y = x even if the
first sum results in an overflow, our approach works with negligible changes. Finally,
we can also work modulo some large prime (so that vaues fit within a memory word)
to enforce graceful overflow. These variations have negligible effects on the analysis.
However, we point out that in many settings (except in the case where we may have
duplicate copies of the same key-value pair), we can use XORs in place of sums in our
algorithms, and avoid overflow issues entirely.
2.1
Operations Supported
Our structure supports the following operations:
• INSERT(x, y): insert the key-value pair, (x, y), into B. This operation always
succeeds, assuming that all keys are distinct.
• DELETE(x, y): delete the key-value pair, (x, y), from B. This operation always
succeeds, provided (x, y) ∈ B, which we assume for the rest of this section.
• GET(x): return the value y such that there is a key-value pair, (x, y), in B. If
y = null is returned, then (x, y) 6∈ B for any value of y. With low (but constant)
probability, this operation may fail, returning a “not found” error condition. In
this case there may or may not be a key-value pair (x, y) in B.
• LIST E NTRIES(): list all the key-value pairs being stored in B. With low (inverse
polynomial in t) probability, this operation may return a partial list along with an
6
“list-incomplete” error condition.
When an IBLT B is first created, it initializes a lookup table T of m cells. Each
of the cells in T stores a constant number of fields, each of which corresponds to a
single memory word. We emphasize that an important feature of the data structure is
that at times the number of key-value pairs in B can be much larger than m, but the
space used for B remains O(m) words. (We discuss potential issues with word-value
overflow where appropriate.) The INSERT and DELETE methods never fail, whereas the
GET and LIST E NTRIES methods, on the other hand, only guarantee good probabilistic
success when n ≤ t. For our structures we shall generally have m = O(t), and often
we can give quite tight analyses on the constants required, as we shall see below.
2.2
Data Structure Architecture
Like a standard Bloom filter, an IBLT uses a set of k random3 hash functions, h1 , h2 ,
. . ., hk , to determine where key-value pairs are stored. In our case, each key-value
pair, (x, y), is placed into cells T [h1 (x)], T [h2 (x)], . . . T [ht (x)]. In what follows,
for technical reasons4 , we assume that the hashes yield distinct locations. This can
be accomplished in various ways, with one standard approach being to split the m
cells into k subtables each of size m/k, and having each hash function choose one cell
(uniformly) from each subtable. Such splitting does not affect the asymptotic behavior
in our analysis and can yield other benefits, including ease of parallelization of reads
and writes into the hash table. (Another approach would be to select the first k distinct
hash values from a specific sequence of random hash functions.)
Each cell contains three fields:
• a count field, which counts the number of entries that have been mapped to this
cell,
• a keySum field, which is the sum of all the keys that have been mapped to this
cell,
• a valueSum field, which is the sum of all the values that have been mapped to
this cell.
Given these fields, which are initially 0, performing the update operations is fairly
straightforward:
• INSERT(x, y):
for each (distinct) hi (x), for i = 1, . . . , k do
3 We assume, for the sake of simplicity in our analysis, that the hash functions are fully random, but
this does not appear strictly required. For example, the techniques of [24] can be applied if the data has a
sufficient amount of entropy. For worst-case data, we are not aware of any results regarding the 2-core of
a random hypergraph where the vertices for each edge are chosen according to hash functions with limited
independence, which, as we will see, would be needed for such a result. Similar graph problems with limited
independence have recently been studied in [1]. It is an interesting theoretical question to obtain better
bounds on the randomness needed for our proposed IBLT data structure.
4 Incidentally, this same technicality can be used to correct a small deficiency in the paper of Eppstein and
Goodrich [17].
7
add 1 to T [hi (x)].count
add x to T [hi (x)].keySum
add y to T [hi (x)].valueSum
end for
• DELETE(x, y):
for each (distinct) hi (x), for i = 1, . . . , k do
subtract 1 from T [hi (x)].count
subtract x from T [hi (x)].keySum
subtract y from T [hi (x)].valueSum
end for
2.3
Data Lookups
We perform the GET operation in a manner similar to how membership queries are
done in a standard Bloom filter. The details are as follows:
• GET(x):
for each (distinct) hi (x), for i = 1, . . . , k do
if T [hi (x)].count = 0 then
return null
else if T [hi (x)].count = 1 then
if T [hi (x)].keySum = x then
return T [hi (x)].valueSum
else
return null
end if
end if
end for
return “not found”
Recall that for now we assume that all insertions and deletions are done correctly,
that is, no insert will be done for an existing key in B and no delete will be performed
for a key-value pair not already in B. With this assumption, if the above operation
returns a value y or the null value, then this is the correct response. This method may
fail, returning “not found,” if it can find no cell that x maps to that holds only one entry.
Also, as a value is returned only if the count is 1, overflow of the sum fields is not a
concern.
For a key x that is in B, consider the probability p0 that each of its hash locations
contains no other item. Using the standard analysis for Bloom filters (e.g., see [8]), we
find p0 is:
(n−1)
k
≈ e−kn/m .
p0 = 1 −
m
That is, assuming the table is split into k subtables of size m/k (one for each hash
function), each of the other n − 1 keys misses the location independently with probability 1 − k/m. One nice interpretation of this is that the number of keys that hash to
8
the cell is approximately a Poisson random variable with mean kn/m, and e−kn/m is
the corresponding probability a cell is empty. The probability that a GET for a key that
is in B returns “not found” is therefore approximately
k
(1 − p0 )k ≈ 1 − e−kn/m ,
which corresponds to the false-positive rate for a standard Bloom filter. As is standard
for these arguments, these approximations can be readily replaced by tight concentration results [8].
The probability that a GET for a key that is not in B returns“not found” instead of
null can be found similarly. Here, however, note that every cell hashed to by that key
must be hashed to by at least two other keys from B. This is because an empty cell
would lead to a null return value, and a cell with just one key hashed to it would yield
the corresponding true key value, and hence also lead to a null return value for a key
not in B. Using the same approximation – specifically, that the number of keys from
B that land in a cell is approximately distributed as a discrete Poisson random variable
with mean kn/m – we find this probability is
2.4
1 − e−kn/m −
kn −kn/m
e
m
k
.
Listing Set Entries
Let us next consider the method for listing the contents of B. We describe this method
in a destructive fashion—if one wants a non-destructive method, then one should first
create a copy of B as a backup.
• LIST E NTRIES():
while there’s an i ∈ [1, m] with T [i].count = 1 do
add the pair (T [i].keySum , T [i].valueSum) to the output list
call DELETE(T [i].keySum , T [i].valueSum)
end while
It is a fairly straightforward exercise to implement this method in O(m) time, say, by
using a link-list-based priority queue of cells in T indexed by their count fields and
modifying the DELETE method to update this queue each time it deletes an entry from
B.
If at the end of the while-loop all the entries in T are empty, then we say that the
method succeeded and we can confirm that the output list is the entire set of entries
in B. If, on the other hand, there are some cells in T with non-zero counts, then the
method only outputs a partial list of the key-value pairs in B.
This process should appear entirely familiar to those who work with random graphs
and hypergraphs. It is exactly the same procedure used to find the 2-core of a random
hypergraph (e.g., see [16, 25]). To make the connection, think of the cells as being vertices in the hypergraph, and the key-value pairs as being hyperedges, with the vertices
for an edge corresponding to the hash locations for the key. The 2-core is the largest
9
sub-hypergraph that has minimum degree at least 2. The standard “peeling process”
finds the 2-core: while there exists a vertex with degree 1, delete it and the corresponding hyperedge. The equivalence between the peeling process and the scheme for
LIST E NTRIES is immediate. We note that this peeling process is similarly used for various erasure-correcting codes, such as Tornado codes and its derivatives (e.g., see [23]),
that have, in some ways, the same flavor as this construction5 .
Assuming that the cells associated with a key are chosen uniformly at random, we
use known results on 2-cores of random hypergraphs. In particular, tight thresholds
are known; when the number of hash values k of each is at least 2, there are constants
ck > 1 such that if m > (ck + )n for any constant > 0, LIST E NTRIES succeeds
with high probability, that is with probability 1 − o(1). Similarly, if m < (ck − )n for
any constant > 0, LIST E NTRIES succeeds with probability o(1). Hence t = m/ck
is (approximately) the design threshold for the IBLT. As can be found in for example
[16, 25], these values are given by
n
o
−kαxk−1
c−1
=
sup
α
:
0
<
α
<
1;
∀x
∈
(0,
1),
1
−
e
<
x
.
k
It is easy to check from this definition that ck ≤ k, as for α = 1/k we immediately
k−1
have 1 − e−x
< x. In fact ck grows much mor slowly with k, as shown in Table 1,
which gives numerical values for these thresholds for 3 ≤ k ≤ 7. Here we are not
truly concerned with the exact values ck ; it is enough that only linear space is required.
It is worthwhile noting that ck is generally close to 1, while to obtain successful GET
operations we require a number of cells which is a significant constant factor larger
than n. Therefore, in practice the choice of the size of the IBLT will generally be
determined by the desired probability for a successful GET operation, not the need for
listing. (For applications where GET operations are unimportant and listing is the key
feature, further improvements can be had by using irregular IBLTs.)
k
ck
3
1.222
4
1.295
5
1.425
6
1.570
7
1.721
Table 1: Thresholds for the 2-core rounded to four decimal places.
When we design our IBLT, depending on the application, we may want a target
probability for succeeding in listing entries. Specifically, we may desire failure to
occur with probability O(t−c ) for a chosen constant c (whenever n ≤ t). By choosing
k sufficiently large and m above the 2-core threshold, we can ensure this; indeed,
standard results give that the bottleneck is the possibility of having two edges with the
same collection of vertices, giving a failure probability of O(t−k+2 ). The following
theorem follows from previous work but we provide it for completeness.
Theorem 1: As long as m is chosen so that m > (ck + )t for some > 0, LIST E N TRIES fails with probability O(t−k+2 ) whenever n ≤ t.
5 Following this analogy, one could for example, consider irregular versions of the IBLT, where different
keys utilize a different number of hash values; such a variation could use less space while allowing LIST E N TRIES to succeed, or could be used to allow some keys with more hash locations to obtain a better likelihood
of a successful lookup. These variations are straightforward and we do not consider the details further here.
10
Proof: We describe the result in terms of the 2-core. In what follows we assume
n ≤ t. The probability that j hyperedges form a non-empty 2-core is dominated by the
probability that these edges utilize only jk/2 vertices. This probability is at most
jk
n
m
jk
j
jk/2
2m
≤
n j 2m jk/2 jk jk
e
nj
mjk/2
=
jk
2m
jk/2 j
jke
e
.
m
j
For k constant, m > (ck + )n, and j ≤ γn for some constant γ, the sum of these
probabilities is dominated by the term where j = 2, which corresponds to a failure
probability of O(n−k+2 ). To deal separately with the case of j > γn, we note that
standard analysis of the peeling process shows that, as long as m is above the decoding
threshold, the probability that the peeling process fails before reaching a core of size
δn for any constant δ is asymptotically exponentially small in n. (See, e.g., [14].) By
this argument the case of j > γn adds a vanishing amount to the failure probability,
completing the proof of the theorem.
3
Adding Fault Tolerance to an Invertible Bloom Lookup
Table
For cases where there can be deletions for key-value pairs that are not already in B, or
values can be inserted for keys that are already in B, we require some fault tolerance.
We can utilize a standard approach of adding random checksums to get better fault
tolerance.
Extraneous Deletions Let us first consider a case with extraneous deletions only.
Specifically, we assume a key-value pair might be deleted without a corresponding
insertion; however, in this first setting we still assume each key is associated with a
single value, and is not inserted or deleted multiple times at any instant. This causes
a variety of problems for both the GET and LIST E NTRIES routines. For example, it is
possible for a cell to have an associated count of 1 even if more than one key has hashed
to it, if there are corresponding extraneous deletions; this causes us to re-evaluate our
LIST E NTRIES routine.
To help deal with these issues, we add to our IBLT structure. We assume that each
key x has an additional hash value given by a hash function G1 (x), which in general
we assume will take on uniform random values in a range [1, R]. We then require each
cell has the following additional field:
• a hashkeySum field, which is the sum of the hash values, G1 (x), for all the
keys that have been mapped to this cell.
The hashkeySum field must be of sufficiently many bits and the hash function must be
sufficiently random to make collisions sufficiently unlikely; this is not hard to achieve
11
in practice. Our insertion and deletion operations must now change accordingly, in that
we now must add G1 (x) to each T [hi (x)].hashkeySum on an insertion and subtract
G1 (x) during a deletion. The pseudocode for these and the other operations is given in
Figure 3 at the end of this paper.
The hashkeySum field can serve as an extra check. For example, to check when a
cell has a count of 1 that it corresponds to a cell without extraneous deletions, we check
G1 (x) field against the hashkeySum field. For an error to occur, we must have that
a deletion has caused a count of 1 where the count should be higher, and the hashed
key values must align so that their sum causes a false check. This probability is clearly
at most 1/R (using the standard principle of deferred decisions, the “last hash” must
take on the precise wrong value for a false check). We will generally assume that R
is chosen large enough that we can assume a false match does not occur throughout
the lifetime of the data structure, noting that only O(log n) bits are needed to handle
lifetimes that are polynomial in n. Notice that even if sum fields overflow, as long as
they overflow gracefully, the probability of a false check is still 1/R.
Let us now consider GET operations. The natural approach is to assume that the
hashkeySum field will not lead to a false check, as above. In this case, on a GET
of a key x, if the count field is 0, and the keySum and hashkeySum are also 0, one
should assume that the cell is in fact empty, and return null. Similarly, if the count field
is 1, and the keySum and hashkeySum match x and G1 (x), respectively, then one
should assume the cell has the right key, and return its value. In fact, if the count field
is −1, and after negating keySum and hashkeySum the values match x and G1 (x),
respectively, one should assume the cell has the right key, except that it has been deleted
instead of inserted! In our pseudocode we return the value, although one could also flag
it as an extraneous deletion as well. Note, however, that we can no longer return null if
the count field is 1 but the keySum field does not match x; in this case, there could be,
for example, an additional key inserted and an additional key extraneously deleted from
that cell, which would cause the field to not match even if x was hashed to that cell.
If we let n be the number of keys either inserted or extraneously deleted in the IBLT,
k
then this reduces the probability of returning null for a key not in B to 1 − e−kn/m .
That is, to return null we must have at least one cell with zero key-value pairs from
B hashing to it, which occurs (approximately) with the given probability (using our
Poisson approximation).
For the LIST E NTRIES operation, we again use the hashkeySum field to check
when a cell has a count of 1 that it corresponds to a cell without extraneous deletions.
Note here that an error in this check will cause the entire listing operation to fail, so the
probability of a false check should be made quite low—certainly inverse polynomial
in n. Also note, again, that we can make progress in recovering keys with cells with
a count of −1 as well, if the cell contains only one extraneously deleted key and no
inserted keys. That is, if a cell contains a count of −1, we can negate the count,
keySum, and hashkeySum fields, check the hash value against the key to prevent a
false match, and if that check passes recover the key and remove it (in this case, add it
back in) to the other associated cells. Hence, a cell cannot yield a key during the listing
process only if more than one key, either inserted or deleted, has hashed to that cell.
This is now exactly the same setting as in the original case of no extraneous deletions,
12
and hence (assuming that no false checks occur!) the same analysis applies, with n
representing the number of keys either inserted or extraneously deleted. We give the
revised pseudo-code descriptions in Figure 3.
Multiple Values A more challenging case for fault tolerance occurs when a key can
be inserted multiple times with different values, or inserted and deleted with different
values. If a key is inserted multiple times with different values, not only can that key
not be recovered, but every cell associated with that key has been poisoned, in that it
will not be useful for listing keys, as it cannot have a count of 1 even as other values are
recovered. (A later deletion of a key-value pair could correct this problem, of course,
but the cell is poisoned at the time.) The same is true if a key is inserted and deleted
with different values, and here the problem is potentially even worse: if a single other
key hashes to that cell, the count may be 1 and the keySum and hashkeySum fields
will be correct even though the valueSum field will not match the other key’s value,
causing errors.
Correspondingly, we introduce an additional check for the sum of the values at a
cell, using a hash function G2 (y) for the values, and adding the following field:
• a hashvalueSum field, which is the sum of the hash values G2 (y) for all the
values that have been mapped to this cell.
One can then check that the hash of the keySum and valueSum take on the appropriate
values when the count field of a cell is 1 (or −1) in order to see if listing the key-value
pair is appropriate.
The question remains whether the poisoned cells will prevent recovery of key values. Here we modify the goal of LIST E NTRIES to return all key-value pairs for all valid
keys with high probability—that is, all keys with a single associated value at that time.
We first claim that if the invalid keys make up a constant fraction of the n keys that
this is not possible under our construction with linear space. A constant fraction of the
cells would then be poisoned, and with constant probability each valid key would then
hash solely to poisoned cells, in which case the key could not be recovered.
However, it is useful to consider these probabilities, as in practical settings these
quantities will determine the probability of failure. For example, suppose γn keys are
invalid for some constant γ. By our previous analysis,
the fraction of cells that are
poisoned is concentrated around 1 − e−kγn/m , and hence the probability that any
k
specific valid key has all of its cells poisoned is 1 − e−kγn/m . (While there are
other possible ways a key could not be recovered, for example if two keys have all
but one of their cells poisoned and their remaining cell is the same, this gives a good
first approximation for reasonable values, as other terms will generally be lower order
when these probabilities are small.) For example, in a configuration we use in our
experiments below, we choose k = 5, m/n = 8, and γ = 1/10; in this case, the
probability of a specific valid key being unrecoverable is approximately 8.16 · 10−7 ,
which may be quite suitable for practice.
For a more theoretical asymptotic analysis, suppose instead that there are only n1−β
invalid keys. Then if each key uses at least d1/βe + 1 hash functions, with high probability every valid key will have at least one hash location that does not coincide with
13
a invalid cell. This alone does not guarantee reconstructing the valid key-value pairs,
but we can extend this idea to show LIST E NTRIES can provably successfully obtain all
valid keys even with n1−β invalid keys; by using k = d1/βe + 4 hash functions, we
can guarantee with high probability that every valid key has at least 3 hash locations
without an invalid cell. (One can raise the probability to any inverse polynomial in n
by changing the constant 4 as desired.) Indeed, we can determine the induced distribution for the number of neighbors that are unpoisoned cells for the valid keys, but the
important fact is that the number of keys with k hashes to unpoisoned cells in n − o(n).
It follows from the standard analysis of random cores (e.g., see Molloy [25]) that the
same threshold as for the original setting with k hash functions applies. Hence the
number of cells needed will again be linear in n (with the coefficient dependent on β)
in order to guarantee successful listing of keys with high probability. This yields the
following theorem:
Theorem 2: Suppose there are n1−β invalid keys. Let k = d1/βe + 4. Then if m >
(ck + )n for some > 0, LIST E NTRIES succeeds with high probability.
While this asymptotic analysis provides some useful insights, namely that full recovery is practical, in practice we expect the analysis above based on setting γ so that
the number of invalid keys is γn will prove more useful for parameter setting and predicting performance.
Extensions to Duplicates Interestingly, using the same approach as for extraneous
deletions, our IBLT can handle the setting where the same key-value pair is inserted
multiple times. Essentially, this means the IBLT is robust to duplicates, or can also
be used to obtain a count for key-value pairs that are inserted multiple times. We
again use the additional hashkeySum and valueSum fields. When the count field is
j for the cell, we take the keySum, hashkeySum, and valueSum fields and divide
them by j to obtain the proposed key, value, and corresponding hash. (Here, note
we cannot use XORs in place of sums in our algorithms.) If the key hash matches,
we assume that we have found the right key and return the lookup value or list the
key-value pair accordingly, depending on whether a GET or LIST E NTRIES operation is
being performed. If it is possible to have the same key appear with multiple values, as
above, then we must also make use of the hashvalueSum fields, dividing it by j and
using it to check that the value is correct as well. For the listing operation, the IBLT
deletes j copies of the key-value pair from the other cells.6 The point here is that a
key that appears multiple times, just as a key that is deleted rather than inserted, can be
handled without significant change to the decoding approach.
The one potential issue with duplicate key-value pairs is in the case of word-value
overflow for the memory locations containing the sum; in case of overflow, it may be
that one does not detect that the key hash matches (and similarly for the hashvalueSum fields). In practice this may limit the number of duplicates that can be tolerated;
however, for small numbers of duplicates and suitably sized memory fields, overflow
6 Note that here we are making use of the assumption that the hash locations are distinct for a key;
otherwise, the count for the number of copies at this location might not match the number of copies of the
key in all the other locations.
14
will be a rare occurrence (that would require large numbers of keys to hash to the same
cell, a provably probabilistically unlikely event).
Fault Tolerance to Lost Memory Subblocks We offer one additional way this structure proves resilient to various possible faults. Suppose that the structure is indeed set
up with k different memory subblocks, one for each hash function. Conceivably, one
could lose an entire subblock, and still be able to recover all the keys in a listing with
high probability, with only a reduction in the success probability of a GET (as long as
k − 1 hashes with a smaller range space remains sufficient for listing). In some sense,
this is because the system is arguably overdesigned; obtaining high lookup success
probability when n is less than the threshold t requires a large number of empty cells,
and this space is far more than is needed for decoding.
An Example Application As an example of where these ideas might be used, we
return to our mirror site application. An IBLT B from Alice can be used by Bob to find
filename-checksum (key-value) pairs where his filename has a different checksum than
Alice’s. After deleting all his key-value pairs, he lists out the contents of B to find files
that he or Alice has that the other does not. The IBLT might not be empty at this point,
however, as the listing process might not have been able to complete due to poisoned
cells, where deletions were done for keys with values different than Alice’s values. To
discover these, Bob can re-insert each of his key-value pairs, in turn, to find any that
may unpoison a cell in B (where he immediately deletes ones that don’t lead to a new
unpoisoned cell). If a new unpoisoned cell is found found (using the G1 hash function
as a check), then Bob can then remove a key-value pair with the same key as his but
with a different value (that is, with Alice’s value). Note Bob may then also be able to
possibly perform more listings of keys that might have been previously unrecovered
because of the poisoned cells. Repeating this will discover with high probability all the
key-value pairs where Alice and Bob differ.
4
Space-Saving for an Invertible Bloom Lookup Table
Up to this point, we have not been concerned with minimizing space for our IBLT
structure, noting that it can be done in linear space. Nevertheless, there are a variety
of techniques available for reducing the space required, generally at the expense of
additional computation and shuffling of memory, while still keeping constant amortized
or worst-case time bounds on various operations. Whether such efforts are worthwhile
depends on the setting; in some applications space may be the overriding issue, while
in others speed or even simplicity might be more relevant. We briefly point to some of
the previous work that can offer further insights.
IBLTs, like other Bloom filter structures, often have a great deal of wasted space
corresponding to zero entries that can be compressed or fixed-length space required
for fields like the count field that can be made variable-length depending on the value.
This wasted space can be compressed away using techniques for keeping compressed
forms of arrays, including those for storing arrays of variable-length strings. Such
mechanisms are explored in for example [4, 28, 34] and can be applied here.
15
A simpler, standard approach (saving less space) is to use quotienting, whereby
the hash value for a key determines a bucket and an additional quotient value that can
be stored. Quotienting can naturally be used with the IBLT to reduce the space used
for storing for example the keySum or the hashkeySum values. Also, as previously
mentioned, in settings without multiple copies of the same key-value pair, we can use
XORs in place of sums to save space.
Finally, we recall that the space requirements arise because of the desire for high
accuracy for GET operations, not because of the LIST E NTRIES operation. If one is
willing to give up lookup accuracy—which may be the case if, for example, one expects
the system to be overloaded much of the time—then less space is needed to maintain
successful listing.
5
Simulations and Experiments
We have run a number of simulations to test the IBLT structure and our analysis. In
these experiments we have not focused on running time; a practical implementation
could require significant optimization. Also, we have not concerned ourselves with
issues of word-value overflow. Because of this, there is no need to simulate the data
structure becoming overloaded and then deleting key-value pairs, as the state after deletions would be determined entirely by the key-value pairs in the system. Instead, we
focus on the success probability of the listing of keys and, to a lesser extent, on the
success probability for a GET operation. Overall, we have found that the IBLT works
quite effectively and the performance matches our theoretical analysis. We provide a
few example results. In all of the experiments here, we have chosen to use five hash
functions.
First, our calculated asymptotic thresholds for decoding from Table 1 are quite accurate even for reasonably small values. For example, in the setting where there are no
duplicate keys or extraneous deletions of keys, we repeatedly performed 20,000 simulations with 10,000 keys, and varied the number of cells. Table 1 suggests an asymptotic
threshold for listing all entries near 14,250. As shown in Figure 1(a), around this point
we see a dramatic increase in the average number of key-value pairs recovered when
performing our LIST E NTRIES operation. In fact, at 14,500 cells only two trials failed to
recover all key-value pairs, and with 14,600 cells or more all trials successfully recover
all key-value pairs. We performed an additional 200,000 trials with 14,600 cells, and
again all trials succeeded. In Figure 1(b), we consider 20,000 simulations with 100,000
keys, where the corresponding threshold should be near 142,500. With more keys, we
expect tighter concentration around the threshold, and indeed with 144,000 cells or
more all trials successfully recover all key-value pairs. We performed an additional
200,000 trials with with 144,000 cells, and again all trials succeeded.
We acknowledge that more simulations would be required to obtain detailed bounds
on the probability of failure to recover all key-value pairs for specific values of the
number of key-value pairs and cells. This is equivalent to the well-studied problem
of “finite-length analysis” for related families of error-correcting codes. Dynamic programming techniques, as discussed in [21] and subsequent follow-on work, can be
applied to obtain such bounds.
16
Our next tests of the IBLT allow duplicate keys with the same value and extraneous deletions, but without keys with multiple values. Our analysis suggests this should
work exactly as with no duplicate or extraneous deletions, and our simulations verify
this. In these simulations, we had each key result in a duplicate with probability 1/5,
and each key result in a deletion in place of an insertion with probability 1/5. Using
a check on key and value fields, in 20,000 simulations with 10,000 keys, 80,000 cells,
and 5 hash functions, a complete listing was obtained every time, and GET operations
were successful on average for 97.83 percent of the keys, matching the standard analysis for a Bloom filter. Results were similar with 20,000 runs with 100,000 keys and
800,000 cells, again with complete recovery each time and GET operations successful
on average for 97.83 percent of the keys.
Finally, we tested the IBLT with keys that erroneously obtain multiple values. As
expected, these keys can prevent recovery of other key-value pairs during listing, but
do not impact the success probability of GET operations for other keys. For example,
again using a check on key and value fields, in 20,000 simulations with 10,000 keys of
which 500 had multiple values, 80,000 cells, and 5 hash functions, the 9500 remaining
key-value pairs were recovered 19,996 times; the remaining 4 times all but one of the
9500 key-value pairs was recovered. With 1000 keys with multiple values, the remaining key-value pairs were recovered 19,872 times, and again the remaining 128 times
all but one of the 9000 key-value pairs was recovered. The average success rate for
GET operations remained 97.83 percent on the valid keys in both cases, as would be
expected. We note that with 10,000 keys with 1,000 with multiple values, our previous
back-of-the-envelope calculation showed that each valid key would fail with probability roughly 8.16·10−7 ; hence, with 9,000 other keys, assuming independence we would
estimate the probability of complete recovery at approximately 0.9927, closely matching our experimental results. More detailed results are give in Figures 2(a) and 2(b),
where we vary the number keys with multiple values for two settings: 10,000 keys and
80,000 cells, and 100,000 keys and 800,000 cells. The results are based on 20,000
trials. As can be seen complete recovery is possible with large numbers of multiplevalued keys in both cases, but naturally the probability of complete recovery becomes
worse with larger numbers of keys even if the percentage of invalid keys is the same.
We emphasize that even when complete recovery does not occur in this setting,
generally almost all keys with a single value can be recovered. For example, in Table 2
we consider three experiments. The first is for 10,000 keys, 80,000 cells, and 1,000
keys with duplicate values. The second is the same but with 2,000 keys with duplicate
values. The third is for 100,000 keys, 800,000 cells, and 10,000 keys with duplicate
values. Over all 20,000 trials for each experiment, in no case were more than 3 valid
keys unrecovered. The main point of Table 2 is that with suitable design parameters,
even when complete recovery is not possible because of invalid keys, the degradation is
minor. We suspect this level of robustness may be useful for some applications where
almost-complete recovery is acceptable.
17
Experiment 1
Experiment 2
Experiment 3
0
99.360
83.505
92.800
Unrecovered Keys
1
2
3
0.640 0.000 0.000
14.885 1.520 0.090
6.915 0.265 0.020
Table 2: Percentage of trials where 1, 2, and 3 keys are left unrecovered.
6
Conclusion and Future Work
We have given an extension to the Bloom filter data structure to key-value pairs and
the ability to list out its contents. This structure is deceptively simple, but is able
to achieve functionalities and efficiencies that appear to be unique in many respects,
based on our analysis derived from recent results on 2-cores in hypergraphs. One
possible direction for future work includes whether one can easily include methods for
allowing for multiple values as a natural condition instead of an error.
Acknowledgments
Michael Goodrich was supported in part by the National Science Foundation under
grants 0724806, 0713046, 0847968, and 0953071. Michael Mitzenmacher was supported in part by the National Science Foundation under grants IIS-0964473, CCF0915922 and CNS-0721491, and in part by grants from Yahoo! Research, Cisco, Inc.,
and Google.
References
[1] N. Alon and A. Nussboim. k-wise independent random graphs. In Proceedings
of the 49th Annual IEEE Symposium on Foundations of Computer Science,
pages 813–822, 2008.
[2] M. Bawa, R. J. Bayardo, Jr, R. Agrawal, and J. Vaidya. Privacy-preserving
indexing of documents on the network. The VLDB Journal, 18(4):837–856,
2009.
[3] S. M. Bellovin and W. R. Cheswick. Privacy-enhanced searches using encrypted
Bloom filters. Technical Report CUCS-034-07, Columbia University, Dept. of
Computer Science, 2007.
[4] D. Blandford and G. Blelloch. Compact dictionaries for variable-length keys and
data with applications. ACM Transactions on Algorithms (TALG), 4(2):1–25,
2008.
[5] B. Bloom. Space/time trade-offs in hash coding with allowable errors.
Communications of the ACM, 13(7):422–426, 1970.
18
[6] F. Bonomi, M. Mitzenmacher, R. Panigrahy, S. Singh, and G. Varghese. Beyond
Bloom filters: From approximate membership checks to approximate state
machines. ACM SIGCOMM Computer Communication Review, 36(4):326,
2006.
[7] F. Bonomi, M. Mitzenmacher, R. Panigrahy, S. Singh, and G. Varghese. An
improved construction for counting Bloom filters. In Proceedings of the
European Symposium on Algorithms (ESA), volume 4168 of LNCS, pages
684–695, 2006.
[8] A. Broder and M. Mitzenmacher. Network applications of Bloom filters: A
survey. Internet Mathematics, 1(4):485–509, 2004.
[9] B. Chazelle, J. Kilian, R. Rubinfeld, and A. Tal. The Bloomier filter: an efficient
data structure for static support lookup tables. In Proceedings of the Fifteenth
Annual ACM-SIAM Symposium on Discrete Algorithms, pages 30–39, 2004.
[10] S. Chen, R. Wang, X. Wang, and K. Zhang. Side-channel leaks in web
applications: a reality today, a challenge tomorrow. In Proceedings of the 31st
IEEE Symposium on Security and Privacy, 2010.
[11] B. Chor, E. Kushilevitz, O. Goldreich, and M. Sudan. Private information
retrieval. J. ACM, 45:965–981, November 1998.
[12] S. Cohen and Y. Matias. Spectral Bloom filters. In Proceedings of the 2003
ACM SIGMOD International Conference on Management of Data, pages
241–252, 2003.
[13] G. Cormode and S. Muthukrishnan. An improved data stream summary: the
count-min sketch and its applications. J. Algorithms, 55:58–75, April 2005.
[14] R. Darling and J. Norris. Differential equation approximations for Markov
chains. Probability Surveys, 5:37–79, 2008.
[15] L. Devroye and P. Morin. Cuckoo hashing: further analysis. Inf. Process. Lett.,
86(4):215–219, 2003.
[16] M. Dietzfelbinger, A. Goerdt, M. Mitzenmacher, A. Montanari, R. Pagh, and
M. Rink. Tight thresholds for cuckoo hashing via XORSAT. In Proceedings of
ICALP, pages 213–225, 2010.
[17] D. Eppstein and M. T. Goodrich. Straggler identification in round-trip data
streams via Newton’s identities and invertible Bloom filters. IEEE Transactions
on Knowledge and Data Engineering, to appear.
[18] E.-J. Goh. Secure indexes. Cryptology ePrint Archive, Report 2003/216, 2003.
http://eprint.iacr.org/2003/216/.
[19] O. Goldreich and R. Ostrovsky. Software protection and simulation on oblivious
RAMs. J. ACM, 43(3):431–473, 1996.
19
[20] M. T. Goodrich and M. Mitzenmacher. MapReduce parallel cuckoo hashing and
oblivious RAM simulations. CoRR, abs/1007.1259, 2010.
[21] R. Karp, M. Luby, and A. Shokrollahi. Finite length analysis of LT codes. In
Proceedings of the International Symposium on Information Theory, page 39,
2004.
[22] Y. Lu, A. Montanari, B. Prabhakar, S. Dharmapurikar, and A. Kabbani. Counter
braids: a novel counter architecture for per-flow measurement. In Proceedings of
the 2008 ACM SIGMETRICS International Conference on Measurement and
Modeling of Computer Systems, pages 121–132, 2008.
[23] M. Luby, M. Mitzenmacher, M. Shokrollahi, and D. Spielman. Efficient erasure
correcting codes. IEEE Transactions on Information Theory, 47(2):569–584,
2001.
[24] M. Mitzenmacher and S. Vadhan. Why simple hash functions work: exploiting
the entropy in a data stream. In Proceedings of the Nineteenth Annual
ACM-SIAM Symposium on Discrete Algorithms, pages 746–755, 2008.
[25] M. Molloy. The pure literal rule threshold and cores in random hypergraphs. In
Proceedings of the fifteenth annual ACM-SIAM symposium on Discrete
algorithms, pages 672–681. Society for Industrial and Applied Mathematics,
2004.
[26] M. Naor, G. Segev, and U. Wieder. History-independent cuckoo hashing. In
Proceedings of ICALP, pages 631–642, 2008.
[27] R. Nojima and Y. Kadobayashi. Cryptographically secure Bloom-filters. Trans.
Data Privacy, 2(2):131–139, 2009.
[28] A. Pagh, R. Pagh, and S. Rao. An optimal Bloom filter replacement. In
Proceedings of the Sixteenth Annual ACM-SIAM Symposium on Discrete
Algorithms, page 829, 2005.
[29] R. Pagh and F. F. Rodler. Cuckoo hashing. J. Algorithms, 51(2):122–144, 2004.
[30] J. J. Parekh, K. Wang, and S. J. Stolfo. Privacy-preserving payload-based
correlation for accurate malicious traffic detection. In LSAD ’06: Proceedings of
the 2006 SIGCOMM Workshop on Large-scale Attack Defense, pages 99–106,
2006.
[31] B. Pinkas and T. Reinman. Oblivious RAM revisited. In T. Rabin, editor,
Advances in Cryptology (CRYPTO), volume 6223 of Lecture Notes in Computer
Science, pages 502–519. Springer, 2010.
[32] E. Price. Efficient sketches for the set query problem. CoRR, abs/1007.1253,
2010.
[33] F. Putze, P. Sanders, and J. Singler. Cache-, hash-, and space-efficient Bloom
filters. J. Exp. Algorithmics, 14:4.4–4.18, 2009.
20
[34] R. Raman and S. Rao. Succinct dynamic dictionaries and trees. In Proceeding of
ICALP, pages 357–368, 2003.
[35] H. Song, S. Dharmapurikar, J. Turner, and J. Lockwood. Fast hash table lookup
using extended Bloom filter: An aid to network processing. In Proceedings of
SIGCOMM, pages 181–192, 2005.
[36] M. Stonebraker. In search of database consistency. Commun. ACM, 53:8–9,
October 2010.
[37] S. Yekhanin. Private information retrieval. Commun. ACM, 53:68–73, April
2010.
21
Recovery for 10000 Keys
Avg. Percentage Recovered
100
80
60
40
20
0
13800
14000
14200
14400
Size of Table
14600
14800
146000
148000
(a) 10,000 keys
Recovery for 100000 Keys
Avg. Percentage Recovered
100
80
60
40
20
0
138000
140000
142000
144000
Size of Table
(b) 100,000 keys
Figure 1: Percentage of key-value pairs recovered around the threshold. Slightly over
the theoretical asymptotic threshold, we obtain full recovery of all key-value pairs with
LIST E NTRIES on all simulations. Each data point represents the average of 20,000
simulations.
22
Recovery with Damaged Keys
Percentage Incomplete Recovery
1
0.8
0.6
0.4
0.2
0
0
200
400
600
Damaged Keys
800
1000
8000
10000
(a) 10,000 keys
Recovery with Damaged Keys
Percentage Incomplete Recovery
10
8
6
4
2
0
0
2000
4000
6000
Damaged Keys
(b) 100,000 keys
Figure 2: Percentage of trials with incomplete recovery with “damaged” keys that have
multiple values. Each data point represents the average of 20,000 simulations.
23
• INSERT(x, y):
for each hi (x) value, for i = 1, . . . , k do
add 1 to T [hi (x)].count
add x to T [hi (x)].keySum
add y to T [hi (x)].valueSum
add G1 (x) to T [hi (x)].hashkeySum
end for
• DELETE(x, y):
for each hi (x) value, for i = 1, . . . , k do
subtract 1 from T [hi (x)].count
subtract x from T [hi (x)].keySum
subtract y from T [hi (x)].valueSum
subtract G1 (x) to T [hi (x)].hashkeySum
end for
• GET(x):
for each hi (x) value, for i = 1, . . . , k do
if T [hi (x)].count = 0 and T [hi (x)].keySum = 0 and
T [hi (x)].hashkeySum = 0 then
return null
else if T [hi (x)].count = 1 and T [hi (x)].keySum = x and
T [hi (x)].hashkeySum = G1 (x) then
return T [hi (x)].valueSum
else if T [hi (x)].count = −1 and T [hi (x)].keySum = −x and
T [hi (x)].hashkeySum = −G1 (x) then
return −T [hi (x)].valueSum
end if
end for
return “not found”
• LIST E NTRIES():
while there is an i ∈ [1, m] such that T [i].count = 1 or T [i].count = −1
do
if T [hi (x)].count
=
1 and T [hi (x)].hashkeySum
=
G1 (T [hi (x)].keySum) then
add the pair, (T [i].keySum , T [i].valueSum), to the output list
call DELETE(T [i].keySum, T [i].valueSum)
else if T [hi (x)].count = −1 and −T [hi (x)].hashkeySum =
G1 (−T [hi (x)].keySum) then
add the pair, (−T [i].keySum , −T [i].valueSum), to the output list
call INSERT(−T [i].keySum, −T [i].valueSum)
end if
end while
Figure 3: Revised pseudo-code for tolerating extraneous deletions.
24
| 8cs.DS
|
Stochastic Gradient Descent in Continuous Time: A Central Limit
Theorem
Justin Sirignano∗ and Konstantinos Spiliopoulos†‡
arXiv:1710.04273v2 [math.PR] 2 Nov 2017
November 3, 2017
Abstract
Stochastic gradient descent in continuous time (SGDCT) provides a computationally efficient method
for the statistical learning of continuous-time models, which are widely used in science, engineering, and
finance. The SGDCT algorithm follows a (noisy) descent direction along a continuous stream of data.
The parameter updates occur in continuous time and satisfy a stochastic differential equation. This paper
analyzes the asymptotic convergence rate of the SGDCT algorithm by proving a central limit theorem
for strongly convex objective functions and, under slightly stronger conditions, for non-convex objective
functions as well. An Lp convergence rate is also proven for the algorithm in the strongly convex case.
1
Introduction
“Stochastic gradient descent in continuous time” (SGDCT) is a statistical learning algorithm for continuoustime models, which are common in science, engineering, and finance. SGDCT is the continuous-time analog of
the well-known stochastic gradient descent. Given a continuous stream of data, stochastic gradient descent in
continuous time (SGDCT) can estimate unknown parameters or functions in stochastic differential equation
(SDE) models. [3] analyzes the numerical performance of SGDCT for a number of applications in finance
and engineering.
Batch optimization for the statistical estimation of continuous-time models may be impractical for large
datasets where observations occur over a long period of time. Batch optimization takes a sequence of descent
steps for the model error for the entire observed data path. Since each descent step is for the model error
for the entire observed data path, batch optimization is slow (sometimes impractically slow) for long periods
of time or models which are computationally costly to evaluate (e.g., partial differential equations).
SGDCT provides a computationally efficient method for statistical learning over long time periods and for
complex models. SGDCT continuously follows a (noisy) descent direction along the path of the observation;
this results in much more rapid convergence. Parameters are updated online in continuous time, with the
parameter updates θt satisfying a stochastic differential equation.
Consider a diffusion Xt ∈ X = Rm :
dXt = f ∗ (Xt )dt + σdWt .
(1.1)
The goal is to statistically estimate a model f (x, θ) for f ∗ (x) where θ ∈ Rk . The function f ∗ (x) is unknown.
Wt ∈ Rm is a standard Brownian motion. The diffusion term Wt represents any random behavior of the
system or environment. The functions f (x, θ) and f ∗ (x) may be non-convex.
The stochastic gradient descent update in continuous time follows the SDE:
(1.2)
dθt = αt ∇θ f (Xt , θt )(σσ ⊤ )−1 dXt − ∇θ f (Xt , θt )(σσ ⊤ )−1 f (Xt , θt )dt ,
∗ Department of Industrial and Enterprise Systems Engineering, University of Illinois at Urbana-Champaign, Urbana, E-mail:
[email protected]
† Department of Mathematics and Statistics, Boston University, Boston, E-mail: [email protected]
‡ Research of K.S. supported in part by the National Science Foundation (DMS 1550918)
1
α
. The
where ∇θ f (Xt ; θt ) is matrix-valued and αt is the learning rate. For example, αt could equal CC0 +t
parameter update (1.2) can be used for both statistical estimation given previously observed data as well as
online learning (i.e., statistical estimation in real-time as data becomes available).
We assume that Xt is sufficiently ergodic (to be concretely specified later in the paper) and that it has
some well-behaved π(dx) as its unique invariant measure. As a general notation, if h(x, θ) is a generic L1 (π)
function, then we define its average over π(dx) to be
Z
h(x, θ)π(dx).
h̄(θ) =
X
Let us set
g(x, θ) =
E
−1
1
1D
2
f (x, θ) − f ∗ (x), σσ ⊤
(f (x, θ) − f ∗ (x)) .
kf (x, θ) − f ∗ (x)kσσ⊤ =
2
2
The gradient ∇θ g(Xt , θ) cannot be evaluated since f ∗ (x) is unknown. However, dXt = f ∗ (Xt )dt + σdWt
is a noisy estimate of f ∗ (x)dt, which leads to the algorithm (1.2). SGDCT follows a noisy descent direction
along a continuous stream of data produced by Xt .
ḡ is a natural objective function for the estimation of the continuous-time
R dynamics. Heuristically, it
is expected that θt will tend towards the minimum of the function ḡ(θ) = X g(x, θ)π(dx). The data Xt
will be correlated over time, which complicates the mathematical analysis. This differs from the standard
discrete-time version of stochastic gradient descent where the the data is usually considered to be i.i.d. at
every step.
a.s.
The authors previously proved in [3] that k∇ḡ(θt )k → 0. In this paper, we prove several new results
regarding the convergence rate of the algorithm θt .
The first main result of the paper is an Lp convergence rate and a central limit theorem in the case where
p
K
ḡ(θ) is strongly convex. Let θ∗ be the global minimum of ḡ(θ). We prove that E[kθt − θ∗ k ] ≤ (C +t)
p/2 for
0
√
d
α
. The Lp
p ≥ 1 and t(θt − θ∗ ) → N (0, Σ̄) when ḡ(θ) is strongly convex and the learning rate αt is CC0 +t
convergence rate is proven in Theorem 2.7 and the central limit theorem is proven in Theorem 2.8. We prove
these results for models f (x, θ) with up to quadratic growth in θ and polynomial growth in x. Theorems 2.7
and 2.8 do not make use of the results in [3].
The second main result is a central limit theorem when the objective function ḡ(θ) is non-convex. We
√
d
prove that t(θt − θ∗ ) → N (0, Σ̄) when ḡ(θ) is non-convex with a single critical point; see Theorem 2.12.
We prove this result for models f (x, θ) with up to linear growth in θ and polynomial growth in x. As a part
of the proof, we also strengthen the convergence result of [3], which did not allow f (x, θ) to grow in θ; see
Theorem 2.10.
Analogous results for Theorems 2.7, 2.8, and 2.12 hold of course for a general class of learning rates
αt ; see Proposition 2.13. The precise statement of the mathematical results and the technical assumptions
required are presented in Section 2.
These mathematical results are important for two reasons. First, they establish theoretical guarantees
for the rate of convergence of the algorithm. Secondly, they can be used to analyze the effects of different
features such as the learning rate αt , the level of noise σ, and the shape of the objective function ḡ(θ).
We are able to precisely characterize the regime under which the optimal convergence rate exists as well as
characterize the limiting covariance Σ̄. The regime depends entirely upon the choice of the learning
rate.
Rt
h(X
The
proofs
in
this
paper
require
addressing
several
challenges.
First,
fluctuations
of
the
form
α
s , θs )−
s
0
h̄(θs ) ds must be analyzed. We evaluate, and control with rate α2t , these fluctuations using a Poisson partial
differential equation. Second, the model f (x, θ) is allowed to grow with θ. This means that the fluctuations
as well as other terms can grow with θ. Therefore, we must prove an a priori stability estimate for kθt k.
Proving a central limit theorem for the non-convex ḡ(θ) in Theorem 2.12 is challenging since the convergence
speed of θt can become arbitrarily slow in certain regions, and the gradient can even point away from the
global minimum θ∗ . We prove the central limit theorem for the non-convex case by analyzing two regimes
[0, τδ ] and [τδ , ∞), where τδ is defined such that kθt − θ∗ k < δ for all t ≥ τδ . The proof also requires the
analysis of stochastic integrals with anticipative integrands, which is challenging since standard approaches
(such as the Itô Isometry) cannot be directly applied.
2
1.1
Literature Review
The vast majority of the statistical learning, machine learning, and stochastic gradient descent literature
addresses discrete-time algorithms. In contrast, this paper analyzes a statistical learning algorithm in continuous time. Below we review the existing literature that is most relevant to our work. We also comment
on the importance of developing and analyzing continuous-time algorithms for addressing continuous-time
models.
Many discrete-time papers study algorithms for θn without the X-dynamics (for example, stochastic
gradient descent with i.i.d. noise at each step). The inclusion of the X-dynamics makes analysis significantly
more challenging. An L2 convergence rate and central limit theorem result for discrete-time stochastic
gradient descent is presented in [8]. Our setup and assumptions are different from [8]. Our proof leverages
the continuous-time nature of our setting, which is the formulation of interest in many engineering and
financial problems (see [3]).
[4] studies continuous-time stochastic mirror descent in a setting different than ours. In the framework
of [4], the objective function is known. In this paper, we consider the statistical estimation of the unknown
dynamics of a random process (i.e. the X process satisfying (1.1)).
Statisticians and financial engineers have actively studied parameter estimation of SDEs, although typically not with statistical learning or machine learning approaches. The likelihood function will usually be
calculated from the entire observed path of X (i.e., batch optimization) and then maximized to find the
maximum likelihood estimator (MLE). Unlike in this paper, the actual optimization procedure to maximize
the likelihood function is often not analyzed.
Some relevant publications in the financial statistics literature include [1], [2], [7], and [10]. [7] derives
the likelihood function for continuously observed X. The MLE can be calculated via batch optimization.
[1] and [2] consider the case where X is discretely observed and calculate MLEs via a batch optimization
approach. [10] estimates parameters by a Bayesian approach. Readers are referred to [9, 15, 11] for thorough
reviews of classical statistical inference methods for stochastic differential equations.
Continuous-time models are common in engineering and finance. There are often coefficients or functions
in these models which are uncertain or unknown; stochastic gradient descent can be used to learn these
model parameters from data.
It is natural to ask why use SGDCT versus a straightforward approach which (1) discretizes the continuoustime dynamics and then (2) applies traditional stochastic gradient descent. We elaborated in detail in regards
to this issue in [3], where specific examples are provided there to showcase the differences. For completeness,
let us briefly discuss the issues that arise.
SGDCT allows for the application of numerical schemes of choice to the theoretically correct statistical
learning equation for continuous-time models. This can lead to more accurate and more computationally efficient parameter updates. Numerical schemes are always applied to continuous-time dynamics and different
numerical schemes may have different properties for different continuous-time models. A priori performing
a discretization to the system dynamics and then applying a traditional discrete-time stochastic gradient
descent scheme can result in a loss of accuracy, or may not even converge, see [3]. For example, there is no
guarantee that (1) using a higher-order accurate scheme to discretize the system dynamics and then (2) applying traditional stochastic gradient descent will produce a statistical learning scheme which is higher-order
accurate in time. Hence, it makes sense to first develop the continuous-time statistical learning equation,
and then apply the higher-order accurate numerical scheme.
In addition to model estimation, SGDCT can be used to solve continuous-time optimization problems,
such as American options. In [3], SGDCT was combined with a deep neural network to solve American
options in up to 100 dimensions. Alternatively, one could discretize the dynamics and then use the Qlearning algorithm (traditional stochastic gradient descent applied to an approximation of the discrete HJB
equation). However, as we showed in [3], Q-learning is biased while SGDCT is unbiased. Furthermore, in
SDE models with Brownian motions, the Q-learning algorithm can blow up as the time step size ∆ becomes
small; see [3] for details.
1.2
Organization of Paper
In Section 2 we state our assumptions and the main results of this paper. The proof of the Lp convergence
rate for p ≥ 1 is in Section 3. The proof of the CLT in the strongly convex case is in Section 4. Section
3
5 proves the central limit theorem for a class of non-convex models. In Section 6, the convergence rate
results are used to analyze the behavior and dynamics of the stochastic gradient descent in continuous-time
algorithm. Some technical results required for the proofs are presented in Appendix A. Appendix B contains
the proof of Theorem 2.10, which strengthens the convergence result of [3]. In particular, Appendix B
provides the necessary adjustments to the proofs of [3] in order to guarantee convergence in the case where
the model f (x, θ) is allowed to grow with respect to θ.
2
Main Results
We prove three main results. Theorem 2.7 is an Lp convergence for the strongly convex case. Theorem
2.8 is a central limit theorem for the strongly convex case. Theorem 2.12 is a central limit theorem for the
non-convex case with a single critical point.
We say that “the function h(θ) is strongly convex with constant C” if there exists a C > 0 such that
z ⊤ ∆θ h(θ)z ≥ Cz ⊤ z for any non-zero z ∈ Rk . Conditions 2.6 and 2.11 require that CCα > 1 where Cα is
the magnitude of the learning rate and C is the strong convexity constant for the objective function ḡ(θ∗ )
at point θ = θ∗ . This is an important conclusion of the convergence analysis in this paper: the learning rate
needs to be sufficiently large in order to achieve the optimal rate of convergence. This dependence of the
convergence rate on the learning rate is not specific to the algorithm SGDCT in this paper, but applies to
other algorithms (even deterministic gradient descent algorithms). We discuss this in more detail in Section
6.
It should be emphasized that the assumptions for Theorem 2.7 and Theorem 2.8 (the strongly convex
case) allow for the model f (x, θ) to grow up to quadratically in θ. On the other hand, Theorem 2.12 (the
non-convex case) is proven under the assumption that f (x, θ) grows up to linearly in θ. The growth of the
model f (x, θ) is allowed to be polynomial in x for Theorems 2.7, 2.8, and 2.12.
The proofs of Theorems 2.7, 2.8, and 2.12 are in Sections 3, 4, and 5, respectively. Some technical
results required for these proofs are presented in Appendix A. Appendix B contains the changes necessary
to generalize the proof of convergence of [3] from bounded ḡ(θ) to ḡ(θ) that can grow up to quadratically in
θ; see Theorem 2.10 for the corresponding rigorous statement.
Let us now list our conditions. Condition 2.1 guarantees uniqueness and existence of an invariant measure
for the X process. Condition 2.1 and the second part of Condition 2.2 guarantee that equation (1.1) is wellposed.
Condition 2.1. We assume that σσ ⊤ is non-degenerate bounded diffusion matrix and lim|x|→∞ f ∗ (x)·x = −∞
In regards to regularity of the involved functions, we impose the following Condition 2.2.
2
∇θ g
∈ C X , Rk , ∇θ g(·, θ) ∈
Condition 2.2.
1. We assume that ∇θ g(x, ·) ∈ C 2 (Rk ) for all x ∈ X , ∂ ∂x
2
C α (X ) uniformly in θ ∈ Rk for some α ∈ (0, 1).
2. The function f ∗ (x) is C 2+α (X ) with α ∈ (0, 1). Namely, it has two derivatives in x, with all partial
derivatives being Hölder continuous, with exponent α, with respect to x.
3. For every t ≥ 0, equation (1.2) is well-posed.
Conditions 2.3 and 2.4 are imposed in order to guarantee that θt has bounded moments uniformly in
time t.
Condition 2.3. There exists a constant R < ∞ and an almost everywhere positive function κ(x) such that
h−∇θ g(x, θ), θ/ kθki ≤ −κ(x) kθk , for kθk ≥ R.
Condition 2.4. Consider the function
1/2
θ
θ
⊤
.
,
τ (x, θ) = ∇θ f (x, θ)∇θ f (x, θ)
kθk kθk
(2.1)
Then, there exists a function λ(x), growing not faster than polynomially in kxk, such that for any θ1 , θ2 ∈ Rk ,
t > 0 and x ∈ Rm
|τ (x, θ1 ) − τ (x, θ2 )| ≤ |λ(x)|ρ (kθ1 − θ2 k)
4
(2.2)
R
where ρ(ξ) is an increasing function on [0, ∞) with ρ(0) = 0 and ξ>0 ρ−2 (ξ)dξ = ∞.
Notice that Condition 2.4 simplifies considerably when θ is one-dimensional and that it is always true in
the case where ∇θ f (x, θ) is independent of θ (which happens for example when f (x, θ) is affine in θ).
p
The Conditions 2.3 and 2.4 are necessary to prove that supt≥0 E[kθt k ] < K for p ≥ 2. This uniform
bound on moments, which we prove in Appendix A.1, is in turn required for the proofs of the Lp convergence
and central limit theorem in Theorems 2.7 and 2.8. The function g(x, θ) is allowed to grow, which means
that some a priori bounds on the moments of θt are necessary.
In regards to the learning rate we assume
α
Condition 2.5. The learning rate is αt = CC0 +t
where Cα > 0 is a constant.
Condition 2.5 makes the presentation of the results easier. However, as we shall see in Corollary 2.13 the
α
is not necessary as long as αt satisfies certain conditions. We chose to present the
specific form αt = CC0 +t
results under this specific form both for presentation purposes and because this is the usual form that the
learning rate takes in practice. However, we do present the result for a general learning rate in Corollary
2.13.
Condition 2.6.
1. ∇iθ f (x, θ) ≤ K 1 + kxkq + kθk(2−i)∨0 , for i = 0, 1, 2, 3 and for some finite constants
K, q < ∞.
2. ḡ(θ) is strongly convex with constant C.
3. CCα > 1.
4. ḡ(θ) ∈ C 3 .
5.
4−i
, for i = 0, 1, 2, 3 and for some finite constant K < ∞.
∇iθ ḡ(θ) ≤ K 1 + kθk
R t As will be seen from the proof of the Theorems 2.7 and 2.8, one needs to control terms of the form
0 αt (∇ḡ(θs ) − g(Xs , θs ))ds. Due to ergodicity of the X process one expects that such terms are small in
magnitude and go to zero as t → ∞. However, the speed at which they go to zero is what matters here. We
treat such terms by rewriting them equivalently using appropriate Poisson type partial differential equations
(PDE). Conditions 2.1, 2.2 and 2.6 guarantee that these Poisson equations have unique solutions that do
not grow faster than polynomially in the x and θ variables (see Theorem A.1 in Appendix A).
Theorem 2.7. Assume that Conditions 2.1, 2.2, 2.3, 2.4, 2.5 and 2.6 hold. Then, for p ≥ 1, there exists a
constant 0 < K < ∞ such that
K
p
.
E[kθt − θ∗ k ] ≤
(C0 + t)p/2
In order to state the central limit theorem results we need to introduce some notation. Let us denote by
v(x, θ) the solution to the Poisson equation (A.5) with H(x, θ) = ∇θ g(x, θ) − ∇θ ḡ(θ). Let us also set
⊤
Z
⊤ −1
⊤ −1
⊤
h(x, θ) = ∇θ f (x, θ)(σσ ) −∇x v(x, θ) σσ ∇θ f (x, θ)(σσ ) −∇x v(x, θ)
and h̄(θ) = h(x, θ)π(dx).
Theorem 2.8. Assume that Conditions 2.1, 2.2, 2.3, 2.4, 2.5 and 2.6 hold. Then,
√
d
t(θt − θ∗ ) → N (0, Σ̄),
where Σ̄ is defined to be
Σ̄ = Cα2
Z
∞
e−s(Cα ∆ḡ(θ
∗
)−I)
h̄(θ∗ )e−s(Cα (∆ḡ)
⊤
(θ ∗ )−I )
ds.
0
The main result of [3] states that if ḡ(θ) and its derivatives are bounded with respect to θ, then under assumptions of ergodicity for the X process one has convergence of the algorithm, in the sense that
limt→∞ k∇θ ḡ(θ)k = 0. In this paper, we allow growth of ḡ(θ) with respect to kθk. In particular, as we shall
state in Theorem 2.10 and prove in Appendix B, the results of [3] hold true without considerable extra work
if one allows up to linear growth for f (x, θ) with respect to θ, which translates into up to quadratic growth
for ḡ(θ) and up to linear growth of ∇ḡ(θ) with respect to θ. Let us formalize the required assumptions in
the form of Condition 2.9 below.
5
Condition 2.9.
K, q < ∞.
2.
1.
q
(1−i)∨0
, for i = 0, 1, 2 and for some finite constants
∇iθ f (x, θ) ≤ K 1 + kxk + kθk
2−i
, for i = 0, 1, 2 and for some finite constant K < ∞.
∇iθ ḡ(θ) ≤ K 1 + kθk
3. ∇ḡ(θ) is globally Lipschitz.
Theorem 2.10. Assume that Conditions 2.1, 2.2, 2.3, 2.4, 2.5 and 2.9 hold. Then,
lim k∇θ ḡ(θ)k = 0,
t→∞
a.s.
Theorem 2.10 proves convergence for non-convex ḡ(θ) even when f (x, θ) grows at most linearly in θ.
Theorem 2.10 is proven in Appendix B. The proof is based on the uniform bound for the moments of θ
established in Appendix A.
Theorem 2.10 is required for proving a central limit theorem for non-convex ḡ(θ). The central limit
theorem for non-convex ḡ(θ) is proven in Theorem 2.12.
1. ḡ(θ) may be non-convex but has a single critical point θ∗ .
Condition 2.11.
2. There is a δ ∗ > 0 small enough such that ḡ(θ) is strongly convex in the region kθ − θ∗ k < δ ∗ with
constant C and CCα > 1.
3. ḡ(θ) ∈ C 3 .
4. ∇θi ḡ(θ) > 0 if θi − θi∗ > R0 and ∇θi ḡ(θ) < 0 if θi − θi∗ < −R0 for i = 1, 2, . . . , k and for some R0 > 0
large enough.
The fourth part in Condition 2.11 guarantees that ∇θi ḡ(θ) always points inwards towards the global
minimum θ∗ if |θi | is sufficiently large.
Theorem 2.12. Assume that Conditions 2.1, 2.2, 2.3, 2.4, 2.5, 2.9 and 2.11 hold. Then, we have that
√
d
t(θt − θ∗ ) → N (0, Σ̄),
where Σ̄ is defined as in Theorem 2.8.
If f (x, θ) and ∇θ f (x, θ) are uniformly bounded in θ and polynomially bounded in x, Theorem 2.12 is
true without Conditions 2.3 and 2.4.
Proposition 2.13 shows that under certain conditions on the learning rate αt , the specific form of the
learning rate assumed in Condition 2.5 is not necessary. In particular, the convergence rate and central limit
theorem results can be proven for a general learning rate αt . The proof of Proposition 2.13 follows exactly
the same steps as the proofs of Theorems 2.7, 2.8, and 2.12 albeit more tedious algebra and is omitted.
(p)
Rt
Proposition 2.13. Let us denote Ψt,s = e−pC s αρ dρ , for p ≥ 1 and consider the matrix-valued solution
Φ∗t,s to equation (4.1). Theorems 2.7, 2.8, and 2.12 also hold under a general learning rate αt satisfying the
conditions:
Z ∞
Z ∞
Z ∞
αt dt = ∞,
α2t dt < ∞,
|α′s |ds < ∞, ∃p > 0 such that lim αt tp = 0,
t→∞
0
0
0
Z t
Z t
p/2
(1)
1/2
(p)
α2s Ψt,s ds ≤ o(αt )
α2s + |α′s | Ψt,s αsp/2−1 ds ≤ O(αt ), and
∀p ≥ 2
0
0
Z t
5/2 (2)
αs Ψt,s ds ≤ o(αt ),
0
Z t
2
α2s Φ∗t,s ds = O(αt ),
0
∀p ≥ 2
(p)
p/2
(1)
1/2
Ψt,0 ≤ O(αt ) and Ψt,0 ≤ o(αt ).
6
In particular, the statements of Theorems 2.7, 2.8, and 2.12 then take the form
p/2
p
E[kθt − θ∗ k ] ≤ Kαt
and
−1/2
αt
d
(θt − θ∗ ) → N (0, Σ̄).
h
i
k
C2
where now Σ̄ = Σ̄i,j i,j=1 is as in (4.9) but with the bracket term (λm +λ α′ )Cα −1 replaced by
m
lim α−1
t
t→∞
Z
t
α2s e−(λm +λm′ )
0
Rt
s
αu du
ds .
α
as the learning rate, then the conditions that appear in
It is easy to check that if we use αs = CC0 +s
Proposition 2.13 all hold if CCα > 1.
We conclude this section by mentioning that in the bounds that appear in the subsequent sections,
0 < K < ∞ will denote unimportant fixed constants (that do not depend on t or other important parameters).
The constant K may change from line to line but it will always be denoted by the same symbol K. Without
loss of generality and in order to simplify notation in the proofs, we will let C0 = 0, consider t ≥ 1 (i.e., the
initial time is set to t = 1), and let σ be the identity matrix.
3
Theorem 2.7 - Lp convergence rate in strongly convex case
The proofs in this paper will repeatedly make use of two important uniform moment bounds. First, as we
prove in Appendix A.1, we have that
sup E[kθt kp ] < K,
t≥0
for p ≥ 2. Second, it is known from [12] that under the imposed conditions for the X process
p
sup E[kXt k ] < K,
t≥0
for p ≥ 2.
To begin the proof for the Lp convergence rate, re-write the algorithm (1.2) for θt in terms of g(x, θ) and
ḡ(θ).
dθt
= αt ∇θ f (Xt , θt )(dXt − f (Xt , θt )dt)
= αt ∇θ f (Xt , θt )(f ∗ (x) − f (Xt , θt ))dt + αt ∇θ f (Xt , θt )dWt
= −αt ∇θ g(Xt , θt )dt + αt ∇θ f (Xt , θt )dWt
= −αt ∇θ ḡ(θ) + αt ∇θ ḡ(θt ) − ∇θ g(Xt , θt ) dt + αt ∇θ f (Xt , θt )dWt .
(3.1)
A Taylor expansion yields:
∇θ ḡ(θt ) =
=
∇θ ḡ(θ∗ ) + ∆ḡ(θt1 )(θt − θ∗ )
∆ḡ(θt1 )(θt − θ∗ ),
where θt1 is an appropriately chosen point in the segment connecting θt and θ∗ . Substituting this Taylor
expansion into equation (3.1) produces the equation:
d(θt − θ∗ ) = −αt ∆ḡ(θt1 )(θt − θ∗ )dt + αt (∇θ ḡ(θt ) − ∇θ g(Xt , θt ))dt + αt ∇θ f (Xt , θt )dWt .
Let Yt = θt − θ∗ . Then, Yt satisfies the SDE
dYt = −αt ∆ḡ(θt1 )Yt dt + αt (∇θ ḡ(θt ) − ∇θ g(Xt , θt ))dt + αt ∇θ f (Xt , θt )dWt .
7
By Itô formula, we then have for p ≥ 2
X
p
p−2
p−2
d kYt k = p kYt k
Ytk αt (∇θ f (Xt , θt ))k,i dWti − pαt kYt k
Yt , ∆ḡ(θt1 )Yt dt
i,k
+ pαt kYt k
p−2
hYt , ∇θ ḡ(θt ) − ∇θ g(Xt , θt )i dt
!2
2
X
X
X
p 2
p−2
−1
+ αt kYt k
(∇θ f (Xt , θt ))i,k + (p − 2)
kYt k Ytk (∇θ f (Xt , θt ))k,i dt. (3.2)
2
i
i,k
k
Using the strong convexity of ḡ we obtain the inequality
X
d kYt kp ≤ p kYt kp−2
Ytk αt (∇θ f (Xt , θt ))k,i dWti − pCαt kYt kp dt
i,k
p−2
+ pαt kYt k
hYt , ∇θ ḡ(θt ) − ∇θ g(Xt , θt )i dt
!2
2
X
X
X
p 2
(∇θ f (Xt , θt ))i,k + (p − 2)
kYt k−1 Ytk (∇θ f (Xt , θt ))k,i dt.
+ αt kYt kp−2
2
i
i,k
k
Let’s now define the process
Z t
Rt
X
p−2
Mt = p
e−pC s αρ dρ αs kYs k
Ysk (∇θ f (Xs , θs ))k,i dWsi ,
1
i,k
and notice that Mt solves the SDE
dMt = −pCαt Mt dt + p kYt k
p−2
X
Ytk (∇θ f (Xt , θt ))k,i dWti .
i,k
p
Next, if we set Γt = kYt k − Mt we obtain that
p−2
dΓt ≤ −pCαt Γt dt + pαt kYt k
hYt , ∇θ ḡ(θt ) − ∇θ g(Xt , θt )i dt
!2
X
X
X
2
p
(∇θ f (Xt , θt ))i,k + (p − 2)
kYt k−1 Ytk (∇θ f (Xt , θt ))k,i dt.
+ α2t kYt kp−2
2
i
i,k
k
Next, we define the function
(p)
Ψt,s = e−pC
Rt
s
αρ dρ
=
s pCCα
t
,
and the comparison principle gives
Z th
i
(p)
(p)
p
p−2
Γt ≤ Ψt,1 kY1 k + p
αs Ψt,s kYs k
hYs , ∇θ ḡ(θs ) − ∇θ g(Xs , θs )i ds
1
!2
Z t
2
X
X
X
p
(p)
p−2
−1
+
Ψ α2 kYs k
(∇θ f (Xs , θs ))i,k + (p − 2)
kYs k Ysk (∇θ f (Xs , θs ))k,i ds
2 1 t,s s
i
i,k
=
Γ1t
+
Γ2t
+
k
Γ3t .
(3.3)
The next step is to rewrite the second term of (3.3), i.e., Γ2t =
i
Rth
(p)
p−2
αs Ψt,s kYs k
hYs , ∇θ ḡ(θs ) − ∇θ g(Xs , θs )i ds.
1
We construct the corresponding Poisson equation and use its solution to analyze Γ2t . Define G(x, θ) =
hθ − θ∗ , ∇θ ḡ(θ) − ∇θ g(x, θ)i and let v(x, θ) be the solution to the PDE Lx v(x, θ) = G(x, θ). Here Lx is the
8
infinitesimal generator of the X process. Due to Theorem A.1, the Poisson PDE solution has the following
bounds:
kv(x, θ)k + k∇x v(x, θ)k ≤ K(1 + kθkm1 )(1 + kxkm2 ),
X ∂ iv
∂ 2v
∂3v
m
m
(x,
θ)
+
(x, θ) ≤ K(1 + kθk 3 )(1 + kxk 4 )
(x,
θ)
+
i
2
∂θ
∂x∂θ
∂x∂θ
i=1,2
(3.4)
for appropriate, but unimportant for our purposes, constants m1 , m2 , m3 , m4 . By Itô’s formula:
Z t
Z t
v(Xt , θt ) − v(Xs , θs ) =
Lx v(Xu , θu )du +
Lθ v(Xu , θu )du
s
s
Z t
Z t
αu ∇θ v(Xu , θu )∇θ f (Xu , θu )dWu
∇x v(Xu , θu )dWu +
+
s
s
Z
+
t
αu ∇θx v(Xu , θu )∇θ f (Xu , θu )du.
s
Define vt ≡ v(Xt , θt ) and recognize that:
G(Xt , θt )dt
Lx v(Xt , θt )dt
=
dvt − Lθ v(Xt , θt )dt − ∇x v(Xt , θt )dWt − αt ∇θ v(Xt , θt )∇θ f (Xt , θt )dWt
=
−
αt ∇θx v(Xt , θt )∇θ f (Xt , θt )dt.
(3.5)
Using this result, Γ2t can be rewritten as:
Γ2t
Z th
i
(p)
p−2
αs Ψt,s kYs k
hYs , ∇θ ḡ(θs ) − ∇θ g(Xs , θs )i ds
1
Z t
(p)
αs Ψt,s kYs kp−2 dvs
1
Z t
(p)
αs Ψt,s kYs kp−2 ∇x v(Xs , θs )dWs
1
Z t
(p)
αs Ψt,s kYs kp−2 Lθ v(Xs , θs )ds
=
=
−
−
1
Z
−
t
1
Z
−
t
α2s Ψt,s kYs k
(p)
p−2
∇θ v(Xs , θs )∇θ f (Xs , θs )dWs
(p)
p−2
∇θx v(Xs , θs )∇θ f (Xs , θs )ds
α2s Ψt,s kYs k
1
Γ2,1
t
=
2,3
2,4
2,5
+ Γ2,2
t + Γt + Γt + Γt .
(3.6)
(p)
Let’s first rewrite the first term Γ2,1
t . We apply Itô’s formula to αs Ψt,s kYs k
(p)
αt Ψt,t kYt k
p−2
(p)
vt − α1 Ψt,1 kY1 k
+
Z
1
(p)
t
αs
∂Ψt,s
∂s
t
t
p−2
vs :
Cα (p)
p−2
Ψt,s kYs k
vs ds
2
s
1
1
Z t
Z t
h
i
(p)
(p)
p−2
p−2
p−2
kYs k
vs ds +
αs Ψt,s vs d kYs k
+
αs Ψt,s d kYs k
, vs .
p−2
v1 =
Z
(p)
p−2
αs Ψt,s kYs k
1
9
dvs −
Z
1
Then, we have the following representation for Γ2,1
t :
Z t
(p)
p−2
Γ2,1
=
αs Ψt,s kYs k
dvs
t
1
Z t
(p)
∂Ψt,s
Cα (p)
p−2
Ψ
kY
k
v
ds
−
α
kYs kp−2 vs ds
=
−
kY1 k
v1 +
s
s
s
t,s
2
∂s
1
1 s
Z t
Z t
i
h
(p)
(p)
−
αs Ψt,s vs d kYs kp−2 −
αs Ψt,s d kYs kp−2 , vs
1
1
Z t
Z t
(p)
(p)
(p)
(p)
= αt Ψt,t kYt kp−2 vt − α1 Ψt,1 kY1 kp−2 v1 + Cα−1
α2s Ψt,s kYs kp−2 vs ds − pC
α2s Ψt,s kYs kp−2 vs ds
(p)
αt Ψt,t kYt kp−2 vt
(p)
α1 Ψt,1
t
Z
p−2
1
−
Z
1
t
(p)
αs Ψt,s vs d kYs k
p−2
−
Z
t
1
(p)
h
p−2
αs Ψt,s d kYs k
1
i
, vs .
(3.7)
Now, we are ready to put things together. Equation (3.2) with p − 2 in place of ph is then usedi to evaluate
p−2
the second to last term of (3.7) and similarly for the quadratic covariation term d kYs k
, vs of the last
term of (3.7). Plugging (3.7) in (3.6) and that in (3.3), we get that there is an unimportant constant K < ∞
large enough and a matrix-valued function ζ(x, θ) that grows at most polynomially in x and θ such that
(p)
p
(p)
(p)
p−2
Γt ≤ Ψt,1 kY1 k + αt Ψt,t kYt k
vt − α1 Ψt,1 kY1 k
Z t
(p)
p−2
ζ(Xs , θs )ds + M̂t ,
+K
α2s Ψt,s kYs k
p−2
v1
(3.8)
1
where M̂t is a mean zero and square integrable (this follows from the uniform moment bounds on the X and
θ processes) Brownian stochastic intergal.
Recall now that we want to evaluate E kYt kp . Recalling the definition of Γt and taking expectation in
(3.8) we obtain
Z t
i
h
(p)
(p)
(p)
p−2
p
p
p−2
p−2
α2s Ψt,s kYs k
ζ(Xs , θs )ds.
E kYt k ≤ E Ψt,1 kY1 k + αt kYt k
vt − α1 Ψt,1 kY1 k
v1 + KE
1
(p)
Ψt,1
(p)
Ψt,1
≤ t−p . Hence we have obtained that
Recalling that
= t−pCCα and that CCa > 1 we get that
for any p ≥ 2 the following inequality holds
Z t
h
i
(p)
p
p−2
p−2
E kYt k ≤ Kt−p + E αt kYt k
vt + KE
α2s Ψt,s kYs k
ζ(Xs , θs )ds.
(3.9)
1
The next step is to proceed by induction. Using the uniform moment bounds for X and θ together with
the polynomial growth of v(x, θ) and ζ(x, θ), we get for p = 2
Z t
(2)
2
E kYt k ≤ Kt−2 + E [αt vt ] + KE
α2s Ψt,s ζ(Xs , θs )ds
1
Z t
−2
−1
−2CCα
2CCα −2
≤K t +t +t
s
ds
1
≤ Kt−1 ,
where the unimportant constant K may change from line to line. Hence the desired statement is true for
p = 2. Next let us assume that it is true for exponent p − 1 and we want to prove that it is true for exponent
p. Using Hölder inequality with exponents r1 , r2 > 1 such that 1/r1 + 1/r2 = 1 and choosing r1 = p−1
p−2 > 1
we get
1/r1
i
h
(E|vt |r2 )1/r2
E αt kYt kp−2 vt ≤ αt E kYt kp−1
≤ Kt−1 t−
10
p−2
2
= Kt−p/2 ,
and likewise
Z t
Z t
1/r1
(p)
(p)
r 1/r
(p−2)r1
p−2
E
(E kζ(Xs , θs )k 2 ) 2 ds
α2s Ψt,s E kYs k
α2s Ψt,s kYs k
ζ(Xs , θs )ds ≤
1
1
≤K
Z
t
1
t
1/r1
(p)
(p−2)r1
ds
α2s Ψt,s E kYs k
1
(p)
α2s Ψt,s (p−2)/2 ds
s
1
Z t
p
−pCCα
≤ Kt
spCCα −1− 2 ds
1
−p/2
≤K t
− t−pCCα
≤K
Z
≤ Kt−p/2 .
Putting the last two displays together, (3.9) gives
E kYt kp ≤ Kt−p/2 ,
which is the statement of Theorem 2.7 for integer p ≥ 2. The statement for any p ≥ 1 then follows from
Hölder inequality. This concludes the proof of the theorem.
4
Proof of Central Limit Theorem in strongly convex case
To prove the central limit theorem, we use a second-order Taylor expansion:
∇θ ḡ(θt ) =
∇θ ḡ(θ∗ ) + ∆ḡ(θ∗ )(θt − θ∗ ) +
1 ∂ 3 ḡ 1
(θ )(θt − θ∗ )(θt − θ∗ )⊤ ,
2 ∂θ3 t
where θt1 is an appropriate point chosen in the segment that connects θt to θ∗ . Then, the evolution of θt
follows:
d(θt − θ∗ ) =
−αt ∆ḡ(θ∗ )(θt − θ∗ )dt −
αt ∂ 3 ḡ 1
(θ )(θt − θ∗ )(θt − θ∗ )⊤ dt + αt (∇θ ḡ(θt ) − ∇θ g(Xt , θt ))dt
2 ∂θ3 t
+αt ∇θ f (Xt , θt )dWt .
Let Yt = θt − θ∗ . Then, Yt satisfies the SDE
dYt = −αt ∆ḡ(θt∗ )Yt dt −
αt ∂ 3 ḡ 1
(θ )Yt Yt⊤ dt + αt (∇θ ḡ(θt ) − ∇θ g(Xt , θt ))dt + αt ∇θ f (Xt , θt )dWt .
2 ∂θ3 t
Let Φ∗t,s ∈ Rk×k be the fundamental solution satisfying
dΦ∗t,s
Φ∗s,s
=
=
−αt ∆ḡ(θ∗ )Φ∗t,s dt,
I,
(4.1)
where I is the identity matrix. As in Section 3 we set without loss of generality C0 = 0 and assume that the
initial time is at t = 1. Then, Yt can be written in terms of Φ∗t,s :
Z t
Z t
Z t
3
∗
∗ αs ∂ ḡ 1
⊤
∗
Yt = Φt,1 Y1 −
Φt,s
(θ )Ys Ys ds +
Φt,s αs (∇θ ḡ(θs ) − ∇θ g(Xs , θs ))ds +
Φ∗t,s αs ∇θ f (Xs , θs )dWs
2 ∂θ3 s
1
1
1
= Γ1t + Γ2t + Γ3t + Γ4t .
(4.2)
qP
2
Next, a convergence rate for Φ∗t,s must be established in the matrix norm kAk =
i,j Ai,j . Consider
Rt
the change of variables τ (t) = s + s αu du and define Φ̃∗τ,s via Φ̃∗τ,s = Φ∗t,s . For αt = Ctα , we have τ (t) =
s + Cα (log(t) − log(s). Performing this change of variables,
dΦ̃∗τ,s
Φ̃∗s,s
= −∆ḡ(θ∗ )Φ̃∗τ,s dτ,
= I,
11
(4.3)
for τ ≥ s. Note that the columns of the matrix solution Φ̃∗τ,s evolve independently. This makes the analysis
∗
∗ i
∗
much simpler. Define Φ̃∗,j
τ,s as the j-th column of Φ̃τ,s and ∆ḡ(θ ) as the i-th row of the matrix ∆ḡ(θ ).
Consider the one-dimensional differential equation:
2
d
d
∗,j,k 2
∗,j,1 2
∗,j
+ · · · + Φ̃τ,s
Φ̃τ,s
Φ̃τ,s
=
dτ
dτ
∗,j,1
∗ 1 ∗,j
∗,j,k
∗ k ∗,j
= −2 Φ̃τ,s ∆ḡ(θ ) Φ̃τ,s + · · · + Φ̃τ,s ∆ḡ(θ ) Φ̃τ,s
⊤
⊤
∗,j
∗
∗,j
∗,j
Φ̃∗,j
∆ḡ(θ
)
Φ̃
<
−2C
Φ̃
= −2 Φ̃∗,j
τ,s = −2C Φ̃τ,s
τ,s
τ,s
τ,s
2
,
where we have used the strong convexity assumption. This of course yields the convergence rate:
2
Φ̃∗τ,s
≤ Ke−2C(τ −s).
Changing variables again yields the convergence rate in the original time coordinate t:
Φ∗t,s
2
=
Φ̃∗τ,s
2
≤ Ke−2C
≤ Ke−2C(τ −s)
Rt
s
α̃u du
= Kt−2CCα s2CCα .
(4.4)
We recall some important properties of Φ∗t,s (for reference, see Proposition 2.14 in [14]). Φ∗t,s is differentiable in t with the semi-group property Φ∗t,1 = Φ∗t,s Φ∗s,1 . Furthermore, Φ∗t,s is invertible with inverse
= Φ∗s,t . Therefore,
Φ∗,−1
t,s
∂ ∗
Φ
∂s t,s
∂
∗,−1
∗
Φt,1 Φs,1
=
∂s
∗
∂
∗,−1 ∂Φs,1 ∗,−1
∗
= Φ∗t,1 Φ∗,−1
Φ
s,1 = Φt,1 Φs,1
∂s
∂s s,1
= −Φ∗t,s αs ∆ḡ(θ∗ )Φ∗s,1 Φ∗,−1
s,1
= −αs Φ∗t,s ∆ḡ(θ∗ ).
(4.5)
Combining the calculation (4.5) with the bound (4.4) yields:
∂ ∗
Φ
∂s t,s
2
≤ Kαs t−2CCα s2CCα .
√
Now, let’s examine the
process tYt , which represents the fluctuations of θ√
t around the global minimum
√
a.s.
p
θ∗ . First, observe that tΓ1t → 0 due to the bound (4.4). Let’s now examine tΓ2t . Recall that E[kYt k ] ≤
K
from the Lp convergence rate in the previous section. Applying this result with p = 2, 3 we obtain
tp/2
√ 2
tΓt
E
1
Z
√
≤ E t
t
∂ 3 ḡ 1
⊤
(θ
)Y
Y
ds
s
s
∂θ3 s
1
1
Z
t
√
1
1
∗
+ 5/2 ds
≤ K t
Φt,s
s2
s
1
√ −CC Z t CC −2
√
α
≤ K tt
s α ds ≤ K tt−1 = Kt−1/2 .
Φ∗t,s αs
(4.6)
1
√
p
Therefore, tΓ2t → 0.
√
Next, we address the term tΓ3t . We construct the corresponding Poisson equation and use its solution
to analyze Γ3t . Define G(x, θ) = ∇θ ḡ(θ) − ∇θ g(x, θ) and let v(x, θ) be the solution to the PDE Lx v(x, θ) =
12
√
G(x, θ). Proceeding in a fashion similar to equation (3.5) (but with different function v now) we write tΓ3t
as:
√ 3
√ Z t
tΓt =
t
αs Φ∗t,s (∇θ ḡ(θs ) − ∇θ g(Xs , θs ))ds
1
√ Z t
t
=
αs Φ∗t,s dvs
1
√ Z t
t
αs Φ∗t,s ∇x v(Xs , θs )dWs
−
1
√ Z t
−
t
αs Φ∗t,s Lθ v(Xs , θs )ds
1
√ Z t 2 ∗
t
−
αs Φt,s ∇θ v(Xs , θs )∇θ f (Xs , θs )dWs
1
Z
t
√
α2s Φ∗t,s ∇θx v(Xs , θs )∇θ f (Xs , θs )ds
t
−
1
√ 3,1 √ 3,2 √ 3,3 √ 3,4 √ 3,5
=
tΓt + tΓt + tΓt + tΓt + tΓt ,
(4.7)
where v(x, θ) satisfies bounds similar to the ones in (3.4). Following similar steps as in Section 3,
√
L1
3,1
3,3
3,4
3,5
t |Γt | + |Γt | + |Γt | + |Γt | → 0,
√
3,3
3,4
3,5 p
which of course
that t(Γ3,1
t + Γt + Γt + Γt ) → 0.
√ implies
√
4
The term tΓ3,2
t + tΓt is now analyzed. This term will produce the limiting Gaussian random variable.
Recalling the reduction to σ = I, we have
Z t
√ 3,2 √ 4 √
(4.8)
tΓt + tΓt = t
αs Φ∗t,s ∇θ f (Xs , θs ) − ∇x v(Xs , θs ) dWs .
1
⊤
R
Let h(x, θ) = ∇θ f (x, θ) − ∇x v(x, θ) ∇θ f (x, θ) − ∇x v(x, θ)
and h̄(θ) = h(x, θ)π(dx).
q
q
Let
3
H(x, θ) = h(x, θ) − h̄(θ). Recall that k∇θ f (x, θ)k ≤ K(1 + kxk + kθk), k∇x v(x, θ)k ≤ K(1 + kxk + kθk ),
2
q
3
2
q
∂ v
≤ K(1 + kxk + kθk ), and ∂x∂θ
≤ K(1 + kxk + kθk). The latter bounds on v(x, θ)
2 (x, θ)
and its derivatives are from equation (3.4). These bounds imply that the function H(x, θ) and its derivatives have polynomial growth in kxk and kθk. Based on Theorem A.1, the solution w(x, θ) to the PDE
Lx w(x, θ) = H(x, θ) and its derivatives will also have at most polynomial growth in kxk and kθk.
√ 4 d
√
tΓt → N (0, Σ̄) for the appropriate limiting variance-covariance matrix Σ̄.
We will prove that tΓ3,2
t +
The proof will rely upon the Poisson partial differential equation approach using the at most polynomial
growth of its solution and its derivatives together with the uniform boundedness of the moments of Xt and
θt processes to analyze the rate of convergence.
√ 4
√
tΓt is:
The quadratic covariation matrix of tΓ3,2
t +
Z t
Σt = t
α2s Φ∗t,s h(Xs , θs )Φ∗,⊤
t,s ds.
∂ v
∂x∂θ (x, θ)
1
p
It is necessary to show that Σt → Σ̄ as t → ∞. To begin, we show a simpler limit. Consider the process:
Z t
Σ̄t = t
α2s Φ∗t,s h̄(θ∗ )Φ∗,⊤
t,s ds.
1
It will be proven now that Σ̄t converges to a limit Σ̄ as t → ∞. Let us now establish the limit Σ̄ as
t → ∞. Recall that ∆ḡ(θ∗ ) is both strictly positive matrix and symmetric. Therefore, by the eigenvalue
decomposition we can write that
∆ḡ(θ∗ ) = U ΛU ⊤ ,
13
where U = [u1 , · · · , uk ] is the corresponding matrix of orthogonal eigenvectors u1 , . . . , uk ∈ Rk and Λ =
diag(λ1 , · · · , λk ) is the diagonal matrix with the positive eigenvalues λi > 0 for i = 1, · · · , k. Next, we notice
Rt
that a time transformation as in (4.3) allows us to write that for τ (t) = s + s αu du and Φ̃∗τ (t),s = Φ∗t,s
Φ̃∗τ,s = e−∆ḡ(θ
∗
)(τ −s)
.
Combining the latter results we then obtain that
Φ∗t,s = U e−Λ
Rt
s
αu du
U ⊤.
For notational convenience, let us now set
Λt,s = e−
Rt
s
αu duΛ
Let us write λ(t,s),i =
s λk Cα
Rt
Rt
s λ1 Cα
.
,··· ,
= diag e−λ1 s αu du , · · · , e−λk s αu du = diag
t
t
s λi Cα
t
for i = 1, · · · , k. The (i, j)-th element of the matrix Φ∗t,s takes the form
Φ∗t,s,i,j
k
X
=
λ(t,s),m ui,m uj,m .
m=1
P
Naturally (Φ∗,⊤ )t,s,k′ ,j = Φ∗t,s,j,k′ = km′ =1 λ(t,s),m′ uj,m′ uk′ ,m′ . Hence, the (i, j)-th element of the matrix
Σ̄t takes the form
#
Z t "X
k
X
∗
2
∗,⊤
∗
Φt,s,i,n
αs
Σ̄t,i,j = t
h̄(θ )n,k′ (Φ )t,s,k′ ,j ds
1
=t
Z
1
=
n
t
α2s
k′ =1
"
X X
n
λ(t,s),m ui,m un,m
m
!
X
∗
h̄(θ )n,k′
X
λ(t,s),m′ uj,m′ uk′ ,m′
m′
k′
!#
ds
XXXX Z t
t
α2s λ(t,s),m ui,m un,m h̄(θ∗ )n,k′ λ(t,s),m′ uj,m′ uk′ ,m′ ds
n
m
k′
m′
1
X X X X Z t s (λm +λm′ )Cα
ds ui,m un,m h̄(θ∗ )n,k′ uj,m′ uk′ ,m′
=
t
α2s
t
1
n m k ′ m′
XXXX
Cα2
1 − t1−(λm +λm′ )Cα ui,m un,m h̄(θ∗ )n,k′ uj,m′ uk′ ,m′ .
=
(λm + λm′ ) Cα − 1
′
′
n m
k
m
By the Rayleigh-Ritz theorem, λ1 , . . . , λk ≥ C. Therefore, Cα (λm + λk′ ) ≥ 2Cα C > 1. Then, we get
that the (i, j)-th element of the limiting quadratic covariation matrix Σ̄ takes the form
XXXX
Cα2
Σ̄i,j = lim Σ̄t,i,j =
ui,m un,m h̄(θ∗ )n,k′ uj,m′ uk′ ,m′
t→∞
′
(λ
+
λ
)
C
−
1
m
m
α
n m k ′ m′
XX
X
X
Cα2
=
(4.9)
uj,m′ uk′ ,m′ .
ui,m un,m
h̄(θ∗ )n,k′
(λm + λm′ ) Cα − 1
′
′
n m
m
k
Notice that we can conveniently write that
Z ∞
⊤
∗
∗
Σ̄ = Cα2
e−s(Cα ∆ḡ(θ )−I) h̄(θ∗ )e−s(Cα (∆ḡ) (θ )−I ) ds.
(4.10)
0
It remains to show that E Σt − Σ̄t
E Σt − Σ̄
Then, E Σt − Σ̄
1
1
→ 0 as t → ∞. If this is true, then the triangle inequality yields
≤ E Σt − Σ̄t
1
p
1
→ 0 would imply that Σt → Σ̄.
14
+ E Σ̄t − Σ̄
1
→ 0.
To prove that E Σt − Σ̄t
→ 0, we begin by defining
Z t
V̄t = t
α2s Φ∗t,s h̄(θs )Φ∗,⊤
t,s ds.
1
1
≤ Σt − V̄t 1 + V̄t − Σ̄t 1 . We first address the second term:
Z t
∗,⊤
∗
∗
.
(4.11)
t
α2s Φ∗t,s h̄(θs )Φ∗,⊤
t,s − Φt,s h̄(θ )Φt,s ds
By the triangle inequality, Σt − Σ̄t
V̄t − Σ̄t
1
=
1
1
1
The (i, j)-th element of the matrix V̄t − Σ̄t is:
Z t X
k
k
X
X
X
∗,⊤
∗,⊤
2
∗
∗
∗
αs
V̄t − Σ̄t i,j = t
Φt,s,i,n
h̄(θs )n,k′ Φt,s,k′ ,j −
Φt,s,i,n
h̄(θ )n,k′ Φt,s,k′ ,j ds.
1
=
t
Z
n
t
1
α2s
X
k
X
Φ∗t,s,i,n
n
θs1
n
k′ =1
k′ =1
k′ =1
∗,⊤
∗
∇θ h̄(θs1 )⊤
n,k′ (θs − θ )Φt,s,k′ ,j ds,
where
is an appropriately chosen point in the segment connecting θs and θ∗ . Recall now that v and its
derivatives can grow at most polynomially in kxk and kθk. As a result of the specific growth rates we have
6
∇θ h̄(θ) ≤ K(1 + kθk ).
2
In addition, E[kθs − θ∗ k ] ≤ K
s from the convergence rate in Section 3. Using these facts, the uniform in
time moments bound on θt , and the Cauchy Schwartz inequality,
Z t
k
X
X
∗,⊤
∗
Φ∗t,s,i,n
∇θ h̄(θs1 )⊤
α2s E
E[| V̄t − Σ̄t i,j |] ≤ t
n,k′ (θs − θ )Φt,s,k′ ,j ds
1
n
t
≤ Kt
Z
t
≤ Kt
Z
1
α2s
k′ =1
X
Φ∗t,s,i,n
n
1
k
X
E ∇θ h̄(θs1 )n,k′
k′ =1
5
s− 2 Φ∗t,s
2
2
21
21
2
E kθs − θ∗ k
Φ∗,⊤
t,s,k′ ,j ds
ds.
Therefore, V̄t − Σ̄t 1 → 0 as t → ∞.
Now, let’s address Σt − V̄t 1 using the Poisson equation method.
Z t
∗,⊤
∗
ds.
−
Φ
h̄(θ
)Φ
Σt − V̄t = t
α2s Φ∗t,s h(θs , Xs )Φ∗,⊤
s
t,s
t,s
t,s
1
The (i, j)-th element of the matrix Σt − V̄t is:
Z t X
k
X
2
∗
αs
Φt,s,i,n
V̄t − Σ̄t i,j = t
h(θs , Xs )n,k′ − h̄(θs )n,k′ Φ∗,⊤
t,s,k′ ,j ds.
1
n
k′ =1
Using Itô’s formula, the bounds (A.6) on the Poisson equation Lx w = H, the moment bounds on Xt and
θt , and the Itô Isometry, it can be shown that
E V̄t − Σ̄t i,j → 0,
as t → ∞. This of course implies that E V̄t − Σ̄t
p
1
→ 0. Combining results and using the triangle inequality,
we have the desired result Σt → Σ̄ as t → ∞.
The convergence in probability of the quadratic variation Σt for equation (4.8) implies that (4.8) converges in distribution to a mean zero normal random variable with covariance Σ̄ (see Section 1.2.2 in [15]).
Combining all of the results yields the central limit theorem:
√
d
tYt → N (0, Σ̄),
which is our desired result.
15
5
Proof of Central Limit Theorem in the non-convex case
From Theorem 2.10, we know that k∇ḡ(θt )k → 0 almost surely. Under the imposed conditions this implies
that either kθt − θ∗ k → 0 or kθt k → ∞. We must therefore first show that kθt k remains finite almost surely.
The parameter θt evolves as:
Z t
Z t
Z t
θt = θ1 +
−αs ∇θ ḡ(θs )ds +
αs ∇θ ḡ(θs ) − ∇θ g(θs , Xs ) ds +
αs ∇θ f (Xs , θs )dWs .
(5.1)
1
1
1
Recall from Condition 2.11 that ∇θi ḡ(θ) > 0 if θi − θi∗ > R0 and ∇θi ḡ(θ) < 0 if θi − θi∗ < −R0 for
i = 1, 2, . . . , k. If |θt,i (ω)| → ∞, then either θt,i (ω) → +∞ or θt,i (ω) → −∞ since θt has continuous
paths (i.e., a divergent sequence such as −2n cannot occur). Next, note that the second and third integrals
in equation (5.1) converge to finite random variables almost surely. Suppose that θt,i (ω) → +∞. This
Rt
implies that there exists a T (ω) such that θt,i > θi∗ for all t ≥ T (ω). However, T (ω) −αs ∇θi ḡ(θs )ds < 0.
This, combined with the fact that the second and third terms in (5.1) converge to finite random variables
almost surely, proves that θt,i < +∞ with probability one. A similar argument shows that θt,i > −∞ with
probability one. Therefore, |θt,i − θi∗ | → 0 almost surely.
a.s.
θt → θ∗ implies that, for any δ > 0 sufficiently small, there exists a (random) time τδ such that
kθt − θ∗ k < δ for all t ≥ τδ . Notice that for fixed δ > 0, τδ is almost surely finite, but we also remark that
τδ is not a stopping time. In this section, the parameter δ is chosen sufficiently small such that 0 < δ < δ ∗ ,
where δ ∗ is taken from Condition 2.11. Using a Taylor expansion:
∇θ ḡ(θt ) =
∇θ ḡ(θ∗ ) + ∆ḡ(θt1 )(θt − θ∗ ) = ∆ḡ(θt1 )(θt − θ∗ ),
where θt1 is appropriately chosen in the segment connecting θt to θ∗ . Then we have
d(θt − θ∗ ) = −αt ∆ḡ(θt1 )(θt − θ∗ )dt + αt (∇θ ḡ(θt ) − ∇θ g(Xt , θt ))dt + αt ∇θ f (Xt , θt )dWt .
Let Φt,s satisfy
dΦt,s
=
Φs,s
=
−αt ∆ḡ(θt1 )Φt,s dt,
I,
and notice that for t > s > τδ ,
2
kΦt,s k ≤ Kt−2CCα s2CCα .
We will also later make use of the fact that Φt,s satisfies the semi-group property Φt,s = Φt,τ Φτ,s (for
reference, see Proposition 2.14 in [14]).
Letting Yt = θt − θ∗ , we obtain
Z t
Z t
αs Φt,s (∇θ ḡ(θs ) − ∇θ g(Xs , θs ))ds +
αs Φt,s ∇θ f (Xs , θs )dWs
Yt = Φt,1 Y1 +
1
=
1
Γ1t + Γ2t + Γ3t .
The first term Γ1t is analyzed below.
Lemma 5.1.
√ 1 a.s.
tΓt → 0,
as t → ∞.
16
Proof. For t ≥ τδ we have
√ 1
tΓt
1
√
t Φt,τδ̄ Φτδ̄ ,1 Y1 1
√
≤ K t Φt,τδ̄ Φτδ̄ ,1 Y1
=
1
≤ Kt−CCα + 2 τδ̄CCα Φτδ̄ ,1 Y1
1
= K(τδ̄ )t−CCα + 2 .
where K(τδ̄ ) = τδ̄CCα Φτδ̄ ,1 Y1 is almost surely finite since P[τδ̄ < ∞] = 1. Then, since CCα > 1,
as t → ∞.
√ 1 a.s.
tΓt → 0
Let v(x, θ) satisfy the Poisson PDE Lx v = G(x, θ) where G(x, θ) = ∇θ ḡ(θ) − ∇θ g(x, θ). The solution
v(t, x) and its relevant partial derivatives will be growing at most polynomially in kxk and linearly in kθk
due to the assumptions of Theorem 2.12. Itô’s formula yields the representation
Z t
Γ2t =
αs Φt,s (∇θ ḡ(θs ) − ∇θ g(Xs , θs ))ds
1
Z t
=
αs Φt,s dvs
1
Z t
−
αs Φt,s ∇x v(Xs , θs )dWs
1
Z t
−
αs Φt,s Lθ v(Xs , θs )ds
1
Z t
α2s Φt,s ∇θ v(Xs , θs )∇θ f (Xs , θs )dWs
−
1
Z t
α2s Φt,s ∇θx v(Xs , θs )∇θ f (Xs , θs )ds
−
=
1
Γ2,1
t
2,4
2,5
2,3
+ Γ2,2
t + Γt + Γt + Γt .
In order now to analyze the terms involved in Γ2t and Γ3t we need some intermediate results that we state
now below. For presentation purposes the proof of these Lemmas is deferred to the end of this section.
Lemma 5.2. Let ζ(x, θ) be a (potentially matrix-valued) function that can grow at most polynomially in x
and θ. Then, we have that
√ Rt
a.s.
1. It1 = t 1 α2s Φt,s ζ(Xs , θs )ds → 0, as t → ∞, and
√ Rt
2. It2 = t 1 α2s Φt,s ζ(Xs , θs )dWs → 0, in probability, as t → ∞.
√
(recall√that the operator Lθ has a multiplicative
Lemma 5.2√immediately gives that the terms tΓ2,3
t
converge
almost
surely
to
zero
and
that tΓ2,4
converges in probability to zero, as
factor of αt ), tΓ2,5
t
t
t → ∞. To see this it is enough to notice
that
all
of
these
terms
take
the
form of the It1 and It2 quantities
√ 2,1
mentioned in Lemma 5.2. The term tΓt also converges almost surely to zero by first rewriting using Itô
(p)
formula, as it was done for the corresponding term of Section 3 (refer to (3.7) with p = 2 and replace Ψt,s
by Φt,s ) and then using again Lemma 5.2.
limiting Gaussian random variable will be produced by Γ2,2
and Γ3t . Therefore, it remains to analyze
t
√ The
2,2
3
t(Γt + Γt ). We have
√ Z t
√ 2,2
3
t(Γt + Γt ) = t
αs Φt,s ∇θ f (Xs , θs ) − ∇x v(Xs , θs ) dWs .
1
Using the results from [15], it is sufficient to prove the convergence in probability of
Z t
Σt = t
α2s Φt,s h(Xs , θs )Φ⊤
t,s ds
1
17
⊤
to a deterministic quantity. We recall here that h(x, θ) = ∇θ f (x, θ)−∇x v(x, θ) ∇θ f (x, θ)−∇x v(x, θ) .
As before, let Φ∗t,s be the solution to
dΦ∗t,s
Φ∗s,s
−αt ∆ḡ(θ∗ )Φ∗t,s dt,
I.
=
=
Recall that Φ∗t,s satisfies the bound
Φ∗t,s
2
≤ Kt−2CCα s2CCα .
Define Σ̄t and Σ̄∗t as:
Σ̄t
=
t
Z
t
1
Σ̄∗t
=
t
Z
1
t
α2s Φt,s h̄(θ∗ )Φ⊤
t,s ds.
α2s Φ∗t,s h̄(θ∗ )Φ∗,⊤
t,s ds.
Note that we already proved in the previous section that Σ̄∗t converges in probability to Σ̄ as t → ∞. We
would like to show that Σ̄t − Σ̄∗t → 0 as t → ∞, almost surely.
Lemma 5.3.
Σ̄t − Σ̄∗t → 0, with probability 1, as t → ∞.
To prove that Σt − Σ̄t → 0, we begin by defining
V̄t = t
Z
1
t
α2s Φt,s h̄(θs )Φ⊤
t,s ds
By the triangle inequality, Σt − Σ̄t ≤ Σt − V̄t + V̄t − Σ̄t . We then have the following lemmas.
Lemma 5.4.
V̄t − Σ̄t → 0, with probability 1, as t → ∞.
Lemma 5.5.
Σt − V̄t → 0, in probability, as t → ∞.
By the triangle inequality,
Σt − Σ̄
≤
≤
Σt − Σ̄t + Σ̄t − Σ̄
Σt − V̄t + V̄t − Σ̄t + Σ̄t − Σ̄ .
p
Combining Lemmas 5.5, 5.3, and 5.4, Σt − Σ̄ → 0 as t → ∞. Therefore, using the results from [15],
√ 2,2
d
t(Γt + Γ3t ) → N (0, Σ̄),
as t → ∞.
Combining all of the results yields the central limit theorem:
√
d
tYt → N (0, Σ̄),
which is our desired result.
5.1
Proof of Lemmas 5.2-5.5
In this subsection we give the proof of the lemmas that were used in the proof of the central limit theorem
of the non-convex case.
First we need an intermediate result to properly handle convergence to zero of multidimensional stochastic
integrals. Such a result should be standard in the literature, but because we did not manage to locate an
appropriate statement we present Lemma 5.6.
18
Rt
Lemma 5.6. Let Zt = 1 b(t, s, Xs , θs )dWs . Let p ∈ N be a given integer, and consider a constant c > p − 21
and a matrix E where EE ⊤ is positive definite, such that
Z t
1
p
t−c+p− 2
αps sc b(t, s, Xs , θs )E ⊤ ds → 0,
(5.2)
1
as t → ∞. Then, if
Rt
1
p
p
b(t, s, Xs , θs )b(t, s, Xs , θs )⊤ ds → 0 as t → ∞, Zt → 0 as t → ∞.
Proof of Lemma 5.6. Let η > 0 be arbitrarily chosen and construct the random variable
s
Z
2c − 2p + 1 −c+p− 12 t p c
t
αs s EdWs .
Z̃t = η
Cα2p
1
d
d
From Section 1.2.2 in [15], Zt + Z̃t → N (0, η 2 EE ⊤ ) as t → ∞. In addition, Z̃t → N (0, η 2 EE ⊤ ) as
t → ∞. From the triangle inequality,
k
X
ǫ
2
P Zt,i >
P kZt k > ǫ ≤
k
i=1
r
r
k
k
X
X
ǫ
ǫ
P |Zt,i | >
P |Zt,i + Z̃t,i − Z̃t,i | >
=
k
k
i=1
i=1
r
k
X
ǫ
P |Z̃t,i | + |Zt,i + Z̃t,i | >
k
i=1
r
r
k
X
1 ǫ
1 ǫ
P |Z̃t,i | >
+ P |Zt,i + Z̃t,i | >
.
2 k
2 k
i=1
=
≤
≤
2
(5.3)
For each fixed η > 0, the RHS of the inequality (5.3) converges to a finite quantity as t → ∞ due to the
continuous mapping theorem and the convergence in distribution of Zt,i + Z̃t,i and Z̃t . Furthermore, the
limit of the RHS can be made arbitrarily small by choosing a sufficiently small η. Therefore, for any δ > 0,
there exists a η > 0 such that:
2
lim P kZt k > ǫ < δ.
t→∞
Proof of Lemma 5.2. Let us first prove the first statement of the lemma. Without loss of generality, let
t ≥ τδ . To begin, divide [0, t] into two regimes [1, τδ ∧ t] and [τδ ∧ t, τδ ∨ t]:
√ Z t 2
√ Z τδ ∧t 2
√ Z τδ ∨t 2
1
It = t
αs Φt,s ζ(Xs , θs )ds = t
αs Φt,s ζ(Xs , θs )ds + t
αs Φt,s ζ(Xs , θs )ds.
1
=
It1,1
+
1
τδ ∧t
It1,2 .
(5.4)
Let us first study the second term.
It1,2
1
=
√ Z
t
√
≤K t
τδ ∨t
α2s Φt,s ζ(Xs , θs )ds
τδ ∧t
Z τδ ∨t
τδ ∧t
1
α2s kΦt,s k kζ(Xs , θs )k ds
≤ Kt−CCα+1/2
≤ Kt−CCα+1/2
τδ ∨t
Z
τδ ∧t
Z τδ ∨t
1
m1
+ kθs k
m2
) ds
m1
+ kθs k
m2
) ds.
sCCα −2 (1 + kXs k
sCCα −2 (1 + kXs k
19
Let us now define the quantity
Lt = Kt−CCα +1/2
t
Z
sCCα −2 (1 + kXs k
1
m1
m2
+ kθs k
) ds,
and notice that with probability one we have
lim sup It1,2
t→∞
1
≤ lim sup Lt .
t→∞
For ε > 0, consider now the event At,ε = Lt ≥ tε−1/2 . Using the uniform in time bounds for the
moments of Xs and θs we obtain that
Z t
E |Lt | ≤ Kt−CCα +1/2
sCCα −2 ds ≤ K t−1/2 − t−CCα +1/2 .
1
Then, Markov’s inequality and the fact that CCα > 1 give
P [At,ε ] ≤
E |Lt |
≤ Kt−ε .
tε−1/2
The latter then implies that
X
n∈N
P [A2n ,ε ] < ∞,
which then due to the Borel-Cantelli lemma, guarantees the existence of a finite positive random variable
d(ω) and of some n0 < ∞ such that for every n ≥ n0 ,
L2n ≤ d(ω)2−nε .
For any t ∈ [2n , 2n+1 ] and n ≥ n0 ,
Lt
n(−CCα +1/2)
Z
≤
K2
≤
K2CCα−1/2 L2n+1
≤
≤
1
2n+1
sCCα −2 (1 + kXs km1 + kθs km2 )ds.
K2CCα−1/2 d(ω)2−(n+1)ε
K(ω)t−ε .
a.s.
a.s.
Therefore, Lt → 0, which immediately implies that It1,2 → 0. Next, we analyze the first term on the
RHS of equation (5.4). If t > τδ , the semi-group property yields:
It1,1
=
Therefore, if t > τδ ,
√ Z
t
1
√ R τδ ∧t 2
R τ ∧t
√
t 1
αs Φt,s ζ(Xs , θs )ds = tΦt,τδ 1 δ α2s Φτδ ,s ζ(Xs , θs )ds.
τδ ∧t
α2s Φt,s ζ(Xs , θs )ds
Z
√
≤ K t kΦt,τδ k
≤ C(τδ )t
τδ
1
−CCα +1/2
α2s Φτδ ,s ζ(Xs , θs )ds
.
The constant C(τδ ) is almost surely finite since P[τδ < ∞] = 1 and because kXs k and kθs k are almost
surely finite for s ≤ τδ . Therefore, using the constraint CCα > 1, we have obtained
Z τδ ∧t
√
a.s.
α2s Φt,s ζ(Xs , θs )ds → 0.
t
1
Combining results, the integral (5.4) converges to 0 almost surely as t → ∞.
20
Next, we prove the second statement
√ of the lemma, where we shall use Lemma 5.6. In the notation of
Lemma 5.6, let us set b(t, s, x, θ) = tα2s Φt,s ζ(x, θ) and, with some abuse of notation, let us consider the
quantity
Z t
[It2 ] = t
α4s Φt,s ζ(Xs , θs )ζ T (Xs , θs )ΦTt,s ds.
1
Now [It2 ] is a matrix with elements
[It2 ]i,j = t
=t
Z
t
Z
τδ ∧t
1
1
α4s
X
Φt,s,i,n
n
α4s
= [It2 ]1i,j +
X
k
X
ζζ ⊤
k′ =1
Φt,s,i,n
n
[It2 ]2i,j .
k
X
k′ =1
n,k′
ζζ
⊤
(Xs , θs )Φt,s,k′ ,j ds
n,k′
(Xs , θs )Φt,s,k′ ,j ds + t
Z
τδ ∨t
τδ ∧t
α4s
X
Φt,s,i,n
n
k
X
k′ =1
ζζ ⊤
n,k′
(Xs , θs )Φt,s,k′ ,j ds
Proceeding as in the first part of the lemma shows that for each index i, j, the corresponding element of
the matrix [It2 ] goes to zero, i.e.,
[It2 ]i,j = [It2 ]1i,j + [It2 ]2i,j → 0, a.s., as t → ∞.
To complete the proof, we then need to use Lemma 5.6. Let us define
Z t
√
1
α2+p
sc tΦt,s ζ(Xs , θs )E ⊤ ds
Dt = t−c+p− 2
s
1
Z τδ ∧t
Z τδ ∨t
−c+p
c
⊤
−c+p
=t
α2+p
s
Φ
ζ(X
,
θ
)E
ds
+
t
α2+p
sc Φt,s ζ(Xs , θs )E ⊤ ds.
t,s
s s
s
s
1
τδ ∧t
= Dt1 + Dt2 .
By Lemma 5.6, if we show that for the appropriate choices of p, c and E, Dt goes to zero in probability
as t → ∞, then we would have shown that the second statement of the lemma holds, i.e. that It2 → 0 in
probability as t → ∞.
We pick p = 2, c = 2CCα and E = I to be the identity matrix. Let us first start with Dt2 . Notice that
for the (i, j) element of the matrix we have
Z τδ ∨t
2
−2CCα +2
Dt 1 ≤ Kt
α4s s2CCα kΦt,s ζ(Xs , θs )k ds
≤ Kt−3CCα+2
Z
τδ ∧t
τδ ∨t
τδ ∧t
s3CCα −4 (1 + kXs k
m1
+ kθs k
m2
)ds.
It is clear that from this point the analysis of Dt2 is identical to the analysis of It1,2 . In particular, define
the quantity
Z t
m
m
−3CCα +2
L̂t = Kt
s3CCα −4 (1 + kXs k 1 + kθs k 2 )ds,
1
n
o
and, for ε > 0, consider the event Ât,ε = L̂t ≥ tε−1 . Using Markov’s inequality and the uniform in time
bounds for the moments of Xs and θs we obtain that
Rt
i E L̂t
h
t−3CCα +2 1 s3CCα −4 ds
P Ât,ε ≤ ε−1 ≤ K
t
tε−1
−ε
≤ Kt .
21
From here on, the rest of the argument follows the Borel-Cantelli argument that was used for the proof
a.s.
of the first part of the lemma. This yields that Dt2 → 0 as t → ∞.
Next, using the semi-group property and for τδ < t, Dt1 can be re-written as:
Z τδ
Dt1 = t−2CCα +2
α4s s2CCα Φt,τδ Φτδ ,s ζ(Xs , θs )ds.
1
Z τδ
−2CCα +2
=t
Φt,τδ
α4s s2CCα Φτδ ,s ζ(Xs , θs )ds.
1
Therefore, by similar logic as for the term It1,1 , we obtain
Dt1
≤ C(τδ )t−3CCα +2 ,
a.s.
where C(τδ ) is almost surely finite. Therefore, since CCα > 1, Dt1 → 0 as t → ∞. Consequently, as t → ∞,
we have indeed obtained
a.s.
Dt = Dt1 + Dt2 → 0,
which by Lemma 5.6 implies that the second statement of the lemma is true. This concludes the proof of
the lemma.
Proof of Lemma 5.3. Φt,s can be expressed in terms of Φ∗t,s . To see this, first perform a Taylor expansion:
dΦt,s
=
Φs,s
=
−αt ∆ḡ(θ∗ )Φt,s dt − αt Ct Yt Φt,s dt,
I,
Ct are the third-order partial derivatives of ḡ(θ) and, based on the imposed assumptions, they are uniformly
bounded. Therefore,
Z t
Φt,s = Φ∗t,s −
αu Φ∗t,u Cu Yu Φu,s du.
s
Rt
Define ξt,s = s αu Φ∗t,u Cu Yu Φu,s du. Recall that there exists an almost surely finite random time τδ such
that kYt k < δ for all t ≥ τδ and any δ small enough. From the bounds on Φt,s and Φ∗t,s , we have that for
t > s > τδ where δ small enough,
Z t
2
2
kξt,s k ≤ t
α2u Φ∗t,u Cu Yu Φu,s du
s
Z t
2
2
2
≤ tK
α2u Φ∗t,u kYu k kΦu,s k du
s
Z t
≤ Kδ 2 t−2CCα +1 s2CCα
u−2 du
s
2 −2CCα 2CCα
≤ Kδ t
s
.
We have used the Cauchy-Schwartz inequality in the first inequality above. Now, consider the case where
t > τδ but s < τδ .
Z t
αu Φ∗t,u Cu Yu Φu,s du
ξt,s =
s
Z t
Z τδ
αu Φ∗t,u Cu Yu Φu,s du
αu Φ∗t,u Cu Yu Φu,s du +
=
τδ
s
Z τδ
αu Φ∗τδ ,u Cu Yu Φu,s du + ξt,τδ .
= Φ∗t,τδ
s
22
From the previous bound, kξt,τδ k ≤ Kδt−CCα τδCCα . Therefore, using the fact that τδ is almost surely
finite, the bounds on Φ∗t,τδ and ξt,τδ , and the triangle inequality,
kξt,s k ≤ C(τδ )t−CCα ,
when t > τδ but s < τδ .
We will use these results to derive the limit of Σ̄t as t → ∞.
Z t
Σ̄t = t
α2s Φt,s h̄(θ∗ )Φ⊤
t,s ds
1
Z t
Z t
Z t
∗,⊤
⊤
∗
⊤
∗
2
2 ∗
ds
α2s Φ∗t,s h̄(θ∗ )ξt,s
αs ξt,s h̄(θ )Φt,s ds − t
= t
αs Φt,s h̄(θ )Φt,s ds − t
1
1
1
Z t
Z t
⊤
ds.
α2s Φ∗t,s h̄(θ∗ )ξt,s
α2s ξt,s h̄(θ∗ )Φ⊤
= Σ̄∗t − t
t,s ds − t
(5.5)
1
1
Let’s consider the second term in (5.5). Without loss of generality, assume t > τδ .
t
Z
2
t
1
α2s ξt,s h̄(θ∗ )Φ⊤
t,s ds
Z
≤ t3
≤ t
t
τδ
Z t
3
2
α2s ξt,s h̄(θ∗ )Φ⊤
t,s
≤ Kδ 2 t−4CCα +3
≤ Kδ 2 + C(τδ )t
Z
t
Z
τδ
1
2
α2s ξt,s h̄(θ∗ )Φ⊤
t,s
τδ
ds + t3
ds + t
3
Z
1
τδ
α2s ξt,s h̄(θ∗ )Φ⊤
t,s
2
ds
⊤
α2s ξt,s h̄(θ∗ )Φ⊤
τδ ,s Φt,τδ
2
ds
s4CCα −4 ds + C(τδ )t−4CCα +3
τδ
−4CCα +3
.
(5.6)
C(τδ ) is almost surely finite since τδ is almost surely finite. The third term in (5.5) is similar. Using (5.5)
and the bound (5.6), we have that with probability one:
lim sup Σ̄t − Σ̄∗t
t→∞
≤ Kδ.
This implies that limt→∞ Σ̄t − Σ̄∗t = 0 since δ can be as small as we want. This concludes the proof of the
lemma.
Proof of Lemma 5.4.
V̄t − Σ̄t
=
t
Z
1
t
∗
⊤
α2s Φt,s h̄(θs )Φ⊤
t,s − Φt,s h̄(θ )Φt,s ds .
Using the semi-group property,
Z τδ
∗
⊤
⊤
⊤
⊤
2
αs Φt,τδ Φτδ ,s h̄(θs )Φτδ ,s Φt,τδ − Φt,τδ Φτδ ,s h̄(θ )Φτδ ,s Φt,τδ ds
V̄t − Σ̄t = t
1
Z t
∗
⊤
⊤
2
αs Φt,s h̄(θs )Φt,s − Φt,s h̄(θ )Φt,s ds.
+ t
τδ
The (i, j)-th element of the matrix V̄t − Σ̄t is:
V̄t − Σ̄t
i,j
=
k
X
t
m,ℓ,n,k′ =1
+
t
Z
t
τδ
α2s
Z
X
n,k′
1
τδ
∗
α2s Φt,τδ ,i,m Φτδ ,s,m,ℓ ∇θ h̄(θs1 )⊤
ℓ,n (θs − θ )Φτδ ,s,n,k′ Φt,τδ ,k′ ,j
∗
Φt,s,i,n ∇θ h̄(θs1 )⊤
n,k′ (θs − θ )Φt,s,k′ ,j ds.
23
By the Cauchy-Schwartz inequality and the bound ∇θ h̄(θ) ≤ K(1 + kθk) we have
2
Z t X
1 ⊤
∗
2
αs
Φt,s,i,n ∇θ h̄(θs )n,k′ (θs − θ )Φt,s,k′ ,j ds
t
≤ t3
Z
τδ
n,k′
t
X
τδ
α4s
n,k′
2 −4CCα +3
≤ Kδ t
≤ Kδ 2 .
We also have that:
Z
t
τδ
1
≤ t3
Z
τδ
1
≤ Kt3
k
X
α2s
m,ℓ,n,k′ =1
α4s
m,ℓ,n,k′ =1
≤ C(τδ )t
Z
k
X
ds
s4CCα −4 ds
τδ
(5.7)
−θ
∗
⊤
)Φ⊤
τδ ,s,n,k′ Φt,τδ ,k′ ,j ds
∗
Φt,τδ ,i,m Φτδ ,s,m,ℓ ∇θ h̄(θs1 )⊤
ℓ,n (θs − θ )Φτδ ,s,n,k′ Φt,τδ ,k′ ,j
2 Z
Φt,τδ ,k′ ,j Φt,τδ ,i,m
τδ
1
−4CCα +3
2
t
Φt,τδ ,i,m Φτδ ,s,m,ℓ ∇θ h̄(θs1 )⊤
ℓ,n (θs
m,ℓ,n,k′ =1
k
X
∗
Φt,s,i,n ∇θ h̄(θs1 )⊤
n,k′ (θs − θ )Φt,s,k′ ,j
2
2
ds
2
∗
(θ
−
θ
)
α4s Φ2τδ ,s,m,ℓ ∇θ h̄(θs1 )⊤
Φ2τδ ,s,n,k′ ds
ℓ,n s
.
(5.8)
Since τδ < ∞ with probability 1, C(τδ ) is also finite with probability 1. Combining the results from equations
(5.7) and (5.8),
V̄t − Σ̄t i,j ≤ Kδ 2 + C(τδ )t−4CCα +3 .
Therefore, we have that lim supt→∞ V̄t − Σ̄t ≤ Kδ 2 . Since δ is arbitrarily small, we have that V̄t −
a.s.
Σ̄t i,j → 0, concluding the proof of the lemma.
Proof of Lemma 5.5. Notice that we have
Z t
⊤
h̄(θ
)Φ
−
Φ
α2s Φt,s h(θs , Xs )Φ⊤
Σt − V̄t = t
s
t,s
t,s ds.
t,s
(5.9)
1
and consequently the (i, j)-th element of the matrix Σt − V̄t is:
V̄t − Σ̄t
= t
i,j
Z
t
1
α2s
X
Φt,s,i,n
n
k
X
k′ =1
h(θs , Xs )n,k′ − h̄(θs )n,k′ Φ⊤
t,s,k′ ,j ds.
Let wn,k′ (x, θ) be the solution to the Poisson equation Lwn,k′ (x, θ) = h(θ, x)n,k′ − h̄(θ)n,k′ . The solution
to θ and x due to
wn,k′ (t, θ) and its relevant partial derivatives will grow at most polynomially with respect
the assumptions of Theorem 2.12. The next step is to rewrite the difference V̄t − Σ̄t i,j using that Poisson
equation and Itô’s formula and then to show that each term on the right hand side of the resulting equation
goes to zero.
For example we shall have that
t
Z
t
1
= t
α2s
X
Φt,s,i,n
n
Z
1
t
α2s
X
n
k
X
h(θs , Xs )n,k′ − h̄(θs )n,k′ Φt,s,k′ ,j ds
k′ =1
Φt,s,i,n
k
X
k′ =1
Φt,s,k′ ,j ∇x wn,k′ (Xs , θs )⊤ dWs + (∗),
24
(5.10)
where (∗) is a collection of Riemann integrals resulting from the application of Itô’s formula. Now each
of these terms can be shown to go to zero with an argument exactly parallel to Lemma 5.2. Due to the
similarity of the argument, the details are omitted.
6
Convergence Analysis
The central limit theorem provides an important theoretical guarantee for the performance of the SGDCT
algorithm developed in [3]. Theorem 2.12 is particularly significant since it shows that the asymptotic
1
convergence rate of t− 2 even holds for a certain class of non-convex models. This is important since many
models are non-convex.
In addition, the analysis yields insight into the behavior of the algorithm and provides guidance on
selecting the optimal learning√rate for numerical performance. The regime where the central limit theorem
holds with the optimal rate t is Cα C > 1. Cα is the magnitude of the learning rate. For example, take
α
. Therefore, the learning rate magnitude must be chosen sufficiently large in order to achieve the
αt = CC0 +t
optimal rate of convergence. The larger the constant C is, the steeper the function ḡ(θ) is around the global
minimum θ∗ . The smaller the constant C is, the smaller the function ḡ(θ) is around the global minimum θ∗ .
The “flatter” the region around the global minimum point, the larger the learning rate magnitude must be.
If the region around the global minimum point is steep, the learning rate magnitude can be smaller.
1
The condition of Cα C > 1 to ensure the convergence rate of t− 2 is not specific to the SGDCT algorithm,
but is in general a characteristic of continuous-time statistical learning algorithms. The convergence rate of
any continuous-time gradient descent algorithm with a decaying learning rate will depend upon the learning
rate magnitude Cα . Consider the deterministic gradient descent algorithm
dθt
= −αt ∇θ ḡ(θ).
dt
Let αt =
Cα
C0 +t
and assume ḡ(θ) is strongly convex. Then,
kθt − θ∗ k ≤ Kt−CCα .
Note that the convergence rate depends entirely upon the choice of the learning rate magnitude Cα . If Cα
is very small, the deterministic gradient descent algorithm will even converge at a rate much smaller than
1
t− 2 .
1
t− 2 is the fastest possible convergence rate given that the noise in the system (1.1) is a Brownian motion.
This is due to the quadratic variation of a Brownian motion growing linearly in time. However, other types
1
of noise with variances which grow sub-linearly in time could allow for a faster rate of convergence than t− 2 .
An example of a stochastic process whose variance grows sub-linearly in time is a fractional Brownian motion
with appropriately chosen Hurst parameter. Analyzing the convergence rate under more general types of
noise would be a very interesting topic for future research.
In the central limit theorem result, we are also able to precisely characterize the asymptotic covariance
k
Σ̄ = Σ̄i,j i,j=1 , with
Σ̄i,j =
XX
n
ui,m un,m
m
and
Σ̄ = Cα2
X
∗
h̄(θ )
n,k′
k′
Z
∞
e−s(Cα ∆ḡ(θ
X
m′
∗
)−I)
Cα2
uj,m′ uk′ ,m′
(λm + λm′ ) Cα − 1
h(θ∗ )e−s(Cα (∆ḡ)
⊤
(θ ∗ )−I )
ds.
0
The covariance depends upon the eigenvalues and eigenvectors of the matrix ∆ḡ(θ∗ ), which is the Hessian
matrix ∆ḡ(θ) at the global minimum θ∗ . The larger the eigenvalues, the smaller the variance. This means
that the steeper the function ḡ(θ) is near the global minimum θ∗ , the smaller the asymptotic variance. The
flatter the function ḡ(θ) is near the global minimum, the larger the asymptotic variance. If the function
is very flat, θt ’s drift towards θ∗ is dominated by the fluctuations from the noise Wt . The covariance also
25
depends upon the learning rate magnitude Cα . The larger the learning rate magnitude, the larger the
asymptotic variance Σ̄. Although a sufficiently large learning rate is required to achieve the optimal rate of
1
convergence t− 2 , too large of a learning rate will cause high variance.
7
Conclusion
Stochastic gradient descent in continuous time (SGDCT) provides a computationally efficient method for
the statistical learning of continuous-time models, which are widely used in science, engineering, and finance.
The SGDCT algorithm follows a (noisy) descent direction along a continuous stream of data. The algorithm
updates satisfy a stochastic differential equation. This paper analyzes the asymptotic convergence rate of
the SGDCT algorithm by proving a central limit theorem. An Lp convergence rate is also proven for the
algorithm.
In addition to a theoretical guarantee, the convergence rate analysis provides important insights into the
behavior and dynamics of the algorithm. The asymptotic covariance is precisely characterized and shows
the effects of different features such as the learning rate, the level of noise, and the shape of the objective
function.
Rt
The
proofs in this paper require addressing several challenges. 2First, fluctuations of the form 0 αs h(Xs , θs )−
h̄(θs ) ds must be analyzed. We evaluate, and control with rate αt , these fluctuations using a Poisson partial
differential equation. Secondly, the model f (x, θ) is allowed to grow with θ. This means that the fluctuations
as well as other terms can grow with θ. Therefore, we must prove an a priori stability estimate for kθt k.
Proving a central limit theorem for the non-convex ḡ(θ) in Theorem 2.12 is challenging since the convergence
speed of θt can become arbitrarily slow in certain regions, and the gradient can even point away from the
global minimum θ∗ . We prove the central limit theorem for the non-convex case by analyzing two regimes
[0, τδ ] and [τδ , ∞), where τδ is defined such that kθt − θ∗ k < δ for all t ≥ τδ . The proof also requires the
analysis of stochastic integrals with anticipative integrands, which is challenging since standard approaches
(such as the Itô Isometry) cannot be directly applied.
A
Preliminary Estimates
This section presents two key bounds that are used throughout the paper. Section A.1 proves uniform in
p
time moment bounds for θt . That is, we prove that E[kθt k ] is bounded uniformly in time. Section A.2
presents a bound on the solutions for a class of Poisson partial differential equations. In the paper, we relate
certain equations to the solution of a Poisson partial differential equation and then apply this bound.
A.1
Moment bounds
It is easy to see that for kθk > R (in fact for kθk > 0) one has for kθt k =
q
Pk
i=1
θti
2
n
n θ i θ j ∇ f ∇T f (X , θ )
X
X
∇θ f ∇Tθ f (Xt , θt ) i,i
θ
t t i,j
t
t
hθ
,
∇
g(X
,
θ
)i
θ
t
θ
t
t
dt
+ α2t
d kθt k = −αt
− α2t
3
kθt k
2
kθ
k
t
2
kθ
k
t
i=1
i,j=1
1
+ αt
hθt , ∇θ f (Xt , θt )dWt i
kθt k
T
n
n θ i θ j ∇ f ∇T f (X , θ )
X
X
∇
f
∇
f
(X
,
θ
)
θ
t
t
θ
t
t
hθt , ∇θ g(Xt , θt )i
t t
θ
θ
i,j
i,i
= −αt
+ α2t
dt
− α2t
3
kθt k
2
kθ
k
t
2
kθ
k
t
i=1
i,j=1
+ αt τ (Xt , θt )dW̃t ,
for an independent one-dimensional Brownian motion W̃ . Let us also now consider the process θ̃t which
satisfies
dθ̃t = −αt κ(Xt )θ̃t dt + αt ∇θ f (Xt , θ̃t )dWt ,
26
where κ(x) is from Condition 2.3. Then, we get similarly that for θ̃ > 0 and θ̃t =
d θ̃t = −αt κ(Xt ) θ̃t − α2t
+ αt
1
θ̃t
θ̃ti θ̃tj ∇θ̃ f ∇Tθ̃ f (Xt , θ̃t )
n
X
2 θ̃t
i,j=1
i,j
3
+ α2t
n
X
= −αt κ(Xt ) θ̃t − α2t
2 θ̃t
i,j=1
i,j
3
+ α2t
n
X
θ̃ti
2
∇θ̃ f ∇Tθ̃ f (Xt , θ̃t )
i,i
dt
2 θ̃t
∇θ̃ f ∇Tθ̃ f (Xt , θ̃t )
i,i
dt
2 θ̃t
i=1
θ̃ti θ̃tj ∇θ̃ f ∇Tθ̃ f (Xt , θ̃t )
n
X
i=1
E
D
θ̃t , ∇θ̃ f (Xt , θ̃t )dWt
r
Pk
i=1
+ αt τ (Xt , θ̃t )dW̃t
Due to Conditions 2.3 and 2.4 and continuity of the involved drift and diffusion coefficients for kθk , θ̃ >
R > 0, we may use the comparison theorem (see for example [5]) to obtain that
P kθt k ≤ θ̃t , t ≥ 0 = 1.
(A.1)
It is easy to see that the proof of the comparison theorem 1.1 of [5] goes through almost verbatim,
despite the presence of the term λ(x) in Condition 2.4. The reason is that |λ(x)| is assumed to have at most
polynomial growth in kxk and all moments of Xt are bounded uniformly in t.
Now, notice that θ̃t can be written as the solution to the integral equation
Z t
R
Rt
− 0t αs κ(Xs )ds
θ̃t = θ̃0 e
+
αs e− s αr κ(Xr )dr ∇θ f (Xs , θ̃s )dWs .
0
From the latter representation we obtain, recall that κ(x) is almost surely positive, that for any p ≥ 1
E θ̃t
2p
≤ E θ̃0
≤ E θ̃0
≤ E θ̃0
≤ E θ̃0
≤ E θ̃0
≤ E θ̃0
≤ E θ̃0
2p
+E
t
Z
αs e−
Rt
s
αr κ(Xr )dr
0
2p
+E
2p
+E
0
t
Z
α2s e−2
−2p
α2p
s e
0
2p
t
Z
Rt
s
αr κ(Xr )dr
∇θ f (Xs , θ̃s )
Rt
αr κ(Xr )dr
∇θ f (Xs , θ̃s )
s
t
Z
∇θ f (Xs , θ̃s )dWs
2p
0
2p
+K
Z
0
2p
≤K +K
Z
t
α2p
s ds + K
Z
t
0
α2p
s E θ̃s
2p
0
ds
2p
p
ds
α2p
s E θ̃s
2p
ds
ds
1 1−2p
+K
C0
− (C0 + t)1−2p + K
2p − 1
t
2
2p
α2p
∇θ f (Xs , θ̃s )
ds
s
0
Z t
2p
2pq
+K
α2p
1 + E kXs k + E θ̃s
s
+E
2p
Z
0
t
α2p
s E θ̃s
2p
ds.
ds,
where the unimportant finite constant K < ∞ changes from line to line. Hence, Gronwall lemma then
immediately gives that for any p ≥ 1 there exists a finite constant K < ∞ such that
sup E θ̃t
t>0
27
2p
≤ K.
(A.2)
Combining (A.1) and (A.2) we then obtain that for any p ≥ 1 and for an appropriate finite constant K,
2p
sup E kθt k
t>0
≤ K,
completing the proof of the targeted bound.
A.2
Poisson PDE
We recall the following regularity result from [13] on the Poisson equations in the whole space, appropriately
stated to cover our case of interest.
Theorem A.1. Let Conditions 2.1, 2.2 and 2.6 be satisfied. Assume that H(x, θ) ∈ C α,2 X , Rk ,
Z
H(x, θ)π(dx) = 0,
(A.3)
X
and that for some positive constants K, p1 , p2 , p3 and q,
kH(x, θ)k
∂H
(x, θ)
∂θ
∂2H
(x, θ)
∂θ2
p
q
p
q
p
q
≤
K(1 + kθk 1 )(1 + kxk ),
≤
K(1 + kθk 2 )(1 + kxk ),
≤
K(1 + kθk 3 )(1 + kxk ).
Let Lx be the infinitesimal generator for the X process. Then the Poisson equation
Z
u(x, θ)π(dx) = 0
Lx u(x, θ) = H(x, θ),
(A.4)
(A.5)
X
has a unique solution that satisfies u(x, ·) ∈ C 2 for every x ∈ X , ∂θ2 u ∈ C (X × Rn ) and there exist positive
constants K ′ and m such that
ku(x, θ)k + k∇x u(x, θ)k ≤ K ′ (1 + kθkp1 )(1 + kxkm ),
∂2u
∂u
p
m
(x, θ) +
(x, θ) ≤ K ′ (1 + kθk 2 )(1 + kxk ),
∂θ
∂x∂θ
∂2u
∂3u
p
m
(x,
θ)
+
(x, θ) ≤ K ′ (1 + kθk 3 )(1 + kxk ).
∂θ2
∂x∂θ2
(A.6)
Terms of the form H(x, θ) = h(x, θ) − h̄(θ) must be analyzed and controlled throughout the paper. Note
that H(x, θ) satisfies the centering condition (A.3). For example, the term G(x, θ) = ∇θ g(x, θ) − ∇θ ḡ(θ)
needs to be controlled. G(x, θ) satisfies the centering condition (A.3) and Condition 2.6 implies the bounds
(A.4) with appropriate choices for p1 , p2 , p3 and q. Therefore, the Poisson solution (A.5) associated with
H(x, θ) = G(x, θ) satisfies the bounds (A.6).
B
Proof of Convergence with Linear Growth in f (x, θ)
Under the global Lipschitz assumption on ∇ḡ(θ) and the uniform bounds on the moments of θ derived in
(1)
Appendix A.1 the proof is exactly the same as in [3], except for the term Jt of Lemma 3.1, in Section 3 of
[3] which we re-define below and prove converges almost surely to 0.
(1)
Jt
= αt kv(Xt , θt )k .
Using the bounds (A.6) and results from [12], there is some 0 < K < ∞ (that may change from line to
line below) and 0 < q < ∞ such that for t large enough
(1)
E|Jt |2 ≤ Kα2t E 1 + kXt kq + kθt k2 ≤ Kα2t .
28
o
n
(1)
Consider p > 0 such that limt→∞ α2t t2p = 0 and for any δ ∈ (0, p) define the event At,δ = Jt ≥ tδ−p .
Then we have for t large enough such that α2t t2p ≤ 1
(1)
P (At,δ ) ≤
E|Jt |2
1
α2 t2p
≤ K t2δ ≤ K 2δ .
2(δ−p)
t
t
t
The latter implies that
X
n∈N
P (A2n ,δ ) < ∞.
Therefore, by Borel-Cantelli lemma we have that for every δ ∈ (0, p) there is a finite positive random
variable d(ω) and some n0 < ∞ such that for every n ≥ n0 one has
(1)
J2n ≤
d(ω)
.
2n(p−δ)
Thus for t ∈ [2n , 2n+1 ) and n ≥ n0 one has for some finite constant K < ∞
(1)
Jt
≤ Kα2n+1
sup
s∈(0,2n+1 ]
kv(Xs , θs )k ≤ K
d(ω)
2(n+1)(p−δ)
≤K
d(ω)
.
tp−δ
The latter display then guarantees that for t ≥ 2n0 we have with probability one
(1)
Jt
≤K
d(ω)
→ 0, as t → ∞.
tp−δ
References
[1] Y. Ait-Sahalia, Maximum Likelihood Estimation of Discretely Sampled Diffusions: A Closed-form Approximation Approach, Econometrica, Vol. 70, No. 1, (2002), pp. 223-262.
[2] Y. Ait-Sahalia, Closed-form likelihood expansions for multivariate diffusions, Annals of Statistics, Vol.
36, No. 2, (2008), pp. 906-937.
[3] Justin Sirignano and Konstantinos Spiliopoulos, Stochastic Gradient Descent in Continuous Time, SIAM
Journal on Financial Mathematics, 2017, to appear.
[4] M. Raginsky and J. Bouvrie, Continuous-time stochastic mirror descent on a network: variance reduction, consensus, convergence, IEEE Conference on Decision and Control, 2012.
[5] N. Ikeda and S. Watanabe, A comparison theorem for solutions of stochastic differential equations and
its applications, Osaka J. Math., Vol. 14, (1977), pp. 619-633.
[6] A. Yu. Veretennikov, On polynomial mixing bounds for stochastic differential equations, Stochastic
Processes and their Applications Vol. 70, Issue 1, (1997), pp. 115-127.
[7] I. Basawa and B. Rao, Asymptotic inference for stochastic processes, Stochastic Processes and their
Applications, Vol.10, No. 3, (1980), pp. 221-254.
[8] A. Benveniste, M. Metivier, and P. Priouret, Adaptive Algorithms and Stochastic Approximations.
Springer-Verlag, 2012.
[9] J. P. N. Bishwal, Parameter estimation in stochastic differential equations, in: Lecture Notes in Mathematics, Vol. 1923, Springer Science & Business Media, 2008.
[10] O. Elerian, S. Chib, and N. Shephard, Likelihood inference for discretely observed nonlinear diffusions,
Econometrica, Vol. 69, No. 4, (2001), pp. 959-993.
[11] B. L. S. P. Rao, Statistical inference for diffusion type processes, Arnold, 1999.
29
[12] E. Pardoux and A.Yu. Veretennikov, On Poisson equation and diffusion approximation 1, Annals of
Probability, Vol. 29, No. 3, (2001), pp. 1061-1085.
[13] E. Pardoux and A. Y. Veretennikov, On Poisson equation and diffusion approximation 2, The Annals
of Probability, Vol. 31, No. 3, (2003), 1166-1192.
[14] C. Chicone, Ordinary differential equations with applications, second edition, Springer, New York, 2006.
[15] Y. Kutoyants, Statistical Inference for Ergodic Diffusion Processes. Springer, 2004.
30
| 10math.ST
|
SkyQuery: A Web Service Approach to Federate Databases
Tanu Malik
Alex S. Szalay
Tamas Budavari
Ani R. Thakar
Johns Hopkins University
Baltimore
USA
[email protected], [email protected], [email protected], [email protected]
Abs tract
Traditional science searched for new objects and
phenomena that led to discoveries. Tomorrow's
science will combine together the large pool of
information in scientific archives and make
discoveries. Scientists are currently keen to
federate together the existing scientific
databases. The major challenge in building a
federation
of
these
autonomous
and
heterogeneous databases is system integration.
Ineffective integration will result in defunct
federations and under utilized scientific data.
Astronomy, in particular, has many
autonomous archives spread over the Internet. It
is now seeking to federate these, with minimal
effort, into a Virtual Observatory that will solve
complex distributed computing tasks such as
answering federated spatial join queries.
In this paper, we present SkyQuery, a
successful prototype of an evolving federation of
astronomy archives. It interoperates using the
emerging Web services standard. We describe
the SkyQuery architecture and show how it
efficiently evaluates a probabilistic federated
spatial join query.
1. Introduction
Autonomous astronomy archives operating over the
Internet feel an increasing need to federate with each
other. A federation will enable their users to combine data
in different archives for the same astronomical object and
Permission to copy without fee all or part of this material is granted
provided that the copies are not made or distributed for direct
commercial advantage, the VLDB copyright notice and the title of the
publication and its date appear, and notice is given that copying is by
permission of the Very Large Data Base Endowment. To copy
otherwise, or to republish, requires a fee and/or special permission from
the Endowment
Proceedings of the 28th VLDB Conference,
Hong Kong, China, 2002
thus examine each object in more detail. Similar needs to
federate are also felt in pharmaceutical [UFI01], medical
[Daw00], geographical [Arn98], and other scientific
archives.
A federated database is a collection of cooperating but
autonomous component archives behaving like a single
integrated database [She90]. The challenge in federation
arises from the heterogeneity of the autonomous archives.
Each archive makes independent choices in its hardware,
software, data schema, etc., and invests its resources
accordingly. As a result, differences arise among archives
in the types of network systems, operating systems,
database systems,
programming platforms, and
algorithmic techniques used.
An archive is naturally reluctant to forsake its
autonomy --- even if it isn't, complete restructuring can be
prohibitively expensive. Federated database architecture
is designed to maintain autonomy and yet accomplish
federated tasks. This implies that the interfaces of the
architecture must conform to some common standard fo r
interoperability. Such a standard specifies, primarily, a
communication protocol and a data exchange format. To
be functional, a federation on the Internet needs a
communication protocol that is universally acceptable and
a data exchange format that is simple and extensible.
In this paper, we present SkyQuery [Sky02], a
successful prototype of one of the first proposed
federations of astronomy archives. It supports a complex
federated query called the cross match query, which is a
spatial join query that matches objects between archives,
if they correspond to the same astronomical body.
SkyQuery is implemented using the emerging middleware
standard called Web services [Sim01, XML02]. Its
success in supporting cross matches has implications on
the feasibility of adopting Web Services as a standard for
a full-fledged federation of astronomy archives that also
supports a variety of other federated queries and
transactions.
The rest of the paper is organized as follows. In
Section 2, we give an overview of astronomy archives on
the Internet, mentioning characteristics of astronomical
data and the reasons behind the need to federate. A lot of
effort, including from the geographical information
systems community, has been directed towards standards
for interoperability [Ama96, Arn98]. Some of the
standards that have taken advantage from this community
are CORBA [Obj02], DCOM [Dis98], and Java RMI
[Jav98]. In Section 3, briefly look at the interoperability
issues relevant to Internet federations. We also describe
Web Services and point out its advantages and
disadvantages. Spatial query execution in distributed and
federated environments has recently become a focus of
research. In Section 4, we give an overview of the
existing work in this area. Section 5 is the heart of the
paper. We first describe in detail the architecture of
SkyQuery. We then define a cross match query. Next, we
present an original algorithm designed to efficiently solve
this query. Finally, we present the steps followed by the
various components of SkyQuery in the execution of this
algorithm. In Section 6 we make concluding remarks.
In this paper, we use the terms “archive” and
“database” interchangeably. Also, when we say, “user
query”, we mean, the cross match query specified by the
user.
2. Astronomy and the DataAvalanche: the
Need for Federation
Astronomy is a particularly good domain for building
federations and evaluating techniques for federated spatial
queries [APS99]. Astronomers have conducted more than
a 100 independent surveys [Slo02, NAS02], each
mapping the sky in different wavelengths, but related to
each other by the unique position of each astronomical
body. Typically, such surveys cover 10 – 100 million
objects, with some of the larger ones expected to cover a
billion objects. The archives resulting from these surveys
are autonomous and heterogeneous. In spite of this,
building a federation of these archives has been a goal of
astronomers for a long time as the users will then be able
to run federated spatial join queries. Through these
queries, they would observe the same sky in other
wavelengths (using someone else’s data) and combine the
available observations into a multi-spectral data set. This
would immensely aid in making discoveries faster and
easier [Sch00].
Astronomical data sets are large and contain a widevariety of data types —from raw pixels to fully derived
tables of scientific measurements [Sch00]. In addition this
data is freely available. By studying and providing for the
needs of the world’s 10,000 strong astronomy
community, one can gain valuable insight in developing
federated systems – which can then be scaled to support a
much wider audience.
3. Choice of a Standard for Interoperability
The components of a federated database need to agree
upon a flexible communication standard for supporting
message passing and for the RPC semantics used between
them. To achieve this flexibility, we can either use
proprietary standards or existing distributed object
frameworks like DCOM, CORBA, Java RMI, etc., which
have provided excellent solutions for distributed and
multi-databases [Ozs99, Ama96, Arn98]. Proprietary
solutions provide limited flexibility, as they are usually
tailored to satisfy the needs of a single enterprise where
the kind of heterogeneity is not the same as witnessed
over the Internet. Similarly, distributed object frameworks
for multi-databases are poorly suited to federated database
systems in the following ways:
1. Network protocol dependence: These standards
are integrated closely with a communication
protocol. For example, CORBA 2.0 compliance
requires vendors to support the TCP/IP protocol
suite.
2. Operating system dependence: Most of the
existing standards have platform preferences.
DCOM is tightly integrated with Windows while
CORBA with UNIX-based platforms. For
DCOM, porting the solution can be arduous;
while in CORBA, it is not guaranteed that ORBs
[Obj02] from different vendors are interoperable.
3. Semantic dependence: Standards like DCOM,
and Java RMI are based on the object -oriented
paradigm. Implementing them is easy only in an
object-oriented language.
4. Unwieldy: These standards are reliable, secure,
and scalable, but at the same time, overloaded
with functionality. This implies that solutions
based on them are complex and difficult to
maintain, and so unsuitable for applications not
making use of the extra functionality they
provide.
In the following subsection, we give a brief
description of the emerging Web services standard.
3.1 Web Services
A distributed computing model consists of a message
exchange model, a communication protocol, and
mechanisms for describing, defining, and discovering
services. Web Services use Internet -based applicationlevel protocols like Simple Mail Transfer Protocol
(SMTP) [Smt01] and Hypertext Transfer Protocol
(HTTP)
[Hyp99]
for
communication
between
applications. Simple Object Access Protocol (SOAP)
[Sim01] is the message-exchange protocol; Web Services
Description Language (WSDL) [Web01] is the standard
for describing and defining, services; and Universal
Description, Discovery and Integration (UDDI) [Uni02] is
the standard for discovering services.
SOAP is a simple and lightweight mechanism for
exchanging structured and typed information between
peers in a decentralized, distributed environment. SOAP
supports both one-way and request-response message
exchange paradigms. Further, it is a flexible, easy-t odeploy, loosely coupled, access-protocol-independent,
open standard protocol. It is highly flexible as it uses
XML [Ext96] as the data exchange format. The syntaxindependence of XML allows communicating entities to
specify the syntax of the data (meta-data) along with the
actual payload, making it highly flexible for an evolving
payload. SOAP message processing entities are
lightweight, as it doesn’t define bulky protocol-dependent
headers. HTTP messages containing SOAP need to
specify only one extra field "Soap Action", below custom
HTTP headers, that identifies the SOAP entity on the
destination machine. SOAP is easy-t o-deploy as it
currently uses XML and HTTP, whi ch are ubiquitous
easy-t o-understand de facto Internet standards. Almost all
machines support the HTTP protocol, and XML parsers
are freely available, and being text based, very easy to
understand and debug. SOAP is loosely coupled, as, due
to the flexibility of XML, it doesn’t rely on any
programming model or application semantics. SOAP is
also access-protocol-independent.WSDL consists of two
distinct parts - service definition and service
implementation. Service definition is an XML-style
description of what the service intends to provide, i.e.,
names of messages and their parameters, type of
messages, etc. Service implementation specifies binding
to a particular protocol or data type, i.e., syntax of the
messages exchanged, protocols used to transfer messages,
etc. By dissociating service definition from its
implementation, WSDL allows re-use of the service
description interface by clients that might be using other
programming models to implement the service. Currently
WSDL supports only SOAP and HTTP protocols for
message communication.
Services need a unique service for discovering other
services that may be available on the Internet. This
service should provide a common repository where
services can register themselves and be discovered. UDDI
is the standard architecture for building such repositories.
To summarize, the simplicity and flexibility of XML
and the ubiquity of HTTP make Web Services an
attractive solution for federating archives.
4. Related Work
Federated database systems have been extensively
researched
[Lit90].
Most studies, however, have
concentrated on schema integration, cost estimation for
sub-query execution at individual archives, construction
of query execution plans, transaction management, and
other issues that are not the primary focus of this paper
and were only partially considered while building
SkyQuery.
Spatial joins, which are much more expensive than
regular relational joins, have also received a lot of
research attention [Arg98, Bri93]. But most schemes rely
on the existence of efficient spatial indices to perform
spatial joins. Abel et al. [Abe95] formalise a spatial semijoin operator, which is a generalisation of the
conventional semi-join operator, and present efficient
algorithms for performing distributed spatial joins. They
present two algorithms based, again, on specialized spatial
indices: (1) single dimensional object mappings,
commonly known as space filling curves, for point data,
and (2) the R tree for region data.
These indices are rarely available in autonomous archives.
Furthermore, since their objective is to compute a
distributed join over a LAN, they assume that processing
costs are relatively “more” than transmission costs. This
is not true for federated spatial joins over the Internet, and
so their algorithms aren’t well suited to our purpose.
5. SkyQuery: Architecture and
Implementation
SkyQuery is a prototype of an evolving federation of
geographically separate astronomy databases on the
Internet. It has a web interface that allows astronomers to
query spatial data stored in the databases participating in
the federation. In particular, SkyQuery evaluates a
probabilistic federated spatial join query called the cross
match query. In this section, we describe the algorithms
and the architecture used in SkyQuery. We begin with a
description of the architectural components and their
interaction using Web services. Next, we explain the cross
match query, and finally, we show how this framework
for federation can be used for efficiently evaluating such
queries.
5.1 Architecture
The SkyQuery architecture is based on the wrappermediator architecture (Figure 1), which is common in
federated database systems. It consists of three
components (1) the Clients, (2) the Portal and (3) the
SkyNodes. The Portal is the mediator and the SkyNodes
contain the wrappers around databases. These three
components coordinate with each other to perform two
tasks: allowing databases to join the federation, and
evaluating cross match queries.
The Clients are web interfaces (or similar
applications) that accept user queries and pass them on to
the Portal. The Portal mediates between the Clients and
the SkyNodes. It accepts user queries from the Clients and
and coordinates with the SkyNodes to evaluate the
queries. Primarily, SkyNodes are the heterogeneous,
autonomous databases participating in the federation.
Each SkyNode also implements services that act as
Web services
Interface
Web services
Interface
Databasespecific API
Web
Meta-data
services
Information
Portal
Astronomy
Archive
Sky Node
Query
Registration
Cross-Match
Meta-data
archive
SkyNode
SkyQuery
Meta-data
Web services
Interface
Information
Query
More such SkyNodes are
waiting to join Portal
Cross-match
More such SkyNodes are
waiting to join Portal
Clients
Messages between clients, portal and SkyNodes are exchanged as SOAP
messages (over HTTP protocol)
Figure 1 The SkyQuery Architecture
wrappers and hide its DBMS and other platform specific
details. This presents a uniform view to the Portal.
The Portal provides two functionalities. First, it
implements a Registration service that SkyNodes can use
to join the federation. As part of registration, the Portal
catalogs meta-data-information, which is used in query
decomposition. Secondly, it receives SQL-like queries
from the Client through its SkyQuery service. It
decomposes the queries to generate performance queries
that are used for query optimization. Based on the results
of performance queries, the Portal constructs an optimized
execution plan, which is, essentially, an ordered set of
spatial queries. This order defines the routing of the
spatial query set among the participating databases. Many
federations, based on the wrapper-mediator architecture,
pull results from each database to the Portal. SkyQuery,
instead, moves the partial results of spatial queries from
one SkyNode to the next along a chain that ends at the
Portal. The Portal relays the final result back to the Client.
The SkyNode databases usually have very similar
logical schemas. A primary table stores the unique sky
position for each astronomical object. Other tables store
secondary observations like light intensities in various
wavelengths, spectrograms and spectral lines of selected
objects, etc. The implementation of the schemas,
however, varies, and is an autonomous decision of the
organization hosting the database. Apart from this,
SkyNode databases have libraries that support range
queries. Many astronomy archives support the HTM
library [Hie02]; the HTM is a hierarchical spatial index
based on spherical triangles. It helps in reducing spatial
processing at individual databases.SkyNodes have APIs,
which are database specific that provide access to the data
and meta-data. In addition, each SkyNode has a wrapper,
which implements its specific federation task. This
wrapper interfaces with the APIs, as well as the libraries
implementing the Web services standard. In essence, the
wrapper hides the heterogeneity of the APIs and provides
a uniform view to the Portal. This obviates the need for a
different module at the Portal to communicate with each
SkyNode. (Figure 1).
A SkyNode implements the following four Web
services – Information service, Meta-data service, Query
service and Cross match service. The first three Web
services are the minimum set of services needed to
provide a basic framework for the federation of
astronomy databases. The fourth service performs the
actual federated task, which, in our case, is the execution
of a cross match query. When a SkyNode wishes to join
the SkyQuery federation; it calls the Registration service
of the Portal. The registration request includes
information about services available on the SkyNode. The
Portal, in response to this request, calls the Meta-data
service at the SkyNode. The Meta-data service is
responsible for providing complete schema information to
the Portal, which the Portal catalogs. Once the Portal
successfully recognizes a SkyNode, it calls the
Information service to collect certain astronomy specific
constants of that SkyNode such as the object position
estimation errors, the name of primary table that stores the
position of objects, etc. The Query service is a generalpurpose database querying service. In our case, it is used
by the Portal to answer performance queries. The Cross
match service executes the cross match query. The next
sub-sections explain the cross match query and describe
its evaluation using these services.
In this query, SDSS: Photo_Object, TWOMASS:
Photo_Primary, and FIRST: Primary_Object are tables
from three different federated SkyNodes, namely, SDSS,
TWOMASS, and FIRST. The AREA clause limits the
query to a circular range centered at longitude 185.0 and
latitude -0.5, and with a radius of 4.5 arc seconds.
5.2 Cross Match Queries
A cross match query is a SQL-like query with special
clauses to specify spatial constraints. The spatial clauses
are (1) AREA to specify a spatial range; and (2)
XMATCH to specify a probabilistic (fuzzy) spatial join.
The AREA clause is used to specify a circular range in
the sky. All objects returned by the query have to lie
within this range.
The XMATCH clause is used to match objects in
different archives that are observations of the same
astronomical body (In this sub-section we use the term
"body" to refer to the real astronomical object, and the
term "object" to refer to an observation of such a body
stored in an archive). Each archive measures the unique
position in sky, i.e., right ascension and declination, for
the astronomical bodies it observes. This measured
position would be sufficient for exact matches between
objects in different archives, except that there is always
some unavoidable error in the measurements. The
measured position varies slightly, from archive to archive.
Astronomers assume that this measured position is a
random variable distributed normally around the real
position of the body. The standard deviation σ of the
error is circular and is known for each survey; its value
depends on the accuracy of the survey’s measuring
instruments. Using this knowledge, given a set of objects,
one from each archive, one can compute the probability of
their being a cross match, i.e., of their referring to the
same astronomical body. For a given query, the
XMATCH clause specifies the minimum probability
threshold for a set of objects to be considered a cross
match.
For example, take the following cross match query:
Figure 2 Different sets of objects are selected based on
XMATCH clause specified in user query. The figure
shows that set {aO,aT, aP} gets selected when user
specifies XMATCH(O,P,T) < 3.5 while {b O, bT,bP}
gets selected when the user specifies
The XMATCH clause specifies that only those sets of
objects, one each from the three archives, which are
within 3.5 standard deviations of their mean position,
should be considered. (There is a direct correspondence
from this specification to the specification in terms of the
minimum probability threshold for a set of objects that
was mentioned earlier). For example, in Figure 2, a O, aT,
and a P are the positions for the body a as observed by the
archives O, T, and P. Similarly b O, bT, and, b P are
observations for body b.
Assume that all these
observations satisfy the AREA clause. The mean
positions for the two sets of observations are a m and b m,
respectively. Note, that the mean positions depend on
which observations are being considered for computing
the mean, but for the sake of simplicity, we will assume
that they are always at the positions shown. The figure
also shows, for each body, the range of 3.5 standard
deviations from the mean position. All three observations
for the object a are within this range. So the set {a , aT, aP
O
SELECT
O.object_id, O.right_accession, T.object_id
FROM
SDSS: Photo_Object O,
TWOMASS: Photo_Primary T,
FIRST: Primary_Object PWHERE
AREA (185.0,-0.5,4.5) AND
XMATCH (O,T,P) <3.5 AND
O.type= GALAXY AND (O.i_flux - T.i_flux)>2.
} is a cross match. But the set for body b, {b O, bT, b P },
is not a cross match as b P is out of range.
The XMATCH clause can also specify the
absence of a match. For example, if instead, the clause
was XMATCH (O, T, !P) <3.5, it would imply selection
of only those sets of objects, one each from the first two
archives, that are within 3.5 standard deviations of their
mean position, and which don’t have a matching object in
the last archive within the same error bound. The last
archive is called a drop out. Note that a drop out
specification corresponds to an “exclusive” outer join. In
Figure 2, {aO, a T, a P} is not a cross match for this
specification, but {b O, b T} is. Archives specified in
XMATCH clause that are not drop outs are called
mandatory.
5.3 Query Execution and Optimization
Query execution (see Figure 3) in SkyQuery incurs
processing costs at the individual SkyNodes and
transmission costs in sending partial results from one
SkyNode to the next. We reduce processing costs by
using an efficient algorithm for cross match computation
at each SkyNode, and we reduce transmission costs by
first sending performance queries to estimate the
maximum size of the partial result that each SkyNode will
transmit.
There are methods that involve extensive calculations
for reducing transmission costs, but we use one that is
simple, yet effective. Our method follows the basic
approach of treating component DBMSs as black boxes,
running test queries on them, and finally estimating
transmission costs from the results. This approach has
attracted considerable attention for its simplicity [Du92,
Zhu96]. We call our method the count star approach, as
our test queries are count queries on the component
databases of the federation.
For every user submitted cross match query, the Portal
creates performance queries (Step 2 in Figure 3), one for
each of the mandatory archives contained in the
XMATCH clause. A performance query for a component
database is a SQL query constructed to find an upper
bound on the number of tuples the database will return
Figure 3 The Portal and the SkyNodes coordinate to solve a cross match query submitted by a
Client. The figure shows the order in which the sample query gets executed.
during the cross match computation. For example, the
cross match query discussed earlier will result in the
following three performance queries:
SELECT count(*) FROM SDSS: Photo_Object O
WHERE AREA(185.0,0.5,4,5) AND O.type = GALAXY.
SELECT count(*) FROM TWOMASS: Photo_Primary T
WHERE AREA(185.0,0.5,4,5)
SELECT count(*) FROM FIRST: Primary_Object P
WHERE AREA(185.0,0.5,4,5)
Note, these queries contain clauses that can be evaluated
entirely at their SkyNodes. These performance queries are
passed as asynchronous SOAP messages to the respective
Query services of each SkyNode (Step 3 in Figure 3).
This will often warm the database cache on each SkyNode
with index pages that satisfy the main cross match query,
and thus aid in reducing processing time that would be
incurred while doing the cross match.
When the messages of performance query results with the
number of potentially qualifying tuples arrive at the Portal
(Step 4 in Figure 3), it creates a federated query execution
plan (Step 5 in Figure 3). (De-serialization of these
messages is not an expensive operation as they are single
integers.) The federated query execution plan consists of a
list of ordered pairs, each containing a query and the URL
information of the SkyNode where it would be executed.
The list is in decreasing order of the count star values
returned by the performance queries, with the drop out
archives, if any, at the beginning of the list.
To begin cross match query execution, the Portal
sends a SOAP RPC to the Cross match service of the first
SkyNode on the list, and passes the query execution plan
as a parameter. This Cross match service, in turn, calls
the similar service at the second SkyNode on the list,
again with the plan as a parameter.
This daisy chaining continues until the last SkyNode on
the list, which executes the query meant for it (Step 6 in
Figure 3). The query selects rows that satisfy the AREA
and other clauses. The query result includes spatial
attributes required for computing the cross match. The
SkyNode returns this result, as a serialized XML encoded
SOAP message to its calling SkyNode. (Step 7 in Figure
3).
Now, at this (second last in the list) SkyNode the
encoded results are first de-serialized using a SOAP
translator. Next, the results are cast into a database object
characteristic to the database that the SkyNode hosts. The
Cross match service uses the database specific API at that
SkyNode to insert the values in the database object into a
temporary table. Next, a stored procedure encoding the
cross match algorithm (See Section 5.4) uses this
temporary table and the primary table at this SkyNode to
identify matching objects (with the last SkyNode) that
have a potential to satisfy the XMATCH clause. Note, the
certainty of the XMATCH clause being satisfied is only
known after all the SkyNodes have been consulted. This
procedure, in fact, computes an implicit spatial join. (The
join is akin to an outer join or an inner join based on
whether the archive is a drop out or not). After this, the
Cross match service executes its own (non-spatial) query
from the execution plan and joins - this is an explicit inner
join - with the results of the spatial join. The final results
are returned to the calling SkyNode. The temporary table
is deleted.
The process above is repeated at every SkyNode in the
reverse order of the calls. The first SkyNode returns the
results to the Portal, which relays it to the client. In this
whole process, it is clear that the order based on the count
star values will often decrease the network transmission
costs.
5.4 The Cross Match Algorithm
The AREA clause in a cross match query is implemented
using the range search capabilities of the individual
archives. We assume that each archive has a reasonably
efficient mechanism for implementing a range search.
One such mechanism, which is also used by the SkyQuery
test databases, is the HTM [Hie02]. This mechanism
builds a quad tree on the sky, each node of which
corresponds to a spherical triangle. To retrieve objects in
a given range, triangles that are entirely within or intersect
the range are first computed. All objects in the triangles
that are entirely within the range are in the range too.
Objects that are in intersecting triangles, however, are
again individually tested to check if they are within the
range or not. Objects in other triangles are not in range.
Similarly, the other clauses, except XMATCH, are
implemented using each archive’s ability to evaluate
regular queries.
The XMATCH clause is implemented by computing a
probabilistic spatial join. Let the clause contain N
archives. Assume that none of them are dropouts; the
extension to include dropouts is straightforward. We first
describe, given an N -tuple of objects R = (o1 ,...,oN), one
from each archive, how to compute the log likelihood of
each oi, 1 = ?i = ?N, being an observation of the same
astronomical body. Let the position for object o i (from
archive i) be given by (xi, yi, zi). Let (x, y, z) be the
unknown true position of the astronomical body. The log
likelihood that the observations in R are of this object is
N
1
2
χ 2(x , y, z) = −∑ 2 (x − xi ) + (y − yi )2 + (z − zi )2
i =1 2σ
i
where s i is the standard deviation of the error at archive i.
The best such position (x, y, z) is obtained by minimizing
the following chi -squared expression
[
]
1
∑ 2σ [(x − x ) + ( y − y ) + (z − z ) ]− λ[x
N
2
2
i =1
i
2
i
2
i
1
2
2
]
+ y2 + z 2 − 1
i
where the second term is the Lagrange multiplier that
forces (x, y, z) to be a unit vector. This best position is
along the direction ( a x , a y , a z ) , and the log likelihood at
this position is, − a
N
a=
+ a 2x + a 2y + a z2 , where
N x
N y
N z
1
i
i
i
,
a
=
,
a
=
,
a
=
∑ 2 x ∑ 2 y ∑ 2 z ∑ 2.
i=1 σi
i=1 σi
i=1 σi
i=1 σi
Only those N-tuples satisfy the XMATCH clause that
have a log likelihood value greater than the threshold
specified in the clause. Let the set of satisfying tuples be
S. This set is computed in a distributed fashion as each
archive is at a different site; the computation is done in N
steps, one at each archive. At archive i, the set Si
consisting of tuples of length i that “have the potential” to
satisfy the clause is constructed. Let Ri = (o1,..., oi) be
one such tuple. For each such Ri the corresponding
cumulative values are also computed. Remember that
(ai,x ,ai,y, ai,z) is the direction of the best position (xi, yi,
zi) of the astronomical object if Ri is to be a cross match.
To compute these values, the archive i receives from
archive i-1 tuples of type Ri-1 = (o1, ..., oi-1) and the
corresponding values ai -1, ai-1,x, ai-1, y, ai-1,z. For each
Ri-1, archive i retrieves all objects that are close to the
current best position (xi-1,yi-1,zi-1) (and satisfy the other
clauses in the cross match query) using a range search.
Each such object is appended to Ri -1 to get tuples of type
Ri. The cumulative values for each such Ri are then
computed using the information available. Next, the log
likelihood of Ri being a cross match is computed, and
only if it is larger than the threshold, Ri (and its
cumulative values) is sent to the next archive i+1.
The first archive just needs to send 1-tuples
comprising of objects that satisfy the other clauses in the
query to the second archive. The tuples surviving the Nth
archive are the required cross matches. This XMATCH
scheme is fully symmetric; the particular order of the
archives considered doesn’t matter. We do, however,
impose an order, using the count star approach to reduce
network data transfer costs.
6.
Conclusion
We want to highlight some of our experiences while
developing SkyQuery, talk about extensions, as well as
speculate on what it may imply for the future of
federations (of astronomy archives) on the Internet.
Since SOAP is an evolving standard, it was difficult to
find a fully standard-conforming, well-documented SOAP
implementation. The Web service implementation did not
require any significant programming overhead as, we feel,
would have been required with an alternative technology
like CORBA. We faced problems while trying to match
large number of objects, though. The XML parser at the
SkyNode would run out of memory while parsing SOAP
messages of about 10 MB. We worked around by
dividing large data sets into smaller chunks. SOAP is
considered to be slower than other middleware, like,
CORBA, because of the time spent for serialization and
de-serialization [XML02].
XML and SOAP are technologies that are still
evolving and probably quite far from realizing their peak
in terms of the service and the convenience they provide.
Yet, we found that they were more than sufficient for
building SkyQuery. This makes us want to extend
SkyQuery, using XML and SOAP, to perform other
complex distributed computing tasks for the Virtual
Observatory. We would also like to include a lot more
participating archives. That would effectively test the
interoperability and the system integration capabilities of
Web services. Another extension is to implement
transaction processing for exchange of data between
astronomy archives, and see how the stateless SOAP
handles such complex requirements.
We are still building SkyQuery and extending its
functionality. We want to evaluate spatial queries more
complex than cross matches. For example, queries on
spatial objects that are not just points but objects of
different shapes like polygons. They AREA clause can
also be extended to specify arbitrary polygons rather than
just simple circles. This would imply many more
evaluations of spatial joins and an exchange of a larger
magnitude of XML data.
SkyQuery shows that Web services are a good
medium for publishing data on the Internet.
It
demonstrates that such archives can be federated with
minimal additional software effort and compliance
requirements.
7.
Acknowledgements
The authors thank Jim Gray for inspiring them to write
this paper in the first place as well as for the helpful
suggestions and discussions at all stages of this project.
References
[Abe95] D.J. Abel, B.C. Ooi, K.L. Tan, R. Power, and
J.X. Yu. Spatial Join Strategies in Distributed Spatial
Dbms. Proc. Fourth Int'l Symp. Large Spatial Databases,
pp. 348--367, Aug. 1995.
[Ama96] Bernd Amann. Integrating GIS Components
with Mediators and CORBA. http://citeseer.nj.nec.com/
amann96integrating.html, 1996.
[APS99] APS Federated Database. http://spihr.spa.umn.
edu/FedBase, 1999.
[Arn98] H.-Arno Jacobsen, Agnès Voisard. CORBABased Interoperable Geographic Information Systems.
European Conference on Parallel and Distributed
Systems, 1998.
[Arg98] Lars Arge, Octavian Procopiuc, Sridhar
Ramaswamy, Torsten Suel, Jeffrey Scott Vitter. Scalable
Sweeping-Based Spatial Join. Proc. 24th Int. Conf. Very
Large Data Bases ( VLDB), 1998.
[Bri93] T. Brinkhoff, H.-P. Kriegel, B. Seeger. Efficient
Processing of Spatial Joins Using R-trees. Proceedings of
ACM SIGMOD Conf., 1993.
[Daw00] Steven Dawson, Shelly Qian, and Pierangela
Samarati. Providing Security and Interoperation of
Heterogeneous Systems. Distributed and Parallel
Databases, 8(1):119--145, 2000.
[Dis98] Distributed Component Object Model.
http://www.microsoft.com/com/tech/DCOM.asp, 1998.
[Du92] W. Du, R. Krishnamurthy, M. C. Shan. Query
Optimization in a Heterogeneous DBMS. Proc. 18 th Intl.
Conf. on VLDB, pp 277-291, 1992.
[Ext96] Extensible Markup Language
(XML)
http://www.w3.org/TR/WD-xml-961114.html, 1996.
[Hei85] D. Heimbigner and D. McLeod. A Federated
Architecture for Information Management., ACM
Transactions on Office Information Systems, 3(3): 253 -278, Jul. 1985.
[Hie02] Hierarchical Triangular
sdss.jhu.edu/htm, 2002.
Mesh.
[Sch00] Schade, D., Dowler, P., Zingle, R., Durand, D.,
Gaudet, S., Hill, N., Jaeger, S., & Bohlender, D. 2000, in
ASP Conf. Ser., Vol. 216, Astronomical Data Analysis
Software and Systems IX, Eds. N. Manset, C. Veillet, D.
Crabtree (San Francisco: ASP), 215.
[She90] A. Sheth and J. Larson. Federated database
systems for managing distributed, heterogeneous, and
autonomous databases. ACM Trans. Database System,
22(3), 1990.
[Sim01] Simple Object Access Protocol (SOAP 1.1).
http://www.w3.org/TR/SOAP/, 200.
[Sky02] SkyQuery http://www.skyquery.net, 2002.
[Slo02] Sloan Digital Sky Survey. http://www.sdss. org,
2002.
[20 Smt01] RFC 2821. Simple Mail Transfer Protocol
(SMTP). http://www.ietf.org/rfc/rfc2821.txt, 2001.
http://www.
[Hyp99] RFC 2616. Hypertext Transfer Protocol -HTTP/1.1. http://www.ietf.org/rfc/rfc2616.txt, 1999.
[UFI01] UF Interdisciplinary Bioinformatics Initiative
Clinical Data Subgroup Summary, http://www.biotech.
ufl.edu/BioInformatics/ClinicalDataSummary.htm, 2001.
[18 Uni98] RFC 2396. Uniform Resource Identifiers
(URI):
Generic
Syntax.
http://www.ietf.org/rfc/
rfc2396.txt, 1998.
[Jav98] Java Remote Method Invocation.
http://java.sun.com/, 1998.
[17 Kia00] Kian-Lee Tan, Beng Chin Ooi, David. J. Abel.
Exploiting Spatial Indexes for Semijoin-based Join
Processing in Distributed Spatial Databases. IEEE Trans.
and Data Engineering, 12(6), 2000.
[Lit90] Witold Litwin, Leo Mark, and Nick
Roussopoulos. Interoperability of multiple autonomous
databases. ACM Computing Surveys (CSUR), 22(3): 267 –
293, September 1990.
[Mul96] RFC 2045 Multipurpose Internet Mail
Extensions (MIME) Part One: Format of Internet
Message Bodies.http://www.ietf.org/rfc/rfc2045.txt, 1996.
[NAS02]
NASA
Astronomical
http://adc.gsfc. nasa.gov/, 2002.
[Ozs99] M. T. Ozsu and P. Valduriez. Principles of
Distributed Database Systems. Prentice Hall, Inc., second
edition, 1999.
Data
Center.
[Obj02] Object Management Group. http://www.omg.org,
2002
[Uni02] Universal Description, Discovery and Integration
of Web services. http://www.uddi.org, 2002.
[ Web01] Web Services Description Language (WSDL).
http://www.w3.org/TR/wsdl, 2001
[ XML02] XML as a basis for remote procedure calls
http://www.webservices.org/fullpaper.php, 2002.
[Zhu96] Q. Zhu and P.A. Larson. Solving Local Cost
Estimation Problem for Global Query Optimization in
Multidatabase Systems. Proc. 9th Intl. Conf. on Parallel
and Distributed Computing Systems (PDCS), Sep 1996,
pp. 640-647.
| 5cs.CE
|
Modular Deep Q Networks for Sim-to-real Transfer
of Visuo-motor Policies
1
Introduction
The advent of large datasets and sophisticated machine
learning models, commonly referred to as deep learning, has in recent years created a trend away from handcrafted solutions towards more data-driven ones. Learning techniques have shown significant improvements in
robustness and performance [Krizhevsky et al., 2012],
particularly in the computer vision field. Traditionally robotic reaching approaches have been based on
crafted controllers that combine (heuristic) motion planners with the use of hand-crafted features to localize the
target visually. Recently learning approaches to tackle
this problem have been presented [Zhang et al., 2015;
Levine et al., 2016b; Bateux et al., 2017; Katyal et al.,
B
64 filters
64 filters
Conv1
Conv2
Conv3
Θ
I
Perception Module
fully conn.
9 units
fully conn. + ReLU
300 units
fully conn. + ReLU
LqBN
Task Loss (Lq)
64 filters
5 units
Or
fully conn.
84×84
400 units
Perception Loss (Lp)
C
stride 1
3×3 conv + ReLU
While deep learning has had significant successes in computer vision thanks to the abundance of visual data, collecting sufficiently large
real-world datasets for robot learning can be
costly. To increase the practicality of these
techniques on real robots, we propose a modular deep reinforcement learning method capable
of transferring models trained in simulation to a
real-world robotic task. We introduce a bottleneck between perception and control, enabling
the networks to be trained independently, but
then merged and fine-tuned in an end-to-end
manner to further improve hand-eye coordination. On a canonical, planar visually-guided
robot reaching task a fine-tuned accuracy of
1.6 pixels is achieved, a significant improvement over naive transfer (17.5 pixels), showing
the potential for more complicated and broader
applications. Our method provides a technique
for more efficient learning and transfer of visuomotor policies for real robotic systems without relying entirely on large real-world robot
datasets.
A
stride 2
4×4 conv + ReLU
Abstract
stride 2
7×7 conv + ReLU
arXiv:1610.06781v4 [cs.RO] 19 Dec 2017
Fangyi Zhang, Jürgen Leitner, Michael Milford, Peter Corke
Australian Centre for Robotic Vision (ACRV), Queensland University of Technology (QUT)
Brisbane, Australia
[email protected]
Q-values
BN
FC_c1 FC_c2 FC_c3
Bottleneck
Control Module
Figure 1: We present a technique for efficient learning
and transfer of visuo-motor policies for a planar reaching
task from simulated (A) to real environments (B) using
a modular deep Q network (C).
2017]. However a consistent issue faced by most approaches is the reliance on large amounts of data to
train these models. For example, Google researchers addressed this problem by developing an "arm farm" with
6 to 14 robots collecting data in parallel [Levine et al.,
2016b]. Generalization forms another challenge: many
current systems are brittle when learned models are applied to robotic configurations that differ from those used
in training. This leads to the question: is there a better
way to learn and transfer visuo-motor policies on robots
for tasks such as reaching?
Various approaches have been proposed to address
these problems in a robot learning context: (i) the use of
simulators, simulated and synthetic data [Bateux et al.,
2017; D’Innocente et al., 2017; Tobin et al., 2017; James
et al., 2017]; (ii) methods that transfer the learned
models to real-world scenarios [Fitzgerald et al., 2015;
Tzeng et al., 2016]; (iii) directly learning real-world tasks
by collecting large amounts of data [Levine et al., 2016b;
Pinto and Gupta, 2016].
In this paper, we present a method that connects these
three, usually separately considered, approaches. Vision
and kinematics data is gathered in simulations (cheap)
to decrease the amount of real world collection necessary
(costly). The approach is capable of transferring the
learned models to real-world scenarios with a fraction
of the real-world data typically required for direct realworld learning approaches.
In particular, we propose a modular deep reinforcement learning approach – inspired by DQN for Atari
game playing [Mnih et al., 2015] – to efficiently learn and
transfer visuo-motor policies from simulated to real environments, and benchmark with a visually-guided planar
reaching task for a robotic arm (Figure 1). By introducing a modular approach, the perception skill and the
controller can be transferred individually to a robotic
platform, while retaining the ability to fine-tune them in
an end-to-end fashion to further improve hand-eye coordination on a real robot (in this research a Baxter).
2
Related Work
Data-driven learning approaches have become popular in
computer vision and are starting to replace hand-crafted
solutions also in robotic applications. Especially robotic
vision tasks – robotic tasks based directly on real image data – such as, navigation [Tai et al., 2016], object grasping and manipulation [Levine et al., 2016b;
Pinto and Gupta, 2016; Lenz et al., 2015] have seen
increased interest. The lack of large-scale real-world
datasets, which are expensive, slow to acquire and limit
the general applicability of the approach, has so far limited the broader application. Collecting the datasets
required for deep learning has been sped up by using
many robots operating in parallel [Levine et al., 2016b].
With over 800,000 grasp attempts recorded, a deep network was trained to predict the success probability of
a sequence of motions aiming at grasping on a 7 DoF
robotic manipulator with a 2-finger gripper. Combined
with a simple derivative-free optimization algorithm the
grasping system achieved a success rate of 80%. Another example of dataset collection for grasping is the
approach to self-supervised grasping learning in the real
world where force sensors were used to autonomously label samples [Pinto and Gupta, 2016]. After training with
50,000 real-world trials using a staged leaning method,
a deep convolutional neural network (CNN) achieved a
grasping success rate around 70%. These are impressive results but achieved at high cost in terms of dollars,
space and time.
DeepMind showed that a deep reinforcement learning
system is able to directly synthesize control actions for
computer games from vision data [Mnih et al., 2015].
While this result is an important and exciting breakthrough it does not transfer directly to real robots with
real cameras observing real scenes [Zhang et al., 2015].
In fact very modest image distortions in the simulation environment (small translations, Gaussian noise and
scaling of the RGB color channels) caused the performance of the system to fall dramatically. Introducing a real camera observing the game screen was even
worse [Tow et al., 2016].
There has been increasing interest to create robust
visuo-motor policies for robotic applications, especially
in reaching and grasping. Levine et al. introduced a
CNN-based policy representation architecture with an
added guided policy search (GPS) to learn visuo-motor
policies (from joint angles and camera images to joint
torques) [Levine et al., 2016a], which allow to reduce the
number of real world training by providing an oracle (or
expert’s initial condition to start learning). Impressive
results were achieved in complex tasks, such as hanging a
coat hanger, inserting a block into a toy, and tightening
a bottle cap. Recently it has been proposed to simulate
depth images to learn and then transfer grasping skills
to real-world robotic arms [Viereck et al., 2017], yet no
adaptation in the real-world has been performed.
Transfer learning attempts to develop methods to
transfer knowledge between different tasks [Taylor and
Stone, 2009; Pan and Yang, 2010]. To reduce the amount
of data collected in the real world (expensive), transferring skills from simulation to the real world is an attractive alternative. Progressive neural networks are leveraged to improve transfer and avoid catastrophic forgetting when learning complex sequences of tasks [Rusu
et al., 2017]. Their effectiveness has been validated
on reinforcement learning tasks, such as Atari and 3D
maze game playing. Modular reinforcement learning approaches have shown skill transfer capabilities in simulation [Devin et al., 2017]. However, methods for realworld robotic applications are still scarce and require
manually designed mapping information, e.g. similaritybased approach to skill transfer for robots [Fitzgerald et
al., 2015]. To reduce the number of real-world images required, a method of adapting visual representations from
simulated to real environments was proposed, achieving
a success rate of 79.2% in a “hook loop” task, with 10
times less real-world images [Tzeng et al., 2016].
3
Methodology
Reinforcement learning [Sutton and Barto, 1998a] has
been proposed for agents to learn novel behaviours. One
approach for learning from rewards is Q-learning [Sutton
and Barto, 1998b], which aims to obtain a policy π that
maximizes the expectation of accumulated rewards by
approximating an optimal Q-value function
∞
hX
i
Q∗ (s, a) = max E
γ i rt+i |st = s, at = a, π ,
π
(1)
i=0
Modular Deep Q Networks
Our preliminary studies of deep visuo-motor policies indicate that the convolutional layers focus on perception, i.e., extracting useful information from visual inputs, while the fully connected (FC) layers perform
control [Zhang et al., 2017b]. To make the learning and transfer of perception and control more efficient, we propose to separate the DQN into perception and control modules connected by a bottleneck
layer (Figure 1C). The bottleneck forces the network to
learn a low-dimensional representation, not unlike Autoencoders [Hinton and Salakhutdinov, 2006]. The difference is that we explicitly equate the bottleneck layer
with the minimal scene configuration Θ whose meaning will be further introduced in Section 4. The values
in Θ are normalized to the interval [0, 1].
With the bottleneck, the perception module learns
how to estimate the scene configuration Θ̂ from a rawpixel image I; the control module learns to approximate
the optimal Q-value function as defined in Eq. 1, determining the most appropriate action a∗ given the scene
configuration Θ, i.e., a∗ = max Q(Θ, a).
a
To further improve the performance of a combined
network (perception + control), a weighted end-to-end
fine-tuning method is proposed, since experimental results show that a naive end-to-end fine-tuning using a
straight-forward loss function does not work well for performance improvement (Section 5.3).
3.2
m
Lp =
where rt is the reward at each time step t, when following a behaviour policy π = P (a|s) that determines
which action a to take in each state s. γ is a discount
factor applied to future rewards. A deep neural network
was introduced to approximate the Q-value function,
named Deep Q Network (DQN) [Mnih et al., 2015]. The
state can therefore be represented by a high-dimensional
raw-pixel image, since latent state features can be extracted by the convolutional layers [Krizhevsky et al.,
2012]. However, learned visuo-motor policies with highlevel (raw pixel) input do not transfer directly from simulated to real robots [Zhang et al., 2015].
3.1
with the quadratic loss function
Training Method
Perception
The perception network is trained using supervised
learning – first conducted in simulation, then fine-tuned
with a small number of real samples for skill transfer –
1 X
y(I j ) − Θj
2m j=1
2
,
(2)
where y(I j ) is the prediction of Θj for I j ; m is the number of samples.
Control
The control network is trained using Q-learning, where
weights are updated using the Bellman equation which
is equivalent to the loss function
m
1 X
Q(Θjt , ajt ) − (rtj + γ max
Q(Θjt+1 , ajt+1 ))
Lq =
2m j=1
ajt+1
2
(3)
where Q(Θjt , ajt ) is the Q-value function; γ is the discount factor applied to future rewards.
End-to-end fine-tuning using weighted losses
Aiming for a better hand-eye coordination, an end-toend fine-tuning is conducted for a combined network
(perception + control) after their separate training, using weighted task (Lq ) and perception (Lp ) losses. Here
for end-to-end fine-tuning, Θj is replaced with I j in
Lq (Eq. 3). The control network is updated using only
Lq , while the perception network is updated using the
weighted loss
L = βLp + (1 − β)LBN
q ,
(4)
where LBN
is a pseudo-loss which reflects the loss of Lq
q
in the bottleneck (BN); β ∈ [0, 1] is a balancing weight.
From the backpropagation algorithm [LeCun, 1988], we
, where δL is the
can infer that δL = βδLp + (1 − β)δLBN
q
are
the gradients
gradients resulted by L; δLp and δLBN
q
(equivalent
to
resulting respectively from Lp and LBN
q
that resulting from Lq in the perception module).
4
Benchmark: Robotic Reaching
We use the canonical planar reaching task in [Zhang et
al., 2015] as a benchmark to evaluate the feasibility of
the modular DQN and its training method. The task is
defined as controlling a robot arm so that its end-effector
position x in operational space moves to the position of
a target x∗ ∈ Rm . The robot’s joint configuration is
represented by its joint angles q ∈ Rn . The two spaces
are related by the forward kinematics, i.e., x = K(q).
The reaching controller adjusts the robot configuration
to minimize the error between the robot’s current and
target position, i.e., kx − x∗ k. In this task, we use the
target position x∗ and arm configuration q to represent
the scene configuration Θ. The physical meaning of Θ
guarantees the convenience of collecting labelled training
,
configuration Θ, its outputs are the Q-value estimates
for each of the 9 possible actions.
Networks with a first convolutional layer initialized
with weights from pre-trained GoogLeNet [Szegedy et
al., 2015] (on ImageNet data [Deng et al., 2009]) were
observed to converge faster and achieve higher accuracy.
As GoogLeNet has three input channels (RGB) compared to our single (grey) channel network a weight conversion, based on standard RGB to grey-scale mapping,
is necessary in the first convolutional layer initialization.
The other parts of the networks are initialized with random weights.
Figure 2: A webcam is used to observe the scene, providing visual inputs.
data, as it can directly be measured. We consider a
robotic arm (Figure 1) with 3 degrees of freedom (DoF),
i.e., q ∈ R3 steering its end-effector position in the plane
i.e., x ∈ R2 – ignoring orientation.
Task setup
The real-world task employs a Baxter robot’s left arm
(Figure 1B) to reach (in a vertical plane) for an arbitrarily placed blue target using vision. We control only
three joints, keeping the others fixed. At each time step
one of 9 possible actions a ∈ a is chosen to change the
robot configuration, 3 per joint: increasing or decreasing
by a constant amount (0.04 rad) or leaving it unchanged.
A monocular webcam is placed on a tripod to observe
the scene, providing raw-pixel image inputs (Figure 2).
Simulator
A simple simulator was created that, given a scene configuration Θ = [x∗ q] ∈ R5 , generates the corresponding
image. It creates images using a simplistic representation of a Baxter arm (in configuration q) and the target
(at location x∗ ) represented by a blue disc with a radius
of 3.8 pixels (9 cm) in the image (Figure 1A). A reach
is deemed successful if the robot reaches and keeps its
end-effector within 7 pixels (16 cm) of the target’s centre
for four consecutive actions. Experimental results show
that although the simulator is low-fidelity, and therefore
cheap and fast for data collection, reaching skills can be
learned and transferred to the real robot.
Network architecture
The perception network for the task has an architecture
as shown in Figure 1C, which consists of 3 convolutional
and 1 fully-connected (FC) layer. Images from the simulator or the webcam (RGB, cropped to 160 × 210) are
converted to grey-scale and downsized to 84 × 84 as inputs to the network.
The control network consists of 3 fully-connected layers, with 400 and 300 units in the two hidden layers
(Figure 1C). Input to the control network is the scene
Reward
The reward r for Q learning is determined by the Euclidean distance d = kx − x∗ k between the end-effector
and the target disc’s centre
λ(δ/d − 1), if d > δ
r = 0,
(5)
if d 6 δ, n < N
1,
if d 6 δ, n > N
where δ is a threshold for reaching a target (δ = 0.05m);
λ is a constant discount factor (λ = 10−3 ); n represents the times of d is consecutively smaller than δ and
N = 4 is a threshold that determines task completion.
This reward function will yield negative rewards until
getting close enough to the target. This helps to take
into account temporal costs during policy search, i.e.,
fewer steps are better. By giving positive rewards only
when d is smaller than the threshold δ for more than
N consecutive times, the reward function will guide a
learner to converge to a target rather than just pass
through it. This reward function proves successful for
learning planar reaching, but we do not claim optimality. Designing a good reward function is an active topic
in reinforcement learning.
Guiding Q learning with K-GPS
In Q-learning, the ε-Greedy method is frequently used
for policy search. However, our experiments show that
ε-Greedy works poorly for the planar reaching task when
using multiple DoF (Section 5.2). Therefore, we introduce a kinematics-based controller to guide the policy search (K-GPS), i.e., guide the learning of the operational space controller with a joint-space controller,
which selects actions by
arg min
a
a[q] − K−1 (x∗ ) ,
(6)
where a[] is an operator that returns an updated arm
configuration when executing an action, and K−1 (·) is
the inverse kinematic function.
Algorithm 1 shows the DQN with K-GPS. A replay
memory D is used to store samples of (Θt , at , rt , Θt+1 ).
1
2
3
4
5
6
7
8
9
Initialize replay memory D
Initialize Q-function Q(Θ, a) with random weights
for iteration=1,K do
if previous trial finished then
Start a new trial:
Randomly generate configurations q and q∗
Compute the end-effector position
x∗ = K(q∗ )
end
if rand(0,1) < ε then at = arg min ka[q] − q∗ k
12
13
14
15
160×210
C
640×480
84×84
a
else at = arg max Q(Θt , a)
10
11
B
A
Algorithm 1: DQN with K-GPS
a
Execute at and observe rt and Θt+1
Add the new sample (Θt , at , rt , Θt+1 ) into D
Sample a random mini-batch from D
Update (Q(Θ, a)) using the mini-batch
end
At the beginning of each trial, the arm’s starting configuration and target position are randomly generated. To
guarantee a random target position x∗ is reachable by
the arm, we first randomly select an arm configuration
q∗ , then use the position of its end-effector as the target position. q∗ (= K−1 (x∗ )) is also used as the desired
configuration to guide the policy search. In each iteration, the action will be selected either by the kinematic
controller (with probability ε) or by the control network.
During training, ε decreases linearly from 1 to 0.1, i.e.,
the guidance gets weaker in the process. The newly observed sample (Θt , at , rt , Θt+1 ) is added to D before the
network is updated using a mini-batch randomly selected
from D.
5
Experiments and Results
Perception and control networks were first trained and
evaluated independently under various conditions for the
benchmark reaching task. Then comparisons were made
for different combined (end-to-end) networks, such as
naively combined vs fine-tuned networks. Evaluations
were conducted in both simulated and real scenarios: a
Baxter robot arm reaching observed by a camera.
5.1
Assessing Robot Perception
To understand the effect of adapting perception with real
images, we trained six networks for the planar reaching
task with different training data as shown in Table 1.
SIM was trained from scratch purely in simulation; RW
was trained from scratch using real images; P25–100
were trained by adapting SIM with different percentages
of real images found in the mini-batches.
1418 images were collected on the real robot together
with their ground-truth scene configuration for use in
Figure 3: An image from a webcam (A) is first cropped
and scaled to match the simulator size, a virtual target
is also added (B). Like the simulated images, it is then
converted to grey-scale and scaled to 84 × 84 (C).
training and adaptation. During image collection, the
robot was moved to fixed arm configurations uniformly
distributed in the joint space. The target (blue disc)
is rendered into the image at a random position to create a large number of training samples. Figure 3 shows
a typical scene during data capture and a final dataset
image after scaling, cropping and target addition. To
increase the robustness of the trained network the image dataset was augmented by applying transformations
to the original images (rotation, translation, noise, and
brightness).
In both training and adaptation, RmsProp [Tieleman
and Hinton, 2012] was adopted using a mini-batch size
of 128 and a learning rate between 0.1 and 0.01. The
networks trained from scratch converged after 4 million
update steps (~6 days on a Tesla K40 GPU). In contrast,
those adapted from SIM converged after only 2 million
update steps.
Performance was evaluated using perception error, defined as the Euclidean norm between the predicted and
b − Θ∗ . We
ground-truth scene configuration e = Θ
compared three different scenarios:
[Sim] 400 simulator images, uniformly distributed in the
scene configuration space
[Real] 400 images collected using the real robot but
withheld during training
[Live] 40 different scene configurations during live trials
on Baxter
Results are listed in Table 1 with mean eµ and standard deviation eσ of e. As expected, the perception
network performed well in the scenarios in which they
were trained or adapted but poorly otherwise. The network trained with only simulated images (SIM) had a
small error in simulation but very poor performance in
real scenarios (Real and Live). Similarly, the network
trained (RW) or adapted (P100) with only real images
Table 1: Perception Networks, Conditions and Error Reported
Sim
Nets
Training Conditions
SIM
RW
P25
P50
P75
P100
Train from scratch, simulated images
Train from scratch, real images
Adapt SIM, 25% real, 75% simulated images
Adapt SIM, 50% real, 50% simulated images
Adapt SIM, 75% real, 25% simulated images
Adapt SIM, 100% real images
1.0
0.8
Simulation
e =kbµ¡µ¤ k
Live on Baxter
0.6
0.4
0.2
0.0
P25
P50
P75
Perception Modules
P100
Figure 4: Distance errors for networks P25 to P100
in simulation (blue) and during live trials on Baxter
(green). The circles and diamonds represent outliers.
performed fine in real scenarios but poorly in simulation. In contrast, the networks adapted with a mixture
of simulated and real images coped with all scenarios.
Results for P25 to P100 show that the fraction of real
images in a mini-batch is important for balancing real
and simulated environment performance (Figure 4). The
more real images presented during training the smaller
the real world experiment error became – similarly for
simulation. In particular, P25 had the smallest mean
error eµ in simulation and P100 the smallest eµ for real
world and live scenarios. However, when balancing eµ
and eσ , P75 had the best performance when tested live
on Baxter: it had a smaller eσ and only slightly larger
eµ compared to P100.
Comparing the performance in simulation we see that
the network adapted with no simulated images (P100
in Figure 4) resulted in a much larger error than SIM.
This indicates that the presence of simulated images in
adaptation prevents a network from forgetting the skills
learned. We also observe that, a network adapted using
only real images (P100) had a smaller error than one
trained from scratch (RW). This shows that adaptation
from a pre-trained network leads to better performance
eµ
0.013
0.537
0.012
0.013
0.015
0.498
eσ
0.009
0.191
0.008
0.008
0.010
0.162
Real
eµ
13.92
0.023
0.025
0.024
0.021
0.019
eσ
0.877
0.046
0.044
0.045
0.046
0.049
Live
eµ
13.53
0.308
0.219
0.192
0.135
0.133
eσ
1.436
0.138
0.091
0.109
0.123
0.153
as well as reduces the training time.
For all networks except SIM, errors in live trials on
Baxter were slightly larger than that for the real world
testing set, although the collected real world dataset was
augmented with translations and rotations in training.
This indicates a high sensitivity of the perception networks to variations in camera pose (between capture of
the training/testing images and the live trials). To further test this indication we trained some perception networks without data augmentation, which resulted in significantly poorer performance during live trials.
To check sensible network behaviour, we investigated
the perception networks behaviour when no target was
present. All trained networks output incorrect constant
values (with small variance) for the target position prediction. When images with two targets were presented
to the networks, a random mixture of the two target
positions were output. However in both cases, joint angles were estimated accurately. When part of the robot
body or arm was occluded, as shown in Figure 6, the arm
configurations were still estimated well, although with a
slightly greater error in most cases.
5.2
Assessing Robot Control
We trained 6 control networks in simulation for the planar reaching task with varying degrees of freedom using
ε-Greedy or K-GPS policy search. In the 1 DoF case,
only q2 was active; 2 DoF uses q2 and q3 ; while 3 DoF
controls all three joints. In training, we used a learning
rate between 0.1 and 0.01, and a mini-batch size of 64.
The ε probability decreased from 1 to 0.1 within 1 million training steps for 1 DoF reaching, 2 million steps
for 2 DoF, and 3 million steps for 3 DoF. ε-Greedy and
K-GPS used the same ε.
Figure 5 shows the learning curves indicating the success rate of a network after a certain number of training
steps. For each data point 200 reaching tests were performed using the criteria introduced in Section 4. The
target positions and initial arm configurations were uniformly distributed in the scene configuration space.
For 1 DoF reaching, networks trained using K-GPS
Table 2: Performance of ε-Greedy and K-GPS
DoF
1
2
3
dmed [cm]
ε-Greedy K-GPS
1.0
0.7
4.9
2.8
14.5
3.4
dQ3 [cm]
ε-Greedy K-GPS
2.3
1.1
9.0
3.7
28.5
4.3
α [%]
ε-Greedy K-GPS
100.00
100.00
83.75
99.50
50.25
98.50
we also tested several other control network architectures, varying the number of hidden layers and the number of units in each layer. Qualitative results show that
a network with only one hidden layer was enough for
1 DoF reaching but insufficient for 2 and 3 DoF cases.
The number of units in each layer also influenced the performance. Our proposed architecture worked best for 3
DoF reaching; at least two hidden layers with 200 and
150 units were needed.
5.3
Figure 5: Learning curves showing that K-GPS converges faster than ε-Greedy.
and ε-Greedy both converged to a success rate of 100%
after around 1 million steps (4 hrs on one 2.66GHz 64bit
Intel Xeon processor core). For the 2 DoF case, K-GPS
and ε-Greedy converged to around 100% and 80% and
took 2 million (8 hrs) and 4 million (16 hrs) steps respectively. For 3 DoF reaching, they converged to about
100% and 40% after 4 million and 6 million (24 hrs) steps
respectively. The results show that K-GPS was feasible
for all degrees of freedom, while ε-Greedy worked appropriately only in 1 DoF reaching and degraded as the
number of DoF increased.
For a more detailed comparison, we further analyzed
the error distance d – the Euclidean distance between
the end-effector and target – reached by a converged
network. 400 reaching tests were performed for each
network in simulation. The results are sumarized in Table 2 which shows that K-GPS achieved smaller error
distances for both median dmed and third quartile dQ3
than ε-Greedy in all DoF cases.
To evaluate the performance of a control network in
real scenarios, a K-GPS trained network (3 DoF) was
directly transferred on Baxter. In the test, joint angles
were taken from the robot’s encoders and the target position was set externally. It achieved a median distance
error of 1.3 pixels (3.2 cm) in 20 consecutive reaching trials (CR in Table 3), indicating robustness to real-world
sensing noise.
In addition to the proposed FC network architecture,
End-to-end Network Performance
We evaluated the end-to-end performance of combined
networks in both simulated and real-world scenarios using the metrics of Euclidean distance error d (between
the end-effector and target) and average accumulated reward R̄ (a bigger accumulated reward means a faster and
closer reaching to a target) in 400 simulated trials or 20
real trials. When testing in real scenarios, virtual targets
were rendered into the image stream from the camera for
repeatability and simplicity.
For comparison, we evaluated three networks end-toend: EE1, EE2 and EE2-FT. EE1 is a combined network comprising SIM and CR; EE2 consists of P75 and
CR; EE2-FT is EE2 after end-to-end fine-tuning using
weighted losses. P75 and CR are the perception and
control modules selected in Section 5.1 and Section 5.2,
which have the best performance individually.
The end-to-end fine-tuning was mainly conducted in
simulation. In the fine-tuning, β = 0.8, we used a learning rate between 0.01 and 0.001, a mini-batch size of
64 and 256 for task and perception losses respectively,
and an exploration possibility of 0.1 for K-GPS. These
parameters were empirically selected. To prevent the
perception module from forgetting the skills for real scenarios, the 1418 real samples were also used to obtain
δLp . Similar to P75, 75% samples in a mini-batch for
δLp were from real scenarios, i.e., at each weight updating step, 192 real and 64 simulated samples were used.
The error distances in cm and pixels (in the 84 ×
84 image) are compared. Results are listed in Table 3,
where dmed and dQ3 are the median and third quartile of
d. The CR network, where perfect perception is assumed
is added as baseline.
From the results in simulation, we can see that EE1
and EE2 have similar performance in all metrics. After
Table 3: End-to-end Reaching Performance
Scenario
Nets
EE1 (SIM+CR)
EE2 (P75+CR)
EE2-FT
CR (Control Only Baseline)
EE1
EE2
EE2-FT
CR (Control Only Baseline)
Sim
Real
Occlusion
Occlusion
A
B
dmed
[cm] [pixels]
4.7
2.0
4.6
1.9
3.6
1.5
3.4
1.4
41.8
17.5
4.6
1.9
3.7
1.6
3.2
1.3
C
[cm]
6.7
6.2
4.8
4.3
80.2
6.2
5.2
4.3
dQ3
[pixels]
2.8
2.6
2.0
1.8
33.6
2.6
2.2
1.8
R̄
[\]
0.313
0.319
0.626
0.761
-0.050
0.219
0.628
0.781
Occlusion
D
E
Occlusion
Figure 6: Successful reaching with real targets (A) and occlusions (B-E) which were not present during training.
end-to-end fine-tuning, EE2-FT achieved a much better
performance (21.7% smaller dmed and 96.2% bigger R̄)
than EE2. The fine-tuned performance is very close to
that of the control module (CR) which controls the arm
using ground-truth Θ as sensing inputs. This indicates
the proposed fine-tuning approach significantly improved
the hand-eye coordination.
In the real world, as expected, EE1 worked poorly,
since the perception network had not experienced real
scenarios. In contrast, EE2 and EE2-FT worked well and
achieved comparable performance to that in simulation.
Note that due to the cost of real world experiments, only
20 trials each were run (compared to 400 in simulation).
Similar to the results in simulation, benefiting from the
end-to-end fine-tuning, EE2-FT achieved a smaller median distance error (3.7 cm, 1.6 pixels) than EE2. This
shows that the adaptation to real scenarios can be kept
by presenting (a mix of simulated and) real samples to
compute the perception loss. All networks (except EE1
in the real scenario) achieved a success rate between 98%
and 100%.
Apart from the weighted end-to-end fine-tuning approach, we also tried naively fine-tuning combined networks only using the task loss Lq . It did not work well
for performance improvement (making the performance
even worse), although many efforts were made in searching appropriate hyper-parameters.
To see the combined networks’ robustness to a real
target and occlusions, we tested EE2-FT in the setups
shown in Figure 6. In the case without occlusion (A),
real targets can be reached (although only virtual targets
were used for training). Occlusions had not been experienced by the network during training, yet we see that
in cases B, C and E, most targets can be reached but
with larger distance errors (about 2 times larger than in
case A). In case D, only a few targets could be reached
with a yet increased error across all cases, as shown in
the attached video1 .
6
Conclusion
In this paper, we demonstrated reliable vision-based planar reaching on a real robot using a modular deep Q
network (DQN), trained in simulation, with transference to a real robot utilizing only a small number of real
world images. The proposed end-to-end fine-tuning approach using weighted losses significantly improved the
hand-eye coordination of the naively combined network
(EE2). Through fine-tuning, its (EE2-FT) reaching accuracy was improved by 21.7%. This work has led to the
following observations:
Value of a modular structure and end-to-end
fine-tuning:
The significant performance improvement (hand-eye coordination) and relatively low real world data require1
https://goo.gl/vtLuVV
ments show the feasibility of the modular structure and
end-to-end fine-tuning for low-cost transfer of visuomotor policies. Through fine-tuning with weighted
losses, a combined network comprising perception and
control modules trained independently can even achieve
performance very close to the control network alone (indicating the performance upper-limit). Scaling the proposed techniques to more complicated tasks and networks will likely be achievable with an appropriate scene
configuration representation.
Perception adaptation:
A small number of real-world images are sufficient to
adapt a pre-trained perception network from simulated
to real scenarios in the benchmark task, even with a
simulator of only modest visual fidelity. The percentage of real images in a mini-batch plays a role in balancing the performance in real and simulated environments. The presence of simulated images in fine-tuning
prevents a network from forgetting pre-mastered skills.
The adapted perception network also has some interesting robustness properties: it can still estimate the robot
configuration even in the presence of occlusions it has
not directly experienced or when there is/are zero or
multiple targets.
Control training with K-GPS:
With guidance from a kinematic controller K-GPS leads
to better policies (smaller error distance) in a shorter
time than ε-Greedy, producing a trained control network
that is robust to real-world sensing noise. However KGPS does assume that we already have some knowledge
of the task to learn, i.e., a model of the task.
We believe the architecture presented here: introducing a bottleneck between perception and control,
training networks independently then merging and finetuning, is a promising line of investigation for robotic
visual servoing and manipulation tasks. In current and
future work we are scaling up the complexity of the robot
tasks and further characterizing the performance of this
approach. Promising results have been obtained in tabletop object reaching in clutter using a 7 DoF robotic arm
in velocity control mode [Zhang et al., 2017a].
Acknowledgements
This research was conducted by the Australian Research
Council Centre of Excellence for Robotic Vision (project
number CE140100016). Additional computational resources and services were provided by the HPC and Research Support Group at the Queensland University of
Technology, Brisbane, Australia.
References
[Bateux et al., 2017] Quentin Bateux, Eric Marchand,
Jürgen Leitner, Francois Chaumette, and Peter Corke.
Visual servoing from deep neural networks. In New
Frontiers for Deep Learning in Robotics Workshop
at Robotics: Science and Systems Conference (RSS),
2017.
[Deng et al., 2009] Jia Deng, Wei Dong, Richard Socher,
Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A largescale hierarchical image database. In IEEE Conference on Computer Vision and Pattern Recognition
(CVPR), 2009.
[Devin et al., 2017] Coline Devin, Abhishek Gupta,
Trevor Darrell, Pieter Abbeel, and Sergey Levine.
Learning modular neural network policies for multitask and multi-robot transfer. In IEEE International
Conference on Robotics and Automation (ICRA),
2017.
[D’Innocente et al., 2017] Antonio
D’Innocente,
Fabio Maria Carlucci, Mirco Colosi, and Barbara Caputo. Bridging between computer and robot
vision through data augmentation: a case study on
object recognition. In International Conference on
Computer Vision Systems (ICVS), 2017.
[Fitzgerald et al., 2015] Tesca Fitzgerald, Ashok Goel,
and Andrea Thomaz. A similarity-based approach
to skill transfer. In Women in Robotics Workshop
at Robotics: Science and Systems Conference (RSS),
2015.
[Hinton and Salakhutdinov, 2006] Geoffrey E Hinton
and Ruslan R Salakhutdinov. Reducing the dimensionality of data with neural networks. science,
313(5786):504–507, 2006.
[James et al., 2017] Stephen James, Andrew J Davison,
and Edward Johns. Transferring end-to-end visuomotor control from simulation to real world for a multistage task. In 1st Annual Conference on Robot Learning (CoRL), 2017.
[Katyal et al., 2017] Kapil Katyal, I-Jeng Wang, and
Philippe Burli. Leveraging deep reinforcement learning for reaching robotic tasks. In The IEEE Conference on Computer Vision and Pattern Recognition
(CVPR) Workshops, 2017.
[Krizhevsky et al., 2012] Alex
Krizhevsky,
Ilya
Sutskever, and Geoffrey E Hinton.
Imagenet
classification with deep convolutional neural networks. In Advances in Neural Information Processing
Systems (NIPS), pages 1097–1105, 2012.
[LeCun, 1988] Y. LeCun. A theoretical framework for
back-propagation. In D. Touretzky, G. Hinton, and
T. Sejnowski, editors, Proceedings of the 1988 Connectionist Models Summer School, pages 21–28, CMU,
Pittsburgh, Pa, 1988. Morgan Kaufmann.
[Lenz et al., 2015] Ian Lenz, Ross Knepper, and
Ashutosh Saxena. Deepmpc: Learning deep latent
features for model predictive control. In Robotics:
Science and Systems (RSS), 2015.
[Taylor and Stone, 2009] Matthew E Taylor and Peter
Stone. Transfer learning for reinforcement learning
domains: A survey. The Journal of Machine Learning
Research, 10:1633–1685, 2009.
[Levine et al., 2016a] Sergey Levine, Chelsea Finn,
Trevor Darrell, and Pieter Abbeel. End-to-end training of deep visuomotor policies. Journal of Machine
Learning Research, 17(39):1–40, 2016.
[Tieleman and Hinton, 2012] Tijmen Tieleman and Geoffrey Hinton. Lecture 6.5-rmsprop: Divide the gradient by a running average of its recent magnitude.
COURSERA: Neural Networks for Machine Learning,
4(2), 2012.
[Levine et al., 2016b] Sergey Levine, Peter Pastor
Sampedro, Alex Krizhevsky, and Deirdre Quillen.
Learning hand-eye coordination for robotic grasping
with deep learning and large-scale data collection. In
International Symposium on Experimental Robotics
(ISER), 2016.
[Mnih et al., 2015] Volodymyr
Mnih,
Koray
Kavukcuoglu, David Silver, Andrei A Rusu, Joel
Veness, Marc G Bellemare, Alex Graves, Martin
Riedmiller, Andreas K Fidjeland, Georg Ostrovski,
et al. Human-level control through deep reinforcement
learning. Nature, 518(7540):529–533, 2015.
[Pan and Yang, 2010] Sinno Jialin Pan and Qiang Yang.
A survey on transfer learning. IEEE Transactions on
Knowledge and Data Engineering, 22(10):1345–1359,
2010.
[Pinto and Gupta, 2016] Lerrel Pinto and Abhinav
Gupta. Supersizing self-supervision: Learning to
grasp from 50k tries and 700 robot hours. In IEEE
International Conference on Robotics and Automation
(ICRA), 2016.
[Rusu et al., 2017] Andrei A Rusu, Matej Vecerik,
Thomas Rothörl, Nicolas Heess, Razvan Pascanu, and
Raia Hadsell. Sim-to-real robot learning from pixels
with progressive nets. In 1st Annual Conference on
Robot Learning (CoRL), 2017.
[Sutton and Barto, 1998a] Rich Sutton and Andrew G
Barto. Reinforcement learning: An introduction. MIT
Press, 1998.
[Sutton and Barto, 1998b] R.S. Sutton and A.G. Barto.
Reinforcement learning: An introduction. Cambridge
Univ Press, 1998.
[Szegedy et al., 2015] Christian Szegedy, Wei Liu,
Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir
Anguelov, Dumitru Erhan, Vincent Vanhoucke, and
Andrew Rabinovich. Going deeper with convolutions.
In IEEE Conference on Computer Vision and Pattern
Recognition (CVPR), pages 1–9, 2015.
[Tai et al., 2016] Lei Tai, Shaohua Li, and Ming Liu.
A deep-network solution towards model-less obstacle
avoidance. In IEEE/RSJ International Conference on
Intelligent Robots and Systems (IROS), 2016.
[Tobin et al., 2017] Josh Tobin, Rachel Fong, Alex Ray,
Jonas Schneider, Wojciech Zaremba, and Pieter
Abbeel. Domain randomization for transferring deep
neural networks from simulation to the real world.
In IEEE/RSJ International Conference on Intelligent
Robots and Systems (IROS), 2017.
[Tow et al., 2016] Adam W Tow, Sareh Shirazi, Jürgen
Leitner, Niko Sünderhauf, Michael Milford, and Ben
Upcroft. A robustness analysis of deep q networks. In
Australasian Conference on Robotics and Automation
(ACRA), 2016.
[Tzeng et al., 2016] Eric Tzeng, Coline Devin, Judy
Hoffman, Chelsea Finn, Pieter Abbeel, Sergey Levine,
Kate Saenko, and Trevor Darrell. Adapting deep
visuomotor representations with weak pairwise constraints. In Workshop on the Algorithmic Foundations
of Robotics (WAFR), 2016.
[Viereck et al., 2017] Ulrich Viereck, Andreas ten Pas,
Kate Saenko, and Robert Platt. Learning a visuomotor controller for real world robotic grasping using
simulated depth images. In 1st Annual Conference on
Robot Learning (CoRL), 2017.
[Zhang et al., 2015] Fangyi Zhang, Jürgen Leitner,
Michael Milford, Ben Upcroft, and Peter Corke. Towards vision-based deep reinforcement learning for
robotic motion control. In Australasian Conference on
Robotics and Automation (ACRA), December 2015.
[Zhang et al., 2017a] Fangyi Zhang, Jürgen Leitner,
Michael Milford, and Peter Corke. Sim-to-real transfer of visuo-motor policies for reaching in clutter:
Domain randomization and adaptation with modular
networks. Technical report, Queensland University of
Technology, 2017.
[Zhang et al., 2017b] Fangyi Zhang, Jurgen Leitner,
Michael Milford, and Peter I. Corke. Tuning modular networks with weighted losses for hand-eye coordination. In The IEEE Conference on Computer
Vision and Pattern Recognition (CVPR) Workshops,
July 2017.
| 2cs.AI
|
The Sparse Awakens: Streaming Algorithms for Matching Size
Estimation in Sparse Graphs
arXiv:1608.03118v2 [cs.DS] 14 Nov 2016
Graham Cormode∗
Hossein Jowhari†
Morteza Monemizadeh‡
S. Muthukrishnan §
November 15, 2016
Abstract
Estimating the size of the maximum matching is a canonical problem in graph algorithms, and one
that has attracted extensive study over a range of different computational models. We present improved
streaming algorithms for approximating the size of maximum matching with sparse (bounded arboricity)
graphs.
• (Insert-Only Streams) We present a one-pass algorithm that takes O(c log2 n) space and approximates the size of the maximum matching in graphs with arboricity c within a factor of O(c). This
improves significantly upon the state-of-the-art Õ(cn2/3 )-space streaming algorithms.
• (Dynamic Streams) Given a dynamic graph stream (i.e., inserts and deletes) of edges of an underlying c-bounded arboricity graph, we present an one-pass algorithm that uses space Õ(c10/3 n2/3 )
and returns an O(c)-estimator for the size of the maximum matching. This algorithm improves
the state-of-the-art Õ(cn4/5 )-space algorithms, where the Õ(.) notation hides logarithmic in n
dependencies.
In contrast to the previous works, our results take more advantage of the streaming access to the input
and characterize the matching size based on the ordering of the edges in the stream in addition to the
degree distributions and structural properties of the sparse graphs.
∗
Department of Computer Science, University of Warwick, UK. Supported in part by European Research Council grant ERC2014-CoG 647557 and a Royal Society Wolfson Research Merit Award. [email protected].
†
Department of Computer Science, University of Warwick, UK. Supported by European Research Council grant ERC-2014CoG 647557. [email protected] .
‡
Rutgers University, Piscataway, NJ, USA. [email protected].
§
Rutgers University, Piscataway, NJ, USA. [email protected].
1 Introduction
Problems related to (maximum) matchings in graph have a long history in Combinatorics and Computer
Science. They arise in many contexts, from choosing which advertisements to display to online users,
to characterizing properties of chemical compounds. Stable matchings have a suite of applications, from
assigning students to universities, to arranging organ donations. These have been addressed in a variety of
different computation models, from the traditional RAM model, to more recent sublinear (property testing)
and external memory (MapReduce) models. Matching has also been studied for a number of classes of input
graph: including general graphs, bipartite graphs, weighted graphs, and those with some sparsity structure.
We focus on the streaming case, where each edge is seen once only, and we are restricted to space
sublinear in the size of the graph (ie., no. of its vertices). In this case, the objective is to find (approximately)
the size of the matching. Even here, results for general graphs are either weak or make assumptions about
the input or the stream order. In this work, we seek to improve the guarantees by restricting to graphs that
have some measure of sparsity – bounded arboricity, or bounded degree. This aligns with reality, where
most massive graphs have asymptotically fewer than O(n2 ) edges.
Recently, Kapralov, Khanna, and Sudan [19] developed a streaming algorithm which computes an estimate of matching size for general graphs within a factor of O(polylog(n)) in the random-order streaming
model using O(polylog(n)) space. In the random-order model, the input stream is assumed to be chosen
uniformly at random from the set of all possible permutations of the edges. Esfandiari et al. [10] were the
first to study streaming algorithms for estimating the size of matching in bounded arboricity graphs in the
adversarial-order streaming model, where the algorithm is required to provide a good approximation for
any ordering of edges. Graph arboricity is a measure to quantify the density of a given graph. A graph
G(V, E) has arboricity c if the set E of its edges can be partitioned into at most c forests. Since a forest on
n nodes has at most n − 1 edges, a graph with arboricity c can have at most c(n − 1) edges. Indeed, by a
result of Nash-Williams [28, 29] this holds for any subgraph of a c-bounded arboricity graph G. Formally,
the Nash-Williams Theorem [28, 29] states that c = maxU ⊆V {|E(U )|/(|U | − 1)}, where |U | and |E(U )|
are the number of nodes and edges in the subgraph with nodes U , respectively. Several important families
of graphs have constant arboricity. Examples include planar graphs (that have arboricity c = 3), bounded
genus graphs, bounded treewidth graphs, and more generally, graphs that exclude a fixed minor.1
The important observation in [10] is that the size of matching in bounded arboricity graphs can be
approximately characterized by the number of high degree vertices (vertices with degree above a fixed
threshold) and the number of so called shallow edges (edges with both low degree endpoints). This characterization allows for estimation of the matching size in sublinear space by taking samples from the vertices
and edges of the graph. The work of [10] implements the characterization in Õ(cn2/3 ) space giving a O(c)
approximation of the matching size. Subsequent works [2, 26] consider alternative characterizations and
improve upon the approximation factor however they do not result in major space improvements.
1.1 Our Contributions
We present major improvements in the space usage of streaming algorithms for sparse graphs (c-bounded
Arboricity Graphs). Our main result is a polylog space algorithm that beats the nε space bound of prior
algorithms. More precisely, we show:
Theorem 1. Let G(V, E) be a graph with arboricity bounded by c. Let S be an (adversarial order) insertiononly stream of the edges of the underlying graph G. Let M ∗ be the size of the maximum matching of G (or S
interchangeably). Then, there is a randomized 1-pass streaming algorithm that outputs a (22.5c+ 6)(1+ ε)approximation to M ∗ with probability at least 1 − δ and takes O( εc2 log( 1ε ) log cn log n) space.
1
√
It can be shown that for an H-minor-free graph, the arboricity number is O(h h) where h is the number of vertices of H. [22]
1
For the case of dynamic streams (i.e, streams of inserts and deletes of edges), we design a different algorithm using Õ(c10/3 n2/3 ) space which improves the Õ(cn4/5 )-space dynamic (insertion/deletion) streaming
algorithms of [2, 3]. The following theorem states this result (proved in Section 3.3).
Theorem 2. Let G(V, E) be a graph with the arboricity bounded by c. Let M ∗ be the size of the maximum
matching of G. Let S be a dynamic stream of edge insertions and deletions of the underlying graph G of
length at most O(cn) . Let β = µ((2µ)/(µ − 2c + 1) + 1) where µ > 2c. Then, there exists a streaming
4/3
2/3
) space in expectation and outputs a (1 + ε)β approximation of M ∗ with
algorithm that takes O( β ε(nc)
4/3
probability at least 0.86.
Our algorithms for bounded arboricity graphs are based on two novel streaming-friendly characterizations of the maximum maching size. The first characterization is a modification of the characterization in
[9] which approximates the size of the maximum matching by hµ + sµ where hµ is defined as the number
of high degree vertices (vertices with degree more than a threshold µ) and sµ is the number of shallow
edges (edges with low degree endpoints). While hµ can be easily approximated by sampling the vertices
and checking if they are high degree or not, approximating sµ in sublinear space is a challenge because in
one pass we cannot determine if a sampled edge is shallow or not. The work of [9] resolves this issue by
sampling the edges at a high rate and manages to implement their characterization in Õ(cn2/3 ) space for
adversarial insert-only streams.
To bring the space usage down to Õ(c2.5 n1/2 ) (for insert-only streams), we modify the formulation of
the above characterization. We still need to approximate hµ but instead of sµ we approximate nL the number
of non-isolated vertices in the induced subgraph GL defined over the low degree vertices. Note that sµ is the
number of edges in GL . This subtle change of definition turns out to be immensely helpful. Similar to hµ we
only need to sample the nodes and check if their degrees are below a certain threshold or not, however we
carry the additional constraint that we have to avoid counting the nodes in GL that are isolated (have only
high degree nodes as neighbors). To satisfy this additional constraint, our algorithm stores the neighbors
of the sampled vertices along with a counter for each that maintains their degree in the rest of the stream.
Although we only obtain a lower bound on the degree of the neighbors, as it turns out the lower bound
information on the degree is still useful because we can ensure the number of false positives that contribute
to our estimate is within a certain limit. As result, we can approximate hµ + nL using Õ(cn1/2 ) space
which gives a (2c + 1)(2c + 2) approximation of the maximum matching size after choosing appropriate
values for µ and other parameters. This characterization is of particular importances, as it can be adapted to
work under edge deletions as well as long as the number of deletions is bounded by O(cn). Details of the
characterization and the associated algorithms are given in Lemma 4 and Section 3.2.
To obtain a polylog(n) space algorithm (and prove the claim of Theorem 1), we give a totally new
characterization. This characterization, unlike the previous ones that only depend on the parameters of the
graph, also takes the ordering of the edges in the stream into account. Roughly speaking, we characterize
the size of a maximum matching by the number of edges in the stream that have few neighbor edges in
the rest of the stream. To understand the connection with maximum matching, consider the following
simplistic special case. Suppose the input graph G is a forest composed of k disjoint stars. Observe that the
maximum matching on this graph is just to pick one edge from each star. We relate this to a combinatorial
characterization that arises from the sequence of edges in the stream: no matter how we order the edges of
G in the stream, from each star there is exactly one edge that has no neighboring edges in the remainder of
the stream (in other words, the last edge of the star in the stream). Our characterization generalizes this idea
to graphs with arboricity bounded by c by counting the α-good edges, i.e. edges that have at most α = 6c
neighbors in the remainder of the stream. We prove this characterization gives an O(c) approximation
of the maximum matching size. More important, a nice feature of this characterization is that it can be
implemented in polylog(n) space if one allows a 1 + ε approximation. The implementation adapts an idea
2
Reference
Graph class
Stream
Approx. Factor ∗
Space Bound∗∗
[19]
General
Random Order
O(polylog(n))
O(polylog(n))
[9]
Arboricity ≤ c
Insert-Only
5c + 9
Õ(cn2/3 )
Arboricity ≤ c
Insert-Only
c+2
Õ(cn2/3 )
Arboricity ≤ c
Insert/Delete
O(c)
Arboricity ≤ c
Insert-Only
(2c + 1)(2c + 2)
Õ(cn4/5 )
√
Õ(c2.5 n)
Insert/Delete
(2c + 1)(2c + 2)
Õ(c10/3 n2/3 )
Arboricity ≤ c
Insert-Only
22.5c + 6
Õ(c log2 n)
[26]
[2, 3]
This paper
This paper
This paper
Arboricity ≤ c
Table 1: Known results for estimating the size of a maximum matching in data streams. (*) In some entries,
a (1 + ε) multiplicative factor has been suppressed for concision. (**) The Õ(.) notation hides logarithmic
in n dependencies.
from the well-known L0 sampling algorithm. It runs O(log n) parallel threads each sampling the stream at
a different rate. At the end, a thread “wins” that has sampled roughly Θ(log n) elements from the α-good
edges (samples the edges with a rate of logk n where k is the number of α-good edges). The threads that
under-sample will end up with few edges or nothing while the ones that have oversampled will keep too
many α-good edges and will be terminated as result.
Table 1 summarizes the known and new results for estimating the size of a maximum matching.
1.2 Further Related Streaming Work
In the classical offline model, where we assume we have enough space to store all vertices and edges of a
graph G = (V, E), the problem of computing the maximum matching of G has been extensively studied.
The best result in this model is the 30-years-old algorithm due to Micali and Vazirani [27] with running
√
time O(m n), where n = |V | and m = |E|. A matching of size within (1 − ε) factor of a maximum
cardinality matching can be found in O(m/ε) time [17, 27]. Very recently, Duan and Pettie [5] develop a
(1 − ε)-approximate maximum weighted matching algorithm in time O(m/ε).
The question of approximating the maximum cardinality matching has been extensively studied in the
streaming model. An O(n)-space greedy algorithm trivially obtains a maximal matching, which is a 2approximation for the maximum cardinality matching [11]. A natural question is whether one can beat
the approximation factor of the greedy algorithm with O(n polylog(n)) space. Recently, it was shown
e
≃ 1.58 in one pass requires n1+Ω(1/ log log n) space
that obtaining an approximation factor better than e−1
[12, 18], even in bipartite graphs and in the vertex-arrival model, where the vertices arrive in the stream
together with their incident edges. This setting has also been studied in the context of online algorithms,
where each arriving vertex has to be either matched or discarded irrevocably upon arrival. Seminal work due
e
approximation factor in this model.
to Karp, Vazirani and Vazirani [20] gives an online algorithm with e−1
e
Closing the gap between the upper bound of 2 and the lower bound of e−1
remains one of the most
appealing open problems in the graph streaming area (see [31]). The factor of 2 can be improved on if
one either considers the random-order model or allows for two passes [21]. By allowing even more passes,
the approximation factor can be improved to multiplicative (1 − ǫ)-approximation via finding and applying
augmenting paths with successive passes [24, 25, 6, 7, 1].
Another line of research [11, 24, 32, 8] has explored the question of approximating the maximum-weight
matching in one pass and O(n polylog(n)) space. Currently, the best known approximation factor equals
4 + ε (for any positive constant ε) [4].
3
2 Preliminaries and Notations
Let G(V, E) be an undirected unweighted graph with n = |V | vertices and m = |E| edges. For a vertex
v ∈ V , let degG (v) denote the degree of vertex v in G. A matching M of G is a set of pairwise non-adjacent
edges, i.e., no two edges share a common edge. Edges in M are called matched edges; the other edges are
called unmatched. A maximum matching of graph G(V, E) is a matching of maximum size. Throughout
the paper, when we fix a maximum matching of G(V, E), we denote it by M ∗ . A matching M of G is
maximal if it is not a proper subset of any other matching in graph G. Abusing the notation, we sometimes
use M ∗ and M for the size the maximum and maximal matching, respectively. It is well-known (see for
example [23]) that the size of a maximal matching is at least half of the size of a maximum matching, i.e.,
M ≥ M ∗ /2. Thus, we say a maximal matching is a 2-approximation of a maximum matching of G. It
is known [23] that the simple greedy algorithm, where we include the newly arrived edge if none of its
endpoint are already matched, returns a maximal matching.
3 Algorithms for Bounded Arboricity Graphs
Throughout this section, hµ denotes the number of vertices in graph G = (V, E) that have degree above µ.
Let GL = (V ′ , E ′ ) be an induced subgraph of G where V ′ = {v| degG (v) ≤ µ} and (u, v) ∈ E ′ iff u and
v are both in V ′ . Note that GL might have isolated vertices. In the following we let Mµ denote the size of
maximum matching in GL .
3.1 Characterization lemmas
Lemma 3 ([9]). For a c-bounded arboricity graph G(V, E) and µ > 2c, we have hµ ≤
2µ
∗
µ−2c+1 M .
Lemma 4. For a c-bounded arboricity graph G(V, E) and µ > 2c, we have
2µ
∗
+ 1 M∗ .
M ≤ hµ + Mµ ≤
µ − 2c + 1
Proof. The lower bound is easy to see: every edge of a maximum matching either has an endpoint with
degree more than µ or both of its endpoints are vertices with degree at most µ. The number of matched
edges of the first type are bounded by hµ whereas the number of matched edges of the second type are
bounded by Mµ .
To prove the upper bound, we use the fact Mµ ≤ M ∗ and Lemma 3.
Definition 5. Let S = (e1 , . . . , em ) be a sequence of edges. We say the edge ei = (u, v) is α-good with
respect to S if max{di (u), di (v)} ≤ α where di (x) is defined as the number of the neighbors of x that
appear after the i-th location in the stream.
Lemma 6. Let µ > 2c be a (large enough) parameter. Let Eα be the set of α-good edges in an edge stream
for a graph with arboricity at most c. We have:
c
1
−
2 µ+1
M
∗
≤ |Eα | ≤
5
α + 2 M ∗,
4
4c(µ+1)
where α = max{µ − 1, µ+1−2c
}. In particular for µ = 6c − 1, we have
M ∗ ≤ 3|E6c | ≤ (22.5c + 6)M ∗
Proof. First we prove the lower bound on |Eα |. In particular we show a relation involving the number of
edges where both endpoints have low degree, sµ = |{e = (u, v)|e ∈ E, deg(u) ≤ µ, deg(v) ≤ µ}|:
4
1
c
−
hµ + sµ ≤ |Eα |.
2 µ+1
The claim in the lemma follows from the relatively loose bound that M ∗ ≤ hµ + sµ . Let H be the set of
vertices in the graph with degree above µ and let L = V \ H. Recall that hµ = |H|. Let H0 be the vertices
in H that have no neighbor in L, and let H1 = H \ H0 . First we notice that |H1 | ≥ (1 − 2c
µ )|H|. To see
this, let E ′ be the edges with at least one endpoint in H0 . By definition, every node in H0 has degree at least
µ + 1, so we have |E ′ | ≥ µ+1
2 |H0 |. At the same time, the total number of edges in the subgraph induced by
the nodes H is at most c(|H| − 1), using the arboricity assumption. Therefore,
µ+1
|H0 |
c(|H| − 1) ≥ |E ′ | ≥
2
2c
It follows that |H0 | ≤ µ+1
(|H| − 1) which further implies that
2c
2c
|H1 | ≥ (1 −
)|H| = (1 −
)hµ .
(1)
µ+1
µ+1
P
Now let dH (v) be the degree of v in the induced subgraph H. We have v∈H1 dH (v) ≤ 2c|H|,
again using the arboricity bound and the fact that summing over degrees counts each edge at most twice.
Therefore, taking the average over nodes in H1 ,
2c
dH (v) ≤
2c
1 − µ+1
for v ∈ H1 . Consequently, at least half of the vertices in H1 have their dH bounded by 4c(µ+1)
µ+1−2c (via the
′
′
Markov inequality). Let H1 be those vertices. For each v ∈ H1 we find an α-good edge. Let e∗ = (v, u)
be the last edge in the stream where u ∈ L. Then, there cannot be too many edges that neighbor (v, u)
and come after it in the stream: the total number of edges that share an endpoint with e∗ in the rest of
4c(µ+1)
the stream is bounded by max{µ − 1, 4c(µ+1)
µ+1−2c }. Consequently, for α = max{µ − 1, µ+1−2c }, we have
c
|Eα | ≥ ( 12 − µ+1
)hµ , based on the set of |H1 |/2 edges in H1′ and using (1). For α ≥ µ, Eα also contains
the disjoint set of edges from L × L, which are all guaranteed to be α-good since both their endpoints have
c
degree bounded by µ. Therefore |Eα | ≥ sµ + ( 21 − µ+1
)hµ as claimed.
To prove the upper bound on |Eα |, we notice that the subgraph containing only the edges in Eα has
4|Eα |
[14]. It follows that |Eα | ≤
degree at most α + 1. Such a graph has a matching size of at least 5(α+1)+3
5α+8
∗
4 M . This finishes the proof of the lemma.
Last, we note that in the special case of trees (or more generally, graph streams which represent forests),
a tighter approximation bound follows (for which our algorithms specified below will also apply).
Lemma 7. For trees we have M ∗ ≤ |E1 | ≤ 2M ∗ .
Proof. Let T = (V, E) be a tree with maximum matching size M ∗ . The upper bound follows by considering
E1 : the subgraph of T formed by E1 has degree at most 2, and since we are considering trees, the set E1
can have no cycles and so consists of paths. Hence, |E1 | ≤ 2M ∗ .
For the lower bound, we use induction on the number of nodes. Suppose the claim |E1 | ≥ M ∗ is true
for all trees on n nodes. We want to show that the claim remains true for trees on n + 1 nodes. The base
case n = 2 is trivially true. Given a tree T = (V, E) with n + 1 nodes, there is always a leaf w ∈ V , that
is connected to a node u where u has at most one non-leaf neighbor. If we remove the edge (u, w) from the
tree, we get a tree T ′ with n nodes and by our induction hypothesis, |E1 (T ′ )| ≥ M ∗ (T ′ ) no matter how the
stream is ordered. Fix some ordering of the stream for T ′ . We claim after inserting the edge (u, w) in the
stream (anywhere) we will have |E1 (T )| ≥ M ∗ (T ). Why? We have two cases to consider.
(1) w has no sibling in T . In this case E1 (T ) = E1 (T ′ ) ∪ {(u, w)}. This is because u must have been a
5
leaf in T ′ and as result adding (u, w) does not cause any other edge to lose the 1-goodness property.
If follows that the size of E1 increases by 1 while M ∗ increases by at most 1.
(2) w has a sibling. In this case for sure M ∗ does not increase. Although there may be a concern that the
size of E1 could drop, below we show that adding a leaf to the stream of the edges of a tree does not
cause the size of the set E1 to drop. This is enough to show that in this case as well |E1 (T )| ≥ M ∗ (T ).
To see why adding a leaf to the stream of edges cannot reduce |E1 |, assume we insert an edge e = (u, w)
in the stream where w is a newly added leaf. If u has no 1-good edges incident on it, then E1 remains as
it was. If u has one 1-good edge on it and adding e = (u, w) causes it to be kicked out of E1 , it means
e is admitted as a new member of E1 . So the loss is accounted for. If u has two 1-good edges on it,
say e1 and e2 , we show adding e cannot cause them to be ejected from E1 . To see this, assume to the
contrary that it could, and suppose (without loss of generality) the edges come in the following order in the
stream: . . . , e2 , . . . , e1 , . . . , e, . . .. The edge e2 clearly cannot be part of E1 . But e1 must have a neighboring
edge e3 incident on u, that follow it. But that means e2 was already out before adding e to the stream. A
contradiction. Finally we note that u cannot have more than two 1-good edges on it. This finishes the
proof.
√
3.2 Õ( n) space algorithm for insert-only streams
In this section, first we present Algorithm 1 that estimates Mµ + hµ and prove the following theorem.
Theorem 8. Let G(V, E) be a graph with the arboricity bounded by c. Let S be an (adversarial order)
insertion-only stream of the edges of the underlying graph G. Let β = µ((2µ)/(µ − 2c + 1) +√1) where
µ > 2c. Then, there exists an insertion-only streaming algorithm (Algorithm 1) that takes O( β εcn log n)
space in expectation and outputs a (1 + ε)β approximation of M ∗ with probability at least 0.86, where M ∗
is a maximum matching of G.
Algorithm 1: Estimate-Mµ + hµ
Initialization: Each node is sampled to set S with probability p (determined below).
Stream Processing:
forall the edges e = (u, v) in the stream do
if u ∈ S or v ∈ S then
store e in H if u ∈ S then increment d(u) else increment l(u) if numv ∈ S then increment
d(v) else increment l(v)
Post Processing:
Let S1 = {v ∈ S|d(v) ≤ µ, ∃w ∈ Γ(v) : d(w) + l(w) ≤ µ}
Let S2 be the set of vertices {v|v ∈ S, d(v) > µ}
return s = (|S1 | + |S2 |)/p
For each w in Γ(S) (the set of neighbors of nodes in S), the algorithm maintains l(w), the number
of occurrences of w observed since (the first) v ∈ S such that w ∈ Γ({v}) was added. Note that in this
algorithm, l(w) is a lower bound on the degree of w. For the output, S1 is the subset of nodes in S whose
degree is bounded by µ and additionally there is at least one neighbor of v, w, whose observed degree (d(w)
or l(w)) is at most µ. Meanwhile, S2 is the “high degree” nodes in S.
2µ
+ 1). Algorithm 1 outputs s where (1 − ε)M ∗ ≤ s ≤
Lemma 9. Let ε ∈ (0, 1) and β = µ( µ−2c+1
(1 + ε)βM ∗ with probability at least 1 − e
−ε2 M ∗ p
4β 2
.
6
Proof. First we prove the following bounds on E(s). Mµ + hµ ≤ E(s) ≤ µ(Mµ + hµ ). Let L be the set of
vertices in G that have degree at most µ and let GL be the induced graph on L. Let H = V \ L. Note that
GL might have isolated vertices. Let N be the non-isolated vertices in GL . It is clear that if the algorithm
samples v ∈ N , v will be in S1 . Likewise, if it samples a vertex w ∈ H, w will be in S2 . Given the fact that
|H| = hµ and |N | ≥ Mµ , this proves the lower bound on E(s).
The expectation may be above Mu , as the algorithm may pick an isolated vertex in GL (a vertex that is
only connected to the high-degree vertices) and include it in S1 because one of its high-degree neighbours w
was identified as low degree, i.e., w ∈ Γ(S) and l(w) ≤ µ but w ∈ H. Let u ∈ H and let U = {a1 , . . . , aµ }
be the last µ neighbours of u according to the ordering of the edges in the stream. The algorithm can only
identify u as low degree when it picks a sample from U and no samples from Γ(u) \ U . This restricts the
number of unwanted isolated vertices to at most µhµ . Together with the fact that |N | ≤ µMµ , it establishes
the upper bound on E(s).
Now using a Chernoff bound, Pr[|s − E(s)| ≤ λE(s)] ≤ e
probability at least 1 − e
Setting λ =
ε
β
−λ2 (Mµ +hµ )p
4
−λ2 M ∗ p
4
≤ e
−λ2 M ∗ p
4
. Therefore with
,
(Mµ + hµ ) − λµ(hµ + Mµ ) ≤ s ≤ µ(1 + λ)(Mµ + hµ )
and putting this and Lemma 4 together, we derive the statement of the lemma.
Algorithm 2: Estimate-M ∗
(2)
√
Initialization: Let ε ∈ (0, 1) and t = ⌈ β ε8nc ⌉ where β is as defined in Lemma 9.
Stream Processing: Do the following tasks in parallel:
(1) Greedily keep a maximal matching of size at most r ≤ t (and terminate this task if this size bound is
exceeded).
(2) Run the Estimate-(Mµ + hµ ) procedure (Algorithm 1) with parameter p ≥
8
λ2 t
where λ = βε .
Post processing: If r < t then output 2r as the estimate for M ∗ , otherwise output the result of the
Estimate-(Mµ + hµ ) procedure.
Proof of Theorem 8: Suppose M ∗ < t. Clearly the size of the maximal matching r obtained by the first
task will be less than t. In this case, M ∗ ≤ M ′ ≤ 2M ∗ . Now suppose M ∗ ≥ t. By Lemma 4, we will have
Mµ + hµ ≥ t and hence by Lemma 9, with probability at least 1 − e−2 ≥ 0.86, the output of the algorithm
will be within the promised bounds. The expected space of the algorithm is O((t + pnc) log
n). Setting
√
√
β cn
t = β 8nc/ε to balance the space costs, the space complexity of the algorithm will be O( ε log n) as
claimed.
✷
3.3 O(n2/3 ) space algorithm for insertion/deletion streams
Algorithms 1 and 2 form the basis of our solution in the more general case where the stream contains
deletions of edges as well. In the case of Algorithm 1, the algorithm has to maintain the induced subgraph
on S and the edges of the cut (S, Γ(S)). However if we allow arbitrary number of insertions and deletions,
the size of the cut (S, Γ(S)) can grow as large as O(n) even when |S| = 1. This is because each node at
some intermediate point could become high degree and then lose its neighbours because of the subsequent
deletion of edges. Therefore here in order to limit the space usage of the algorithm, we make the assumptions
that number of deletions is bounded by O(cn). Since the processed graph has arboricity at most c this forces
the number of insertions to be O(cn) as well. Under this assumption, if we pick a random vertex, still, in
expectation the number of neighbours is bounded by O(c).
7
Another complication arises from the fact that, with edge deletions, a vertex added to Γ(S) might become isolated at some point. In this case, we discard it from Γ(S). Additionally for each vertex in S ∪ Γ(S),
the counters d(v) (or l(v) depending on if it belongs to S or Γ(S)) can be maintained as before. The space
complexity of the algorithm remains O(pnc log n) in expectation as long as the arboricity factor remains
within O(c) in the intermediate graphs. In the case of Algorithm 2, we need to keep a maximal matching
)1/3 to
of size O(t). This can be done in O(t2 ) space using a randomized algorithm [3]. Setting t at ( 8βnc
ε2
rebalance the space costs, we obtain the result of Theorem 2.
3.4 The O(log2 n) space algorithm for insert-only streams
In this section we present our polylog space algorithm by presenting an algorithm for estimating |Eα | within
(1 + ε) factor. Our algorithm is similar in spirit to the known L0 sampling strategy. It runs O(log n) parallel
threads each sampling the stream at a different rate. At the end, a thread “wins” that has sampled roughly
n
Θ(log n) elements from |Eα | (samples the edges with a rate of log
|Eα | ). The threads that under-sample will
end up with few edges or nothing while the ones that have oversampled will keep too many elements of Eα
and will be aborted as result.
First we give a simple procedure (Algorithm 3) that is the building block of the algorithm.
Algorithm 3: The α-good test
Initialization: given the edge e = (u, v) in the stream, let r(u) = 0 and r(v) = 0.
forall the subsequent edges e′ = (t, w) do
if t = u or w = u then increment r(u) if t = v or w = v then increment r(v) if
max{r(u), r(v)} > α then terminate and report NOT α-good
Now we present the main algorithm (Algorithm 4) followed by its analysis.
Algorithm 4: An algorithm for approximating |Eα |
Initialization: ∀i.Xi = ∅
⊲ Xi represents the current set of sampled α-good edges.
Stream Processing:
forall the levels i ∈ {0, 1, . . . , [⌊log1+ε m⌋} in parallel do
forall the edges e do
Feed e to the active α-good tests and update Xi
1
With probability pi = (1+ε)
i add e to Xi and start a α-good test for e.
Let |Xi | be the number of active α-good tests within this level.
2
n
if |Xi | > τ = 64αcεlog
then terminate level i
2
Post processing:
if |X0 | ≤ τ then
return |X0 |
else
⊲ |X0 | > τ
and the j-th level was not terminated;
let j be the smallest integer such that Xj ≤ 8 log n(1+ε)
ε2
Xj
if there is no such j then return FAIL else return pj
Lemma 10. With high probability, Algorithm 4 outputs a 1 ± O(ε) approximation of |Eα |.
Proof. It is clear that if |X0 | ≤ τ then X0 = Eα and the algorithm makes no error. In case |X0 | > τ , we
claim that |Eα | > 2αc 2 τ . To prove this let t be the time step where |X0 | exceeds τ and let Gt = (V, E (t) )
8
be the graph where E (t) = {e1 , . . . , et }. Clearly M ∗ (G) ≥ M ∗ (Gt ) because the size of matching only
increases. Abusing the notation, let Eα (Gt ) denote the set of α-good edges at time t. By Lemma 6, we have
5
2µ
2α2
τ < |Eα (Gt )| ≤
α + 2 M ∗ (Gt ) ≤ 4αM ∗ (G) ≤
4α|Eα | ≤
|Eα |.
4
µ − 2c
c
This proves the claim.
n
and let i∗ be the integer such
Therefore in the following we assume that |Eα | > 2αc 2 τ . Let τ ′ = 8 log
ε2
∗
∗
′
i
−1
′
i
that (1 + ε)
τ ≤ |Eα | ≤ (1 + ε) τ .
τ′
≤ E[|Xi∗ |] ≤ τ ′ . By a
Assuming the i∗ -th level does not terminate before the end, we have (1+ε)
Chernoff bound, for each i we have (again assuming we do not terminate the corresponding level)
2
ε pi |Eα |
.
Pr[||Xi | − E(|Xi |)| ≥ εE(|Xi |)] ≤ exp −
4
2
ε |Eα |
2 log n
−1 ).
Therefore, Pr[||Xi∗ | − E(|Xi∗ |)| ≥ εE(|Xi∗ |)] ≤ exp (− 2(1+ε)
i∗ ) ≤ exp ( 1+ε ) ≤ O(n
As a result, with high probability |Xi∗ | ≤ 8 log n(1+ε)
. Moreover for all i < i∗ − 1, the corresponding
ε2
levels either terminate prematurely or in the end we will have |Xi | > 8 log n(1+ε)
with high probability.
ε2
∗
∗
∗
∗
Consequently j ∈ {i , i − 1}. It remains to prove that runs corresponding to i and i − 1 will survive until
the end with high probability. We prove this for i∗ . The case of i∗ − 1 is similar.
(t)
Consider a fixed time t in the stream and let Xi∗ be the set of sampled α-good edges at time t corre(t)
sponding to the i∗ -th level. Note that Xi∗ contains the a subset of α-good edges with respect to the stream
St = (e1 , . . . , et ). From the definition of i∗ and our earlier observations we have
2α2 |Eα |
2α2 τ ′
|Eα (Gt )|
(t)
≤
≤
E[|Xi∗ |] =
∗
∗
(1 + ε)i
c(1 + ε)i
c
By the Chernoff inequality for δ ≥ 1,
i
h
−δ
(t)
(t)
(t)
Pr |Xi∗ | ≥ (1 + δ)E(|Xi∗ |) = τ ≤ exp
E(|Xi∗ |) .
3
From δ =
τ
(t)
E(|Xi∗ |)
∗
−1 =
τ (1+ε)i
|Eα (Gt )|
− 1, we get
−τ
|Eα (Gt )|
2α2 τ ′
−τ
≥ τ ≤ exp
Pr
+
+
≤ exp
3
(1 + ε)i∗
3
c
8α2 τ ′
For τ ≥ c , the term inside the exponent is smaller than −2 log n. It also satisfies δ ≥ 1. After
h
(t)
|Xi∗ |
i
(t)
applying the union bound, for all t the size of Xi∗ is bounded by τ =
finished the proof of the lemma.
64α2 log n
cε2
with high probability. This
Next, putting everything together, we prove Theorem 1.
Proof of Theorem 1: The theorem follows from Lemmas 6 and 10 and taking α = µ + 1 = 6c. Observe
that the space cost of Algorithm 4 can be bounded: we have log1+ε m levels where each level runs at
most τ concurrent α-good tests otherwise it will be terminated. Each α-good test keeps an edge and two
counters and as result it occupies O(1) space. Consequently the space usage of the algorithm is bounded by
O(τ log1+ε m). The space bound in the theorem follows from the facts that τ = O( εc2 log n) for α = 6c
and m ≤ cn.
✷
References
[1] K. J. Ahn, S. Guha, and A. McGregor. Analyzing graph structure via linear measurements. In Proceedings of the 23rd Annual ACM-SIAM Symposium on Discrete Algorithms (SODA), pages 459–467,
2012.
9
[2] M. Bury and C. Schwiegelshohn. Sublinear estimation of weighted matchings in dynamic data streams.
In Proceedings of the 23rd Annual European Symposium on Algorithms (ESA), pages 263–274, 2015.
[3] R. Chitnis, G. Cormode, H. Esfandiari, M.T. Hajiaghayi, A. McGregor, M. Monemizadeh, and
S. Vorotnikova. Kernelization via sampling with applications to finding matchings and related problems in dynamic graph streams. In Proceedings of the 27th Annual ACM-SIAM Symposium on Discrete
Algorithms (SODA), pages 1326–1344, 1326-1344.
[4] M. Crouch and D. S. Stubbs. Improved streaming algorithms for weighted matching, via unweighted
matching. In Proceedings of the 17th International Workshop on Randomization and Approximation
Techniques in Computer Science (RANDOM), pages 96–104, 2014.
[5] R. Duan and S. Pettie. Linear-time approximation for maximum weight matchings. Journal of the
ACM, 61(1):1–23, 2014.
[6] S. Eggert, L. Kliemann, P. Munstermann, and A. Srivastav. Bipartite graph matchings in the semistreaming model. Algorithmica, 63(1-2):490–508, 2012.
[7] Sebastian Eggert, Lasse Kliemann, Peter Munstermann, and Anand Srivastav. Bipartite matching in
the semi-streaming model. Algorithmica, 63(1-2):490–508, 2012.
[8] Leah Epstein, Asaf Levin, Julián Mestre, and Danny Segev. Improved approximation guarantees for
weighted matching in the semi-streaming model. SIAM J. Discrete Math., 25(3):1251–1265, 2011.
[9] H. Esfandiari, M.T. Hajiaghyi, V. Liaghat, M. Monemizadeh, and K. Onak. Streaming algorithms
for estimating the matching size in planar graphs and beyond. In Proceedings of the 26th Annual
ACM-SIAM Symposium on Discrete Algorithms (SODA), 2015.
[10] Hossein Esfandiari, Mohammad Taghi Hajiaghayi, Vahid Liaghat, Morteza Monemizadeh, and
Krzysztof Onak. Streaming algorithms for estimating the matching size in planar graphs and beyond. In Proceedings of the Twenty-Sixth Annual ACM-SIAM Symposium on Discrete Algorithms,
SODA 2015, San Diego, CA, USA, January 4-6, 2015, pages 1217–1233, 2015.
[11] J. Feigenbaum, S. Kannan, A. McGregor, S. Suri, and J. Zhang. On graph problems in a semi-streaming
model. Theoretical Computer Science, 348(2):207–216, 2005.
[12] Ashish Goel, Michael Kapralov, and Sanjeev Khanna. On the communication and streaming complexity of maximum bipartite matching. In Proceedings of the 23rd Annual ACM-SIAM Symposium on
Discrete Algorithms (SODA), pages 468–485, 2012.
[13] Oded Goldreich and Dana Ron. Property testing in bounded degree graphs. Algorithmica, 32(2):302–
343, 2002.
[14] Yijie Han. Matching for graphs of bounded degree. In Frontiers in Algorithmics, Second Annual
International Workshop, FAW 2008, Changsha, China, June 19-21, 2008, Proceeedings, pages 171–
173, 2008.
[15] Yijie Han. Matching for graphs of bounded degree. In Frontiers in Algorithmics, pages 171–173.
Springer, 2008.
[16] A. Hassidim, J. A. Kelner, H. N. Nguyen, and K. Onak. Local graph partitions for approximation and
testing. In Proceedings of the 50th IEEE Symposium on Foundations of Computer Science (FOCS),
pages 22–31, 2009.
10
[17] John E. Hopcroft and Richard M. Karp. An n5/2 algorithm for maximum matchings in bipartite graphs.
SIAM J. Comput., 2(4):225–231, 1973.
[18] M. Kapralov. Better bounds for matchings in the streaming model. Proceedings of the 23rd Annual
ACM-SIAM Symposium on Discrete Algorithms (SODA), pages 1679–1697, 2013.
[19] M. Kapralov, S. Khanna, and M. Sudan. Approximating matching size from random streams. Proceedings of the 25th Annual ACM-SIAM Symposium on Discrete Algorithms (SODA), pages 734–751,
2014.
[20] R. M. Karp, U. V. Vazirani, and V. V. Vazirani. An optimal algorithm for on-line bipartite matching.
Proceedings of the 22nd Annual ACM Symposium on Theory of Computing (STOC), pages 352–358,
1990.
[21] C. Konrad, F. Magniez, and C. Mathieu. Maximum matching in semi-streaming with few passes. In
Proceedings of the 11th International Workshop on Randomization and Approximation Techniques in
Computer Science (RANDOM), pages 231–242, 2012.
[22] Alexandr V. Kostochka. Lower bound of the hadwiger number of graphs by their average degree.
Combinatorica, 4(4):307–316, 1984.
[23] L. Lovasz and M.D. Plummer. Matching theory. In North-Holland, Amsterdam-New York, 1986.
[24] A. McGregor. Finding graph matchings in data streams. In Proceedings of the 8th International
Workshop on Randomization and Approximation Techniques in Computer Science (RANDOM), pages
170–181, 2005.
[25] A. McGregor. Graph mining on streams. In Encyclopedia of Database Systems, pages 1271–1275.
Springer, 2009.
[26] A. McGregor and S. Vorotnikova. Planar matching in streams revisited. In Proceedings of the 19th
International Workshop on Approximation Algorithms for Combinatorial Optimization Problems (APPROX), 2016.
p
[27] S. Micali and V. V. Vazirani. An o( |V ||e|) algorithm for finding maximum matching in general
graphs. Proceedings of the 21st IEEE Symposium on Foundations of Computer Science (FOCS), pages
17–27, 1980.
[28] C. St. J. A. Nash-Williams. Edge-disjoint spanning trees of finite graphs. Journal of the London
Mathematical Society, 36(1):445–450, 1961.
[29] C. St. J. A. Nash-Williams. Decomposition of finite graphs into forests. Journal of the London Mathematical Society, 39(1):12, 1964.
[30] H. N. Nguyen and K. Onak. Constant-time approximation algorithms via local improvements. In
Proceedings of the 49th IEEE Symposium on Foundations of Computer Science (FOCS), pages 327–
336, 2008.
[31] List of open problems in sublinear algorithms: Problem 60. http://sublinear.info/60.
[32] Mariano Zelke. Weighted matching in the semi-streaming model. Algorithmica, 62(1-2):1–20, 2012.
11
| 8cs.DS
|
ARXIV PREPUBLICATION
1
Regularized maximum likelihood estimation of
covariance matrices of elliptical distributions
arXiv:1611.10266v1 [math.ST] 30 Nov 2016
Christophe Culan , Claude Adnet∗ ,
Advanced Radar Concepts division, Thales, Limours 91338, France
Abstract—The maximum likelihood principle is widely used in
statistics, and the associated estimators often display good properties. indeed maximum likelihood estimators are guaranteed to
be asymptotically efficient under mild conditions.
However in some settings, one has too few samples to get a good
estimation. It then becomes desirable to take into account prior
information about the distribution one wants to estimate.
One possible approach is to extend the maximum likelihood
principle in a bayesian context, which then becomes a maximum
a posteriori estimate; however this requires a distribution model
on the distribution parameters.
We shall therefore concentrate on the alternative approach of
regularized estimators in this paper; we will show how they
can be naturally introduced in the framework of maximum
likelihood estimation, and how they can be extended to form
robust estimators which can reject outliers.
A. Notations and conventions
In the following development, the following notations and
conventions shall be observed:
•
•
Index Terms—Maximum likelihood estimation, KL-divergence,
entropy, elliptical distributions, complex elliptical distributions,
adaptive detection, iterative algorithm, outlier detection, outlier
rejection, regularized estimators
I. I NTRODUCTION : REGULARIZED ESTIMATION
Many statistical estimators are shown to be asymptotically
well behaved; in particular maximum likelihood estimators are
consistent and efficient under relatively mild conditions, and
often have good convergence properties, thus making them
a popular choice for parametric estimation [21][25][12][1].
However in a few sample regime, their performances can be
greatly degraded, and in some cases one may not even be able
to compute these estimators [20]; this is notably the case for
covariance estimations when the number of samples is of the
same order or lower than the dimension of the space [13][29].
In such situations, one has to add additional information
in the estimation procedure, other than available samples.
Hence a popular approach is to use regularized estimators,
which are estimators purposely biased towards a prior distribution representing an initial knowledge of the model; the
effect generally is a smoothing of the estimated parameter,
although this largely depends upon the shape of the prior
distribution; moreover this prior is not necessarily explicitly
known. Various regularization methods have been proposed
[30][4][2][9][16][6][11]. A popular approach is to derive
such estimates by a penalized maximum likelihood approach
[31][22][29][17].
The applied penalties are often chosen for convenience of
computation; we introduce in this article a way to define such
penalties which are in agreement with the theory of maximum
likelihood.
Manuscript received October XX, 2016; revised October XX, 2016. Corresponding author: M. Culan (email: [email protected]).
•
•
•
•
•
•
•
For any topological space R1 and R2 , the topological
space R1 × R2 is the product of R1 and R2 , whereas
R1 ∨ R2 is the disjoint union of R1 and R2 . One also
notes, for a topological space R, RN to be the product
of N copies of R and ∨N R to be the disjoint union of
N copies of R.
Let R1 , ..., RN be topological spaces, and let µn be
a measure defined onQ Rn for 1 ≤ n ≤ N . The
N
measure µ1 . . . µn = n=1 µn is the product measure
QN
of (µn )1≤n≤N defined on n=1 Rn , whereas
WN
WN
µ1 ∨ · · · ∨ µn = n=1 µn is the measure of n=1 Rn
such that its restriction to Rn is µn for 1 ≤ n ≤ N and
is the joint measure of (µn )1≤n≤N .
Moreover for a topological space R and a measure µ
defined on R, µN is the product measure of N copies of
µ defined on RN , whereas ∨N µ is the joint measure of
N copies of µ defined on N · R.
The vector space Cd is canonically identified to R2d [23].
Hd (R) is the set of symmetric matrices of size (d, d);
Hd (C) is the set of hermitian matrices of size (d, d).
Hd+ (K) is the subset of matrices of Hd+ (K) which are
+
positive; HP +
d (K) is the subset of matrices of Hd (K)
of unit determinant.
The vector space Cd is canonically identified to R2d [23].
X † is the transpose conjugate of any matrix or vector X.
µRd is the Lebesgues measure of the space Rd . Similarly,
µCd is the Lebesgues measure of Cd ' R2d .
Sd−1 is the (d − 1)-sphere, which is identified as the following part of Rd : Sd−1 ' x ∈ Rd ; x† x = 1 . Similarly
S2d−1 is identified to x ∈ Cd ; x† x = 1 in Cd .
sd−1 shall denote the probability distribution of Sd−1
isotropic for the canonical scalar product of Cd , defined
by:
d−1
X
hy, xi =
yk xk
(1)
k=0
R
• δx
•
is the Dirac delta distribution centered in x in space
R.
One shall note:
R
Z
dδ0
(t) dδ0R (t)
(2)
hδ = H δ0R |µR =
log
dt
t∈R
It is the entropy of the Dirac distribution relative to
the Lebesgues measure of R due to its translational
ARXIV PREPUBLICATION
2
symmetry, which is positive and infinite. It can be used
to express different other entropies in higher dimensional
settings:
d
H δxR |µRd = dhδ
(3)
d
H δxC |µCd = 2dhδ
II. M AXIMUM LIKELIHOOD WITH PRIOR INFORMATION
A. General theory
Let us first recall that likelihood can be understood in terms
of information; indeed it can be expressed as the relative
entropy (or KL-divergence [15]) of the sampling distribution
S relative to the model distribution P [7]:
Z
l(P |S) = −H(S|P ) =
log
x∈R
dP
(x) dS(x)
dS
(4)
Under the usual hypothesis in which one has N i.i.d samples
(xn )1≤n≤N , the usual sampling distribution is defined on a
space ∨N B, with B being some base space, and is given by the
disjoint union of Dirac distributions centered on each sample
[7]:
N
1 _ B
S=
δ
N n=1 xn
The model distribution is given by
being some distribution of B.
(5)
1
N
∨N PB , with PB
Thus the likelihood can be expressed under this hypothesis
by:
l(P |S) = −
N
1 X
H(δxBn |PB )
N n=1
(6)
Obviously this particular instantiation of the maximum
likelihood problem is not well suited for our current problem,
as we would like to take into account some prior information
in addition to the information provided by samples.
We shall focus on a situation in which we have an existing
prior estimate, and we want to take into account this prior in
our estimation procedure [29][17].
Fortunately the entropic formulation of the maximum
likelihood is quite flexible; in particular one can use different
models for the sampling distribution, as is done for example
in [7] to take into account the circular symmetry of the
distributions.
We therefore propose to take into account the prior knowledge
as part of the sampling distribution. The sampling distribution
under the i.i.d hypothesis, with additional prior information
can then be modeled as:
S=
with:
N
α _
Sn
N n=1
!
∨ ((1 − α)Pprior )
(7)
Sn the sample distribution associated to the n-th sample,
defined on a base space B; it would generally be given
by δxBn .
• Pprior the prior distribution. It is not necessarily defined
on the same space as the sample distributions depending
on the information it carries, and might be defined on a
quotiented image of the base space, which shall be noted
as [B].
• α ∈ [0; 1] the integration factor, which represents how
much ones trusts the prior estimate compared to the newly
acquired samples.
The distribution model one wants to fit can on the other
hand be expressed as:
α
∨N P ∨ ((1 − α)[P ])
(8)
N
with:
• P the distribution model
• [P ] the distribution model resulting from P on the
quotiented space [B]
The corresponding likelihood can therefore be expressed as:
•
N
α X
H(Sn |P ) − (1 − α)H(Pprior |[P ])
l(P |S) = −
N n=1
(9)
Therefore this results in a penalized version of maximum
likelihood , the penalty being defined from the relative entropy
of the prior distribution for the model distribution [17][29].
III. R EGULARIZED COVARIANCE ESTIMATORS
We shall now apply the regularized maximum likelihood
equation to different covariance estimation problems. We’ll
focus on the two following cases:
• The distribution which we try to estimate is elliptically
symmetric (ES/CES) [5][18]. No hypothesis is made on
the radial distribution of this distribution; this corresponds
to the conditions for which Tyler’s estimator is the maximum likelihood estimator for i.i.d samples [7][27][19].
• The distribution is gaussian, or circular gaussian for complex distributions. Note that the corresponding maximum
likelihood estimators under the i.i.d hypothesis are given
in [7].
A. Regularized Tyler’s estimator
Let us first recall that an elliptical distribution can be
expressed in general by [7]:
√
1
dP (x) = p dQ
x† R−1 x dsd−1
|R|
L(R)−1 x
√
x† R−1 x
(10)
with the radial distribution Q being any probability measure
on R+ , and the correlation matrix R being any matrix of
HP +
d (K) [7].
Supposing here that one has no a priori on the radial
distribution, the prior distribution one wants to use should
be defined on the unit sphere, and corresponds to the unique
ARXIV PREPUBLICATION
3
probability distribution of the unit sphere which is isotropic
Rprior
for the corresponding correlation matrix Rprior , noted as sd−1
.
The entropy between two such distributions isotropic for
correlation matrices R1 and R2 shall be noted H(R1 |R2 ).
Thus the concentrated likelihood for samples (xn )1≤n≤N in
Rd under i.i.d hypothesis is given, after maximization over
the radial distribution, by [7]:
l(R|S) = − dhδ −
N
α(d − 1) X
log xn † R−1 xn
2N
n=1
1:
function REG T YLER(Rprior ,(xn )1≤n≤N ,α,)
2:
R ← Rprior
3:
repeat
√
R 12 ←
4:
5:
R− 12 ← R 12 −1
6:
S ← (1 − α)anscm R− 21 Rprior R− 21
(11)
+
− (1 − α)H(Rprior |R)
H(Rprior |R) however represents a challenge to compute.
Fortunately, one can characterize its differential with respect
to R (see A for more details):
1
1
1
d − 1 1
tr R 2 anscm(R− 2 Rprior R− 2 )R 2 d R−1
2
(12)
d
tr(S) S
S←
8:
R ← R 21 exp(S − I)R 12
R
R ← tr(R)
2
until tr (R−1 R − I) ≤
10:
=
N
X
α
xn xn †
R− 21
R 1
†
N
x R− 21 R− 21 xn − 2
n=1 n
7:
9:
dH(Rprior |R)
R
return R− 12
11:
12:
end function
In the case of a distribution in Cd under the circularity
hypothesis, the maximum likelihood equation under i.i.d hypothesis is slightly different; indeed the sampling distribution
is then given by:
with anscm(Σ) being the expected value of the normalized
sample covariance estimator (NSCM) for an elliptical distribution of correlation matrix proportional to Σ.
Unfortunately there is no explicit way of computing this matrix
in the general case to the authors’ knowledge and one has
to resort to numerical methods to extract its eigenvalues.
However it has been characterized for complex circularly
symmetric distribution in [3].
By deriving the concentrated likelihood in a manner similar
to [7] and differentiating it, on gets the following equation:
In any case this leads to the following maximum likelihood
equation:
tr λR
S=
−
Without any structural constrain on R, this simplifies to:
1
1
1
1
1
R=
(1 − α)R 2 anscm R− 2 Rprior R− 2 R 2
2λ(d − 1)
N
α X xn xn †
+
N n=1 xn † R−1 xn
(14)
This can be solved by using the following numerical
procedure:
(15)
1
1
1
1
1
− (1 − α) d −
R 2 anscm R− 2 Rprior R− 2 R 2
2
N
α(d − 1) X xn xn †
−1
−
d R
=0
N
x † R−1 xn
n=1 n
(16)
tr λR
1
1
1
(1 − α)(d − 1) 1
R 2 anscm R− 2 Rprior R− 2 R 2
2
N
α(d − 1) X xn xn †
−1
−
d
R
=0
2N
x R−1 xn
n=1 n
(13)
N
α _ C
Rprior
δ
+ (1 − α)sd−1
N n=1 [xn ]
This is almost the same equation, although each sample
actually carries less information compared to the prior
distribution for a same α parameter.
The equation can thus be solved numerically by calling the
same algorithm REG T YLER with a different α parameter:
REG
T YLER(Rprior ,(xn )1≤n≤N , d−α(d−1)
,)
1
(1+α)
2
Note however that the computation of the ansm function
differs in the real and complex circular cases (see annex).
These regularized Tyler estimators are more complicated
to compute than those found in the existing litterature by
applying different penalties (specifically ones maintaining the
convexity of the problem) and offer no theoretical guaranty
ARXIV PREPUBLICATION
4
of convergence [28][17][26]. However they use a penalty
which is consistent with the definition of the likelihood as an
information theoretic quantity, which is the aim of this paper.
B. Regularized estimators for gaussian distributions
A subclass of interest of elliptical models are estimations
constrained on a subclass of models corresponding to a fixed
scaled radial distribution, of the from:
L(R)−1 x
(17)
dP (x) = σdQ0 (σr)dsd−1 √
x† R−1 x
Indeed most M-estimators can be expressed as maximum
likelihood estimators for one of these subclasses of models
[7].
However the existence of an easy to compute regularized
estimator then depends on the expression of the gradient
of H(Σ1 |Σ2 ) with respect to Σ2 for two distributions with
the same base radial distribution Q0 . Unfortunately there is
no analytical expression of this gradient except in specific
cases [14]. We shall therefore focus on one of these specific
cases: that of gaussian distributions, which corresponds to Q0
following a χ2 (cK d) distribution, with cK = 1 for K = R,
and cK = 2 for K = C.
The relative entropy of two such gaussian models is given
by [10]:
cK
tr Σ2 −1 Σ1 + log |Σ2 | − log |Σ1 | − d
2
(18)
with Σ1 and Σ2 being the covariance matrices of said
models. In the standard case of gaussian distributions in Rd ,
the resulting likelihood is thus given up to constant additive
terms by:
H(Σ1 |Σ2 ) =
N
α X † −1
xn Σ xn
2N n=1
(19)
1
1−α
−1
tr Σ Σprior − log |Σ|
−
2
2
The corresponding maximum likelihood estimator is then
given by a weighted mean of the covariance of the prior
distribution and that of the sample covariance matrix of the
new samples (SCM):
This can be differentiated, leading to the following maximum likelihood equation:
Σ=
1
α ((1 − α)Σprior
1 − 2d
N
α X
1
+
1−
xn xn † )
N n=1
2xn † Σ−1 xn
(22)
A solution can be found numerically by adapting the
algorithm described in [7]:
CG COV
1:
2:
3:
function REG CG COV(Σprior , (xn )1≤n≤N , α, , Kmax )
Σ ← (1 − α)Σprior +
repeat
√
4:
Σ 21 ←
5:
Σ− 12 ← Σ 21 −1
6:
S←
7:
8:
9:
10:
α
N
N
P
xn xn †
n=1
Σ
1
α
(1− 2d
)
((1 − α)Σprior
N
1
α X
1−
+
xn xn † )
N n=1
2xn † Σ−1 xn
R ← Σ 12 exp Σ− 12 SΣ− 12 − I Σ 21
2
≤
until tr Σ− 21 ΣΣ− 12 − I
return Σ− 21
end function
Thus in the specific case of gaussian and complex circular
gaussian variables, the regularized maximum likelihood estimators can be computed efficiently.
l(Σ|S) = −
Σ = (1 − α)Σprior +
N
α X
xn xn †
N n=1
(20)
In the complex circular case, one can take into account the
phase symmetry of the sampling distribution, in a manner
similar to Tyler’s estimator; this results in the following
likelihood, up to constant additive terms:
N
α X
1
† −1
† −1
l(Σ|S) = −
xn Σ xn − log xn Σ xn
N n=1
2
α
− (1 − α)tr Σ−1 Σprior − 1 −
log |Σ|
2d
(21)
IV. PARTIAL ESTIMATION AND REGULARIZED ESTIMATORS
We have thus successfully extended maximum likelihood
estimators of covariance matrices to a regularized estimation
setting. We shall now discuss a further extension of these
regularized estimators in order to improve their robustness to
cases of contamination by outliers. In particular, we would
like to extend the partial maximum likelihood estimation
method developed in [8] to these estimators.
Let us recall that partial estimation as introduced in [8]
consists in the maximization of a partial likelihood:
dP
(x) dS(x)
l|X (P |S) = −H|X (P |S) =
log
dS
x∈X
(23)
under the constrain that X can be any measurable set such
that S(X) ≥ p with p ∈ [0; 1] being the order of the partial
estimation.
Z
Unfortunately this definition is impractical in the case of
regularized estimators for which R = ∨N B ∨ [B]. Indeed one
ARXIV PREPUBLICATION
5
would have to consider partial entropy integrals between the
prior and model distribution of the form:
dPprior
H|X∩[B] (Pprior |P ) =
log
(x) dPprior (x)
dP
x∈X∩[B]
(24)
We therefore propose to restrict the possible partial domains
to subsets X of the form X ∨ [B], with X ⊂ ∨N B and such
that S(X) ≥ (1 − α) + αp. Thus the corresponding restricted
partial likelihood, maximized on the partial domain X is now
given by [8]:
Z
11:
R ← R 21 exp(S − I)R 12
12:
R
R ← tr(R)
2
if tr (R−1 R − I) ≤ then
13:
14:
15:
16:
17:
18:
19:
lp (P |S) = − Hp (P |S)
Z
=
sup
S(X)≥αp+(1−α)
log
x∈X
dP
(25)
(x) dS(x)
dS
This is the restricted partial likelihood of order p, conditional to the fact that the prior distribution Pprior contains no
outliers.
This can be recast as:
dpN e
α X B
H δxo(n) |P
lp (P |S) = −
dpN e n=1
This is a penalized form of the partial likelihood as expressed in [8]. Thus the partial estimation procedure introduced in [8] can be directly extended to such regularized
estimators.
A. Restricted partial maximum likelihood principle for regularized estimators
We shall now apply the partial estimation procedure to the
different regularized estimators introduced so far.
Outlined below is the algorithmic procedures corresponding
to the partial version of the regularized Tyler estimator for a
distribution in Rd :
function REG P T YLER(Rprior , (xn )1≤n≤N , α, p, , Kmax )
p
2:
R 21 ← Rprior
3:
R− 12 ← R 21 −1
4:
for k from 1 to Kmax do
5:
6:
7:
8:
9:
for n from 1 to N do
τn ← xn † R− 12 R− 12 xn
end for
o ← argsort↑ (τn )1≤n≤N
S ← (1 − α)anscm R− 12 Rprior R− 12
dpN e
X xo(n) xo(n) †
α
1
+
R− 2
R− 21
dpN e
τo(n)
n=1
10:
S←
d
tr(S) S
else
R 12 ←
√
R
R− 21 ← R 12 −1
end if
end for
20:
return R− 21
21:
end function
The corresponding partial procedure for a distribution in
Cd is the same as Tyler but using a different α parameter:
(26)
− (1 − α)H(Pprior |[P ])
1:
break
REG P T YLER (Rprior ,(xn )1≤n≤N ,
α(d−1)
,p,)
d− 21 (1+α)
As stated in III-B, it is in general quite difficult to derive
regularized estimators for M-estimators which are consistent
with maximum likelihood theory as stated in II-Aexcept
for some specific cases. The same remains valid for partial
regularized estimates, for which we shall again concentrate
on the specific case of gaussian models. The corresponding
partial regularized estimators, corresponding to the partial
versions of the regularized sample covariance in the real case
and the REG CG COV algorithm in the complex circular case
are outlined below:
1:
function REG PARTIAL SCM(Σprior , (xn )1≤n≤N , α, Kmax )
2:
Σ ← Σprior
3:
for n from 1 to N do
4:
τn ← xn † Σ−1 xn
5:
end for
6:
o0 ← argsort↑ (xn )1≤n≤N
7:
for k from 1 to Kmax do
α
dpN e
8:
Σ ← (1 − α)Σprior +
9:
for n from 1 to N do
10:
dpN
Pe
n=1
τn ← xn † Σ−1 xn
11:
end for
12:
o1 ← argsort↑ (τn )1≤n≤N
xo0 (n) xo0 (n) †
ARXIV PREPUBLICATION
13:
14:
15:
16:
17:
6
if o0 = o1 then
the background noise is generated as a white gaussian noise
of unit variance.
The target signal is generated as a complex centered circular
1-dimensional gaussian signal aligned with the test signal s,
whose variance σ is such that:
break
else
o0 ← o1
end if
18:
end for
19:
return Σ
20:
end function
10 log10 (σ) = SiNR
In the complex circular case the corresponding algorithm
is given by:
1:
2:
3:
function REG PCG COV(Σprior (xn )1≤n≤N , α, , Kmax )
Σ ← (1 − α)Σprior +
√
Σ 12 ← Σ
α
N
N
P
Σ− 12 ← Σ 12 −1
5:
for k from 1 to Kmax do
7:
for n from 1 to N do
end for
9:
o ← argsort↑
11:
12:
13:
14:
15:
16:
17:
S←
1
α
(1− 2d
)
τn −
1
2
log (τn ) 1≤n≤N
This covariance/correlation matrix is then used in an
adaptive detection scheme.
(1 − α)Σprior
dpN e
α X
1
†
+
xo(n) xo(n)
1−
dpN e n=1
2τo(n)
R ← Σ 12 exp Σ− 12 SΣ− 12 Σ 12
2
if tr Σ− 21 ΣΣ− 12 − I
≤ then
The performances are considered in the asymptotic regime
(meaning that Σm follows the same distribution as Σm−1 for
any m ∈ Z). The performances are shown for N = 11, p = 34
for partial estimators, with various values of α; the results are
also compared to the non regularized versions of the estimators
for N = 22 [7][8].
break
else
√
Σ 21 ← Σ
Σ− 12 ← Σ 21 −1
end if
18:
end for
19:
return Σ− 21
20:
In this scenario, we have an infinite stream of learning
samples (xn,m )1≤n≤N,m∈Z . These learning samples are used
to learn the covariance/correlation matrix of the signal sequentially, using the following pattern:
Σm = reg(Σm−1 , (xn,m )1≤n≤N )
τn ← xn † Σ−1 xn
8:
10:
xn xn †
n=1
4:
6:
The detection thresholds are defined to have a false alarm
rate of 10−4 ; they are learned on clean training sets, that is
that they contain no outliers.
end function
V. S IMULATIONS
We shall now show some simulation results of adaptive
detectors using various estimators introduced in this article,
using the detection tests introduced in [7][24].
The simulation results are shown in a single channel scenario of dimension d = 8, as a function of the SiNR of the
target signal.
Fig. 1. Detection capabilities of Tyler-NMF and pTyler-NMF
ARXIV PREPUBLICATION
7
presence of outliers, by using the partial procedure defined in
[8].
Finally the potential gain in terms of detection performances
in adaptive detection schemes is shown on some simulations.
A PPENDIX A
O N THE RELATIVE ENTROPY OF ANGULAR MODELS
The relative entropy of two angular models of Rd , characterized by their correlation matrix R0 and R (with the convention
that |R0 | = |R| = 1), is given by:
H(R0 |R) =
† −1
Z
x R x
d−1
log
dsd−1
2
x† R0 −1 x
x∈Sd−1
Fig. 2. Detection capabilities of reg SCM-MF and reg pSCM-MF
L(R )−1 x
p 0
x† R0 −1 x
!
(27)
Although the exact computation of this integral represents
a challenge, what is of interest to us is its differential with
respect to the matrix R:
dH(R0 |R) =
Z
d−1
tr
2
xx
†
−1
x† R−1 x
dsd−1
x∈Sd−1
L(R ) x
p 0
x† R0 −1 x
!
d(R−1 )
(28)
Thus we have to compute:
xx†
dsd−1
x† R−1 x
Z
Σ=
x∈Sd−1
L(R )−1 x
p 0
x† R0 −1 x
!
(29)
1
By making a change of variable y = R− 2 x, this integral
can be recast as:
†
− 12 −1
− 21
yy
R
R
)
y
L(R
0
R 12
Σ=R
dsd−1 q
†
1
1
y∈Sd−1 y y
y † R− 2 R0 R− 2 y
(30)
Let us note for any symmetric positive matrix Σ:
1
2
As can be seen on figures 1, 2 and V, the use of a regularized
estimate in this case allows to greatly improve the detection
capabilities of adaptive tests. Indeed the performances quickly
approach the asymptotic limit in which the covariance matrix
is perfectly estimated when a slower integration factor is
chosen.
Moreover the degradation of performances of partial estimators can be almost completely negated for a slow enough
integration (α ≤ 14 ), thus resulting in estimators which offer
performances close to non-adaptive detectors, while retaining
the robustness to outliers offered by partial estimators.
VI. C ONCLUSION
We have presented a theoretical background for a likelihood
based approach to regularized estimations. This background
allows to define penalties used in penalized maximum likelihood schemes in a way which is consistent with the maximum
likelihood principle.
This maximum likelihood principle for regularized estimation
is then further extended to allow for robust estimators in the
Z
Z
anscm(Σ) =
y∈Sd−1
yy †
dsd−1
y† y
L(Σ)−1 y
p
y † Σ−1 y
!
(31)
†
This is the integral of yy
on the unit ellipse defined by the
y† y
correlation matrix Σ, and therefore corresponds to the expected
value of the normalized sample covariance estimator (NSCM)
for an elliptical distribution of correlation matrix proportional
to Σ. The structure of this matrix has been studied thoroughly
in [3] for complex circular distributions, in the case in which
all eigenvalues are distinct:
anscm(Σ) = U diag(µ)U †
(32)
with:
•
Σ = U diag(λ)U † with U U † = I being the eigen
decomposition of Σ
ARXIV PREPUBLICATION
•
µ = Ec with:
E
=
c=
8
λ
log( λj )
i
λj
λi −1
−
λi
λj
if i 6= j
if i = j
1≤i,j≤d
0
!
1
λk
k6=j 1− λj
Q
1≤j≤d
Note that in a practical implementation, the limitation of
having distinct eigenvalues is not restrictive, as it is almost
surely the case. However one should take steps to insure a
decent numerical stability. We have on our end opted for
enforcing a minimum spacing between eigenvalues determined
by a numerical tolerance, as given by the following algorithm:
1:
2:
function R ESPACE
((λk )1≤k≤d
, )
o ← argsort↑ (λk )1≤k≤d
3:
(µk )1≤k≤d ← (λk )1≤k≤d
4:
for k from 1 to d − 1 do
if
5:
µo(k+1)
µo(k)
− 1 < then
for m from k + 1 to d do
6:
µo(m) ← (1 + )µo(m−1) − µo(m)
7:
end for
8:
10:
end if
end for
11:
return (µk )1≤k≤d
9:
12:
end function
Unfortunately we do not have a similar result in the general
case of distributions in Rd , in which case one has to resort
to numerical integration methods. Indeed the matrix can be
expressed as anscm U diag(λ)U † = U diag(µ)U † , with:
µk =
ZZ
2 Γ d2
√
π Γ d−1
2
t∈[0;1]
x∈Sd−2
d
t2 (1 − t2 ) 2
P
t2 + (1 − t2 )
−1
m6=k
λm
2
λk xk
dsd−2 (x)dt
(33)
There is no known analytical expression of these eigenvalues to the authors’ knowledge.
R EFERENCES
[1] John Aldrich et al. Ra fisher and the making of maximum likelihood
1912-1922. Statistical Science, 12(3):162–176, 1997.
[2] Frédéric Barbaresco. Super-resolution spectrum analysis regularization:
Burg, capon & ago-antagonistic algorithms. In European Signal Processing Conference, 1996. EUSIPCO 1996. 8th, pages 1–4. IEEE, 1996.
[3] Sebastien Bausson, Frederic Pascal, Philippe Forster, Jean-Philippe
Ovarlez, and Pascal Larzabal. First-and second-order moments of the
normalized sample covariance matrix of spherically invariant random
vectors. IEEE Signal Processing Letters, 14(6):425–428, 2007.
[4] Peter J Bickel and Elizaveta Levina. Covariance regularization by
thresholding. The Annals of Statistics, pages 2577–2604, 2008.
[5] MA Chmielewski. Elliptically symmetric distributions: a review and
bibliography. International Statistical Review/Revue Internationale de
Statistique, pages 67–74, 1981.
[6] Philippe Ciuciu, Jérôme Idier, and Jean-François Giovannelli. Regularized estimation of mixed spectra using a circular Gibbs-Markov model.
IEEE Transactions on Signal Processing, 49(10):2201–2213, oct. 2001.
[7] Christophe Culan and Claude Adnet. Maximum likelihood estimation of covariances of elliptically symmetric distributions. eprint,
arXiv:1611.04365, 2016.
[8] Christophe Culan and Claude Adnet. Partial estimators and application
to covariance estimation of gaussian and elliptical distributions. eprint,
arXiv:1611.07418, 2016.
[9] Junbo Duan, Charles Soussen, David Brie, Jérôme Idier, Mingxi Wan,
and Yu-Ping Wang. Generalized LASSO with under-determined regularization matrices. Signal Processing, 127:239–246, 2016.
[10] John Duchi. Derivations for linear algebra and optimization. Berkeley,
California, 2007.
[11] Jean-François Giovannelli, Jérôme Idier, Guy Desodt, and Daniel Muller.
Regularized adaptive long autoregressive spectral analysis. IEEE Transactions on Geoscience and Remote Sensing, 39(10):2194–2202, oct.
2001.
[12] Anders Hald. On the history of maximum likelihood in relation to
inverse probability and least squares. Statistical Science, pages 214–
222, 1999.
[13] Joseph P Hoffbeck and David A Landgrebe. Covariance matrix estimation and classification with limited training data. IEEE Transactions on
Pattern Analysis and Machine Intelligence, 18(7):763–767, 1996.
[14] Reshad Hosseini, Suvrit Sra, Lucas Theis, and Matthias Bethge. Inference and mixture modeling with the elliptical gamma distribution.
Computational Statistics & Data Analysis, 101:29–43, 2016.
[15] Solomon Kullback and Richard A Leibler. On information and sufficiency. The annals of mathematical statistics, 22(1):79–86, 1951.
[16] Ali Mohammad-Djafari, Jean-François Giovannelli, Guy Demoment,
and Jérôme Idier. Regularization, maximum entropy and probabilistic
methods in mass spectrometry data processing problems. Int. Journal
of Mass Spectrometry, 215(1-3):175–193, avr. 2002.
[17] Esa Ollila and David E Tyler. Regularized-estimators of scatter matrix.
IEEE Transactions on Signal Processing, 62(22):6059–6070, 2014.
[18] Esa Ollila, David E Tyler, Visa Koivunen, and H Vincent Poor. Complex
elliptically symmetric distributions: Survey, new results and applications.
Signal Processing, IEEE Transactions on, 60(11):5597–5625, 2012.
[19] Frédéric Pascal, Yacine Chitour, Jean-Philippe Ovarlez, Philippe Forster,
and Pascal Larzabal. Covariance structure maximum-likelihood estimates in compound gaussian noise: Existence and algorithm analysis.
Signal Processing, IEEE Transactions on, 56(1):34–48, 2008.
[20] Johann Pfanzagl. Parametric statistical theory. Walter de Gruyter, 1994.
[21] John W Pratt. Fy edgeworth and ra fisher on the efficiency of maximum
likelihood estimation. The Annals of Statistics, pages 501–514, 1976.
[22] Pradeep Ravikumar, Martin J Wainwright, Garvesh Raskutti, Bin Yu,
et al. High-dimensional covariance estimation by minimizing 1penalized log-determinant divergence. Electronic Journal of Statistics,
5:935–980, 2011.
[23] Louis L Scharf. Statistical signal processing, volume 98. AddisonWesley Reading, MA, 1991.
[24] Louis L Scharf and Benjamin Friedlander. Matched subspace detectors.
Signal Processing, IEEE Transactions on, 42(8):2146–2157, 1994.
[25] Stephen M Stigler. Statistics on the table: The history of statistical
concepts and methods. Harvard University Press, 2002.
[26] Ying Sun, Prabhu Babu, and Daniel P Palomar. Regularized tyler’s scatter estimator: Existence, uniqueness, and algorithms. IEEE Transactions
on Signal Processing, 62(19):5143–5156, 2014.
[27] David E Tyler. Statistical analysis for the angular central gaussian
distribution on the sphere. Biometrika, 74(3):579–589, 1987.
[28] Ami Wiesel. Geodesic convexity and covariance estimation. Signal
Processing, IEEE Transactions on, 60(12):6182–6189, 2012.
[29] Ami Wiesel. Unified framework to regularized covariance estimation
in scaled gaussian models. IEEE Transactions on Signal Processing,
60(1):29–38, 2012.
[30] Daniela M Witten and Robert Tibshirani. Covariance-regularized regression and classification for high dimensional problems. Journal of the
Royal Statistical Society: Series B (Statistical Methodology), 71(3):615–
636, 2009.
[31] Lingzhou Xue, Shiqian Ma, and Hui Zou. Positive-definite 1-penalized
estimation of large covariance matrices. Journal of the American
Statistical Association, 107(500):1480–1491, 2012.
ARXIV PREPUBLICATION
Christophe Culan was born in Villeneuve-St-Georges, France, on November
1st , 1988. He received jointly the engineering degree from Ecole Centrale de
Paris (ECP), France and a master’s degree in Physico-informatics from Keio
University, Japan, in 2013.
He was a researcher in applied physics in Itoh laboratory from 2013 to
2014, specialized in quantum information and quantum computing, and has
contributed to several publications related to these subjects.
He currently holds a position as a research engineer in Thales Air Systems,
Limours, France, in the Advanced Radar Concepts division. His current research interests include statistical signal and data processing, robust statistics,
machine learning and information geometry.
Claude Adnet was born in A, France, in 1961. He received the DEA degree
in signal processing and Phd degree in 1988 and 1991 respectively, from the
Institut National Polytechnique de Grenoble (INPG), Grenoble France. Since
then, he has been working for THALES Group, where he is now Senior
Scientist. His research interests include radar signal processing and radar data
processing.
9
| 10math.ST
|
M ULTI - MODAL G EOLOCATION E STIMATION
D EEP N EURAL N ETWORKS
USING
arXiv:1712.09458v1 [cs.CV] 26 Dec 2017
Jesse M. Johns, Jeremiah Rounds & Michael J. Henry
Pacific Northwest National Laboratory
Richland, WA 99352, USA
{jesse.johns,jeremiah.rounds,michael.j.henry}@pnnl.gov
A BSTRACT
Estimating the location where an image was taken based solely on the contents of
the image is a challenging task, even for humans, as properly labeling an image
in such a fashion relies heavily on contextual information, and is not as simple as
identifying a single object in the image. Thus any methods which attempt to do so
must somehow account for these complexities, and no single model to date is completely capable of addressing all challenges. This work contributes to the state of
research in image geolocation inferencing by introducing a novel global meshing
strategy, outlining a variety of training procedures to overcome the considerable
data limitations when training these models, and demonstrating how incorporating additional information can be used to improve the overall performance of a
geolocation inference model. In this work, it is shown that Delaunay triangles are
an effective type of mesh for geolocation in relatively low volume scenarios when
compared to results from state of the art models which use quad trees and an order
of magnitude more training data. In addition, the time of posting, learned user
albuming, and other meta data are easily incorporated to improve geolocation by
up to 11% for country-level (750 km) locality accuracy to 3% for city-level (25
km) localities.
1
I NTRODUCTION
Advancements in deep learning over the past several years have pushed the bounds of what is possible with machine learning beyond typical image classification or object localization. Whereas prior
work in computer vision and machine learning has focused primarily on determining the contents
of the image–“is there a dog in this photo?”—more recent methods and more complicated models
have enabled deeper examinations of the contextual information behind an image. This deeper exploration allows a researcher to ask more challenging questions of its models. One such question is,
“where in the world was this picture taken?”
However, estimating the geographic origin of a ground-level image is a challenging task due to a
number of different factors. Firstly, the volume of available images with geographic information
associated with them is not evenly distributed across the globe. This uneven distribution increases
the complexity of the model chosen and complicates the design of the model itself. Furthermore,
there are additional challenges associated with geo-tagged image data, such as the potential for
conflicting data (eg. incorrect geolabels and replica landmarks - St Peters Basilica in Nikko, Japan)
and the ambiguity of geographic terms (eg. imagery of similar and ambiguous geological features,
such as beaches, screenshots of websites, or generally, plates of food).
The work presented herein focuses on the task of content-based image geolocation: the process of
identify the geographic origin of an image taken at ground level. Given the significant growth in the
production of data, and trends in social data to move away from pure text-based platforms to more
images, videos, and mixed-media, tackling the question of inferring geographic context behind an
image becomes a significant and relevant problem. Whereas social textual data may be enriched
with geolocation information, images may or may not contain such information; EXIF data is often
stripped from images or may be present but incomplete. Within the context of mixed-media data,
using the geolocation information associated with a linked object may be unreliable or not even
1
applicable. For example, a user on Twitter may have geolocation information enabled for their
account, but may post a link to an image taken from a location different from the one where they
made their post.
1.1
R ELATED W ORK
A wide variety of approaches have been considered for geolocation from image content, please
consult (Brejcha & Čadı́k, 2017) for a review. The approach taken in this paper builds on recent
work in global geolocation from ground-based imagery (Weyand et al., 2016) for global scale geoinferencing. Weyand et al. (2016) utilize a multi-class approach with a one-hot encoding. A common
approach, as well, is to perform instance-level scene retrieval in order to perform geolocation of
imagery (Hays & Efros, 2008; 2015; Vo et al., 2017). These works query on previously geotagged
imagery and assign a geolabel based on the similarity of the image query to the database. Further,
Vo et al. (2017) builds on work by Hays & Efros (2008; 2015); Weyand et al. (2016) by utilizing the
feature maps for the mesh-based classifier for the features of their nearest-neighbors scene retrieval
approach.
Prior work exists for data sampling strategies for large-scale classification problems for social media
applications. Kordopatis-Zilos et al. (2016) considers weighted sampling of minority classes. In
this work, the class selection is biased during training of the deep learning models. Biasing class
selection with random noising (often called image regularization) is a well known way to allow a
model to see more examples of rare classes, but also there are additional concerns related specifically
to social media applications. For example, researchers in Kordopatis-Zilos et al. (2016) consider the
case of sampling such that an individual user is only seen once per a training epoch. In this work,
sampling is performed without respect to user, so that images are selected in epochs completely at
random, but the broader influence of latent variables, other than user, and communities is of concern
in social media geolocation.
1.2
C ONTRIBUTIONS
The first type of model considered in this work is purely geolocation for image content (M1). The
use of time information forms a second set of models within this work (M2). User-album inputs
form a third set of models within this work (M3). Our work contributes meaningful consideration of
the use of an alternative mesh for geolocation in M1 and demonstrates the meaningful use of time
and user information to improve geolocation (M2 and M3).
2
G EOLABELED DATA
Collected data holdings are derived from YFCC100M (Thomee et al., 2016), where training and
validation was performed on a randomly selected 14.9M (12.2M training/2.7M validation) of the
48.4M geolabeled imagery. PlaNet, in comparison, used 125M (91M training/34M validation) images (Weyand et al., 2016) for model developemt. In using the data, it is assumed that the ground
truth GPS location is exactly accurate, which is a reasonable approximation based on the results of
Hauff (2013).
Every YFCC100M image has associated meta data. Importantly, these data contain user-id and
posted-time. User-id is a unique token that groups images by account. Posted time is the time that
the user uploaded data to the website that ultimately created the combined storage of all images.
The most likely true GPS location for an image from varies by the time the image is uploaded to a
website, as shown in Figure 1.
3
3.1
M ETHODOLOGY
CNN C LASSIFICATION (M ODEL M1)
The approach described in this work for the global-scale geolocation model is similar to the classification approach taken by PlaNet (Weyand et al., 2016). Spatially the globe is subdivied into a grid,
forming a mesh, decribed in Section 3.1.1, of classification regions. An image classifier is to trained
2
Figure 1: True Longitude versus Time of Posting. In these boxplots, we limit users to one image
selected at random to decouple any user effects from time of day. What this image implies is that
the prior distribution for image longitude differs by the time of day that an image is posted. As an
example, consider longitudes near 0. It would take less evidence to predict a longitude close to zero
for an image posted around 01:00 UTC than to make a similar prediction at 21:00 UTC because
of the observed priors. At 21:00 UTC, the boxplot indicates images are typically as far from zero
longitude as will been seen in the data.
to recognize the imagery whose ground truth GPS is contained in a cell and hence, during inference,
the mesh ascribes a probability distribution of geo-labels for the input imagery.
3.1.1
M ESH
The classification structure is generated using a Delaunay triangle-based meshing architecture. This
differs from the PlaNet approach, which utilizes a quad-tree mesh. Similarly to PlaNet, the mesh
cells are generated such that they conserve surface area, but the approach is not as simple as with a
quad-tree. The triangular mesh was deployed under the hypothesis that the Delaunay triangles would
be more adaptive to the geometric features of the Earth. Since triangular meshes are unstructures,
they can more easily capture water/land interfaces without the additional refinement needed by quadtree meshes. However, the triangular mesh loses refinement level information which comes with a
structured quad-tree approach, which would allow for granularity to be controlled more simplistically, based which cells contain other cells. In order to control mesh refinment, cells are adapatively
refined, or divided, when the cell contains more than some number of examples (refinement limit) or
the mesh cell is dropped from being classified by the model (the contained imagery is also dropped
from training) if the cell contains less than a number of samples (minimum examples). The options
used for generating the three meshes in this paper are shown in Table 1. Parameters for the initialization of the mesh were not explored; each mesh was initialized with a 31 x 31 structured grid with
equal surface area in each triangle.
The mesh for the a) coarse mesh and b) fine mesh are shown in Figure 3.1.1.
Table 1: The geolocation classification mesh can be tuned by modifying the refinement level to
adjust the maximum and minimum number of cells in each mesh cell. This table shows the meshing
parameters used for the three meshes studied in this work. These meshes were selected to cover
a reasonable range of mesh structures, with fineP meant to replicate the mesh parameters utilized
by PlaNet. However, it should be noted that this isn’t direcly comparable since the fineP mesh is
generated with both a different methodology and a different dataset.
Mesh
Coarse
Fine
FineP
Refinement Limit
Minimum Examples
8000
5000
10000
1000
500
50
3
Cell Count
538
6565
4771
Figure 2: Coarse mesh and fine mesh classification structure for classification models developed in
this paper. The red triangles indicate regions where the number of images per cell meets the criteria
for training. The coarse mesh was generated with an early dataset of 2M YFCC100M images,
but final training was performed by populating those cells with all training data without additional
refinement. The fine and fineP meshes were generated with all available imagery (14M). The fine
mesh does a better job at representing geographic regions, for example, the water/land interface of
Spain and Portugal becomes evident. However, this is masked with the coarser mesh, which lacks
fidelity to natural geographic boundaries.
3.1.2
M ODEL
The Inception v4 convolutional neural network architecture proposed in (Szegedy et al., 2016) is
deployed to develop the mesh-based classification geolocation model (Model M1) presented in this
work, which differs from the Inception v3 architecture used in PlaNet (Weyand et al., 2016). A
softmax classification is utilized, similar to the PlaNet approach. One further difference between the
work presented here and that of PlaNet is that cells are labeled based the cell centroid, in latitudelongitude. Here an alternate approach is used where the “center-of-mass” of the training data in
a containing cell is computed, whereby the geolocation for each cell was specified as the lat/lon
centroid of the image population. Significant improvements are expected for coarse meshes, as it
can be noticed in Figure 3.1.1 for example, that the cell centroid on the coast of Spain is out in the
ocean. Therefore, any beach image, as an example, will have an intrinsically higher error that would
be otherwise be captured by a finer mesh. This is expecially true for high density population regions.
3.1.3
M ODEL E VALUATION
Models are evaluated by calculating the distance between the predicted cell and the ground-truth
GPS coordinate using a great-circle distance:
d = R × arcos (sin(x1 )sin(x2 ) + cos(x1 )cos(x2 )cos(y2 − y1 ))
where xi and yi is the latitude and longitude, respectively, in radians, and R is the radius of the Earth
(assumed to be 6372.795 km). An error threshold here is defined as the number of images which are
classified within a distance of d. Error thresholds of 1 km, 25 km, 200 km, 750 km, 2500 km are
utilized to represent street, city, region, country, and continent localities to remain consistent with
previous work (Hays & Efros, 2008; 2015; Weyand et al., 2016; Vo et al., 2017). As an example, if
1 out of 10 images are geolabeled with a distance of less than 200 km, then the 200 km threshold
would be 10%.
3.2
G EOLOCATION I MPROVEMENT WITH T IME M ETA DATA (M ODEL M2)
Every YFCC100M image has associated meta data; importantly these data consist of user id and
posted time. Posting time is utilized in model M2, Figure 3. Let Zik be the one-hot encoding of
the ith image for the kth class, so that it takes on value 1 if image i belongs to the kth geolocation
class. P (Zik = 1|Xi ) is assumed not equal to P (Zik = 1|Xi , ti ), where ti is the time of posting
of the ith image. Evidence for this fact is of the form of observed true GPS longitude shown as
changing distribution with respect to posting time (Hour of Day UTC) as shown in Figure 1. Figure
1 is only strong evidence for P (Zik = 1|ti ) 6= P (Zik = 1), and it could still be the case that
P (Zij = 1|Xi ) = P (Zij = 1|Xj , tj ). Which is to say, conditioned on the content of an image
there could be no dependence on time, but it seems prudent with the evidence in this figure to
proceed under the assumption that there is time dependence. The operational research hypothesis
for this model (M2) is that there remains a time-dependence after conditioning on image content.
4
Table 2: M2: Time-Adjusting Geolocation Model. N is the length of p̂j , which takes on two values:
538 (coarse mesh) and 6565 (fine mesh). All layers except the last are dense layers with 50% dropout and batch normalization.
Layer Type
Number Neurons Activation
1
Input
N+4
ReLU
2
Hidden 542*3
Sigmoid
3
Hidden 542*3
ReLU
4
Hidden 542*3
Sigmoid
5
Output N
Softmax
To incorporate time, related variables are appended to the output of the geolocation model (M1) to
form a new input for M2. Every image has a vector of fit probabilities p̂i from the softmax layer of
M1. p̂i is filtered so that only the top 10 maximum entries
P remain, and the other entries are set to
0 (p̂0i ). This vector is normalized, so that , p̂1ik = p̂0ik / l p̂0il . The L1-norm (denominator) in the
normalization |p̂0i | is appended as a feature. Time of posting is turned into a tuple of numeric input
as Ti = (Hour(ti ) − 11.5)/24, (Day of Week(ti ) − 3.5)/7, (Month(ti ) − 6.5)/12, where ti is the
time of publishing, Hour(ti ) is a function that returns hour of the day (0 to 23), Day of Week is a
function that returns 1 to 7 for numbered days of the week starting with Sunday as 1, and Month(ti )
is a function that returns the month of year starting with January as 1. Geolocation uses as input
for each image, the concatenated vector (p̂1i , |p̂0i |, Ti ). Year is specifically omitted from time inputs
because this would not generalize to new data posted in the future.
3.3
U SER -A LBUM G EOLOCATION R EFINEMENT (M ODEL M3)
Model M3, Figure 3, simultaneously geolocates many images from a single user with an LSTM
model. The idea here is to borrow information from other images a user has posted to help aid
geolocation.The Bidirectional LSTMs capitalizes on correlations in a single user’s images. LSTMs
were also considered by Weyand et al. (2016), but in PlaNet the albums were created and organized
by humans. When a human organizes images they may do so by topic or location. In M3, all
images by a single user are organized sequentially in time with no particular further organization.
The related research question is: does the success observed by Weyand et al. (2016) extend to this
less informative organization of images? All images from a user are grouped into albums of size 24.
If there are not enough images to fill an album then albums are 0 padded and masking is utilized.
During training, a user was limited to a single random album per epoch.
Album averaging was considered by Weyand et al. (2016) (PlaNet). In album averaging, the images
in an album are assigned a mesh cell based on the highest average probability across all images. This
method increases accuracy by borrowing information across related images. As a control, a similar
idea is applied to user images, in which, the location of an image is determined as the maximum
average probability across all images associated with the posting user. This result assumes that all
images from a user are from the same mesh grid cells. In addition with user-averaging, there is no
optimization that controls class frequencies to be unbiased.
Finally, LSTM on a time-ordered sequence of images was considered (without respect to user).
However, we were unable to improve performance significantly past that gained by just adding time
to the model, so albums without user are not further considered in this paper.
4
R ESULTS
4.1
4.1.1
M ODEL M1 C LASSIFICATION E XPERIMENTS
M ESH C OMPARISON
Meshing parameters are investigated to understand the sensitivity to mesh adaptation. The results
for each mesh is shown in Table 3. There is an apparent trade-off between fine-grain and coarsegrain geolocation. The coarse mesh demonstrates improved large granularity geolocation and the
fine mesh performs better at finer-granularities, as would be expected. This observation was also
noticed in (Vo et al., 2017). In addition, the impact of classifying on the centroid of the training data
5
Figure 3: Album model with component details added. Operator 1 is a filtering and re-normalization
of M1: Geolocation output. At operator (1) the output of M1 is filtered to output only the top 10
mesh cell probabilities (making it mostly sparse) re-normalized to sum to 1. Training of M2 and
M3 was only done on the validation data of M1 (using a new random split of test-validation). Time
inputs are concatenated to filtered and normalized outputs at (1) and a new training step is implied.
A small abuse of notation is present: M3 (Time-Albums) is properly described as M2 concatenated
with M3 layers.
is compared to utilizing the cell centroid for class labels. A dramatic improvement is noticed for the
coarse mesh, with only a modest improvement for the fine mesh.
Table 3: Geolocation locality error thresholds on the YFCC100M validation hold-out set comparing
both imagery centroid and cell centroid classification structures.
Mesh
Coarse
Coarse
Fine
Fine
FineP
FineP
4.1.2
Imagery Centroid
Street
1 km
City
25 km
Region
200 km
Country
750 km
False
True
False
True
False
True
0.42
1.25
3.20
4.44
1.86
3.23
6.89
10.57
9.56
10.49
8.23
9.47
21.22
23.99
16.64
16.63
16.23
16.69
38.53
38.72
29.49
29.57
30.98
31.07
Continent
2500 km
56.64
56.66
49.17
49.13
50.97
50.98
O UTDOOR I MAGERY
An experiment was conducted to determine if the performance of the M1 model could be improved
by applying the geolocation model on outdoor imagery only. From Thomee et al. (2016) it would be
expected that approximately 50% of the geolabeled YFCC100M imagery contains outdoor imagery.
A Place365 (Zhou et al., 2016) model was used in conjunction with indoor-outdoor label deliminations to filter geolocation inference only on outdoor imagery. Note that the geolocation model
was not re-trained on the “outdoor” imagery, this is only conducted as a filtering operation during
inference. Results are shown in Table 4. In general, the improvement is quite good, about a 4-8%
improvement in accuracy for region/country localities, with a more modest boost in smaller-scale
regions.
4.1.3
IM2GPS T EST R ESULTS
The Im2GPS testing data is utilized to test the model on the 237 images provided by the work
of Hays & Efros (2008). Results are tabluated in Table 5 for all of the meshes. The imagery
centroid classification labels are generated with the YFCC100M training data, yet the performance
6
Table 4: Geolocation locality error thresholds on the YFCC100M validation hold-out set comparing
inference on outdoor imagery only.
Mesh
Coarse
Coarse
Fine
Fine
FineP
FineP
Outdoor Only
Street
1 km
City
25 km
Region
200 km
Country
750 km
False
True
False
True
False
True
1.25
1.46
4.44
4.90
3.23
3.57
10.57
12.15
10.49
12.17
9.47
10.82
23.99
29.22
16.63
19.96
16.69
19.85
38.72
46.23
29.57
34.84
31.07
36.2
Continent
2500 km
56.66
63.50
49.13
54.33
50.98
55.94
is still greatly improved when applied to the Im2GPS testing set, demostrating generality with the
approach. The performance of the M1 classification model is comparable to the performance of
(Weyand et al., 2016) with a factor of ten less training data and far fewer classification regions (6k
compared to 21k); the coarse mesh M1 model exceeds the performance of PlaNet for large regions
(750 and 2500 km).
Table 5: Geolocation locality error thresholds on the Im2GPS testing data comparing both imagery
centroid and cell centroid classification structures. Bold indicates the best of the models developed in this work. The work of Weyand et al. (2016) used 125M images for model development
(91M training/34M validation) and Vo et al. (2017) is an instance retrieval based method, utilizing a
PlaNet-inspired mesh-based classifier for feature extraction which was trained on 9M Flickr images.
Mesh
Coarse
Fine
FineP
Coarse
Fine
FineP
Hays & Efros (2008)
Hays & Efros (2015)
Weyand et al. (2016)
Vo et al. (2017)
4.2
Imagery Centroid
Street
1 km
City
25 km
Region
200 km
Country
750 km
FALSE
FALSE
FALSE
TRUE
TRUE
TRUE
0.71
4.57
3.27
2.89
7.53
4.75
11.35
20.41
16.79
18.47
22.42
18.73
12.0
21.9
24.5
33.3
29.32
32.78
28.87
35.53
33.82
30.20
15.0
32.1
37.6
44.3
55.62
49.95
46.01
56.17
50.96
46.23
23.0
35.4
53.6
57.4
2.5
8.4
12.2
Continent
2500 km
72.02
67.06
64.04
71.99
67.07
64.10
47.0
51.9
71.3
71.3
T IME A DJUSTED G EOLOCATION AND A LBUM R ESULTS (M2 AND M3)
Use of time improves geolocation in two ways. There is a slight gain of accuracy for using time
as indicated in Tables 4 and 6. 24.20% of images are geolocated to within 200 km as opposed to
23.99% without using time with the coarse mesh, and 12.28% versus 10.49% are within 25 km
using the fine mesh. This small persistent advantage can be seen across all error calculations and is
statistically significant.
There is a measurable difference between the error of the coarse mesh using time (M2) and not using
time (M1). There exist a matched pair of errors for each image: e1i and e2i , where i is a validation
image index. The first superscript being the M1 error and the second superscript being M2 error for
image i in km. A null hypothesis is µ1 = µ2 where µ1 is the mean of the unknown distribution of e1i
and likewise for µ2 . This hypothesis is tested with a Wilcoxon-Signed-Test for paired observations,
which makes a minimal number of distributional assumptions. Specifically, normality of the errors
is not assumed. The difference µ̂2 − µ̂1 = 381 is highly significant (p-value < 10−16 ) in favor of
using time-inputs, so even though the effect of M2 is small, it is not explained by chance alone.
It is the case that the distribution of errors is mean shifted, but it is not uniformly shifted to lower
error, nor is it it the case that images are universally predicted closer to the truth. The median of
coarse mesh e1 is 1627 km while the median of e2 is 1262 km (Table 7).
7
Time input models appear to have lower-bias class probabilities. Cross-entropy was optimized
in training for both the classification model (M1) and time-inputs models (M2). In each training
method the goal is to minimize these class biases. KL-Divergence(p̂, q̂) is calculated for each model,
where p is the observed class proportions for the true labels in validation, and q is the observed class
proportions for the model predicted labels in validation (in both cases, 1 is added to the counts prior
to calculating proportions as an adjustment for 0 count issues). The KL-divergence of the model
output class frequencies compared to the true class frequencies in validation are in Table 7.
“User-Averaging” is incorporated into results because it is a simple method that appears to be more
accurate then predicting individual images with M1 or M2; however, it biases cell count frequency
(Table 7). In general when using the average probability vector to predict a user’s image, there is no
guarantee that the class frequencies are distributed similarly to the truth; thus, improved accuracy
can come with higher bias which is what is observed. Albums are a much better approach to borrow
information across users because built into the training method is a bias reducing cross-entropy
optimization, and indeed LSTMs on user albums had the lowest class bias of any model considered.
Table 6: Percentage accuracy to specific resolution: Researched models compared at various spatial
resolutions. Coarse and fine mesh have 538 and 6565 triangles across the globe (cells), respectively.
”Time inputs” indicates time meta information has been concatenated with M1 output. Albums are
created using user-id and contain 24 images. Bold is the best observed accuracy in column. Coarse
Mesh and Fine Mesh Best Possible are not actual models, but rather the best possible accuracy
if every image was given exactly the right class. In the fine mesh the best possible accuracy is
incidental, but in the coarse mesh it is a severe limitation for street level accuracy.
Model
Coarse Mesh Time Inputs (M2)
Coarse Mesh Time Inputs (M2)
Coarse Mesh User Averaging
Coarse Mesh User Albums (M3)
Coarse Mesh User Albums (M3)
Coarse Mesh Best Possible (cell centroid)
Coarse Mesh Best Possible (imagery centroid)
Fine Mesh Time Inputs (M2)
Fine Mesh Time Inputs (M2)
Fine Mesh Best Possible (cell centroid)
Fine Mesh Best Possible (imagery centroid)
Outdoor Only
False
True
False
False
True
False
False
False
True
False
False
Street
1 km
1.20
1.36
0.84
1.25
1.39
1.57
4.10
4.82
4.95
21.92
31.22
City
25 km
10.47
11.99
11.34
11.25
12.52
25.33
43.38
12.28
13.44
73.13
83.23
Region
200 km
24.20
29.28
27.30
25.65
30.24
81.97
93.86
18.44
21.17
97.26
99.09
Country
750 km
40.12
47.60
44.74
42.38
49.17
99.99
99.98
31.39
35.85
99.99
99.99
Continent
2500 km
60.66
66.69
62.84
63.17
68.61
100.00
100.00
52.25
56.31
100.00
100.00
Table 7: KL-Divergence of various models from observed class proportions in validation data.
Lower KL-Divergence is better, and 0 indicates distributional equivalence between model class frequencies and true class frequencies. Bold is best model observed by these measures. Entropy has
no ”best”, so it is not bolded.
Model
KL
Divergence
Shannon
Entropy
Mean
Average Error (km)
Median
Image Error (km)
0.6718
0.4998
0.9702
0.4991
0.0000
1.7120
0.3543
0.0000
5.6360
5.2890
5.0760
5.1900
6.0180
7.0460
8.2770
8.6090
3897
3516
3192
3320
62
4561
4311
17
1627
1262
1051
1124
33
2705
2140
4
Coarse Mesh Inception (M1)
Coarse Mesh Time Inputs (M2)
Coarse Mesh User Averaging
Coarse Mesh User Albums Time Inputs (M3)
Coarse Mesh Best Possible
Fine Mesh Inception (M1)
Fine Mesh Time Inputs (M2)
Fine Mesh Best Possible
5
C ONCLUSIONS
Conditioning on latent variables can only improve geolocation models. Universally using time of
day in the models was observed to increase accuracy and lower bias. Time of day is a weak addition
8
to Inception-like results, but it is useful to be as accurate as possible, and it makes statistically
significant improvements to geolocation. Both meshes were improved by using time information
(M2). This is a result that is not surprising, and as a research approach, can be applied to any
number of meta data variables that might accompany images.
Accounting for indoor/outdoor scenes in images explained variation in validation accuracy. Outdoor
only results are better than results for all images. We suggest as future work that the probability
that an image is outdoor could be concatenated to the input of M2. The accompanying research
hypothesis is that in the case that an image is indoors, perhaps the model will learn to weight time
or other meta data results more heavily, or otherwise be able to use that information optimally.
Increasing the granularity of a grid reduces accuracy at the country and regional level, while improving accuracy at street and city level accuracy. To be clear though, street level geoinferencing is
not practical with a course mesh. This is shown by the best possible accuracy in Table 6, and so it is
expected that a fine mesh would do better. On the other hand, there is no reason to assume that a fine
mesh has to be better for geolocation at larger resolutions than 25 km, nor is there any explicit way
to prove a fine mesh should do no worse than a course mesh. What we observe is that a course mesh
is a superior grid for 200 km resolutions. Furthermore, we show that for both the coarse mesh and
the fine mesh, using a Delaunay triangle-based mesh provides the ability to train accurate models
with far fewer training examples than what was previously published.
ACKNOWLEDGMENTS
This research was performed at Pacific Northwest National Laboratory, a multi-program national
laboratory operated by Battelle for the U.S. Department of Energy.
R EFERENCES
Yoshua Bengio and Jean-Sbastien Sencal. Quick Training of Probabilistic Neural Nets by Importance Sampling. 2003.
Jan Brejcha and Martin Čadı́k. State-of-the-art in visual geo-localization. Pattern Analysis and
Applications, 20(3):613–637, Aug 2017. ISSN 1433-755X. doi: 10.1007/s10044-017-0611-1.
URL https://doi.org/10.1007/s10044-017-0611-1.
Rich Caruana, Steve Lawrence, and C Lee Giles. Overfitting in neural nets: Backpropagation,
conjugate gradient, and early stopping. In Advances in neural information processing systems,
pp. 402–408, 2001.
John Duchi, Elad Hazan, and Yoram Singer. Adaptive Subgradient Methods for Online Learning
and Stochastic Optimization. J. Mach. Learn. Res., 12:2121–2159, July 2011. ISSN 1532-4435.
URL http://dl.acm.org/citation.cfm?id=1953048.2021068.
Yoav Freund and Robert E Schapire. A desicion-theoretic generalization of on-line learning and an
application to boosting. In European conference on computational learning theory, pp. 23–37.
Springer, 1995.
Claudia Hauff. A Study on the Accuracy of Flickr’s Geotag Data. In Proceedings of the 36th
International ACM SIGIR Conference on Research and Development in Information Retrieval,
SIGIR ’13, pp. 1037–1040, New York, NY, USA, 2013. ACM. ISBN 978-1-4503-2034-4. doi: 10.
1145/2484028.2484154. URL http://doi.acm.org/10.1145/2484028.2484154.
James Hays and Alexei A Efros. Im2gps: estimating geographic information from a single image.
In Computer Vision and Pattern Recognition, 2008. CVPR 2008. IEEE Conference on, pp. 1–8.
IEEE, 2008.
James Hays and Alexei A. Efros. Large-Scale Image Geolocalization. In Jaeyoung Choi and
Gerald Friedland (eds.), Multimodal Location Estimation of Videos and Images, pp. 41–62.
Springer International Publishing, 2015. ISBN 978-3-319-09860-9 978-3-319-09861-6. URL
http://link.springer.com/chapter/10.1007/978-3-319-09861-6_3. DOI:
10.1007/978-3-319-09861-6 3.
9
Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing
human-level performance on imagenet classification. CoRR, abs/1502.01852, 2015. URL http:
//arxiv.org/abs/1502.01852.
Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint
arXiv:1412.6980, 2014.
Giorgos Kordopatis-Zilos, Symeon Papadopoulos, and Yiannis Kompatsiaris. In-depth exploration
of geotagging performance using sampling strategies on yfcc100m. In Proceedings of the 2016
ACM Workshop on Multimedia COMMONS, pp. 3–10. ACM, 2016.
Andriy Mnih and Geoffrey E Hinton.
A Scalable Hierarchical Distributed Language Model.
In D. Koller, D. Schuurmans, Y. Bengio, and L. Bottou (eds.),
Advances in Neural Information Processing Systems 21, pp. 1081–1088. Curran Associates,
Inc.,
2009.
URL http://papers.nips.cc/paper/
3583-a-scalable-hierarchical-distributed-language-model.pdf.
Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng
Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei.
ImageNet Large Scale Visual Recognition Challenge. arXiv:1409.0575 [cs], September 2014.
URL http://arxiv.org/abs/1409.0575. arXiv: 1409.0575.
Yikang Shen, Shawn Tan, Chrisopher Pal, and Aaron Courville. Self-organized Hierarchical Softmax. arXiv:1707.08588 [cs], July 2017. URL http://arxiv.org/abs/1707.08588.
arXiv: 1707.08588.
Christian Szegedy, Sergey Ioffe, Vincent Vanhoucke, and Alex Alemi. Inception-v4, InceptionResNet and the Impact of Residual Connections on Learning. arXiv:1602.07261 [cs], February
2016. URL http://arxiv.org/abs/1602.07261. arXiv: 1602.07261.
Bart Thomee, David A. Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland,
Damian Borth, and Li-Jia Li. YFCC100m: The New Data in Multimedia Research. Communications of the ACM, 59(2):64–73, January 2016. ISSN 00010782. doi: 10.1145/2812802. URL
http://arxiv.org/abs/1503.01817. arXiv: 1503.01817.
Nam Vo, Nathan Jacobs, and James Hays. Revisiting IM2gps in the Deep Learning Era.
arXiv:1705.04838 [cs], May 2017. URL http://arxiv.org/abs/1705.04838. arXiv:
1705.04838.
Tobias Weyand, Ilya Kostrikov, and James Philbin. Planet-photo geolocation with convolutional
neural networks. In European Conference on Computer Vision, pp. 37–55. Springer, 2016.
Bolei Zhou, Aditya Khosla, Agata Lapedriza, Antonio Torralba, and Aude Oliva. Places: An Image
Database for Deep Scene Understanding. arXiv:1610.02055 [cs], October 2016. URL http:
//arxiv.org/abs/1610.02055. arXiv: 1610.02055.
10
A
T RAINING P ROCEDURE
Images were divided at random into training and validation sets of 12.2M and 2.7M images and
associated metadata, respectively. Validation data used for M1 was further sub-divided at random
into training and validations sets for training time-based models (M2 and M3), so that no data used
to train M1 was used to also train M2 and M3.
A.1
M ODEL M1
Historically, softmax classification is shown to perform quite poorly when the number of output
classes is large (Bengio & Sencal, 2003; Mnih & Hinton, 2009; Shen et al., 2017). During initial experiments with large meshes (> 4,000 mesh regions), a training procedure was developed
to circumvent these challenges. Empirically, this procedure worked for training the large models
presented in this paper; however, it is not demostrated to be ideal or that all steps are necessary for
model training. This approach started by training the model, pre-trained on ImageNet (Russakovsky
et al., 2014), with Adagrad (Duchi et al., 2011). Second, the number of training examples were
increased each epoch by 6%, with the initial number of examples equal to the number of cells times
200. Third, the classes were biased by oversampling minority classes, such that all classes were initially equally represented. A consequence of this approach, however, is that the minority classes are
seen repeatedly and therefore the majority classes have significantly more learned diversity. Fourth,
the class bias was reduced after each model had completed a training cycle — previous weights
loaded and the model re-trained with a reduced bias. The final model was trained with SGD, using a
linearly decreasing learning rate reduced at 4% with each epoch, without class biasing and with the
full dataset per epoch. The initial value of the learning rate varied for each model (between 0.004
and 0.02). The values of those hyperparameters were empirically determined.
A.2
M ODEL M2 AND M3
The layers of M2 are described in Table 2. M2 is trained using He intializations ((He et al., 2015)),
initial iterations of Adaboost(Freund & Schapire, 1995), followed by ADAM at learning rates of
0.005 and 0.001 (Kingma & Ba, 2014). Early stopping is used to detect a sustained decrease in
validation accuracy (Caruana et al., 2001).
B
M1 F EATURE - BASED I MAGE R ETRIEVAL
The generality of the M1 classification model is demonstrated by performing a query-by-example
on the 2K random Im2GPS. An example of an image of the Church of the Savior on Spilled Blood
is shown in Figure 4. By manual inspection (querying on the bounding box by GPS location), this
church was not present in the training data nor in the 2K random dataset (Hays & Efros, 2008).
C
C LASS B IAS A NALYSIS
Each image is given categorical indicator variable zik = 1 if the ith image is in the kth class,
0 otherwise. There exist a latent class distribution pi = P (Zi = 1) which is assumed constant between training, testing, and application. An estimate of this unknown distribution is
1 P
p̂z =
I(zij = 1). The second to last layer in all trained networks is assumed to
N j∈Training
be a fit logit for class i: Li = (L1i , ..., LKi ) for the ith image where K is the number
P of classes
in the mesh grid. The last layer output from networks is a softmax p̂m
=
exp(L
)/
ik
ik
v exp(Liv ),
where the mth index indicates model output given appropriate
image
input.
Optimization
is done
P
m
by minimizing the cross entropy between p̂i and p̂m
as
−
p̂
log
p̂
.
Images
are
given
the
class
k
i
ik
k
Ŷi = arg maxv p̂m
,
ẑ
=
I(
Ŷ
=
k).
ik
i
v
observed distribution of class frequencies in validation is q̂k
=
P
1
I(ẑik = 1). As a diagnostic we investigated how close the class
|Validation| k∈Validation
frequencies are to each other when both are calculated from the validation data. In general if two
The
11
Figure 4: Fine-mesh based query-by-example on Im2GPS 2K random testing dataset for a qualitative
assement for model generalization. The L2 distance is computed and shown below the top six
queried results.
models are compared we prefer the most accurate, but may also tilt toward unbiased models in classification distribution. If training has been done well, it should be the case that the KL-divergence
P
p̂k
between q and p is low: k p̂k log . As a matter of completeness we also consider the entropy of
q̂k
p and q.
12
| 1cs.CV
|
arXiv:1710.07198v1 [cs.CV] 19 Oct 2017
Dress like a Star: Retrieving Fashion Products from Videos
Noa Garcia
George Vogiatzis
Aston University, UK
Aston University, UK
[email protected]
[email protected]
Abstract
This work proposes a system for retrieving clothing and
fashion products from video content. Although films and
television are the perfect showcase for fashion brands to
promote their products, spectators are not always aware of
where to buy the latest trends they see on screen. Here,
a framework for breaking the gap between fashion products shown on videos and users is presented. By relating
clothing items and video frames in an indexed database and
performing frame retrieval with temporal aggregation and
fast indexing techniques, we can find fashion products from
videos in a simple and non-intrusive way. Experiments in a
large-scale dataset conducted here show that, by using the
proposed framework, memory requirements can be reduced
by 42.5X with respect to linear search, whereas accuracy is
maintained at around 90%.
Figure 1: Fashion items in a popular TV show.
difficult and it involves time-consuming searches.
To help in the task of finding fashion products that appear in multimedia content, some websites, such as Film
Grab4 or Worn on TV 5 , provide catalogs of items that can
be seen on films and TV shows, respectively. These websites, although helpful, still require some effort before actually buying the fashion product: users need to actively
remember items from videos they have previously seen and
navigate through the platform until they find them. On the
contrary, we propose an effortless, non-intrusive and fast
computer vision tool for searching fashion items in videos.
This work proposes a system for retrieving clothing
products from large-scale collections of videos. By taking
a picture of the playback device during video playback (i.e.
an image of the cinema screen, laptop monitor, tablet, etc.),
the system identifies the corresponding frame of the video
sequence and returns that image augmented with the fashion
items in the scene. In this way, users can find a product as
soon as they see it by simple taking a photo of the video sequence where the product appears. Figure 1 shows a frame
1. Introduction
Films and TV shows are a powerful marketing tool for
the fashion industry, since they can reach thousands of millions of people all over the world and impact on fashion
trends. Spectators may find clothing appearing in movies
and television appealing and people’s personal style is often
influenced by the multimedia industry. Also, online videosharing websites, such as YouTube1 , have millions of users
generating billions of views every day2 and famous youtubers3 are often promoting the latest threads in their videos.
Fashion brands are interested in selling the products that
are advertised in movies, television or YouTube. However,
buying clothes from videos is not straightforward. Even
when a user is willing to buy a fancy dress or a trendy pair
of shoes that appear in the latest blockbuster movie, there
is often not enough information to complete the purchase.
Finding the item and where to buy it is, most of the times,
1 https://www.youtube.com/
2 https://www.youtube.com/yt/press/statistics.html
4 http://filmgarb.com/
3 Users
5 https://wornontv.net
who have gained popularity from their videos on YouTube
1
of a popular TV show along with its fashion items.
Recently, many computer vision applications for clothing retrieval [14, 11] and style recommendation [24, 21]
have been proposed. Our system is close to clothing retrieval approaches in that the aim is to retrieve a product
from an image. Instead of retrieving products directly as
in standard clothing retrieval, we propose to first retrieve
frames from the video collection. There are several reasons
for that. Firstly, in standard clothing retrieval users usually
provide representative images of the object of interest (e.g.
dresses in front view, high-heeled shoes in side view, etc.).
In a movie, the view of the object of interest cannot be chosen, and items might be partially or almost completely occluded, such as the red-boxed dress in Figure 1. Secondly,
standard clothing retrieval requires to select a bounding box
around the object of interest. This is undesirable in clothing
video retrieval as it may distract user’s attention from the
original video content. Finally, performing frame retrieval
instead of product retrieval in videos allows users to get
the complete list of fashion items in a scene, which usually
matches the character’s style, including small accessories
such as earrings, watches or belts. Some of these items are
usually very small, sometimes almost invisible, and would
be impossible to detect and recognize using standard object
retrieval techniques.
Retrieving frames from videos is a challenging task in
terms of scalability. An average movie of two hours duration may contain more than 200,000 frames. That means
that with only five movies in the video dataset, the number
of images in the collection might be over a million. To overcome scalability issues, we propose the combination of two
techniques. Firstly, frame redundancy is exploited by tracking and summarizing local binary features [7, 20] into a key
feature. Secondly, key features are indexed in a kd-tree for
fast search of frames. Once the scene the query image belongs to is identified, the clothing items associated to that
scene are returned to the user.
The contributions of this paper are:
• Introduction of the video clothing retrieval task and
collection of a dataset of videos for evaluation.
• Proposal of a fast and scalable video clothing retrieval
framework based on frame retrieval and indexing algorithms.
• Exhaustive evaluation of the framework, showing that
similar accuracy to linear search can be achieved while
the memory requirements are reduced by a factor of
42.5.
This paper is structured as follows: related work is summarized in Section 2; the details of the proposed system
are explained in Section 3; experiment setup and results are
detailed in Section 4 and 5, respectively; finally, the conclusions are summarized in Section 6.
2. Related Work
Clothing Retrieval. Clothing retrieval is the field concerned with finding similar fashion products given a visual
query. In the last few years, it has become a popular field
within the computer vision community. Proposed methods
[14, 25, 13, 11, 12, 15] are mainly focused on matching realworld images against online catalog photos. In this crossscenario problem, some methods [25, 13, 15] train a set of
classifiers to find items with similar attributes. Other approaches [14, 12, 11] model the differences across domains
by using a transfer learning matrix [14] or a deep learning
architecture [12, 11]. All these methods, however, perform
clothing retrieval on static images, in which a representative
and recognizable part of the item of interest is contained. In
the scenario of clothing retrieval from films and television,
fashion items are not always shown in a representative and
recognizable way across the whole duration of the video
due to object occlusions and changes in camera viewpoints.
Given a snapshot of a video, an alternative solution is to
first, identify the frame it belongs to and then, find the products associated to it.
Scene Retrieval. Scene retrieval consists on finding similar frames or scenes from a video collection according to a
query image. Early work, such as Video Google [22] and
others [18, 9, 8], retrieve frames from video datasets by applying image retrieval methods and processing each frame
independently. More recent scene retrieval approaches use
temporal aggregation methods to improve scalability and
reduce memory requirements. [2, 5] extract hand-crafted
local features from frames, track them along time and aggregate them into a single vector. Other proposals [26, 3, 4]
produce a single scene descriptor, which improves efficiency but involves a more difficult comparison against the
static query image. Recently, Araujo et al. [4] showed that,
in scene retrieval, methods based on hand-crafted features
outperform convolutional neural networks.
Similarly to [2], our approach is based on the aggregation
of local features along temporal tracks. Instead of using
SIFT features [16], we provide evidence that binary features
[7, 20] are more appropriate for the task and that they can
be easily indexed in a kd-tree for fast search.
3. Video Clothing Retrieval Framework
To retrieve fashion products from movies, television and
YouTube, we propose a framework based on frame retrieval
and fast indexing techniques. Since the number of frames
explodes with the number of movies available in a collection, a key feature of this framework is its scalability. The
system overview is shown in Figure 2. There are three main
modules: product indexing, training phase and query phase.
The product indexing and the training phase are done offline, whereas the query phase is performed online.
Product Indexing
Offline
Training phase
Fashion Items
Indexed Database
Video Collection
Feature
Extraction
Feature
Tracking
Shot
Detection
Key Feature
Computation
Key Feature
Indexing
Video Frames
Online
Output
Query phase
Retrieved Frame
Feature
Extraction
Key Feature
Search
Shot
Search
Fashion Items
Frame
Search
Query Image
Figure 2: System overview. First, during the product indexing, fashion items and frames are related in an indexed database.
Then, frames are indexed by using key features. Finally, in the query phase, query images are used to retrieve frames and find
associated fashion items.
In the product indexing, fashion items and frames from
the video collection are related in an indexed database. This
process can be done manually (e.g. with Amazon Mechanical Truck6 ) or semi-automatically with the support of a standard clothing retrieval algorithm. In the training phase, features are extracted and tracked along time. Tracks are used
to detect shot boundaries and compute aggregated key features, which are indexed in a kd-tree. In the query phase,
features extracted from the query image are used to find
their nearest key features in the kd-tree. With those key
features, the shot and the frame the query image belong to
are retrieved. Finally, the fashion products associated with
the retrieved frame are returned by the system.
3.1. Feature Extraction and Tracking
Local features are extracted from every frame in the
video collection and tracked across time by applying descriptor and spatial filters. The tracking is performed in a
bidirectional way so features within a track are unique (i.e.
each feature can only be matched with up to two features:
one in the previous frame and one in the following frame).
6 https://www.mturk.com
As local features, instead of the popular SIFT [16], we
use binary features for two main reasons. Firstly, because
Hamming distance for binary features is faster to compute
than Euclidean distance for floating-points vectors. Secondly, because we find that binary features are more stable
over time than SIFT, as can be seen in Figure 3. Convolutional neural network (CNN) features from an intermediate
layer of a pre-trained network were also studied. However,
their results were not satisfactory as the resulting tracked
features quickly started to get confused when the dataset
increased.
3.2. Shot Detection and Key Feature Computation
Consecutive frames that share visual similarities are
grouped into shots. The boundaries of different shots are
detected when two consecutive frames have no common
tracks. Each shot contains a set of tracks, each track representing the trajectory of a particular feature along time.
We define a key feature as the aggregation of all the features
in the same track into a single vector. Subsequently, each
shot is then represented by a set of key features, similarly
to how frames are represented by a set of features. For each
track, a key feature is computed by using majorities [10].
Figure 4: Visual similarities between frames. Left: Query
image. Middle: Ground truth frame. Right: Retrieved
frame similar to ground truth frame, thus, Visual Match.
FIFO queue we first explore the closest nodes to the root,
because the corresponding bits exhibit more variability by
construction of the tree.
3.4. Search
Figure 3: Trajectories of sample tracks along a sequence of
frames. Left: Binary features. Right: SIFT features. Binary
features are more constant over time than SIFT features.
3.3. Key Feature Indexing
A popular method for searching in binary feature space
is FLANN [17]. FLANN uses multiple, randomly generated hierarchical structures and it is not obvious how it can
be deployed across multiple CPUs in order to provide the
scalability we are looking for. On the other hand, kd-trees
are a formalism that has been shown to be highly parallelizable in [1]. In this work, we modify the basic kd-tree to
handle binary features and index our key features.
In a kd-tree each decision node has an associated dimension, a splitting value and two child nodes. If the value of a
query vector in the associated dimension is greater than the
splitting value, the vector is assigned to the left child. Otherwise, the vector is associated to the right child. The process
is repeated at each node during the query phase until a leaf
node is reached. In our case, as we are dealing with binary
features, each decision node has an associated dimension,
dim, such that all query vectors, v, with v[dim] = 1 belong
to the left child, and all vectors with v[dim] = 0 belong to
the right child. The value dim is chosen such that the training data is split more evenly in that node, i.e. its entropy
is maximum. Note that this criterion is similar to the one
used in the ID3 algorithm [19] for the creation of decision
trees, but where the splitting attribute is chosen as the one
with smallest entropy. Leaf nodes have as many as SL indices pointing to the features that ended up in that node. A
first-in first-out (FIFO) queue keeps record of the already
visited nodes to backtrack B times and explore them later.
We use a FIFO queue to ensure that even if some of the bits
in a query vector are wrong, the vector can reach its closest neighbours by exploring unvisited nodes latter. With the
In the query phase, binary features are extracted from an
input image and assigned to its nearest set of key features by
searching down the kd-tree. Each key feature votes for the
shot it belongs to. The set of frames contained in the most
voted shot are compared against the input image by brute
force, i.e. distances between descriptors in the query image
and descriptors in the candidate frames are computed. The
frame with minimum distance is retrieved. Shots are commonly groups of a few hundreds of frames, thus the computation can be performed very rapidly when applying the
Hamming distance. Finally, all the fashion vectors associated with the retrieved frame are returned to the user.
4. Experiments
Experimental Details. To evaluate our system a collection of 40 movies with more than 80 hours of video and up
to 7 million frames is used. Query images are captured by
a webcam (Logitech HD Pro Webcam C920) while movies
are being played on a laptop screen. To provide a ground
truth for our experiments, the frame number of each captured image is saved in a text file. Movies and queries have
different resolutions and aspect ratios, thus all the frames
and images are scaled down to 720 pixels in width. Binary features are computed by using ORB detector [20] and
BRIEF extractor [7]. In the tracking module, only matches
with a Hamming distance less than 20 and a spatial distance less than 100 pixels are considered, whereas in the
key feature computation algorithm, only tracks longer than
7 frames are used. The default values for the kd-tree are set
at SL = 100 and B = 50.
Evaluation criteria. The aim of the system is to find
the fashion products in the query scene, so we only retrieve
one frame per query image. The returned frame may not be
the same frame as the one in the annotated ground truth but
one visually similar. As long as the retrieved frame shares
strong similarities with the ground truth frame, we consider
it as a Visual Match, as shown in Figure 4. To measure the
visual similarities between two frames, we match SURF [6]
BF
KT
KF
Ours
85M
85M
25M
2M
2.53GB
2.53GB
762MB
61MB
0.90
0.91
0.92
0.94
0.92
0.93
B = 100
0.96
0.93
0.94
B = 250
0.97
0.93
0.94
Indexed
Features
Memory
B = 10
(b)
Figure 5: Precision of the evaluation method: (a) TPR and
FPR curves; (b) Three ground truth frames (left column)
and retrieved frames (right column) along with their score.
features. The linear comparison between two frames that
do not present any noise or perspective distortion is an easy
task that almost all kinds of features can perform correctly.
If the matching score between SURF features of two dataset
frames is greater than a threshold, τ , they are considered to
be visually similar.
To measure the precision of this evaluation method, we
manually annotate either if a pair of frames (i.e. ground
truth frame and retrieved frame) is a Visual Match or not,
along with its score. For different values of τ the True Positive Rate (TPR) as well as the False Positive Rate (FPR) are
computed as:
TPR =
TP
TP + FN
FPR =
FP
FP + TN
(1)
where T P is the number of true positives (Visual Match
with score > τ ), F P is the number of false positives (No
Match with score > τ ), T N is the number of true negatives
(No Match with score ≤ τ ) and F N is the number of false
negatives (Visual Match with score ≤ τ ). Figure 5 shows
both the TPR and the FPR computed with the annotations of
615 pairs of frames and 406 different values of τ . Among all
the possible values, we chose τ = 0.15, with T P R = 0.98
and F P R = 0. Finally, the accuracy for a set of query
images is computed as:
Acc =
No. Visual Matches
Total No. Queries
(2)
5. Experimental Results
To evaluate the proposed framework, two different experiments are performed. In the first one, our frame retrieval
system is compared against other retrieval systems. In the
second one, the performance of the frame retrieval method
when scaling up the video collection is evaluated in terms
of accuracy.
5.1. Retrieval Performance
First, we compare our system against other retrieval
frameworks.
Accuracy
(a)
B = 50
0.98
Table 1: Comparison between different systems on a single
movie. Accuracy is shown for four backtracking steps, B.
Our system achieves comparable results by using 42.5x less
memory than BF and KT and 12.5x less memory than KF.
BF. Brute Force matcher, in which query images are
matched against all frames and all features in the database.
Brute Force system is only used as an accuracy benchmark,
since each query take, in average, 46 minutes to be processed. Temporal information is not used.
KT. Kd-Tree search, in which all features from all
frames are indexed using a kd-tree structure, similarly to
[1]. Temporal information is not used.
KF. Key Frame extraction method [23], in which temporal information is used to reduce the amount of frames
of each shot into a smaller set of key frames. Key frames
are chosen as the ones at the peaks of the distance curve
between frames and a reference image computed for each
shot. For each key frame, features are extracted and indexed
in a kd-tree structure.
We compare these three systems along with the method
proposed here, using a single movie, The Devil Wears
Prada, consisting of 196,572 frames with a total duration
of 1 hour 49 minutes and 20 seconds. To be fair in our comparison, we use binary features in each method. The results
of this experiment are detailed in Table 1. The performance
is similar for the four systems. However, the amount of
processed data and the memory requirements are drastically
reduced when the temporal information of the video collection is used. In the case of our system, by exploiting temporal redundancy between frames, the memory is reduced
by 42.5 times with respect to BF and KT and by 12.5 times
with respect to the key frame extraction technique. Theoretically, that means that when implemented in a distributed
kd-tree system as the one in [1], where the authors were
able to process up to 100 million images, our system might
be able to deal with 4,250 million frames, i.e. more than
Title
N. Frames
N. Features
N. Shots
N. Key Features
N. Queries
Accuracy
The Help
210387
101M
1726
2.2M
813
0.98
Intolerable Cruelty
179234
86M
1306
2M
544
0.97
Casablanca
147483
71M
881
1.5M
565
0.96
Witching & Bitching
163069
66M
4193
0.8M
588
0.74
Pirates of the Caribbean 3
241127
108M
3695
1.7M
881
0.74
Captain Phillips
190496
59M
7578
0.6M
618
0.67
7M
3040M
116307
58M
25142
0.87
Total
Table 2: Summary of the results from the experiments on the dataset containing 40 movies. Due to space restrictions, only
three of the best and worst accuracies are shown, along with the total results for the whole database.
Pirates of the Caribbean 3 perform worst, as fewer descriptors can be found in those kinds of dimly lit images.
Figure 6 shows the evolution of accuracy when the size
of the database increases for five different movies. It can
be seen that most of the movies are not drastically affected
when the number of frames in the database is increased
from 200,000 to 7 million. For example, both Intolerable
Cruelty and 12 Years a Slave movies maintain almost a constant accuracy for different sizes of the collection. Even
in the worst case scenario, The Devil Wears Prada movie,
the loss in accuracy is less than a 8.5%. This suggests that
our frame retrieval system is enough robust to handle largescale video collections without an appreciable loss in performance.
Figure 6: Accuracy vs Database size for 5 different movies.
20,000 movies and 40,000 hours of video.
5.2. Large-scale Dataset
In this experiment, we explore the scalability of our
framework by increasing the dataset up to 40 movies and 7
million frames. The collection is diverse and it contains different movie genres such as animation, fantasy, adventure,
comedy or drama, to ensure that a wide range of movies can
be indexed and retrieved with the proposed algorithm.
Table 2 shows the results of this experiment. It can be
seen that, by using our temporal aggregation method, the
amount of data is reduced from 7 million frames and 3,040
million features to only 116,307 shots and 58 million key
features. Even so, the total number of key features in the 40
movie collection is still smaller than the 80 million features
that, in average, a single movie contains. The total accuracy
over the 40 movies is 0.87, reaching values of 0.98 and 0.97
in The Help and Intolerable Cruelty movies, respectively.
Movies with very dark scenes such as Captain Phillips and
6. Conclusions
This work proposes a framework to perform video clothing retrieval. That is, given a snapshot of a video, identify
the video frame and retrieve the fashion products that appear in that scene. This task would help users to easily find
and purchase items shown in movies, television or YouTube
videos. We propose a system based on frame retrieval and
fast indexing. Experiments show that, by using our temporal aggregation method, the amount of data to be processed
is reduced by a factor of 42.5 with respect to linear search
and accuracy is maintained at similar levels. The proposed
system scales well when the number of frames is increased
from 200,000 to 7 million. Similar to [1] our system can
easily be parallelised across multiple CPUs. Therefore, the
encouraging experimental results shown here indicate that
our method has the potential to index fashion products from
thousands of movies with high accuracy. In future work we
intend to further increase the size of our dataset in order
to identify the true limits of our approach. Furthermore it
is our intention to make all data publicly available so that
other researchers can investigate this interesting retrieval
problem.
References
[1] M. Aly, M. Munich, and P. Perona. Distributed kd-trees for
retrieval from very large image collections. British Machine
Vision Conference, 2011. 4, 5, 6
[2] A. Anjulan and N. Canagarajah. Object based video retrieval
with local region tracking. Signal Processing: Image Communication, 22(7), 2007. 2
[3] A. Araujo, J. Chaves, R. Angst, and B. Girod. Temporal aggregation for large-scale query-by-image video retrieval. In
Image Processing (ICIP), 2015 IEEE International Conference on, pages 1519–1522. IEEE, 2015. 2
[4] A. Araujo and B. Girod. Large-scale video retrieval using
image queries. IEEE Transactions on Circuits and Systems
for Video Technology, 2017. 2
[5] A. Araujo, M. Makar, V. Chandrasekhar, D. Chen, S. Tsai,
H. Chen, R. Angst, and B. Girod. Efficient video search
using image queries. In Image Processing (ICIP), 2014 IEEE
International Conference on, pages 3082–3086. IEEE, 2014.
2
[6] H. Bay, T. Tuytelaars, and L. Van Gool. SURF : Speeded Up
Robust Features. European Conference on Computer Vision,
pages 404–417, 2006. 4
[7] M. Calonder, V. Lepetit, C. Strecha, and P. Fua. BRIEF : Binary Robust Independent Elementary Features. ECCV, 2010.
2, 4
[8] D. Chen, N.-M. Cheung, S. Tsai, V. Chandrasekhar,
G. Takacs, R. Vedantham, R. Grzeszczuk, and B. Girod.
Dynamic selection of a feature-rich query frame for mobile
video retrieval. In Image Processing (ICIP), 2010 17th IEEE
International Conference on, pages 1017–1020. IEEE, 2010.
2
[9] O. Chum, J. Philbin, M. Isard, and A. Zisserman. Scalable
near identical image and shot detection. In Proceedings of
the 6th ACM international conference on Image and video
retrieval, pages 549–556. ACM, 2007. 2
[10] C. Grana, D. Borghesani, M. Manfredi, and R. Cucchiara.
A fast approach for integrating orb descriptors in the bag of
words model. In IS&T/SPIE Electronic Imaging. International Society for Optics and Photonics, 2013. 3
[11] M. Hadi Kiapour, X. Han, S. Lazebnik, A. C. Berg, and T. L.
Berg. Where to buy it: Matching street clothing photos in
online shops. In ICCV, 2015. 2
[12] J. Huang, R. S. Feris, Q. Chen, and S. Yan. Cross-domain image retrieval with a dual attribute-aware ranking network. In
Proceedings of the IEEE International Conference on Computer Vision, pages 1062–1070, 2015. 2
[13] Y. Kalantidis, L. Kennedy, and L.-J. Li. Getting the look:
clothing recognition and segmentation for automatic product
suggestions in everyday photos. In Proceedings of the 3rd
ACM conference on International conference on multimedia
retrieval, pages 105–112. ACM, 2013. 2
[14] S. Liu, Z. Song, G. Liu, C. Xu, H. Lu, and S. Yan. Street-toshop: Cross-scenario clothing retrieval via parts alignment
and auxiliary set. In Computer Vision and Pattern Recognition (CVPR), 2012 IEEE Conference on, pages 3330–3337.
IEEE, 2012. 2
[15] Z. Liu, P. Luo, S. Qiu, X. Wang, and X. Tang. Deepfashion:
Powering robust clothes recognition and retrieval with rich
annotations. In CVPR, 2016. 2
[16] D. G. Lowe. Distinctive image features from scale-invariant
keypoints. International Journal of Computer Vision, 60(2),
2004. 2, 3
[17] M. Muja and D. G. Lowe. Fast matching of binary features.
Proceedings of the 2012 9th Conference on Computer and
Robot Vision, 2012. 4
[18] D. Nistér and H. Stewénius. Scalable recognition with a vocabulary tree. CVPR, 2006. 2
[19] J. R. Quinlan. Induction of decision trees. Machine learning,
1(1):81–106, 1986. 4
[20] E. Rublee, V. Rabaud, K. Konolige, and G. Bradski. ORB:
an efficient alternative to SIFT or SURF. In ICCV, 2011. 2,
4
[21] E. Simo-Serra, S. Fidler, F. Moreno-Noguer, and R. Urtasun. Neuroaesthetics in fashion: Modeling the perception of
fashionability. In Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition, pages 869–877,
2015. 2
[22] J. Sivic and A. Zisserman. Video Google: a text retrieval
approach to object matching in videos. ICCV, 2003. 2
[23] Z. Sun, K. Jia, and H. Chen. Video key frame extraction based on spatial-temporal color distribution. In Intelligent Information Hiding and Multimedia Signal Processing,
2008. IIHMSP’08 International Conference on, pages 196–
199. IEEE, 2008. 5
[24] A. Veit, B. Kovacs, S. Bell, J. McAuley, K. Bala, and S. Belongie. Learning visual clothing style with heterogeneous
dyadic co-occurrences. In Proceedings of the IEEE International Conference on Computer Vision, pages 4642–4650,
2015. 2
[25] D. Wei, W. Catherine, B. Anurag, P.-m. Robinson, and
S. Neel. Style finder: fine-grained clothing style recognition
and retrieval, 2013. 2
[26] C.-Z. Zhu and S. Satoh. Large vocabulary quantization
for searching instances from videos. In Proceedings of the
2nd ACM International Conference on Multimedia Retrieval,
page 52. ACM, 2012. 2
| 7cs.IT
|
arXiv:1702.08615v1 [math.ST] 28 Feb 2017
Bridging Finite and Super Population Causal Inference
Peng Ding∗, Xinran Li†, and Luke W. Miratrix‡
Abstract
There are two general views in causal analysis of experimental data: the super population
view that the units are an independent sample from some hypothetical infinite populations, and
the finite population view that the potential outcomes of the experimental units are fixed and the
randomness comes solely from the physical randomization of the treatment assignment. These
two views differs conceptually and mathematically, resulting in different sampling variances of
the usual difference-in-means estimator of the average causal effect. Practically, however, these
two views result in identical variance estimators. By recalling a variance decomposition and
exploiting a completeness-type argument, we establish a connection between these two views in
completely randomized experiments. This alternative formulation could serve as a template for
bridging finite and super population causal inference in other scenarios.
Keywords: Completeness; Finite population correction; Potential outcome; Simple random
sample; Variance of individual causal effects
1.
Introduction
Neyman (1923, 1935) defined causal effects in terms of potential outcomes, and proposed an inferential framework viewing all potential outcomes of a finite population as fixed and the treatment
assignment as the only source of randomness. This finite population view allows for easy interpretation free of any hypothetical data generating process of the outcomes, and is used in a variety
of contexts (e.g., Kempthorne 1952; Hinkelmann and Kempthorne 2008; Copas 1973; Rosenbaum
∗
Department of Statistics, University of California,
Berkeley, California 94720 USA. Email:
[email protected]
†
Department of Statistics, Harvard University, Cambridge, Massachusetts 02138 USA. Email:
[email protected]
‡
Graduate School of Education and Department of Statistics, Harvard University, Cambridge, Massachusetts
02138 USA. Email: [email protected]
1
2002; Imai 2008; Freedman 2008a,b; Rosenbaum 2010; Aronow and Middleton 2013; Aronow et al.
2014; Abadie et al. 2014; Miratrix et al. 2013; Ding 2016; Lin 2013; Middleton and Aronow 2015;
Imbens and Rubin 2015; Chiba 2015; Rigdon and Hudgens 2015; Li and Ding 2016). This approach
is considered desirable because, in particular, it does not assume the data is somehow a representative sample of some larger (usually infinite) population.
Alternative approaches, also using the potential outcomes framework, assume that the potential
outcomes are independent and identical draws from a hypothetical infinite population. Mathematical derivations under this approach are generally simpler, but the approach itself can be criticized
because of this typically untenable sampling assumption. Furthermore, this approach appears to
ignore the treatment assignment mechanism.
That being said, it is well known that the final variance formulae from either approach tend to be
quite similar. For example, while the variance of the difference-in-means estimator for a treatmentcontrol experiment under an infinite population model is different from the one under Neyman
(1923)’s finite population formulation, this difference is easily represented as a function of the variance of the individual causal effects. Furthermore, this difference term is unidentifiable and is often
assumed away under a constant causal effect model (Neyman 1923, 1935; Hodges and Lehmann
1964; Rubin 1990; Reichardt and Gollob 1999; Freedman et al. 2007), or by appeals to the final
estimators being “conservative.”
Strictly speaking, the infinite population variance estimate gives a conservative (overly large)
estimate of the finite population variance for the difference in means. As deriving infinite population variance expressions, relative to finite population variance expressions, tends to be more
mathematically straightforward, we might naturally wonder if we could use infinite population expressions as conservative forms of finite population expressions more generally. In this work we
show that in fact we can assume an infinite population model as an assumption of convenience, and
derive formula from this perspective. This shows that we can thus consider the resulting formula
as focused on the treatment assignment mechanism and not on a hypothetical sampling mechanism, i.e., we show variance derivations under the infinite population framework can be used as
conservative estimators in a finite context.
Mathematically, this result comes from a variance decomposition and a completeness-style argument that characterizes the connection and the difference between these two views. The vari2
ance decomposition we use has previously appeared in Imai (2008), Imbens and Rubin (2015), and
Balzer et al. (2016). The completeness-style argument, which we believe is novel in this domain,
then sharpens the variance decomposition by moving from an expression on an overall average
relationship to one that holds for any specific sample. Our overall goal is simple: we wish to
demonstrate that if one uses variance formula derived from assuming an infinite population sampling model, then the resulting inference one obtains will be correct with regards to the analogous
samples-specific treatment effects (although it could be potentially conservative in that the standard
errors may be overly large) regardless of the existence of any sampling mechanism.
2.
Super Population, Finite Population, and Samples
Assume that random variables (Y (1), Y (0)) represent the pair of potential outcomes of an infinite super population, from which we take an independent and identically distributed (IID) finite
population of size n:
S = {(Yi (1), Yi (0)) : i = 1, . . . , n}.
We first discuss completely randomized experiments, and comment on other experiments in Section
4. For a completely randomized experiment, we randomly assign n1 units to receive treatment,
leaving the remaining n0 = n − n1 units to receive control. Let Z = (Z1 , . . . , Zn ) ∈ {0, 1}n be the
treatment assignment vector, which takes a particular value (z1 , . . . , zn ) ∈ {0, 1}n with probability
P
n1 !n0 !/n!, for any (z1 , . . . , zn ) ∈ {0, 1}n with ni=1 zi = n1 . The observed outcome for unit i is
then Yiobs = Zi Yi (1) + (1 − Zi )Yi (0).
At the super population level, the average potential outcomes are E{Y (1)} and E{Y (0)}, and
the average causal effect is
τ = E{Y (1) − Y (0)}.
The population variances of the potential outcomes and individual causal effect are
V1 = Var{Y (1)},
V0 = Var{Y (0)},
Vτ = Var{Y (1) − Y (0)}.
At the finite population level, i.e., for a fixed sample S, the average potential outcomes and
3
average causal effect are
n
Ȳ1 =
1X
Yi (1),
n
n
Ȳ0 =
i=1
1X
Yi (0),
n
τS = Ȳ1 − Ȳ0 .
i=1
The corresponding finite population variances of the potential outcomes and individual causal
effects are
n
S12 =
1 X
{Yi (1) − Ȳ1 }2 ,
n−1
i=1
n
S02 =
1 X
{Yi (0) − Ȳ0 }2 ,
n−1
n
Sτ2 =
i=1
1 X
(τi − τ )2 ,
n−1
i=1
where we use the divisor n − 1 following the tradition of survey sampling (Cochran 1977). All these
quantities are fully dependent on S. In classical casual inference (Neyman 1923), the potential
outcomes of these n experimental units, S, are treated as fixed numbers. Equivalently, we can
consider such causal inference to be conducted conditional on S (e.g., Copas 1973; Rosenbaum
2002; Imbens and Rubin 2015; Rigdon and Hudgens 2015; Chiba 2015).
Regardless, we have two parameters—the population average treatment effect τ , and the sample
average treatment effect τS . After collecting the data, we would want to draw inference about τ or
τS .
Our primary statistics are the averages of the observed outcomes and the difference-in-means
estimator:
Ȳ1obs =
n
1 X
Zi Yiobs ,
n1
Ȳ0obs =
i=1
n
1 X
(1 − Zi )Yiobs ,
n0
i=1
τb = Ȳ1obs − Ȳ0obs .
We also observe the sample variances of the outcomes under treatment and control using
n
s21
1 X
=
Zi (Yiobs − Ȳ1obs )2 ,
n1 − 1
n
s20
i=1
1 X
=
(1 − Zi )(Yiobs − Ȳ0obs )2 .
n0 − 1
i=1
We do not have the sample analogue of Sτ2 or Vτ because Yi (1) and Yi (0) are never jointly observed
for any unit i in the sample.
We summarize the infinite population, finite population and sample quantities in Table 1.
4
Table 1: Means and variances at the super population, finite population and sample levels.
super population
finite population
sample
3.
treatment
E{Y (1)}
Ȳ1
Ȳ1obs
means
control
effect
E{Y (0)} τ = E{Y (1) − Y (0)}
Ȳ0
τS = Ȳ1 − Ȳ0
Ȳ0obs
τb = Ȳ1obs − Ȳ0obs
variances
treatment control
V1
V0
S12
S02
s21
s20
effect
Vτ
Sτ2
−
Deriving Complete Randomization Results with an Independent Sampling Model
The three levels of quantities in Table 1 are connected via independent sampling and complete
randomization. Neyman (1923), without reference to any infinite population and by using the
assignment mechanism as the only source of randomness, represented the assignment mechanism
via an urn model, and found
Var(b
τ | S) =
S12 S02 Sτ2
+
−
.
n1
n0
n
(1)
He then observed that the final term was unidentifiable but nonnegative, and thus if we dropped
it we would obtain an upper bound of the estimator’s uncertainty.
We next derive this result by assuming a hypothetical sampling mechanism from some assumed
infinite super-population model of convenience. This alternative derivation of the above result,
which can be extended to other assignment mechanisms, shows how we can interpret formulae
based on super-population derivations as conservative formulae for finite-sample inference.
3.1.
Sampling and randomization
To begin, note that IID sampling of S from the super population implies three things: first, the
finite population average causal effect τS satisfies E(τS ) = τ , second
#
n
Vτ
1X
{Yi (1) − Yi (0)} =
,
Var(τS ) = Var
n
n
"
i=1
5
(2)
and third, the sample variances are unbiased for the true variances:
E(S12 ) = V1 ,
E(S02 ) = V0 ,
E(Sτ2 ) = Vτ .
(3)
Conditional on S, randomization of the treatment Z is the only source of randomness. In a
completely randomized experiment, the outcomes in the treatment group form a simple random
sample of size n1 from {Yi (1) : i = 1, . . . , n}, and the outcomes in the control group form a simple
random sample of size n0 from {Yi (0) : i = 1, . . . , n}. Therefore, classical survey sampling theory
(Cochran 1977) for the sample mean and variance gives
E(Ȳ1obs | S) = Ȳ1 ,
E(Ȳ0obs | S) = Ȳ0 ,
E(b
τ | S) = τS ,
E(s21 | S) = S12 ,
E(s20 | S) = S02 .
(4)
We do not use the notation ES (·), because, depending on the contexts, such notation could either
indicate expectation conditional on S or expectation averaged over S. We therefore use conditional
expectations and conditional variances explicitly.
If we do not condition on S, then the independence induced by the assignment mechanism means
the outcomes under treatment are IID samples of Y (1) and the outcomes under control are IID
samples of Y (0), and furthermore these samples are independent of each other. This independence
makes it straightforward to show that τb is unbiased for τ with super population variance
Var(b
τ ) = Var(Ȳ1obs ) + Var(Ȳ0obs ) =
V1
V0
+ .
n1 n0
(5)
This is the classic infinite population variance formula for the two sample difference-in-means
statistic. We could use it to obtain standard errors by plugging in s21 and s20 for the two variances.
This derivation, from (2)–(5), is motivated by the sampling assumption: the assignment mechanism
makes the two samples independent, but it is the IID assumption and classic sampling theory
that gives this result (along with asymptotical normality which allows for associated testing and
confidence interval generation).
6
3.2.
Connecting the finite and infinite population inference with a variance
decomposition
We will now extend the above to indirectly derive the result on the variance of τb in finite popula-
tion inference without explicitly enumerating the potential outcomes. The variance decomposition
formula implies
Var(b
τ ) = E{Var(b
τ | S)} + Var{E(b
τ | S)}
(6)
= E{Var(b
τ | S)} + Var(τS )
= E{Var(b
τ | S)} +
Vτ
,
n
(7)
which further implies that the finite population variance of τb satisfies (using (3))
E {Var(b
τ | S)} = Var(b
τ) −
Vτ
V1
V0 Vτ
=
+
−
=E
n
n1 n0
n
S12 S02 Sτ2
+
−
n1
n0
n
.
(8)
Compare to the classic variance expression (1), which is this without the expectation. Here we
have that on average our classic variance expression holds. Now, because this is true for any
infinite population, as it is purely a consequence of the IID sampling mechanism and complete
randomization, we can close the gap between (1) and (8). Informally speaking, because (8) holds as
an average over many hypothetical super populations, it should also hold for any finite population
at hand, and indeed it does, as we next show using a “completeness” concept from statistics
(Lehmann and Romano 2008).
3.3.
A “Completeness” argument
First, define
f (S) ≡
S12 S02 Sτ2
+
−
− Var(b
τ | S),
n1
n0
n
(9)
a function of a fixed finite sample S, as the difference of the hypothesized finite sample variance
formula and the actual finite sample variance. Formula (8) shows E{f (S)} = 0. Now we are going
to show the stronger result f (S) = 0.
7
For any given sample S, we have fixed sample quantities: U = Y1 (1) − Ȳ1 , . . . , Yn (1) − Ȳ1 ,
and W = Y1 (0) − Ȳ0 , . . . , Yn (0) − Ȳ0 . Some algebra gives
τb − τS =
n
n
1 X
1 X
Zi Ui −
(1 − Zi )Wi .
n1
n0
i=1
(10)
i=1
According to (10), the finite population variance Var(τ̂ | S) = E{(τ̂ − τS )2 | S} must be a quadratic
form of (U, W ). Therefore, from (9), f (S) must be a quadratic form of (U, W ), and therefore must
have the following expression:
f (S) =
n
n X
X
aij Ui Uj +
n
n X
X
bij Wi Wj +
cij Ui Wj .
i=1 j=1
i=1 j=1
i=1 j=1
n
n X
X
for some aij , bij and cij . f (S) is the same regardless of the ordering of units, so by symmetry
the constants aij , bij and cij must be the same regardless of i and j. This gives aii = a11 , aij =
a12 , bii = b11 , bij = b12 , cii = c11 , cij = c12 , for all i 6= j. Thus
f (S) =a11
n
X
Uj2
+ a12
i=1
Because
XX
Ui Uj + b11
i=1 Ui
Ui2
+
i=1
n
X
i=1
Pn
Wi2
i=1 Wi
+
+ b12
XX
Wi Wj + c11
n
X
Ui Wi + c12
i=1
i6=j
XX
Ui W j .
i6=j
= 0, we have
n
X
Similarly, as
Wi2
i=1
i6=j
Pn
n
X
XX
Ui Uj =
n
X
Ui
i=1
i6=j
!2
= 0.
= 0, we have
XX
Wi Wj = 0,
n
X
Ui Wi +
i=1
i6=j
XX
Ui W j =
n
X
i=1
i6=j
Ui
!
n
X
i=1
Wi
!
= 0.
Use the above to replace our cross terms of i 6= j with single index terms to write f (S) as
f (S) =a
n
X
i=1
Ui2 + b
n
X
i=1
Wi2 + c
n
X
Ui Wi ,
i=1
where a = a11 − a12 , b = b11 − b12 , and c = c11 − c12 . Because E{f (S)} = 0, we have aE(U12 ) +
8
aE(W12 ) + cE(U1 W1 ) = 0, where
E(U12 ) =
n−1
n−1
n−1
Var{Y (1)}, E(W12 ) =
Var{Y (0)}, E(U1 W1 )2 =
Cov{Y (1), Y (0)}.
n
n
n
Thus,
aVar{Y (1)} + bVar{Y (0)} + cCov{Y (1), Y (0)} = 0.
(11)
Because (11) holds for any populations regardless of its values of Var{Y (1)}, Var{Y (0)} and Cov{Y (1), Y (0)},
it must be true that a = b = c = 0, which implies f (S) = 0.
4.
Discussion
Equation (8) relies on the assumption that the hypothetical infinite population exists, but Equation (1) does not. However, the completeness-style argument allowed us to make our sampling
assumption only for convenience in order to prove (1) by, in effect, dropping the expectation on
both sides of (8). Similar argument exists in the classical statistics literature; see Efron and Morris
(1973) for the empirical Bayes view of Stein’s estimator. While the final result is, of course, not
new, we offer it as it gives an alternative derivation that does not rely on asymptotics such as a
growing superpopulation or a focus on the properties of the treatment assignment mechanism.
Using Freedman (2008a)’s results, Aronow et al. (2014) considered a super population with N
units, with the finite population being a simple random sample of size n. Letting N → ∞, we
can obtain similar results. We go in the other direction: we use the variance decomposition (6) to
derive the finite population variance from the super population one.
This decomposition approach also holds for other types of experiments. First, for a stratified
experiment, each stratum is essentially a completely randomized experiment. Apply the result to
each stratum, and then average over all strata to obtain results for a stratified experiment. Second, because a matched-pair experiment is a special case of a stratified experiment with two units
within each stratum, we can derive the Neyman-type variance (cf. Imai 2008; Imbens and Rubin
2015) directly from that of a stratified experiment. Third, a cluster-randomized experiment is a
completely randomized experiment on the clusters. If the causal parameters can be expressed as
9
cluster-level outcomes, then the result can be straightforwardly applied (cf. Aronow and Middleton
2013; Middleton and Aronow 2015). Fourth, for general experimental designs, the variance decomposition in (6) still holds, and therefore we can modify the derivation of the finite population
variance according to different forms of (4) and (5).
In a completely randomized experiment, the finite population sampling variance of τb in (1) de-
pends on three terms: the first two can be unbiasedly estimated by s21 /n1 and s20 /n0 , but the third
term Sτ2 /n is unidentifiable from the data. Assuming a constant causal effect model, Sτ2 = 0 and
the variance estimators under both finite and super population inference coincide. However, Robins
(1988), Aronow et al. (2014) and Ding and Dasgupta (2016) demonstrated that the treatment variation term Sτ2 has a sharp lower bound that may be larger than 0, which allows for more precise
variance estimators under the finite population view. This demonstrates that we can indeed make
better inference conditional on the sample we have. On the other hand, in this work, we showed
that assuming an infinite population, while not necessarily giving the tightest variance expressions,
nonetheless gives valid (conservative) variance expressions from a finite-population perspective.
We offer this approach as a possible method of proof that could ease derivations for more complex
designs. More broadly, it is a step towards establishing that infinite population derivations for randomized experiments can be generally thought of as pertaining to their finite population analogs.
Also see Samii and Aronow (2012) and Lin (2013), who provide alternative discussions of super
population regression-based variance estimators under the finite population framework.
Our discussion is based on the frequentists’ repeated sampling evaluations of the difference-inmeans estimator for the average causal effect. In contrast, Fisher (1935) proposed the randomization test against the sharp null hypothesis that Yi (1) = Yi (0) for all units, which is numerically the same as the permutation test for exchangeable units sampled from an infinite population
(Lehmann and Romano 2008). This connection becomes more apparent when only the ranks of the
outcomes are used to construct the test statistics, as discussed extensively by Lehmann (1975).
Acknowledgments
We thank the Associate Editor, Dr. Peter Aronow, and three anonymous reviewers for helpful
comments.
10
References
Abadie, A., Athey, S., Imbens, G. W., and Wooldridge, J. M. (2014). Finite population causal
standard errors. Technical report, National Bureau of Economic Research.
Aronow, P. M., Green, D. P., and Lee, D. K. K. (2014). Sharp bounds on the variance in randomized
experiments. The Annals of Statistics, 42:850–871.
Aronow, P. M. and Middleton, J. A. (2013). A class of unbiased estimators of the average treatment
effect in randomized experiments. Journal of Causal Inference, 1:135–154.
Balzer, L. B., Petersen, M. L., and Laan, M. J. (2016). Targeted estimation and inference for the
sample average treatment effect in trials with and without pair-matching. Statistics in Medicine,
35:3717–3732.
Chiba, Y. (2015). Exact tests for the weak causal null hypothesis on a binary out come in randomized trials. Journal of Biometrics and Biostatistics, 6:doi:10.4172/2155–6180.1000244.
Cochran, W. G. (1977). Sampling Techniques. John Wiley & Sons, 3rd edition.
Copas, J. (1973). Randomization models for the matched and unmatched 2 × 2 tables. Biometrika,
60:467–476.
Ding, P. (2016). A paradox from randomization-based causal inference (with discussion). Statistical
Science, in press.
Ding, P. and Dasgupta, T. (2016). A potential tale of two by two tables from completely randomized
experiments. Journal of American Statistical Association, 111:157–168.
Efron, B. and Morris, C. (1973). Stein’s estimation rule and its competitors—an empirical bayes
approach. Journal of the American Statistical Association, 68:117–130.
Fisher, R. A. (1935). The Design of Experiments. Edinburgh, London: Oliver and Boyd, 1st edition.
Freedman, D., Pisani, R., and Purves, R. (2007). Statistics. W. W. Norton & Company, 4th edition.
Freedman, D. A. (2008a). On regression adjustments in experiments with several treatments. The
Annals of Applied Statistics, 2:176–196.
11
Freedman, D. A. (2008b). Randomization does not justify logistic regression. Statistical Science,
23:237–249.
Hinkelmann, K. and Kempthorne, O. (2008). Design and Analysis of Experiments, Volume 1:
Introduction to Experimental Design. New Jersey: John Wiley & Sons, Inc., 2nd edition.
Hodges, J. L. J. and Lehmann, E. L. (1964). Basic Concepts of Probability and Statistics. San
Francisco: Holden-Day.
Imai, K. (2008). Variance identification and efficiency analysis in randomized experiments under
the matched-pair design. Statistics in Medicine, 27:4857–4873.
Imbens, G. W. and Rubin, D. B. (2015). Causal Inference for Statistics, Social and Biometrical
Sciences: An Introduction. Cambridge: Cambridge University Press.
Kempthorne, O. (1952). The Design and Analysis of Experiments. New York: John Wiley and
Sons.
Lehmann, E. L. (1975). Nonparametrics: Statistical Methods Based on Ranks. San Francisco:
Holden-Day, 1st edition.
Lehmann, E. L. and Romano, J. P. (2008). Testing Statistical Hypotheses. New York : Wiley, 3rd
edition.
Li, X. and Ding, P. (2016). Exact confidence intervals for the average causal effect on a binary
outcome. Statistics in Medicine, 35:957–960.
Lin, W. (2013). Agnostic notes on regression adjustments to experimental data: Reexamining
Freedman’s critique. The Annals of Applied Statistics, 7:295–318.
Middleton, J. A. and Aronow, P. M. (2015). Unbiased estimation of the average treatment effect
in cluster-randomized experiments. Statistics, Politics and Policy, 6:39–75.
Miratrix, L. W., Sekhon, J. S., and Yu, B. (2013). Adjusting treatment effect estimates by poststratification in randomized experiments. Journal of the Royal Statistical Society: Series B
(Statistical Methodology), 75:369–396.
12
Neyman, J. (1923). On the application of probability theory to agricultural experiments. essay on
principles (with discussion). section 9 (translated). reprinted ed. Statistical Science, 5:465–472.
Neyman, J. (1935). Statistical problems in agricultural experimentation (with discussion). Supplement to the Journal of the Royal Statistical Society, 2:107–180.
Reichardt, C. S. and Gollob, H. F. (1999). Justifying the use and increasing the power of a t test
for a randomized experiment with a convenience sample. Psychological Methods, 4:117–128.
Rigdon, J. and Hudgens, M. G. (2015). Randomization inference for treatment effects on a binary
outcome. Statistics in Medicine, 34:924–935.
Robins, J. M. (1988). Confidence intervals for causal parameters. Statistics in Medicine, 7:773–785.
Rosenbaum, P. R. (2002). Observational Studies. New York: Springer, 2nd edition.
Rosenbaum, P. R. (2010). Design of Observational Studies. New York: Springer.
Rubin, D. B. (1990). Comment: Neyman (1923) and causal inference in experiments and observational studies. Statistical Science, 5:472–480.
Samii, C. and Aronow, P. M. (2012). On equivalencies between design-based and regression-based
variance estimators for randomized experiments. Statistics and Probability Letters, 82:365–370.
13
| 10math.ST
|
1
Superadditivity in trade-off capacities of quantum
channels
arXiv:1708.04314v2 [quant-ph] 16 Aug 2017
Elton Yechao Zhu, Quntao Zhuang, Min-Hsiu Hsieh, Senior Member, IEEE, and Peter W. Shor
Abstract—In this article, we investigate the additivity phenomenon in the dynamic capacity of a quantum channel for
trading classical communication, quantum communication and
entanglement. Understanding such additivity property is important if we want to optimally use a quantum channel for general
communication purpose. However, in a lot of cases, the channel
one will be using only has an additive single or double resource
capacity, and it is largely unknown if this could lead to an
superadditive double or triple resource capacity. For example,
if a channel has an additive classical and quantum capacity, can
the classical-quantum capacity be superadditive? In this work,
we answer such questions affirmatively.
We give proof-of-principle requirements for these channels to
exist. In most cases, we can provide an explicit construction of
these quantum channels. The existence of these superadditive
phenomena is surprising in contrast to the result that the
additivity of both classical-entanglement and classical-quantum
capacity regions imply the additivity of the triple capacity region.
Index Terms—Additivity; Quantum Channel Capacity; Tradeoff Capacity Regions; Quantum Shannon theory.
I. I NTRODUCTION
N studying classical communication, Shannon developed
powerful probabilistic tools that connect the theoretic
throughput of a channel to an entropic quantity defined on
a single use of the channel [1]. Shannon’s noiseless channel
coding theorem involves a random coding strategy to prove
achievability and entropic inequalities that show optimality,
i.e., the converse. This methodology has now become standard
in proving finite or asymptotic optimal resource conversions
in information theory.
Quantum Shannon information starts by mimicking classical information theory: typical sets can be generalized to
typical subspaces to prove achievability while various entropic
inequalities, such as the quantum data processing inequality,
can be used to prove the converse. However, the differences
between quantum and classical Shannon information are also
significant. On one hand, additional resources available in the
quantum domain diversify the allowable capacities, resulting in
I
E.Y. Zhu is with the Center of Theoretical Physics and Department of
Physics, Massachusetts Institute of Technology, Cambridge, MA, 02139 USA
(e-mail: [email protected]).
Q. Zhuang is with the Research Laboratory of Electronics and Department
of Physics, Massachusetts Institute of Technology, Cambridge, MA, 02139
USA (e-mail: [email protected]).
Min-Hsiu Hsieh is with the Centre of Quantum Software and Information (UTS |QSIi), University of Technology Sydney, Australia (e-mail: [email protected]).
P.W. Shor is with the Department of Mathematics, Massachusetts Institute
of Technology, Cambridge, MA, 02139 USA (e-mail: [email protected]).
trade-off regions for the resources that are consumed or generated [2]–[4]. The most common, and useful, quantum resource
in communication settings is quantum entanglement. Unlike
classical shared randomness, which does not increase a classical channel’s capability to send more messages, preshared
quantum entanglement will generally increase the throughput
of a quantum channel for sending classical messages or
quantum messages or both [2], [5]–[9]. It thus makes sense
to consider the trade-off capacity regions among these three
useful resources: entanglement, classical communication, and
quantum communication, and this was done in Ref. [4]. The
result in Ref. [4] further shows that a coding strategy that
exploits the channel coding of these three resources as a whole
performs better than strategies that do not take advantage of
channel coding.
On the other hand, single-lettered channel capacity formulas in the classical regime generally become intractable
regularized capacity formulas in the quantum regime [10]–
[14]. In other words, evaluation of these capacity quantities
requires optimizing channel inputs over an arbitrary finite
number of uses of a given channel. This largely blocks our
understanding of how quantum channels behave. An extreme
example shows the existence of two quantum channels that
cannot be used to send a quantum message individually but
will have a positive channel capacity when both are used
simultaneously [15]. However, there are also several examples
showing that when additional resources are used to assist,
the corresponding assisted capacity will also become additive.
The classical capacity over quantum channels is generally
superadditive; however, when assisted by a sufficient amount
of entanglement, the entanglement-assisted capacity becomes
additive [6], [16]. The quantum capacity also exhibits similar
properties. When assisted by either entanglement [2], [3] or an
unbounded symmetric side channel [17], its assisted quantum
capacity becomes additive.
This superadditive property of quantum channel capacities
has accordingly attracted significant attention. Hastings [18]
proved that the classical capacity over quantum channels is not
additive, a result built upon earlier developments by HaydenWinter [19] and Shor [20]. Recently, three of us showed a
rather perplexing result [21]: when assisted by an insufficient
amount of entanglement, a channel’s classical capacity could
be superadditive regardless of whether the unassisted classical
capacity is additive or not. Further, the additive property of
the entanglement-assisted classical capacity shows a form of
phase transition. Even if the channel is additive when assisted
by a sufficient amount of entanglement or no entanglement at
all, it can still be superadditive when assisted with an insuf-
2
ficient amount of entanglement. This phenomenon indicates
that quantum channels behave fundamentally differently from
classical channels, and our understanding of it is still quite
limited.
This paper is inspired by, and aims to extend Ref. [21]. Will
additivity of single or double resource capacities always lead
to additivity of a general resource trade-off region? We will
study superadditi vity in a general framework that considers
the three most common resources of: entanglement, noiseless
classical communication and quantum communication. Our
results show that (i) additivity of single resource capacities
of a quantum channel does not generally imply additivity of
double resource capacities, except for the known result [2] that
an additive quantum capacity yields an additive entanglementassisted quantum capacity region (see Table I); and (ii) additive double resource capacities does not generally imply an
additive triple resource capacity, except for the known case
[8] that additive classical-entanglement and classical-quantum
capacity regions yield an additive triple dynamic capacity (see
Table II). These results again demonstrate how complex a
quantum channel can be, and further investigation is required.
The paper is structured as follows. Section II introduces the
various definitions, notations and previous results on the triple
resource quantum Shannon theory. Section III summarizes the
various superadditivity results that we establish in the paper.
Section IV establishes the switch channel that we use for all
our constructions, and how this reduces the triple resource
trade-off formula. Section V gives a detailed construction of
all the possible superadditivity phenomena.
II. P RELIMINARIES
In this section, we give definitions of basic entropic quantities used in the paper. We also describe the dynamic capacity
theorem. Special cases of this include the various single and
double resource capacities. Finally, we define the elementary
channels that will be used in our explicit constructions.
A bipartite quantum state σAB is a positive semi-definite
matrix in Hilbert space H A ⊗ H B with trace one. We define
the von Neumann entropy, conherent information and quantum
mutual information of σAB , respectively, as follows:
S(AB)σ
= −Tr [σAB log σAB ],
I(AiB)σ
=
S(B)σ − S(AB)σ,
I(A; B)σ
=
S(A)σ + I(AiB)σ,
where S(A)σ is the von Neumann entropy of the reduced state
σA = TrB [σAB ].
x }
For an ensemble {p(x), σAB
x ∈X , let
Õ
x
σX AB =
p(x)|xihx|X ⊗ σAB
,
x ∈X
where {|xi} forms a fixed orthonormal (computational) basis
in Hilbert Space HX . We need the following information
quantities as well:
Õ
I(AiBX)σ =
p(x)I(AiB)σx ,
(1)
I(A; B|X)σ
=
x
Õ
p(x)I(A; B)σx ,
(2)
I(X; B)σ + I(A; B|X)σ,
(3)
x
I(AX; B)σ
=
where I(AiBX)σ and I(A; B|X) in Eqs. (1) and (2) are the
conditional coherent information and the conditional mutual
information, respectively. I(X; B)σ in Eq. (3) is the Holevo
information of σX B = Tr A[σX AB ].
A quantum channel N is a completely positive and tracepreserving map. With it, we can transmit either classical or
quantum information or both with possible entanglement assistance between the sender and the receiver [8]. More generally,
the authors in Ref. [4] proved the following capacity theorem
that involves a noisy quantum channel N and the three
resources mentioned above; namely, classical communication
(C), quantum communication (Q) and quantum entanglement
(E).
Theorem 1 (CQE trade-off [4]): The dynamic capacity
region CCQE (N ) of a quantum channel N is equal to the
following expression:
CCQE (N ) =
∞
Ø
1 (1)
CCQE N ⊗k ,
k
k=1
where the overbar indicates the closure of a set. The region
(1)
CCQE
(N ) is equal to the union of the state-dependent regions
(1)
CCQE,σ
(N ):
(1)
CCQE
(N ) ≡
The state-dependent region
C, Q and E, such that
Ø
σ
(1)
CCQE,σ
(N ) .
(1)
CCQE,σ
(N ) is the set of all rates
C + 2Q ≤ I(AX; B)σ,
(4)
Q + E ≤ I(AiBX)σ,
(5)
C + Q + E ≤ I(X; B)σ + I(AiBX)σ .
(6)
The above entropic quantities are with respect to a classicalquantum state (cq state) σX AB , where
Õ
σX AB ≡
p(x) |xi hx| X ⊗ NA0 →B φ xAA0 ,
(7)
x
φ xAA0
and the states
are pure.
We say that the dynamic capacity of a channel N is additive
if
(1)
CCQE (N ) = CCQE
(8)
(N ) .
The dynamic capacity region CCQE (N ) in Theorem 1
allows us to recover known capacity theorems by choosing
certain (C, Q, E) in Eqs. (4)-(6) as follows:
• the classical capacity CC (N ) when choosing Q = E = 0
[10], [11];
• the quantum capacity CQ (N ) when choosing C = E = 0
[12]–[14];
• the classical and quantum capacity CCQ (N ) when choosing E = 0 (CQ trade-off) [22];
• the entanglement assisted classical capacity CCE (N )
when choosing Q = 0 (CE trade-off) [7], [16];
• the entanglement assisted quantum capacity CQE (N )
when choosing C = 0 (QE trade-off) [2], [3];
Additivity of these special cases follows similarly from Eq. (8).
We note that the dynamic capacity region is concave, as a
convex combination of any two points in the region can be
3
achieved by a time-sharing strategy, i.e., using the channel for
a fraction of uses to achieve one point, and using it for the
other fraction to achieve the second point.
Below we will briefly describe a few channels which we
will repeatedly use.
Definition 2: A Hadamard channel is a quantum channel
whose complementary channel is entanglement breaking. Suppose Ψ A0 →B is a Hadamard channel, with the complementary
channel ΨcA0 →E . Then there is a degrading map DB→E such
that
ΨcA0 →E = DB→E ◦ Ψ A0 →B .
Moreover, D can be decomposed as
1
2
DB→E = DY→E
,
◦ DB→Y
where Y is a classical variable.
A Hadamard channel has an additive quantum dynamic capacity region, when tensored with an arbitrary quantum channel
[23]. Examples of Hadamard channels include the qubit dephasing channel, 1 → N cloning channels, and the Unruh
channel. We’ll define the qubit dephasing channel below, but
refrain from giving definitions of other Hadamard channels,
since their exact forms are not needed for understanding this
work. We refer the interested readers to Ref. [23] for more
details and properties of these channels.
dph
Definition 3: The qubit dephasing channel Ψη , with dephasing probability η, is defined as
dph
Ψη (ρ) = (1 − η)ρ + ηZ ρZ.
Imply the additivity of
CE
CQ
QE
N [21]
N [25]
N [25]
N (SecV-C)
N (SecV-C)
Y [3]
N (SecV-B)
N (SecV-D)
Y[3]
TABLE I
S UMMARY OF RESULTS FOR DOUBLE RESOURCES .“N” STANDS FOR “ DOES
NOT IMPLY ADDITIVITY ”, WHILE “Y” MEANS “ IMPLIES ADDITIVITY ”.
Additive capacities
C
Q
C,Q⇔ C,QE
Imply the additivity of
Additive capacities
CQE
QE
N (SecV-C)
CQ
N (SecV-G)
CE
N (SecV-E)
CE,Q⇔CE,QE
N (SecV-F)
CE,CQ
Y[8]
TABLE II
S UMMARY OF RESULTS FOR TRIPLE RESOURCES . “N” STANDS FOR “ DOES
NOT IMPLY ADDITIVITY ”, WHILE “Y” MEANS “ IMPLIES ADDITIVITY ”.
III. S UMMARY OF R ESULTS
We summarize all of our results here. We will denote the
single capacity region by a single letter, e.g. C for CC (N ).
We will also use short notation for double and triple tradeoff regions, e.g. CE for CCE (N ) and CQE for CCQE (N ). We
will use the arrow notation, with “→” meaning additivity of
the left-hand side capacity implies additivity of the right-hand
side capacity, and “6→” meaning additivity of the left-hand
side capacity does not imply additivity of the right-hand side
capacity.
dpo
Definition 4: The qubit depolarizing channel Ψp , with
depolarizing probability p, is defined as
I
dpo
Ψp (ρ) = (1 − p)ρ + p .
2
The qubit depolarizing channel is known to have an additive
classical capacity [24], but a superadditive quantum capacity
[25].
Definition 5: A random orthogonal channel Ψro is defined
as
D
Õ
|
Ψro (ρ) =
Pi Oi ρOi ,
i=1
where Oi are chosen from the orthogonal group and the
probabilities Pi are roughly equal.
For 1 D N, with N the input dimension, such a channel
will have a subadditive minimum output entropy with high
probability [18].
Definition 6: Consider an arbitrary channel ΨC→B . Append
a register R to the input, with a set of orthonormal bases {| ji}
and |R| = |B| 2 . We define its unitally extended channel [20],
[26] ΦRC→B as
Õ
ΦRC→B (ρRC ) =
X( j)ΨC→B (h j | ρRC | ji R ) X( j)†, (9)
A. Double resources (see table I)
1) CE:
a) C 6→ CE [21]: There exists a quantum channel N ,
such that its classical capacity is additive, but its CE
trade-off capacity region is superadditive. We will give
a simplified construction in Sec V-A.
b) C, Q 6→ CE: There exists a quantum channel N ,
such that its classical and quantum capacities are
both additive, but its CE trade-off capacity region is
superadditive, i.e.,, ∃ a quantum channel N s.t.
CC (N ) = CC(1) (N )
and
(1)
CQ (N ) = CQ
(N )
but
(1)
CCE (N ) ) CCE
(N ) .
j
where {X( j) : j ∈ {1, . . . , |R|}} are the Heisenberg-Weyl
operators.
The unital extension of a random orthogonal channel will have
a superadditive classical capacity with high probability [20].
An explicit construction of N is given in Sec V-B.
2) Q → QE [3]: For all quantum channels N , if its quantum
capacity is additive, then its QE trade-off capacity region
is always additive.
4
3) CQ:
a) C 6→ CQ [25]: There exists a quantum channel N , such
that its classical capacity is additive, but its CQ tradeoff capacity region is superadditive. The depolarizing
channel has a superadditive quantum capacity and
hence a superadditive CQ trade-off capacity, while its
classical capacity is additive.
b) Q 6→ CQ: There exists a quantum channel N , such
that its quantum capacity is additive, but its CQ tradeoff capacity region is superadditive, i.e., ∃ a quantum
channel N s.t.
(1)
CQ (N ) = CQ
(N )
but
(1)
CCQ (N ) ) CCQ
(N ) .
A construction of this example quantum channel is
given in Sec V-C.
c) C, Q 6→ CQ: Moreover, there exists a quantum channel
N , such that its classical and quantum capacities
are additive, but its CQ trade-off capacity region is
superadditive, i.e., ∃ a quantum channel N s.t.
CC (N ) = CC(1) (N )
and
but
(1)
CQ (N ) = CQ
(N ) ,
(1)
CCQ (N ) ) CCQ
(N ) .
A construction of this example quantum channel is
given in Sec V-D.
B. Triple resources (see table II)
1) CE 6→ CQE: There exists a quantum channel N such
that its CE trade-off capacity region is additive, but
its dynamic capacity region is superadditive, i.e., ∃ a
quantum channel N s.t.
(1)
CCE (N ) = CCE
(N )
but
(1)
CCQE (N ) ) CCQE
(N ) .
An example is constructed in Sec V-E.
2) CE, Q 6→ CQE: There exists a quantum channel N such
that its quantum capacity and its CE trade-off capacity
region are additive, but its dynamic capacity region is
superadditive, i.e., ∃ a quantum channel N s.t.
(1)
CQ (N ) = CQ
(N )
and
but
(1)
CCE (N ) = CCE
(N ) ,
(1)
CCQE (N ) ) CCQE
(N ) .
An example is constructed in Sec V-F.
3) CQ 6→ CQE: There exists a quantum channel N such
that its CQ trade-off capacity region is additive, but
its dynamic capacity region is superadditive, i.e., ∃ a
quantum channel N s.t.
(1)
CCQ (N ) = CCQ
(N )
but
(1)
CCQE (N ) ) CCQE
(N ) .
An example is given in Sec V-G.
4) CE, CQ → CQE [8]: If a quantum channel N has
additive CE and CQ trade-off capacity regions, then its
dynamic capacity region is also additive. This statement
is first observed in Ref. [8], and an explicit argument can
be found in Ref. [23].
IV. F RAMEWORK
This section presents technical tools that we require for
demonstration of superadditivity in trade-off capacities. We
first define the concept of switch channels.
0
Definition 7: A switch channel NMC→B between NC→B
1
and NC→B with M being a 1-bit switch register is defined as
NMC→B (ρ MC )
0
1
=NC→B
(h0| ρ MC |0i M ) + NC→B
(h1| ρ MC |1i M ) .
In quantum information theory, switch channels were first
used in Ref. [7] to demonstrate the existence of quantum
channels such that the quantum capacity is nonzero, but for
which pre-shared entanglement does not improve the classical
capacity. Subsequently, they are used in Ref. [27] to show
the superadditivity of private information, with an alternative
definition. Recently, they are also used in Ref. [21] to show
the superadditivity of the classical capacity with limited entanglement assistance.
One immediate difficulty is that, even if N 0 and N 1 are
well-studied, the dynamic capacity region of N may not
always have a simple expression in terms of those of N 0 and
N 1 . This is due to the fact that the switch register M can be
in a superposition state. However, if N 0 and N 1 are unitally
extended channels, then the dynamic capacity region of N
does have a simple expression.
Lemma 8: Consider a switch channel NA0 →B between
0
1
NRC→B
and NRC→B
, with input partition A0 = M RC and M
0
1
being a switch register. Here NRC→B
and NRC→B
are unital
0
1
extensions of ΨC→B and ΨC→B respectively. Then
(1)
(1)
(1)
CCQE
(N ) = Conv CCQE N 0 , CCQE N 1 ,
where Conv denotes the convex hull of points from the two
sets.
If the quantum dynamic capacity region for N 0 ⊗ Ψ is
additive for any Ψ, then we also have
CCQE (N ) = Conv CCQE N 0 , CCQE N 1 .
The rest of this section is devoted to the proof of this lemma.
Firstly, we note that switch channels and unitally extended
channels fall under a broader class of channels that we call
partial classical-quantum channels (partial cq channels).
5
Definition 9: A channel ΨRC→B is a partial cq channel if
there exists a noiseless classical channel ΠR→R with orthonormal basis {| ji R }, such that
ΨRC→B = ΨRC→B ◦ ΠR→R .
x, j
(12)
with its output state
Õ
x
p(x) |xi hx| X ⊗ ς AB
,
x
x = Ψ 0
x
where ς AB
A →B φ AA0 , there exists a corresponding state
ρX AA0 , in the form of Eq. (11), which can achieve the same
rate, if not better.
In fact, the state ρX AA0 can be obtained by applying ΠR→R
on %X AA0 and expanding its classical register X. This can be
achieved by the following quantum instrument T : R → RXR ,
Õ
T (ψR ) :=
h j | ψR | ji | ji h j | R ⊗ | ji h j | XR
so that
=
=
0)
T (%
Õ X AA
xj
p(x, j) |x, ji hx, j | X ⊗ | ji h j | R ⊗ φ AC (13)
x, j
where we abuse the notation X to denote X XR in Eq. (13),
p(x, j) ≡ p(x)p( j |x), and p( j |x) = Tr[| ji h j | φ xAC ].
Let σX AB = Ψ A0 →B (ρX AA0 ). Then
Õ
xj
σX AB =
p(x, j) |x, ji hx, j | X ⊗ σAB
x, j
where
≥
Õ
p(x)I(AiB)ς x
x
= I(AiBX)ς ,
(15)
where the inequality is due to Eq. (14) and the convexity
of coherent information with respect to inputs.
2) Now consider I(AX; B)σ . Similarly,
I(AX; B)σ = S(B)σ + I(Bi AX)σ
≥ S(B)ς + I(Bi AX)ς
= I(AX; B)ς ,
where the inequality is due to σB = ςB and Eq. (15).
3) Finally consider I(X; B)σ . Writing |x, ji as |xi | ji, it can
be shown
I(X; B)σ ≥ I(X; B)ς
Lemma 11: The optimal trade-off surface of the 1-shot quantum dynamic capacity region of a unitally extended channel
can always be achieved with σX AB such that S(B)σ = log(|B|).
This extends similarly to higher shots.
Proof. Suppose ΦRC→B is unitally extended from ΨC→B .
Since a unitally extended channel ΦRC→B is a partial cq
channel, by Lemma 10, we can consider states of the form
Õ
xj
%X AA0 =
p(x, j) |x, ji hx, j | X ⊗ | ji h j | R ⊗ φ AC .
x, j
xj
p( j |x)σAB .
x jk
xj
xj
xj
where ς AB = X(k)ς AB X(k)† and ς AB = ΨC→B φ AC .
We can construct another state of the form in Eq. (11):
Õ
xj
ρX 0 AA0 =
p(x, j, k) |x, j, ki hx, j, k | X 0 ⊗ |ki hk | R ⊗ φ AC ,
x, j,k
(16)
where p(x, j, k) = p(x, j)/|R|, and σX 0 AB = ΦRC→B (ρX 0 AA0 ):
Õ
x jk
σX 0 AB =
p(x, j, k) |x, j, ki hx, j, k | X 0 ⊗ σAB ,
xj
xj
xj
where
= X(k)σAB X(k)† and σAB = ΨC→B φ AC . The
state σX 0 AB satisfies
x jk
σAB
It follows that
Õ
x, j
x, j,k
xj
xj
σAB = Ψ A0 →B | ji h j | R ⊗ φ AC .
x
ς AB
=
p(x)p( j |x)I(AiB)σ x j
x, j
Let ςX AB = ΦRC→B (%X AA0 ) with A0 ≡ RC. Then
Õ
xjj
ςX AB =
p(x, j) |x, ji hx, j | X ⊗ ς AB ,
j
ρX AA0
=
Õ
using the data processing inequality when we apply the
partial trace map |xi hx| ⊗ | ji h j | → |xi hx| to σX B .
x
ςX AB ≡ Ψ A0 →B (%X AA0 ) =
x, j
(10)
If there is no register C, then such channels are classicalquantum channels (cq channels).
For partial cq channels, one can always assume inputs are
cq states with respect to the input partition R and C for the
purpose of evaluating capacities, as we show in Lemma 10
below.
Lemma 10: If Ψ A0 →B is a partial cq channel with partition
A0 = RC, then the optimal trade-off surface of the 1-shot
(1)
dynamic capacity region CCQE
(Ψ) can be achieved with
respect to cq states σX AB = Ψ A0 →B (ρX AA0 ), where ρX AA0 is
of the form
Õ
xj
(11)
ρX AA0 =
p(x, j) |x, ji hx, j | X ⊗ | ji h j | R ⊗ φ AC .
Proof. We will show that, for any input state
Õ
%X AA0 =
p(x) |xi hx| X ⊗ φ xAA0,
1) First consider I(AiBX)σ .
Õ
I(AiBX)σ =
p(x, j)I(AiB)σ x j
(14)
j
Since the dynamic capacity region is fully determined by the
three entropic quantities I(AX; B)σ , I(AiBX)σ and I(X; B)σ
in Eqs. (4)-(6), it suffices to show that all three entropic
quantities evaluated on ρX AA0 are greater than those evaluated
on %X AA0 .
©Õ
x jk ª
S(B)σ = S
p(x, j, k)σB ®
«x, j,k
¬
!
Õ
1 Õ x jk
σ
≥
p(x, j)S
|R| k B
x, j
= log(|B|),
6
where we’ve used the qudit twirl formula [28]
1 Õ x jk
1 Õ
1
xj
σB =
X(k)σB X(k)† =
IB .
|R| k
|R| k
|B|
Likewise,
(17)
x, j,k
=
p(x, j)I(AiB)ς x j j = I(AiBX)ς
=
1
Õ
p(x, m, k)I(Bi A)σ x mk
pm I(AX; B)σ m
m=0
and
I(X; B)σ =
(18)
x, j
1
Õ
pm I(X; B)σ m .
m=0
I(AX 0; B)σ = S(B)σ +
Õ
p(x, j, k)I(Bi A)σ x j k
x, j,k
= log(|B|) +
Õ
p(x, j)I(Bi A)ς x j j ≥ I(AX; B)ς
x, j
I(X 0; B)σ
1 Õ
Õ
m=0 x,k
One can verify that the dynamic capacity region with σX 0 AB
is larger than that with ςX AB as follows:
Õ
I(AiBX 0)σ =
p(x, j, k)I(AiB)σ x j k
Õ
I(AX; B)σ = log(|B|) +
= S(B)σ −
Õ
p(x, j, k)S(B)σ x j k
(19)
x, j,k
= log(|B|) −
Õ
p(x, j)S(B)ς x j j ≥ I(X; B)ς .
This means if we consider inputs of the form (21), the
triple rate for using N can always be expressed as a linear
combination of the triple rates of N 0 and N 1 . It is also clear
that any linear combination is achievable by the time-sharing
principle. Since using states of the form (21) is optimal, we
have
Ø
(1)
(1)
(1)
CCQE
pCCQE
N 0 + (1 − p)CCQE
N1
(N ) =
0≤p ≤1
x, j
(1)
(1)
= Conv CCQE
N 0 , CCQE
N1 .
(20)
The key property used in the above equations is, for any
Heisenberg-Weyl operator X(k),
S(σB ) = S(X(k)σB X(k)† ).
Proof of lemma 2. Following from Lemma 11 and Eq. (16),
we only need to consider states of the form
ρX AA0 =
1
Õ
pm |mi hm| M ⊗ ρm
X ARC
(21)
m=0
Here, addition means Minkowski sum1 . Similarly, we have
(1)
(1)
CCQE (N ⊗ N ) =Conv CCQE
N0 ⊗ N0 ,
(1)
(1)
CCQE
N 0 ⊗ N 1 , CCQE
N1 ⊗ N1 .
If the quantum dynamic capacity region is additive for N 0 ⊗Ψ,
for any Ψ, then
(1)
(1)
(1)
CCQE
N 0 ⊗ N 1 = CCQE
N 0 + CCQE
N1 .
(24)
Í
where pm = x,k p(x, m, k) and
Õ p(x, m, k)
=
ρm
|x, m, ki hx, m, k | X ⊗ |ki hk | R ⊗ φ xm
X ARC
AC ,
p
m
x,k
In this case the 1-shot quantum dynamic capacity region for
N ⊗ N can be greatly simplified to
(1)
(1)
CCQE
(N ⊗ N ) = Conv 2CCQE N 0 , CCQE N 1 ⊗ N 1 .
with p(x, m, k) = p(x, m, k 0) for all k, k 0 and m ∈ {0, 1}.
The corresponding channel output is
Similarly,
σX AB =
1
Õ
pm σXmAB
(22)
m=0
where
σXmAB =
Õ p(x, m, k)
xmk
|x, m, ki hx, m, k | X ⊗ σAB
p
m
x,k
(23)
and
xmk
m
†
σAB
= X(k)ΨC→B
φ xm
AC X(k) .
Then all three of the entropic quantities evaluated on σX AB
in Eq. (22) can be decomposed to the corresponding ones
evaluated on σXmAB given in Eq. (23):
I(AiBX)σ =
1 Õ
Õ
p(x, m, k)I(AiB)σ x mk
m=0 x,k
=
1
Õ
m=0
pm I(AiBX)σ m .
(1)
CCQE
N ⊗k
⊗k−1
⊗k
(1)
(1)
=Conv CCQE
N1
, CCQE
N0 ⊗ N1
,
⊗k
⊗k−1
(1)
(1)
.
· · · , CCQE
N0
⊗ N 1 , CCQE
N0
⊗m
⊗k−m
(1)
N0
⊗ N1
, 0 ≤ m ≤ k, can be
Each term CCQE
upper bounded as
⊗k−m
⊗m
(1)
0
⊗ N1
CCQE N
⊗k−m
(1)
=mCCQE N 0 + CCQE
N1
⊆mCCQE N 0 + (k − m)CCQE N 1
⊆kConv CCQE N 0 , CCQE N 1 .
1 For two sets of position vectors A and B in Euclidean space, their
Minkowski sum A + B is obtained by adding each vector in A to each vector
in B, i.e., A + B = {a + b |a ∈ A, b ∈ B } [29].
7
Here the second line follows from the addivity of the dynamic
capacity region of N 0 . The third line follows from the definition of CCQE . The fourth line follows from the definition of
(1)
convex hull. Thus CCQE
N ⊗k can also be upper bounded as
(1)
CCQE
N ⊗k ⊆ kConv CCQE N 0 , CCQE N 1 .
and
∞
Ø
1 (1)
CCQE N ⊗k
CCQE (N ) =
k
k=1
⊆Conv CCQE N 0 , CCQE N 1
=Conv CCQE N 0 , CCQE N 1 .
The last equality follows because of the topology of the
dynamic capacity region, as we show in Appendix D.
By a time-sharing protocol, it is obvious that
CCQE (N ) ⊇ Conv CCQE N 0 , CCQE N 1 .
In each construction, we first state the properties that N 0
and N 1 need to satisfy, in addition to Property (U). We then
show how the desired superadditivity of the switch channel
N follows from these properties. In the end, we explicitly
construct channels that satisfy the properties we required.
Before we start, we first propose two families of unital
extended channels that satisfy (U). Many of our explicit
constructions of N 0 will be chosen from these candidates. The
first family comes from unital extensions of Hadamard channels. The following lemma shows that the dynamic capacity
of the unitally extended Hadamard channels is also additive.
Lemma 12: The dynamic capacity region is additive for Φ0
and any other channel Ψ1 , if Φ0 is a unital extension of a
Hadamard channel Ψ0 .
The second family is unital extensions of classical channels.
Lemma 13: If Ψ0 is a classical channel, then the dynamic
capacity region is additive for Ψ0 ⊗ Ψ1 , for arbitrary Ψ1 . The
same holds for a unital extension of a classical channel.
The proofs of the above lemmas are left to the Appendices,
as they are not essential in understanding the construction.
Hence
CCQE (N ) = Conv CCQE N 0 , CCQE N 1 .
Note that unital extensions are not unique, and we only
used the unitarity of Heisenberg-Weyl operators and the twirl
formula Eq. (17) in proving the above lemmas. Hence, as long
as we have K unitaries {Uk } ∈ U(d) that satisfy the twirl
formula
I
1 Õ
Uk AUk† = Tr(A)
(25)
K k
d
for any d × d matrix A, one has a valid unital extension, and
lemmas 8 and 11 will hold. 2
Moreover, unital extensions are preserved under tensor
product of channels: if Φ1 is a unital extension of Ψ1 , and Φ2 is
a unital extension of Ψ2 , then Φ1 ⊗Φ2 is also a unital extension
of Ψ1 ⊗Ψ2 . This follows from the fact that if {U j } ∈ U(d1 ) and
{Vk } ∈ U(d2 ) both satisfy Eq. (25), then {U j ⊗ Vk } ∈ U(d1 d2 )
also satisfies Eq. (25).
V. E XPLICIT C ONSTRUCTION OF VARIOUS
S UPERADDITIVITY P HENOMENA
With the tools developed in Sec IV, we can now explicitly
construct channels that satisfy the superadditivity properties
stated in Sec III. All our constructions utilize the switch
channel idea. We always assume that N is a switch channel
of two unitally extended channels N 0 and N 1 . Further, we
assume that
(U) N 0 has an additive dynamic capacity region, when tensored with another arbitrary channel.
In this setting, we can use Lemma 8 and its reduction to
various single-resource and two-resource capacities.
2 Note that we do not even require N 0 and N 1 to have the same unital
extension. However, to ensure the input dimensions of N 0 and N 1 are the
same, their unital extensions must involve the same number of unitaries. For
this reason, we stick with the Heisenberg-Weyl operators most of the time.
A. Additive C, Superadditive CE
Here we review the original argument in [21] and recast it
in the current framework.
We use CP (N ) when we view C (N ) as a function of the
amount of entanglement assistance P, where (C (N ) , P) are
points on the CE trade-off curve of N . When P = 0, we
return to the classical capacity CC (N ). When P is maximal,
we arrive at the classical capacity with unlimited entanglement
assistance CE (N ). CP(1) (N ) denotes the 1-shot case.
We require N 0 and N 1 to have the following properties:
(A1) CC N 0 = CC N 1 .
(A2) N 1 has a superadditive CE trade-off capacity region, i.e.,,
(1)
CCE N 1 ) CCE
N1 ,
and CCE N 1 is strictly concave and superadditive at a
boundary point of the trade-off region with entanglement
consumption
P̄.
(A3) CCE N 0 ( CCE N 1 in the sense the CE trade-off
capacity region of N 0 is strictly smaller than that of N 1
when entanglement consumption is at P̄.
In the
property (A2) means at P = P̄,
CP notation,
CP N 1 > CP(1) N 1 and CP N 1 is strictly concave3 in P
at P = P̄ . Property (A3) implies that CP N 0 < CP N 1 at
P = P̄.
Note that these properties are weaker than the ones required
in Ref. [21].
These three properties (A1)-(A3), together with (U), will
guarantee that (i) the classical capacity of N is additive; and
(ii) the CE trade-off capacity region of N is superadditive at
entanglement consumption rate P̄.
3 Here by saying a function f is strictly concave at y, we mean f (y) >
(1 − p) f (v) + p f (w) for all v < y < w satisfying (1 − p)v + pw = y, with
p ∈ (0, 1).
8
Combining property (A1) with (U) yields statement (i):
n
o
CC (N ) = max CC N 0 , CC N 1 = CC N 0
n
o
= max CC(1) N 0 , CC(1) N 1 = CC(1) (N ) ,
where Lemma 8 is used in the first equality.
Property (A3) ensures that
CCE (N ) = Conv CCE N 0 , CCE N 1
= CCE N 1 .
(N ) = Conv CCE N
0
(1)
, CCE
N
1
CP(1) (Φro ) ≤ CC(1) (Φro ) + P,
(28)
CE (Φro ) ≤ CC (Φro ) + log(N) − .
This implies dCP (Φro ) /dP cannot always be 1. Thus there
exists P̄ ∈ [0, log(N)) such that
dCP (Φro ) /dP = 1, ∀ 0 ≤ P ≤ P̄
and
(26)
Since
(1)
CCE
Since
,
there exists P0, P1 ≥ 0 and p ∈ [0, 1] such that pP0 +(1−p)P1 =
P̄ and
CP̄(1) (N ) = pCP0 N 0 + (1 − p)CP(1)1 N 1 .
Statement (ii) follows after considering three different cases.
1) p = 0.
CP̄(1) (N ) = CP̄(1) N 1 < CP̄ N 1 = CP̄ (N ) ,
where the inequality follows from the superadditivity part
of property (A2). The second equality follows from Eq.
(26).
2) 0 < p < 1.
CP̄(1) (N ) = pCP(1)0 N 0 + (1 − p)CP(1)1 N 1
≤ pCP0 N 1 + (1 − p)CP1 N 1
< CP̄ N 1 = CP̄ (N )
where the first inequality follows from Property (A3). The
second inequality follows from the strict concavity part
of property (A2). The last equality follows from Eq. (26).
3) p = 1. Then
CP̄(1) (N ) = CP̄ N 0 < CP̄ N 1 = CP̄ (N ) .
dCP (Φro ) /dP < 1, ∀P > P̄.
Next we discuss different cases of P̄.
1) P̄ > 0. Then CP (Φro ) is strictly concave at P̄. Furthermore, CP̄ (Φro ) − CP̄(1) (Φro ) ≥ since CP̄ (Φro ) =
CC (Φro ) + P̄ but CP̄(1) (Φro ) ≤ CC(1) (Φro ) + P̄. Thus
N 1 = Φro satisfies (A2).
dph
dph
2) P̄ = 0. Let N 1 = Φro ⊗ Φη , where Φη is the unital
extension of the qubit dephasing channel.
Since dCP (Φro ) /dP|0+ < 1, choose η > 0 small such that
dph
dCP Φη /dP|1− > dCP (Φro ) /dP|0+ . This is possible,
dph
dph
dph
as CP Φη
= CP Ψη
and dCP Ψη /dP|1− → 1
as η → 0. This ensures that when 0 < P ≤ 1,
dph
(29)
CP N 1 = CC (Φro ) + CP Φη ,
where we’ve also used Lemma 12.
dph
dph
For Φη , it can be shown that CP Φη
is strictly
concave in P when η < 1/2 (see Appendix C). Hence
CP N 1 is also strictly concave with respect to P, for
0 < P ≤ 1. Also, when P < ,
dph
CP N 1 > CC (Φro ) + CC Φη
dph
> CC(1) (Φro ) + CC Φη + P ≥ CP(1) N 1 .
Here the first inequality
comes from Eq. (29) and
dph
dph
CP Φη
> CC Φη
when P > 0. The second
inequality comes from our assumption P < and Eq.
(27). The last inequality comes
from Eq. (28).
This ensures that CP N 1 is superadditive.
Thus when
0 < P < min{1, }, CP N 1 is strictly concave and
superadditive.
For N 0 , as long as
extension of a classical
it is a unital
channel with CC N 0 = CC N 1 , it will automatically satisfy
property (A3).
Here the first equality follows from additivity of the CE
trade-off capacity region for N 0 . The inequality follows
from property (A3). The last equality follows from Eq.
(26).
Explicit Construction of N : We quote the following property about concave functions [30]: A concave function u(y)
is continuous, differentiable from the left and from the right.
The derivative is decreasing, i.e., for x < y we have u 0(x−) ≥
B. Additive C and Q, Superadditive CE
u 0(x+) ≥ u 0(y−) ≥ u 0(y+). We use “±” to denote the right and
In Section V-A, we constructed a channel N with an
left derivatives when needed.
1
ro
additive
classical capacity, but a superadditive CE trade-off
We first construct N . Choose Ψ to be a random orthogcapacity
region. It’s unclear if our construction N has an
onal channel with a subadditive minimum output entropy, and
ro
ro
additive
quantum
capacity. To extend the argument, we need
Ψ has input dimension N. This is unitally extended to Φ .
to
make
some
modifications
to the original construction.
Due to Lemma 10, the useful entanglement assistance is at
In
addition
to
properties
(A1)-(A3),
the channels N 0 and
most log(N). Thus we restrict to 0 ≤ P ≤ log(N).
1
N need to satisfy
Let
(1)
ro
ro
= CC (Φ ) − CC (Φ ) > 0.
(27) (B1) CQ N 0 ≥ CQ N 1 .
9
then (B1) is automatically satisfied. Hence we will focus on
the case where
CQ N 0 < CQ N 1 .
In this case, we call these two channels Φ0 and Φ1 respectively.
We will construct two new channels N 0 and N 1 that satisfy
properties (A1)-(A3) and (B1).
We will use the qubit dephasing channel and 1 → N cloning
channel. To make the argument work, we will modify them in
the following manner.
For the 1 → N cloning channel Ψ1→N , we always tensor an
appropriate classical channel, such that the resulting channel
has its classical capacity equal to 1, and the output dimension
is the same as the input dimension. We denote the resulting
channel Ψ N .
For the dephasing channel, we will tensor a complete
depolarizing channel, so that its input and output dimensions
match those of Ψ N . Since tensoring a complete depolarizing
channel does not modify the dynamic capacity region of the
dph
qubit dephasing channel, we will continue using Ψη to
denote it.
Based on results in Ref. [23], we can obtain the trade-off
dph
capacities of the qubit dephasing channel Ψη and modified
N
1 → N cloning channel Ψ . We observe that for η = 0.2
and N = 15, their trade-off capacities satisfy the following
properties (see Fig. 1)
dph
CQ Ψη
> CQ Ψ N .
and
dph
CCE Ψη
( CCE Ψ N ,
(30)
in the sense that Ψ N achieves a strictly better classical comdph
munication rate than Ψη , if we have any non-zero amount
of entanglement
assistance. In the CP notation, it means
dph
CP Ψη
< CP Ψ N for all P > 0.
Since unital extensions do not change the CE and CQ tradeoff capacity regions of these two channels (see Appendix C),
dph
the above properties hold if we replace Ψη and Ψ N by their
dph
N
unital extensions Φη and Φ respectively.
Since
dph
CQ Φη > CQ Φ N ,
let n be large enough so that
dph
nCQ Φη + CQ Φ0 ≥ nCQ Φ N + CQ Φ1 .
0.8
2.0
Ψ N,
N = 15
Ψ ηdph , η = 0.2
0.6
0.4
0.2
0.0
0.0
0.2
0.4
0.6
0.8
1.0
Classical Communication Rate
Classical Communication Rate
Explicit Construction of N : We take the channels N 0 and
1
N that were constructed in Sec V-A, and compare their
quantum capacities. Since CQ N 0 = 0, we can only have
CQ N 0 ≤ CQ N 1 . If
CQ N 0 = CQ N 1 ,
1.0
Quantum Communication Rate
This ensures that the quantum capacity of N is also additive:
n
o
CQ (N ) = max CQ N 0 , CQ N 1 = CQ N 0
n
o
(1)
(1)
= max CQ
N 0 , CQ
N 1 = CQ (N ) .
1.8
Ψ N,
N = 15
Ψ ηdph , η = 0.2
1.6
1.4
1.2
1.0
0.0
0.2
0.4
0.6
0.8
1.0
Entanglement Consumption Rate
(a)
(b)
Fig. 1. Comparison of trade-off curves between qubit dephasing channel
dph
Ψη and modified 1 → N cloning channel Ψ N , when η = 0.2 and N = 15.
(a) CQ trade-off. (b) CE trade-off.
Define
and
⊗n
dph
N 0 = Φη
⊗ Φ0
⊗n
N 1 = ΦN
⊗ Φ1 .
Our choice of n ensures that
CQ N 0 ≥ CQ N 1 .
We also need to ensure our newly constructed N 0 and N 1
still satisfy properties (A1)-(A3).
As
dph
CC Φη = CC Φ N = 1
and
CC Φ0 = CC Φ1 ,
we immediately have
CC N 0 = CC N 1
and property (A1) is satisfied.
The CE trade-off curve of Ψ1→N is strictly concave for
N , 1 [23], hence property (A2) is also satisfied for N 1 .
Property (A3) is satisfied due to Eq. (30).
C. Additive Q, Superadditive CQ
We require N 0 and N 1 to have the following properties:
(C1) CQ N 0 ≥ CQ N 1 .
(1)
(C2) CC N 1 > CC N 1 .
(C3) CC N 0 < CC N 1 .
These properties (C1)-(C3) will allow us to show that (i)
(1)
(1)
CQ (N ) = CQ
(N ); and (ii) CCQ (N ) ) CCQ (N ) .
Statement (i) follows from property (C1) and (U) that N 0
has an additive quantum capacity:
n
o
CQ (N ) = max CQ N 0 , CQ N 1 = CQ N 0
n
o
(1)
(1)
(1)
= max CQ
N 0 , CQ
N 1 = CQ
(N ) .
Properties (C2) and (C3) together ensure
n
o
CC (N ) = max CC N 0 , CC N 1 = CC N 1
n
o
> max CC(1) N 0 , CC(1) N 1 = CC(1) (N ) ,
10
i.e., the classical capacity of N is superadditive; hence statement (ii) follows.
Explicit Construction of N : Next we construct N 0 and N 1
that satisfy the above properties.
Let Ψro be a random orthogonal channel, such that its
unital extension has a superadditive classical capacity. For
convenience, we also assume Ψro has the input dimension
dph
N = 2n . Choose η for
the qubit dephasing channel Ψη such
dph
that CQ (Ψro ) + CQ Ψη = m for some integer m.
Define
dph
N 1 = Φro ⊗ Φη ,
dph
where Φro is a unital extension of Ψro and Φη is a unital
dph
extension of Ψη . N 1 has the property that its quantum
capacity is CQ N 1 = m, whereas its classical capacity is
superadditive, and greater than m.
Define
⊗m
⊗n+1−m
dpo
N 0 = ΦI
⊗ Φ1
,
where Φ I is a unital extension of the noiseless qubit channel,
dpo
and Φ1 is a unital extension of the complete qubit depolarizing channel.
0
It’s clear
that N 0 has its classical and quantum capacity as
0
CC N = CQ N = m, thus fulfiling the properties (C1)
and (C3) above.
D. Additive C and Q, Superadditive CQ
We require N 0 and N 1 to satisfy the following properties:
(D1) CC N 0 ≤ CC N 1 = CC(1) N 1 and CQ N 0 =
CQ N 1 .
(D2) N 1 has a superadditive CQ trade-off capacity region,
meaning
(1)
CCQ N 1 ) CCQ
N1 .
CCQ N 1 is strictly concave and superadditive at a
boundary point with classical
communication rate C̄.
(D3) CCQ N 0 ( CCQ N 1 in the sense the CQ trade-off
capacity region of N 0 is strictly smaller than that of N 1
when classical communication rate is at C̄.
With these properties, we can show that (i) CC (N ) =
(1)
CC(1) (N ); (ii) CQ (N ) = CQ
(N ); and (iii) CCQ (N ) )
(1)
CCQ
(N ) .
We’ll focus on the CQ trade-off curve. Same as in Section
V-A, we use a simplified notation QC (N ) when we view
Q (N ) as a function of C (N ). In the 1-shot scenario, it is
(1)
denoted by QC
(N ). We’ll show there exists C̄ , 0 such that
(1)
QC̄ (N ) > QC̄ (N ).
In the QC notation,
means at C = C̄,
property (D2)
(1)
QC N 1 > QC
N 1 and QC N 1 is strictly concave in C
at C = C̄. Property (D3) implies that QC N 0 < QC N 1 at
C = C̄.
Property (D1) and (U) that N 0 has an additive quantum
capacity ensure that
n
o
CC (N ) = max CC N 0 , CC N 1 = CC N 1
n
o
= max CC(1) N 0 , CC(1) N 1 = CC(1) (N )
and
n
o
CQ (N ) = max CQ N 0 , CQ N 1 = CQ N 0
n
o
(1)
(1)
(1)
= max CQ
N 0 , CQ
N 1 = CQ
(N ) ,
i.e., N has an additive classical and quantum capacity.
By property (D3), we have
CCQ (N ) = Conv CCQ N 0 , CCQ N 1
= CCQ N 1 .
(31)
Since
(1)
(1)
CCQ
(N ) = Conv CCQ N 0 , CCQ N 1 ,
there exists C0, C1 and p ∈ [0, 1] such that pC0 + (1 − p)C1 = C̄
and
(1)
(1)
QC̄
(N ) = pQC0 N 0 + (1 − p)QC1 N 1 .
Now consider three different cases.
1) p = 0.
(1)
(1)
QC̄
(N ) = QC̄ N 1 < QC̄ N 1 = QC̄ (N ) ,
where the inequality follows from property (D2). The
second equality follows from Eq. (31).
2) 0 < p < 1.
(1)
(1)
QC̄
(N ) =pQC0 N 0 + (1 − p)QC1 N 1
≤pQC0 N 1 + (1 − p)QC1 N 1
<QC̄ N 1 = QC̄ (N ) .
Here the first inequality follows from the definition of
CCQ and property (D3). The second inequality follows
from the strict concavity part of property (D2). The last
equality follows from Eq. (31).
3) p = 1. Then
(1)
QC̄
(N ) = QC̄ N 0 < QC̄ N 1 = QC̄ (N ) .
Here the inequality follows from property (D3). The last
equality follows from Eq. (31).
Hence statement (iii) follows.
Explicit Construction: Now we explicitly construct N 0 and
N 1.
dpo
Choose p such that the qubit depolarizing channel Ψp is
known to have a superadditive quantum capacity. Consider
its
dpo
dpo
unital extension Φ p . Note that the gradient dQC Φ p /dC
of the CQ trade-off curve cannot always stay at 0 for the choice
dpo
of Ψp with a positive
quantum capacity. It means there exists
dpo
0 ≤ C̄ < CC Φ p
such that
dpo
dQC Φ p /dC = 0, ∀0 ≤ C ≤ C̄−
and
dpo
dpo
dQC Φ p /dC < 0, ∀C̄+ ≤ C ≤ CC Φ p .
(32)
11
dpo
1) C̄ > 0. In this case, we know QC Φ p
is strictly
concave at C̄. Also
dpo
dpo
dpo
dpo
(1)
QC̄ Φ p = Q0 Φ p > Q(1)
Φ p ≥ QC̄
Φp .
0
Here the equality follows from Eq. (32). The first inequaldpo
ity follows because Ψp has a superadditive quantum
(1)
capacity, as both CQ and CQ
remain unchanged after a
unital extension, and QC reduces to the quantum capacity
at C = 0. The second inequality follows as the rate of
quantum communication along the CQ trade-off curve
must not exceed the quantum capacity.
Choose the noise parameter η for the qubit dephasing
dph
channel Ψη appropriately such that
dph
dpo
CQ Ψη = 1 − CQ Ψp .
Here we construct a channel that has an additive CE tradeoff capacity region, but a superadditive quantum capacity,
hence a superadditive quantum dynamic capacity region.
Let Ψ0 be a classical channel and Ψ1 be the depolarizing
dpo
dpo
channel Ψp . p is chosen such that Ψp has a superadditive
quantum capacity. Also, we require
CC Ψ0 > CE Ψ1 .
(33)
Now consider the switch channel N , consisting of N 0 and
which are unital extensions of Ψ0 and Ψ1 . It can be
easily shown that unital extension does not change the classical
capacity with umlimited entanglement assistance of the qubit
depolarizing channel. Thus Eq. (33) implies
(1)
CCE N 0 ⊇ CCE N 1 ⊇ CCE
N1 .
(34)
N 1,
Hence
Define
N =
1
It’s clear that
E. Additive CE, Superadditive Q and CQE
dpo
Φp
⊗
dph
Φη .
dpo
Ψp ⊗
N1
is a unitally extended channel
of
dpo
dph
1
and has CQ N = CQ Ψp ⊗ Ψη = 1. The CQ
trade-off curve is strictly concave and superadditive at C̄.
The corresponding Ψ0 is
dph
Ψη
dpo
Ψ0 = I ⊗ Ψ1 ,
i.e., a noiseless channel tensor a complete qubit depolarizing channel. N 0 is a unital extension of Ψ0 .
2) C̄ = 0. Choose η1 close to 1/2 such that
dpo
dph
dQC Φ p
dQC Φη1
.
>
dph
dC
dC
CC Ψη
0+
1
−
Let
dph
dpo
dph
N 1 = Φη1 ⊗ Φ p ⊗ Φη2 ,
where η2 is chosen such that
dph
dpo
dph
CQ N 1 = CQ Ψη1 ⊗ Ψp ⊗ Ψη2
dph
dpo
dph
=CQ Ψη1 + CQ Ψp + CQ Ψη2 = 1.
dph
dpo
By our choice of η1 , QC Φη1 ⊗ Φ p is strictly concave
dph
in C for 0 < C < 1. QC Φη2 is also strictly concave in
C. Thus QC N 1 is strictly concave in C, for 0 < C < 1.
In this case, the corresponding Ψ0 is
dpo
Ψ0 = I ⊗ Ψ1
⊗2
,
i.e., a noiseless channel tensor two copies of the complete
qubit depolarizing channel. N 0 is a unital extension of
Ψ0 .
CCE (N ) = Conv CCE
= Conv CCE
N 0 , CCE
(1)
N 0 , CCE
N1
= CCE N 0
N1
(1)
= CCE
(N ) ,
i.e., its CE trade-off
capacity region is additive.
Since CQ N 0 = 0, it is clear that the quantum capacity of
N is the same as that of N 1 , which is superadditive.
Note that N is a unitally extended channel. This fact will
be implicitly used in Section V-F.
F. Additive CE and Q, Superadditive CQE
Previously in Section V-D, we give an example of a channel
with an additive classical and quantum capacity, but whose CQ
trade-off curve is superadditive. It is unclear if the channel has
an additive CE trade-off capacity region, because the CE tradeoff capacity region of the depolarizing channel has not been
shown to be additive. This is itself an interesting question but
we’ll not explore it here.
dpo
We replace Ψp in the original argument of Section V-D by
the channel constructed in Section V-E. It’s clear that the rest
of the argument is not changed and N still has a superadditive
CQ trade-off capacity region.
Now both N 0 and N 1 have an additive CE trade-off capacity
region. It’s clear that
CCE (N ) = Conv CCE N 0 , CCE N 1
(1)
(1)
(1)
= Conv CCE
N 0 , CCE
N 1 = CCE
(N ) ,
i.e., the CE trade-off capacity region of N is additive.
G. Additive CQ, Superadditive CQE
Our construction in Section V-A has a superadditive CE
trade-off capacity region. But most likely its CQ trade-off
capacity region is also superadditive. This is because in
Section V-A, N 0 is the unital extension of a classical channel,
and its CQ trade-off capacity region is trivial. Hence the CQ
trade-off capacity region of N is given by that of N 1 , which
is most likely superadditive as well.
12
To achieve an additive CQ trade-off capacity region, we
have to substitute N 0 with a channel that has a non-trivial CQ
trade-off capacity region.
Recall that our construction in Section V-A requires N 0
and N 1 to have properties (A1)-(A3). These three properties
ensure that N will have a superadditive CE trade-off capacity
region, while its classical capacity is still additive.
In extending to a channel with an additive CQ trade-off
capacity region, the additional properties we need are
(G1) CCQ N 0 ⊇ CCQ N 1 .
Property (G1) and (U) ensure the CQ trade-off capacity
region of N is additive, as
CCQ (N ) = Conv CCQ N 0 , CCQ N 1 = CCQ N 0
(1)
(1)
(1)
= CCQ
N 0 = Conv CCQ
N 0 , CCQ
N1
(1)
= CCQ
(N ) .
Unfortunately, we cannot find quantum channels N 0 and
that satisfy all the properties. Hence we do not have an
explicit construction in this case. This is because there are
very few channels that we understand their dynamic capacity
regions. This leaves us with a limited choice of candidates
for N 0 . However, in principle there is no obstacle and the
construction will be readily available once we have a better
understanding of quantum channels.
N1
VI. C ONCLUSION
Unlike previous studies on additivity of single resource
channel capacity, our work aimed to understand how additivity of single or double resource capacity regions will
effect additivity of a general resource trade-off capacity. In
contrast to the two known results in the literature; namely,
(i) additivity of the quantum capacity implies additivity of
the entanglement-assisted quantum capacity region and (ii)
additivity of classical-quantum and classical-entanglement capacity regions implies additivity of the three resource capacity
region, the additivity of all the remaining situations does not
hold. In this work, we identified all possible occurrences where
superadditivity could occur in the trade-off quantum dynamic
capacity. Furthermore, we provided an explicit construction of
quantum channels for most instances. Our main technical tool
combines properties of switch channels and unital extension
of known quantum channels.
An obvious open question is an explicit construction of
a quantum channel whose classical-quantum capacity region
is additive, but its triple trade-off capacity is superadditive.
Moreover, there are other triple resource trade-off capacity
regions [4], [31]. Could similar statements made in this work
hold in these scenarios as well?
ACKNOWLEDGMENT
EYZ and PWS are supported by the National Science
Foundation under grant Contract Number CCF-1525130. QZ
is supported by the Claude E. Shannon Research Assistantship.
MH is supported by an ARC Future Fellowship under Grant
FT140100574. PWS is supported by the NSF through the STC
for Science of Information under grant number CCF0-939370.
A PPENDIX A
P ROOF OF L EMMA 12
Proof. Consider Φ0RC→B0 and Ψ1A1 →B1 , where Φ0 is a unital
0
1 is an
extension of a Hadamard channel ΨC→B
0 , and Ψ
arbitrary channel.
The result follows if both the CQ and CE trade-off capacity
regions of Φ0 are additive [4]. To show that the CQ trade-off
capacity region is additive for Φ0 , it was shown in Ref. [23]
it suffices to prove that
fλ Φ0 ⊗ Ψ1 = fλ Φ0 + fλ Ψ1
(35)
for any channel Ψ1 , where
fλ (N ) = max I(X; B)σ + λI(AiBX)σ .
(36)
ρ
The state σ is the channel output state with ρ being the input
state (see, e.g., Theorem
we will only
1). In the following,
show that fλ Φ0 ⊗ Ψ1 ≤ fλ Φ0 + fλ Ψ1 because the other
direction is trivial from its definition.
Since Φ0 ⊗ Ψ1 : CRA1 → B0 B1 is a partial cq channel, then
by the same argument as that in Lemma 10, fλ Φ0 ⊗ Ψ1 can
be achieved with input states of the following form
Õ p(x)
|x, ji hx, j | X ⊗ | ji h j | R ⊗ φ xAC A1 ,
ρX R AC A1 =
|R|
x, j
with output states
σX AB0 B1 =
where
Õ p(x)
xj
|x, ji hx, j | X ⊗ σAB0 B1 ,
|R|
x, j
(37)
xj
σAB0 B1 = Φ0 ⊗ Ψ1 | ji h j | R ⊗ φ xAC A1 .
0
1
Let UC→B
0 E 0 and U A1 →B 1 E 1 be the isometric extensions of
1
and Ψ , and let
Õ
%X AC A1 =
p(x) |xi hx| X ⊗ φ xAC A1
Ψ0
x
ωX AA1 B0 E 0
ςX AB0 B1 E 0 E 1
†
= U 0 ⊗ I %X AC A1 U 0 ⊗ I
†
= U 0 ⊗ U 1 %X AC A1 U 0 ⊗ U 1 .
Moreover, let
θ XY AB1 E 0 E 1 = DB1 0 →Y (ςX AB0 B1 E 0 E 1 ) ,
2
1
where DY→E
0 ◦ D B 0 →Y = D B 0 →E 0 is a degrading map for the
Hadamard channel Ψ0 .
For any state σX AB0 B1 in Eq. (37), we have
fλ Φ0 ⊗ Ψ1
=I X; B0 B1 + λI AiB0 B1 X
σ
σh
i
=S B0 B1 + (λ − 1)S B0 B1 |X − λS AB0 B1 |X
σ
σ
σ
=S B0 B1 + (λ − 1)S B0 B1 |X − λS AB0 B1 |X
,
ς
ς
ς
13
where the last equality follows from the same argument used
in Eqs. (18) and (20). Then subadditivity of the von Neumann
entropy and chain rule yield
≤S B0 + (λ − 1)S B0 |X − λS E 0 |X
ς
ς
ς
1
1 0
1 0
+S B
+ (λ − 1)S B |B X − λS E |E X
ς
ς
ς
0
0
0
≤S B
+ (λ − 1)S B |X − λS E |X
ς
ς
ς
1
1
1
+S B
+ (λ − 1)S B |XY − λS E |XY
θ
θ
θ
1
where the last inequality uses the fact that S B |B0 X ς ≤
S B1 |Y X θ due to the existence of D 1 and S E 1 |E 0 X ς ≥
S E 1 |Y X θ due to the existence of D 2 . Finally,
= I X; B0 + λI AA1 iB0 X
ω
ω
1
0
1
+ I XY ; B
+ λI AE iB XY
θ
θ
0
1
≤ fλ Φ + fλ Ψ
because S(E 0|X)ς = S(AA1 B0 |X)ω and S E 1 |XY θ =
S AB1 E 0 |XY θ .
To prove that the CE trade-off capacity region of the channel
Φ0 is additive is equivalent to showing that [23]:
gλ Φ0 ⊗ Ψ1 = gλ Φ0 + gλ Ψ1 ,
(38)
where 0 ≤ λ < 1,
gλ (N ) = max I(AX; B)σ − λS(A|X)σ
σ
(39)
and σ is of the form given in Eq. (7).
However this proof proceeds similarly; hence, we will omit
it.
where
is now a product state with respect to A0 and B0 .
The three entropic quantities of interest can be simplied
when evaluated with respect to σX A0 B0 , as
Õ
I A0 X; B0 = S B0 −
p(x, j)S B0 x j ≤ CC Ψ0 ,
σ
I A0 iB0 X
C + Q + E ≤ CC Ψ0 ,
where CC Ψ0 is the classical capacity of Ψ0 .
The same holds for a unital extension of a classical channel.
Proof. Consider the 1-shot dynamic
capacity region of
(1)
Ψ0A00 →B0 . By Lemma 10, CCQE
Ψ0 can be achieved with
respect to cq states σX A0 B0 = Ψ0 ρX A0 A00 , where ρX A0 A00 is
of the form
Õ
xj
ρX A0 A00 =
p(x, j) |x, ji hx, j | X ⊗ | ji h j | A00 ⊗ φ A0 .
x, j
Thus
σX A0 B0 =
Õ
x, j
xj
p(x, j) |x, ji hx, j | X ⊗ σA0 B0 ,
σ
σ
σ
=−
Õ
σ
x, j
p(x, j)S A0
x, j
σx j
≤ 0,
≤ CC Ψ0 .
It’s also clear
that those inequalities can be achieved. Thus
(1)
CCQE
Ψ0 is described by
C + 2Q ≤ CC Ψ0 ,
Q + E ≤ 0,
C + Q + E ≤ CC Ψ0 .
Since the classical capacity of a classical channel is additive,
the dynamic capacity region of Ψ0 is additive and is described
by the same set of inequalities.
Next we show that the dynamic capacity region is additive
for Ψ0 and Ψ1 , with Ψ1 arbitrary.
Since Ψ0A00 →B0 ⊗ Ψ1A10 →B1 is a partial cq channel, its
(1)
Ψ0 ⊗ Ψ1 can be
1-shot dynamic capacity region CCQE
achieved with respect to cq states σX AB0 B1 = Ψ0A00 →B0 ⊗
Ψ1A10 →B1 ρX AA00 A10 , where ρX AA00 A10 is of the form
Õ
xj
ρX AA00 A10 =
p(x, j) |x, ji hx, j | X ⊗ | ji h j | A00 ⊗ φ AA10 .
x, j
For ρ
Q + E ≤ 0,
I X; B0
A PPENDIX B
P ROOF OF L EMMA 13
Lemma 6: If Ψ0 is a classical channel, then the dynamic
capacity region is additive for Ψ0 ⊗ Ψ1 , for arbitrary Ψ1 .
Moreover, the dynamic capacity region of Ψ0 is described by
the following relation
C + 2Q ≤ CC Ψ0 ,
xj
xj
σA0 B0 = Ψ0A00 →B0 | ji h j | A00 ⊗ φ A0
of this form, σX AB0 B1 is of the form
Õ
xj
σX AB0 B1 =
p(x, j) |x, ji hx, j | X ⊗ σAB0 B1 ,
0
0
X AA0 A1
x, j
with
xj
xj
σAB0 B1 = Ψ0A00 →B0 ⊗ Ψ1A10 →B1 | ji h j | A00 ⊗ φ AA10 .
For such σX AB0 B1 , each of the three entropic quantities have
simple upper bounds,
I AX; B0 B1 ≤ I X; B0 + I AX; B1 ,
σ
σ
σ
1
0 1
I AiB B X = I AiB X ,
σ
σ
I X; B0 B1 ≤ I X; B0 + I X; B1 ,
σ
σ
σ
where we’ve used subadditivity of the von Neumann entropy.
Thus the 1-shot dynamic capacity region of Ψ0 ⊗ Ψ1 has a
simple upper bound
(1)
(1)
(1)
CCQE
Ψ0 ⊗ Ψ1 ⊆ CCQE
Ψ0 + CCQE
Ψ1 .
It’s trivial to extend it to the dynamic capacity region of Ψ0 ⊗
Ψ1
CCQE Ψ0 ⊗ Ψ1 ⊆ CCQE Ψ0 + CCQE Ψ1 .
14
Since the other direction of inclusion is obvious, we have
CCQE Ψ0 ⊗ Ψ1 = CCQE Ψ0 + CCQE Ψ1 .
A PPENDIX D
C ONVEX HULL
Here we show that4
For unital extensions of a classical channel, we observe that,
if the Heisenberg-Weyl operators are defined on the standard
basis for the output of the channel, then the resulting channel
is also a classical channel. Hence the above result applies.
A PPENDIX C
UNITAL EXTENSION OF THE QUBIT DEPHASING CHANNEL
AND 1 → N CLONING CHANNEL
Lemma 14: The CE and CQ trade-off curve of the qubit
dephasing channel and 1 → N cloning channels are unchanged
after a unital extension.
dph
Proof. Consider the qubit dephasing channel Ψη and a 1 →
dph
1→N
N cloning channel Ψ
, and their unital extensions Φη and
1→N
Φ
. The statement of this lemma is equivalent to showing
that
for
fλ (Ψ) = fλ (Φ)
∀λ ≥ 1,
gλ (Ψ) = gλ (Φ)
∀0 ≤ λ < 1.
dph
dph
(Ψ, Φ) = Ψη , Φη , Ψ1→N , Φ1→N .
In Lemma 11, we have argued that the 1-shot dynamic
capacity region of a unitally extended channel can be achieved
with input of the form in Eq. (16). Evaluating fλ (Φ) on such
states, one obtains
Conv CCQE N 0 , CCQE N 1
=Conv CCQE N 0 , CCQE N 1 .
We quote a few properties about convex hull and Minkowski
addition that we will use [29]: (i) For two closed sets A and
B in Rk , if A is bounded, then A + B is closed. (ii) For two
sets A and B in Rk , Conv(A + B) = Conv(A) + Conv(B) . (iii)
The convex hull of a bounded set in Rk is also bounded.
First, we note that, by Ref. [4], all points in the 1-shot
dynamic capacity region can be achieved by the classically
enhanced father protocol, combined with unit protocols, i.e.,
Ø
(1)
(1)
CCQE
CCQE,CEF
(N ) =
(N )σ + CCQE,unit,
σ
where
1
1
(1)
CCQE,CEF
(N )σ = {I(X; B)σ, I(A; B|X)σ, − I(A; E |X)σ }
2
2
is the rate achieved using the classically enhanced father
protocol, and σ is of the form in Eq. (7). CCQE,unit are all
the rates achieved by the unit protocols. Clearly CCQE,unit is
convex and closed. Define
Ø
(1)
(1)
CCQE,CEF
CCQE,CEF
(N )σ .
(N ) =
σ
(1)
CCQE,CEF
(N )
Clearly
fλ (Φ) = log(|B|) + (λ − 1)S (B|X 0)σ − λS (AB|X 0)σ
dimensions
of N .
Õ
= log(|B|) +
p(x, j, k) [(λ − 1)S(B)σ x j k − λS(AB)σ x j k ] Then
x, j,k
= log(|B|) +
Õ
CCQE (N ) =
p(x, j) [(λ − 1)S(B)σ x j − λS(AB)σ x j ]
x, j
≤ log(|B|) + max [(λ − 1)S (B)σ − λS (AB)σ ] ,
σ
where
σAB = ΨC→B (φ AC ) .
(41)
For such a σAB = Ψ (φ AC ) that achieves Eq. (40), one can
construct
1
|ki hk | X ⊗ |ki hk | R ⊗ φ AC .
(42)
ρX AA0 =
|R|
This state will saturate the above inequality.
dph
For Ψη and Ψ1→N , it can be verified [23] that their fλ
have the same form, i.e.,
fλ (Ψ) = log(|B|) + max [(λ − 1)S (B)σ − λS (AB)σ ] ,
σ
=
(40)
(43)
with σ of the form given in Eq. (41).
The same argument also applies to gλ .
The CQ trade-off curve of the qubit dephasing channel
was computed in Ref. [22], and the CE trade-off curve was
computed in Ref. [8]. The CE and CQ trade-off curves of the
1 → N cloning channel were given in Ref. [23]. Other than
the special cases (η = 0, 1/2 for the dephasing channel, N = 1
for the 1 → N cloning channel), it can be verified that their
CE and CQ trade-off curves are strictly concave at every point.
By Lemma 14, this property is true for their unital extensions.
is bounded by the input and output
Ø1
(1)
CCQE,CEF
N ⊗k + CCQE,unit
k
k=1
Ø1
(1)
N ⊗k + CCQE,unit .
CCQE,CEF
k
k=1
Denote
A=
Ø1
(1)
CCQE,CEF
N ⊗k ,
k
k=1
B = CCQE,unit .
Since A is bounded, B is closed, by (i) and (iii), A + B is also
closed.
Since A + B ⊆ A + B, and A + B is closed, we have
A + B ⊆ A + B.
It is also obvious that
A + B ⊇ A + B,
hence
A + B = A + B.
Denote
CCQE,CEF (N ) =
4 N0
Ø1
(1)
CCQE,CEF
N ⊗k .
k
k=1
and N 1 are assumed to be finite-dimensional.
15
Then by the above arguments,
CCQE (N ) = CCQE,CEF (N ) + CCQE,unit .
Now we apply the above result to N 0 and N 1 .
Conv CCQE N 0 , CCQE N 1
=Conv CCQE N 0 ∪ CCQE N 1
=Conv CCQE,CEF N 0 ∪ CCQE,CEF N 1 + CCQE,unit
=Conv CCQE,CEF N 0 ∪ CCQE,CEF N 1 + CCQE,unit .
In the last line, we used (ii).
Since CCQE,CEF (N ) is closed and bounded for any
finite dimensional quantum channel N , the same
must
0
1 . Hence
be true for CCQE,CEF
N
∪
C
N
CQE,CEF
Conv CCQE,CEF N 0 ∪ CCQE,CEF N 1
is closed and
bounded. Thus Conv CCQE,CEF N 0 ∪ CCQE,CEF N 1 +
CCQE,unit is also closed.
R EFERENCES
[1] C. E. Shannon, “A mathematical theory of communication,” The Bell
System Technical Journal, vol. 27, no. 3, pp. 379–423, July 1948.
[2] I. Devetak, A. W. Harrow, and A. Winter, “A family of quantum
protocols,” Phys. Rev. Lett., vol. 93, p. 230504, Dec 2004.
[3] I. Devetak, A. W. Harrow, and A. J. Winter, “A resource framework for
quantum shannon theory,” IEEE Transactions on Information Theory,
vol. 54, no. 10, pp. 4587–4618, Oct 2008.
[4] M.-H. Hsieh and M. M. Wilde, “Trading classical communication,
quantum communication, and entanglement in quantum shannon theory,”
IEEE Transactions on Information Theory, vol. 56, no. 9, pp. 4705–
4730, Sept 2010.
[5] C. H. Bennett and S. J. Wiesner, “Communication via one- and
two-particle operators on einstein-podolsky-rosen states,” Phys. Rev.
Lett., vol. 69, pp. 2881–2884, Nov 1992.
[6] C. H. Bennett, P. W. Shor, J. A. Smolin, and A. V. Thapliyal,
“Entanglement-assisted classical capacity of noisy quantum channels,”
Phys. Rev. Lett., vol. 83, pp. 3081–3084, Oct 1999.
[7] C. H. Bennett, P. W. Shor, J. A. Smolin, and A. V. Thapliyal,
“Entanglement-assisted capacity of a quantum channel and the reverse
shannon theorem,” IEEE Transactions on Information Theory, vol. 48,
no. 10, pp. 2637–2655, Oct 2002.
[8] M.-H. Hsieh and M. M. Wilde, “Entanglement-assisted communication
of classical and quantum information,” IEEE Transactions on Information Theory, vol. 56, no. 9, pp. 4682–4704, Sept 2010.
[9] N. Datta and M. H. Hsieh, “One-shot entanglement-assisted quantum
and classical communication,” IEEE Transactions on Information Theory, vol. 59, no. 3, pp. 1929–1939, March 2013.
[10] A. S. Holevo, “The capacity of the quantum channel with general signal
states,” IEEE Transactions on Information Theory, vol. 44, no. 1, pp.
269–273, Jan 1998.
[11] B. Schumacher and M. D. Westmoreland, “Sending classical information
via noisy quantum channels,” Phys. Rev. A, vol. 56, pp. 131–138, Jul
1997.
[12] S. Lloyd, “Capacity of the noisy quantum channel,” Phys. Rev. A,
vol. 55, pp. 1613–1622, Mar 1997.
[13] P. W. Shor, “The quantum channel capacity and coherent information,”
MSRI Workshop on Quantum Computation, 2002.
[14] I. Devetak, “The private classical capacity and quantum capacity of a
quantum channel,” IEEE Transactions on Information Theory, vol. 51,
no. 1, pp. 44–55, Jan 2005.
[15] G. Smith and J. Yard, “Quantum communication with zero-capacity
channels,” Science, vol. 321, no. 5897, pp. 1812–1815, 2008.
[16] P. W. Shor, “The classical capacity achievable by a quantum
channel assisted by a limited entanglement.” Quantum Information and
Computation, vol. 4, no. 6, pp. 537 – 545, 2004.
[17] G. Smith, J. A. Smolin, and A. Winter, “The quantum capacity with
symmetric side channels,” IEEE Transactions on Information Theory,
vol. 54, no. 9, pp. 4208–4217, Sept 2008.
[18] M. B. Hastings, “Superadditivity of communication capacity using
entangled inputs,” Nat Phys, vol. 5, no. 4, pp. 255–257, 04 2009.
[19] P. Hayden and A. Winter, “Counterexamples to the maximal pnorm multiplicativity conjecture for all p > 1,” Communications in
Mathematical Physics, vol. 284, no. 1, pp. 263–280, 2008.
[20] P. W. Shor, “Equivalence of additivity questions in quantum information
theory,” Communications in Mathematical Physics, vol. 246, no. 3, pp.
453–472, 2004.
[21] E. Y. Zhu, Q. Zhuang, and P. W. Shor, “Superadditivity of the classical
capacity with limited entanglement assistance,” Phys. Rev. Lett., vol.
119, p. 040503, Jul 2017.
[22] I. Devetak and P. W. Shor, “The capacity of a quantum channel
for simultaneous transmission of classical and quantum information,”
Communications in Mathematical Physics, vol. 256, no. 2, pp. 287–303,
2005.
[23] K. Brádler, P. Hayden, D. Touchette, and M. M. Wilde, “Trade-off
capacities of the quantum hadamard channels,” Phys. Rev. A, vol. 81,
p. 062312, Jun 2010.
[24] C. King, “Additivity for unital qubit channels,” Journal of Mathematical
Physics, vol. 43, no. 10, pp. 4641–4653, 2002.
[25] D. P. DiVincenzo, P. W. Shor, and J. A. Smolin, “Quantum-channel
capacity of very noisy channels,” Phys. Rev. A, vol. 57, pp. 830–839,
Feb 1998.
[26] M. Fukuda, “Simplification of additivity conjecture in quantum
information theory,” Quantum Information Processing, vol. 6, no. 3,
pp. 179–186, 2007.
[27] D. Elkouss and S. Strelchuk, “Superadditivity of private information
for any number of uses of the channel,” Phys. Rev. Lett., vol. 115, p.
040501, Jul 2015.
[28] M. M. Wilde, Quantum Information Theory. Cambridge University
Press, 2013.
[29] R. Schneider, Convex Bodies: The Brunn-Minkowski Theory (Encyclopedia of Mathematics and its Applications). Cambridge University Press,
1993.
[30] W. Rudin, Principles of Mathematical Analysis. McGraw-Hill
Education, 1976.
[31] M. M. Wilde and M.-H. Hsieh, “The quantum dynamic capacity
formula of a quantum channel,” Quantum Information Processing,
vol. 11, no. 6, pp. 1431–1463, 2012.
| 7cs.IT
|
Fast Algorithms for Convolutional Neural Networks
Andrew Lavin
Scott Gray
Nervana Systems
[email protected]
arXiv:1509.09308v2 [cs.NE] 10 Nov 2015
[email protected]
Abstract
cause they achieve better accuracy with fewer weights than
shallow networks with larger filters [11, 7].
Deep convolutional neural networks take GPU days of
compute time to train on large data sets. Pedestrian detection for self driving cars requires very low latency. Image recognition for mobile phones is constrained by limited
processing resources. The success of convolutional neural
networks in these situations is limited by how fast we can
compute them. Conventional FFT based convolution is fast
for large filters, but state of the art convolutional neural
networks use small, 3 × 3 filters. We introduce a new class
of fast algorithms for convolutional neural networks using
Winograd’s minimal filtering algorithms. The algorithms
compute minimal complexity convolution over small tiles,
which makes them fast with small filters and small batch
sizes. We benchmark a GPU implementation of our algorithm with the VGG network and show state of the art
throughput at batch sizes from 1 to 64.
Therefore there is a strong need for fast convnet algorithms for small batch sizes and small filters. However
conventional convnet libraries require large batch sizes and
large filters for fast operation.
This paper introduces a new class of fast algorithms for
convolutional neural networks based on the minimal filtering algorithms pioneered by Winograd [13]. The algorithms
can reduce the arithmetic complexity of a convnet layer by
up to a factor of 4 compared to direct convolution. Almost
all of the arithmetic is performed by dense matrix multiplies of sufficient dimensions to be computed efficiently,
even when the batch size is very small. The memory requirements are also light compared to the conventional FFT
convolution algorithm. These factors make practical implementations possible. Our implementation for NVIDIA
Maxwell GPUs achieves state of the art throughput for all
batch sizes measured, from 1 to 64, while using at most
16MB of workspace memory.
1. Introduction
Deep convolutional neural networks (convnets) achieve
state of the art results on image recognition problems [11][7]. The networks take several days of GPU time
to train and require significant compute resources during
classification as well. Larger data sets and models lead to
better accuracy but also increase computation time. Therefore progress in deep neural networks is limited by how fast
the networks can be computed.
Likewise the application of convnets to low latency inference problems, such as pedestrian detection in self driving car video imagery, is limited by how fast a small set of
images, possibly a single image, can be classified.
Distributed training of convnets can be achieved by partitioning each batch of examples across the nodes of a cluster
and accumulating weight updates across the nodes. Large
batch sizes adversely affect convergence of the network, so
the minimum batch size that can be computed efficiently
places an upper limit on cluster size [8, 6].
State of the art convnet architectures for image recognition use deep networks of 3 × 3 convolutional layers, be-
2. Related Work
The FFT and convolution theorem have been used to reduce the arithmetic complexity of convnet layers, first by
Mathieu et al. [10], then refined by Visalache et al. [12],
and then implemented in the NVIDIA cuDNN library [1].
The Strassen algorithm for fast matrix multiplication was
used by Cong and Xiao [3] to reduce the number of convolutions in a convnet layer, thereby reducing its total arithmetic complexity. The authors also suggested that more
techniques from arithmetic complexity theory might be applicable to convnets.
Various approaches have been attempted to reduce the
complexity of convnets by quantizing or otherwise approximating the convolutional layer. We consider these approaches as orthogonal and complementary to those that exploit algebraic structure, and therefore declare them outside
the scope of this paper.
1
3. Convolutional Neural Networks
where
A convnet layer correlates a bank of K filters with C
channels and size R × S against a minibatch of N images
with C channels and size H × W . We denote filter elements
as Gk,c,u,v and image elements as Di,c,x,y .
The computation of a single covnnet layer output Yi,k,x,y
is given by the formula:
Yi,k,x,y =
C X
R X
S
X
Di,c,x+u,y+v Gk,c,u,v
(1)
c=1 v=1 u=1
and we can write the output of an entire image/filter pair as
Yi,k =
C
X
Di,c ∗ Gk,c
(2)
c=1
where ∗ denotes 2D correlation.
4. Fast Algorithms
It has been known since at least 1980 that the minimal
filtering algorithm for computing m outputs with an r-tap
FIR filter, which we call F (m, r), requires
µ(F (m, r)) = m + r − 1
(3)
multiplications [13, p. 39]. Also, we can nest minimal 1D
algorithms F (m, r) and F (n, s) to form minimal 2D algorithms for computing m × n outputs with an r × s filter,
which we call F (m × n, r × s). These require
µ(F (m × n, r × s)) = µ(F (m, r))µ(F (n, s))
= (m + r − 1)(n + s − 1)
(4)
multiplications [14]. We can continue to nest 1D algorithms
to form algorithms for multi-dimensional FIR filters.
It is interesting to note that in 1D, 2D, and multidimensions, the minimal algorithm requires a number of
multiplications equal to the number of inputs. In other
words, to compute F (m, r) we must access an interval of
m + r − 1 data values, and to compute F (m × n, r × s)
we must access a tile of (m + r − 1) × (n + s − 1) data
values. Therefore the minimal filtering algorithm requires
one multiplication per input.
4.1. F(2x2,3x3)
The standard algorithm for F (2, 3) uses 2 × 3 = 6 multiplications. Winograd [13, p. 43] documented the following
minimal algorithm:
g0
d
d1
d2
m1 + m2 + m3
F (2, 3) = 0
g1 =
d1
d2
d3
m2 − m3 − m4
g2
(5)
m1 = (d0 − d2 )g0
m4 = (d1 − d3 )g2
g0 + g1 + g2
2
g0 − g1 + g2
m3 = (d2 − d1 )
2
m2 = (d1 + d2 )
This algorithm uses just 4 multiplications and is therefore minimal by the formula µ(F (2, 3)) = 2 + 3 − 1 = 4.
It also uses 4 additions involving the data, 3 additions and
2 multiplications by a constant involving the filter (the sum
g0 + g2 can be computed just once), and 4 additions to reduce the products to the final result.
Fast filtering algorithms can be written in matrix form
as:
Y = AT (Gg) ⊙ (B T d)
(6)
where ⊙ indicates element-wise multiplication.
F (2, 3), the matrices are:
1
0
−1
0
0
1
1
0
BT =
0
−1
1
0
0
1
0
−1
1
0
0
1
1
1
2
2
2
G=
1
1
− 21
2
2
0
0
1
1
1
1
0
AT =
0
1
−1
−1
T
g = g0 g1 g2
T
d = d0 d1 d2 d3
For
(7)
A minimal 1D algorithm F (m, r) is nested with itself to
obtain a minimal 2D algorithm, F (m × m, r × r) like so:
(8)
Y = AT [GgGT ] ⊙ [B T dB] A
where now g is an r × r filter and d is an (m + r − 1) ×
(m + r − 1) image tile. The nesting technique can be generalized for non-square filters and outputs, F (m × n, r × s),
by nesting an algorithm for F (m, r) with an algorithm for
F (n, s).
F (2 × 2, 3 × 3) uses 4 × 4 = 16 multiplications, whereas
the standard algorithm uses 2 × 2 × 3 × 3 = 36. This
is an arithmetic complexity reduction of 36
16 = 2.25. The
data transform uses 32 additions, the filter transform uses
28 floating point instructions, and the inverse transform uses
24 additions.
Algorithms for F (m × m, r × r) can be used to compute
convnet layers with r × r kernels. Each image channel is
divided into tiles of size (m + r − 1) × (m + r − 1), with r −
1 elements of overlap between neighboring tiles, yielding
P = ⌈H/m⌉⌈W/m⌉ tiles per channel, C. F (m × m, r × r)
is then computed for each tile and filter combination in each
channel, and the results are summed over all channels.
Substituting U = GgGT and V = B T dB, we have:
(9)
Y = AT U ⊙ V A
Labeling tile coordinates as (e
x, ye), we rewrite the convnet layer formula (2) for a single image i, filter k, and tile
coordinate (e
x, ye) as:
Yi,k,ex,ey =
C
X
Di,c,ex,ey ∗ Gk,c
c=1
=
Uk,c ⊙ Vc,i,ex,ey A
X
C
Uk,c ⊙ Vc,i,ex,ey A
T
A
c=1
= AT
C
X
c=1
(10)
Thus we can reduce over C channels in transform space,
and only then apply the inverse transform A to the sum.
This amortizes the cost of the inverse transform over the
number of channels.
We examine the sum
Mk,i,ex,ey =
C
X
Uk,c ⊙ Vc,i,ex,ey
(11)
Algorithm 1 Compute Convnet Layer with Winograd Minimal Filtering Algorithm F (m × m, r × r)
P = N ⌈H/m⌉⌈W/m⌉ is the number of image tiles.
α = m + r − 1 is the input tile size.
Neighboring tiles overlap by r − 1.
dc,b ∈ Rα×α is input tile b in channel c.
gk,c ∈ Rr×r is filter k in channel c.
G, B T , and AT are filter, data, and inverse transforms.
Yk,b ∈ Rm×m is output tile b in filter k.
for k = 0 to K do
for c = 0 to C do
u = Ggk,c GT ∈ Rα×α
(ξ,ν)
Scatter u to matrices U: Uk,c = uξ,ν
for b = 0 to P do
for c = 0 to C do
v = B T dc,b B ∈ Rα×α
(ξ,ν)
Scatter v to matrices V: Vc,b = vξ,ν
for ξ = 0 to α do
for ν = 0 to α do
M (ξ,ν) = U (ξ,ν) V (ξ,ν)
for k = 0 to K do
for b = 0 to P do
(ξ,ν)
Gather m from matrices M: mξ,ν = Mk,b
Yk,b = AT mA
c=1
and simplify the notation by collapsing the image/tile coordinates (i, x
e, ye) down to a single dimension, b. We also label each component of the element-wise multiplication separately, as (ξ, ν), yielding:
(ξ,ν)
Mk,b
=
C
X
(ξ,ν)
(ξ,ν)
Uk,c Vc,b
4.2. F(3x3,2x2)
(12)
c=1
This equation is just a matrix multiplication, so we can
write:
M (ξ,ν) = U (ξ,ν) V (ξ,ν)
(13)
Matrix multiply has efficient implementations on CPU,
GPU, and FPGA platforms, owing to its high computational
intensity. Thus we have arrived at the practical implementation for the fast algorithm listed in Algorithm 1.
Winograd documented a technique for generating the
minimal filtering algorithm F (m, r) for any choice of m
and r. The construction uses the Chinese remainder theorem to produce a minimal algorithm for linear convolution, which is equivalent to polynomial multiplication, then
transposes the linear convolution algorithm to yield a minimal filtering algorithm. The reader is referred to Winograd’s seminal book [13], or Blahut’s book [2] for a modern treatment of the subject. We provide derivations of the
specific algorithms used in this paper in the supplementary
material.
Training a network using stochastic gradient descent requires computation of the gradients with respect to the inputs and weights. For a convnet layer, the gradient with respect to the inputs is a convolution of the next layer’s backpropagated error, of dimension N × K × H × W , with a
flipped version of the layer’s R × S filters. Therefore it
can be computed using the same algorithm that is used for
forward propagation.
The gradient with respect to the weights is a convolution
of the layer inputs with the backpropagated errors, producing R × S outputs per filter and channel. Therefore we need
to compute the convolution F (R×S, H ×W ), which is impractical because H ×W is much too large for our fast algorithms. Instead we decompose this convolution into a direct
sum of smaller convolutions, for example F (3 × 3, 2 × 2).
Here the algorithm’s 4 × 4 tiles are overlapped by 2 pixels
in each dimension, and the 3 × 3 outputs are summed over
all tiles to form F (3 × 3, H × W ).
The transforms for F (3 × 3, 2 × 2) are given by:
1
0 −1 0
0
1
1 0
BT =
1
0
0 −1
1 0
1
1
2
0 −1
0 1 , G = 21
− 21
2
1 1
1 0
0
1
AT = 0 1 −1 0
0 1
1 1
(14)
4.4. Fast Fourier Transform
With (3 + 2 − 1)2 = 16 multiplies versus direct convolution’s 3 × 3 × 2 × 2 = 36 multiplies, it achieves the same
36/16 = 2.25 arithmetic complexity reduction as the corresponding forward propagation algorithm.
4.3. F(4x4,3x3)
A minimal algorithm for F (4, 3) has the form:
4
0 −5
0 1 0
0 −4 −4
1 1 0
0
4
−4
−1
1 0
T
B =
2 1 0
0 −2 −1
0
2 −1 −2 1 0
0
4
0 −5 0 1
1
0
0
4
−1
− 61 − 16
61
1
−
− 16
6
6
G= 1
1
1
12
6
24
1
1 −1
24
12
6
0
0
1
1 1
1 1
1 0
0 1 −1 2 −2 0
AT =
0 1
1 4
4 0
0 1 −1 8 −8 1
tiles, the transforms cannot be computed accurately [13,
p. 28].
Convnets require surprisingly little numeric precision [4,
5]. This implies that we can sacrifice some numeric accuracy in the filtering computation without affecting the
accuracy of the convnet. We examine the possibility of
F (6 × 6, 3 × 3) in the supplementary material.
(15)
The data transform uses 13 floating point instructions,
the filter transform uses 8, and the inverse transform uses
10.
Applying the nesting formula yields a minimal algorithm
for F (4 × 4, 3 × 3) that uses 6 × 6 = 36 multiplies, while
the standard algorithm uses 4 × 4 × 3 × 3 = 144. This is an
arithmetic complexity reduction of 4.
The 2D data transform uses 13(6 + 6) = 156 floating
point instructions, the filter transform uses 8(3 + 6) = 72,
and the inverse transform uses 10(6 + 4) = 100.
The number of additions and constant multiplications
required by the minimal Winograd transforms increases
quadratically with the tile size [9, p. 211]. Thus for large
tiles, the complexity of the transforms will overwhelm any
savings in the number of multiplications.
The magnitude of the transform matrix elements also increases with increasing tile size. This effectively reduces
the numeric accuracy of the computation, so that for large
The Fast Fourier Transform (FFT) can be used to produce a tiled convolution algorithm that has the same form
as Algorithm 1. The main difference is that the transform matrices are replaced with FFT and inverse FFT,
and point-wise multiplication of complex FFT components
yields cyclic convolution. Only m × n components of the
(m + r − 1) × (n + s − 1) cyclic convolution are valid, the
rest must be discarded, and the tiles must be overlapped by
r − 1 and s − 1 in order to recompute the discarded outputs.
This technique is referred to as overlap and save [2, p. 195].
The similarity of overlap and save to our approach makes
for an easy comparison. With FFT based convolution, the
multiply stage still uses 1 multiply per input, but now the
operands are complex numbers. Direct multiplication of
complex numbers requires 4 real multiplications. Thankfully, a couple of tricks reduce the complexity further.
The Fourier transform of a real signal has Hermitian
symmetry, which reduces the number of unique products in
each U ⊙V by almost half. FFT based convnet implementations have exploited this property [10, 12]. Specifically, the
discrete Fourier transform of a α × α array of real values
can be represented with an array of α × (⌊ α2 ⌋ + 1) complex
values. Furthermore U H V H = (U V )H , so the products
of the missing values can be reconstructed simply by taking
the complex conjugate of the computed values. Thus the
multiply stage of the FFT convnet algorithm with tile size
α
H
⌉⌈ W
α = m + r − 1 requires N ⌈ m
m ⌉CKα(⌊ 2 ⌋ + 1) comα
plex multiplications, or (⌊ 2 ⌋+1)/α complex multiplies per
input.
Using the standard algorithm for multiplying complex
numbers, this equals 4(⌊ α2 ⌋ + 1)/α > 2 real multiplies per
input.
Another technique, which to our knowledge has not been
used in convnets, is to use a fast algorithm to multiply complex numbers with 3 real multiplications [13]:
(x0 + ix1 )(y0 + iy1 ) = [x0 y0 − x1 y1 , i(x0 y1 + x1 y0 )]
= [uc va + ua vc , i(ua vc − ub vb )]
(16)
where
va = y0
ua = x0
ub = x0 + x1 , vb = y1
uc = x1 − x0 vc = y0 + y1
(17)
Tile
3
4
5
6
8
16
32
64
128
256
α′
9.00
4.00
2.78
2.25
1.78
Winograd
β′
γ′
2.00 1.75
3.60 2.24
4.33 2.00
6.50 2.23
δ′
1.50
2.24
2.78
4.38
α′
FFT
β ′ , γ ′, δ′
4.44
2.94
2.42
2.20
2.10
2.05
2.42
4.23
6.24
8.30
10.37
12.42
Table 1. Multiply (α′ ), data transform (β ′ ), filter transform (γ ′ ),
and inverse transform (δ ′ ) normalized arithmetic complexity versus tile size, for both Winograd and FFT based convolution.
F(4x4,3x3) has tile size 6. Direct convolutions has tile size 3.
An FFT based convnet algorithm can incorporate this by
modifying the FFT transforms of the filter and data to output
the the real valued matrices (Ua , Ub , Uc ) and (Va , Vb , Vc )
instead of the complex valued matrices U and V . This adds
2 floating point instructions per output to the filter transform, and 1 to the data transform. It also increases the memory footprint of each matrix by half.
Then we can calculate M = U V using 3 calls to a standard real matrix multiply function (e.g. SGEMM):
T = Ua Vc
M0 = Uc Va + T
M1 = −Ub Vb + T, M = (M0 , iM1 )
(18)
The accumulation of temporary matrix T is performed
using regular SGEMM with β = 1 and C = T , at the cost
of adding 2 floating point instructions per output. We can
think of these instructions as adding to the inverse transform
cost. The temporary matrix T increases memory use by
half, so that the total workspace size is approximately twice
that of FFT based convolution with direct CGEMM.
Combining Hermitian symmetry with fast CGEMM
gives us a multiplication stage with 3(⌊ α2 ⌋ + 1)/α > 1.5
real multiplies per input. Recall that the multiply stage of
the Winograd algorithms is always 1 real multiply per input. Thus even with fast CGEMM, FFT base convolution
must use a significantly larger tile size in order to rival the
arithmetic complexity of the Winograd algorithms.
For the FFT transform itself, we consider the split-radix
FFT algorithm, which is the minimal practical FFT algorithm when N is a power of 2 [9, p. 150]. We assume the
2D FFT transform is constructed using row-column composition, and borrow the complexity figures from the DSP
Handbook [9, pp. 173,175] for Table 1.
Tile
8
16
32
64
128
256
FFT with Fast CGEMM
α′
β′
γ′
δ′
3.33
3.77
4.30
4.30
2.20
6.23
6.82
6.82
1.81
8.94
9.57
9.57
1.65 11.72 12.36 12.36
1.57 14.48 15.14 15.14
1.54 17.22 17.88 17.88
Table 2. Normalized arithmetic complexity for FFT filtering with
fast CGEMM. Fast CGEMM uses 3 multiplies per complex multiply instead of 4, but has slightly greater transform overhead and
uses more memory.
5. Arithmetic Complexity Analysis
In our model of fast convnets, the arithmetic complexity
of the multiplication stage is:
M = N ⌈H/m⌉⌈W/n⌉CK(m + R − 1)(n + S − 1) (19)
When m = n = 1, the formula equals the arithmetic
complexity of direct convolution. Therefore direct convolution is the minimal algorithm for F (1 × 1, R × S)
Although our analysis employs minimal convolutions,
the convnet layer itself is still not minimal because it performs more convolutions than are strictly necessary. We
could reduce the number of convolutions by employing
Strassen recursions as in [3], but each recursion reduces
all 3 dimensions of our matrices by half while providing
only an 87 reduction in arithmetic complexity. The matrix
multiplications cannot be computed efficiently if C or K
is too small. Fast convolution alone provides a 2.25 or
larger arithmetic complexity reduction while shrinking only
the largest dimension of the matrix, P . Still, for layers
with large C, K, and P , it may be worthwhile to perform
Strassen recursions in addition to fast convolution. We leave
this as an area for further research.
In order to simplify the equations, we will henceforth
assume that W/m and H/n have no remainders. We also
assume square filters and blocks, R = S and m = n.
The multiplication complexity can be rewritten as:
M = (m + R − 1)2 /m2 N HW CK
= α′ N HW CK
(20)
where α = (m + R − 1)2 and α′ = α/m2
The total arithmetic complexities of the data, filter, and
inverse transforms can be written as:
T (D) = β/m2 N HW C
T (F ) = γCK
2
T (I) = δ/m N HW K
(21)
where β, γ, and δ are the number of floating point instructions used by the corresponding transforms for single tiles.
Dividing the complexity of each transform by M yields
its relative complexity:
T (D)/M = β/(Kα2 ) = β ′ /K
T (F )/M = γ/(N HW α2 /m2 )
= γ/(P α2 ) = γ ′ /P
2
6. GPU Implementation
(22)
′
T (I)/M = δ/(Cα ) = δ /C
We call β ′ , γ ′ , and δ ′ the normalized arithmetic complexities of the data, filter, and inverse transforms, respectively.
P = N HW/m2 is the number of tiles per channel.
Adding the terms for each stage gives the total arithmetic
complexity of the convnet layer:
L = α′ (1 + β ′ /K + γ ′ /P + δ ′ /C)N HW CK
large enough dimensions so that the multiply stage is efficient. This is problematic for current GPUs, which have a
limited amount of on chip memory. CPUs have large caches
and might therefore compute FFT based convolution more
efficiently.
(23)
In order to achieve a large speedup, the multiplication
complexity α′ must be as small as possible, and the transform complexities β ′ , γ ′ , and δ ′ must each be small compared with K, P , and C, respectively.
For direct convolution, α′ = α2 = R2 and β ′ = γ ′ =
δ ′ = 0. Therefore the maximum speedup of a fast algorithm
versus direct convolution is R2 /α′ .
We list the normalized transform complexity for different tile sizes and algorithms in Tables 1 and 2. Due to its
similarity to our approach, FFT based convolution complexity can also be measured with Equation 23.
FFT based convnet layers with direct CGEMM must use
tile size at least 64 × 64 to equal the multiplication stage
complexity of Winograd F (4 × 4, 3 × 3) and its 6 × 6 tile,
but then it incurs much greater transform overhead. Also
a 64 × 64 tile will waste computation on many unwanted
pixels for images with sizes that are not close to a multiple
of 62 × 62. Even for moderate size layers, a moderate to
large minibatch must be used, or there will be too few tiles
to compute the CGEMM efficiently. Finally, the memory
used by a single transformed filter channel is 64 × 64 =
4096 units, which is a large expansion of the 3 × 3 = 9 unit
filter. The 6x6 tile of F (4 × 4) expands the same filter to
6 × 6 = 36 units.
FFT based convnet layers with fast CGEMM can be
much more competitive with Winograd algorithms. They
have multiplication stage parity with tile size 16, and reasonable transform complexity. Also tile size 16 generates a
reasonably large number of tiles with large convnet layers
or moderate batch size.
Even with fast CGEMM, the larger tile size compared to
Winograd means FFT based convnet implementations must
have a large memory workspace to hold transformed data.
A decent amount of transformed data must be held in order
to amortize transform cost and to generate matrices with
We implemented F (2 × 2, 3 × 3) for NVIDIA Maxwell
GPUs and tested on the NVIDIA Titan X model.
The small 4 × 4 tile size and light weight transforms of
F (2 × 2, 3 × 3) make possible a fused implementation of
the algorithm stages, where the the data and filter transform,
16 batched matrix multiplies (GEMMs), and inverse transform are all computed in the same block. Another resource
limit is the instruction cache, which can only fit about 720
instructions. Our main loop is larger than this, but aligning
the start of the loop with the 128 byte instruction cache-line
boundary helps mitigate the cost of a cache miss.
The 16 batched GEMMs compute 32×32 outputs, which
enables us to fit the workspace in the registers and shared
memory of a single block and still have 2 active blocks per
SM for latency hiding. Zero padding is implicit through use
of predicates. If the predicate deselects a global image load,
the zero value is loaded with a dual issued I2I instruction.
Image data is stored in CHWN order to facilitate contiguous and aligned memory loads, significantly reducing
over-fetch. We employ a “super blocking” strategy to load
32 tiles of size 4 × 4 from a configurable number of images,
rows, and columns. For N >= 32, we load tiles from 32
separate images. For N < 32, we load a super block of
X × Y = 32/N tiles per image. This strategy facilitates
efficient loads with small batch sizes, as the W × N dimensions of the input data are contiguous in memory. Furthermore, the 2 pixel overlap between adjacent tiles causes high
L1 cache hit rates when using several tiles in a super block.
We also employ L2 cache blocking to increase the re-use
of overlapping blocks. Since the number of image tiles is
typically much larger than the number of filters, our block
mapping iterates over a group of up to 128 filters in the inner
loop, and then iterates over all image tiles in the second
loop. All channels of the filter group fit in L2 cache, so
each filter will only be loaded once from DDR memory, and
each image tile will be loaded ⌈K/128⌉ times as we iterate
over the filter groups. This strategy reduces DDR memory
bandwidth by almost half.
We implemented a version of our kernel that loads fp16
data, which decreases global memory bandwidth. We also
implemented a variant that we call “FX” that runs a filter
transform kernel first and stores the result in a workspace
buffer. The convolution kernel loads transformed filter
values from the workspace as needed. The size of the
workspace is only 16KC units of memory, which equals
just 16MB when K = C = 512 and data is fp32.
Layer
conv 1.1
conv 1.2
conv 2.1
conv 2.2
conv 3.1
conv 3.2
conv 4.1
conv 4.2
conv 5
Total
Depth
1
1
1
1
1
3
1
3
4
C ×H ×W
3 × 224 × 224
64 × 224 × 224
64 × 112 × 112
128 × 112 × 112
128 × 56 × 56
256 × 56 × 56
256 × 28 × 28
512 × 28 × 28
512 × 14 × 14
K
64
64
128
128
256
256
512
512
512
GFLOPs
0.17
3.70
1.85
3.70
1.85
11.10
1.85
11.10
3.70
39.02
Table 3. Convolution layers of VGG network E. All layers uses
3 × 3 filters. Depth indicates the number of times a given layer
shape occurs in the network. GFLOPs is weighted by depth and
assumes N=1.
7. Experiments
We ran accuracy and speed experiments with VGG Network E [11]. This is a deep network that uses 3×3 filters exclusively in the convolution layers, which are summarized
in Table 3.
We tested the accuracy of our fast algorithms with both
single precision (fp32) and half precision (fp16) data and
filters. In all tests we used fp32 arithmetic instructions. We
used random data and filters from the uniform distribution
[−1, 1] and measured absolute element error. Ground truth
was computed by direct convolution using a double precision accumulator for reductions.
We measured the speed of our GPU implementation of
F (2 × 2, 3 × 3) and compared with cuDNN v3 [1] on a superclocked NVIDIA Titan X GPU. We disabled boost clock
and observed a maximum clock rate of 1126MHz. The
GPU has 3072 cores, yielding a device peak throughput of
2 × 3072 × 1126 = 6.96 TFLOPS.
Speed for a given layer was calculated by dividing the
number of GFLOPs of computation required by direct convolution, as tabulated in 3, by the run time in milliseconds to
yield Effective TFLOPS. The reduction of arithmetic complexity allows fast algorithms to have Effective TFLOPS
that can exceed device peak throughput.
Total GFLOPs and run time were calculated by weighting the GFLOPs and run time for each layer by its depth,
and total throughput was calculated as the ratio of the two.
8. Results
Table 4 shows the numeric accuracy of the different convolution layer algorithms tested with single precision (fp32)
and half precision (fp16) input data and filters.
F (2 × 2, 3 × 3) is actually slightly more accurate than
direct convolution. Its simple transforms do not lose much
precision, and its multiplication stage performs a reduction
over C channels, rather than the RSC filter elements reduced by direct convolution. F (4 × 4, 3 × 3) has a larger
error, but it is still more accurate than direct convolution
with fp16 data.
All tested algorithms are equally accurate with fp16 data.
Here accuracy is limited by the precision of the inputs. Because direct convolution is accurate enough for training and
inference with low precision data [4, 5], we conclude that
F (4 × 4, 3 × 3) is too.
Table 5 and Table 6 show the total throughput for VGG
Network E layers for cuDNN and our F (2×2, 3×3) implementation for fp32 and fp16 data for different batch sizes.
For fp32 data, F (2 × 2, 3 × 3) is 1.48X at N = 64 and
2.26X as fast at N = 1. The throughput at N = 16 is 9.49
TFLOPS. For fp16 data, F (2×2, 3×3) extends its lead over
cuDNN, recording 10.28 TFLOPS throughput for N = 64.
N = 8 performance is still very good at 9.57 TFLOPS.
Figure 1 shows throughput by layer. Hatch marks indicate the layers where cuDNN used the FFT algorithm, otherwise direct convolution was used. For F (2 × 2, 3 × 3),
hatch marks indicate that the external filter transform (FX)
was used, otherwise the fused transform was faster.
cuDNN appears to erroneously select its FFT algorithm
for intermediate values of N despite the fact that it performs
very poorly, under 2 TFLOPS. While this is probably just a
bug, it is revealing. Low performance at moderate values of
N suggests that the FFT convolution implementation either
uses large tiles, or possibly just a single tile per image, as
in [12], which leads to inefficient multiplication stages unless N is large. At large N , cuDNN FFT performs much
better, but stays well under 8 TFLOPS.
F (2×2, 3×3) performs better than cuDNN at every layer
and batch size, except layer conv1.1, which contributes less
than 0.5% of the total network computation.
In general, we found that the FX variant of our implementation performed best unless the number of filters and
channels was very large. Computing the filter transform is
heavily memory bound, therefore transforming a larger filter bank decreases computational efficiency.
The worst F (2 × 2, 3 × 3) performance occurs for the
14×14 layers when N = 1. In this case the 8×4 superblock
runs over the image boundary and computes unwanted pixels. Throughput on this layer configuration is still over 5
TFLOPS, where cuDNN performance is just 1.6 TFLOPS.
cuDNN FFT uses a global memory workspace up to 2.6
GB in our experiments. By contrast, our fused F (2 × 2, 3 ×
3) implementation does not use any global workspace, and
the FX variant uses no more than 16 MB.
F (2 × 2, 3 × 3) performance shows new capabilities for
high throughput and small batch size with state of the art
convolutional neural networks. We expect performance to
increase again when F (4 × 4, 3 × 3) is implemented.
Layer
1.2
2.2
3.2
4.2
5
Direct
4.01E-05
8.01E-05
1.53E-04
3.20E-04
3.43E-04
fp32
F(2x2,3x3)
1.53E-05
2.86E-05
5.34E-05
5.34E-05
4.20E-05
F(4x4,3x3)
2.84E-04
5.41E-04
9.06E-04
1.04E-03
1.08E-03
vgg.conv1.1
12
10
8
6
4
2
0
fp16
1.14E-02
1.45E-02
1.99E-02
3.17E-02
2.61E-02
vgg.conv1.2
12
10
8
6
4
2
0
vgg.conv2.1
12
10
8
6
4
2
0
Table 4. Maximum element error on VGG network layers. With
fp32 data, F (2×2, 3×3) is more accurate than direct convolution.
With fp16 data, all algorithms are equally accurate.
vgg.conv2.2
N
1
2
4
8
16
32
64
cuDNN
msec TFLOPS
12.52
3.12
20.36
3.83
104.70
1.49
241.21
1.29
203.09
3.07
237.05
5.27
394.05
6.34
F(2x2,3x3)
msec TFLOPS
5.55
7.03
9.89
7.89
17.72
8.81
33.11
9.43
65.79
9.49
132.36
9.43
266.48
9.37
Speedup
2.26X
2.06X
5.91X
7.28X
3.09X
1.79X
1.48X
Table 5. cuDNN versus F (2 × 2, 3 × 3) performance on VGG
Network E with fp32 data. Throughput is measured in Effective
TFLOPS, the ratio of direct algorithm GFLOPs to run time.
N
1
2
4
8
16
32
64
cuDNN
msec TFLOPS
14.58
2.68
20.94
3.73
104.19
1.50
241.87
1.29
204.01
3.06
236.13
5.29
395.93
6.31
F(2x2,3x3)
msec TFLOPS
5.53
7.06
9.83
7.94
17.50
8.92
32.61
9.57
62.93
9.92
123.12
10.14
242.98
10.28
Effective TFLOPS
12
10
8
6
4
2
0
vgg.conv3.1
12
10
8
6
4
2
0
vgg.conv3.2
12
10
8
6
4
2
0
vgg.conv4.1
12
10
8
6
4
2
0
vgg.conv4.2
12
10
8
6
4
2
0
Speedup
2.64X
2.13X
5.95X
7.42X
3.24X
1.92X
1.63X
Table 6. cuDNN versus F (2 × 2, 3 × 3) performance on VGG
Network E with fp16 data.
12
10
8
6
4
2
0
vgg.conv5
1
2
4
cuDNN
cuDNN FFT
cudNN fp16
cudNN FFT fp16
8
Batch Size
16
32
64
F(2x2,3x3)
F(2x2,3x3) FX
F(2x2,3x3) fp16
F(2x2,3x3) FX fp16
Figure 1. VGG net Effective TFLOPS vs. batch size for cuDNN
and F (2 × 2, 3 × 3) on a 6.96 TFLOPS NVIDIA Titan X GPU.
References
[1] cuDNN. https://developer.nvidia.com/cudnn.
Accessed: 2015-11-01. 1, 7
[2] Richard E Blahut. Fast algorithms for signal processing. Cambridge University Press, 2010. 3, 4
[3] Jason Cong and Bingjun Xiao. Minimizing computation in convolutional neural networks. In Artificial Neural Networks and Machine Learning–ICANN
2014, pages 281–290. Springer, 2014. 1, 5
[4] Matthieu Courbariaux, Yoshua Bengio, and JeanPierre David. Low precision arithmetic for deep learning. CoRR, abs/1412.7024, 2014. 4, 7
[5] Suyog Gupta, Ankur Agrawal, Kailash Gopalakrishnan, and Pritish Narayanan.
Deep learning
with limited numerical precision. arXiv preprint
arXiv:1502.02551, 2015. 4, 7
[6] Suyog Gupta, Wei Zhang, and Josh Milthrope. Model
accuracy and runtime tradeoff in distributed deep
learning. arXiv preprint arXiv:1509.04210, 2015. 1
[7] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by
reducing internal covariate shift. arXiv preprint
arXiv:1502.03167, 2015. 1
[8] Alex Krizhevsky. One weird trick for parallelizing convolutional neural networks. arXiv preprint
arXiv:1404.5997, 2014. 1
[9] V. Madisetti. The Digital Signal Processing Handbook. Number v. 2 in Electrical engineering handbook
series. CRC, 2010. 4, 5
[10] Michaël Mathieu, Mikael Henaff, and Yann LeCun.
Fast training of convolutional networks through ffts.
CoRR, abs/1312.5851, 2013. 1, 4
[11] Karen Simonyan and Andrew Zisserman. Very deep
convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 1, 7
[12] Nicolas Vasilache, Jeff Johnson, Michaël Mathieu,
Soumith Chintala, Serkan Piantino, and Yann LeCun.
Fast convolutional nets with fbfft: A GPU performance evaluation. CoRR, abs/1412.7580, 2014. 1,
4, 7
[13] Shmuel Winograd. Arithmetic complexity of computations, volume 33. Siam, 1980. 1, 2, 3, 4
[14] Shmuel Winograd. On multiplication of polynomials
modulo a polynomial. SIAM Journal on Computing,
9(2):225–229, 1980. 2
| 9cs.NE
|
Under consideration for publication in Theory and Practice of Logic Programming
1
arXiv:1210.2282v2 [cs.PL] 9 Oct 2012
Towards Multi-Threaded Local Tabling
Using a Common Table Space
MIGUEL AREIAS and RICARDO ROCHA
CRACS & INESC TEC, Faculty of Sciences, University of Porto
Rua do Campo Alegre, 1021/1055, 4169-007 Porto, Portugal
(e-mail: {miguel-areias,ricroc}@dcc.fc.up.pt)
submitted 1 January 2003; revised 1 January 2003; accepted 1 January 2003
Abstract
Multi-threading is currently supported by several well-known Prolog systems providing a
highly portable solution for applications that can benefit from concurrency. When multithreading is combined with tabling, we can exploit the power of higher procedural control
and declarative semantics. However, despite the availability of both threads and tabling
in some Prolog systems, the implementation of these two features implies complex ties
to each other and to the underlying engine. Until now, XSB was the only Prolog system
combining multi-threading with tabling. In XSB, tables may be either private or shared
between threads. While thread-private tables are easier to implement, shared tables have
all the associated issues of locking, synchronization and potential deadlocks. In this paper,
we propose an alternative view to XSB’s approach. In our proposal, each thread views its
tables as private but, at the engine level, we use a common table space where tables are
shared among all threads. We present three designs for our common table space approach:
No-Sharing (NS) (similar to XSB’s private tables), Subgoal-Sharing (SS) and Full-Sharing
(FS). The primary goal of this work was to reduce the memory usage for the table space
but, our experimental results, using the YapTab tabling system with a local evaluation
strategy, show that we can also achieve significant reductions on running time. To appear
in Theory and Practice of Logic Programming.
KEYWORDS: Tabling, Multi-Threading, Implementation.
1 Introduction
Tabling (Chen and Warren 1996) is a recognized and powerful implementation technique that overcomes some limitations of traditional Prolog systems in dealing with
recursion and redundant sub-computations. In a nutshell, tabling is a refinement
of SLD resolution that stems from one simple idea: save intermediate answers from
past computations so that they can be reused when a similar call appears during
the resolution process1 . Tabling based models are able to reduce the search space,
1
We can distinguish two main approaches to determine similarity between tabled subgoal calls:
variant-based tabling and subsumption-based tabling.
2
Miguel Areias and Ricardo Rocha
avoid looping, and always terminate for programs with the bounded term-size property (Chen and Warren 1996). Work on tabling, as initially implemented in the XSB
system (Sagonas and Swift 1998), proved its viability for application areas such as
natural language processing, knowledge based systems, model checking, program
analysis, among others. Currently, tabling is widely available in systems like XSB,
Yap, B-Prolog, ALS-Prolog, Mercury and Ciao.
Nowadays, the increasing availability of computing systems with multiple cores
sharing the main memory is already a standardized, high-performance and viable
alternative to the traditional (and often expensive) shared memory architectures.
The number of cores per processor is expected to continue to increase, further expanding the potential for taking advantage of multi-threading support. The ISO
Prolog multi-threading standardization proposal (Moura 2008) is currently implemented in several Prolog systems including XSB, Yap, Ciao and SWI-Prolog, providing a highly portable solution given the number of operating systems supported
by these systems. Multi-threading in Prolog is the ability to concurrently perform
multiple computations, in which each computation runs independently but shares
the database (clauses).
When multi-threading is combined with tabling, we have the best of both worlds,
since we can exploit the combination of higher procedural control with higher declarative semantics. In a multi-threaded tabling system, tables may be either private
or shared between threads. While thread-private tables are easier to implement,
shared tables have all the associated issues of locking, synchronization and potential deadlocks. Here, the problem is even more complex because we need to
ensure the correctness and completeness of the answers found and stored in the
shared tables. Thus, despite the availability of both threads and tabling in Prolog
compilers such as XSB, Yap, and Ciao, the implementation of these two features
such that they work together seamlessly implies complex ties to one another and
to the underlying engine. Until now, XSB was the only system combining tabling
with multi-threading, for both private and shared tables (Marques and Swift 2008;
Swift and Warren 2012). For shared tables, XSB uses a semi-naive approach that,
when a set of subgoals computed by different threads is mutually dependent, then
a usurpation operation (Marques 2007; Marques et al. 2010) synchronizes threads
and a single thread assumes the computation of all subgoals, turning the remaining
threads into consumer threads.
The basis for our work is also on multi-threaded tabling using private tables, but
we propose an alternative view to XSB’s approach. In our proposal, each thread
has its own tables, i.e., from the thread point of view the tables are private, but at
the engine level we use a common table space, i.e., from the implementation point
of view the tables are shared among all threads. We present three designs for our
common table space approach: No-Sharing (NS) (similar to XSB with private tables), Subgoal-Sharing (SS) and Full-Sharing (FS). Experimental results, using the
YapTab tabling system (Rocha et al. 2005) with a local evaluation strategy (Freire
et al. 1995), show that the FS design can achieve significant reductions on memory
usage and on execution time, compared to the NS and SS designs, for a set of worst
case scenarios where all threads start with the same query goal.
Towards Multi-Threaded Local Tabling Using a Common Table Space
3
The remainder of the paper is organized as follows. First, we describe YapTab’s
table space organization and XSB’s approach for multi-threaded tabling. Next, we
introduce our three designs and discuss important implementation details. We then
present some experimental results and outline some conclusions.
2 Basic Concepts
In this section, we introduce some background needed for the following sections.
We begin by describing the actual YapTab’s table space organization, and then we
briefly present XSB’s approach for supporting multi-threaded tabling.
2.1 YapTab’s Table Space Organization
The basic idea behind tabling is straightforward: programs are evaluated by storing
answers for tabled subgoals in an appropriate data space, called the table space.
Similar calls to tabled subgoals are not re-evaluated against the program clauses,
instead they are resolved by consuming the answers already stored in their table
entries. During this process, as further new answers are found, they are stored in
their tables and later returned to all similar calls.
A critical component in the implementation of an efficient tabling system is thus
the design of the data structures and algorithms to access and manipulate tabled
data. The most successful data structure for tabling is tries (Ramakrishnan et al.
1999). Tries are trees in which common prefixes are represented only once. The trie
data structure provides complete discrimination for terms and permits look up and
possibly insertion to be performed in a single pass through a term, hence resulting
in a very efficient and compact data structure for term representation. Figure 1
shows the general table space organization for a tabled predicate in YapTab.
Tabled Predicate
Compliled Code
Table Entry
Subgoal Trie Structure
Subgoal
Frame
call_1
Subgoal
Frame
call_2
. . .
Subgoal
Frame
call_n
Answer
Trie
Structure
Answer
Trie
Structure
. . .
Answer
Trie
Structure
Fig. 1. YapTab’s table space organization
At the entry point we have the table entry data structure. This structure is
4
Miguel Areias and Ricardo Rocha
allocated when a tabled predicate is being compiled, so that a pointer to the table
entry can be included in its compiled code. This guarantees that further calls to
the predicate will access the table space starting from the same point. Below the
table entry, we have the subgoal trie structure. Each different tabled subgoal call
to the predicate at hand corresponds to a unique path through the subgoal trie
structure, always starting from the table entry, passing by several subgoal trie data
units, the subgoal trie nodes, and reaching a leaf data structure, the subgoal frame.
The subgoal frame stores additional information about the subgoal and acts like an
entry point to the answer trie structure. Each unique path through the answer trie
data units, the answer trie nodes, corresponds to a different answer to the entry
subgoal.
2.2 XSB’s Approach to Multi-Threaded Tabling
XSB offers two types of models for supporting multi-threaded tabling: private tables
and shared tables (Swift and Warren 2012).
For private tables, each thread keeps its own copy of the table space. On one hand,
this avoids concurrency over the tables but, on the other hand, the same table can
be computed by several threads, thus increasing the memory usage necessary to
represent the table space.
For shared tables, the running threads store only once the same table, even if multiple threads use it. This model can be viewed as a variation of the table-parallelism
proposal (Freire et al. 1995), where a tabled computation can be decomposed into
a set of smaller sub-computations, each being performed by a different thread.
Each tabled subgoal is computed independently by the first thread calling it, the
generator thread, and each generator is the sole responsible for fully exploiting and
obtaining the complete set of answers for the subgoal. Similar calls by other threads
are resolved by consuming the answers stored by the generator thread.
In a tabled evaluation, there are several points where we may have to choose
between continuing forward execution, backtracking, consuming answers from the
table, or completing subgoals. The decision on which operation to perform is determined by the evaluation strategy. The two most successful strategies are batched
evaluation and local evaluation (Freire et al. 1996). Batched evaluation favors forward execution first, backtracking next, and consuming answers or completion last.
It thus tries to delay the need to move around the search tree by batching the return
of answers. When new answers are found for a particular tabled subgoal, they are
added to the table space and the evaluation continues. On the other hand, local
evaluation tries to complete subgoals as soon as possible. When new answers are
found, they are added to the table space and the evaluation fails. Answers are only
returned when all program clauses for the subgoal at hand were resolved.
Based on these two strategies, XSB supports two types of concurrent evaluations:
concurrent local evaluation and concurrent batched evaluation. In the concurrent local evaluation, similar calls by other threads are resolved by consuming the answers
stored by the generator thread, but a consumer thread suspends execution until
the table is completed. In the concurrent batched evaluation, new answers are con-
Towards Multi-Threaded Local Tabling Using a Common Table Space
5
sumed as they are found, leading to more complex dependencies between threads.
In both evaluation strategies, when a set of subgoals computed by different threads
is mutually dependent, then a usurpation operation (Marques et al. 2010) synchronizes threads and a single thread assumes the computation of all subgoals, turning
the remaining threads into consumer threads.
3 Our Approach
Yap implements a SWI-Prolog compatible multi-threading library (Wielemaker
2003). Like in SWI-Prolog, Yap’s threads have their own execution stacks and
only share the code area where predicates, records, flags and other global nonbacktrackable data are stored. Our approach for multi-threaded tabling is still based
on this idea in which each computational thread runs independently. This means
that each tabled evaluation depends only on the computations being performed by
the thread itself, i.e., there isn’t the notion of being a consumer thread since, from
each thread point of view, a thread is always the generator for all of its subgoal
calls. We next introduce the three alternative designs for our approach: No-Sharing
(NS), Subgoal-Sharing (SS) and Full-Sharing (FS). In what follows, we assume a
local evaluation strategy.
3.1 No-Sharing
The starting point of our work is the situation where each thread allocates fully
private tables for each new subgoal called during its computation. Figure 2 shows
the configuration of the table space if several different threads call the same tabled
subgoal call i. One can observe that the table entry data structure still stores the
common information for the predicate (such as the arity or the evaluation strategy),
and then each thread t has its own cell Tt inside a bucket array which points to
the private data structures. The subgoal trie structure, the subgoal frames and the
answer trie structures are private to each thread and they are removed when the
thread finishes execution.
The memory usage for this design for a particular tabled predicate P , assuming that all running threads N T have completely evaluated the same number N S
of subgoals, is sizeof (T EP ) + sizeof (BAP ) + [sizeof (ST SP ) + [sizeof (SFP ) +
sizeof (AT SP )] ∗ N S] ∗ N T , where T EP and BAP represent the common table entry and bucket array data structures, ST SP and AT SP represent the nodes inside
the subgoal and answer trie structures, and SFP represents the subgoal frames.
3.2 Subgoal-Sharing
In our second design, the threads share part of the table space. Figure 3 shows again
the configuration of the table space if several different threads call the same tabled
subgoal call i. In this design, the subgoal trie structure is now shared among the
threads and the leaf data structures in each subgoal trie path, instead of pointing
to a subgoal frame, they now point to a bucket array. Each thread t has its own
6
Miguel Areias and Ricardo Rocha
Table Entry
. . .
Tk-3 Tk-2 Tk-1
Subgoal
Trie
Structure
Subgoal
Trie
Structure
Subgoal
Trie
Structure
Subgoal
Frame
call_i
Subgoal
Frame
call_i
Subgoal
Frame
call_i
Answer
Trie
Structure
Answer
Trie
Structure
Answer
Trie
Structure
T0
T1
T2
Fig. 2. Table space organization for the NS design
cell Tt inside the bucket array which then points to a private subgoal frame and
answer trie structure.
In this design, concurrency among threads is restricted to the allocation of new
entries on the subgoal trie structure. Whenever a thread finishes execution, its
private structures are removed, but the shared part remains present as it can be in
use or be further used by other threads. Assuming again that all running threads
N T have completely evaluated the same number N S of subgoals, the memory usage
for this design for a particular tabled predicate P is sizeof (T EP )+sizeof (ST SP )+
[sizeof (BAP ) + [sizeof (SFP ) + sizeof (AT SP )] ∗ N T ] ∗ N S, where BAP represents
the bucket array pointing to the private data structures.
Table Entry
Subgoal Trie Structure
. . .
Tk-3 Tk-2 Tk-1
Subgoal
Frame
call_i
Subgoal
Frame
call_i
Subgoal
Frame
call_i
Answer
Trie
Structure
Answer
Trie
Structure
Answer
Trie
Structure
T0
T1
T2
Fig. 3. Table space organization for the SS design
Towards Multi-Threaded Local Tabling Using a Common Table Space
7
3.3 Full-Sharing
Our third design is the most sophisticated among three. Figure 4 shows its table
space organization if considering several different threads calling the same tabled
subgoal call i. In this design, part of the subgoal frame information (the subgoal
entry data structure in Fig. 4) and the answer trie structure are now also shared
among all threads. The previous subgoal frame data structure was split into two: the
subgoal entry stores common information for the subgoal call (such as the pointer
to the shared answer trie structure); the remaining information (the subgoal frame
data structure in Fig. 4) remains private to each thread.
Table Entry
Subgoal Trie Structure
Subgoal Entry call_i
T0
Answer
Trie
Structure
. . .
Tk-3 Tk-2 Tk-1
Subgoal
Frame
call_i
Subgoal
Frame
call_i
Subgoal
Frame
call_i
Sg_Entry
Sg_Entry
Sg_Entry
T1
T2
Fig. 4. Table space organization for the FS design
The subgoal entry also includes a bucket array, in which each cell Tt points to
the private subgoal frame of each thread t. The private subgoal frames include an
extra field which is a back pointer to the common subgoal entry. This is important
because, with that, we can keep unaltered all the tabling data structures that point
to subgoal frames. To access the private information on the subgoal frames there is
no extra cost (we still use a direct pointer), and only for the common information
on the subgoal entry we pay the extra cost of following an indirect pointer.
Again, assuming that all running threads N T have completely evaluated the same
number N S of subgoals, the memory usage for this design for a particular tabled
predicate P is sizeof (T EP ) + sizeof (ST SP ) + [sizeof (SEP ) + sizeof (BAP ) +
sizeof (AT SP ) + sizeof (SFP ) ∗ N T ] ∗ N S, where SEP and SFP represent, respectively, the shared subgoal entry and the private subgoal frame data structures.
In this design, concurrency among threads now also includes the access to the
subgoal entry data structure and the allocation of new entries on the answer trie
structures. However, this latest design has two major advantages. First, memory
usage is reduced to a minimum. The only memory overhead, when compared with a
single threaded evaluation, is the bucket array associated with each subgoal entry,
and apart from the split on the subgoal frame data structure, all the remaining
structures remain unchanged. Second, since threads are sharing the same answer
8
Miguel Areias and Ricardo Rocha
trie structures, answers inserted by a thread for a particular subgoal call are automatically made available to all other threads when they call the same subgoal. As
we will see in section 5, this can lead to reductions on the execution time.
4 Implementation
In this section, we discuss some low level details regarding the implementation
of the three designs. We begin by describing the expansion of the table space to
efficiently support multiple threads, next we discuss the locking schemes used to
ensure mutual exclusion over the table space, and then we discuss how the most
important tabling operations were extended for multi-threaded tabling support.
4.1 Efficient Support for Multiple Threads
Our proposals already include support for any number of threads working on the
same table. For that, we extended the original table data structures with bucket
arrays. For example, for the NS design, we introduced a bucket array in the table
entry (see Fig. 2), for the SS design, the bucket array follows a subgoal trie path
(see Fig. 3), and for the FS design, the bucket array is part of the new subgoal
entry data structure (see Fig. 4).
These bucket arrays contain as much entry cells as the maximum number of
threads that can be created in Yap (currently 1024). However, in practice, this
solution is highly inefficient and memory consuming, as we must always allocate
this huge bucket array even when only one thread will use it.
To solve this problem, we introduce a kind of inode pointer structure, where the
bucket array is split into direct bucket cells and indirect bucket cells. The direct
bucket cells are used as before and the indirect bucket cells are only allocated as
needed. This new structure applies to all bucket arrays in the three designs. Figure 5
shows an example on how this new structure is used in the FS design.
Subgoal Entry call_i
T0
Subgoal
Frame
call_i
Sg_Entry
T2
...
Ts-1
B0
Ts
Ts+1
Ts+2
...
Bu-3 Bu-2 Bu-1
Subgoal
Frame
call_i
Sg_Entry
. . .
Answer
Trie
Structure
T1
. . .
Tk-3
Tk-2
Tk-1
Subgoal
Frame
call_i
Sg_Entry
Fig. 5. Using direct and indirect bucket cells in the FS design
A bucket array has now two operating modes. If it is being used by a thread
Towards Multi-Threaded Local Tabling Using a Common Table Space
9
with an identification number t lower than a default starting size s (32 in our
implementation), then the buckets are used as before, meaning that the entry cell
Tt still points to the private information of the corresponding thread. But now, if a
thread with an identification number equal or higher than s appears, the thread is
mapped into one of the u undirected buckets (entry cells B0 until Bu−1 in Fig. 5),
which becomes a pointer to a second level bucket array that will now contain the
entry cells pointing to the private thread information. Given a thread t (t ≥ s), its
index in the first and in the second level bucket arrays is given by the division and
the remainder of (t − s) by u, respectively.
4.2 Table Locking Schemes
Remember that the SS and FS designs introduce concurrency among threads when
accessing shared resources of the table space. Here, we discuss how we use locking
schemes to ensure mutual exclusion when manipulating such shared resources.
We can say that there are two critical issues that determine the efficiency of a
locking scheme. One is the lock duration, that is, the amount of time a data structure
is locked. The other is the lock grain, that is, the amount of data structures that
are protected through a single lock request. It is the balance between lock duration
and lock grain that compromises the efficiency of different locking schemes.
The or-parallel tabling engine of Yap (Rocha et al. 2005) already implements
four alternative locking schemes to deal with concurrent table accesses: the Table
Lock at Entry Level (TLEL) scheme, the Table Lock at Node Level (TLNL) scheme,
the Table Lock at Write Level (TLWL) scheme, and the Table Lock at Write Level
- Allocate Before Check (TLWL-ABC) scheme. Currently, the first three are also
available on our multi-threaded engine. However, in what follows, we will focus our
attention only on the TLWL locking scheme, since its performance showed to be
clearly better than the other two (Rocha et al. 2004).
The TLWL scheme allows a single writer per chain of sibling nodes that represent
alternative paths from a common parent node (see Fig. 6). This means that each
node in the subgoal/answer trie structures is expanded with a locking field that,
once activated, synchronizes updates to the chain of sibling nodes, meaning that
only one thread at a time can be inserting a new child node starting from the same
parent node.
With the TLWL scheme, the process of check/insert a term t in a chain of sibling
nodes works as follows. Initially, the working thread starts by searching for t in the
available child nodes (the non-critical region) and only if the term is not found, it
will enter the critical region in order to insert it on the chain. At that point, it waits
until the lock be available, which can cause a delay proportional to the number of
threads that are accessing the same critical region at the same time.
In order to reduce the lock duration to a minimum, we have improved the original
TLWL scheme to use trylocks instead of traditional locks. With trylocks, when a
thread fails to get access to the lock, instead of waiting, it returns to the non-critical
region, i.e., it traverses the newly inserted nodes, if any, checking if t was, in the
meantime, inserted in the chain by another thread. If t is not found, the process
10
Miguel Areias and Ricardo Rocha
Table Entry
Parent
Child
Lock L1
Child
Child
Child
Locked by L1
Locked by L1
Locked by L1
Subgoal
Frame
Parent
Child
Lock L2
Child
Child
Child
Locked by L2
Locked by L2
Locked by L2
Fig. 6. The TLWL locking scheme
repeats until the thread get access to the lock, in order to insert t, or until t be
found. Figure 7 shows the pseudo-code for the implementation of this procedure
using the TLWL scheme with trylocks.
trie_node_check_insert(term T, parent trie node P)
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
14.
15.
16.
17.
18.
19.
20.
21.
22.
23.
last_child = NULL // used to mark the last child to be checked
do { // non-critical region
first_child = TrNode_first_child(P)
child = first_child
while (child != last_child) // traverse the chain of sibling nodes ...
if (TrNode_term(child) == T) // ... searching for T
return child
child = TrNode_sibling(child)
last_child = first_child
} while (! trylock(TrNode_lock(P)))
// critical region, lock is set
child = TrNode_first_child(P)
while (child != last_child) // traverse the chain of sibling nodes ...
if (TrNode_entry(child) == T) // ... searching for T
unlock(TrNode_lock(P)) // unlocking before return
return child
child = TrNode_sibling(child)
// create a new node to represent T
child = new_trie_node(T)
TrNode_sibling(child) = TrNode_first_child(P)
TrNode_first_child(P) = child
unlock(TrNode_lock(P)) // unlocking before return
return child
Fig. 7. Pseudo-code for the trie node check/insert operation
Initially, the procedure traverses the chain of sibling nodes, that represent al-
Towards Multi-Threaded Local Tabling Using a Common Table Space
11
ternative paths from the given parent node P, and checks for one representing the
given term T. If such a node is found (line 6) then execution is stopped and the
node returned (line 7). Otherwise, this process repeats (lines 3 to 10) until the
working thread gets access to the lock field of the parent node P. In each round,
the last child auxiliary variable marks the last node to be checked. It is initially
set to NULL (line 1) and then updated, at the end of each round, to the new first
child of the current round (line 9).
Otherwise, the thread gets access to the lock and enters the critical region (lines
12 to 23). Here, it first checks if T was, in the meantime, inserted in the chain
by another thread (lines 13 to 17). If this is not the case, then a new trie node
representing T is allocated (line 19) and inserted in the beginning of the chain
(lines 20 and 21). The procedure then unlocks the parent node (line 22) and ends
returning the newly allocated child node (line 23).
4.3 Tabling Operations
In YapTab, programs using tabling are compiled to include tabling operations that
enable the tabling engine to properly schedule the evaluation process. One of the
most important operations is the tabled subgoal call. This operation inspects the
table space looking for a subgoal similar to the current subgoal being called. If a
similar subgoal is found, then the corresponding subgoal frame is returned. Otherwise, if no such subgoal exists, it inserts a new path into the subgoal trie structure,
representing the current subgoal, and allocates a new subgoal frame as a leaf of the
new inserted path. Figure 8 shows how we have extended the tabled subgoal call
operation for multi-threaded tabling support.
tabled_subgoal_call(table entry TE, subgoal call SC, thread id TI)
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
14.
15.
16.
17.
18.
19.
20.
root = get_subgoal_trie_root_node(TE, TI)
leaf = check_insert_subgoal_trie(root, SC)
if (NS_design)
sg_fr = get_subgoal_frame(leaf)
if (not_exists(sg_fr))
sg_fr = new_subgoal_frame(leaf)
return sg_fr
else if (SS_design)
bucket = get_bucket_array(leaf)
if (not_exists(bucket))
bucket = new_bucket_array(leaf)
else if (FS_design)
sg_entry = get_subgoal_entry(leaf)
if (not_exists(sg_entry))
sg_entry = new_subgoal_entry(leaf)
bucket = get_bucket_array(sg_entry)
sg_fr = get_subgoal_frame(bucket)
if (not_exists(sg_fr))
sg_fr = new_subgoal_frame(bucket)
return sg_fr
Fig. 8. Pseudo-code for the tabled subgoal call operation
12
Miguel Areias and Ricardo Rocha
The procedure receives three arguments: the table entry for the predicate at
hand (TE), the current subgoal being called (SC), and the id of the working thread
(TI). The NS design, SS design and FS design macros define which table design
is enabled.
The procedure starts by getting the root trie node for the subgoal trie structure
that matches with the given thread id (line 1). Next, it checks/inserts the given
SC into the subgoal trie structure, which will return the leaf node for the path
representing SC (line 2). Then, if the NS design is enable, it uses the leaf node
to obtain the corresponding subgoal frame (line 4). If the subgoal call is new, no
subgoal frame still exists and a new one is created (line 6). Then, the procedure
ends by returning the subgoal frame (line 7). This code sequence corresponds to
the usual tabled subgoal call operation.
Otherwise, for the SS design, it follows the leaf node to obtain the bucket array
(line 9). If the subgoal call is new, no bucket exists and a new one is created (line
11). On the other hand, for the FS design, it follows the leaf node to obtain the
subgoal entry (line 13) and, again, if the subgoal call is new, no subgoal entry
exists and a new one is created (line 15). From the subgoal entry, it then obtains
the bucket array (line 16).
Finally, for both SS and FS designs, the bucket array is then used to obtain the
subgoal frame (line 17) and one more time, if the given subgoal call is new, a new
subgoal frame needs to be created (line 19). The procedure ends by returning the
subgoal frame (line 20). Note that, for the sake of simplicity, we omitted some of
the low level details in manipulating the bucket arrays, such as in computing the
bucket cells or in expanding the indirect bucket cels.
Another important tabling operation is the new answer. This operation checks
whether a newly found answer is already in the corresponding answer trie structure
and, if not, inserts it. Remember from section 2.2 that, with local evaluation, the
new answer operation always fails, regardless of the answer being new or repeated,
and that, with batched evaluation, when new answers are inserted the evaluation
should continue, failing otherwise. With the FS design, the answer trie structures are
shared. Thus, when several threads are inserting answers in the same trie structure,
it may be not possible to determine when an answer is new or repeated for a certain
thread. This is the reason why the FS design can be only safely used with local
evaluation. We are currently studying how to bypass this constraint in order to also
support the FS design with batched evaluation.
5 Experimental Results
In this section, we present some experimental results obtained for the three proposed
table designs using the TLWL scheme with traditional locks and with trylocks. The
environment for our experiments was a machine with 4 Six-Core AMD Opteron
(tm) Processor 8425 HE (24 cores in total) with 64 GBytes of main memory and
running the Linux kernel 2.6.34.9-69.fc13.x86 64 with Yap 6.3. To put our results
in perspective, we make a comparison with the multi-threaded implementation of
XSB, version 3.3.6, using thread-private tables.
Towards Multi-Threaded Local Tabling Using a Common Table Space
13
We used five sets of benchmarks. The Large Joins and WordNet sets were
obtained from the OpenRuleBench project2 ; the Model Checking set includes
three different specifications and transition relation graphs usually used in model
checking applications; the Path Left and Path Right sets implement two recursive
definitions of the well-known path/2 predicate, that computes the transitive closure
in a graph, using several different configurations of edge/2 facts (Fig. 9 shows an
example for each configuration). We experimented the BTree configuration with
depth 18, the Pyramid and Cycle configurations with depth 2000 and the Grid
configuration with depth 35. All benchmarks find all the solutions for the problem.
BTree
(depth 2)
Pyramid
(depth 4)
Cycle
(depth 4)
Grid
(depth 4)
Fig. 9. Edge configurations
Table 1 shows the execution time, in milliseconds, when running 1 working thread
with local scheduling, for our three table designs, using the TLWL scheme with
traditional locks (columns NS and FS) and with trylocks (columns SST and FST )3 ,
and for XSB. In parentheses, it also shows the respective overhead ratios when
compared with the NS design. The running times are the average of five runs. The
ratios marked with n.c. for XSB mean that we are not considering them in the
average results (we opted to do that since they correspond to running times much
higher than the other designs, which may suggest that something was wrong).
One can observe that, on average, the SS design and the XSB implementation
have a lower overhead ratio (around 10%) than the FS and FST designs (around
20%). For the SS and, mainly, for the FS approaches, this can be explained by the
higher complexity of the implementation and, in particular, by the cost incurred
with the extra code necessary to implement the TLWL locking scheme. Note that,
even with a single working thread, this code has to be executed.
Starting from these base results, Table 2 shows the overhead ratios, when compared with the NS design with 1 thread, for our table designs and XSB, when
running 16 and 24 working threads (the results are the average of five runs).
In order to create a worst case scenario that stresses the trie data structures,
we ran all threads starting with the same query goal. By doing this, it is expected
2
3
Available from http://rulebench.projects.semwebcentral.org. We also have results for the
other benchmarks proposed by the OpenRuleBench project (Liang et al. 2009) but, due to lack
of space, here we only include these two sets.
In general, for this set of benchmarks, the SS design presented similar results with traditional
locks and with trylocks and, thus, here we only show the results with trylocks.
14
Miguel Areias and Ricardo Rocha
Table 1. Execution time, in milliseconds, when running 1 working thread with local
scheduling, for the NS, SST , FS and FST designs and for XSB, and the respective
overhead ratios when compared with the NS design
Bench
NS
Large Joins
Join2
3,419
Mondial
730
Average
WordNet
Clusters
789
Hypo
1,488
Holo
694
Hyper
1,386
Tropo
598
Mero
678
Average
Model Checking
IProto
2,517
Leader
3,726
Sieve
23,645
Average
SST
FS
FST
3,418 (1.00)
725 (0.99)
(1.00)
3,868 (1.13)
856 (1.17)
(1.15)
3,842 (1.12)
887 (1.21)
(1.17)
990
1,671
902
1,587
784
892
981
1,728
881
1,565
763
869
982
1,720
884
1,576
762
864
(1.26)
(1.12)
(1.30)
(1.15)
(1.31)
(1.32)
(1.24)
2,449 (0.97)
3,800 (1.02)
24,402 (1.03)
(1.01)
(1.24)
(1.16)
(1.27)
(1.13)
(1.28)
(1.28)
(1.23)
2,816 (1.12)
3,830 (1.03)
24,479 (1.04)
(1.06)
XSB
3,444 (1.01)
1,637 (2.24)
(1.62)
(1.24)
(1.16)
(1.27)
(1.14)
(1.27)
(1.27)
(1.23)
549 (0.70)
992,388 (n.c.)
425 (0.61)
1,320 (0.95)
271 (0.45)
131,830 (n.c.)
(0.68)
2,828 (1.12)
3,897 (1.05)
25,201 (1.07)
(1.08)
3,675 (1.46)
10,354 (2.78)
27,136 (1.15)
(1.80)
Path Left
BTree
2,966
Pyramid
3,085
Cycle
3,828
Grid
1,743
Average
2,998
3,159
3,921
1,791
(1.01)
(1.02)
(1.02)
(1.03)
(1.02)
3,826
3,256
3,775
2,280
(1.29)
(1.06)
(0.99)
(1.31)
(1.16)
3,864
3,256
3,798
2,293
(1.30)
(1.06)
(0.99)
(1.32)
(1.17)
2,798
2,928
3,357
2,034
(0.94)
(0.95)
(0.88)
(1.17)
(0.98)
Path Right
BTree
4,568
Pyramid
2,520
Cycle
2,761
Grid
2,109
Average
5,048
2,531
2,773
2,110
(1.11)
(1.00)
(1.00)
(1.00)
(1.03)
5,673
3,664
3,994
3,097
(1.24)
(1.45)
(1.45)
(1.47)
(1.40)
5,701
3,673
3,992
3,117
(1.25)
(1.46)
(1.45)
(1.48)
(1.41)
3,551
2,350
2,817
2,462
(0.78)
(0.93)
(1.02)
(1.17)
(0.97)
Total Average
(1.08)
(1.21)
(1.22)
(1.12)
that they will access the table space, to check/insert for subgoals and answers, at
similar times, thus causing a huge stress on the same critical regions. In particular,
for this set of benchmarks, this will be specially the case for the answer tries (and
thus, for the FS and FST designs), since the number of answers clearly exceeds the
number of subgoals. Analyzing the general picture of Table 2, one can observe that,
on average, the NS and SST designs show very poor results for 16 and 24 threads. In
Towards Multi-Threaded Local Tabling Using a Common Table Space
15
Table 2. Overhead ratios, when compared with the NS design with 1 thread, for
the NS, SST , FS and FST designs and for XSB, when running 16 and 24 working
threads with local scheduling (best ratios are in bold)
Bench
NS
16 Threads
SST
FS
FST
XSB
NS
24 Threads
SST
FS
FST
XSB
Large Joins
Join2
Mondial
Average
7.96
1.05
4.51
8.05
1.07
4.56
3.14
1.46
2.30
3.14
1.53
2.34
5.74
2.43
4.08
24.78
1.13
12.96
24.84
1.13
12.98
3.77
1.60
2.68
3.76
1.64
2.70
8.64
2.53
5.58
WordNet
Clusters
Hypo
Holo
Hyper
Tropo
Mero
Average
6.29
5.33
6.15
8.03
6.03
4.90
6.12
5.61
5.09
5.41
7.65
4.96
4.92
5.61
3.92
4.56
3.73
3.57
3.93
3.90
3.93
3.94
2.99
3.72
2.94
3.95
3.71
3.54
2.82
n.c.
2.77
4.26
2.93
n.c.
3.19
12.23
9.20
10.92
21.34
13.46
8.93
12.68
8.67
8.33
9.87
16.82
8.44
7.96
10.02
4.52
5.21
4.67
4.59
5.64
4.59
4.87
4.55
4.15
4.55
3.34
5.68
4.44
4.45
4.87
n.c.
4.37
7.14
4.69
n.c.
5.27
Model Checking
IProto
4.15
Leader
1.02
Sieve
1.01
Average
2.06
4.20
1.04
1.04
2.09
1.60
1.05
1.05
1.24
1.55
1.07
1.08
1.23
1.92
2.80
1.15
1.95
7.16
1.02
1.02
3.07
7.31
1.04
1.04
3.13
1.71
1.05
1.06
1.27
1.63
1.07
1.08
1.26
2.14
2.79
1.15
2.03
Path Left
BTree
Pyramid
Cycle
Grid
Average
9.85
7.67
7.32
5.99
7.71
9.78
7.79
7.38
6.00
7.74
6.88
3.74
3.73
3.77
4.53
4.81
3.40
3.25
3.15
3.65
5.11
4.40
4.36
2.41
4.07
25.65
24.92
22.39
19.82
23.20
25.42
24.88
23.05
19.80
23.29
8.03
5.86
5.95
4.65
6.12
5.97
4.48
4.08
4.46
4.75
8.09
7.02
6.99
5.30
6.85
Path Right
BTree
Pyramid
Cycle
Grid
Average
13.82
17.09
17.96
9.52
14.60
13.13
17.00
18.17
9.48
14.44
10.57
14.85
17.05
7.13
12.40
5.54
8.15
8.36
5.53
6.90
6.33
5.94
6.63
3.75
5.66
29.53
46.25
47.89
26.58
37.56
27.36
45.31
47.60
27.80
37.02
Total Average
7.43
7.25
5.24
3.78
3.87
18.64
17.72
10.16 6.76 10.38
10.86 10.42 10.31
11.49 10.76 10.99
7.50
6.96 6.41
10.00 8.73
9.52
5.42
4.73
6.11
particular, these bad results are more clear in the benchmarks that allocate a higher
number of trie nodes. The explanation for this is the fact that we are using Yap’s
memory allocator, that is based on Linux system’s malloc, which can be a problem,
when making a lot of memory requests, since these requests require synchronization
at the low level implementation.
16
Miguel Areias and Ricardo Rocha
For the FS and FST designs, the results are significantly better and, in particular
for FST , the results show that its trylock implementation is quite effective in reducing contention and, consequently, the running times for most of the experiments.
Regarding XSB, for 16 threads, the results are similar to the FST design (3.87 for
XSB and 3.78 for FST , on average) but, for 24 threads, the FST is noticeable better
(6.11 for XSB and 4.73 for FST , on average). These results are more important
since XSB shows base execution times (with 1 thread) lower than FST (please revisit Table 1) and since FST also pays the cost of using Yap’s memory allocator
based on Linux system’s malloc.
We can say that there are two main reasons for the good results of the FS design.
The first, and most important, is that the FS design can effectively reduce the
memory usage of the table space, almost linearly in the number of threads4 , which
has the collateral effect of also reducing the impact of Yap’s memory allocator. The
second reason is that, since threads are sharing the same answer trie structures,
answers inserted by a thread are automatically made available to all other threads
when they call the same subgoal. We observed that this collateral effect can also
lead to unexpected reductions on the execution time.
6 Conclusions
We have presented a new approach to multi-threaded tabled evaluation of logic
programs using a local evaluation strategy. In our proposal, each thread views its
tables as private but, at the engine level, the tables are shared among all threads.
The primary goal of our work was, in fact, to reduce the memory table space but,
our experimental results, showed that we can also significantly reduce the running
times. Since our implementation achieved very encouraging results on worst case
scenario tests, it should keep at least the same level of efficiency on any other tests.
Moreover, we believe that there is still considerable space for improvements, mainly
related to the low-level issues of Yap’s memory allocator for multi-threaded support.
The goal would be to implement strategies that pre-allocate bunches of memory in
order to minimize the performance degradation that the system suffers, when it is
exposed to simultaneous memory requests made by multiple threads. Further work
will also include extending the FS design to support batched evaluation.
Acknowledgments
This work is partially funded by the ERDF (European Regional Development Fund)
through the COMPETE Programme and by FCT (Portuguese Foundation for Science and Technology) within projects HORUS (PTDC/EIA-EIA/100897/2008) and
LEAP (PTDC/EIA-CCO/112158/2009). Miguel Areias is funded by the FCT grant
SFRH/BD/69673/2010.
4
We have experimental results confirming the memory usage formulas introduced on section 3
but, due to lack of space, we are not including them here.
Towards Multi-Threaded Local Tabling Using a Common Table Space
17
References
Chen, W. and Warren, D. S. 1996. Tabled Evaluation with Delaying for General Logic
Programs. Journal of the ACM 43, 1, 20–74.
Freire, J., Hu, R., Swift, T., and Warren, D. S. 1995. Exploiting Parallelism in
Tabled Evaluations. In International Symposium on Programming Languages: Implementations, Logics and Programs. Number 982 in LNCS. Springer-Verlag, 115–132.
Freire, J., Swift, T., and Warren, D. S. 1996. Beyond Depth-First: Improving Tabled
Logic Programs through Alternative Scheduling Strategies. In International Symposium
on Programming Language Implementation and Logic Programming. Number 1140 in
LNCS. Springer-Verlag, 243–258.
Liang, S., P.Fodor, Wan, H., and M.Kifer. 2009. OpenRuleBench: An Analysis of
the Performance of Rule Engines. In Internacional World Wide Web Conference. ACM
Press, 601–610.
Marques, R. 2007. Concurrent Tabling: Algorithms and Implementation. Ph.D. thesis,
Department of Computer Science, New University of Lisbon.
Marques, R. and Swift, T. 2008. Concurrent and Local Evaluation of Normal Programs.
In International Conference on Logic Programming. Number 5366 in LNCS. SpringerVerlag, 206–222.
Marques, R., Swift, T., and Cunha, J. C. 2010. A Simple and Efficient Implementation of Concurrent Local Tabling. In International Symposium on Practical Aspects of
Declarative Languages. Number 5937 in LNCS. Springer-Verlag, 264–278.
Moura, P. 2008. ISO/IEC DTR 13211–5:2007 Prolog Multi-threading Predicates. Available from http://logtalk.org/plstd/threads.pdf.
Ramakrishnan, I. V., Rao, P., Sagonas, K., Swift, T., and Warren, D. S. 1999.
Efficient Access Mechanisms for Tabled Logic Programs. Journal of Logic Programming 38, 1, 31–54.
Rocha, R., Silva, F., and Santos Costa, V. 2004. Concurrent Table Accesses in
Parallel Tabled Logic Programs. In International Euro-Par Conference. Number 3149
in LNCS. Springer-Verlag, 662–670.
Rocha, R., Silva, F., and Santos Costa, V. 2005. On applying or-parallelism and
tabling to logic programs. Theory and Practice of Logic Programming 5, 1 & 2, 161–
205.
Sagonas, K. and Swift, T. 1998. An Abstract Machine for Tabled Execution of FixedOrder Stratified Logic Programs. ACM Transactions on Programming Languages and
Systems 20, 3, 586–634.
Swift, T. and Warren, D. S. 2012. XSB: Extending Prolog with Tabled Logic Programming. Theory and Practice of Logic Programming 12, 1 & 2, 157–187.
Wielemaker, J. 2003. Native Preemptive Threads in SWI-Prolog. In International
Conference on Logic Programming. Number 2916 in LNCS. Springer-Verlag, 331–345.
| 6cs.PL
|
SAMUEL COMPACTIFICATIONS OF AUTOMORPHISM GROUPS
DANA BARTOŠOVÁ AND ANDY ZUCKER
arXiv:1802.02513v1 [math.DS] 7 Feb 2018
1. Introduction
In this paper, we are interested in the automorphism groups of countable first-order structures and the Samuel compactifications of these groups. We will address a variety of questions about the algebraic structure of the Samuel compactification and exhibit connections
between this algebraic structure and the combinatorics of the first-order structures at hand.
Let G be a topological group; all topological groups and spaces will be assumed Hausdorff.
The group G comes with a natural uniform structure, the left uniformity, whose entourages
are of the form {(g, h) ∈ G × G : g −1 h ∈ V } where V ranges over open symmetric neighborhoods of the identity. Every uniform space U admits a Samuel comactifiation, the Gelfand
space of the algebra of bounded uniformly continuous functions on U (see [Sa] or [U]). We
denote by S(G) the Samuel compactification of the group G with its left uniform structure.
In addition to being a compact Hausdorff space, the space S(G) can also be endowed with
algebraic structure. A G-flow is a compact Hausdorff space X equipped with a continuous
right G-action a : X × G → X. Typically the action a is understood, and we write x · g or
xg for a(x, g). We can give S(G) the structure of a G-flow; indeed, for each g ∈ G, the rightmultiplication map h → hg is left-uniformly continuous, so can be continuously extended
to S(G). With some extra work, it can be shown that the evaluation S(G) × G → S(G) is
continuous.
If X and Y are G-flows, a G-map is a continuous map ϕ : X → Y which respects the
G-action. A G-ambit is a pair (X, x0 ), where X is a G-flow and x0 ∈ X has a dense orbit. If
(X, x0 ) and (Y, y0 ) are ambits, then a map of ambits is a G-map ϕ : X → Y with ϕ(x0 ) = y0 .
Notice that there is at most one map of ambits from (X, x0 ) to (Y, y0 ). By identifying G as
embedded into S(G) and by considering the orbit of 1G , we turn (S(G), 1G ) into an ambit.
It turns out that this is the greatest ambit; for any G-ambit (X, x0 ), there is a map of ambits
ϕ : (S(G), 1G ) → (X, x0 ).
We can use this universal property to endow S(G) with yet more structure. A compact
left-topological semigroup is a semigroup S with a compact Hausdorff topology in which the
left multiplication maps t → st are continuous for each s ∈ S. Now let x ∈ S(G); then the
pair (x · G, x) is a G-ambit, so there is a unique G-map λx : S(G) → x · G with λx (1G ) = x.
We can endow S(G) with the structure of a compact left-topological semigroup by setting
xy := λx (y). It is not hard to show that this operation is associative.
Another consequence of the universal property is the existence of universal minimal flows.
Let X be a G-flow. A subflow is any closed Y ⊆ X which is invariant under the G-action.
The G-flow X is minimal if X 6= ∅ and every orbit is dense; equivalently, X is minimal if X
contains no proper subflows. An easy Zorn’s Lemma argument shows that every flow contains
a minimal subflow. The G-flow X is universal if there is a G-map from X onto any minimal
flow. Let M ⊆ S(G) be any minimal subflow, and let Y be a minimal flow. Pick y0 ∈ Y
arbitrarily, making (Y, y0 ) an ambit. Then there is a map of ambits ϕ : (S(G), 1G ) → (Y, y0 ),
1
2
D. BARTOŠOVÁ AND A. ZUCKER
and ϕ|M : M → Y is a G-map. We have just shown the existence of a universal minimal
flow. By using some techniques from the theory of compact left-topological semigroups,
it can be shown that there is a unique universal minimal flow up to G-flow isomorphism,
denoted M (G).
The existence and uniqueness of the universal minimal flow suggests another “canonical”
G-ambit we can construct. If X is a G-flow, we can view each g ∈ G as the function
ρg : X → X. Form the product space X X , and set E(X) = {ρg : g ∈ G}. It will be useful in
this instance to write functions on the right, so if f ∈ X X , we write x·f or xf instead of f (x).
The group G acts on E(X) via x · (f · g) = (x · f ) · g. Notice that ρg · h = ρgh , so (E(X), ρ1G )
is an ambit. We can also give E(X) a compact left-topological semigroup structure; rather
than a universal property, it is the fact that members of E(X) are functions that allows us
to do this. Indeed, the product is given by composition, which with our notation means that
for f1 , f2 ∈ E(X), we define x · (f1 · f2 ) = (x · f1 ) · f2 . The ambit E(X) (the distinguished
point being understood) is called the enveloping semigroup of X; we will be particularly
interested in E(M (G)), the enveloping semigroup of the universal minimal flow.
It is worth pointing out that we could have avoided some of this notational awkwardness by
switching the roles of left and right throughout, i.e. working with left G-actions and compact
right-topological semigroups. The reason we work with our left-right conventions is due to
the specific groups that we will be working with, i.e. automorphism groups of countable
first-order structures. We will point out later how a left-right switch could be made. Also
note that several of the references use the opposite left-right conventions, in particular [HS]
and [Ba].
Robert Ellis (see [E]) first proved the existence and uniqueness of M (G), and was the first
to consider the two canonical ambits S(G) and E(M (G)). As S(G) is the greatest ambit,
there is a map of ambits ϕ : S(G) → E(M (G)) (when referring to S(G) and enveloping semigroups, we will suppress the distinguished point unless there is possible confusion). He posed
the following very natural question: is ϕ : S(G) → E(M (G)) an isomorphism? Vladimir
Pestov (see [P]) observed that the existence of extremely amenable groups, groups where
M (G) is a singleton, provides a negative answer to Ellis’s question. Pestov also constructed
many other examples of groups G where S(G) and E(M (G)) were not isomorphic. The
diversity of counterexamples to Ellis’s question led Pestov to make the following conjecture.
Conjecture 1.1 (Pestov). Let G be a topological group. Then the canonical map ϕ : S(G) →
E(M (G)) is an isomorphism iff G is precompact.
Here, G is said to be precompact if the completion of its left uniformity is compact. If this
is the case, then all of S(G), M (G), and E(M (G)) are isomorphic to the left completion.
Aside from the initial work of Pestov, very little work has been done on Conjecture 1.1.
Glasner and Weiss (see [GW1]) have shown that S(Z) and E(M (Z)) are not isomorphic.
Their proof is rather difficult and uses some deep results in ergodic theory (see [F]); as such,
their methods are unlikely to generalize even to countable amenable groups.
In this paper, we address Conjecture 1.1 for groups of the form G = Aut(K) where K is
a countable first-order structure. We endow G with the topology of pointwise convergence,
turning G into a Polish group. In a mild abuse of terminology, we will call groups of this
form automorphism groups. When K is a countable set with no additional structure, we
have Aut(K) = S∞ , the group of all permutations of a countable set. More generally, automorphism groups are exactly the closed subgroups of S∞ . The work of Kechris, Pestov, and
Todorcevic [KPT] provides explicit computations of M (G) for many automorphism groups.
SAMUEL COMPACTIFICATIONS
3
Having an explicit representation of M (G) aids in analyzing the properties of E(M (G)).
Along with an explicit representation of S(G) for automorphism groups (see [Z]), this allows us to address Conjecture 1.1 for some of these groups. Our first main theorem is the
following.
Theorem 1.2. Let K be any of the following:
• a countable set without structure,
• the random Kn -free graph,
• the random r-uniform hypergraph.
Then for G = Aut(K), we have S(G) ∼
6= E(M (G)).
We then turn to finding the extent to which S(G) and E(M (G)) differ. Any minimal
subflow M ⊆ S(G) is isomorphic to M (G), and it turns out that S(G) admits a retraction
onto M , i.e. a G-map ϕ : S(G) → M with ϕ|M the identity. Pestov has shown (see [P1]) that
S(G) ∼
= E(M (G)) iff the retractions of S(G) onto a minimal subflow M ⊆ S(G) separate
the points of S(G). So if S(G) ∼
6= E(M (G)), it makes sense to ask which pairs of points
cannot be separated; this will not depend on the choice of minimal subflow M ⊆ S(G).
Given x, y ∈ S(G), we say they can be separated by retractions if there is a retraction
ϕ : S(G) → M with ϕ(x) 6= ϕ(y).
Every compact left-topological semigroup S admits a smallest two-sided ideal, denoted
K(S). Our second main theorem is the following.
Theorem 1.3. There are x 6= y ∈ K(S(S∞ )) which cannot be separated by retractions.
On the way to proving Theorem 1.3, we prove some theorems of independent interest
both for general topological groups G and for S∞ . By a well-known theorem of Ellis, every
compact left-topological semigroup S contains an idempotent, an element u ∈ S which
satisfies u · u = u (see [E]). Given Y ⊆ S, write J(Y ) for the set of idempotents in Y . Our
route to proving Theorem 1.3 involves a careful understanding of when the product of two
idempotents is or is not an idempotent.
In the case G = S∞ , we are able to find large semigroups of idempotents; this is what
allows us to prove Theorem 1.3.
Theorem 1.4. There are two minimal subflows M 6= N ⊆ S(S∞ ) so that J(M ) ∪ J(N ) is
a semigroup.
It is worth noting that any minimal subflow M ⊆ S(G) is a compact subsemigroup of
S(G), so J(M ) 6= ∅.
There are some cases when it is clear that K(S(G)) contains sufficiently large semigroups
of idempotents. Given a G-flow X, recall that a pair of points x, y ∈ X is called proximal
if there is p ∈ E(X) with xp = yp; the pair (x, y) is called distal if it is not proximal. A
G-flow X is proximal if every pair from X is proximal, and X is called distal if every pair
x 6= y ∈ X is distal. If M (G) is proximal, then whenever M ⊆ S(G) is a minimal subflow, we
have J(M ) = M . If M (G) is distal and M ⊆ S(G) is a minimal subflow, then J(M ) = {u},
a single idempotent. So long as S(G) contains at least two minimal right ideals, which is
always the case when G is Polish (see [Ba]), then E(M (G)) ∼
6= S(G) in these cases. We will
discuss examples of groups with M (G) proximal and provide a partial characterization of
Polish groups G with M (G) distal.
4
D. BARTOŠOVÁ AND A. ZUCKER
Theorem 1.5. Let G be a Polish group. The following are equivalent.
(1) M (G) is distal and metrizable.
i
ϕ
(2) There is a short exact sequence of groups 1 → H →
− G−
→ K → 1 with H extremely
amenable and K compact metrizable.
Furthermore, in item (2), K ∼
= M (G), and the G-action is given by k · g = kϕ(g).
2. Countable first-order structures and the Samuel compactification
In this section, we provide the necessary background on countable structures and provide
an explicit construction of the Samuel compactification of an automorphism group. The
presentation here is largely taken from [Z1].
Recall that S∞ is the group of all permutations of ω := {0, 1, 2, ...}. We can endow S∞
with the topology of pointwise convergence; a typical basic open neighborhood of the identity
is {g ∈ S∞ : g(k) = k for every k < n} for some n < ω. Notice that each of these basic open
neighborhoods is in fact a clopen subgroup.
Fix now G a closed subgroup of S∞ . A convenient way to describe the G-orbits of finite
tuples from ω is given by the notions of a Fraı̈ssé class and structure. A relational language
L = {Ri : i ∈ I} is a collection of relation symbols. Each relation symbol Ri has an arity
ni ∈ N. An L-structure A = hA, RiA i consists of a set A and relations RiA ⊆ Ani ; we say
that A is an L-structure on A. If A, B are L-structures, then g : A → B is an embedding
if g is a map from A to B such that RiA (x1 , ..., xni ) ⇔ RiB (g(x1 ), ..., g(xni )) for all relations.
We write Emb(A, B) for the set of embeddings from A to B. We say that B embeds A
and write A ≤ B if Emb(A, B) 6= ∅. An isomorphism is a bijective embedding, and an
automorphism is an isomorphism between a structure and itself. If A ⊆ B, then we say that
A is a substructure of B, written A ⊆ B, if the inclusion map is an embedding. A is finite,
countable, etc. if A is.
Definition 2.1. Let L be a relational language. A Fraı̈ssé class K is a class of L-structures
with the following four properties.
(1) K contains only finite structures, contains structures of arbitrarily large finite cardinality, and is closed under isomorphism.
(2) K has the Hereditary Property (HP): if B ∈ K and A ⊆ B, then A ∈ K.
(3) K has the Joint Embedding Property (JEP): if A, B ∈ K, then there is C which
embeds both A and B.
(4) K has the Amalgamation Property (AP): if A, B, C ∈ K and f : A → B and
g : A → C are embeddings, there is D ∈ K and embeddings r : B → D and
s : C → D with r ◦ f = s ◦ g.
If K is a countably infinite L-structure (which we will typically assume has underlying set
ω), we write Age(K) for the class of finite L-structures which embed into K. The following
is the major fact about Fraı̈ssé classes.
SAMUEL COMPACTIFICATIONS
5
Fact 2.2. If K is a Fraı̈ssé class, there is up to isomorphism a unique countably infinte
L-structure K with Age(K) = K satisfying one of the following two equivalent conditions.
(1) K is ultrahomogeneous: if f : A → B is an isomorphism between finite substructures
of K, then there is an automorphism of K extending f .
(2) K satisfies the Extension Property: if B ∈ K, A ⊆ B, and f : A → K is an
embedding, there is an embedding h : B → K extending f .
Conversely, if K is a countably infinite L-structure satisfying 1 or 2, then Age(K) is a
Fraı̈ssé class.
Given a Fraı̈ssé class K, we write Flim(K), the Fraı̈ssé limit of K, for the unique structure
K as above. We say that K is a Fraı̈ssé structure if K ∼
= Flim(K) for some Fraı̈ssé class.
Our interest in Fraı̈ssé structures stems from the following result.
Fact 2.3. For any Fraı̈ssé structure K, Aut(K) is isomorphic to a closed subgroup of S∞ .
Conversely, any closed subgroup of S∞ is isomorphic to Aut(K) for some Fraı̈ssé structure
K.
Fix S
a Fraı̈ssé class K with Fraı̈ssé limit K. Set G = Aut(K). We also fix an exhaustion
K = n AnS
, with each An ∈ K, |An | = n, and Am ⊆ An for m ≤ n. Whenever we
write K = n An , it will be assumed that the right side is an exhaustion of K. Write
Hn = {gGn : g ∈ G}, where Gn = G ∩ NAn is the pointwise stabilizer of An . We can
identify Hn with Emb(An , K),Sthe set of embeddings of An into K. Note that under this
identification, we have Hn = N ≥n Emb(An , AN ). For g ∈ G, we often write g|n for gGn ,
and we write in for Gn . The group G acts on Hn on the left; if x ∈ Hn and g ∈ G, we have
g · x = g ◦ x. For m ≤ n, we let inm ∈ Emb(Am , An ) be the inclusion embedding.
Each f ∈ Emb(Am , An ) gives rise to a dual map fˆ : Hn → Hm given by fˆ(x) = x◦f . Note
that we must specify the range of f for the dual map to make sense, but this will usually be
clear from context.
Proposition 2.4.
(1) For f ∈ Emb(Am , An ), the dual map fˆ : Hn → Hm is surjective.
(2) For every f ∈ Emb(Am , An ), there is N ≥ n and h ∈ Emb(An , AN ) with h ◦ f = iN
m.
Proof. Item 1 is an immediate consequence of the extension property. For item 2, use ultrahomogeneity to find g ∈ G with g ◦ f = im . Let N ≥ n be large enough so that ran(g|n ) ⊆ AN ,
and set h = g|n .
We now proceed with an explicit construction of S(G). First, if X is a discrete space, we
let βX be the space of ultrafilters on X. We topologize βX by declaring a typical basic open
neighborhood to be of the form {p ∈ βX : A ∈ p}, where A ⊆ X. We view X as a subset of
βX by identifying x ∈ X with the ultrafilter {A ⊆ X : x ∈ A}. If Y is a compact Hausdorff
space and ϕ : X → Y is any map, there is a unique continuous extension ϕ̃ : βX → Y .
Now let f ∈ Emb(Am , An ). The dual map fˆ extends to a continuous map f˜ : βHn → βHm .
If p ∈ βHn and f ∈ Emb(Am , An ), we will sometimes write p · f for f˜(p). Form the inverse
6
D. BARTOŠOVÁ AND A. ZUCKER
limit lim βHn along the maps ı̃nm . We can identify G with a dense subspace of lim βHn by
←−
←−
associating to each g ∈ G the sequence of ultrafilters principal on g|n . The space lim βHn
←−
turns out to be the Samuel compactification S(G) (see Corollary 3.3 in [P]).
To see that S(G) is the greatest ambit, we need to exhibit a right G-action on S(G).
This might seem unnatural at first; after all, the left G-action on each Hn extends to a left
G-action on βHn , giving us a left G-action on S(G). The problem is that the left action is
not continuous when G is given its Polish topology. The right action we describe doesn’t
“live” on any one level of the inverse limit lim βHn ; we need to understand how the various
←−
levels interact.
Let πn : lim βHn → βHn be the projection map. We often write α(n) := πn (α). For
←−
α ∈ lim βHn , g ∈ G, m ∈ N, and S ⊆ Hm , we have
←−
S ∈ αg(m) ⇔ {x ∈ Hn : x ◦ g|m ∈ S} ∈ α(n)
where n ≥ m is large enough so that ran(g|m ) ⊆ An . Notice that if g|m = h|m = f ,
then αg(m) = αh(m) := α · f := λαm (f ). By distinguishing the point 1 ∈ lim βHn with
←−
1(m) principal on im , we endow S(G) with the structure of a G-ambit, and (S(G), 1) is the
greatest ambit (see Theorem 6.3 in [Z]).
Using the universal property of the greatest ambit, we can define a left-topological semigroup structure on S(G): Given α and γ in lim βHn , m ∈ N, and S ⊆ Hm , we have
←−
S ∈ αγ(m) ⇔ {f ∈ Hm : S ∈ α · f } ∈ γ(m).
If α ∈ S(G) and S ⊆ Hm , a useful shorthand is to put
α−1 (S) = {f ∈ Hm : S ∈ α · f }.
Then the semigroup multiplication can be written as
S ∈ αγ(m) ⇔ α−1 (S) ∈ γ(m).
Notice that for fixed α, αγ(m) depends only on γ(m); indeed, if α ∈ lim βHn , p ∈ βHm , and
←−
S ⊆ Hm , we have S ∈ α · p iff α−1 (S) ∈ p. In fact, α · p = λ̃αm (p), where the map λ̃αm is the
continuous extention of λαm to βHm .
As promised in the introduction, we now explain the reason behind our left-right conventions. The primary reason behind considering right G-flows is because for G = Aut(K), the
left uniformity is very natural to describe. Namely, every entourage contains an entourage
of the form {(g, h) ∈ G × G : g|m = h|m }. This leads naturally to considering the embeddings from Am to K. If we wanted to consider the right uniformity, we would instead be
considering partial isomorphisms of K with range Am , which are less easily described.
3. KPT correspondence
In this section, we provide a brief review of KPT correspondence. For proofs of the results
in this section, see [KPT], [NVT], or [Z].
Let L be a relational language and L∗ = L ∪ S, where S = {Si : i ∈ N} and the Si
are new relational symbols of arity ni . If A is an L∗ -structure, write A|L for the structure
SAMUEL COMPACTIFICATIONS
7
obtained by throwing away the interpretations of the relational symbols in L∗ \ L. If K∗
is a class of L∗ -structures, set K∗ |L = {A∗ |L : A∗ ∈ K∗ }. If K = K∗ |L and K∗ is closed
under isomorphism, we say that K∗ is an expansion of K. If A∗ ∈ K∗ and A∗ |L = A, then
we say that A∗ is an expansion of A, and we write K∗ (A) for the set of expansions of A
in K∗ . If f ∈ Emb(A, B) and B∗ ∈ K∗ (B), we let B∗ · f be the unique expansion of A
so that f ∈ Emb(B∗ · f, B∗ ). The expansion K∗ is precompact if for each A ∈ K, the set
{A∗ ∈ K∗ : A∗ |L = A} is finite.
If K∗ is an expansion of the Fraı̈ssé class K, we say that the pair (K∗ , K) is reasonable if for
any A, B ∈ K, embedding f : A → B, and expansion A∗ of A, then there is an expansion
B∗ of B with f : A∗ → B∗ an embedding. When K∗ is also a Fraı̈ssé class, we have the
following equivalent definition.
Proposition 3.1. Let K∗ be a Fraı̈ssé expansion class of the Fraı̈ssé class K with Fraı̈ssé
limits K∗ , K respectively. Then the pair (K∗ , K) is reasonable iff K∗ |L ∼
= K.
Set Fin(K) = {A ∈ K : A ⊆ K}. Suppose (K∗ , K) is reasonable and precompact. Set
~ : hA, S|
~ A i ∈ K∗ whenever A ∈ Fin(K)}.
XK∗ := {hK, Si
We topologize this space by declaring the basic open neighborhoods to be of the form
N (A∗ ) := {K0 ∈ XK∗ : A∗ ⊆ K0 }, where A∗ is an expansion of some A ∈ Fin(K). We
can view XK∗ as a closed subspace of
Y
{A∗ : A∗ ∈ K∗ (A)}.
A∈Fin(K)
Notice that since (K∗ , K) is precompact, XK∗ is compact. If
compatible metric is given by
S
n
An = K is an exhaustion, a
~ hK, T~ i = 1/k(S,
~ T~ ),
d(hK, Si,
~ T~ ) is the largest k for which hAk , S|
~A i∼
where k(S,
= hAk , T~ |Ak i.
k
We can now form the (right) logic action of G = Aut(K) on XK∗ by setting K0 · g to be
the structure where for each relation symbol S ∈ S, we have
0
0
S (K ·g) (x1 , ..., xn ) ⇔ S K (g(x1 ), ..., g(xn )).
This action is jointly continuous, turning XK∗ into a G-flow. For readers used to left logic
actions, acting on the right by g is the same as acting on the left by g −1 .
First let us consider when XK∗ is a minimal G-flow.
Definition 3.2. We say that the pair (K∗ , K) has the Expansion Property (ExpP) when for
any A∗ ∈ K∗ , there is B ∈ K such that for any expansion B∗ of B, there is an embedding
f : A∗ → B ∗ .
Proposition 3.3. Let K∗ be a reasonable, precompact Fraı̈ssé expansion class of the Fraı̈ssé
class K with Fraı̈ssé limits K∗ , K respectively. Let G = Aut(K). Then the G-flow XK∗ is
minimal iff the pair (K∗ , K) has the ExpP.
8
D. BARTOŠOVÁ AND A. ZUCKER
Expansion classes are particularly inetesting when K∗ has the following combinatorial
property.
Definition 3.4. Let C be a class of finite structures.
(1) We say that A ∈ C is a Ramsey object if for any r ≥ 2 and any B ∈ C with A ≤ B,
there is C ∈ C with B ≤ C so that for any coloring c : Emb(A, C) → r, there is
h ∈ Emb(B, C) with |c(h ◦ Emb(A, B))| = 1.
(2) We say that C has the Ramsey Property (RP) if every A ∈ C is a Ramsey object.
The following is one of the major theorems in [KPT]. This theorem in its full generality
is proven in [NVT].
Theorem 3.5. Let K∗ be a reasonable, precompact Fraı̈ssé expansion class of the Fraı̈ssé
class K with Fraı̈ssé limits K∗ , K, respectively. Let G = Aut(K). Then XK∗ ∼
= M (G) iff the
pair (K∗ , K) has the ExpP and K∗ has the RP.
Pairs (K∗ , K) of Fraı̈ssé classes which are reasonable, precompact, satisfy the ExpP, and
where K∗ has the RP are called excellent. In particular, if K = Flim(K), G = Aut(K), and
there is an expansion class K∗ so that (K∗ , K) is excellent, then M (G) is metrizable. The
following converse is one of the major theorems of [Z].
Theorem 3.6. Let K be a Fraı̈ssé class with K = Flim(K) and G = Aut(K). If M (G) is
metrizable, then there is an expansion class K∗ so that (K∗ , K) is excellent.
4. Ellis’s problem for random relational structures
In S
this section, we prove Theorem 1.2. Let G = Aut(K) for some Fraı̈ssé structure
K = n An . If T ⊆ Hm , n ≥ m, and f ∈ Emb(Am , An ), set
f (T ) = {s ∈ Hn : s ◦ inm ∈ T }.
This is a minor abuse of notation, since f also denotes a map from Am to An ; however, this
map induces a continuous embedding of the Boolean algebra P(Hm ) into P(Hn ).
We will freely identify P(Hm ) with 2Hm ; in particular, G acts on 2Hm by right shift, where
for ϕ ∈ 2Hm , f ∈ Hm , and g ∈ G, we have ϕ · g(f ) = ϕ(g · f ).
Definition 4.1. We call a subset S ⊆ Hm minimal if the flow χS · G ⊆ 2Hm is minimal.
The formulation of Ellis’s problem we will work with is the one concerning retractions
given by Pestov. We will be interested in whether every pair x 6= y ∈ S(G) can be separated
by retractions. A characterization of when this occurs for discrete groups can be found in
[Ba] (see Proposition 11). We first prove a similar characterization for automorphism groups
in the next two lemmas.
SAMUEL COMPACTIFICATIONS
9
Before proceeding, a quick remark on notation is in order. If X is a G-flow, then there
is a unique map of ambits ϕ : S(G) → E(X). If x ∈ X and p ∈ S(G), we write x · p for
x · ϕ(p).
Lemma 4.2. Suppose α, γ ∈ S(G) cannot be separated by retractions, and let S ⊆ Hm be
minimal. Then S ∈ α(m) ⇔ S ∈ γ(m).
Proof. Let M ⊆ S(G) be a minimal subflow, and consider the non-empty closed subsemigroup {p ∈ M : χS · p = χS }. By Ellis’s theorem, let u ∈ M be an idempotent with
χS · u = χS . As the left multiplication λu : S(G) → M is a retraction, we must have
u · α = u · γ. Therefore χS · α = χS · γ, and α−1 (S) = γ −1 (S) := T . It follows that im ∈ T
iff S ∈ α(m) iff S ∈ γ(m).
For each m < ω, let Bm ⊆ P(Hm ) be the Boolean algebra generated by the minimal
0
subsets of Hm . Let Bm
be the Boolean algebra {T ⊆ Hm : ∃n ≥ m(inm (T ) ∈ Bn )}.
Lemma 4.3. Fix M ⊆ S(G) a minimal. The following are equivalent.
(1) Retractions of S(G) onto M separate points of S(G),
0
= P(Hm ).
(2) For every m < ω, we have Bm
Proof. Suppose that there are α 6= γ ∈ S(G) which cannot be separated by retractions. Find
m < ω with α(m) 6= γ(m), and find T ⊆ Hm with T ∈ α(m), T 6∈ γ(m). Note that for
every n ≥ m, we have inm (T ) ∈ α(n) and inm (T ) 6∈ γ(n). Towards a contradiction, suppose
for some n ≥ m that inm (T ) was a Boolean combination of minimal set A1 , ..., Ak ⊆ Hn .
By our assumption, α(n) and γ(n) agree on the membership of each Ai , hence also on the
membership of T , a contradiction.
0
0
) denote
. Let S(Bm
Conversely, suppose that T ⊆ Hm is not in the Boolean algebra Bm
0
0
0
the Stone space of Bm ; form the inverse limit lim S(Bn ), and let p ∈ lim S(Bn ) be chosen
←−
←−
so that p(m) doesn’t decide T . Then find α, γ ∈ lim βHn with T ∈ α(m), T 6∈ γ(m) which
←−
both extend p. Then α and γ cannot be separated by retractions.
Notice that item (2) of Lemma 4.3 does not depend on M . In general, the relation of
whether x 6= y ∈ S(G) can be separated by retractions does not depend on the minimal
subflow of S(G) chosen, but we postpone this discussion until the end of section 6 (see the
discussion after Theorem 5.11).
Now suppose that (K∗ , K) is an excellent pair of Fraı̈ssé classes. Given a set of expansions
E ⊆ K∗ (Am ) and K0 ∈ XK∗ , let Hm (E, K0 ) = {f ∈ Hm : Am (f, K0 ) ∈ E}. If n ≥ m and
A0n ∈ K∗ (An ), we set Emb(E, A0n ) = {f ∈ Emb(Am , An ) : Am (f, A0n ) ∈ E}.
Proposition 4.4. S ⊆ Hm is minimal iff there is K0 ∈ XK∗ and E ⊆ K∗ (Am ) so that
S = Hm (E, K0 )
Proof. One direction is easy once we note that given E ⊆ K∗ (Am ), the map ϕE : XK∗ → 2Hm
given by ϕE (K0 ) = Hm (E, K0 ) is a map of G-flows. In the other direction, let S ⊆ Hm be
minimal. Then Y := χS · G ⊆ 2Hm is a minimal G-flow, so fix a G-map ϕ : XK∗ → Y . Note
that ϕ(K∗ ) must be a G∗ -fixed point, so by ultrahomogeneity of K∗ it must be of the form
10
D. BARTOŠOVÁ AND A. ZUCKER
Hm (E, K∗ ) for some E ⊆ K∗ (Am ). It follows that ϕ = ϕE , so in particular S = Hm (E, K0 )
for some K0 ∈ XK∗ .
The main tool allowing us to prove Theorem 1.2 is an explicit characterization of M (G)
for certain autormorphism groups G. The following facts can be found in [KPT].
Fact 4.5. Let K = Age(K), where K is any of the structures in the statement of Theorem
1.2. Let K∗ be the class of linearly ordered members of K. Then (K∗ , K) is an excellent pair.
Setting G = Aut(K), then M (G) = XK∗ is the space of linear orders of K.
The next theorem is the simplest case of Theorem 1.2. The following notion will be useful
in the proof. Given T ⊆ Hm and N ≥ m, an N -pattern of T is a set S ⊆ Emb(Am , AN ) so
that there is y ∈ HN with S = {f ∈ Emb(Am , AN ) : y ◦ √
f ∈ T }. We Will be using Stirling’s
formula which states that asymptotically n! is equal to 2πn( ne )n .
Theorem 4.6. Let K = {xn : n < ω} be a countable set with no structure (so G ∼
= S∞ ),
0
⊆ 2Hm is meager. In particular, any
and set Am = {xi : i < m}. Then for every m ≥ 2, Bm
0
.
T ∈ Hm whose orbit is dense is not in Bm
Proof. Let T ⊆ Hm have dense orbit. So for every N ≥ m, every S ⊆ Emb(Am , AN ) is an
N -pattern of T . Towards a contradiction, suppose for some n ≥ m that inm (T ) was a Boolean
combination of minimal sets B1 , ..., Bk ⊆ Hn . Let N >> n; we will obtain a contradiction
N
m
by counting the number of N -patterns in inm (T ), which by assumption is 2m!(m) ≥ 2N /2 .
Since |K∗ (An )| = n! and since there are N ! linear orders on AN , this gives us 2n! N !
possible N -patterns for each Bi by Proposition 4.4. Therefore any N -pattern of T must be
a Boolean combination of some k of these N -patterns. Each choice of k patterns results
k
in at most 22 Boolean combinations, so the total number of possible patterns is at most
k
22 (2n! N !)k patterns. Noting that n and k remain fixed as we let N grow large, we have
that asymptotically there are CN · ekN possible N -patterns of inm (T ), which is far less than
m
2N /2 , a contradiction.
We now consider the case where K = Flim(K) is the random r-uniform hypergraph for
some r ≥ 2. In order to generalize the arguments
in the proof of Theorem 4.6, we will
S
need some control over the exhaustion K = n An . We will do this by not specifying an
exhaustion in advance, but instead determining parts of it as we proceed.
We will need the following notion. With K as above, let A ⊆ B ∈ K, and let C ⊆ D ∈ K.
We say that D extends C along A ⊆ B if for any f ∈ Emb(A, C), there is an h ∈ Emb(B, D)
with h|A = f .
Given C ∈ K, write |C| for the number of vertices in C.
Lemma 4.7. Let K be the class of r-uniform hypergraphs for some r ≥ 2. Let e ⊆ B ∈ K,
where e ∈ K is the hypergraph on r vertices consisting of an edge, and let C ∈ K with
|C| = N . Then there is D ∈ K extending C along e ⊆ B with |D| ≤ cN r−1 for some
constant c depending only on |B|.
SAMUEL COMPACTIFICATIONS
11
Proof. Recall that given an r-uniform hypergraph C, a matching is a subset of the edges
of C so that each vertex is included in at most one edge. By Baranyai’s theorem [B],
the edge set
into M1 , ..., M` with each Mi a matching and with
of C can be partitioned
rdN/re
r−1
`≤
/dN/re) ≈ c0 N . For each i ≤ ` and j ≤ r!, let Dij be a set of |B| − r new
r
S S
vertices. We will define the hypergraph D on vertex set C ∪ i≤` j≤r! Dij . First add edges
to D so that Dij ∼
= B \ e. For each e0 ∈ Mi , enumerate the embeddings fj : e → C with
0
range e . Add edges to D so that each fj extends to an embedding hj : B → D with range
e0 ∪ Dij . This is possible since each Mi is a matching. The hypergraph D has |D| ≤ cN r−1
as desired.
Theorem 4.8. Let K be the class of r-uniform hypergraphs for r ≥ 2, with K = Flim(K).
Let Ar ⊆ K be an edge on r vertices. Then if T ⊆ Hr has dense orbit, then T 6∈ Br0 .
Remark. Though we are not specifying an exhaustion in advance, we will still use some of the
associated notation. In particular, when we write Am for some m < ω, we mean a subgraph
of K on m vertices.
Proof. Suppose towards a contradiction that there were some graph An ⊇ Ar so that inr (T )
was a Boolean combination of minimal sets B1 , ..., Bk ⊆ Hn . Let N n, and fix a graph
AN ⊇ An with at least Nr /2 edges. Let AN 0 ⊇ AN extend AN along Ar ⊆ An with
N 0 ≈ cN r−1 as guaranteed by Lemma 4.7. We will obtain a contradiction by counting the
number of N 0 -patterns of inr (T ). Exactly as in the proof of Theorem 4.6, there are at most
k
0
r−1
22 (2n! N 0 !)k ≈ CN 0 · ekN ≈ CN r−1 · ekN
many N 0 -patterns. But since T has dense orbit,
N
r
there must be at least 2( r )/2 ≈ 2cN N 0 -patterns of inr (T ), a contradiction.
We next turn to the class K of Kr -free graphs for some r ≥ 3. We will need a result similar
to Lemma 4.7, but the given proof will not work as the construction doesn’t preserve being
Kr -free. Recall that the Ramsey number R(r, n) of r and n is the smallest integer such that
any graph on R(r, n) vertices either contains a clique of size r or an independent set of size
n.
Lemma 4.9. Let K be the class of Kr -free graphs for some r ≥ 3. Let e ⊆ B ∈ K, where
e is an edge, and let C ∈ K with |C| = N . Then there is D ∈ K extending C along e ⊆ B
with |D| ≤ cN 2(r−1)/(r−2) .
Proof. Let R(r, n) be the Ramsey number of r and n. In [AKS], it is shown that R(r, n) =
o(nr−1 ). Since C is Kr -free, this implies that C has an independent set of size at least
N 1/(r−1) ; by repeatedly removing independent sets, we see that the chromatic number of C
is at most ` ≈ N (r−2)/(r−1) . Write C = C1 t · · · t C` so that each Ci is an independent set.
For every ordered pair (i, j) of distinct indices with i, j ≤S`, let D(i,j) be a set of |B| − 2 new
vertices. We will define the graph D on vertex set C ∪ {i,j}∈[`]2 (D(i,j) ∪ D(j,i) ). First add
edges to D so that D(i,j) ∼
= B \ e; fix h0 : B \ e → D(i,j) an isomorphism. Write e = {a, b}; if
f : e → C with f (a) = i and f (b) = j, then add edges to D so that h0 ∪ f := h : B → D is an
embedding with range f (e)∪D(i,j) . The hypergraph D is Kr -free and has |D| ≤ cN 2(r−2)/(r−1)
as desired.
12
D. BARTOŠOVÁ AND A. ZUCKER
Theorem 4.10. Let K be the class of Kr -free graphs for some r ≥ 3, with K = Flim(K).
Let A2 ⊆ K be an edge. Then if T ⊆ H2 has dense orbit, then T 6∈ B20 .
As in the proof of Theorem 4.8, we will not specify an exhaustion in advance, but we will
still use some of the notational conventions.
Proof. Suppose towards a contradiction that there were some graph An ⊇ A2 so that in2 (T )
was a Boolean combination of minimal sets B1 , ..., Bk ⊆ Hn . Let N n, and fix a graph
AN ⊇ An with at least n2 /r edges. Let AN 0 ⊇ AN extend AN along A2 ⊆ An with
N 0 ≈ cN 2(r−2)/(r−1) as guaranteed by Lemma 4.9. We now obtain a contradiction by counting
0
2(r−1)/(r−2)
N 0 -patterns. Once again, there are at most CN 0 · ekN ≈ CN 2(r−2)/(r−1) · ekN
many
N
/r
0
n
cN 2
(
)
2
N -patterns in i2 (T ), which contradicts the fact that there are at least 2
≈2
many
N 0 -patterns.
We end this section with a conjecture. While it is a strict sub-conjecture of Conjecture
1.1, we think it might be more easily approached.
Conjecture 4.11. Let G be a closed, non-compact subgroup of S∞ with metrizable universal
minimal flow. Then S(G) 6∼
= E(M (G)).
5. A closer look at S∞
In this section, we take a closer look at S(S∞ ), with an eye towards understanding which
pairs of points x 6= y ∈ S(S∞ ) can be separated by retractions. We view S∞ as the group of
permutations of ω. We can realize ω as a Fraı̈ssé structure in the empty language. We set
An = n, so that Hn is the set of all injections from n into ω, and for m ≤ n, Emb(Am , An )
is the set of all injections from m into n. We will often abuse notation and write s ∈ Hm as
the tuple (s0 , ..., sm−1 ), where si = s(i).
We start by developing some notions for any automorphism group. Let f ∈ Emb(Am , An ).
If F ⊆ P(Hm ) is a filter, then we write f (F) for the filter generated by {f (T ) : T ∈ F}.
If H ⊆ P(Hn ) is a filter, then f˜(H) is the push-forward filter {T ⊆ Hm : f (T ) ∈ H}. This
may seem like a conflict of notation since f˜ : βHn → βHm is the extended dual map of f .
We canTjustify this notation as follows. To each filter H on Hn , we associate the closed set
XH := A∈H A ⊆ βHn . Conversely, given a closed set X ⊆ βHn , we can form the filter of
clopen neighborhoods FX := {A ⊆ Hn : X ⊆ A}. Then we obtain the identity
Xf˜(H) = f˜(XH ).
A similar identity holds given a filter F on Hm :
Xf (F ) = f˜−1 (XF ).
Let Y ⊆ S(G) be closed. Let πm : S(G) → βHm be the projection map. Then πm (Y )
Y
is a closed subset of βHm . Write Fm
for Fπm (Y ) . For n ≥ m, the filter FnY extends the
n
Y
n
Y
Y
filter im (Fm ) and ı̃m (Fn ) = Fm . Conversely, given filters Fm on Hm for every m < ω such
SAMUEL COMPACTIFICATIONS
13
that Fn extends inm (Fm ) and with ı̃nm (Fn ) = Fm , there is a unique closed Y ⊆ S(G) with
Y
Fm = Fm
for each m < ω. We will call such a sequence of filters compatible.
M
We will need to understand the filters Fm
when M ⊆ S(G) is a minimal subflow. It turns
out that these filters are characterized by a certain property of their members.
Definition 5.1. Given T ⊆ Hm , we say that T is thick if either of the following equivalent
items hold (see [Z1]).
(1) T ⊆ Hm is thick iff χHm ∈ χT · G.
(2) T ⊆ Hm is thick iff for every n ≥ m, there is s ∈ Hn with s ◦ Emb(Am , An ) ⊆ T .
We can now state the following fact from [Z1].
Theorem 5.2. Let G be an automorphism group, and let M ⊆ S(G) be closed. Then M is
M
a minimal subflow iff each Fm
is a maximal filter of thick sets.
Another observation is the following.
Y
and f ∈ Emb(Am , An ).
Proposition 5.3. Say Y ⊆ S(G) is a subflow, and let T ∈ Fm
Y
Then f (T ) ∈ Fn .
Proof. Pick g ∈ G with g|m = f . Then for any α ∈ S(G), we have T ∈ αg(m) iff f (T ) ∈ α(n).
As Y is G-invariant, the result follows.
We now turn our attention to G = S∞ . Let {σi : i < m!} T
list the permutations of m, i.e.
M
. Call S ⊆ Hm
the members of Emb(Am , Am ). Then by Proposition 5.3, i σi (T ) ∈ Fm
saturated if whenever (a0 , ..., am−1 ) ∈ S and σ is a permutation, then (aσ(0) , ..., aσ(m−1) ) ∈ S.
M
has a base of saturated sets.
We have just shown that Fm
Let us recall Ramsey’s theorem: for every k, `, r < ω with k ≤ `, there is m ≥ ` so that
for every coloring c : [m]k → r, there is A ⊆ m with |A| = ` and |c([A]k )| = 1. Ramsey’s
theorem will
consequences for the collection of thick subsets of Hm .
F
F have interesting
n
Let ϕ : n Hn → n [ω] be the order forgetful map, i.e. for (y0 , ..., ym−1 ) ∈ Hm , we set
ϕ(y0 , ..., ym−1 ) = {y0 , ..., ym−1 } ∈ [ω]m . Any filter F on Hm pushes forward to a filter ϕ(F)
on [ω]m . We can define a thick subset of [ω]m in a very similar fashion to a thick subset of
Hm ; more precisely, we say T ⊆ [ω]m is thick iff for every n ≥ m, there is s ∈ [ω]n with
[s]m ⊆ T . Call S ⊆ [ω]m thin if it is not thick. We now have the following crucial corollary
of Ramsey’s theorem: if T ⊆ [ω]m is thick and T = T0 ∪ · · · ∪ Tk , then some Ti is thick. In
particular, if H is a thick filter on [X]m , i.e. a filter containing only thick sets, then we can
extend H to a thick ultrafilter. It also follows that for every m < ω, the collection of thin
subsets of [ω]m forms an ideal.
M
Theorem 5.4. Let M ⊆ S(S∞ ) be a minimal right ideal. Then for every m < ω, ϕ(Fm
)
m
−1
is a thick ultrafilter. Conversely, if p ∈ β[ω] is a thick ultrafilter, then {ϕ (T ) : T ∈ p}
generates a maximal thick filter on Hm , hence there is a minimal M ⊆ S(S∞ ) with p =
M
ϕ(Fm
).
14
D. BARTOŠOVÁ AND A. ZUCKER
M
) is a thick filter. Towards a contradiction, suppose it is not an ultrafilProof. Clearly ϕ(Fm
M
M
ter, and extend it to a thick ultrafilter p ∈ β[ω]m . Let T ∈ p \ ϕ(Fm
). Then ϕ−1 (T ) 6∈ Fm
.
−1
M
However, ϕ (T ) ∩ S is thick for every saturated S ∈ Fm . As saturated sets form a base for
M
M
.
, this contradicts the maximality of Fm
Fm
m
Now let p ∈ β[ω] be a thick ultrafilter. Then F := {ϕ−1 (T ) : T ∈ p} generates a thick
filter. Suppose S ⊆ Hm and {S} ∪ F generated a thick filter strictly larger than F. We may
assume S is saturated. Then ϕ(S) ∈ p, so ϕ−1 (ϕ(S)) = S ∈ F, a contradiction.
Notice that if p ∈ β[ω]n is thick and m ≤ n, then there is a unique thick ultrafilter
q ∈ β[ω]m with the property that {a ∈ [ω]n : [a]m ⊆ S} ∈ p for every S ∈ q. Certainly
such a q must be unique. To see that this q exists, suppose [ω]m = S t T . Then the set
n
(p) for this q. If
{a ∈ [ω]n : [a]m ∩ S 6= ∅ and [a]m ∩ T 6= ∅} is not thick. We will write πm
M
n
M
M ⊆ S(G) is a minimal right ideal and p = ϕ(Fn ), then we have πm (p) = ϕ(Fm
).
Let LO(ω) be the space of linear orders on ω. Viewed as a subset of the right shift 2H2 ,
LO(ω) becomes an S∞ -flow. It is known (see [KPT] or [GW]) that LO(ω) ∼
= M (S∞ ). Indeed,
we saw in section 4 that if K is the class of finite sets and K∗ is the class of finite linear orders,
then (K∗ , K) is an excellent pair, and XK∗ ∼
= LO(ω). If M ⊆ S(S∞ ) is a minimal right ideal
and <∈ LO(ω), then the map λ : M → LO(ω) given by λ(α) = < ·α := limgi →α < ·gi is an
S∞ -flow isomorphism. We will often write <α for < ·α, and we will write > for the reverse
linear order of <.
If <0 , <1 ∈ LO(ω) and m ≥ 2, define the set
Am (<0 , <1 ) = {{a0 , ..., am−1 } ∈ [ω]m : ∀i, j < m(ai <0 aj ⇔ ai <1 aj )}
and define Bm (<0 , <1 ) = Am (<0 , >1 ). If s ∈ Hm , we say that <0 and <1 agree on s if
ϕ(s) ∈ Am (<0 , <1 ), and we say that they anti-agree on s if ϕ(s) ∈ Bm (<0 , <1 ). When
m = 2 we often omit the subscript. If M is a minimal right ideal, then ϕ(F2M ) contains
exactly one of A(<0 , <1 ) or B(<0 , <1 ). Let AM ⊆ LO(ω) × LO(ω) be defined AM = {(<0
, <1 ) : A(<0 , <1 ) ∈ ϕ(F2M )}. Then AM is certainly reflexive and symmetric. To see that AM
is an equivalence relation, note that A(<0 , <1 ) ∩ A(<1 , <2 ) ⊆ A(<0 , <2 ). Furthermore, AM
has exactly two equivalence classes; this is because B(<0 , <1 ) ∩ B(<1 , <2 ) ⊆ A(<0 , <2 ).
Lemma 5.5. Let M ⊆ S(S∞ ) be a minimal right ideal, and let (<0 , <1 ) ∈ AM . Then for
M
any m < ω, we have Am (<0 , <1 ) ∈ ϕ(Fm
).
Proof. Let {fi : i < k} enumerate Emb(A2 , Am ). Then
this is exactly the desired set.
T
i
M
fi (ϕ−1 (A(<0 , <1 ))) ∈ Fm
, and
Lemma 5.6. Let M ⊆ S(S∞ ) be a minimal right ideal, and let α ∈ M . Then the following
are equivalent:
(1) α is an idempotent,
(2) For any <∈ LO(ω), we have (<, <α ) ∈ AM ,
(3) There is <∈ LO(ω) with (<, <α ) ∈ AM .
SAMUEL COMPACTIFICATIONS
15
Proof. Suppose α ∈ M is an idempotent, and let <∈ LO(ω). Then considering i2 = (x0 , x1 ) ∈
Emb(A2 , A2 ), we have x0 <α x1 iff {f ∈ H2 : f0 < f1 } ∈ α(2). But since α is an idempotent,
this is equivalent to {f ∈ H2 : f0 <α f1 } ∈ α(2). But this implies that ϕ−1 (A(<, <α )) ∈ α(2),
implying that A(<, <α ) ∈ ϕ(F2M ).
Conversely, suppose α ∈ M and <∈ LO(ω) with (<, <α ) ∈ AM . If f ∈ Emb(A2 , An ), then
we have f0 <α f1 iff {s ∈ Hn : s(f0 ) < s(f1 )} ∈ α(n). By Lemma 5.5, we see that this is iff
{s ∈ Hn : s(f0 ) <α s(f1 )} ∈ α(n). It follows that < ·α · α =< ·α, so α is idempotent.
Theorem 5.7. Let M, N ⊆ S(S∞ ) be minimal right ideals. The following are equivalent.
(1) AM = AN ,
(2) If u ∈ M and v ∈ N are idempotents, then uv ∈ M is also idempotent.
Proof. Suppose AM 6= AN , with (<0 , <1 ) ∈ AN \ AM . Find u ∈ M with <0 ·u = <0 , and
find v ∈ N with <0 ·v = <1 . By Lemma 5.6, u and v are idempotents and uv is not an
idempotent.
Conversely, suppose u ∈ M and v ∈ N are idempotents with uv not idempotent. Find
<0 ∈ LO(ω) with <0 ·u = <0 , and let <1 = <0 ·v. Since v is idempotent, we have by Lemma
5.6 that (<0 , <1 ) ∈ AN ; but since uv is not idempotent, we have (<0 , <1 ) 6∈ AM .
It is easy to construct minimal right ideals M, N ⊆ S(S∞ ) with AM 6= AN . Let <0
m
, <1 ∈ LO(ω) be linear orders so that for every m < ω, there are sm = (sm
0 , ..., sm−1 ) ∈ Hm
m
m
m
m
and t = (t0 , ..., tm−1 ) ∈ Hm so that <0 and <1 agree on s and anti-agree on tm . Let
M ⊆ S(S∞ ) be a minimal subflow with ϕ−1 (A(<0 , <1 )) ∈ F2M , and let N ⊆ S(S∞ ) be a
minimal subflow with ϕ−1 (B(<0 , <1 )) ∈ F2N . Then (<0 , <1 ) ∈ AM \ AN .
We now turn our attention to constructing M 6= N ⊆ S(S∞ ) minimal right ideals with
AM = AN ; this will prove Theorem 1.4 as a corollary of Theorem 5.7. To this end, we will
construct two thick ultrafilters p 6= q ∈ β[ω]3 with π23 (p) = π23 (q), so that whenever M and
N are minimal subflows of S(S∞ ) with ϕ(F3M ) = p and ϕ(F3N ) = q, then ϕ(F2M ) = ϕ(F2N ).
In particular, this implies that AM = AN .
Recall that a selective ultrafilter is an ultrafilter p on ω with the property that for any
finite coloring c : [ω]2 → r, there is a p-large set A ⊆ ω which is monochromatic for c.
Another way of saying this is as follows. Given a set A ⊆ ω, set λA = [A]2 , and if F is a
filter on ω, let λF be the filter generated by {λA : A ∈ F}. Then the ultrafilter p is selective
iff λp is an ultrafilter. The existence of selective ultrafilters is independent of ZFC.
We will be considering the following generalizations of selective ultrafilters. Let m < ω.
If T ⊆ [ω]m , we set λT = {s ∈ [ω]m+1 : [s]m ⊆ T }. If n > m, we set λ(n−m) (T ) = {s ∈ [ω]n :
[s]m ⊆ T }. Notice that the λ(n−m) operation is the same as applying λ (n − m)-many times,
justifying this notation. If F is a filter on [ω]m , we let λ(n−m) F be the filter generated by
{λ(n−m) T : T ∈ F}. It can happen that for some T ∈ F we have λ(n−m) T = ∅. We will
usually be working under assumptions that prevent this from happening. For instance, if
T ⊆ [ω]m is thick, then λ(n−m) T 6= ∅ for every n > m. Even better, if F is a thick filter on
[ω]m , then λ(n−m) F is a thick filter on [ω]n .
16
D. BARTOŠOVÁ AND A. ZUCKER
Definition 5.8. Let p ∈ β[ω]m be a thick ultrafilter. We say that p is (m, n)-selective if
λ(n−m) p is an ultrafilter. We say that p is weakly (m, n)-selective if there is a unique thick
ultrafilter extending the filter λ(n−m) p.
If p ∈ β[ω]m is a thick ultrafilter and q ∈ β[ω]n is a thick ultrafilter extending λ(n−m) p,
n
then we have πm
(q) = p. Therefore to prove Theorem 1.4, it is enough to construct a
thick ultrafilter p ∈ β[ω]2 which is not weakly (2, 3)-selective. Indeed, if p ∈ β[ω]2 is not
weakly (2, 3)-selective, then there are thick ultrafilters q0 6= q1 both extending the filter λp,
so π23 (q0 ) = π23 (q1 ).
Our construction proceeds in two parts. First we define a certain type of pathological
subset of [ω]3 and show that its existence allows us to construct p ∈ β[ω]2 which is not
weakly (2, 3)-selective. Then we show the existence of such a pathological set.
We begin by developing some abstract notions. Let Y be a set, and let I be a proper
ideal on Y . Write S ⊆I T if S \ T ∈ I. Let ψ : P(Y ) → P(Y ) be a map satisfying ψ 2 = ψ,
S ⊆ ψ(S), S ⊆ T ⇒ ψ(S) ⊆ ψ(T ), and ψ(∅) = ∅. Call a set S ⊆ Y ψ-closed or just closed if
ψ(S) = S, and call S near-closed if there is a closed set T with S∆T ∈SI. Call a set S ⊆ Y
(< ℵ0 )-near-closed if there are k < ω and closed T0 , ..., Tk−1 with S∆( i<k Ti ) ∈ I. Notice
that a finite union of near-closed sets is (< ℵ0 )-near-closed.
Now suppose S ⊆ Y is a set which is not (< ℵ0 )-near-closed. If p, q ∈ βY , we say that p
ψ-intertwines q over S modulo I if the following three items all hold:
(1) {S \ T : T is near-closed and T ⊆I S} ⊆ p,
(2) {ψ(T ) \ S : T ∈ p, T ⊆ S} ⊆ q,
(3) p and q extend the filterdual of I.
If ψ, S, and I are understood, we will just say that p intertwines q. Notice in (1) that
if T is near-closed with T ⊆I S, then S ∩ T is also near-closed, so it is enough to consider
near-closed T with T ⊆ S.
Lemma 5.9. Fix S ⊆ Y which is not (< ℵ0 )-near-closed.
(1) If B ⊆ Y with B ∈ I, then B is near-closed. Hence S 6∈ I.
(2) There are p, q ∈ βY so that p intertwines q.
Proof. The first part follows since the empty set is closed.
Since S is not (< ℵ0 )-near-closed, we have that {S \ T : T near-closed and T ⊆I S}
generates a filter F extending the filterdual of I. Let p ∈ βY be any ultrafilter extending
F.
Now let T ∈ p. Then ψ(T ) \ S 6∈ I; otherwise we would have ψ(T ) ⊆I S, so S \ ψ(T ) ∈ p,
contradicting that T ∈ p. Also note by monotonicity of ψ that (ψ(T0 ) ∩ ψ(T1 )) \ S ⊇
ψ(T0 ∩ T1 ) \ S, so the collection {ψ(T ) \ S : T ∈ p} generates a filter H avoiding I; letting q
be any ultrafilter extending both H and the filterdual of I, we see that p intertwines q.
We now apply these ideas. Let Y = [ω]3 , and let I be the thin ideal. Given T ⊆ [ω]3 , view
T as a 3-uniform hypergraph, and form the shadow graph ∂T := {{a, b} ∈ [ω]2 : ∃c({a, b, c} ∈
SAMUEL COMPACTIFICATIONS
17
T )}. Define ψ(T ) = λ∂T . In words, ψ(T ) is the largest hypergraph with ∂ψ(T ) = ∂T . More
generally, we can set Y = [ω]n and let I be the ideal of subsets of [ω]n which are not thick.
If m < n and T ⊆ [ω]n , we set ∂ (n−m) T = {s ∈ [ω]m : ∃t ∈ [ω]n−m (s ∪ t ∈ T )}. Then we can
set ψ(T ) = λ(n−m) ∂ (n−m) T .
Theorem 5.10. Let Y = [ω]n , let I be the thin ideal, and let ψ = λ(n−m) ∂ (n−m) for some
m < n. Suppose S ⊆ [ω]n is not (< ℵ0 )-near-closed, and say p, q ∈ β[ω]n where p intertwines
n
n
q. Then πm
(p) = πm
(q).
n
n
Proof. Suppose towards a contradiction that p0 := πm
(p) 6= q 0 := πm
(q) as witnessed by
m
0
m
0
A ⊆ [ω] with A ∈ p , [ω] \ A ∈ q . Then setting B := {s ∈ [ω]n : [s]m ⊆ A} and
C := {s ∈ [ω]n : [s]m ⊆ [ω]m \ A}, we have B ∈ p, C ∈ q, and B ∩ C = ∅. Note that both B
and C are ψ-closed. Since p and q are intertwined, we have ψ(B ∩ S) \ S ∈ q, so in particular
B \ S ∈ q. But since C ∈ q, this is a contradiction.
The next theorem along with Theorems 5.10 and 5.7 will prove Theorem 1.4.
Theorem 5.11. With I and ψ as in Theorem 5.10, there is S ⊆ [ω]n which is not (< ℵ0 )near-closed.
Proof. The following elegant proof is due to Anton Bernshteyn.
We take S to be the random n-uniform hypergraph. Suppose towards a contradiction that
S was k-near-closed for some k < ω. We write S = S0 ∪ · · · ∪ Sk−1
S with each Si near-closed.
Let Ti ⊆ [ω]n be a ψ-closed set with Si ∆Ti ∈ I, and write T = i<k Ti . So S∆T ∈ I. This
means that there is some ` < ω so that the hypergraph S∆T contains no clique of size `.
We now compute an upper bound on the number of induced subgraphs of S that can
appear on N vertices V := {v0 , ..., vN −1 } ⊆ ω. Since S is the random n-uniform hypergraph,
N
there must be 2( n ) many possibilities. But by assumption, S = T ∆G, where G is some
hypergraph with no cliques of size `. Since an induced subgraph of a ψ-closed graph is
N
ψ-closed, each Ti |V is determined by ∂ (n−m) (Ti |V ), so in particular, there are at most 2(m)
N
many possibilities for each Ti |V , so at most 2k(m) possibilities for T |V . As for G, we need
an estimate on the number of `-free n-uniform hypergraphs on N vertices.
It is a fact that
N
for some constant c > 0 depending only on ` and n, we can find c n subsets of N of size
` which pairwise have intersection smaller than n. By a probabilistic argument, it follows
that the proportion of n-uniform hypergraphs on N vertices which are `-free is at most
`
N
`
N
N
(1 − 2−(n) )c( n ) ≤ 2−c(n)( n ) := 2−d( n ) .
Multiplying together the number of choices for T |V with the number of choices for G|V , we
have that the number of possibilities for S|V is at most
N
N
N
(2(1−d)( n ) )(2k(m) ) 2( n ) .
This shows that S is not (< ℵ0 )-near-closed.
Let us now briefly discuss why Theorem 1.4 implies Theorem 1.3. Recall (see [HS]) that
in any compact left-topological semigroup S, the smallest ideal K(S) is both the union of
the minimal right ideals and the union of the minimal left ideals. The intersection of any
18
D. BARTOŠOVÁ AND A. ZUCKER
minimal right ideal and any minimal left ideal is a group, so in particular contains exactly
one idempotent. More concretely, if M ⊆ S is a minimal right ideal and u ∈ M is an
idempotent, then Su is a minimal left ideal and M u = M ∩ Su. All the groups formed in
this way are algebraically isomorphic. When S = S(G) for some topological group G, we
can interpret this group as aut(M (G)), the group of G-flow isomorphisms of M (G).
Fix M ⊆ S(G) a minimal subflow, and let ϕ : S(G) → M be a G-map. Letting p = ϕ(1G ),
then we must have ϕ = λp . It follows that ϕ is a retraction iff ϕ = λu for some idempotent
u ∈ M . Furthermore, if p ∈ M , then there is a unique idempotent u ∈ M with p = pu ∈ M u.
It follows that for some q ∈ M we have λq ◦ λp = λu .
Now suppose N ⊆ S(G) is another minimal right ideal, and that x 6= y ∈ S(G) can be
separated by a retraction ψ onto N . Pick any p ∈ M and form the G-map λp ◦ ψ. Notice
that λp |N is an isomorphism. For some q ∈ M we have λp ◦ ψ = λq . Then for some r ∈ M ,
we have λr ◦ λq = λu a retraction. It follows that x and y are also separated by λu . Hence
the relation of being separated by a retraction does not depend on the choice of minimal
subflow M ⊆ S(G).
Now let G = S∞ , and let M 6= N be the minimal right ideals found in Theorem 1.4. Let
L be any minimal left ideal, and let u ∈ M ∩ L and v ∈ N ∩ L be idempotents. We will
show that u and v cannot be separated by retractions, so let ϕ : S(G) → M be a retraction.
Then ϕ = λw for some idempotent w ∈ M . Then ϕ(u) = wu = u since idempotents in M
are left identities for M . But now consider ϕ(v) = wv. By our assumption on M and N , wv
is an idempotent. However, we must also have wv ∈ M ∩ L since M and L are respectively
right and left ideals. It follows that wv = u, so ϕ(u) = ϕ(v) as desired.
6. Proximal and Distal
The technique of finding M, N ⊆ S(G) minimal subflows with J(M ) ∪ J(N ) a semigroup
allows for a quick solution to Ellis’s problem for some Polish groups G.
Recall from the introduction that a pair of points x, y in a G-flow X is called proximal if
there is p ∈ E(X) with xp = yp and X is proximal if every pair of points is proximal. Now
suppose that M (G) is proximal, i.e. that M (G) is proximal. Then every element of M is an
idempotent; to see why, notice that it suffices to show that M ∩ L is a singleton whenever
L is a minimal left ideal. Indeed, suppose u 6= p ∈ M ∩ L, with u idempotent. Suppose
that (u, p) were proximal, i.e. that for some q ∈ S(G) we have uq = pq. Since M ∩ L is a
group with identity u, we must have pu = p. Now as M is a minimal right ideal, find r ∈ M
with uqr = u. But then pqr = puqr = pu = p. This is a contradiction, so (u, p) cannot be
proximal.
A G-flow X is distal if every pair of non-equal points is distal, that is, not proximal. A
useful fact (see [A]) is that X is distal iff E(X) is a group. If M (G) is distal and M ⊆ S(G)
is a minimal subflow, then J(M ) is a singleton. To see this, note that if u, v ∈ J(M ), then
uv = vv = v, so (u, v) is a proximal pair. If u ∈ J(M ) is the unique idempotent, then the
map ϕ : E(M ) → M given by p → u · p is a G-flow isomorphism.
For automorphism groups G with M (G) proximal or distal, it follows that the conclusion
of Theorem 1.4 is automatic for any two minimal right ideals M 6= N . The same argument
for S∞ shows that any two idempotents of the same minimal left ideal cannot be separated
by retractions. Of course, we need to know that S(G) contains more than one minimal right
ideal; see ([Ba], Corollary 11) for a proof of this fact.
SAMUEL COMPACTIFICATIONS
19
The following theorem collects some examples of Polish groups G with M (G) proximal.
Theorem 6.1. Let G be either Homeo(2ω ) or the automorphism group of the countablyinfinite-dimensional vector space over a finite field. Then S(G) ∼
6= E(M (G))
We now consider the case when M (G) is distal. The rest of this section is spent proving
Theorem 1.5.
proof of Theorem 1.5. Suppose M (G) is distal and metrizable, and fix minimal M ⊆ S(G).
As M ∼
= E(M ) is a group, we see that Aut(M ), the group of G-flow isomorphisms of M , acts
transitively on M . Therefore by ch. 2, Theorem 13 of [A], M must be equicontinuous, hence
a compact metrizable group. Furthermore, as M must have a comeager orbit by [BYMT],
we have that M is a single orbit.
Let H = {g ∈ G : ug = u}. By [MNT], H is extremely amenable. To show that H is
normal, we show that the map ϕ : G → M given by ψ(g) = ug is a homomorphism. Indeed,
we have ug · uh = (ugu)h = ugh since u is a two-sided identity for M . This shows the
existence of the short exact sequence from item (2) of Theorem 1.5.
i
ϕ
Now suppose 1 → H →
− G−
→ K → 1 is a short exact sequence of groups with H extremely
amenable and K compact metrizable. Notice that the action k · g := k · ϕ(g) turns K into a
G-flow. Notice that ϕ is left uniformly continuous, so we can extend to a map ϕ : S(G) → K
via ϕ(p) = limgi →p ϕ(gi ). Let M ⊆ S(G) be minimal; then ϕ|M is a G-map.
Since H ⊆ G is extremely amenable, let u ∈ M be an H-fixed point. Now viewing K as
the right coset space {Hg : g ∈ G}, we build a G-equivariant map ψ : K → M by setting
ψ(Hg) = ug. As the image of ψ is dense in M , we will be done once we show that ψ is
continuous. To see this, notice that we have the following commutative diagram.
M
ψ
ϕ
K
λϕ(u)
K
Indeed, let Hg ∈ K. Then ϕ(ψ(Hg)) = ϕ(ug)) = ϕ(u) · g = ϕ(u) · Hg. Since λϕ(u) and ϕ
are continuous, we must have that ψ is continuous.
7. Some ultrafilters on [ω]2
This last section includes a short discussion of some ultrafilters motivated by the work in
section 5. The first main theorem of this section provides a counterpoint to Theorem 1.4.
Theorem 7.1. It is consistent with ZFC that there is a minimal subflow M ⊆ S(G) so that
if N ⊆ S(G) is a minimal subflow with J(M ) ∪ J(N ) a semigroup, then M = N .
The second theorem points out a key difference between selective ultrafilters and (2, 3)selective ultrafilters. Recall that if p, q ∈ βω, then we say that q ≥RK p if there is a function
f : ω → ω with f (q) = p. Another characterization of selective ultrafilters is that they are
20
D. BARTOŠOVÁ AND A. ZUCKER
exactly the ultrafilters which are minimal in the Rudin-Keisler order. The next theorem
shows that (2, 3)-selectives can be very far from Rudin-Keisler minimal.
Theorem 7.2. If p ∈ βω, there is a countably closed forcing extension P adding a (2, 3)selective ultrafilter q with q ≥RK p.
As it turns out, these two theorems will both be proven using the same forcing construction.
We define a forcing P which is very similar to a forcing defined by Laflamme [L]. A slightly
more straightforward forcing would suffice for Theorem 7.1 where we don’t refer to a fixed
p ∈ βω, but with a bit more work, we can prove both theorems.
F
Definition 7.3. Fix p ∈ βω. Write ω = n En with |En | = n. We define P = hP, ≤i as
follows.
(1) A condition A ∈ P is a subset of ω so that for every k < ω, we have {n < ω :
|A ∩ En | ≥ k} ∈ p.
(2) We declare that B ≤ A iff B ⊆ A.
If A, B ∈ P, we define B A iff there is k < ω so that {m < ω : |Em ∩ (B \ A)| ≤ k} ∈ p.
It is straightforward to see that hP, i is a separative pre-order which is equivalent to P.
Lemma 7.4. P is countably closed.
Proof. T
First notice that if hAn : n < ωi is a -decreasing sequence in P , then setting
A0n = i≤n Ai , we have that A0n is -equivalent to An . So we may freely work with ≤decreasing sequences.
Suppose hAn : n < ωi is a ≤-decreasing sequence in P . Write S(m, k) = {n < ω :
|Am ∩ En | ≥ k}. Note that S(m, k) ∈ p for every m, k < ω. Also, if m ≤ m0 and k ≤ k 0 ,
then S(m0 , k 0 ) ⊆ S(m, k).
S For m ≥ 1, we define Tm = S(m, m) \ S(m + 1, m + 1). Note that if m < ω, then
n≥m Tm = S(m, m). If m ≥ 1 and n ∈ Tm , then |Am ∩ En | ≥ m. We form B ∈ P by setting
[ [
B=
Am ∩ En .
m≥1 n∈Tm
For each m ≥ 1, we have {n < ω : |B ∩ En | ≥ m} = S(m, m) ∈ p, so B ∈ p. To see that
B Am , we note that {n < ω : B 6⊆ Am } ⊆ ω \ S(m, m).
If A ∈ P, we set à =
S
n [A
∩ En ]2 ⊆ [ω]2 . The next proposition will prove Theorem 7.2.
Proposition 7.5. Let G ⊆ P be generic. Then G̃ := {Ã : A ∈ G} generates a thick ultrafilter
on [ω]2 which is (2, n)-selective for every n. Furthermore, this ultrafilter is RK-above p.
Proof. Set E2 = 1eP , and suppose E2 = S t T . Let A ∈ p. By Ramsey’s theorem, there is
some non-decreasing function k → b(2, k) increasing to infinity so that any 2-coloring of the
complete graph on k vertices has a monochromatic clique of size b(2, k). If |A ∩ EN | = k,
then let XN ⊆ A ∩ EN be chosen so that |XN | = b(2, k) and X˜N ⊆ S or X˜N ⊆ T . Define
S 0 , T 0 ⊆ ω, placing N ∈ S 0 S
or N ∈ T 0 depending on which outcome happens. WLOG suppose
0
S ∈ p. Then letting X = N ∈S 0 XN , we have X ∈ P, X ≤ A, and X decides whether S or
T is in the filter generated by G̃.
SAMUEL COMPACTIFICATIONS
21
The argument that the ultrafilter generated by G̃ is (2, n)-selective is almost the exact
same. By Ramsey’s theorem, there is some non-decreasing function k → b(n, k) increasing
to infinity so that any 2-coloring of the complete n-uniform hypergraph on k-vertices has a
monochromatic clique of size b(n, k). Now letting En = λ(n−2) (E2 ), fix a partition En = StT .
If A ∈ P, we can in a similar fashion find X ≤ A deciding whether S or T is in the filter
λ(n−2) (G̃).
Lastly, let ψ : E2 → ω be so that ψ({x, y}) = n iff {x, y} ⊆ En . Then if U ∈ V [G] is the
ultrafilter generated by G̃, then ψ(U) = p.
We now turn towards the proof of Theorem 7.1. To do this, we use Theorem 5.7. Working
in V [G], let MG ⊆ S(S∞ ) be the unique minimal subflow so that ϕ(F2MG ) is the ultrafilter
generated by G̃. We need to show that {A(<0 , <1 ) : (<0 , <1 ) ∈ AMG } generates G̃. To see
why this is, fix A ∈ P. We may assume that if A ∩ En 6= ∅, then |A ∩ En | ≥ 2. We will
construct linear orders
S <0 and <1 so that A(<0 , <1 ) = Ã.
First write ω = n Xn , where X0 = ω \ A and Xn = A ∩ En . Some of the Xn may
be empty, but this is fine. First define <0 and <1 on X0 to be any linear orders which
completely disagree. Suppose <0 and <1 have been defined on X0 ∪ · · · ∪ Xn−1 . First define
<0 and <1 on Xn so that they agree. Now place Xn <0 -below everything built so far and
also <1 -above everything built so far. Then A(<0 , <1 ) = Ã as desired. This completes the
proof of Theorem 7.1.
The proof of Theorem 7.1 suggests another type of ultrafilter on [ω]2 we can define. If
p ∈ β[ω]2 is thick, define Ap = {(<0 , <1 ) : A(<0 , <1 ) ∈ p}. As we saw in section 5, Ap is an
equivalence relation on LO(ω).
Definition 7.6. Let p ∈ β[ω]2 be a thick ultrafilter. We call p a linear order ultrafilter if
{A(<0 , <1 ) : (<0 , <1 ) ∈ Ap } generates p. Call p a weak linear order ultrafilter if p is the
unique thick ultrafilter containing every A(<0 , <1 ) with (<0 , <1 ) ∈ p.
One can prove that there are thick ultrafilters p ∈ β[ω]2 which are not weak linear order
ultrafilters, providing an alternate proof of Theorem 1.4. The proof is very similar to the
proof that some p ∈ β[ω]2 is not weakly (2, 3)-selective.
We end with some open question about these ultrafilters.
Question 7.7. Does ZFC prove the existence of (2, 3)-selective ultrafilters? Of linear order
ultrafilters?
Question 7.8. Can there exist a weakly (2, 3)-selective ultrafilter which is not (2, 3)-selective?
Same question for linear order ultrafilters.
The last question is motivated by Theorem 7.2. This shows that (2, 3)-selective ultrafilters
can exist arbitrarily high up in the Rudin-Keisler order.
Question 7.9. Is it consistent with ZFC that the (2, 3)-selective ultrafilters are upwards
Rudin-Keisler cofinal?
References
[AKS] M. Ajtai, J. Komlós, and E. Szemerédi, A Note on Ramsey Numbers, Journal of Combinatorial
Theory, 29, (1980) 354–360.
22
D. BARTOŠOVÁ AND A. ZUCKER
[A] J. Auslander, Minimal Flows and Their Extensions, North Holland, 1988.
[B] Z. Baranyai, On the factorization of the complete uniform hypergraph, Colloq. Math. Soc. Janos Bolyai,
10 (1975), 91–108.
[Ba] D. Bartošová, Topological dynamics of automorphism groups of ω-homogeneous structures via near
ultrafilters, Ph.D. Thesis, University of Toronto, 2013.
[BYMT] I. Ben-Yaacov, J. Melleray, and T. Tsankov, Metrizable universal minimal flows of Polish groups
have a comeagre orbit, GAFA, 27(1) (2017), 67–77.
[E] R. Ellis, Lectures on Topological Dynamics, W.A. Benjamin, 1969.
[F] H. Firstenberg, Disjointness in ergodic theory, minimal sets, and a problem in Diophantine approximation, Math. Syst. Theory, 1 (1967), 1–49.
[GW] E. Glasner and B. Weiss, Minimal actions of the group S(Z) of permutations of the integers, Geometric
and Functional Analysis, 12 (2002), 964988.
[GW1] E. Glasner and B. Weiss, Interpolation sets for subalgebras of `∞ (Z), Israel J. Math., 44(4) (1983),
345–360.
[HS] N. Hindman and D. Strauss, Algebra in the Stone-Čech Compactification, 2nd Edition, De Gruyter,
2012.
[KPT] A.S. Kechris, V.G. Pestov, and S. Todorčević, Fraı̈ssé limits, Ramsey theory, and topological dynamics
of automorphism groups, Geometric and Functional Analysis, 15 (2005), 106–189.
[L] C. Laflamme, Forcing with filters and complete combinatorics, Annals of Pure and Applied Logic, 42(2)
(1989), 125–163.
[MNT] J. Melleray, L. Nguyen Van Thé, T. Tsankov, Polish groups with metrizable universal minimal flows,
Int. Math. Res. Not., no. 5 (2016), 1285–1307.
[NVT] L. Nguyen Van Thé, More on the Kechris-Pestov-Todorčević Correspondence: Precompact Expansions, Fund. Math., 222 (2013), 19-47.
[P] V. Pestov, On free actions, minimal flows, and a problem by Ellis, Trans. Amer. Math. Soc., 350 (10),
(1998), 4149–4165.
[P1] V. Pestov, Some universal constructions in abstract topological dynamics, Topological dynamics and
applications, 215 of Contemp. Math. (1998), 8399.
[Sa] Pierre Samuel, Ultrafilters and compactifications of uniform spaces. Trans. Amer. Math. Soc., 64 (1948),
100–132.
[U] V. Uspenskij, Compactifications of topological groups, Proceedings of the ninth Prague topological
symposium (2001), 2002, 331–346
[Z] A. Zucker, Topological dynamics, ultrafilter combinatorics, and the Generic Point Problem. Trans.
Amer. Math. Soc., 368(9), (2016).
[Z1] A. Zucker, Thick, syndetic, and piecewise syndetic subsets of Fraı̈ssé structures. Preprint (2016).
| 4math.GR
|
Distributed Scheduling Algorithms for Optimizing
Information Freshness in Wireless Networks
arXiv:1803.06469v1 [cs.IT] 17 Mar 2018
Rajat Talak, Sertac Karaman, and Eytan Modiano
Abstract—Age of Information (AoI), measures the time elapsed
since the last received information packet was generated at the
source. We consider the problem of AoI minimization for singlehop flows in a wireless network, under pairwise interference constraints and time varying channel. We consider simple, yet broad,
class of distributed scheduling policies, in which a transmission
is attempted over each link with a certain attempt probability.
We obtain an interesting relation between the optimal attempt
probability and the optimal AoI of the link, and its neighboring
links. We then show that the optimal attempt probabilities can
be computed by solving a convex optimization problem, which
can be done distributively.
I. I NTRODUCTION
Timely delivery of information updates is gaining increasing
relevance with the advent of technologies such as autonomous
flying vehicles, internet of things, and other cyber physical
systems. In autonomous flying vehicles, for example, timely
exchange of position, velocity, and other control information
can improve network safety [1], [2].
Packet delay, the traditionally used measure, accounts for
the average time taken for a packet to reach the destination
node, but does not measure the ‘information lag’ at the destination node, especially when the information update generation
can be controlled [3]–[5]. For example, if two packets carry
the same information, and one arrives early at the destination,
then there is no need in accounting for the delay in reception
of the second packet.
Age of information (AoI), a newly proposed metric [3], [6],
is defined to be the time elapsed since the last received information update was generated at the source node. AoI, upon
reception of a new update packet drops to the time elapsed
since generation of the packet, and grows linearly otherwise.
AoI, therefore, more accurately captures the ‘information lag’
at the destination node.
AoI has been analyzed for various queueing models [3], [7]–
[11], but very little attention has been paid to AoI minimization
in wireless networks. A problem of scheduling finitely many
update packets under physical interference constraints was
shown to be NP-hard in [12]. Age for a broadcast network,
where only a single link can be activated at any time, was
studied in [13]–[15]. Some preliminary analysis of age for
a slotted ALOHA like random access was done in [16],
The authors are with the Laboratory for Information and Decision Systems
(LIDS) at the Massachusetts Institute of Technology (MIT), Cambridge, MA.
{talak, sertac, modiano}@mit.edu
Submitted to SPAWC 2018.
This work was supported by NSF Grants AST-1547331, CNS-1713725, and
CNS-1701964, and by Army Research Office (ARO) grant number W911NF17-1-0508.
while multi-hop networks have been considered in [5], [17].
More recent work includes [4], [18], [19], which propose
centralized scheduling policies for AoI minimization, under
general interference constraints and time varying channels.
In a wireless network, it is not always possible to implement
centralized schemes. In this paper, we consider a simple, yet
broad, class of distributed scheduling policies for wireless
networks, in which each link attempts transmission with a
certain probability. We consider time varying channel and a
pairwise interference model, in which each link interferes with
a specified set of links in the network.
We first derive an interesting relation between the optimal
attempt probability, and the optimal age of the given link
and the interfering links. We then characterize the optimal
distributed policy πD , as a solution to a convex optimization
problem. We further show that this convex problem can be
solved distributively.
II. S YSTEM M ODEL
We model a wireless communication network as a graph
G = (V, E), where V is the set of nodes and E is the set of
directed communication links between the nodes. Each link
e ∈ E is associated with a source-destination pair. Time is
slotted, with the duration of each slot set equal to the time
it takes to transmit a single update packet. For simplicity, we
normalize the slot duration to unity.
Transmission of update packets cannot occur simultaneously
over all links due to wireless interference constraints [20]. For
each link e ∈ E, there is a subset of links Ne ⊂ E, that
interfere with link e, i.e., if link e and a link l ∈ Ne attempt
simultaneous transmission, then the transmission over link e
fails due to interference. We call this the pairwise interference
model. Popular models such as 1-hop and 2-hop interference
models [20] are special cases of this model.
Channel uncertainty can also lead to failure in reception
of update packets. We consider a two state channel process
Se (t) ∈ {0, 1} for all time t and links e. Channel state Se (t) =
1 implies that link e’s channel is ON at time t, and a noninterfering transmission over e is successfully received. When
Se (t) = 0, even a non-interfering transmission over link e fails
due to bad channel conditions. We assume the channel process
{Se (t)}t,e to be independent and identically distributed (i.i.d.)
across time t, and only independent across links e. We let
γe = Pr (Se (t) = 1) denote the channel success probability
for link e. We consider the case when the channel statistics,
namely γe for all e ∈ E, are known but the current channel
state Se (t) cannot be observed.
We use Ue (t) to denote scheduling decisions at time t, for
each link e. Ue (t) = 1 when link e attempts transmissions
in slot t, and Ue (t) = 0 otherwise. Thus, a successful
transmission occurs over link e if and only if a transmission
is attempted on link e, the link e’s channel is ON, and
no transmission is attempted over links e0 ∈ Ne . This is
equivalent to
Y
(1 − Ue0 (t)) = 1.
Ûe (t) , Se (t)Ue (t)
(1)
e0 ∈Ne
A. Age of Information
AoI of a network is a function of AoI of each sourcedestination pair. Thus, we define age Ae (t) for each link, to
be the time since the last received update was generated. We
assume source nodes to be active sources, i.e., it can generates
a fresh update packet before every transmission.
When all sources are active, the age Ae (t), is equal to
the time elapsed since the last successful transmission over
e. Thus, the evolution of age Ae (t) can be written as
Ae (t) + 1 if Ûe (t) = 0
Ae (t + 1) =
,
(2)
1
if Ûe (t) = 1
for all e ∈ E. We define average age for link e to be
Aave
e = lim sup
T →∞
T
1X
Ae (t),
T t=1
(3)
whenever the limit exists. We define the peak age to be the
average of all the age peaks:
Ape
Ne (T )
X
1
Ae (Te (i)),
= lim sup
T →∞ Ne (T ) i=1
(4)
where Te (i) denote the ith instance of successful transmission
over link e, i.e., the ith instance when Ûe (t) = 1, and Ne (T )
is the number of successful transmissions over link e until
time T . We define average and peak AoI of the network to be
a weighted sum of link AoI:
X
X
Aave =
we Aave
and Ap =
we Ape .
(5)
e ,
e∈E
e∈E
Our objective is to design distributed scheduling policies
that minimize the network average and peak age.
B. Distributed Stationary Policies
We focus our attention on policies in which scheduling
decisions are made by the source of each link e, and not
by a centralized scheduler. In particular, we consider policies
in which each link e attempts transmission with probability
pe > 0, independent across links and slots. We refer to these
policies as the distributed stationary policies.
The probability that a non-interferingQtransmission is attempted over link e is given by fe = pe e0 ∈Ne (1 − pe0 ). We
will refer to fe as the link activation frequency of link e, and
use f = (fe )e∈E to denote the link activation frequency vector.
Fig. 1. Example of two interfering links, 1 and 2. The yellow region is FD .
Also shown is the set of feasible link activation frequencies for centralized
schedule: F = {(f1 , f2 ) | f1 + f2 ≤ 1 and fi ≥ 0}.
The space of all link activation frequencies f , attainable using
distributed stationary policies, is given by
Y
FD = f ∈ R|E| | fe = pe
(1 − pe0 )
e0 ∈Ne
and 0 ≤ pe ≤ 1 ∀ e ∈ E . (6)
Figure 1 shows the set of feasible link activation frequencies
for distributed policies for the two interfering link example.
Also, shown is the set of all link activation frequencies F
attainable using a centralized scheduler. It can be seen that
the set FD is non-convex, and strictly smaller than F. The
gap between these two sets indicates the price one has to pay
in resorting to distributed scheduling.
III. O PTIMAL P OLICY
In this section, we characterize an age optimal policy, and
propose an implementation that requires only local information. We first show that for any distributed stationary policy,
the peak and average age are equal, and can be written as a
simple convex function in link activation frequencies fe .
Lemma 1: For any distributed stationary policy, the
average and peak age is given by
X we
Aave = Ap =
,
(7)
γe fe
e∈E
Q
where fe = pe e0 ∈Ne (1 − pe0 ) is the link activation
frequency of link e.
Proof: A similar result was proved in [4] which considered a class of centralized policies, that are not necessarily
stationary. Here, we provide a proof for distributed stationary
policies.
Consider a link e. Then, under any distributed stationary
policy, the event that link e is successfully activated in slot
t is independent and identically distributed across slots, with
success probability of γe fe . Thus, the time since last activation
is geometrically distributed with mean γe1fe . It turns out that
the peak and average age for link e are indeed equal to this
quantity, i.e.,
1
p
Aave
,
e = Ae =
γe fe
P
we
and therefore, the network age is
e∈E γe fe . The detailed
proof is given in Appendix A.
An immediate consequence of Lemma 1 is that both peak
and average age minimization problems, over the space of all
distributed stationary policies, can be written as
X we
Minimize
(8)
f ∈FD
γe fe
e∈E
Notice that, although the objective function is convex, the
constraint set FD is non-convex, see Figure 1. Furthermore,
obtaining optimal link activation frequencies f ∗ does not
suffice as the distributed policy is characterized by the attempt
probabilities pe . The optimal link attempt probabilities p∗ , that
solve (10), yield a distributed stationary policy, that is both
peak and average age optimal. We call this policy πD .
The following result characterizes the optimal link attempt
probabilities p∗ in terms of the optimal link age A∗e .
Theorem 1: The attempt probabilities p∗
∈ E), that solve (8), are given by
=
(p∗e |e
p∗e =
we A∗e
+
w A∗
Pe e
e0 :e∈Ne0
we0 A∗e0
,
(9)
−1
Q
is
for all e ∈ E, where A∗e = γe p∗e e0 ∈Ne (1 − p∗e0 )
the optimal peak/average age of link e.
Proof: Substituting FD , from (6), in (8) we get
X we
Minimize
γe fe
p∈[0,1]|E| ,f ∈R|E|
e∈E
Y
subject to
fe = pe
(1 − pe0 ) ∀ e ∈ E
However, Theorem 1 shows that the attempt probability
p∗e = P
A∗e
e0 ∈E
A∗e0
,
(13)
is both peak and average age optimal. This gives us a more
precise answer to the question posed in [16]. Our result,
however, holds under the more general pairwise interference
constraints, and non-equal weights we > 0.
A. Distributed Computation of πD
Although, Theorem 1 gives a characterization of the optimal
attempt probabilities, in terms of the optimal link age A∗e , it
does not provide for a simple method to computation.
We now provide for an alternate characterization of the
network age minimization problem (8). In particular, we show
that p∗ can be obtained by a simpler convex optimization
problem, than the one in (11). We will also see that this simpler
problem is akin to distributed computation of the policy πD .
Theorem 2: The age minimization problem (8) is
equivalent to the convex optimization problem:
!
X
X
λ
e
λe +
P
Maximize
λe0 H
λe ≥0
λe + e0 :e∈Ne0λe0
e∈E
e0 :e∈Ne0
X
we
+
λe 1 + log
, G(λ),
λe
e∈E
(14)
1
where H (p) = p log p1 + (1 − p) log 1−p
is the
∗
entropy function. If λ is the optimal solution to this
problem, then the optimal attempt probabilities p∗ are
given by
λ∗
p∗e = ∗ P e
,
(15)
λe + e0 :e∈Ne0 λ∗e0
for all e ∈ E.
(10)
e0 ∈Ne
fe ≥ 0 for all e ∈ E
Now, substituting qe = 1 − pe , the optimization problem (10)
reduces to
X
w
Q e
Minimize
p≥0,q≥0
γe pe e0 ∈Ne qe0
(11)
e∈E
subject to pe + qe ≤ 1 ∀e ∈ E
This is a convex program in standard form, and can be solved
using KKT conditions to obtain Theorem 1. The details are
given in Appendix B.
Preliminary results on age optimization for a ALOHA type
network model were presented in [16] for a network in which
only one link can be activated at any given time. As a heuristic,
it was argued in [16] that the attempt rates should be
√
1/ γe
P
pe =
(12)
√ .
0
e0 ∈E 1/ γe
Proof: See Appendix C.
We first note that the variable λe is a proxy for the weighted
age we Ae of link e, and the solution λ∗e corresponds to optimal
weighted age we A∗e . This can also be intuited from (15), which
is same as (9).
The objective function G(λ) in (14) is a concave function;
since the problem is convex. Therefore, a simple projected
gradient ascent is guaranteed to converge to the optimal
λ∗ [21]. The gradient of the function G (λ) is given by
X
we
θe
λe0
∂G (λ)
= log
+ log 1+
+
log 1 +
,
∂λe
λe
λe 0
θ e0
e :e∈∈Ne0
P
0
for all e ∈ E, where θe =
e0 :e∈Ne0 λe . Note that the
∂G(λ)
gradient ∂λe depends only on λe and λe0 , for e0 that are
‘extended neighbors’ of e. This space dependence of ∂G(λ)
∂λe
on λ makes the projected gradient descent algorithm akin to
distributed implementation.
We give the projected gradient descent algorithm for distributively computing policy πD in Algorithm 1. In it, we
divide time into frames, each frame of F ≥ 1 slots. For
simplicity, we assume symmetric interference, i.e., e interferes
with e0 if and only if e0 interferes with e which is same as the
condition Ne = {e0 ∈ E | e ∈ Ne0 } for all links e.
Link e, sets its attempt probability to pe (m), in frame
m ≥ 1. The source of link e tracks and updates two variables,
namely, λe (m) and θe (m). We use Π to denote the projection
on the set [, +∞), for a > 0. As seen in Steps 5 and 7 in
Algorithm 1, variables λe (m) and θe (m) need to be exchanged
only between the neighboring links.
Algorithm 1 Distributed Computation of Policy πD
1: pe (m): attempt probability of link e, in frame m
2: λe (m), θe (m): dual variables, in frame m
3: Start: Set λe (0) = 1 and θe (0) = |Ne | for all e ∈ E. Set
pe (0) = 1/2 and m = 0.
4: for each frame m do
5:
Send: θe (m) to all e0 ∈ Ne
6:
Compute λe (m + 1):
"
(
we
λe (m + 1) ← Π λe (m) + ηm log
λe (m)
)#
X
θe (m)
λe0 (m)
+ log 1 +
+
log 1 +
λe (m)
θe0 (m)
0
e ∈Ne
7:
8:
Send: λe (m + 1) to all e0 ∈ Ne
Update θe (m + 1):
X
θe (m + 1) ←
λe0 (m + 1)
e0 ∈Ne
9:
Update pe (m + 1):
pe (m + 1) ←
10:
11:
λe (m + 1)
λe (m + 1) + θe (m + 1)
m←m+1
end for
IV. C ONCLUSION
We considered age minimization in a wireless network, with
pairwise interference constraints, time varying channels, and
single-hop flows. We considered a simple class of distributed
scheduling policies, in which each link attempts transmission
with probability pe . We showed an interesting relationship
between the optimal attempt probability for a link, and the
optimal age of the link and it’s neighboring links. We then
showed that the optimal link attempt probabilities can be
obtained by solving a convex optimization problem, which
can be done distributively using the projected gradient ascent
algorithm.
R EFERENCES
[1] O. K. Sahingoz, “Networking models in flying ad-hoc networks
(FANETs): Concepts and challenges,” J. Intell. Robotics Syst., vol. 74,
pp. 513–527, Apr. 2014.
[2] R. Talak, S. Karaman, and E. Modiano, “Speed limits in autonomous
vehicular networks due to communication constraints,” in Proc. CDC,
pp. 4998–5003, Dec. 2016.
[3] S. Kaul, R. Yates, and M. Gruteser, “Real-time status: How often should
one update?,” in Proc. INFOCOM, pp. 2731–2735, Mar. 2012.
[4] R. Talak, S. Karaman, and E. Modiano, “Optimizing information freshness in wireless networks under general interference constraints,” in
submited to Mobihoc, Jun. 2018.
[5] R. Talak, S. Karaman, and E. Modiano, “Minimizing age-of-information
in multi-hop wireless networks,” in Proc. Allerton, Oct. 2017.
[6] S. Kaul, M. Gruteser, V. Rai, and J. Kenney, “Minimizing age of
information in vehicular networks,” in Proc. SECON, pp. 350–358, Jun.
2011.
[7] L. Huang and E. Modiano, “Optimizing age-of-information in a multiclass queueing system,” in Proc. ISIT, pp. 1681–1685, Jun. 2015.
[8] C. Kam, S. Kompella, and A. Ephremides, “Age of information under
random updates,” in Proc. ISIT, pp. 66–70, Jul. 2013.
[9] S. K. Kaul, R. D. Yates, and M. Gruteser, “Status updates through
queues,” in Proc. CISS, pp. 1–6, Mar. 2012.
[10] E. Najm and R. Nasser, “Age of information: The gamma awakening,”
in Proc. ISIT, pp. 2574–2578, Jul. 2016.
[11] M. Costa, M. Codreanu, and A. Ephremides, “Age of information with
packet management,” in Proc. ISIT, pp. 1583–1587, Jun. 2014.
[12] Q. He, D. Yuan, and A. Ephremides, “Optimizing freshness of information: On minimum age link scheduling in wireless systems,” in Proc.
WiOpt, pp. 1–8, May 2016.
[13] I. Kadota, E. Uysal-Biyikoglu, R. Singh, and E. Modiano, “Minimizing
the age of information in broadcast wireless networks,” in Proc. Allerton,
pp. 844–851, Sep. 2016.
[14] Y.-P. Hsu, E. Modiano, and L. Duan, “Age of information: Design and
analysis of optimal scheduling algorithms,” in Proc. ISIT, pp. 1–5, Jun.
2017.
[15] I. Kadota, A. Sinha, and E. Modiano, “Optimizing age of information
in wireless networks with throughput constraints,” in to appear in Proc.
INFOCOM, Apr. 2018.
[16] S. K. Kaul and R. D. Yates, “Status updates over unreliable multiaccess
channels,” ArXiv e-prints, May 2017.
[17] A. M. Bedewy, Y. Sun, and N. B. Shroff, “Age-optimal information
updates in multihop networks,” in Proc. ISIT, pp. 576–580, Jun. 2017.
[18] R. Talak, I. Kadota, S. Karaman, and E. Modiano, “Scheduling policies
for age minimization in wireless networks with unknown channel state,”
in submitted to ISIT, Jun. 2018.
[19] R. Talak, S. Karaman, and E. Modiano, “Optimizing age of information
in wireless networks with perfect channel state information,” in to
appear in Proc. WiOpt, May 2018.
[20] A. Kumar, D. Manjunath, and J. Kuri, Wireless Networking. Morgan
Kaufmann, 2008.
[21] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge
University Press, 2004.
[22] R. W. Wolff, Stochastic Modeling and the Theory of Queues. Prentice
Hall, 1 ed., 1989.
A PPENDIX
A. Proof of Lemma 1
Consider a distributed stationary policy with link activation
frequencies fe > 0 for each e ∈ E. For a link e, let Te (i)
be the ith instance when link e was successfully activated,
i.e., the ith instance when Ûe (t) = 1. Let Xe (i) = Te (i) −
Te (i−1) to be the ith inter-(successful) activation time for link
e. Since all the processes involved, namely Se (t) and Ue (t),
are i.i.d. across time
h t, Xe (i)i is geometrically distributed with
rate equal to P Ûe (t) = 1 = γe fe , i.e., P [Xe (i) = k] =
k−1
γe fe (1 − γe fe )
, for all k ∈ {1, 2, . . .}.
We now compute the average age Aave
e for link e. Note that
the age Ae (t) over slots t ∈ {Te (i)+1, . . . Te (i+1)} increases
PTe (i+1)
from 1 to Xe (i) in steps of 1. Therefore, t=T
Ae (t) =
e (i)+1
PXe (i)
m for all i. Then, using renewal theory [22], we can
derive the average age to be
PN PXe (i)
T
1X
ave
i=1
m=1 m
P
Ae (t) = lim
,
Ae = lim sup
m
N →∞
T →∞ T t=1
i=1 Xe (i)
PN 1
i=1 2 Xe (i) (Xe (i) + 1)
= lim
,
PN
N →∞
i=1 Xe (i)
1 E [Xe (i) (Xe (i) + 1)]
=
a.s..
(16)
2
E [Xe (i)]
m=1
1
Computing the moments of Xe (i) we get Aave
e = γe fe a.s..
For peak age, note that the ith peak is equal to the ith inter(successful) activation time Xe (i), i.e., Ae (Te (i)) = Xe (i).
Therefore, we have
Ape = lim sup
T →∞
N (t)
N
1 X
1 X
Ae (Te (i)) = lim
Xe (i),
N →∞ N
N (T ) i=1
i=1
which is equal to E [Xe (1)] =
1
γe fe
a.s.. This proves the result.
B. Proof of Theorem 1
This follows by first making the the substitution pe = ePe ,
qe = eQe , and fe = ehe , and then taking log in each of
the constraints. This is a standard technique in geometric
programming [21]. However, unlike geometric programming,
we don’t need to apply log function to the objective as it is
already convex, and also separable in variables he .
The problem (20) is convex [21], and satisfies slater’s
conditions. As a consequence, the duality gap is zero, and
the problem (20) is equivalent to its Lagrangian dual problem.
Define the Lagrangian function of the optimization problem (20) to be
X
X we
e−he +
νe log ePe + eQe
L (h, P, Q, λ, ν) =
γe
e∈E
e∈E
X
X
+
λe he − Pe −
Qe0 , (21)
e0 ∈Ne
e∈E
for λ ≥ 0 and ν ≥ 0. The dual problem for (20) is given by
Maximize g (λ, ν) ,
(22)
λ≥0,ν≥0
where g (λ, ν) is the dual objective function defines as
Consider the Lagrangian dual
X
X
w
Q e
L(p, q, λ) =
+
λe (pe + qe − 1) .
γe pe e0 ∈Ne qe0
e∈E
e∈E
(17)
The problem (11) satisfies Slater’s conditions, and thus, the
KKT conditions are both necessary and sufficient. Setting
partial derivatives of L with respect to pe and qe to zero,
we obtain
X
pe λe = we Ae , and qe λe =
we0 Ae0 ,
(18)
e0 :e∈Ne0
for all e ∈ E, where Ae = γe pe Q 10
is the age of
e ∈Ne qe0
link e. Equations (18) imply that λe cannot be zero. By
complementary slackness criteria we must have pe + qe = 1.
This, with (18), gives pe = we Ae +P w0 e Ae w 0 A 0 and λe =
e
e
e :e∈N 0
e
P
we Ae + e0 :e∈Ne0 we0 Ae0 . This proves the result.
C. Proof of Theorem 2
The age minimization problem, given in (10), can be written
as by substituting qe = 1 − pe :
X we
Minimize
f ≥0,p≥0,q≥0
γe fe
e∈E
Y
(19)
subject to
fe ≤ p e
qe0 ∀ e ∈ E
e0 ∈Ne
pe + qe ≤ 1 ∀ e ∈ E
Now, substituting Pe = log pe , Qe = log qe , and he = log fe ,
the problem reduces to
X we
Minimize
exp {−he }
h,P,Q
γe
e∈E
X
(20)
subject to he − Pe −
Qe0 ≤ 0 ∀ e ∈ E
g (λ, ν) =Minimize L (h, P, Q, λ, ν) .
h,P,Q
We will now show that the the optimization problem in
Theorem 2 is in fact the dual problem (22).
First note that the Lagrangian function L (h, P, Q, λ, ν) is
strictly convex in h, P, and Q [21]. Thus, first order conditions
∇h,P,Q L = 0 yield the optimal h, P, and Q. Setting these
partial derivatives to 0, we get
X
νe eQe
we −he
νe ePe
,
λe0 = Pe
, and λe =
e
,
λe = Pe
Q
Q
e
e
e +e
e +e
γe
e0 :e∈Ne0
(24)
for all e ∈ E. Adding and dividing the first two equations
in (24) also yields
!
X
λe
νe = λ e +
λe0 , and Pe − Qe = log P
,
0
e0 :e∈Ne0 λe
e0 :e∈Ne0
(25)
respectively, for all e. Substituting (24) and (25) in L, in (21),
we recover g(λ, µ) = G(λ), the objective function in (14),
and therefore, the dual problem (22) is same as the optimization problem in Theorem 2. This proves the first part.
We now show that if λ∗ is the solution to this problem,
then the optimal attempt probability is given by (15). Let f ∗ ,
p∗ , q∗ be the solution to (19), h∗ , P∗ , Q∗ be the solution
to (20), and λ∗ be the solution to the problem in Theorem 2.
Note
link e ∈ E, we must have fe∗ =
Q that,∗ for each
∗
∗
pe e0 ∈Ne qe and pe + qe∗ = 1; as otherwise the objective
function can be reduced by increasing fe∗ and/or p∗e . Since
∗
∗
∗
∗
p∗e = ePe and qe∗ = eQe , we have ePe + eQe = 1. Substituting
this in (24), and using (25), we obtain
λ∗e
∗
p∗e = ePe =
e0 ∈Ne
log (exp {Pe } + exp {Qe }) ≤ 0 ∀ e ∈ E
(23)
which proves the result.
λ∗e +
P
e0 :e∈Ne0
λ∗e0
,
(26)
| 7cs.IT
|
Pedestrian-Robot Interaction Experiments in an Exit Corridor
arXiv:1802.05730v1 [cs.RO] 15 Feb 2018
Zhuo Chen, Chao Jiang and Yi Guo
Abstract— The study of human-robot interaction (HRI) has
received increasing research attention for robot navigation in
pedestrian crowds. In this paper, we present empirical study of
pedestrian-robot interaction in an uni-directional exit corridor.
We deploy a mobile robot moving in a direction perpendicular
to that of the pedestrian flow, and install a pedestrian motion
tracking system to record the collective motion. We analyze both
individual and collective motion of pedestrians, and measure
the effect of the robot motion on the overall pedestrian flow.
The experimental results show the effect of passive HRI,
where the pedestrians’ overall speed is slowed down in the
presence of the robot, and the faster the robot moves, the lower
the average pedestrian velocity becomes. Experiment results
show qualitative consistency of the collective HRI effect with
simulation results that was previously reported. The study can
be used to guide future design of robot-assisted pedestrian
evacuation algorithms.
I. I NTRODUCTION
With the rapid advances in the development of autonomous
mobile robot technologies, the applications of service and
socially assistive robots have been woven into people’s daily
life. As robots are performing tasks where human beings
are involved, the problem of human-robot interaction (HRI)
arises and has received considerable attention during the
past decade. Understanding the effect of HRI on human’s
behavior, especially in human and robot collective motion,
is of great significance for robot control or decision-making.
In one of the recent works, HRI is studied and considered
for human-aware robot navigation, where traditional motion
planning approach is amended to respect the effect of HRI
on human behavior in the presence of mobile robot [1]. In
addition to considering human being as moving obstacles,
criteria such as human comfort [2, 3], natural motion [4, 5]
and socially-adaptive motion [6, 7] are taken in to account
for robot motion planning.
In the studies of robot-assisted guidance where multiple
robots work collaboratively to escort a group of people, HRI
is commonly modeled as repulsive or attractive force that is
embedded in the social force model which governs human
motion dynamics [8]. The robot performs either as a leader
who directs a group of people towards the destination, or
as a shepherd who regroups people escaping from the group
formation. Thus, human’s motion is controlled through the
effect of attractive and repulsive force from the leader and
shepherd robot, respectively. In [9], the parameters of the
social force based HRI model were found by analyzing the
This work was partially supported by the US National Science Foundation
under Grant CMMI-1527016.
The authors are with the Department of Electrical and Computer Engineering, Stevens Institute of Technology, Hoboken, NJ 07030, USA
{zchen39,cjiang6,yi.guo}@stevens.edu
data obtained from real-world experiment using an interactive learning approach.
HRI is also studied for robot motion planning in largescale human crowds. It has been found interesting how
implicit interaction between humans and mobile robots affects human collective motion, so that desired human crowd
behaviors can be achieved by adjusting the motion of robot in
human crowd. Specially, inspired by self-organized phenomena in human collective motion, attempts have been made to
improve human flow efficiency by introducing mobile robots
in human crowds. The large-scale dynamics of human crowds
can be modified as human crowds implicitly interact with the
robot [10, 11]. Previous study [12] has found that placing
a mobile robot in crossing human flows can help to create
diagonal strip pattern in human flow as the human agents try
to avoid the collision with the robot. The strip pattern is a
desirable phenomenon for preventing congestion in crossing
crowds. By adjusting their motion, the interacting robots can
improve human flow efficiency under different flow densities.
In our early work [13], simulations based on social force
models were carried out to characterize HRI in an unidirectional exit corridor. An interacting robot was placed and
programmed to move in a direction that is perpendicular to
that of the flow direction to regulate the flow velocity in an
evacuation situation in a uni-directional corridor.
The focus of this paper is to empirically study passive HRI
in an exit corridor for the purpose of robot-assisted pedestrian
flow regulation. We conduct real robot experiments with
a group of pedestrians walking towards an exit door, and
measure the effect of robot motion on the overall pedestrian
flow. Our experimental results show that in an exit corridor
environment, a robot moving in a direction perpendicular to
that of the uni-directional pedestrian flow can slow down
the uni-directional flow, and the faster the robot moves, the
lower the average pedestrian velocity becomes. Furthermore,
the effect of the robot on the pedestrian velocity is more
significant when people walk at a faster speed. This empirical
evidence is qualitatively consistent with simulation results
reported in our early paper [13], and can be used for future
development on robot-assisted pedestrian flow regulation.
The rest of the paper is organized as follows: Section
II introduces the the passive pedestrian-robot interaction
studied in the paper, and present the experimental setup and
procedure. Section III presents the experimental results, and
discuss comparisons with simulation results reported before.
The conclusion and future work are presented in Section IV.
Kinect sensors
Pedestrian flow direction
4m
Robot
𝒚
𝒙
𝑶
10m
Fig. 1: The schematic diagram of pedestrian flow in an indoor
corridor with robot-assisted regulation
Fig. 2: The customized Adept Pioneer P3-DX mobile robot.
II. PASSIVE P EDESTRIAN -ROBOT I NTERACTION AND
E XPERIMENTAL S ETUP
A. Passive Pedestrian-Robot Interaction
We empirically study the regulation of human flow velocity in an exit corridor environment using an interacting robot.
As shown in Fig. 1, a single interacting robot is added to the
pedestrian flow, and moves in a direction perpendicular to
that of the pedestrian flow with a predesignated speed. Since
the robot’s movement is in the way of the pedestrians towards
the exit, the pedestrians see the robot and avoid potential
collisions by either changing the speed or alter the trajectory
of their own. The pedestrian flow is regulated through the
effect of passive HRI. That is, the robot moves in a predesignated trajectory with a predesignated speed, and the
pedestrians avoid potential collisions when passing through
the robot thus passively interact with the robot. Note that we
assume normal situations that the pedestrians can clearly see
the robot and their movement behaviors are rational. We do
not consider panic situations or nonrational behaviors.
Our hypothesis is that the speed of the robot moving in a
direction perpendicular to that of a uni-directional pedestrian
flow will affect the average speed of the pedestrians going
through the corridor. The goal of the study is to test this
hypothesis, measure the effect of the passive HRI by adjusting the robot’s speed on the pre-designated trajectory, and
compare the effect with the case without the robot.
B. Experimental Setup
1) The Environment: The experiment was conducted in
the Griffith Building of Stevens Institute of Technology on
June 22, 2016. A pedestrian tracking system deployed in
the building covers a 10 meter by 4 meter tracking area as
shown in Fig. 1 with an exit door at the end of the area. The
boundaries of the tracking area are marked with red duct
tapes. The tracking system consists of 5 Microsoft Kinect
sensors and 6 computers connected to a local area network.
A pedestrian detection and tracking software OpenPTrack
[14] is used.
OpenPTrack is a real-time distributed people detection and
tracking software capable of utilizing RGB-D images from
the Kinect network. OpenPTrack runs on Robot Operating
System (ROS) in Ubuntu 14.04. During the experiment, each
of the 5 Kinect sensors is connected to a computer and the
image stream obtained by the Kinect sensor is fed into a
OpenPTrack detection program running on the computer. The
OpenPTrack detection program on each computer processes
the images and transmits the detection results to a central
computer over the local area network. The central computer
fuses the received data and records the tracking results.
2) The Mobile Robot: We used a customized Adept
Pioneer P3-DX mobile robot, as shown in Fig. 2. The robot
motion dynamics is described by a single integrator model,
that is
ẋr = ur
(1)
where xr = [xxr , xyr ]T ⊆ R2 denotes the robot position and
ur = [uxr , uyr ]T denotes robot velocity control. The robot
motion is controlled to follow a pre-designed trajectory that
is perpendicular to the pedestrian flow direction. The robot
velocity in the y direction is controlled to track a sinusoidal
signal, and the velocity control in the x direction remains 0.
That is,
uxr (t) = 0
(2a)
uyr (t) = Aω sin (ωt)
(2b)
where A is the amplitude of the simple harmonic motion,
which is set to 2m (half the corridor width); ω is the robot
motion frequency. This control law makes sure that the robot
moves within the boundary of the area, and the upper limit
of the controlled speed is Aω.
3) Tracking System Calibration: After the sensors are
mounted in place, both intrinsic and extrinsic parameters
of the tracking system need to be determined. Default
intrinsic parameters for the Kinect sensors were used without performing any intrinsic calibration. To calibrate the
extrinsic parameters of the tracking system, we printed a
6-by-7 checkerboard with the cell width and height both
being 75mm. The networking capabilities of ROS enable the
Kinect sensors’ poses to be estimated by: Step (1) letting
a pair of sensors both see the checkerboard to calibrate
their relative pose; Step (2) finding one calibrated sensor
and one uncalibrated sensor, and doing Step (1) with this
pair of sensors; Step (3) repeating Step (2) until all sensors
are calibrated; Step (4) placing the checkerboard on the
ground where it can be seen by at least 1 sensor to specify
the checkerboard reference frame; and Step (5) performing
calibration refinement with people detections [14].
By performing the 5 steps above, sensor poses are estimated with respect to the checkerboard reference frame
5
Low-velocity position centroid
Grid vertex ground truth position
3
4
y position (m)
y position (m)
2
1
0
3
2
1
-1
0
-2
-1
-8
-6
-4
-2
0
0
x position (m)
2
4
6
8
10
x position (m)
(a)
(b)
(c)
Fig. 3: Tracking system calibration: estimation of the coordinate transformation from checkerboard frame to ground frame. (a)
Walking Pattern in checkerboard reference frame; (b) Comparison between measurement and ground truth; (c) Measurement
error magnitude histogram.
(a) t=4.0s
(b) t=6.0s
(c) t=9.0s
(d) t=24.0s
Fig. 4: Experimental procedure. (a) People start to walk in the bounded area; (b) People walk toward the robot; (c) People
interact with the moving robot; (d) People finish one round and walk back to the starting line.
defined by the pose of the checkerboard in Step (4), which
does not necessarily coincide with the ground reference
frame as shown in Fig. 1. Here we define the transformation
T : P2 (R) → P2 (R) that maps the homogeneous coordinates
of points in checkerboard reference frame to those in ground
reference frame as
cos θ
0
∆x
sin θ ∆y
(3)
T= 0
0
0
1
where (∆x, ∆y) represents translation while θ represents
rotation.
To determine the transformation T, the vertices of a unitdistance square grid graph are marked on the tracking area
ground. In Fig. 3b, ground truth positions of the 55 vertices
marked with circles in Fig. 3b are expressed in checkerboard
reference frame as (xgt,k , ygt,k ), k = 1, 2, ..., 55. A person
is asked to walk through every marked vertex of the grid,
stop at each mark and remain still for a second or two before
proceeding to the next one. Fig. 3a shows an example pattern
of trajectory recorded by the tracking system, starting from
(4, 0) and ending at (10, 0) in ground frame coordinates.
Once the trajectory is recorded, the measured positions
where the subject stops, or low-velocity points, are obtained
by filtering out those points in the trajectory whose velocity is greater than 0.2m/s. These low-velocity positions
are clustered and the centroid of each cluster is marked
with a “×” in Fig. 3b. The low-velocity position centroid
expressed in checkerboard reference frame is denoted by
(xct,k , yct,k ), k = 1, 2, ..., 55. We solve for the transformation T by minimizing the sum of squared errors between lowvelocity position centroids and their corresponding ground
truth positions, that is,
2
55
xct,k
xgt,k
X
(4)
T · yct,k − ygt,k
e (T) =
0
0
k=1
C. Experiment Procedure
After the tracking system was set up and calibrated, we
conducted experiments with 6 cases as listed in Table I. We
compare HRI in two testing scenarios with people walking
at a normal speed (Cases #1, #2, and #3) and a faster
speed (Cases #4, #5, and #6). In each of the two scenarios,
we compare pedestrian motions with: 1) no robot, 2) a
slow-moving robot, and 3) a fast-moving robot. The mobile
robot, whose motion is subject to (1), has a pre-set angular
frequency ω = 0.1rad/s for the slow-moving cases (Cases #2
and #5), and ω = 0.4rad/s in the fast-moving cases (Cases
#3 and #6).
In each case, there are 11 participants who have no
knowledge of the aim or nature of this experiment. As
shown in Table I, in Cases #1 and #4, pedestrians walk
without robot intervention, while in the rest of the cases,
a mobile robot whose motion is subject to (1) is introduced
in the environment. The participants are instructed to walk
within the boundaries of the tracking area at two different
walking speed, i.e., normal or faster. As shown in Fig. 4,
the participants waiting on the left side of the tracking area
receives an instruction to walk, and they enter the tracking
area. Once they exit the tracking area from the right side of
the corridor, they walk back through another pathway next
to the tracking area and get ready for the next run. Each case
consists of 3 repeated runs. Fig. 4 shows the snapshots of a
complete run of Case #3.
TABLE I: Experimental scenarios and parameters
Case #
1
2
3
4
5
6
Human
Instructed
# of
Walking
People
Speed
11
Normal
11
Normal
11
Normal
11
Faster
11
Faster
11
Faster
Robot
Angular
Start
Freq.
Position
ω(rad/s)
N/A
N/A
0.1
(6.6, 0)
0.4
(6.6, 0)
N/A
N/A
0.1
(6.6, 0)
0.4
(6.6, 0)
We present the results from collected data in the next
section.
III. E XPERIMENTAL R ESULTS
In this section, we present the observed HRI behaviors,
summary statistics for collected HRI, and compare with
numerical simulation results reported in our previous work
[13].
A. Individual HRI Behaviors
We have recorded pedestrian trajectories using the
OpenPTrack software and a video recording system. We
observe the HRI behaviors when people get close to the
robot. It appears that the pedestrian motion behavior changes
according to their anticipation of potential collision with the
robot, and the density of the area and whether there’s room to
move around. We observe that the speed change of the robot
affects humans’ motion. We first describe individual HRI
behaviors in this subsection, and the HRI effect on collective
motion and the average flow velocity is presented in the next
subsection.
Fig. 5 shows the temporal sequence of experiment snapshots in Case #2, where the pedestrians walk at normal speed.
It can be seen that the pedestrians approach the interacting
robot at t = 46s and start to avoid the robot at t = 47s. At
t = 48s and t = 49s, some pedestrians adjust their walking
directions to avoid collision with the robot. Fig. 6 shows
the temporal sequence of experiment snapshots in Case #5,
where the pedestrians walk at faster speed.
To understand the motion behavior of pedestrians while
interacting with the robot, we further investigate the change
of trajectory and velocity of individual pedestrians. Fig. 7a
and Fig. 7b show the trajectories of the pedestrians and
the interacting robot in Case #2 and Case #3, respectively.
The circles and stars represent the position of the pedestrian
and the robot at every 1 second. The arrow represents
the velocity with the length proportional to the velocity
magnitude. From Fig. 7a, one can see that the pedestrian
adjusts his/her walking direction at around x = 4.5m to
avoid the robot as he/she foresees potential collision with
the robot. Meanwhile, the velocity of the pedestrian changes
slightly. One can see from Fig. 7b that the pedestrian adjusts
his/her walking direction at around x = 4.6m and x = 5.8m
respectively, and meanwhile slows down to avoid collision
with the robot. Fig. 7c shows the trajectories of the pedestrian
and the interacting robot in Case #6, from which one can see
the pedestrian adjusts his/her walking direction at around
x = 4.8m. The results demonstrate that both the trajectory
and velocity of the pedestrians can be affected by HRI.
B. Summary Statistics for Collective HRI
We have collected pedestrian trajectory data for all 6 experimental cases listed in Table I. One set of such trajectories
is plotted in Fig. 8a for 3 runs of Case #2. Each data point
on the trajectory has a timestamp. The velocity of each
pedestrian is calculated by performing a backward difference
on the position signal and then applying a moving average
filter to the resulting velocity signal.
To understand the effect of HRI on the collective motion of
pedestrians, we plot the average pedestrian velocity profile,
which is obtained by averaging all individual pedestrians’
velocities along the x axis for each case. Fig. 8b shows
the testing scenario of normal walking speed of pedestrians
(i.e., Cases #1, #2 and #3), and Fig. 8c shows the case
of the fast walking speed (i.e., Cases #4, #5 and #6). We
can see that the average pedestrian velocities are lowered
with the presence of robot; and the faster the robot moves,
the lower the average pedestrian velocity becomes. This
phenomenon is more visible around the robot area (i.e., the
vertical dashed line of robot positions shown in the figure)
than other areas (e.g., the beginning and the end of the
pedestrian trajectories), which indicates an HRI region exists
around the robot positions.
To further understand the collective pedestrian motion, we
plot the velocity distribution of pedestrians in each case in
the box plot of Fig. 9, where the central rectangle spans
the first quartile to the third quartile, the segment inside the
rectangle shows the median, and “whiskers” above and below
the box show the locations of the minimum and maximum.
We can clearly see the trends of HRI, that is, the first quartile,
the median, and the third quartile consistently show that: 1)
pedestrian velocities are lowered with the presence of robot,
and 2) the faster the robot moves, the lower the average
pedestrian velocity becomes.
C. Comparison with Simulations and Discussions
To compare the experimental results with the numerical simulation results presented in our previous work [13]
(Figures 4 and 5 therein), we calculate the pedestrian flow
velocity by averaging the velocities over all pedestrians in
each case in an HRI region, which is defined the same as in
the simulations (i.e., the rectangular regions within 2 meters
sideways from the robot path). Fig. 9c shows the average
(a) t=46.0s
(b) t=47.0s
(c) t=48.0s
(d) t=49.0s
Fig. 5: HRI behavior with normal walking speed of pedestrians (Case #2).
(a) t=39.0s
(b) t=39.6s
(c) t=40.2s
(d) t=40.8s
Fig. 6: HRI behavior with fast walking speed of pedestrians (Case #5).
3
0
0
1
2
3
4
5
2
Pedestrian trajectory
Pedestrian position at every 1 second
Pedestrian velocity vector
Robot trajectory
Robot position at every 1 second
Robot velocity vector
1
0
6
7
8
y position (m)
y position (m)
y position (m)
Pedestrian trajectory
Pedestrian position at every 1 second
Pedestrian velocity vector
Robot trajectory
Robot position at every 1 second
Robot velocity vector
1
Pedestrian trajectory
Pedestrian position at every 1 second
Pedestrian velocity vector
Robot trajectory
Robot position at every 1 second
Robot velocity vector
4
3
2
9
0
1
2
x position (m)
3
4
3
2
1
5
6
7
8
9
0
1
2
3
x position (m)
(a)
4
5
6
7
8
9
x position (m)
(b)
(c)
Fig. 7: Single pedestrian and robot interacting trajectories. (a) Case #2, run #2, pedestrian #10, t is from 47.1s to 51.1s; (b)
Case #3, run #2, pedestrian #1, t is from 40.0s to 44.0s; (c) Case #6, run #2, pedestrian #7, t is from 36.7s to 39.7s.
y position (m)
4
3
2
Case #1: No robot
Case #2: Slow-moving robot
Case #3: Fast-moving robot
2.5
5
velocity (m/s)
velocity (m/s)
3
2
1.5
Robot x position→
1
1
0
0
1
2
3
4
5
6
x position (m)
7
8
9
10
0
2
4
6
8
x position (m)
10
Case #4: No robot
Case #5: Slow-moving robot
Case #6: Fast-moving robot
4
3
Robot x position→
2
0
2
4
6
8
10
x position (m)
(a)
(b)
(c)
Fig. 8: (a)Trajectories recorded from all 3 runs of Case #2. The trajectory followed by each pedestrian is represented by a
solid curve. The vertical dotted line denotes the robot trajectory; Average pedestrian velocities vs. x position for (a) normal
walking speed of pedestrians, and (b) faster walking speed of pedestrians.
pedestrian speed with no-robot and in the presence of a
slow- and fast- moving robot in the two cases of normal
pedestrian speed and faster pedestrian speed. We can see that
the slow-down effect of the pedestrian-robot interaction is
more significant for the group of faster-moving pedestrians,
comparing with slower-moving pedestrians. In comparison
with simulations reported in our previous work [13], we can
clearly see that the trends of HRI match. Specifically, the
following claims are supported by both simulations and our
HRI experiments:
•
•
•
The pedestrian speeds are affected by the presence of a
moving robot;
A moving robot slows down a uni-directional flow. The
faster the robot moves, the lower the average pedestrian
velocity becomes;
The effect of the robot on the pedestrian velocity is
2
1.8
1.6
1.4
1.2
3.5
3
2.5
No Robot Slow-moving Fast-moving
robot
robot
No Robot Slow-moving Fast-moving
robot
robot
(a)
(b)
Average pedestrian velocity (m/s)
Pedestrian velocity (m/s)
Pedestrian velocities (m/s)
2.2
4
3.5
Normal walking speed
Faster walking speed
3
2.5
2
1.5
1
No Robot Slow-moving Fast-moving
robot
robot
(c)
Fig. 9: Box plot of pedestrian velocities for each case. (a) Normal walking speed; (b) Faster walking speed; (c) Average
pedestrian velocity comparison with and without the presence of a moving robot.
more significant when people walk at a faster speed.
For the tested scenarios, it clearly shows qualitative
agreement of HRI in an exit corridor in comparison with
numerical simulations based on social force models. The
results indicate that it is promising to use a moving robot
to regulate pedestrian flows, specifically, slowing down the
traffic to a desired average speed. Note that in evacuation
scenarios, being able to slow down the traffic and avoid the
faster-is-slower effect is often desirable.
With more time and resources, we would have collected
more experimental data. For example, we may vary the
pedestrian density of the area, and compare the HRI effect
with different density setups. Nevertheless, the obtained data
can be used to tune simulation models for a better match with
real world situations, which will be in the scope of our future
research. Also, we plan to conduct future HRI experiments to
validate learning-based robot-assisted pedestrian regulation
schemes that was presented in our early work [13].
IV. C ONCLUSION AND F UTURE W ORK
In this paper, we presented a human-robot interaction
experiment in an exit corridor. We investigated the pedestrian
behavior when they interact with a robot moving in a direction perpendicular to that of the uni-directional pedestrian
flow. Using the experimental data, we not only studied the
individual HRI behavior, but also analyzed collective HRI
and the effect of robot on the average pedestrian flow. We
compared the HRI effect on the pedestrian flow with that obtained from numerical simulations based on the social force
models that were reported in our early work [13]. We found
that the experimental HRI effect on the collective pedestrian
flow is qualitatively consistent with numerical simulations.
Future work includes calibrating simulation models using
experimentally obtained data, and developing learning-based
robot motion planner to regulate the collective speed of the
pedestrians to a desired level.
ACKNOWLEDGMENT
The authors would like to thank the undergraduate students, Peter Smith and Randall Devitt, and the graduate
student, Muhammad Fahad, for their assistance in collecting
data during the human robot experiments.
R EFERENCES
[1] T. Kruse, A. K. Pandey, R. Alami, and A. Kirsch, “Human-aware robot
navigation: A survey,” Robotics and Autonomous Systems, vol. 61,
no. 12, pp. 1726–1743, 2013.
[2] M. Luber, L. Spinello, J. Silva, and K. O. Arras, “Socially-aware
robot navigation: A learning approach,” in IEEE/RSJ International
Conference on Intelligent Robots and Systems, pp. 902–907, 2012.
[3] D. G. Macharet and D. A. Florencio, “Learning how to increase
the chance of human-robot engagement,” in IEEE/RSJ International
Conference on Intelligent Robots and Systems, pp. 2173–2179, 2013.
[4] M. Bennewitz, W. Burgard, G. Cielniak, and S. Thrun, “Learning motion patterns of people for compliant robot motion,” The International
Journal of Robotics Research, vol. 24, no. 1, pp. 31–48, 2005.
[5] P. Henry, C. Vollmer, B. Ferris, and D. Fox, “Learning to navigate
through crowded environments,” in IEEE International Conference on
Robotics and Automation, pp. 981–986, 2010.
[6] A. K. Pandey and R. Alami, “A framework for adapting social
conventions in a mobile robot motion in human-centered environment,”
in International Conference on Advanced Robotics, pp. 1–8, 2009.
[7] B. Kim and J. Pineau, “Socially adaptive path planning in human environments using inverse reinforcement learning,” International Journal
of Social Robotics, vol. 8, no. 1, pp. 51–66, 2016.
[8] A. Garrell and A. Sanfeliu, “Local optimization of cooperative robot
movements for guiding and regrouping people in a guiding mission,” in
IEEE/RSJ International Conference on Intelligent Robots and Systems,
pp. 3294–3299, 2010.
[9] G. Ferrer, A. Garrell, and A. Sanfeliu, “Robot companion: A socialforce based approach with human awareness-navigation in crowded
environments,” in IEEE/RSJ International Conference on Intelligent
robots and systems, pp. 1688–1694, 2013.
[10] J. A. Kirkland and A. A. Maciejewski, “A simulation of attempts
to influence crowd dynamics,” in IEEE International Conference on
Systems, Man and Cybernetics, vol. 5, pp. 4328–4333, 2003.
[11] B. D. Eldridge and A. A. Maciejewski, “Using genetic algorithms to
optimize social robot behavior for improved pedestrian flow,” in IEEE
International Conference on Systems, Man and Cybernetics, vol. 1,
pp. 524–529, 2005.
[12] K. Yamamoto and M. Okada, “Control of swarm behavior in crossing pedestrians based on temporal/spatial frequencies,” Robotics and
Autonomous Systems, vol. 61, no. 9, pp. 1036–1048, 2013.
[13] C. Jiang, Z. Ni, Y. Guo, and H. He, “Robot-assisted pedestrian
regulation in an exit corridor,” in IEEE/RSJ International Conference
on Intelligent Robots and Systems, pp. 815–822, 2016.
[14] M. Munaro, F. Basso, and E. Menegatti, “Openptrack: Open source
multi-camera calibration and people tracking for RGB-D camera
networks,” Robotics and Autonomous Systems, vol. 75, pp. 525–538,
2016.
| 3cs.SY
|
1
Efficient Systematic Encoding of
Non-binary VT Codes
arXiv:1708.04071v2 [cs.IT] 30 Aug 2017
Mahed Abroshan, Ramji Venkataramanan and Albert Guillén i Fàbregas
Abstract
Varshamov-Tenengolts (VT) codes are a class of codes which can correct a single deletion or
insertion with a linear-time decoder. This paper addresses the problem of efficient encoding of nonbinary VT codes, defined over an alphabet of size q > 2. We propose a simple linear-time encoding
method to systematically map binary message sequences onto VT codewords. The method provides a
new lower bound on the size of q-ary VT codes of length n.
I. I NTRODUCTION
Designing codes for correcting deletions or insertions is well known to be a challenging
problem; see, e.g., [1]–[8]. For the special case of correcting one insertion or deletion, there
exists an elegant class of codes called Varshamov-Tenengolts (VT) codes. Binary VT codes
were first introduced by Varshamov and Tenengolts in [9] for channels with asymmetric errors.
Later, Levenshtein [10] showed that they can be used for correcting a single deletion or insertion
with a simple decoding algorithm whose complexity is linear in the code length [1]. Tenengolts
subsequently introduced a non-binary version of VT codes, defined over a q-ary alphabet for any
q > 2 [11]. The q-ary VT codes retain many of the attractive properties of the binary codes. In
particular, they can correct deletion or insertion of a single symbol from a q-ary VT codeword
with a linear-time decoder.
M. Abroshan and R. Venkataramanan are with the Department of Engineering, University of Cambridge, UK,
([email protected], [email protected]).
A. Guillén i Fàbregas is with the Department of Information and Communication Technologies, Universitat Pompeu Fabra,
Barcelona 08018, Spain, also with the Institució Catalana de Recerca i Estudis Avançats (ICREA), Barcelona 08010, Spain, and
also with the Department of Engineering, University of Cambridge, Cambridge CB2 1PZ, U.K. (e-mail: [email protected]).
This work was supported in part by the European Research Council under Grants 259663 and 725411, in part by the Spanish
Ministry of Economy and Competitiveness under Grant Grant TEC2016-78434-C3-1-R.
2
Given the simplicity of VT decoding, a natural question is: can one construct a linear-time
encoder to efficiently map binary message sequences onto VT codewords? For binary VT codes,
such an encoder was proposed by Abdel-Ghaffar and Ferriera [12]. (A similar encoder was also
described in [13].) However, the issue of efficient encoding for non-binary VT codes has not
been addressed previously, to the best of our knowledge. In this paper, we propose an efficient
systematic encoder for non-binary VT codes. The encoder has complexity that is linear in the
code length, and is systematic in the sense that the message bits are assigned to pre-specified
positions in the codeword. The encoder also yields a new lower bound on the size of q-ary VT
codes, for q > 2.
VT codes are a key ingredient of recent code constructions for channels with segmented
deletions and insertions [14], [15], and the proposed VT encoder can be applied to these
constructions. VT codes have also recently been used in algorithms for synchronization from
deletions and insertions, e.g., [16]–[19].
In [11, Sec. 5], Tenengolts introduced a systematic non-binary code that can correct a single
deletion or insertion. However, this code is not strictly a VT code as its codewords do not
necessarily share the same VT parameters. (Formal definitions of VT codes and their parameters
are given in the next two sections.) In this paper, we propose an encoder for VT codes defined
in the standard way, noting that using standard VT codes is a key requirement in some of the
code constructions mentioned above.
Notation: Sequences are denoted by capital letters, and scalars by lower-case letters. Throughout, we use n for the code length and k for the number of message bits mapped to the codeword.
The set Zq = {0, 1, · · · , q − 1} is the finite integer ring of size q. We consider the natural order
for the elements of Zq , i.e., 0 < 1 . . . < (q − 1). The term dyadic index will be used to refer to
an index that is a power of 2.
The rest of the paper is organized as follows. In the next section, we formally define binary
VT codes and briefly review the systematic encoder from [12]. In Section III, we define the
q-ary VT codes, and describe the systematic encoding method and the resulting lower bound on
the size of the codes.
3
II. B INARY VT
CODES
The VT syndrome of a binary sequence S = s1 s2 · · · sn ∈ Zn2 is defined as
syn(S) ,
n
X
i si
mod (n + 1).
(1)
i=1
For positive integers n and 0 ≤ a ≤ n, the VT code of length n and syndrome a, is defined as
VTa (n) = S ∈ Zn2 : syn(S) = a ,
(2)
i.e., the set of binary sequences S of length n that satisfy syn(S) = a. For example, the VT
code of length 3 and syndrome 2 is
3
n
o
X
3
VT2 (3) = s1 s2 s3 ∈ Z2 :
j sj = 2 mod 4
j=1
(3)
= {010, 111}.
Each of the sets VTa (n), 0 ≤ a ≤ n, is a code that can correct a single deletion or insertion
with a decoder whose complexity is linear in n. The details of the decoding algorithm can be
found in [1].
The (n + 1) sets VTa (n), 0 ≤ a ≤ n, partition the set of all binary sequences of length n, i.e.,
each sequence S ∈ Zn2 belongs to exactly one of the sets. Therefore, the smallest of the codes
VTa (n) will have at most
2n
n+1
min
sequences. Hence
0≤a≤n
1
1
log2 |VTa (n)| ≤ 1 − log2 (n + 1).
n
n
(4)
An exact formula for the size |VTa (n)| was given by Ginzburg [20]. The formula [1, Theorem
2.2] does not give an analytical expression for general n, but it shows that if (n + 1) is a power
of 2, then |VTa (n)| = 2n /(n + 1), for 0 ≤ a ≤ n. Moreover, for general n, the formula can be
used to deduce that the sizes of the codes VTa (n) are all approximately 2n /(n + 1). In particular,
2n
2n
− 2(n+1)/3 ≤ |VTa (n)| ≤
+ 2(n+1)/3 ,
(n + 1)
(n + 1)
for a ∈ {0, . . . , n}.
(5)
Abdel-Ghaffar and Ferriera [12] proposed a systematic encoder to map k-bit message sequences onto codewords in |VTa (n)|, where k = n − ⌈log2 (n + 1)⌉. We briefly review the
encoding procedure as it is an ingredient of our systematic q-ary VT encoder. Consider a k-bit
message M = m1 · · · mk to be encoded into a codeword C = c1 · · · cn ∈ VTa (n), for some
a ∈ {0, 1, · · · , n}. The number of “parity” bits is denoted by t = n − k = ⌈log2 (n + 1)⌉. The
4
idea is to use the code bits in dyadic positions, i.e., c2i , for 0 ≤ i ≤ (t − 1), to ensure that
syn(C) = a. The encoding steps are:
1) Denote the first k non-dyadic indices by {j1 , · · · , jk }, where the indices are in ascending
order, i.e., j1 = 3, j2 = 5, . . . We set cji equal to the message bit mi , for 1 ≤ i ≤ k.
2) First set the bits in all the dyadic positions to be zero and denote the resulting sequence
by C ′ = c′1 · · · c′n (so that we have c′2i = 0 for 0 ≤ i ≤ t − 1 and c′jl = ml for 1 ≤ l ≤ k =
(n − t)). Define the deficiency d as the difference between the desired syndrome a and the
syndrome of C ′ . That is,
d = a − syn(C ′ ) mod (n + 1).
3) Let the binary representation of d be dt−1 . . . d1 d0 , i.e., d =
0 ≤ i ≤ (t − 1), to obtain C.
(6)
Pt−1
i=0
2i di . Set c2i = di , for
The rate of this systematic encoder is R = 1 − n1 ⌈log2 (n + 1)⌉, regardless of the syndrome
a ∈ {0, . . . , n}. Comparing with (4), we observe that the rate loss for the smallest VT code of
length n is less than
1
.
n
On the other hand, if (n + 1) is not a power of two, the rate loss for
the larger VT codes may be higher due to codewords that are unused by the encoder. However
this rate loss is unavoidable with any systematic encoder [12].
We remark that the dyadic positions are not the only set of positions that can be used for
syndrome bits. For instance, the following set of indices also produce all syndromes:
{ci0 , · · · , cit−1 } where
ij = −2j
(mod n + 1) for 0 ≤ j ≤ t − 1.
This can be helpful in some applications (see e.g. [15]) where some code bits are already reserved
for prefixes or suffixes, and thus cannot be used as syndrome bits. In general, a set of positions
{p1 , p2 , · · · , pr } can be used for syndrome bits if for each syndrome a ∈ {0, · · · , n}, there exists
a subset P ⊆ {p1 , p2 , · · · , pr } such that
X
pj = a (mod n + 1).
(7)
j∈P
In other words, for each a ∈ {0, · · · , n}, there should exist binary coefficients b1 , · · · , br such
that
r
X
j=1
bj pj = a
(mod n + 1).
(8)
5
III. E FFICIENT E NCODING
FOR
N ON - BINARY VT
CODES
For any code length n, the VT codes over Zq , for q > 2 are defined as follows [11]. For each
q-ary sequence S = s0 s1 · · · sn−1 ∈ Znq , define a corresponding length (n − 1) auxiliary binary
sequence AS = α1 α2 . . . αn−1 as follows1. For 1 ≤ i ≤ n − 1,
1
if si ≥ si−1
αi =
0
if si < si−1 .
(9)
We also define the modular sum of S as
sum(S) =
n−1
X
si
(mod q).
(10)
i=0
For 0 ≤ a ≤ n − 1 and b ∈ Zq , the q-ary VT code with length n and parameters (a, b) is defined
as
VTa,b (n) = S ∈ Znq : syn(AS ) = a, sum(S) = b .
(11)
Each of the sets VTa,b (n) is a code that can correct deletion or insertion of a single symbol
with a decoder whose complexity is linear in the code length n. The details of the decoding
algorithm can be found in [11, Sec. II].
Similarly to the binary case, the codes VTa,b (n), for 0 ≤ a ≤ n − 1 and b ∈ Zq , partition the
space Znq of all q-ary sequences of length n. For a given n, there are nq of these codes, and
hence the smallest of them will have no more than
qn
nq
sequences. Let Rmin be the rate of the
smallest of these codes, i.e.,
Rmin , min
a,b
log2 |VTa,b (n)|
,
n
(12)
where the minimum is over 0 ≤ a ≤ n − 1 and b ∈ Zq . We then have the bound
Rmin ≤ log2 q −
1
1
log2 n − log2 q bits/symbol.
n
n
(13)
The encoding procedure described below yields a lower bound on the size of |VTa,b (n)| (see
Proposition 1), which shows that for q ≥ 4,
1
1
Rmin ≥ log2 q− ⌈log2 n⌉(3 log2 q−2 log2 (q−1))− (5 log2 (q−1)−3 log 2 q) bits/symbol. (14)
n
n
1
For non-binary sequences, we start the indexing from 0 as this makes it convenient to describe the encoding procedure in
Section III.
6
Kulkarni and Kiyavash [21] have shown that the size of any single deletion correcting q-ary
code of length n is bounded by
q n −q
.
(q−1)(n−1)
This yields a rate upper bound Rmax for any single
deletion correcting code, where
Rmax ≤ log2 q −
log2 (n − 1) log2 (q − 1)
−
.
n
n
(15)
We now describe the encoder to map a sequence of message bits to a codeword of the q-ary
VT code VTa,b (n). For simplicity, we first assume that q is a power of two, and address the case
of general q at the end of this section. We will map a k-bit message M = m1 m2 · · · mk to a
codeword in VTa,b (n), where
k = (n − 3t + 3) log2 q + (t − 3)(2 log2 q − 1) + (log2 q − 1)
= n log2 q − t(log2 q + 1) − 2(log2 q − 1),
(16)
(17)
with t = ⌈log2 n⌉. Therefore, the rate of our encoding scheme is
R = log2 q −
⌈log2 n⌉(log2 q + 1) 2 log2 q − 2
−
bits/symbol.
n
n
(18)
Our encoding method gives a lower bound on the size of any non-binary VT code of length
n. An immediate lower bound on the size is 2k , with k given by (17). The proposition below
gives a slightly better bound, which is obtained by modifying the encoding method to map q-ary
message sequences to q-ary VT codewords, rather than a binary message sequence to a q-ary
VT codeword.
Proposition 1. For n ≥ 6, q ≥ 4, and any 0 ≤ a ≤ n, and b ∈ Zq , we have
|VTa,b (n)| ≥ (q − 1)2t−5 q n−3t+3 ,
= q n(1− n [3 log2 q−2 log2 (q−1)]− n [5 log2 (q−1)−3 log2 q])
t
1
where t = ⌈log2 n⌉.
The proof of the proposition is given at the end of this section, after describing the encoding
procedure. We emphasize that we use t = ⌈log2 n⌉ throughout this section (as opposed to
⌈log2 (n + 1)⌉ used for binary VT encoding) because the binary auxiliary sequence has length
(n − 1).
7
A. Encoding procedure
The high level idea for mapping a k-bit message to a codeword C ∈ VTa,b (n) is the following.
Similarly to the binary case, we reserve the t dyadic positions in the binary auxiliary sequence
AC to ensure that syn(AC ) = a. Recall from (9) that each bit of AC is determined by comparing
two adjacent symbols of the q-ary sequence C. Therefore, to ensure that syn(AC ) = a we not
only reserve the symbols in the dyadic positions of C, but also place some restrictions on the
symbols adjacent to the dyadic positions. Finally, we use the first three symbols of C to ensure
that sum(C) = b. We explain the method in six steps with the help of the following running
example.
Example 1. Let q = 8, n = 16, and suppose that we wish to encode a binary message M to a
codeword C in VT0,1 (16). We have t = ⌈log2 n⌉ = 4 and log2 q = 3. Therefore, from (17) the
length of M is k = 3(n − 3t + 3) + 5(t − 3) + 2 = 25 bits. Let
M = 110 001 000 111 010 101 000 11100 11.
(19)
Step 1. Let S be the set of pairs of symbols adjacent to a dyadic symbol, i.e.,
S = {(c2j −1 , c2j +1 ), for 2 ≤ j ≤ (t − 1)}.
(20)
There are |S| = (t − 2) pairs of symbols in S. Excluding c0 , the number of symbols in C that
are neither in dyadic positions nor in S is
(n − 1) − 2|S| − t = (n − 3t + 3).
(21)
Assign the first (n − 3t + 3) log2 q bits of the message M to these symbols, by converting each
set of log2 q bits to a q-ary symbol. This corresponds to the first term in (16).
In Example 1, (n − 3t + 3) log2 q = 18, and the representation of first 18 bits of M in Z8 is
6 1 0 7 2 5 0. Therefore the sequence C is
C = c0 c1 c2 c3 c4 c5 6 c7 c8 c9 1 0 7 2 5 0.
(22)
Step 2. In this step, we assign the remaining bits of the message to the symbols in S. For
a given dyadic position c2j , j = 2, 3, · · · , (t − 1), we constrain the pair of adjacent symbols
(c2j −1 , c2j +1 ) to belong to the following set
T = {(r, l) ∈ Zq × Zq : r 6= 0, l 6= (r − 1)}.
(23)
8
Via (23), we enforce c2j −1 6= 0 because if c2j −1 were 0, then we necessarily have c2j ≥ c2j −1
which constrains the value of a2j to 1. However, a2j needs to be unconstrained in order to
guarantee that any desired syndrome can be generated. Furthermore, we will see in Step 5 that if
c2j +1 = c2j −1 −1, then we may be unable find a suitable symbol c2j compliant with the restrictions
induced by the auxiliary sequence. We therefore enforce the constraint c2j +1 6= c2j −1 − 1 using
(23). It is easy to see that |T | = (q − 1)2 .
Excluding the pair (c3 , c5 ), there are (t − 3) pairs in S. If we were encoding q-ary message
symbols, each of these (t − 3) pairs could take any pair of symbols in T . Since we are encoding
message bits, we use a look up table to map ⌊log2 |T |⌋ = 2 log2 q − 1 bits to each of the pairs in
S excluding (c3 , c5 ). We thus map (t − 3)(2 log2 q − 1) bits to the pairs in S excluding (c3 , c5 ).
This corresponds to the second term in (16).
Next, set c3 = q − 1. This choice is important as it will facilitate step 6. Since (c3 , c5 ) ∈ T ,
when c3 = q − 1, then c5 has to be such that c5 6= q − 2. Hence, there are q − 1 possible values
for c5 . As we are encoding a binary message, we map ⌊log2 (q − 1)⌋ = log2 q − 1 bits to c5 using
a look-up table. This corresponds to the third term in (16). We note that the two look-up tables
used in this step have sizes at most (q − 1)2 and q, respectively. Thus, in steps one and two in
total we have mapped the claimed k message bits to the symbols of C.
In Example 1, as seen from (22), (c7 , c9 ) is the only pair in S different from (c3 , c5 ). We can
assign 2 log2 q − 1 = 5 bits to (c7 , c9 ). After the first 18 message bits mapped in Step 1, the next
five bits in M are 11100. Suppose that in our look-up table these bits correspond to the pair
(3, 5). We then have (c7 , c9 ) = (3, 5). Also, we fix c3 = q − 1 = 7, and the last two message
bits determine c5 . The last two message bits are 11. Suppose that 3 is the corresponding symbol
in the look-up table. We therefore set c5 = 3. Therefore, we have
C = c0 c1 c2 7 c4 3 6 3 c8 5 1 0 7 2 5 0.
(24)
Up to this point, we have mapped our k message bits to a partially filled q-ary sequence. In the
following steps we ensure that the resulting sequence lies in the correct VT code by carefully
choosing remaining (t+1) symbols to obtain the auxiliary sequence syndrome a and the modular
sum b.
Step 3. In this step, we specify the bits in the non-dyadic locations of the auxiliary sequence
AC . Notice that according to (9), in order to define a2j +1 , the value of c2j should be known. This
is not the case here as the dyadic positions in C have been reserved to generate the required
9
syndrome. To circumvent this issue, we determine a2j +1 (for 1 < j < t) by comparing c2j +1
with c2j −1 as follows:
a2j +1 =
1
0
if c2j +1 ≥ c2j −1 ,
(25)
if c2j +1 < c2j −1 .
As we shall show in step 5, we will be able to make these choices for the auxiliary sequence
compatible with the definition of a valid auxiliary sequence in (9).
Next, since we have chosen c3 = q − 1, from the rule in (9) we have a3 = 1, regardless of
what c2 is. The other non-dyadic positions of the auxiliary sequence AC can be filled in using
(9), i.e., ai = 1 if ci ≥ ci−1 , and 0 otherwise.
For our example with C shown in (24), at the end of this step we have
AC = a1 a2 1 a4 0 1 0 a8 1 0 0 1 0 1 0.
(26)
Step 4. In this step, we use the binary encoding method described in Section II to find the bits
in the dyadic positions a20 , · · · , a2t−1 such that syn(AC ) = a. With this, the auxiliary sequence
AC is fully determined.
In the example, we need to find a1 , a2 , a4 and a8 such that syn(AC ) = 0. First, we set the
syndrome bits a1 = a2 = a4 = a8 = 0, and denote the resulting sequence by
A′C = 0 0 1 0 0 1 0 0 1 0 0 1 0 1 0.
(27)
Now, syn(A′C ) = 12, and the deficiency d = 0 − 12 (mod 6) = 4. The binary representation of
d is d3 d2 d1 d0 = 0100. Hence, a1 = a2 = a8 = 0 and a4 = 1 will produce the desired syndrome.
Summarizing, we have the following auxiliary sequence
AC = 0 0 1 1 0 1 0 0 1 0 0 1 0 1 0.
(28)
Step 5. In this step, we specify the symbols of C in the dyadic positions (except c1 and c2 ).
This will be done by ensuring that AC is a valid auxiliary sequence consistent with the definition
in (9). In particular, the choice of c2j for j = 2, · · · , t − 1, should be consistent with a2j +1 and
a2j . We ensure this by choosing c2j (for 1 < j < t) as follows:
c2j −1 − 1
if a2j = 0,
c2j =
c2j −1
if a2j = 1.
(29)
From the definition in (9), this choice is consistent with a2j . Now we show that it is also
consistent with a2j +1 . If a2j +1 = 1, then according to (25), c2j +1 ≥ c2j −1 ; then the choice of c2j
10
in (29) always guarantees that c2j −1 ≥ c2j , and thus c2j +1 ≥ c2j . Next suppose that a2j +1 = 0.
Then according to (25), c2j +1 < c2j −1 . We need to verify that c2j +1 < c2j Now, if a2j = 1, then
c2j = c2j −1 and c2j +1 < c2j . Also, if a2j = 0, from (29) we have c2j = c2j −1 − 1. Since symbols
adjacent to dyadic positions (c2j −1 , c2j +1 ) are chosen from T (see step 2), then c2j +1 6= c2j −1 −1.
Thus, we have that c2j +1 < c2j −1 − 1 = c2j . Therefore, in either case the choice is consistent
with (9).
For the example, using (29) and (28) we obtain
C = c0 c1 c2 7 7 3 6 3 2 5 1 0 7 2 5 0.
(30)
Step 6. Finally, we need to find c0 , c1 and c2 that are compatible with a1 , a2 , a3 (the first three
bits of the auxiliary sequence), and such that sum(C) = b. Let
n
X
w ,b−
ci (mod q).
(31)
i=3
Hence we need c0 + c1 + c2 = w (mod q). We will show that when q ≥ 4, we can find three
distinct integers (x, y, z) such that 0 ≤ x < y < z < q and x + y + z = w (mod q). We will
assign these numbers to c0 , c1 and c2 . Also recall that we set c3 = q − 1; hence we always have
x, y, z ≤ c3 , which is consistent with a3 = 1.
The triplet with smallest numbers that we can choose is x = 0, y = 1, z = 2. For this choice,
w = 0 + 1 + 2 = 3 (mod q). By increasing z from 2 to q − 1 with x = 0 and y = 1, we can
produce any value of w from 3 to q − 1 as well as w = 0. Finally, the only remaining values
are w = 1, 2. To obtain these values, we choose x, y, z as follows.
1) w = 1: Choose x = 0, y = 2, z = q − 1.
2) w = 2: Choose x = 1, y = 2, z = q − 1.
Hence, we have shown that for q ≥ 4, any w ∈ Zq can be expressed as the (mod q) sum
of three distinct elements of Zq . Assigning these elements to c0 , c1 , c2 in the order required
by the auxiliary sequence completes the encoding procedure. We now have sum(C) = b and
syn(AC ) = a, and thus C ∈ VTa,b (n) as required.
In our example, from (30) we have
15
X
ci = 48 = 0 (mod 8),
(32)
i=3
and b = 1. Therefore we need c0 + c1 + c2 = 1 (mod 8). We have a1 = a2 = 0 so c0 > c1 > c2
is the correct order. We therefore assign c0 = 7, c1 = 2, and c2 = 0 to obtain the codeword.
C = 7 2 0 7 7 3 6 3 2 5 1 0 7 2 5 0 ∈ VT0,1 (16).
(33)
11
It can be verified that sum(C) = 1, and the auxiliary sequence syndrome syn(AC ) = 0.
B. The case where q is not a power of two
When log2 q is not an integer, the main difference is that we map longer sequences of bits to
sequences of q-ary symbols. Recall that in step 1, we determine (n − 3t + 3) symbols of the
q-ary codeword. One can map ⌊(n − 3t + 3) log2 q⌋ bits to these (n − 3t + 3) symbols using
standard methods to convert an integer expressed in base 2 into base q. In the second step,
as described earlier we can map ⌊log2 (q − 1)2 ⌋ bits to (t − 3) pairs in S (excluding (c3 , c5 )).
Moreover ⌊log2 (q − 1)⌋ bits can be mapped to c5 . Therefore, in total we can map k bits to a
q-ary VT codeword of length n, where
k = ⌊(n − 3t + 3) log2 q⌋ + (t − 3)⌊log2 (q − 1)2 ⌋ + ⌊log2 (q − 1)⌋
≥ n log2 q − t(log2 q + 2) − (2 log2 q − 4).
(34)
(35)
For q ≥ 4, the remaining steps are identical to the case where q is a power of two. The case of
q = 3 is slightly different and it is discussed in Appendix A.
C. Proof of Proposition 2
The result can be directly derived from steps one and two of our encoding method by mapping
sequences of q-ary message symbols (rather than sequences of message bits) to distinct codewords
in |VTa,b (n)|. In step 1, we can assign (n − 3t + 3) arbitrary symbols to positions that are neither
dyadic nor in S. There are q n−3t+3 ways to choose these symbols. Then in step two, we can
choose (q − 1)2 pairs for each of the (t − 3) specified pairs of positions; furthermore, there are
(q − 1) choices for c5 . According to steps 3 to 6, we can always choose the remaining symbols
such that resulting codeword lies in VTa,b (n). Therefore, we can map q n−3t+3 (q −1)2t−5 different
sequences of message symbols to distinct codewords in VTa,b (n). This yields the lower bound
on |VTa,b (n)|.
A PPENDIX
A. Encoding for q = 3
For q = 3, we need to slightly modify the proposed algorithm. The first step is as described in
Section III-B. The difference in the second step is that we do not embed data in c5 and simply
12
choose c5 = c3 = 2. Steps three to five remain the same. In the sixth step, we compute w as in
(31), and choose c0 , c1 , c2 as follows depending on the values of a0 and a1 :
1) a1 = a2 = 1: Choose c2 = c1 = 2 and c0 = w − 4 (mod 3).
2) a1 = 1, a2 = 0: Choose c2 = 1 and c1 = 2 and c0 = w − 4 (mod 3).
3) a1 = 0, a2 = 1: Choose c2 = 2. If w = 1, then c1 = 0, c0 = 2. If w = 0, then c1 = 0, c0 = 1.
If w = 2, then c1 = 1, c0 = 2.
The only remaining case is when a1 = a2 = 0. For this case, we need to change c3 and c4 , and
also the first three bits of AC . Since c3 has been set to 2, the first three bits of AC in this case
are 001. If we change these three bits to 110, syn(AC ) will remain unchanged. We therefore set
a1 = a2 = 1 and a3 = 0. Now we update c0 , c1 , c2 , c3 to be compatible with the new auxiliary
sequence. Set c3 = 1, recall that c5 = 2 so we still have c5 ≥ c3 and hence this change will not
affect a5 . Update c4 according to (29). Set c2 = c1 = 2, and c0 = w − 4 (mod 3). Now we have
c3 < c2 which is consistent with a3 = 0. Also c2 ≥ c1 ≥ c0 is consistent with a1 = a2 = 1.
Hence, for q = 3, we have mapped k = ⌊log2 3(n−3t+ 3)⌋+ 2(t−3) bits to a q-ary codeword
C. This induces following rate:
⌊log2 3 (n − 3⌈log2 n⌉ + 3)⌋ 2(⌈log2 n⌉ − 3)
+
n
n
2.76 ⌈log2 n⌉ 2.25
−
≥ log2 3 −
n
n
R=
(36)
(37)
Similarly to Proposition 2, we can show that for q = 3 there are at least 22(t−3) 3n−3t+3 codewords
in each of the VT codes.
Acknowledgement
The authors thank Andreas Lenz for pointing them to the references [12] and [13].
R EFERENCES
[1] N. J. A. Sloane, “On single-deletion-correcting codes,” in Codes and Designs, Ohio State University (Ray-Chaudhuri
Festschrift), pp. 273–291, 2000. Online: https://arxiv.org/abs/math/0207197.
[2] M. C. Davey and D. J. C. MacKay, “Reliable communication over channels with insertions, deletions, and substitutions,”
IEEE Trans. Inf. Theory, vol. 47, no. 2, pp. 687–698, 2001.
[3] E. A. Ratzer, “Marker codes for channels with insertions and deletions,” Ann. Telecommn., vol. 60, no. 1, pp. 29–44, 2005.
[4] K. A. Abdel-Ghaffar, F. Paluncic, H. C. Ferreira, and W. A. Clarke, “On Helberg’s generalization of the Levenshtein code
for multiple deletion/insertion error correction,” IEEE Trans. Inf. Theory, vol. 58, no. 3, pp. 1804–1808, 2012.
[5] T. A. Le and H. D. Nguyen, “New multiple insertion/deletion correcting codes for non-binary alphabets.,” IEEE Trans.
Inf. Theory, vol. 62, no. 5, pp. 2682–2693, 2016.
13
[6] J. Brakensiek, V. Guruswami, and S. Zbarsky, “Efficient low-redundancy codes for correcting multiple deletions,” in Proc.
ACM-SIAM Symp. on Discrete Algorithms, pp. 1884–1892, 2016.
[7] V. Guruswami and C. Wang, “Deletion codes in the high-noise and high-rate regimes,” IEEE Trans Inf. Theory, vol. 63,
no. 4, pp. 1961–1970, 2017.
[8] S. K. Hanna and S. E. Rouayheb, “Guess & check codes for deletions and synchronization,” in Proc. IEEE Int. Symp. Inf.
Theory, 2017.
[9] R. R. Varshamov and G. M. Tenengolts, “Codes which correct single asymmetric errors,” Automatica i Telemekhanica,
vol. 26, no. 2, pp. 288–292, 1965.
[10] V. I. Levenshtein, “Binary codes capable of correcting deletions, insertions and reversals,” Doklady Akademii Nauk SSSR,
vol. 163, no. 4, pp. 845–848, 1965.
[11] G. Tenengolts, “Nonbinary codes, correcting single deletion or insertion,” IEEE Trans. Inf. Theory, vol. 30, no. 5, pp. 766–,
1984.
[12] K. A. S. Abdel-Ghaffar and H. C. Ferreira, “Systematic encoding of the Varshamov-Tenengolts codes and the ConstantinRao codes,” IEEE Trans Inf. Theory, vol. 44, pp. 340–345, Jan 1998.
[13] K. Saowapa, H. Kaneko, and E. Fujiwara, “Systematic deletion/insertion error correcting codes with random error correction
capability,” in Int. Symp. Defect and Fault Tolerance in VLSI Systems, pp. 284–292, Nov. 1999.
[14] Z. Liu and M. Mitzenmacher, “Codes for deletion and insertion channels with segmented errors,” IEEE Trans Inf. Theory,
vol. 56, no. 1, pp. 224–232, 2010.
[15] M. Abroshan, R. Venkataramanan, and A. Guillén i Fàbregas, “Codes for channels with segmented edits,” in Proc. IEEE
Int. Symp. Inf. Theory, 2017. Extended version: http://arxiv.org/abs/1701.06341.
[16] R. Venkataramanan, V. N. Swamy, and K. Ramchandran, “Low-complexity interactive algorithms for synchronization from
deletions, insertions, and substitutions,” IEEE Trans Inf. Theory, vol. 61, no. 10, pp. 5670–5689, 2015.
[17] S. M. S. Tabatabaei Yazdi and L. Dolecek, “Synchronization from deletions through interactive communication,” IEEE
Trans. Inf. Theory, vol. 60, pp. 397–409, Jan. 2014.
[18] F. Sala, C. Schoeny, N. Bitouzé, and L. Dolecek, “Synchronizing files from a large number of insertions and deletions,”
IEEE Trans. Commun., vol. 64, no. 6, pp. 2258–2273, 2016.
[19] M. Abroshan, R. Venkataramanan, and A. Guillén i Fàbregas, “Multilayer codes for synchronization from deletions,” in
To appear in Proc. IEEE Inf. Theory Workshop, 2017. Online: https://arxiv.org/abs/1705.06670.
[20] B. D. Ginzburg, “A number-theoretic function with an application in the theory of coding,” Problemy Kibernet., pp. 249–
252, 1967.
[21] A. A. Kulkarni and N. Kiyavash, “Nonasymptotic upper bounds for deletion correcting codes,” IEEE Trans Inf. Theory,
vol. 59, pp. 5115–5130, Aug. 2013.
| 7cs.IT
|
arXiv:1701.00821v1 [cs.DS] 31 Dec 2016
Partially Recursive Acceptance Rejection
Mark Huber
Claremont McKenna College
[email protected]
Version: January 5, 2017
Abstract
Generating random variates from high-dimensional distributions is often
done approximately using Markov chain Monte Carlo. In certain cases, perfect
simulation algorithms exist that allow one to draw exactly from the stationary
distribution, but most require O(n ln(n)) time, where n measures the size of the
input. In this work a new protocol for creating perfect simulation algorithms
that runs in O(n) time for a wider range of parameters on several models (such
as Strauss, Ising, and random cluster) than was known previously. This work
represents an extension of the popping algorithms due to Wilson.
1
Introduction
Partially recursive acceptance rejection (PRAR) is a new protocol for creating algorithms for exactly generating random variates from high dimensional distributions.
The method is simple to implement efficiently, and results in algorithms that can be
proved to have an expected running time that is linear in problem size over a wider
range of parameters than known previously for several problems of interest.
Consider a distribution defined either by an unnormalized weight function w(x)
for discrete spaces, or an unnormalized density function f (x) for continuous spaces.
In the past, techniques for perfect simulation from these distributions have been very
different depending on whether the spaces were continuous or discrete. PRAR operates
the same way in both situations.
The problems where PRAR is useful often use approximate sampling via Markov
chain Monte Carlo (MCMC). However, these are not true algorithms unless the mixing
time of the Markov chain being used can be somehow bounded. One such technique
for bounding the mixing time of a Markov chain is Dobrushin uniqueness [3] (see
also the work on path coupling of Bubley and Dyer [2]). PRAR requires a condition
very similar to Dobrushin uniqueness, the difference being that even if Dobrushin
uniqueness cannot be shown mathematically, it can be verified algorithmically.
One can prove (up to an arbitrarily small chance of error) that a Markov chain is
slowly mixing through computer experiment, however, there is no generally effective
way to show that a Markov chain is rapidly mixing. However, with perfect simulation
protocols like PRAR, it is possible through computer experiment to verify that the
resulting output comes exactly from the target distribution, thereby making them
more useful in practice for certain problems than the Markov chain approach.
Our concern here is with fast algorithms that use an expected number of random
choices that is linear in the size of the problem. There are currently two protocols for
generating samples in expected linear time from these types of problems. The first
method is the Clan of Ancestors method of [5]. This method is an extension of the
Coupling from the Past [16] protocol or Propp and Wilson. It looks backwards in time
at the events that could affect the state in the present. The set of backwards events
is called the Clan of Ancestors, which gives the name to the method. This approach
has rarely been implemented (see [10] for one such implementation) because unless
special care is taken it can be very computationally expensive to keep track of how
clans arising from different dimensions interact. One of the advantages of PRAR is
that such interactions cannot occur with this approach, making the implementation
much easier.
The second linear time method is the Randomness Recycler method of [11, 6]. Like
PRAR (and unlike Clan of Ancestors) it is straightforward to implement, however, the
range of parameters for which the algorithm is provably linear time is in all known
cases more restricted than with PRAR.
The PRAR protocol is illustrated with the following applications.
• For independent sets of a graph with n nodes, maximum degree ∆, and parameter λ, the method allows O(n) expected sampling time when λ < 1.13/(∆ − 2),
beating the previously best known linear time algorithm which required λ <
1/(∆ − 1).
• For the continuous state space autonormal model on a graph with n nodes
and parameter β, the method allows O(n) expected sampling time when β <
(∆−1) exp(−β), beating the previously best known linear time algorithm which
required β < ∆ exp(−β).
• For the random cluster model on a graph with n nodes and maximum degree
∆, and parameters p and q, the method allows O(n) expected sampling time
when p < 1/∆.
Of course, if the expected running time is allowed to be O(n ln(n)) rather than
linear, other methods with wider parameter ranges exist. For instance, in [11] it
was shown using bounding chains and coupling from the past [16] how to sample
independent sets in O(n ln(n)) time when λ < 2/(∆ − 2). But for O(n) algorithms,
PRAR appears to have the widest bounds.
2
The rest of the paper is organized as follows. The next section presents the applications, followed by a section describing the theory of the protocol. Then Section 4
applies the protocol to the different applications, proving the running time results
listed above. Section 5 then shows how this method can be viewed as a generalization
of the popping method of Wilson for sampling directed rooting spanning trees of a
graph.
2
Applications
Each of the applications will be described as an unnormalized density w(x) with
respect to a base finite measure µ from which it is easy to sample from.
2.1
Independent Sets of a Graph
Consider coloring the nodes of a graph (V, E) with either 0 or 1, so the state space
is Ω = {0, 1}V . The underlying measure we use has a parameter λ ≥ 0 that controls
the average number of nodes in a randomly drawn coloring:
µ({x}) = λ
P
i
x(i)
.
(1)
Say that X is a Bernoulli random variable with mean p (write X ∼ Bern(p)) if P(X =
1) = p and P(X = 0) = 1 − p. To sample from µ, generate X(v) ∼ Bern(λ/(1 + λ))
independently for all v ∈ V .
We use the convention 00 = 1 so that when λ = 0 the only state with positive
measure is the all 0 coloring, while as λ grows the measure favors labelings with more
1’s.
An independent set of a graph is a labeling such that no two adjacent nodes receive
a 1. Hence the density of interest is
Y
w(x) =
(1 − x(i)x(j)).
{i,j}∈E
Note that for all x ∈ Ω, w(x) ∈ {0, 1} as well.
2.2
Strauss process
The Strauss process [17] extends the independent set model by allowing two adjacent
nodes to both be labeled 1, but assigns a penalty factor of α ∈ [0, 1] when this occurs.
The underlying measure µ still is defined by (1), now the density becomes
Y
αx(i)x(j) .
w(x) =
{i,j}∈E
With the convention that 00 = 1, then when α = 0 this is just the independent set
density.
3
2.3
Autonormal
The autonormal model [1] is a continuous extension of the Ising model [13] used for
modeling images and other spatial experiments. For a graph G = (V, E), the state
space is Ω = [0, 1]V , the underlying measure µ has density
!
X
(2)
f (x) = exp −J
(x(i) − y(i))2
i
with respect to Lebesgue measure. Here J ≥ 0 and y(i) are known parameters. To
sample from µ, each node label X(i) is independently normal with mean y(i), variance
(2J)−1 , and conditioned to lie in [0, 1]. The density is
X
w(x) = exp −
β · (x(i) − x(j)2 ,
{i,j}∈E
where β is another constant.
2.4
Random Cluster
The random cluster model [8] is another way of viewing the Potts model [15], which
is itself a different extension of the Ising model [13]. Given a graph G = (V, E) the
state space is now a coloring of edges with 0 and 1. Given x ∈ Ω = {0, 1}E , and
parameter p, the underlying measure µ is
µ({x}) = p
P
e∈E
x(e)
(1 − p)
P
e∈E (1−x(e))
.
Again this measure µ is easy to sample from: draw (for all e ∈ E) X(e) ∼ Bern(p)
independently.
A second parameter q controls the density w through the use of c(x), which is the
number of connected components in the graph using only the edges with x(e) = 1.
w(x) = q c(x) .
(3)
When q = 2, a sample from the random cluster model can be transformed into a
sample from the Ising model in linear time [8], and for q an integer greater than 2, it
can be transformed in linear time into a sample from the Potts model.
3
Theory
The well-known acceptance rejection method requires two properties to hold:
1. It is easy to sample from the underlying measure µ.
4
Figure 1: An illustration of the random cluster model on a 2 by 2 square lattice.
Here the state has measure p(1 − p)3 since one edge is included and three edges are
not. The density of this state is q 3 since the edges of the state break the nodes into
3 connected components. With respect to counting measure, this state has density
p(1 − p)3 q 3 .
2. The target density is bounded by a known constant M.
When these hold (and they do for each of the applications of the previous section),
the acceptance rejection (AR) technique is as follows.
AR
1)
2)
3)
4)
5)
Output: X
Randomly draw X ← µ
Draw U uniformly from [0, 1]
If U < w(X)/M output X and quit
Else
X ← AR, output X and quit
It is well-known [7, 12] that the output is a draw from density w(x) with respect to
measure µ.
Now suppose that x ∈ C D for color set C and dimension set D (as it is for all
of our applications.) Let (D1 , D2 ) be a partition of the dimensions D. Note that for
each of our applications, the product nature of w(x) means that we can write
w(x) = w1 (x(D1 ))w2 (x(D2 )w12 (x(D1 ∪ D2 )),
For instance, for the independent sets model, if (V1 , V2 ) is a partition of the nodes V ,
then
"
#
Y
Y
Y
(1 − x(i)x(j))
w(x) =
(1 − x(i)x(j))
(1 − x(i)x(j))
{i,j}∈V1
{i,j}∈V2
i∈V1 ,j∈V2
and w(x) has been nicely factored into w1 , w2 , and w12 .
When for D1 and D2 there is an M such that w12 (x) ≤ M, fully recursive AR is
as follows.
5
Fully Recusrive AR
1)
2)
3)
4)
5)
6)
Input: S ⊆ D
Output: X(S)
Parition S into D1 and D2
X(D1 ) ← Fully Recursive AR(D1 ), X(D2 ) ← Fully Recursive AR(D2 )
Draw U uniformly from [0, 1]
If U < w12 (X)/M output X and quit
Else
X ← Fully Recursive AR(S), output X(S) and quit
Lemma 1. Fully recursive AR generates output X according to w with respect to µ.
This is similar to the self-reducibility of Jerrum, Valiant, and Vazirani [14]. More
recently, this approach has also been formalized in [12]. The proof follows immediately
from the correctness of basic AR and an induction on the number of times line 2 has
been used.
3.1
The PRAR protocol
Suppose we desire only to know X(v) for a single node v. Then first generate X(v)
and X(V \ {v}) separately, and then choose to accept or reject the combination. The
key idea of PRAR is that sometimes this does not require that we generate X(V \{v}).
For example, in the independent set model, if X(v) = 0, then we accept no matter
what X(V \ {v}) is. Even if X(v) = 1, we do not need to know the entirely of
X(V \ {v}), only those neighbors of v.
The general protocol looks like this.
Partially Recursive AR
Input: S and D where S ⊆ D
Output: [X(S ′ ), S ′] where X is a draw from w over dimensions D and S ⊆ S ′
1) If S = ∅ then output [∅, ∅]
2) Else
3)
Let d be any element of S, let D1 ← {d}
4)
Draw X(d) randomly from w1 over µ
5)
Draw U uniformly from [0, 1]
6)
If U ≤ minx(D\{d}) w12 (X(D1 ), x(D \ D1 ))/M
7)
Draw [X(S ′′ ), S ′′ ] ← Partially Recursive AR(S \ {d}, D \ {d})
8)
S ′ ← S ′′ ∪ {d}. Output [X(S ′ ), S ′ ] and quit
9)
Else
10)
Let D3 be the smallest set such that w12 (X(D1 ∪ D3 )) determines w12 (X)
11)
[X(S ′′ ), S ′′] ← Partially Recursive AR(D3 , D \ {d})
12)
If U ≤ w12 (X(D1 ∪ D3 ))/M, then S ′ ← S ′′ ∪ {d}, output [X(S ′ ), S ′]
13)
Else let [X(S ′ ), S] ← Partially Recursive AR(S, D) and output [X(S ′ ), S ′]
6
Lemma 2. Suppose Partially Recursive AR(S, D) terminates with probability 1
with output (X(S), D). Then X(S) has the same distribution as the coordinates of S
from X where X is a complete draw from density w with respect to µ over dimensions
D.
Proof. The Fundamental Theorem of Perfect Simulation [12, 9] says that if the algorithm terminates with probability 1, then we assume that lines 1, 6, 11, and 13 have
the correct output when proving that the overall algorithm is correct.
With that assumption, lines 7 and 11 are drawing a state X, and then reporting
X(S ′′ ) for some set of dimensions S ′′ that contain S. So the algorithm’s output
has the same distribution as Fully Recursive AR, but not all coordinates of X are
reported in the output.
It is the analysis of the running time where we see a Dobrushin like condition.
Lemma 3. Let c1 < 1 be the probability that line 13 is reached. Let p(d) denote the
probability that line 10 is reached given D1 = d. Suppose for all d ∈ D, there exists B
with E[#(D3 )|d] ≤ B, and that E[#(D3 )|d]p(d) ≤ c2 < 1. Given S then the expected
number of times line 4 is called (or equivalently line 5) is at most (1 − c1 )−1 (1 −
c2 )−1 #(S).
Proof. Consider two temporary changes to the algorithm. First, at line 13, change it
to Else, output [∅, ∅] and quit. Second, at line 3, add S ← S \{d} to the end. Because
of our previous change to line 13, this does not alter the algorithm any further.
Consider the expected time needed to either reach our new line 13, or to quit.
Then there might be several recursive calls to Paritall Recursive AR over the run
of the algorithm. Let Nt be the number of dimensions in the union of the S sets after
t recursive calls to the algorithm. Then N0 = #(S).
Now consider Nt+1 given Nt . Two things can happen. If lines 7 and 8 activate, then
Nt+1 = Nt − 1. If lines 10 and 11 activate, then Nt+1 = Nt − 1 + #(D3 ). This second
event happens with probability p(d). So E[Nt+1 |Nt , d] = Nt − 1 + p(d)E[#(D3 )|d].
When this right hand side is at most Nt − (1 − c2 ), standard martingale theory
(see for instance [4]) gives that the expected amount of time for the Nt process to
reach 0 is N0 /(1 − c2 ).
Now consider what happens if line 13 is changed back. Then the expected time
between reaching line 13 is #(S)(1 − c2 )−1 . Each time through the algorithm there
is at least a 1 − c1 chance line 13 is not executed and the algorithm does not recurse
further. Therefore, the total expected number of steps taken is at most (1 − c1 )−1 (1 −
c2 )−1 #(S).
7
4
4.1
Applications
Independent sets of a graph
Independent sets of a graph label each node either 0 or 1. In this case where there
are only two colors, the general PRAR can be simplified. If the label of a node is 0,
then the node is immediately accepted. But if the label is 1, then all the neighbors
of the node must also be 0. If a neighbor is 1, then we can immediately track down
its neighbors to ensure that they are all 0, and so on.
So if a neighbor of the end of the backbone is labeled 1, that increases the length
of the backbone by 1. On the other hand, if all the neighbors of the end of the
backbone at labeled 0, then the node that is the end of the backbone accepts its label
of 1. If the backbone had length 1, then we are done. Otherwise, that meant the the
second to the end node in the backbone rejects–because it comes into conflict with
the end of the backbone that is labeled 1. Therefore the last two nodes in backbone
are removed from the backbone and return to being unresolved. The process then
starts over again.
This gives us a connected sequence of nodes labeled 1 that grows and shrinks
in length. Call this sequence the backbone of the current state, and the method
Backbone PRAR. See Figure 2 for an illustration.
Step 1
Step 2
Step 3
Step 4
Step 5
Step 6
Step 7
Step 8
Step 9
Step 10
Step 11
Step 12
Figure 2: Backbone PRAR for independent sets. Here white stands for label 0 and
black for label 1. For instance, at step 7 the bottom center node was labeled 0. This
meant that the bottom right node was accepted, which then meant that the right
center node was rejected, so both were removed at Step 8. By step 12 the backbone
has resolved and the upper left labeling of 1 is accepted.
This idea of a backbone works on other applications with two colors, such as the
Ising model and the Strauss model.
8
Lemma 4. The backbone method for independent sets only requires an independent,
identically distributed stream of Bernoulli random variables with mean λ/(1 + λ). Let
∆ be the maximum degree of the graph and
i−1
∆−1
λ
λ X
1−
.
γ=
1 + λ i=1
(1 + λ)∆
If γ < 1, then the expected number of Bernoulli’s needed to resolve one node is bounded
above by C = 1 + λ∆(∆ − 1)/(1 − γ). Since this is a constant with respect to the size
of the graph, the time needed to generate a sample over the entire graph is at most
Cn.
Proof. When the backbone resolves, it resolves either as an acceptance or a rejection.
The number of resolutions until acceptance is stochastically bounded above by a
geometric random variable with parameter 1/(1+λ). Therefore, the expected number
of resolutions of the backbone is at most (1 + λ) − 1 = λ.
How many draws does it take to resolve a node? It certainly takes one to determine
the label for the node. So if T is the resolution time, then E[T ] ≤ 1 + ∆R, where
R is an upper bound on the expected resolution time for subsequent nodes. These
subsequent nodes have maximum degree at most ∆ − 1. The first neighbor takes at
most R expected time.
On to the second neighbor. Note that the second neighbor only needs to be
resolved if the first neighbor resolved to a 0. If the first neighbor had resolved to a
1, then the original node is rejected, there is no need for further action. The chance
that a neighbor resolves to 1 is at least c = [λ/(1 + λ)][1/(1 + λ)]∆−1 . So the second
neighbor is only activated with probability 1 − c.
Similarly, the third neighbor can also take time distributed as R, but only if the
first two neighbors fail. Adding this up over the (up to ∆ − 1) neighbors gives
∆−1 !i−1
∆−1
λ X
λ
1
R≤1+
R 1−
,
1 + λ i=1
1+λ 1+λ
which gives the result.
It is straightforward to verify that for λ < 1.13/(∆ − 2) we have γ < 1. This gives
the result presented in Section 1.
Call the value of λ where γ = 1 the critical value of λ, and denote it λc . For
λ < λc , the algorithm is guaranteed to generate samples in polynomial time. For
λ ≥ λc , the algorithm might operate in polynomial time, or it might not.
4.2
Strauss process
Now consider what happens in the Strauss process. Again consider drawing an initial
node and accepting or rejecting based upon the neighbors of the node.
9
∆
3
4
5
λc for RR
λc for CoA
λc for PRAR
0.2
0.142857. . .
0.111111. . .
0.5
0.333. . .
0.25
1.13224. . .
0.57833. . .
0.29315. . .
Figure 3: Range of critical λ compared to ∆. Since expected sampling time is guaranteed linear for λ < λc , higher is better. Column for λc for the Randomness Recycler
(RR) comes from Theorem 1 of [6]. Column for λc for Clan of Ancestors (CoA) comes
from a branching process analysis.
If the node is labeled 0, then we always accept as before. If the node is labeled 1,
then the chance of accepting is α raised to the power of each of the neighbors of the
node labeled 1.
Another way to view this, is to, for each edge adjacent to the original node, draw
a Bern(α) random variable. If this variable is 1, then the neighbor does not matter.
If this variable is 0, then the neighbor does matter, and recursion needs to be used
to find out its value.
This changes the expected number of neighbors to be considered from ∆ − 1 to
(∆ − 1)α. The chance of a node being labeled 1 is strictly greater in the recursion,
therefore, the new value of γα is at most
λ
α(∆ − 1).
1+λ
Note that γα < 1 is equivalent to λ < 1/[α(∆ − 1) − 1], so a similar argument to the
previous section gives the following result.
γα =
Lemma 5. When λ < [α(∆ − 1) − 1]−1 , then PRAR generates a sample from the
Strauss process using O(n) expected number of Bernoulli draws.
4.3
Autonormal
As with the independent set density, the idea is to consider what happens when a
single node is drawn, the rest of the graph is drawn, and the combination is either
accepted or rejected. The value that U from line 6 in the general protocol must fall
below is the product of the chance of rejecting because of each of the neighboring
edges.
That is, in order to determine if node v can be combined with the state of x(V \
{v}), it is necessary to independently draw a uniform [0, 1] random variable for each
of the neighbors w of v. Only if every edge accepts can the entire state be said to
accept.
If the single node v is assigned value x(v), the chance of rejecting based on w is
at most
exp(−β(x(v) − x(w))2 ) ≤ exp(−β max{1 − x(i), x(i)}2 ),
10
where the upper bound on the right hand side does not depend on x(w)!
Therefore, only if Uw ≥ exp(−β max{1 − x(i), x(i)}2 ) does the value x(w) need to
be determined. Therefore, the expected number of neighbors of the original node v
that need to be found out is bounded above by
∆(1 − exp(−β max{1 − x(i), x(i)}2 )).
The original node v might have had up to ∆ neighbors, but subsequent nodes will only
have at most ∆−1 neighbors. Therefore, if (∆−1)(1−exp(−β max{1−x(i), x(i)}2 )) <
1, then on average the number of new nodes to consider generated by a node will be
negative, and PRAR will terminate after a finite number of steps. This argument
yields the following lemma.
Lemma 6. Let γ = (∆ − 1) exp(−β). If γ < 1, then PRAR can generate a sample
from the Autonormal model in O(n) random choices on a graph with n nodes and
maximum degree ∆.
A similar analysis for the Clan of Ancestors approach requires ∆ exp(−β) < 1,
therefore the parameter range with guaranteed performance here is slightly wider.
4.4
Random Cluster
Suppose that q > 1 in the random cluster model with density given by (3). Then the
density of a configuration gains a factor of q for each connected component in the
graph.
Viewed as a penalty, this means that for a particular edge, there is a penalty of
1/q if it connects two previously disconnected components in the rest of the graph.
In other words, if c = (1/q)n , where n is the number of nodes in the graph, then the
density can be written as
w2 (x) = cw(x) = (1/q)n−c(x)) .
Now consider a single edge {i, j} chosen from the underlying measure. If the
edge is labeled 0 (which happens with probability 1 − p), then the edge is accepted
regardless of the rest of the components. If the edge is labeled 1 (which happens with
probability p), then the edge is accepted with probability 1/q if the rest of the state
does not connect nodes i and j, and accepted with probability 1 otherwise.
Suppose that x({i, j}) is randomly chosen to be 1. Then draw U uniformly over
[0, 1]. If U < 1/q, then the edge {i, j} is accepted regardless of the rest of the state
and the edge value becomes known. Otherwise, recursion must be used to determine
enough of the rest of the state in order to determine if the nodes {i, j} are connected
or not.
The first edge {i, j} can be connected to as many as 2(∆ − 1) different edges, each
of which must be considered to determine if x({i, j}) = 1. Let e be one of these edges
(either of the form {i, k} or {j, k} for some nodes k.)
11
If e is chosen to be labeled 0 in the recursion, then it is removed from the “needs
a value” list. Otherwise, it is adjacent to at most ∆ − 1 new edges that might need a
value. Hence the expected change of the size of the number of edges that need values
is bounded above by
(−1)(1 − p) + p(∆ − 1) = ∆p − 1.
If this quantity is less than 0, then the number of edges to be considered drops (on
average) at each step, and the expected number of steps is bounded. This gives the
following lemma.
Lemma 7. Let G be a graph with n nodes, m edges, and maximum degree ∆. If
p < 1/∆, then the expected number of random choices made in the algorithm is
O(m).
5
Connection to popping
This PRAR protocol can be viewed as an extension of the cycle popping algorithm of
Wilson for uniformly generating from rooted trees in a directed graph.
Consider a graph (V, E), and construct a directed graph by taking each edge {i, j}
and adding directed arcs (i, j) and (j, i). Designate a special node r ∈ V known as
the root of the graph. Then let a configuration consist of a {0, 1} labeling of arcs
such that every node v 6= r has exactly one outgoing edge (v, w) labeled 1.
Let µ be the underlying measure that is uniform over all such labelings. Then it
is easy to sample from µ: independently for each node i select a neighbor j uniformly
at random and label that outgoing arc 1, and all other outgoing arcs from i label 0.
The density w(x) then is 1 if every node has a directed path using edges labeled 1
to the root, and is 0 otherwise. When w(x) is 1, say that x encodes a directed rooted
tree in the original graph.
Then Wilson [18] presented a simple algorithm for generating from w(x) over µ.
Start with an arbitrary node i, and uniformly choose a neighbor j. From j choose
neighbor k, and continue until one of two things happens.
If the directed path reaches a node already examined, then there is a loop. Erase
the loop and continue onwards. For instance, if the choices where a → b → c → d → b,
then (b, c, d, b) forms a loop, and the next step would start with a → b. The other
thing that can happen is that the directed path reaches r. In this case, fix the labeling
for all nodes along the path.
Now choose another node in the graph and repeat, choosing random neighbors,
erasing loops as they form, and stopping when reaching either r or a previously fixed
node.
This algorithm is just PRAR with the backbone method applied to µ and w! At
each step of the algorithm we are recursively moving deeper into the graph in order
to determine if acceptance should occur. If a loop forms, then rejection occurs, and
all the edges in the loop are eliminated. But if the path should encounter a previously
12
fixed path, then acceptance occurs not only for the original edge but all along the
path as well.
Wilson gave a proof tailored for his algorithm. Because this is also a PRAR
algorithm, we immediately have correctness for the loop-erased random walk method
of uniformly generating rooted trees.
Acknowledgements
An earlier version of this technique appears in [12] (without the backbone method,
connection to popping, and the general running time bound given here.) The recent
development of this method was supported by NSF grant DMS-1418495.
References
[1] J. Besag. Spatial interaction and the statistical analysis of lattice systems (with
discussion). J. R. Statist. Soc. Ser. B Stat. Methodol., 36:192–236, 1974.
[2] R. Bubley and M. Dyer. Path coupling: a technique for proving rapid mixing
in Markov chains. In Proc. 38th Sympos. on Foundations of Computer Science,
pages 223–231, 1997.
[3] P. L. Dobruschin. The description of a random field by means of conditional
probabilities and conditions of its regularity. Theor. Prob. Appl., 13(2):197–224,
1968.
[4] R. Durrett. Probability: Theory and Examples, Third edition. Brooks/Cole,
2005.
[5] P.A. Ferrari, R. Fernández, and N. L. Garcia. Perfect simulation for interacting point processes, loss networks and ising models. Stochastic Process. Appl.,
102(1):63–68, 2002.
[6] J. A. Fill and M. L. Huber. The Randomness Recycler: A new approach to
perfect sampling. In Proc. 41st Sympos. on Foundations of Comp. Sci., pages
503–511, 2000.
[7] G. S. Fishman. Choosing sample path length and number of sample paths when
starting in the steady state. Oper. Res. Letters, 16:209–219, 1994.
[8] C.M. Fortuin and P.W. Kasteleyn. On the random cluster model I: Introduction
and relation to other models. Phys., 57:536–564, 1972.
[9] M. Huber. The Fundamental Theorem of perfect simulation. 2016. Submitted.
13
[10] M. Huber, E. Villella, D. Rozenfeld, and J. Xu. Bounds on the artificial phase
transition for perfect simulation of the hard core Gibbs processes. Involve,
5(3):247–255, 2012.
[11] M. L. Huber. A faster method for sampling independent sets. In Proc. 11th
ACM-SIAM Sympos. on Discrete Algorithms, pages 625–626, 2000.
[12] Mark L. Huber. Perfect Simulation. Number 148 in Chapman & Hall/CRC
Monographs on Statistics & Applied Probability. CRC Press, 2015.
[13] E. Ising. Beitrag zur theorie des ferromagnetismus. Z. Phys., 31:253–258, 1925.
[14] M. Jerrum, L. Valiant, and V. Vazirani. Random generation of combinatorial
structures from a uniform distribution. Theoret. Comput. Sci., 43:169–188, 1986.
[15] R. B. Potts. Some generalised order-disorder transformations. Proc. Cambridge
Philosophical Soc., 48:106–109, 1952.
[16] J. G. Propp and D. B. Wilson. Exact sampling with coupled Markov chains
and applications to statistical mechanics. Random Structures Algorithms, 9(1–
2):223–252, 1996.
[17] D. J. Strauss. A model for clustering. Biometrika, 63:467–475, 1975.
[18] D. B. Wilson. Generating random spanning trees more quickly than the cover
time. In Proc. 28th ACM Sympos. on the Theory of Computing, pages 296–303,
1996.
14
| 8cs.DS
|
Coordinated Charging and Discharging Strategies for Plug-in Electric Bus Fast
Charging Station with Energy Storage System
Huimiao Chen, Zechun Hu *, Hongcai Zhang, Haocheng Luo
Department of Electrical Engineering, Tsinghua University, Beijing, 100084, China
*
[email protected]
Abstract: Plug-in electric bus (PEB) is an environmentally friendly mode of public transportation and plug-in electric bus fast
charging stations (PEBFCSs) play an essential role in the operation of PEBs. Under effective control, deploying an energy
storage system (ESS) within a PEBFCS can reduce the peak charging loads and the electricity purchase costs. To deal with the
(integrated) scheduling problem of (PEBs charging and) ESS charging and discharging, in this study, we propose an optimal
real-time coordinated charging and discharging strategy for a PEBFCS with ESS to achieve maximum economic benefits.
According to whether the PEB charging loads are controllable, the corresponding mathematical models are respectively
established under two scenarios, i.e., coordinated PEB charging scenario and uncoordinated PEB charging scenario. The price
and lifespan of ESS, the capacity charge of PEBFCS and the electricity price arbitrage are considered in the models. Further,
under the coordinated PEB charging scenario, a heuristics-based method is developed to get the approximately optimal
strategy with computation efficiency dramatically enhanced. Finally, we validate the effectiveness of the proposed strategies,
interpret the effect of ESS prices on the usage of ESS, and provide the sensitivity analysis of ESS capacity through the case
studies.
1. Nomenclature
Indices and Sets
n
Index of PEB.
m
Index of fast charging port.
Index of time interval in the following
k
optimising time horizon.
Index of parking of PEB n in the
in , jn
following optimising time horizon.
Subscript of charging/discharging.
c/d
N
Index set of PEBs.
Index set of fast charging ports.
M
Index set of time intervals in the following
K
optimising time horizon.
Index set of time intervals during parking
P(in )
in of PEB n .
Index set of parking of PEB n in the
I ( n)
following optimising time horizon.
Index set of time intervals in the following
P( I (n))
optimising time horizon when PEB n is
expected to be parking.
card( X ) The number of elements in set X .
Parameters and Variables
SnPEB
Battery capacity of PEB n (kWh).
ESS
Energy capacity of ESS (kWh).
S
PcPEB
Rated charging power of PEBs (kW).
Ppeak
Peak load of PEBFCS (kW).
ESS
Pc,max
Maximum charging power of ESS (kW).
ESS
Pd,max
Maximum discharging power of ESS (kW).
cPEB
cESS
Charging efficiency of PEBs.
dESS
Discharging efficiency of ESS.
Charging efficiency of ESS.
ESS
SOCmin
Minimum state of charger (SOC) for PEB
batteries.
Minimum SOC for ESS.
SOCnPEB
,in
SOC of in th arrival of PEB n .
PEB
SOCmin
ESS
k
SOC
SOCnPEB
,in
an ,in
SOC of ESS at the beginning of time
interval k .
SOC difference of PEB n between in th
departure and the next return.
Time interval of in th expected return of
PEB n ( an,i K ).
n
ln ,in
Time interval of in th expected departure of
PEB n ( ln,i K ).
Lk
Power of other loads excluding PEB
charging loads in time interval k (kW).
Duration of a time interval (min).
Electricity price in time interval k
(RMB/kWh).
Price of ESS (RMB/kWh).
Capacity charge of PEBFCS (RMB/kW).
The number of charge-discharge cycle of
ESS.
Discount rate of the capacity charge (%).
Life cycle of PEBFCS (year).
Charging state matrix of PEBs
(dimensions: card( N ) card( K ) ).
Charging state matrix of fast charging ports
(dimensions: card(M ) card( K ) ).
Vector of charging power of ESS
(dimensions: 1 card( K ) ).
Vector of discharging power of ESS
(dimensions: 1 card( K ) ).
n
t
kTOU
ESS
Cap
n ESS
C PEB
C FCP
PcESS
PdESS
cnPEB
,k
Element of CPEB , binary variable, 1: on
1
cmFCP
,k
EES
Pc,k
charge; 0: off charge.
Element of C FCP , binary variable, 1: on
state; 0: off state.
Element of PcESS , the charging power of
P
ESS in time interval k (kW).
Element of PdESS , the discharging power of
un, k , vn , k
ESS in time interval k (kW).
Auxiliary variables.
EES
d,k
2. Introduction
The wide use of fossil energy has resulted in global
warming and severe environmental pollution [1]. Plug-in
electric vehicles (PEVs) have incomparable advantage over
fuel-powered vehicles in environmental protection and
sustainable development [2], [3]. With development and
popularization of PEVs, a large-scale of PEVs will be
connected to the public power grid in the future. The
incremental charging loads of PEVs will have a massive
impact on the existing power system [4], [5]. For example,
difference between load peak and off-peak may increase;
power quality could be deteriorated; distribution networks
will face new challenges, including increasing of network
losses [6], [7], overloading of transformers [8], [9],
excessively heavy line loads and larger voltage deviations
[10], [11], etc.
Present research has shown coordinated charging of
PEVs is able to effectively reduce the negative impact of
PEVs’ charging loads on the power system [12]-[17]. The
optimisation objectives and methods of PEV coordinated
charging are various in the literature. Under the time-of-use
(TOU) prices, reference [12] proposes a cost-optimal control
strategy for multiple PEV aggregators to guarantee that the
distribution system runs within the security limits. In
deregulated electricity market, authors of [13] present an
optimal charging control method for PEVs to provide
ancillary services based on the forecast of future electricity
prices. In [14], a two-stage optimisation method is developed:
Firstly, in order to achieve peak shaving and valley filling,
PEVs are allocated appropriate charging periods according to
the urgency degree; then the charging sites are optimised to
minimise transmission losses. In [15], a threshold admission
and greedy scheduling policy is proposed to maximise the
revenue of charging services for large-scale electric vehicles.
Taking vehicle to grid into account, reference [16] builds an
optimal scheduling model to minimise the energy
consumption and carbon emission, whereas reference [17]
formulates a dynamic charging control strategy for providing
frequency regulation services.
Though a lot of research on PEV coordinated
charging has been done, most of the relevant works
concentrate on studies of slow or normal charging mode of
private PEVs, which are not applicable for a plug-in electric
bus fast charging station (PEBFCS). As a special type of
PEVs, plug-in electric bus (PEB) is an electric bus which is
powered by electricity and can be recharged from an external
source of electricity. In [18], an effective charging strategy
for PEBFCSs is proposed to minimise the power purchase
costs by responding to the TOU prices, and as the result, the
peak loads are mitigated as well. But the work is not suitable
for a PEBFCS with ESS. Nowadays, with the rapid
development of energy storage technology, installing ESS in
the charging station can achieve better demand response [19].
However, only a few published literature focuses on charging
stations with ESS. Reference [20] proposes a control strategy
for PEV fast charging station equipped with a flywheel ESS,
which is able to work without any digital communication
between the grid-tied and flywheel ESS converters.
Reference [21] provides a method to schedule PEV charging
with energy storage and shows that aggregator’s revenue
varies as the number of PEVs and the number of energy
storage units change. Authors of [22] present a coordinated
control strategy for ESS to reduce the electricity purchase
costs and flatten the charging load profile. However, the
investment costs of energy storage are not taken into account
both in [21] and [22]. Besides, the original load curve is given
and fixed in [22] so that the elasticity of PEV charging loads
cannot play a role. In [19], the value of ESS in a PEBFCS is
discussed as the core problem and the control strategy of
PEBs is not concerned.
To the best of our knowledge, there is no existing
papers which study the optimal load scheduling method for a
PEBFCS with ESS. Thus, herein, we aim to develop an
integrated control strategy for both ESS and PEB loads in a
PEBFCS with ESS in order to achieve maximum economic
benefits. It is worthy to note that a PEBFCS with ESS is a
valuable research object for following reasons: 1) PEB is a
green public means of transportation, which is convenient for
centralized control and management. Moreover, a large
number of PEB lines have been in commercial operation or
demonstrational operation in some cities (e.g. Shenzhen and
Chongqing, China). 2) Fast charging stations are regarded as
the promising providers of public PEVs’ charging service in
the future because they can provide large charging power and
meet urgent charging demands. 3) As the technology of ESS
advances, the efficiency and lifespan of ESS are expected to
be improved and its price declines. Thus ESS is an effective
supplement for a PEBFCS to reduce the high capacity charge
for the grid integration as well as to reduce the charging costs
through arbitraging the price differences under TOU price
scheme.
Based on the above considerations and motivations,
the main procedures and contributions of the paper are
summarized below.:
1) A coordinated charging strategy for PEBs without
considering ESS is formulated as the baseline strategy.
Additionally, under the coordinated PEB charging scenario
(PEB charging loads are controllable), an optimal
coordinated charging and discharging strategy involving
PEBs and ESS is proposed. The control of ESS and PEBs is
optimised in an integrated way and the combined control
strategy achieves the best optimality.
2) Under the uncoordinated PEB charging scenario
(PEB charging loads are uncontrollable), an optimal
coordinated charging and discharging strategy of ESS is
presented.
3) To enhance the computation efficiency, under the
coordinated PEB charging scenario, a heuristics-based
method is further developed to get the approximately optimal
control strategy of ESS and PEBs.
4) Operation costs, load profiles and some other
important indices of a given PEBFCS with ESS are simulated
and compared with the ones without ESS under both two
scenarios, i.e., coordinated PEB charging scenario and
uncoordinated PEB charging scenario, to verify the
2
effectiveness of the proposed control strategies. The impacts
of ESS capacity on economic benefits are also analyzed.
The remainder of the paper is organized as follows.
Section 3 describes the scenario for the proposed strategies.
The details and the mathematical formulations of the
strategies are presented in Section 4. Section 5 shows case
studies and complements our analysis. Finally, Section 6
concludes.
3. Scenario Descriptions
In order to ensure the security of power grid, the power
capacity for a PEBFCS is usually adequate for simultaneous
charging of all the fast charging piles, otherwise the total
power of chargers might exceed the capacity of distribution
transformer and the line overload might occur. However, in
the actual operation, the total charging power of the PEBFCS
seldom hits the upper limit [19]. For this reason, installing
ESS rather than a distribution transformer with overlarge
capacity could be a more economical way by reducing the
grid connection fee, i.e., the capacity charge for a PEBFCS.
The configuration of a PEBFCS with ESS is illustrated
in Fig. 1. The network supplies power to the station through
the local distribution transformer. And ESS, PEB charging
piles and the appliances of nearby residential or commercial
areas (other loads) are connected to the secondary side of the
distribution transformer. If the station has exclusive
distribution transformer, the ratio of the power of other loads
will be approximately zero. In this text, it is assumed that
PEBFCS purchases electricity from the utility at TOU
electricity prices and provides fast charging service to PEBs.
Note that in deregulated electricity markets, the forecasting
electricity prices can take place of the TOU electricity prices
and the control strategy proposed in the following still works.
For a PEBFCS, we suppose that the number of PEBs,
i.e., card( M ) , the number of fast charging piles, i.e.,
card( N ) , and the PEB departure time-table are given. The
control system of PEBFCS is able to acquire SOCs of PEBs
when they arrive at the station through built-in battery
management systems on PEBs. And SOCnPEB
,in can be used to
forecast the following arrival SOCs. In actual operation,
SOCnPEB
,in can be obtained based on historical data. Note that
the proposed control strategy in this paper is to some degree
resistant to the influence of the SOC forecasting errors
because the real-time control is updated periodically and the
negative effect caused by the forecasting errors will be
gradually mitigated (see Section 4 for the details of the
proposed control strategy). In order to reduce the negative
effect on battery lifespan caused by excessive discharging,
the minimum SOCs for PEB batteries and ESS are set, i.e.,
PEB
ESS
SOCmin
and SOCmin . Also, due to the harmful effect of fast
charging start-stop on chargers and batteries, continuous
(uninterruptible) charging is adopted.
For the control strategies, the optimising time horizon
is discretely divided into card( K ) time intervals equally and
the duration of each time interval is t . Charging power of
PEB, and charging and discharging powers of ESS are
regarded as constants in each time interval. Intuitively, larger
card( K ) (smaller t ) makes the results more accurate, but
the computation burden will be heavier. In practice, the value
of card( K ) and t can be adjusted according to the
accuracy requirements and the computational performances.
Based on the above scenario descriptions, the
coordinated charging and discharging strategies for PEBFCS
with ESS is computed through optimisation models to meet
multiple constraints, such as charging demands of PEBs and
continuous fast charging of PEB batteries, and to improve the
economic benefit of PEBFCS, detailed in the next section.
Substation
Distribution
Transformer
ESS
Power
Time
Plug-in Electric Bus
Fast Charging Station
Other Loads
Fig. 1. Schematic illustration of a PEBFCS with ESS
4. Strategies and Mathematical Formulations
4.1. Control Strategy Overview
In this subsection, we outline the proposed control
strategies. A rolling horizon optimisation method is applied
to implementation of the proposed control strategy. Every
time, a real-time control strategy is formulated according to
information of the following optimising time horizon (from
the next time interval to the card( K ) th time interval), and to
ensure the real-time control performance, each control
strategy will be only executed for one time interval and then
a recalculated strategy will substitute for it for the next time
interval.
For the control strategy under coordinated PEB
charging scenario, it is operated through a three-step serial
processing procedure at each time interval. When a new time
interval begins, the control system will orderly 1) implement
the strategy generated in the last time interval by controlling
the on-off states of fast charging piles and the charging or
discharging power of ESS; 2) make sure all the necessary
data ready and then calculate the new charging and
d i s c h ar g i n g s tr a te g y o f P E B s a n d E S S (d e ta i le d
subsequently), which will be implemented in the next time
interval; 3) prepare data for the next cycle. Here, data
preparation in a time interval include that: 1) if a PEB arrives,
the control system will acquire the number of the PEB and
read its current SOC, i.e., its SOCnPEB
,1 ; 2) the next arrival
SOCs of the rest PEBs, i.e., the other SOCnPEB
,1 , are estimated
through the last departure SOCs (if there is communication
3
New time interval?
No
Yes
Implement
the strategy
generated in
the last time
interval.
Calculate the
next strategy.
Get C PEB , P cE S S and P dE S S.
PEB
Convert C
to C FC P.
Remain the
system
unchanged
and enter a
delay of 10s.
No
New time interval?
Yes
Calculate the next strategy.
Read or estimate
all the SOCnPEB
,1 .
Implement
the strategy
generated in
the last time
interval.
Update the PEB
departure time-table.
Get P cE S S and P dE S S .
Forecast the future PEB
charging load profile.
Read or forecast the
other related parameters.
Read or forecast the other
related parameters.
a
Fig. 2. Flow charts of the control strategies
(a) Under coordinated PEB charging scenario, (b) Under uncoordinated PEB charging scenario
between PEBFCS and PEBs on road, the current SOCs of all
the PEBs will be available) and SOCnPEB
,in ; 3) the PEB
b
min EPC ECC
(1)
where
departure time-table is updated according to the current
information of PEBs; 4) the control system gets the other
necessary information, which can be extracted from the
database (e.g. the battery capacity SnPEB , the electricity price
PEB
TOU
EPC cnPEB
, k Pc
t k
k K nN
Cap
ECC Ppeak ,
kTOU ) or be predicted based on historical data (e.g. the power
of other loads rk ). The corresponding flow chart is shown in
Fig. 2 (a).
As for the control strategy under uncoordinated PEB
charging scenario, the processing procedure is similar but
becomes less complicated without regard to the control of
PEB charging loads. At the beginning of a new time interval,
the control system will orderly 1) implement the strategy
generated in the last time interval by controlling the charging
or discharging power of ESS; 2) calculate the new charging
and discharging strategy of ESS (detailed subsequently); 3)
as the basis for the strategy formulation in the next cycle,
TOU
prepare the necessary data (e.g. the electricity price k , the
power of other loads rk ) and forecast the PEB charging load
profile in the following optimising time horizon. The
corresponding flow chart is shown in Fig. 2 (b).
4.2. Formulations of Optimisation Models
Mathematical models in this section aim at optimising
the economic benefits of PEBFCS, i.e., minimising the
equivalent operation costs of the following optimising time
horizon.
4.2.1 Optimisation Model for Coordinated PEB
Charging Scenario without ESS: Here, the optimisation
model (Model A) is utilized to compute the coordinated PEB
charging strategy, which is used as the reference. The
objective function is to minimise the electricity purchase
costs ( EPC ) and the equivalent capacity charge ( ECC ) in
the following optimising time horizon. The whole model is
formulated as (1)-(10).
Remain the
system
unchanged
and enter a
delay of 10s.
card K t
(2)
1
365 24 60
,
1
1
subject to:
PEB
Ppeak cnPEB
Lk , k K
, k Pc
c
nN
card( M ), k K
cnPEB
, k 0, k K \ P I (n) , n N
PEB
n, k
un, k c
c
, n N , k K \ 1
PEB
n, k
vn, k c
c
, n N,
PEB
n , k -1
PEB
n , k 1
k K \ card K
un,k , vn,k 0,1
(4)
nN
PEB
n, k
kP in
(3)
un, k
kP in
(5)
(6)
(7)
(8)
(9)
vn, k 1, in I n , n N
(10)
PEB
PEB
PEB
PEB
SOCn,in SOCn,1 SOCmin Sn
in jn
PEB PEB PEB
cn, k Pc c t
in jn k P in
PEB
PEB
PEB
SOCn ,in 1 SOCn ,in Sn ,
in jn 1
(11)
jn I n , n N .
In the above model, equations (2) calculates the
electricity purchase costs by summing the costs in all the time
intervals. In (2), cnPEB
, k denotes the charging state of PEB n at
time interval k ; PcPEB denotes the rated charging power of
PEBs; t denotes the duration of a time interval; kTOU is
4
the electricity price at time interval k . Equation (3) calculates
the equivalent capacity fees of PEBFCS. In (3), is capital
recovery factor ( is the discount rate of the capacity charge
and is the life cycle of PEBFCS), which converts the initial
investment costs into a stream of equal annual payments over
years, and the equivalent annual capacity charge is
multiplied by a ratio , which is the proportion of the
duration of optimisation time horizon to a year (see the
second equation in (3)), to proportionally count the capacity
charge for the following optimising time horizon. Cap is the
capacity charge of PEBFCS and the peak load Ppeak can
calculated by (12), where Lk is the total power of other loads
excluding PEV charging loads in time interval k . In the
model, constraints (4) are linearized expressions to describe
Ppeak instead of (12) without any sacrifice of optimality,
because the optimal solution must meet equation (12) to
achieve the lowest costs.
PEB
Ppeak max cnPEB
Lk
(12)
, k Pc
kK
nN
For constraints (5)-(11), (5) are the upper charging pile
number constraints; constraints (6) describe that only parking
PEBs can be on charge, where P( I (n)) is the index set of
time intervals in the following optimising time horizon when
PEB n is expected to be parking and P( I (n))
in I ( n )
P(in ) an,1 , , ln,1∪∪ an,card( I n) , , ln,card( I n )
,
n N ; constraints (7)-(10) ensure the continues charging of
PEBs by introducing two auxiliary variables, i.e., un , k and
vn , k . Note that un , k and vn , k guarantee the continuity of PEB
charging by restricting the change of PEB charging state,
which can be easily proved, and the introduce of un , k and vn , k
makes the constraints linear; constraints (10) describe the
charging demand constraints of PEBs. The recharged energy
of PEBs during each charging between the minimum
charging demands and the available battery capacity, as
shown in (13).
SOC
PEB
n , in
k P in
SOC
t 1 SOC
PEB PEB PEB
n,k
c
c
c
S
SOC
PEB
min
P
PEB
n , in
PEB
n
PEB
n , in
S
PEB
n
,
(13)
Based on (13), we utilize all the SOCnPEB
and
,1
SOC
PEB
n ,in
to estimate the other SOC
4.2.2 Optimisation Model for Coordinated PEB
Charging Scenario with ESS: Here, the optimisation
model (Model B) is utilized when ESS is installed and
charging loads of PEBs are controllable. To ensure the global
optimality of the control strategy, PEB charging strategy and
ESS charging and discharging strategy are optimised
integratedly in the model. The objective function is to
minimise the electricity purchase costs ( EPC ), the life
expenditure costs of ESS ( ESSC ) and the equivalent capacity
charge ( ECC ) in the following optimising time horizon. The
whole model is formulated as (14)-(24).
min EPC + ESSC + ECC
(14)
where ECC is calculated by (3) and
PEB
TOU
EPC cnPEB
, k Pc
t k
k K nN
ESS
TOU
Pc,ESS
k Pd, k t k
k K
kK
ESS
π
ESS
ESSC= ESS Pc,ESS
t
k c
n kK
subject to:
(5)-(11)
PEB
ESS
Ppeak cnPEB
Pc,ESS
, k Pc
k Pd, k Lk ,
. Then, (13) can be
rewritten in form of accumulation, i.e., (11). In (11), in jn
means parking in occurs no later than parking jn . And the
middle part of each inequality in (11) represents the total
recharged energy of PEB n during jn times parking. The
first and last parts of each inequality in (11) respectively
represent the lower and upper limits of total recharged energy
of PEB n after jn times charging processes. Note that (13)
PEB
PEB
+SOCn,i
1, n N .
and (11) both require SOCmix
n
The optimal charging state matrix of PEBs CPEB can
be obtained by solving Model A. And CPEB should be
(15)
(16)
(17)
nN
kK
ESS
0 Pc,ESS
k Pc,max , k K
0 P
ESS
d, k
ESS
k
SOC
SOC
P
ESS
d,max
(18)
(19)
,k K
P
ESS ESS
c, k
c
ESS
k 1
(20)
t P
ESS
d, k
t /
ESS
c, k
P
ESS
d, k
+P
P
ESS
d
, k K \ card K
ESS
SOCmin
SOCkESS 1, k K
kK
in I n , n N
PEB
n ,in
converted into the corresponding charging state matrix of
charging piles CFCP before implementing the control strategy.
This step can be achieved by checking the numbers of each
PEB and the charging pile the PEB connected to.
max P
ESS ESS
c, k
c
ESS
c, k
ESS
d, k
+P
/
ESS
d, k
,P
ESS
d
(21)
(22)
, k K
=0
(23)
(24)
In the above model, equation (15) calculates the
electricity purchase costs for PEBFCS with ESS. In (15),
EES
EES
and Pd,k
denote the charging power and discharging
Pc,k
power of ESS at time interval k , respectively. Equation (16)
approximately calculates the ESS life expenditure costs
during the optimising time horizon through the recharged
energy, where, ESS , n ESS and cESS are the unit price of
ESS, the number of charge-discharge life cycle of ESS and
the charging efficiency of ESS, respectively. The peak load
Ppeak here is constrained by (18) and the optimal solution
must make (25) hold (similar to constraint (4) in Model A).
PEB
ESS
Ppeak max cnPEB
Pc,ESS
(25)
, k Pc
k Pd, k Lk
k K
nN
For constraints (19)-(24), (19) and (20) restrict the
charging power and discharging power of ESS within the
maximum powers, respectively; (21) describe the energy
state transition of ESS; (22) are the SOC range constraints for
5
ESS; (23) describe the mutual exclusion of charging and
discharging states of ESS, i.e., avoid simultaneous charging
and discharging of ESS; (24) is the energy balance constraint
of ESS, since it is usually expected that the initial and final
SOCs are same.
In above constraints, nonlinear constraints (23) can be
deleted from Model B without any sacrifice of optimality due
to the charging and discharging efficiency. The detailed proof
is omitted here and interested reader may need to refer to the
appendix in [19], which is similar.
Solving Model B, the optimal charging state matrix of
PEBs CPEB , the optimal vector of charging power of ESS
PcESS and the optimal vector of discharging power of ESS
PdESS are accordingly obtained.
4.2.3 Optimisation Model for Uncoordinated PEB
Charging Scenario with ESS: In this subsection’s
optimisation model (Model C), ESS is taken into account and
charging loads of PEBs cannot be scheduled. The expression
of the objective function is same as (14), but the charging
PEB
profile of PEBs, i.e., cnPEB
, is input parameters here.
, k Pc
nN
The problem is modeled as (26) and (27).
min EPC + ESSC + ECC
(26)
where EPC , ESSC and ECC are calculated by (15), (16)
and (3), respectively.
subject to:
(18)-(22), (24).
(27)
The optimal vectors of charging and discharging
powers of ESS, i.e., PcESS and PdESS , can be obtained by
solving Model C.
vn , k . To fix this, herein, we further propose a heuristics-based
method.
Firstly, we solve Model A without regard to constraints
(7) and (8) and obtain a charging state matrix of PEBs CPEB
PEB
(elements: cnPEB
and the PEB departure
, k ). According to C
time table, we can extract approximate charging time of all
the PEBs during their each parking, denoted by CTn,in , in the
following optimising time horizon (see (28)). Also, we treat
ˆ PEB sum(CPEB ) (elements: cˆ PEB ; sum(X) is a row vector
C
k
whose elements are the sum of every column of X ) as the
guideline of the number of PEB on charge at each time
interval.
ln ,in 1
k ai ,in 1
cnPEB
,k
determined.
LX n, k ln,in CTn,in k , an,in k ln,in
Based on Tn ,in , cˆ
PEB
k
(28)
Secondly, inspired by [23] and [24], we define the
laxity LX n, k (see (29)), which is used to describe the
(29)
and LX n, k , a heuristic algorithm
is then developed to dispatch the PEB charging with the
continuous charging constraints involved and generate the
final charging state matrix of PEBs CPEB , given in Algorithm
1.
Algorithm 1. PEB Charging Dispatching
1: Initialization: Set CPEB 0 .
2: for k 1 to K do
3:
Find all the parked PEBs that are waiting for
charging and compute their LX n, k .
4:
5:
6:
7:
if LX k , n 0 then
PEB
PEB
Set cnPEB
, k cn , k 1 cn , k CT
n ,in
1
1 .
end if
Sort PEBs according to the increasing LX n, k 0
and for PEBs with same LX n, k , sort them
according to the decreasing CTn,in .
8:
9:
ˆ PEB then
if sum(CPEB ) C
ˆ PEB sum(CPEB ) PEBs according to the
Select C
PEB
order in line 7 and set cnPEB
, k cn , k 1
n
In the previous text, all the formulated models, i.e.,
Model A, Model B, and Model C, are mixed integer linear
programming (MILP) models. However, different from
Model C, Model A and Model B bear the extra computation
burdens caused by the introduced auxiliary variables un , k and
n parks at the PEBFCS, and an ,in and ln ,in are therefore
cnPEB
, k CTn ,i
4.3. Heuristics-based Method for Coordinated
PEB Charging Scenario with ESS
CTn,in
defined for parked PEBs that wait for charging, otherwise (if
the PEB is on charge or has been charged), the laxity is
invalid; 2) for a new time interval, laxity values should be
updated; 3) for a given k , a specific in can be found if PEB
1
1 .
10:
end if
11: end for
Armed with CPEB got by Algorithm 1, we are then
able to skip auxiliary variables un , k and vn , k and seek for the
ESS charging and discharging strategy through Model C.
Note that, in the heuristics-based method, the control
strategies of PEBs and ESS are generated separately and the
PEB coordinated charging is dispatched heuristically so that
the final strategy is not necessarily the optimal strategy.
5. Case Studies
5.1. Parameter Settings
The proposed strategies are tested on a PEBFCS with
10 fast charging piles, which provides charging service to a
loop PEB line with 24 PEBs. The settings of the PEBFCS are
on the basis of a practical PEBFCS in Chongqing, China. ESS
in the PEBFCS is composed of lithium titanate batteries, and
the discount rate, the life cycle, and the capacity charge for
the station are respectively set as 5% , 50 year and
Cap 14847 RMB/kW [25]. The circle length of the PEB
route is 50 km. According to the central limit theorem, we
assume the average speeds of PEBs (km/h) follow Gaussian
distribution N 50,52 , and all SOCnPEB
,in (kWh) follow
flexibility of the PEB charging. Note that 1) the laxity is
6
Table 1 Parameters settings of PEBs and ESS
SnPEB
(kWh)
cPEB
PcPEB
(kW)
PEB
SOCmin
S ESS
(kWh)
ESS
Pc,max
ESS
Pd,max
(kW)
(kW)
324
0.92
117
0.2
800
1000
1000
cESS
dESS
n ESS
ESS
SOCmin
0.92
0.92
15000 [26]
0.2
Table 2 PEB departure time-table
Time
Number*
Interval (min)
12
5
06:00-07:00
12
5
07:00-08:00
12
5
08:00-09:00
12
5
09:00-10:00
12
5
10:00-11:00
12
5
11:00-12:00
12
5
12:00-13:00
12
5
13:00-14:00
6
10
14:00-15:00
* The number of PEBs depart the PEBFCS at each time interval.
Table 3 TOU electricity prices
Time
Valley
Shoulder
High
Peak
Price (RMB/kWh)
23:00-7:00
7:00-10:00,
15:00-18:00,
21:00-23:00
10:00-15:00,
18:00-21:00
11:00-13:00,
20:00-21:00
0.3818
0.8395
1.3222
1.4409
Gaussian distribution N 70, 72 . Specific parameters of
PEBs and ESS are listed in Table 1, where the PEB
parameters partially refer to those of BYD K9 pure electric
bus [27]. The PEB departure time-table is listed in Table 2
and the TOU electricity tariffs are given in Table 3 [28]. The
typically load profile in [29] is selected as the other load curve
in the case. The optimising time horizon is set as 24h and
card K is 288, i.e., t=5min .
Based on the parameters settings above, we carry out
the simulations when the unit price of ESS, i.e., ESS , is
respectively 8000 RMB/kWh, 6000 RMB/kWh, 4000
RMB/kWh, and 2000 RMB/kWh.
5.2. Results and Analysis
Numerical simulation results under uncoordinated and
coordinated PEB charging scenarios are respectively
summarized in Tables 4 and 5 (including the optimal and
heuristics-based strategies), where AOC is the abbreviation
of annual operation costs (including the ESS expenditure
costs, equivalent capacity charge and electricity bills per
year). Typical daily load profile, SOC curve of ESS and
charging/discharging power curves of ESS under
uncoordinated PEB charging scenario are given in Figs. 3 and
4, and the corresponding profiles under coordinated PEB
charging scenario are shown in Figs. 5 and 6. All the
Time
Number*
Interval (min)
15:00-16:00
16:00-17:00
17:00-18:00
18:00-19:00
19:00-20:00
20:00-21:00
21:00-22:00
22:00-23:00
Other
6
12
12
4
4
4
2
1
0
10
5
5
15
15
15
30
60
-
problems were solved by the CPLEX package [30] on a
laptop with an Intel Core i5 processor and 4 GB randomaccess memory.
From Tables 4 and 5, by comparing with the results
without ESS, it can be seen that the proposed control
strategies effectively improve AOC of PEBFCS and peak
loads both under coordinated and uncoordinated PEB
charging scenarios. And as the price of ESS falls, AOC
decreases. According to whether charging loads of PEBs are
controllable and whether ESS is considered, there are four
different AOCs under each price of ESS. These four AOCs
demonstrate that the combination of coordination of PEB and
ESS achieves the best optimality. Besides, when loads of PEB
and ESS are both coordinated, it is observed that the
controlled peak load first decreases and then rebounds as the
price of ESS declines from 8000 RMB/kWh to 2000
RMB/kWh (both using optimal and heuristics-based
strategies). The reason is that electricity price arbitrage makes
more profits than to decrease the capacity charge when the
price of ESS is sufficiently low. While under the
uncoordinated PEB charging scenario, the controlled peak
load remains constant with the change of the ESS price
because all the capacity of ESS is used to shave the peak PEB
charging loads during the high and peak TOU price periods,
which brings larger benefits than ESS costs. Regarding the
different results of optimal and heuristics-based strategies in
Table 5, it can be observed that the heuristics-based strategy
gives rise to a slight increasing of AOC and peak loads, but
has a distinct advantage in computation speed. In practice, if
t=5min , both optimal and heuristics-based strategies are
applicable for a PEBFCS of such size, because all the
computation time is much shorter than t . If the PEBFCS
size increases or t decreases, the computation time could
matter and the heuristics-based strategy is probably a more
appropriate choice. For example, we double the PEBFCS size,
i.e., the PEB number increases to 48 and the departure
numbers in Table 2 are all doubled, and let t=3min , i.e.,
card K is 480, then the computation time of optimal
strategy exceeds
7
.
Table 4 Numerical simulation results under the uncoordinated PEB charging scenario
AOC
Peak Load
Reduction of The Average
Price of ESS
AOC (with Reduction of
Peak Load (with
(without
(without ESS)
Peak Load Computation
(RMB/kWh)
ESS) (RMB) AOC (%)
ESS) (kW)
ESS) (RMB)
(kW)
(%)
Time (s)
8000
3474289
8.58
1045.1
23.51
5.18
6000
3436731
9.57
1045.1
23.51
9.12
3800453
1366.4
4000
3374498
11.21
1045.1
23.51
9.34
2000
3312229
12.85
1045.1
23.51
9.23
Table 5 Numerical simulation results under the coordinated PEB charging scenario (optimal/heuristics-based)
AOC
Peak Load
Reduction of The Average
Price of ESS
AOC (with Reduction of
Peak Load (with
(without
(without ESS)
Peak Load Computation
(RMB/kWh)
ESS) (RMB) AOC (%)
ESS) (kW)
ESS) (RMB)
(kW)
(%)
Time (s)
1314657
8000
6.12/5.48
1057.1/1057.1
19.01/19.01
19.02/0.19
/1324911
1284581
6000
8.33/7.75
1009.2/1023.0
22.68/21.63
20.31/0.19
/1293059
1401710
1305.3
1236876
4000
11.76/11.24
1009.2/1023.2
22.68/21.61
22.24/0.21
/1243737
1181943
2000
15.68/15.31
1015.6/1024.7
22.19/21.50
24.51/0.21
/1187108
1400
1400
Without ESS
With ESS
1000
800
600
1000
800
600
400
400
200
200
0
0
4
8
Without ESS
With ESS
1200
Power(kW)
Power(kW)
1200
12
16
20
0
0
0
4
8
12
16
20
0
Time(h)
Time(h)
a
b
Fig. 3. Typical daily load profiles of the local distribution transformer under uncoordinated PEB charging scenario
(a) Price of ESS: 8000 RMB/kWh, (b) Price of ESS: 6000 RMB/kWh, 4000 RMB/kWh, and 2000 RMB/kWh
1000
1000
1
1
800
800
400
0.4
Charging Power of ESS
Discharging Power of ESS
SOC
200
0
0
4
8
12
Time(h)
16
0.2
20
0
0
Power(kW)
0.6
SOC
Power(kW)
600
600
0.6
400
0.4
Charging Power of ESS
Discharging Power of ESS
SOC
200
0
SOC
0.8
0.8
0
4
8
12
Time(h)
16
0.2
20
0
0
a
b
Fig. 4. Typical daily SOC curves and charging/discharging power curves of ESS under uncoordinated PEB charging scenario
(a) Price of ESS: 8000 RMB/kWh, (b) Price of ESS: 6000 RMB/kWh, 4000 RMB/kWh, and 2000 RMB/kWh
8
1400
1400
Without ESS
With ESS (Optimal)
With ESS (Heuristic)
1000
800
600
1000
800
600
400
400
200
200
0
0
4
8
12
16
20
Without ESS
With ESS (Optimal)
With ESS (Heuristic)
1200
Power(kW)
Power(kW)
1200
0
0
0
4
8
a
20
0
1400
Without ESS
With ESS (Optimal)
With ESS (Heuristic)
1000
800
600
1000
800
600
400
400
200
200
0
4
8
12
16
20
Without ESS
With ESS (Optimal)
With ESS (Heuristic)
1200
Power(kW)
1200
Power(kW)
16
b
1400
0
12
Time(h)
Time(h)
0
0
0
4
8
12
16
20
0
Time(h)
Time(h)
c
d
Fig. 5. Typical daily load profiles of the local distribution transformer under coordinated PEB charging scenario
(a) Price of ESS: 8000 RMB/kWh, (b) Price of ESS: 6000 RMB/kWh, (c) Price of ESS: 4000 RMB/kWh, (d) Price of ESS:
2000 RMB/kWh
1000
Charging Power of ESS
1
Discharging Power of ESS
SOC
0.8
0.6
400
0.4
200
4
8
12
Time(h)
16
20
0.6
400
0.4
200
0.2
0
600
0
0
0
0.2
0
4
8
a
20
0
0
1000
Charging Power of ESS
1
Discharging Power of ESS
SOC
0.8
0.6
400
0.4
200
0.2
0
4
8
12
Time(h)
16
20
0
0
Power(kW)
600
Charging Power of ESS
1
Discharging Power of ESS
0.8
SOC
800
SOC
800
Power(kW)
Time(h)
16
b
1000
0
12
600
0.6
400
0.4
200
0
SOC
0
Power(kW)
600
Charging Power of ESS
1
Discharging Power of ESS
SOC
0.8
800
SOC
Power(kW)
800
SOC
1000
0.2
0
4
8
12
Time(h)
16
20
0
c
d
Fig. 6. Typical daily SOC curves of ESS and charging/discharging power curves of ESS under coordinated PEB charging
scenario (optimal)
9
(a) Price of ESS: 8000 RMB/kWh, (b) Price of ESS: 6000 RMB/kWh, (c) Price of ESS: 4000 RMB/kWh, (d) Price of ESS:
2000 RMB/kWh
1000
Charging Power of ESS
1
Discharging Power of ESS
SOC
0.8
0.6
400
0.4
200
4
8
12
Time(h)
16
20
0.6
400
0.4
200
0.2
0
600
0
0
0
0.2
0
4
8
a
20
0
0
1000
Charging Power of ESS
1
Discharging Power of ESS
SOC
0.8
0.6
400
0.4
200
0.2
0
4
8
12
Time(h)
16
20
0
0
Power(kW)
600
Charging Power of ESS
1
Discharging Power of ESS
SOC
0.8
800
SOC
800
Power(kW)
16
b
1000
0
12
Time(h)
600
0.6
400
0.4
200
0
SOC
0
Power(kW)
600
Charging Power of ESS
1
Discharging Power of ESS
SOC
0.8
800
SOC
Power(kW)
800
SOC
1000
0.2
0
4
8
12
Time(h)
16
20
0
0
c
d
Fig. 7. Typical daily SOC curves of ESS and charging/discharging power curves of ESS under coordinated PEB charging
scenario (heuristics-based)
(a) Price of ESS: 8000 RMB/kWh, (b) Price of ESS: 6000 RMB/kWh, (c) Price of ESS: 4000 RMB/kWh, (d) Price of ESS:
2000 RMB/kWh
t=3min but the time of heuristics-based strategy is within
5s. Thus, under this scenario with larger station size and more
frequent strategy update rate, heuristics-based strategy still
works while the optimal strategy becomes unpractical due to
over-long calculation time.
Comparing the load profiles in Figs. 3 and 5, we can
see that 1) when there is no ESS, the coordinated PEB
charging shifts the peak loads from day (high electricity
prices) to night (low electricity prices); 2) when ESS is taken
into account, the control strategies, including the optimal and
heuristics-based (only for coordinated PEB charging scenario)
strategies, smooth the load profiles under both two scenarios.
Recall that under uncoordinated PEB charging
scenario, ESS is made full use of and the main restriction of
further costs decreasing is the capacity. Due to this, daily load
profile, SOC curve of ESS and charging/discharging power
curves of ESS are identical when the price of ESS is
respectively 6000 RMB/kWh, 4000 RMB/kWh and 2000
RMB/kWh (see Figs. 3 and 4). The reason, why the profiles
under the ESS price 8000 RMB/kWh differ, is that charging
ESS in the shoulder TOU price period and discharging ESS
in the peak TOU price period are not profitable at such an
ESS price (can be observed by comparing subfigures in Figs.
3 and 4). In other words, there is a threshold value of ESS
price, which is between 6000 RMB/kWh and 8000
RMB/kWh, and PEBFCS can make profits through electricity
price arbitrage between shoulder and peak TOU prices if the
ESS price is lower than the value, otherwise larger electricity
price difference, e.g., the price difference between peak and
valley TOU prices, is needed.
Under coordinated PEB charging scenario, the
subfigures in Fig. 5 illustrate that the total load profile of
PEBFCS with ESS becomes more and more flatter as the
price of ESS falls, and the subfigures in Figs. 6 and 7 show
that the usage frequency of ESS trends to increase as the price
of ESS falls. Besides, it is observed that charging and
discharging of ESS both occur in the valley period of
electricity price (see Figs. 6 and 7). And as a result, the night
peak loads are further flattened, which implies that economic
losses caused by energy consumption during the charging and
discharging process are less than the reduction of capacity
charge. Also, Figs. 5-7 illustrate that the load profiles, SOC
curves of ESS, and charging/discharging power curves of
ESS are similar when using optimal and heuristics-based
strategies.
5.3. Sensitivity Analysis of ESS Capacity
We study the impact of the ESS capacity on AOC
when using the optimal control strategy under PEB
coordinated charging scenario (the impacts are similar for
different strategies and scenarios). The ESS prices 6000
RMB/kWh and 4000 RMB/kWh are selected for the
sensitivity analysis, and the results are given in Fig. 8. It can
10
be seen that AOC decreases and converges to a constant as
the ESS capacity increases. However, when the AOC stops
declining, the extra capacity of idle ESS will lead to
unnecessary investment, which is not included in EOA. So,
in practice, the ESS capacity at the stop point of AOC
decreasing is most appropriate for installation. In Fig. 8, the
best ESS capacity at ESS prices 6000 RMB/kWh and 4000
RMB/kWh are both between 1000 kWh and 1200 kWh.
x 10
[8]
[9]
6
1.4
AOC (RMB)
[7]
ESS Price: 6000 RMB/kWh
ESS Price: 4000 RMB/kWh
[10]
1.35
[11]
1.3
[12]
[13]
1.25
1.22
[14]
0
400
800
1200
1600
ESS Capacity (kWh)
Fig. 8. Curves of sensitivity analysis of ESS capacity
[15]
6. Conclusion
This paper proposes coordinated charging and
discharging strategies for a PEBFCS with ESS to optimise the
economic benefits. The mathematical models are respectively
formulated when the PEB charging loads are controllable or
not. And when PEB charging loads are controllable, i.e.,
under coordinated PEB charging scenario, a heuristics-based
strategy is further proposed to enhance the computation
efficiency with a little sacrifice of optimality. We validate the
effectiveness of the proposed strategies under multiple ESS
prices through case studies, and analyze the impacts of ESS
capacity on the PEBFCS operation costs. Further research
includes the cooperation strategy for several PEBFCSs with
ESS.
7. Acknowledgments
This work was supported in part by the National
Natural Science Foundation of China (51477082) and the
National Key Research and Development Program
(2016YFB0900103).
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
8. References
[1]
[2]
[3]
[4]
[5]
[6]
Boulanger, A. G., Chu, A. C., Maxx, S., et al.: ‘Vehicle electrification:
status and issues’, Proceedings of the IEEE, 2011, 99, (6), pp. 1116–
1138.
Deng, J., Shi, J., Liu, Y. et al.: ‘Application of a hybrid energy storage
system in the fast charging station of electric vehicles.’ IET Generation,
Transmission & Distribution, 2016, 10, (4), pp. 1092-1097.
Tang, Y., Zhong J., Bollen, M.: ‘Aggregated optimal charging and
vehicle-to-grid control for electric vehicles under large electric vehicle
population’, IET Generation, Transmission & Distribution, 2016, 10,
(8), pp. 2012-2018.
Lopes, J. A. P., Soares, F. J., Almeida, P. M. R.: ‘Integration of electric
vehicles in the electric power system’, Proceedings of the IEEE, 2010,
99, (1), pp. 168-183.
Vicini, R., Micheloud, O., Kumar H., et al.: ‘Transformer and home
energy management systems to lessen electrical vehicle impact on the
grid’, IET Generation, Transmission & Distribution, 2012, 6, (12) pp.
1202-1208.
Fernandez, L. P., Román, T. G. S., Cossent, R., et al.: ‘Assessment of
the impact of plug-in electric vehicles on distribution networks’, IEEE
Transactions on Power Systems, 2011, 26, (1), pp. 206–213.
[24]
[25]
[26]
[27]
[28]
[29]
[30]
Amini, M. H., Boroojeni, K. G., Wang, C. J., et al.: ‘Effect of electric
vehicle parking lots’ charging demand as dispatchable loads on power
systems loss’, IEEE International Conference on Electro Information
Technology, Grand Forks, North Dakota, USA, May, 2016.
Dow, L., Marshall, M., Xu, L., et al.: ‘A novel approach for evaluating
the impact of electric vehicles on the power distribution system’, IEEE
Power and Energy Society General Meeting, Minneapolis, Minnesota,
USA, July, 2010.
Staats, P. T., Grady, W. M., Arapostathis, A., et al.: ‘A procedure for
derating a substation transformer in the presence of widespread electric
vehicle battery charging’, IEEE Transactions on Power Delivery, 1997,
12, (4), pp. 1562–1568.
Singh, M., Kar, I., Kumar, P.: ‘Influence of EV on grid power quality
and optimizing the charging schedule to mitigate voltage imbalance
and reduce power loss’, Power Electronics and Motion Control
Conference, Ohrid, Macedonia, Sep., 2010.
Putrus, G. A., Suwanapingkarl, P., Johnston, D., et al.: ‘Impact of
electric vehicles on power distribution networks’, IEEE Vehicle Power
and Propulsion Conference, Dearborn, Michigan, USA, Sep., 2009.
Xu, Z., Hu, Z., Song, Y., et al.: ‘Coordination of PEVs charging across
multiple aggregators’, Applied Energy, 2014, 136, pp. 582–589.
Rotering, N., Ilic, M.: ‘Optimal charge control of plug-in hybrid
electric vehicles in deregulated electricity markets’, IEEE Transactions
on Power Systems, 2011, 26, (3), pp. 1021–1029.
Masoum, A. S., Deilami S., Moses, P. S., et al.: ‘Smart load
management of plug-in electric vehicles in distribution and residential
networks with charging stations for peak shaving and loss minimisation
considering voltage regulation’, IET Generation, Transmission &
Distribution, 2011, 5, (8), pp. 877–888.
Yu, Z., Chen, S., Tong, L.: ‘An intelligent energy management system
for large-scale charging of electric vehicles’, CSEE Journal of Power
and Energy Systems, 2016, 2, (1), pp. 47-53.
Yao, W., Zhao, J., Wen, F., et al.: ‘A hierarchical decomposition
approach for coordinated dispatch of plug-in electric vehicles’, IEEE
Transactions on Power Systems, 2013, 28, (3), pp. 2768–2778.
Han, S., Han, S., Sezaki, K.: ‘Development of an optimal vehicle-togrid aggregator for frequency regulation’, IEEE Transactions on Smart
Grid, 2010, 1, (1), pp. 65–72.
Chen, H., Hu, Z., Xu, Z., et al.: ‘Coordinated charging strategies for
electric bus fast charging stations’, IEEE Power Energy Society AsiaPacific Power and Energy Engineering Conference, Xi’an, Shaanxi
China, Oct., 2016.
Ding, H., Hu, Z., Song, Y.: ‘Value of the energy storage system in an
electric bus fast charging station’, Applied Energy, 2015, 157, pp. 630–
639.
Sun, B., Dragičević, T., Freijedo, F. D., et al.: ‘A control algorithm for
electric vehicle fast charging stations equipped with flywheel energy
storage systems’, IEEE Transactions on Power Electronics, 2016, 31,
(9), pp. 6685–6674.
Ding, H., Hu, Z., Song, Y., et al.: ‘Coordinated control strategy of
energy storage system with electric vehicle charging station’, IEEE
Transportation Electrification Conference and Expo, Asia-Pacific,
Beijing, China, Aug., 2014.
Jin, C., Tang J., Ghosh, P.: ‘Optimizing electric vehicle charging with
energy storage in the electricity market’, IEEE Transactions on Smart
Grid, 2013, 4, (1), pp. 311–320.
Petersen, M. K., Edlund, K., Hansen, L. H., et al.: ‘A taxonomy for
modeling flexibility and a computationally efficient algorithm for
dispatch in smart grids’, American Control Conference, Washington,
D.C., USA, June, 2013, pp. 1150-1156.
Xu, Y., Pan, F., Tong, L.: ‘Dynamic scheduling for charging electric
vehicles: A priority rule’, IEEE Transactions on Automatic Control,
2016, 61, (12), pp. 4094-4099.
Luo, Z., Hu, Z., Song, Y., et al.: ‘Optimal coordination of plug-in
electric vehicles in power grids with cost-benefit analysis—Part I:
Enabling techniques’, IEEE Transactions on Power Systems, 2013, 28,
(4), pp. 3546-3555.
‘AceOn Group’ http://www.aceongroup.com/divisions, accessed 27
Oct. 2017.
‘BYD K9-CPTDB Wiki’, https://cptdb.ca/wiki/index.php/BYD_Auto
_K9, accessed 17 Feb. 2017.
‘The Industrial Time-of-use Electricity Prices in Beijing’,
http://www.docin.com/p-1140670657.html, accessed 17 Feb. 2017.
Xu, Z., Hu, Z., Song, Y., et al.: ‘Coordinated charging strategy for
PEVs charging stations’, IEEE Power and Energy Society General
Meeting, San Diego, California, USA, July, 2012.
‘ILOG CPLEX Homepage’, http://www.ilog.com, accessed 17 Feb.
2017.
11
12
| 3cs.SY
|
CRYSTALS AND TREES: QUASI-KASHIWARA
OPERATORS, MONOIDS OF BINARY TREES, AND
ROBINSON–SCHENSTED-TYPE CORRESPONDENCES
arXiv:1702.02998v2 [math.CO] 1 Feb 2018
ALAN J. CAIN AND ANTÓNIO MALHEIRO
Abstract. Kashiwara’s crystal graphs have a natural monoid structure
that arises by identifying words labelling vertices that appear in the same
position of isomorphic components. The celebrated plactic monoid (the
monoid of Young tableaux), arises in this way from the crystal graph
for the q-analogue of the general linear Lie algebra gln , and the socalled Kashiwara operators interact beautifully with the combinatorics
of Young tableaux and with the Robinson–Schensted–Knuth correspondence. The authors previously constructed an analogous ‘quasi-crystal’
structure for the related hypoplactic monoid (the monoid of quasi-ribbon
tableaux), which has similarly neat combinatorial properties. This paper constructs an analogous ‘crystal-type’ structure for the sylvester and
Baxter monoids (the monoids of binary search trees and pairs of twin binary search trees, respectively). Both monoids are shown to arise from
this structure just as the plactic monoid does from the usual crystal
graph. The interaction of the structure with the sylvester and Baxter
versions of the Robinson–Schensted–Knuth correspondence is studied.
The structure is then applied to prove results on the number of factorizations of elements of these monoids, and to prove that both monoids
satisfy non-trivial identities.
1. Introduction
Kashiwara’s notion of a crystal basis [Kas90, Kas91] (informally, a basis
of a representation for a suitable algebra on which the generators have a
particularly straightforward action), gives rise, via tensor products, to the
crystal graph. This graph has a natural monoid structure where the vertices
are viewed as words in the free monoid, and words are identified when they
lie in the same places in isomorphic components of this graph. In the case of
the q-analogue of the general linear Lie algebra gln , the monoid obtained is
the celebrated plactic monoid, whose elements can be identified with Young
tableaux and which appears in such diverse contexts as symmetric functions
[Mac08], representation theory [Gre06], algebraic combinatorics [Lot02], algebraic geometry [Ful97], Kostka–Foulkes polynomials [LS81, LS78], Schubert polynomials [LS85, LS90], and musical theory [Jed11]. (Indeed, since
The first author was supported by an Investigador FCT fellowship
(IF/01622/2013/CP1161/CT0001).
For both authors, this work was partially supported by by the Fundação para a
Ciência e a Tecnologia (Portuguese Foundation for Science and Technology) through the
project UID/MAT/00297/2013 (Centro de Matemática e Aplicações), and the project
PTDC/MHC-FIL/2583/2014.
The authors thank the anonymous referee for carefully reading the paper and for valuable comments, and, in particular, for making the observation contained in Remark 3.4.
1
2
ALAN J. CAIN AND ANTÓNIO MALHEIRO
Lascoux and Schützenberger’s [LS81] seminal study, so many connections
have emerged that Schützenberger proclaimed it ‘one of the most fundamental monoids in algebra’ [Sch97].) The beautiful interaction of the crystal graph and the associated Kashiwara operators with the combinatorics
of Young tableaux and the Robinson–Schensted correspondence is such an
important and powerful combinatorial tool that in this context the Kashiwara operators and crystal graph are sometimes simply referred to as the
‘coplactic’ operators and ‘coplactic’ graph (see, for instance, [Lot02, § 5.5]
and [vL01]).
In a previous paper [CM17], the present authors constructed an analogue of this crystal structure for the monoid of quasi-ribbon tableaux: the
so-called hypoplactic monoid. The hypoplactic monoid emerged from the
theory of non-commutative symmetric functions and quasi-symmetric functions [KT97, KT99, Nov00]. Quasi-ribbon functions form a basis for the
ring of quasi-symmetric functions, just as the Schur polynomials form a basis for the ring of symmetric polynomials. The quasi-ribbon functions are
indexed by the so-called quasi-ribbon tableaux, which can be identified with
the elements of the hypoplactic monoid. As shown in the authors’ previous paper, much of the elegant interaction of the crystal graph, Kashiwara
operators, Young tableaux, and the plactic monoid is echoed in the interaction of the analoguous quasi-crystal graph, quasi-Kashiwara operators,
quasi-ribbon tableaux, and the hypoplactic monoid.
Thus both plactic and hypoplactic monoids can be defined by factoring
the free monoid by a relation that can be defined in three equivalent ways:
1. Defining relations: the relation is the smallest congruence containing
a given set of defining relations.
2. Insertion: the relation is the kernel of the map that takes a word to
a ‘tableau-type’ object computed by an insertion algorithm.
3. Crystals: the relation arises from isomophisms of connected components of a ‘crystal-type’ graph.
The main aim of this paper is to develop an analoguous of the crystaltype structure for the sylvester [HNT05] and Baxter [Gir12] monoids, whose
elements are respectively right strict binary search trees and pairs of twin
binary search trees. To be more precise: the existing literature contains
equivalent definitions for these monoid using defining relations and insertions
algorithms (approaches 1 and 2 above). This paper develops and applies the
crystalline definition (approach 3 above).
In fact, the crystal-type graph for the sylvester and Baxter monoids are
the same as for the hypoplactic monoid: it is the notion of isomorphism
that has to be modified. This paper develops an abstract framework for
this modified notion of isomorphism (Section 5), and then applies it to the
sylvester and Baxter monoids (Sections 6 and 7).
This framework is developed from a slightly different perspective from the
authors’ earlier paper on the hypoplactic monoid. Essentially, the difference
is that rather than having finitely many Kashiwara and ‘quasi-Kashiwara’
operators, countably many are allowed, and the resulting monoids have infinite rank rather than finite rank. This is a step away from the representationtheoretic origin of crystal bases, but is actually a much more natural context
CRYSTALS AND TREES
3
for a combinatorially-focussed study. More importantly, it seems to be the
correct starting-point to crystallize other monoids connected with combinatorics, such as sylvester and Baxter monoids. In order to discuss the
advantages of these modified graphs and to clarify what properties a ‘crystal structure’ for the sylvester and Baxter monoids should enjoy, Section 3
briefly recapitulates some of the existing theory. Although some of this material is in existing literature, the important differences between the finiteand infinite-rank cases are often glossed over.
The resulting crystal-type structure for the sylvester and Baxter monoids
turns out to interact neatly with the sylvester and Baxter versions of the
Robinson–Schensted correspondence, and yield new results.
2. Preliminaries and notation
Much of the notation used in this paper is drawn from [CM17], to which
this paper is a sequel. For definitions and notation for alphabets, words, presentations, standard words, standardization, partitions, compositions, and
weight see [CM17, § 2]. For Young tableaux, the Schensted insertion algorithm, and the definition of the plactic monoid plac and rank-n plactic
monoid placn using these, see [CM17, § 3.1]; note however that the present
paper uses Pplac (u) and Qplac (u) respectively for the Young tableau and
recording tableau computed from u ∈ A∗ and calls these the plactic P- and
Q-symbols of u. Similarly Phypo (u) and Qhypo (u) are respectively the quasiribbon tableau and recording ribbon computed from u ∈ A∗ and calls these
the hypoplactic P- and Q-symbols of u. This is to keep notation uniform
with the analogous notions for the sylvester and Baxter monoids.
Note that the defining relations Rplac given in [CM17, § 1] (known as
the Knuth relations) are the reverse of the ones given in [CGM]. This is
because, in the context of crystal bases, the convention for tensor products
gives rise to a ‘plactic monoid’ that is actually anti-isomorphic to the usual
notion of plactic monoid. Since [CM17] and the present paper are mainly
concerned with combinatorics, rather than representation theory, they follow
Shimozono [Shi05] in using the convention that is compatible with the usual
notions of Young tableaux and the Robinston–Schensted correspondence.
For quasi-ribbon tableaux and the Schensted insertion algorithm, and the
definition of the hypoplactic monoid hypo and rank-n hypoplactic monoid
hypon using these, see [CM17, § 4].
3. Infinite-rank crystals and quasi-crystals
This section recalls the definitions of the Kashiwara and quasi-Kashiwara
operators, defines the infinite-rank crystal and quasi-crystal graphs, and
shows that isomorphisms between components of these graphs give rise,
respectively to the plactic and hypoplactic monoids. Many of the proofs in
this section are not given in full, since they are straightforward modifications
of the proofs for finite-rank crystal and quasi-crystal graphs given in [CM17,
§§ 3 and 5].
4
ALAN J. CAIN AND ANTÓNIO MALHEIRO
3.1. Crystal graph and the plactic monoid. Let i ∈ N and define the
Kashiwara operators ẽi and f˜i on A as follows:
ẽi (i + 1) = i,
f˜i (i) = i + 1,
ẽi (x) is undefined for x 6= i + 1;
f˜i (x) is undefined for x =
6 i.
The definition is extended to A∗ \ A by the recursion
(
ẽi (u) v if ǫ̃i (u) > φ̃i (v);
ẽi (uv) =
u ẽi (v) if ǫ̃i (u) ≤ φ̃i (v),
(
f˜i (u) v if ǫ̃i (u) ≥ φ̃i (v);
f˜i (uv) =
u f˜i (v) if ǫ̃i (u) < φ̃i (v),
where ǫ̃i and φ̃i are auxiliary maps defined by
ǫ̃i (w) = max k ∈ N ∪ {0} : ẽi · · · ẽi (w) is defined ;
| {z }
k times
φ̃i (w) = max k ∈ N ∪ {0} : f˜i · · · f˜i (w) is defined .
| {z }
k times
(Note that this definition is in a sense the mirror image of [KN94, Theorem 1.14], because of the choice of definition for readings of tableaux used
in this paper. Thus the definition of ẽi and f˜i is the same as in [CM17] and
[Shi05, p. 8].)
The operators ẽi and f˜i respectively increase and decrease weight whenever they are defined, in the sense that if ẽi (u) is defined, then wt(ẽi (u)) >
wt(u), and if f˜i (u) is defined, then wt(f˜i (u)) < wt(u). Thus ẽi and f˜i are
respectively called the Kashiwara raising and lowering operators. Furthermore, a word on which none of the ẽi is defined is said to be highest-weight.
The crystal graph for plac, denoted Γ(plac), is the directed labelled graph
with vertex set A∗ and, for u, v ∈ A∗ , an edge from u to v labelled by i if and
only if v = f˜i (u) (or, equivalently, u = ẽi (v)). Figure 1 shows part of the
crystal graph Γ(plac). For any w ∈ A∗ , let Γ(plac, w) denote the connected
component of Γ(plac) that contains the vertex w. For n ∈ N, the crystal
graph for placn , denoted Γ(placn ), is the subgraph of Γ(plac) induced by A∗n .
Notice that edge labels in Γ(placn ) must lie in {1, . . . , n − 1}, since for i ≥ n,
if f˜i (u) is defined, then at least one of u and f˜i (u) does not lie in A∗n .
Define a relation ≈ on A∗ as follows: u ≈ v if and only if there is a labelled
digraph isomorphism θ : Γ(plac, u) → Γ(plac, v) with θ(u) = v. That is,
u ≈ v if u and v lie in corresponding places in isomorphic components of
Γ(plac).
It is worth emphasizing that the definition of ≈ is the crucial point of
divergence from the standard approach to crystals: normally, as in [CGM,
§ 2.4], ≈ is defined in terms of weight-preserving isomorphisms.
Proposition 3.1 ([CGM, Proposition 2.6]). The relation ≈ is a congruence
on A∗ .
(Strictly speaking, the preceding result follows by modifying the proof of
[CGM, Proposition 2.6] by replacing An with A and ignoring the discussion
of weight.)
CRYSTALS AND TREES
5
2111
1
2
3111
1
2112
3
3112
1
3122
1
3222
1
2
4111
3
1
4112
3
1
4122
2113
4
2
5111
2
4
4113
2
1
5112
6111
3
3113
1
5
2122
3123
1
2
2114
3
2
3114
2123
4
1
2115
2
3
2133
2124
Figure 1. Part of the connected component Γ(plac, 2111).
Theorem 3.2 ([Lot02, Theorem 5.5.1]). Let u, v ∈ A∗ .
(1) Qplac (u) = Qplac (ẽi (u)) if ẽi (u) is defined and similarly for f˜i .
(2) Suppose u ≡plac v. Then ẽi (u) is defined if and only if ẽi (v) is
defined, in which case ẽi (u) ≡plac ẽi (v) and similarly for f˜i .
(To be strict, [Lot02, Theorem 5.5.1] is concerned only with words in A∗n
and i ∈ {1, . . . , n − 1}. To obtain the result as stated here, one simply
chooses n to be bigger than the maximum symbol that appears in u or v.)
Proposition 3.3. The relations ≡plac and ≈ coincide.
Proof. Let u ∈ A∗ and i ∈ N. By iterated application of Theorem 3.2(2), if
u, v ∈ A∗ are such that u ≡plac v, then there is a labelled digraph isomorphism θ : Γ(plac, u) → Γ(plac, v) with θ(u) = v.
For the converse, let u, v ∈ A∗ and suppose there is a labelled digraph
isomorphism θ : Γ(plac, u) → Γ(plac, v) with θ(u) = v. If ẽi (u) is defined for
some i ∈ N, so is ẽi (v) (since θ is a labelled digraph isomorphism); thus one
can replace u and v with ẽi (u) and ẽi (v). Eventually one obtains highest
weight words u′ and v ′ . By the proof of [Lot02, Proposition 5.5.2], u′ and
v ′ are Yamanouchi words.
Suppose n is the largest symbol appearing in u′ . Then |u′ |n is the length of
the longest directed path in Γ(plac) made up of edges labelled by n starting at
|u′ |
u′ , or equivalently, φ̃n (u′ ). Furthermore, |u′ |n−1 = |f˜n n (u′ )|n−1 is equal to
|u′ |
φ̃n−1 (f˜n n (u′ )). Continuing in this way, one sees that for each i ∈ {1, . . . , n}
′
′
|u′ |
|u′ |
|u′ |i = |f˜i+1i+1 · · · f˜n|u |n (u′ )|i = φ̃i f˜i+1i+1 · · · f˜n|u |n (u′ ) .
Hence wt(u′ ) is determined by the connected component Γ(plac, u′ ). Similarly, wt(v ′ ) is determined by the connected component Γ(plac, v ′ ).
6
ALAN J. CAIN AND ANTÓNIO MALHEIRO
Since the components Γ(plac, u′ ) and Γ(plac, v ′ ) are isomorphic, wt(u′ ) =
wt(v ′ ). Thus, since u′ and v ′ are Yamanouchi words, Pplac (u′ ) and Pplac (v ′ )
are both equal to the tableaux of shape wt(u′ ) = wt(v ′ ) containing only symbols i on row i. Hence u′ ≡plac v ′ . By iterated application of Theorem 3.2(2),
w ≡plac θ(w) for all w ∈ Γ(plac, u).
In the finite rank case, the analogy of Proposition 3.3 does not hold: for
example, the connected components Γ(placn , ε) and Γ(placn , n(n − 1) · · · 21)
both consist of single vertices and thus are isomorphic as labelled digraphs.
(Thus [Lot02, Proposition 5.5.2] is technically false as stated.) This is the
crucial difference between the infinite-rank and finite-rank cases: In the
infinite-rank case, all necessary information about weights is contained in
the abstract graph Γ(plac) and so it suffices to consider labelled digraph
isomorphisms between connected components. However, in the finite-rank
case, the abstract graph Γ(placn ) does not contain sufficient information to
define the plactic monoid placn , and so it is necessary to restrict to weightpreserving labelled digraph isomorphisms.
Remark 3.4. In fact, the abstract graph Γ(placn ) is the crystal graph
for sln . This corresponds to the fact that quotienting by the full column
n(n − 1) · · · 21 projects from isomorphism classes of gln crystal graphs to
isomorphism classes of sln crystal graphs. The only extra information one
requires to reverse this projection is how many full columns are present.
3.2. Quasi-crystal graph and the hypoplactic monoid. Let i ∈ N.
Define the quasi-Kashiwara operators ëi and f¨i on A∗ as follows: Let u ∈ A∗ .
• If u contains a subsequence (i + 1)i, both ëi (u) and f¨i (u) are undefined.
• If u does not contain a subsequence (i + 1)i, but u contains at least
one symbol i+1, then ëi (u) is the word obtained from u by replacing
the left-most symbol i + 1 by i; if u contains no symbol i + 1, then
ëi (u) is undefined.
• If u does not contain a subsequence (i + 1)i, but u contains at least
one symbol i, then f¨i (u) is the word obtained from u by replacing
the right-most symbol i by i + 1; if u contains no symbol i, then
f¨i (u) is undefined.
(As in [CM17], a subsequence may be made up of non-consecutive letters;
thus the phrase ‘contains a subsequence (i + 1)i’ is equivalent to ‘contains a
symbol i + 1 somewhere to the left of a symbol i’.) For example,
ë2 (3123) is undefined since 3123 contains the subsequence 32;
f¨2 (3131) is undefined since 3131 does not contains a symbol 2;
f¨1 (3113) = 3123.
Lemma 3.5 ([CM17, Lemma 1]). For all i ∈ N, the operators ëi and f¨i are
mutually inverse, in the sense that if ëi (u) is defined, u = f¨i (ëi (u)), and if
f¨i (u) is defined, u = ëi (f¨i (u)).
The operators ëi and f¨i respectively increase and decrease weight whenever they are defined, in the sense that if ëi (u) is defined, then wt(ëi (u)) >
CRYSTALS AND TREES
7
1212
2
1213
3
2
1214
4
1313
2
1215
5
1314
2
1216
6
1217
1
3
4
1323
1315
2
5
1316
1
3
1414
3
1
1324
3
1
1
4
1325
4
1415
2323
3
1424
1
3
2324
Figure 2. Part of the connected component Γ(hypo, 1212).
wt(u), and if f¨i (u) is defined, then wt(f¨i (u)) < wt(u). Thus ëi and f¨i
are respectively called the quasi-Kashiwara raising and lowering operators.
Furthermore, a word on which none of the ëi is defined is said to be highestweight. (Strictly speaking, it is necessary to distinguish words that are
highest-weight with respect to the operators ëi and words that are highestweight with respect to the operators ẽi . Since the ẽi are not mentioned in
the remainder of the paper, the term ‘highest-weight’ will always be with
respect to the operators ëi .)
The quasi-crystal graph for hypo, denoted Γ(hypo), is the directed labelled
graph with vertex set A∗ and, for u, v ∈ A∗ , an edge from u to v labelled by
i if and only if v = f˜i (u) (or, equivalently, u = ẽi (v)). Figure 2 shows part of
the crystal graph Γ(hypo). For any w ∈ A∗ , let Γ(hypo, w) denote the connected component of Γ(hypo) that contains the vertex w. The quasi-crystal
graph for hypon , denoted Γ(hypon ), is the subgraph of Γ(hypo) induced by
A∗n . Notice that edge labels in Γ(hypon ) must lie in {1, . . . , n − 1}, since for
i ≥ n, if f¨i (u) is defined, then at least one of u and f¨i (u) does not lie in A∗n .
Define a relation ∼ on A∗ as follows: u ∼ v if and only if there is a labelled
digraph isomorphism θ : Γ(hypo, u) → Γ(hypo, v) with θ(u) = v. That is,
u ∼ v if u and v lie in corresponding places in isomorphic components of
Γ(hypo). For instance, as can be seen in Figure 3, 1241 ∼ 2141.
Again, note that the definition of ∼ is the point of divergence from the previous definition in [CM17], where ∼ is defined in terms of weight-preserving
isomorphisms.
8
ALAN J. CAIN AND ANTÓNIO MALHEIRO
By the proof of [CM17, Proposition 4], replacing An with A and ignoring
the discussion of weight, the relation ∼ is a congruence on A∗ .
Proposition 3.6 ([CM17, Proposition 5]). Let u ∈ A∗ and i ∈ N.
(1) If ëi (u) is defined, then std(ëi (u)) = std(u).
(2) If f¨i (u) is defined, then std(f¨i (u)) = std(u).
In contrast to Proposition 3.6, the usual Kashiwara operators do not preserve standardization: ẽ2 (1322) = 1323, but std(1322) = 1423 6= 1324 =
std(1323).
Proposition 3.7. The relations ≡hypo and ∼ coincide.
Proof. Let u, v ∈ A∗ .
Suppose that u ≡hypo v. Using the reasoning that leads to the proof of
[CM17, Proposition 8], but replacing An by A, proves that u ∼ v.
Now suppose that there is a labelled digraph isomorphism θ : Γ(hypo, u) →
Γ(hypo, v) with θ(u) = v. Then there is a sequence of operators ëi1 , . . . , ëik
such that the words u′ = ëi1 · · · ëik (u) and v ′ = ëi1 · · · ëik (v) are highestweight. By the same reasoning as in Proposition 3.3, wt(u′ ) and wt(v ′ )
are determined by the isomorphic connected components Γ(hypo, u) and
Γ(hypo, v), and thus wt(u′ ) = wt(v ′ ). Since u′ ≡hypo C(Phypo (u′ )), and
v ′ ≡hypo C(Phypo (v ′ )) by [CM17, Lemmata 5 and 6] (replacing An by A),
and C(Phypo (u′ )) and C(Phypo (v ′ )) are highest-weight (since ≡hypo ⊆ ∼hypo
by the first part) and have the same weight, C(Phypo (u′ )) = C(Phypo (v ′ )) by
[CM17, Corollary 3]. Hence u′ ≡hypo C(Phypo (u′ )) = C(Phypo (v ′ )) ≡hypo v ′ .
By [CM17, Lemma 6(2)], applying f¨i to ≡hypo -related words yields ≡hypo related words. Hence u ≡hypo v.
In the finite rank case, the analogy of Proposition 3.7 does not hold: the
connected components Γ(hypon , ε) and Γ(hypon , n(n−1) · · · 21) both consist
of single vertices and thus are isomorphic as labelled digraphs. This is again
the crucial difference between the infinite-rank and finite-rank cases: In the
infinite-rank case, all necessary information about weights is contained in
the abstract graph Γ(hypo) and so it suffices to consider labelled digraph
isomorphisms between connected components. However, in the finite-rank
case, the abstract graph Γ(hypon ) does not contain sufficient information,
and so it is necessary to restrict to weight-preserving labelled digraph isomorphisms.
There is a very neat interaction between the structure of Γ(hypo) and the
hypoplactic analogue of the Robinson–Schensted correspondence; the proof
for the finite-rank case applies in the infinite-rank case:
Theorem 3.8 ([CM17, Theorem 2]). Let u, v ∈ A∗ . The words u and v lie in
the same connected component of Γ(hypo) if and only if Qhypo (u) = Qhypo (v).
Thus the recording ribbons Qhypo (·) produced by the insertion algorithm
(see the exposition in [CM17, § 4]) index connected components of Γ(hypo).
However, in the infinite-rank case, connected components are also indexed
by standard words, as shown in Corollary 3.10 below.
Lemma 3.9. For any u ∈ A∗ , the equality Qhypo (std(u)) = Qhypo (u) holds.
CRYSTALS AND TREES
Γ(hypo, 1212)
}|
z
1212
{
2
2
1215
3
2
3
1314
1
4
1323
2151
1221
{
1231
2
2141
1313
Γ(hypo, 1221)
}|
z
2
2131
1214
4
2121
{
2
1213
3
Γ(hypo, 2121)
}|
z
9
3
3131
2
3
3141
2
1241
1
4
3231
1251
1331
2
3
1341
1
1332
Figure 3. The isomorphic components Γ(hypo, 1212),
Γ(hypo, 2121), and Γ(hypo, 1221) of the quasi-crystal graph
Γ(hypo).
Proof. Let u = u1 · · · uk and std(u) = s1 · · · sk , where uh , sh ∈ A. Since
std(u1 · · · uk ) = std(s1 · · · sk ), it follows that std(u1 · · · uh ) = std(s1 · · · sh )
for all h (this is immediate from the definition of standardization [Nov00,
Lemma 2.2]). Thus, by [Nov00, Theorems 4.12 and 4.16], Qhypo (u1 · · · uh )
and Qhypo (s1 · · · sh ) both have shape DC(std(u1 · · · uh )−1 ) for all h. The
sequence of these shapes determines where new symbols are inserted during
the computation of Phypo (u) and Qhypo (std(u)) by [CM17, Algorithm 4], and
so Qhypo (u) = Qhypo (std(u)).
Corollary 3.10. In every connected component of Γ(hypo), there is exactly
one standard word, and so for all u, v ∈ A∗ , the words u and v lie in the
same connected component if and only if std(u) = std(v).
Proof. By Lemma 3.9 and Theorem 3.8, u and std(u) are in the same component of Γ(hypo), so each connected component contains at least one standard
word. Suppose u′ is a standard word in the same component of Γ(hypo) as
u. Then it is possible to apply a sequence of quasi-Kashiwara operators to
u and obtain u′ . By Proposition 3.6, quasi-Kashiwara operators preserve
standardization; thus std(u) = std(u′ ) and so u = u′ since u and u′ are
standard words. The last claim in the statement follows immediately.
Remark 3.11. As noted in [CM17], a related notion of ‘quasi-crystal’ is
found in Krob and Thibon [KT99]; see also [Hiv00]. However, the Krob–
Thibon quasi-crystal describes the restriction to quasi-ribbon tableaux (or
rather corresponding to quasi-ribbon tableaux) of the action of the usual
Kashiwara operators: it does not apply to all words and so does not yield a
definition of ≡hypo via isomorphisms of components.
10
ALAN J. CAIN AND ANTÓNIO MALHEIRO
4. Binary trees and the sylvester and Baxter monoids
This section gathers the necessary definitions and background on binary search trees, the sylvester and Baxter monoids, and the analogues
of the Robinson–Schensted correspondence. For further background on the
sylvester monoid, see [HNT05]; for the Baxter monoid, see [Gir12].
4.1. Binary trees and insertion. A left strict binary search tree is a labelled rooted binary tree where the label of each node is strictly greater
than the label of every node in its left subtree, and less or equal to than the
label of every node in its right subtree. A right strict binary search tree is
a labelled rooted binary tree where the label of each node is greater than
or equal to the label of every node in its left subtree, and strictly less than
the label of every node in its right subtree. The following are examples of,
respectively, left strict and right strict binary search trees:
5
4
4
(4.1)
1
5
4
2
7
1
6
2
1
5
4
1
5
.
6
5
7
5
A binary search tree is standard if the labels of its nodes are exactly the
integers from 1 to the number of nodes in the tree. The shape of a binary
search tree T , denoted Sh(T ), is simply its underlying unlabelled rooted
binary tree. The notion of shape will be important later in the paper.
A decreasing tree is a labelled rooted binary tree where the label of each
node is greater than the label of its children. An increasing tree is a labelled
rooted binary tree where the label of each node is less than the label of its
children. The following are examples of, respectively, standard increasing
and standard decreasing trees:
1
10
2
(4.2)
4
3
10
7
6
9
9
5
7
4
8
2
3
1
.
6
5
8
The left-to-right infix traversal (or simply the infix traversal) of a rooted
binary tree T is the sequence that ‘visits’ every node in the tree as follows:
it recursively performs the infix traversal of the left subtree of the root of T ,
then visits the root of T , then recursively performs the infix traversal of the
right subtree of the root of T . Thus the infix traversal of any binary tree
with the same shape as the right-hand tree in (4.1) visits nodes as follows:
(4.3)
.
The left-to-right postfix traversal, or simply the postfix traversal, of a rooted
binary tree T is the sequence that ‘visits’ every node in the tree as follows:
CRYSTALS AND TREES
11
it recursively perform the postfix traversal of the left subtree of the root
of T , then recursively perform the postfix traversal of the right subtree of
the root of T , then visits the root of T . Thus the postfix traversal of any
binary tree with the same shape as the right-hand tree in (4.1) visits nodes
as follows:
(4.4)
.
The infix reading In(T ) (respectively, postfix reading Post(T )) of a labelled
binary search tree T is defined to be the word obtained by listing the labels
of the nodes visited during the infix (respectively, postfix) traversal. For
example, the infix and postfix readings of the right-hand tree in (4.1) are,
respectively, 1124455567 and 1142557654. Note that the weights of the
infix and postfix readings of a given binary search tree are the same; both
weights are simply the tuple that describes the number of each symbol in
A appearing in the tree. Thus define the weight of a labelled binary search
tree to be the weight of these readings.
The following result is immediate from the definition of a binary search
tree, but it is used frequently:
Proposition 4.1. For any (left or right strict) binary search tree T , the
infix reading In(T ) is always the unique weakly increasing word with the
same weight as T .
Let T be a rooted binary tree and let x1 , . . . , xm be the nodes of T in the
order visited in an infix traversal. Let i1 , . . . , ik−1 be such that the xih are
precisely the nodes of T that have non-empty left subtrees. The left interval
partition of T is the partition
o
n
{x1 , . . . , xi1 −1 }, {xi1 , . . . , xi2 −1 }, . . . , {xik−2 , . . . , xik−1 −1 }, {xk−1 , . . . , xm } ;
{z
} |
{z
}
{z
}
|
{z
} |
|
1st left interval
2nd left interval
k − 1-th left interval
k-th left interval
the h-th part of this partition is the h-th left interval of T . Let j1 , . . . , jℓ−1
be such that the xjh are precisely the nodes of T that have non-empty right
subtrees. The right interval partition of T is the partition
o
n
{x1 , . . . , xj1 } , {xj1 +1 , . . . , xj2 }, . . . , {xjℓ−2 +1 , . . . , xjℓ−1 }, {xℓ−1 , . . . , xm } ;
{z
}
|
{z
} |
{z
}
{z
} |
|
1st right interval
2nd right interval
ℓ − 1-th right interval
ℓ-th right interval
the h-th part of this partition is the h-th right interval of T . Consider the
left and right intervals of a binary tree with the same shape as the right-hand
tree in (4.1). Then the nodes lying in the left intervals and right intervals
of this tree are, respectively, those linked by solid lines in the left-hand and
right-hand diagrams below:
4th
3rd
(4.5)
6th
2nd
1st
4th
5th
1st
2nd
3rd
5th
12
ALAN J. CAIN AND ANTÓNIO MALHEIRO
In a sense, the lengths of the left and right intervals of a binary seach tree
are analogous to the lengths of the rows in a quasi-ribbon tableau. However,
the lengths of left and right intervals do not determine the shape of the tree,
for the trees
and
both have left interval length 2, 2 and right interval lengths 1, 2, 1:
;
.
The canopy of a binary tree T , denoted cnp(T ) is the word over {0, 1} obtained by doing an infix traversal of the nodes of T and outputting a 1 when
an empty left subtree is encountered and 0 when an empty right subtree is
encountered, then omitting the 1 from the start and the 0 from the end of
the resulting word. (These symbols correspond to the empty left subtree
of the leftmost node and the empty right subtree of the rightmost node.)
For example, the canopies of the two binary trees in (4.1) are, respectively,
110101100 and 001010011:
5
4
(4.6)
1
1
1
4
1
4
6
2
0 1
2
7
0
1
1
5
5
0
1
1
0
0 1
5
4
0
1
5
5
0
0 1
.
6
1
7
0
Let TL and TR be binary trees. Then (TL , TR ) is a pair of twin binary
trees if cnp(TL ) and cnp(TR ) are complementary, in the sense that for all i,
the i-th symbols of cnp(TL ) and cnp(TR ) are unequal.
Now let TL be a left strict binary search tree and let TR be a right strict
binary search tree. Then (TL , TR ) is a pair of twin binary search trees if
In(TL ) = In(TR ) and cnp(TL ) and cnp(TR ) are complementary (that is, the
underlying binary trees form a pair of twin binary trees). Note that the
binary search trees in (4.1) form a pair of twin binary search trees, since
they have the same infix reading (that is, 1124455567), and complementary
canopies, as shown in (4.6). The shape of a pair of twin binary search trees
(TL , TR ), denoted Sh(TL , TR ), is simply the underlying pair of unlabelled
rooted binary trees.
The insertion algorithms for right (respectively, left) strict binary search
trees adds the new symbol as a leaf node in the unique place that maintains
the property of being a right (respectively, left) strict binary search tree.
Algorithm 4.2 (Left strict leaf insertion).
Input: A left strict binary search tree T and a symbol a ∈ A.
Output: A left strict binary search tree T ← a.
Method: If T is empty, create a node and label it a. If T is non-empty,
examine the label x of the root node; if a ≥ x, recursively insert a into the
right subtree of the root node; otherwise recursively insert a into the left
subtree of the root note. Output the resulting tree.
CRYSTALS AND TREES
13
Algorithm 4.3 (Right strict leaf insertion).
Input: A right strict binary search tree T and a symbol a ∈ A.
Output: A right strict binary search tree a → T .
Method: If T is empty, create a node and label it a. If T is non-empty,
examine the label x of the root node; if a ≤ x, recursively insert a into the
left subtree of the root node; otherwise recursively insert a into the right
subtree of the root note. Output the resulting tree.
Using the leaf insertion algorithms described above, we can compute from
a word in A∗ a left (respectively, right) binary search tree.
Algorithm 4.4 (Left strict insertion).
Input: A word a1 · · · ak , where ai ∈ A.
Output: A left strict binary search tree TL (a1 · · · ak ) and a standard increasing tree TRecL (a1 · · · ak ).
Method: Start with the empty binary search tree T0 and an empty increasing tree D0 . For each i = 1, . . . , k, insert ai into Ti−1 as per Algorithm 4.2;
let Ti be the resulting binary search tree. Build a increasing tree Di that
has the same shape as Ti by adding a node labelled by i to the tree Di−1 in
the same place as ai was inserted into Ti−1 .
Output Tk for TL (a1 · · · ak ) and Dk for TRecL (a1 · · · ak ).
Algorithm 4.5 (Right strict insertion).
Input: A word a1 · · · ak , where ai ∈ A.
Output: A right strict binary search tree TR (a1 · · · ak ) and a standard
decreasing tree TRecR (a1 · · · ak ).
Method: Start with the empty binary search tree T0 and an empty decreasing tree D0 . For each i = 1, . . . , k, insert ak−i+1 into Ti−1 as per
Algorithm 4.3; let Ti be the resulting binary search tree. Build a decreasing
tree Di that has the same shape as Ti by adding a node labelled by k − i + 1
to the tree Di−1 in the same place as ak−i+1 was inserted into Ti−1 .
Output Tk for TR (a1 · · · ak ) and Dk for TRecR (a1 · · · ak ).
For example, if u = 5451761524, then TL (u) and TR (u) are respectively
the left- and right-hand trees in (4.1), and TRecL (u) and TRecR (u) are
respectively the left- and right-hand trees in (4.2). Note that Algorithm 4.4
proceeds through the word from left to right, while Algorithm 4.5 proceeds
through the word from right to left.
Let u ∈ A∗ be a word that contains no repeated symbols. The decreasing
tree of u, denoted DecT(u), is defined as follows: if m is the maximum
symbol that appears in u, so that u = sms′ , then DecT(u) has root labelled
by m, and the left subtree of its root is DecT(s) and the right subtree of the
root is DecT(s′ ). That is,
DecT(u) = DecT sms′ =
m
.
DecT(s)
DecT(s′ )
Similarly, the increasing tree of u, denoted IncT(u), is defined as follows:
if n is the minimum symbol that appears in u, so that u = tnt′ , then IncT(u)
14
ALAN J. CAIN AND ANTÓNIO MALHEIRO
has root labelled by n, and the left subtree of its root is IncT(t) and the
right subtree of the root is IncT(t′ ). That is,
IncT(u) = IncT tnt′ =
n
IncT(t′ )
IncT(t)
Thus if u is 47921013865 (where the underline denotes a single symbol
in A that is not written using a single decimal digit), then IncT(u) and
DecT(u) are respectively the left- and right- hand trees in (4.2).
Remark 4.6. The infix reading of DecT(u) is u and the infix reading of
IncT(u) is u.
Proposition 4.7 ([Gir12, Proposition 4.8]). For any word u ∈ A∗ ,
(1) TRecL (u) = IncT(std(u)−1 );
(2) TRecR (u) = DecT(std(u)−1 );
and TRecL (u), TRecR (u) is a pair of twin binary trees.
(Note that since std(u) is a standard word, it can be viewed as a permutation in one-line notation, and thus its inverse is defined.)
Before proceeding to define the sylvester and Baxter monoids, it is first
necessary to prove two technical lemmata:
Lemma 4.8. Let α, α′ , β, β ′ ∈ A∗ and k ∈ A be such that |α| = |α′ |,
|β| = |β ′ |, the symbol k is greater than every symbol in α, α′ , β, β ′ , and αkβ
and α′ kβ ′ are standard words.
(1) If the decreasing trees DecT(αβ) and DecT(α′ β ′ ) have the same
shape as each other, then the decreasing trees DecT(αkβ) and DecT(α′ kβ ′ )
are of the same shape as each other.
(2) If the increasing trees IncT(αβ) and IncT(α′ β ′ ) have the same shape
as each other, then the increasing trees IncT(αkβ) and IncT(α′ kβ ′ )
are of the same shape as each other.
Proof.
(1) Suppose that the |α|-th symbol in αkβ is c. Now consider
two cases separately:
(a) c is a left child node in DecT(αβ). The infix reading of DecT(αβ)
is αβ, so the left subtree of c must have infix reading α0 , where
α = α0 c, and the right subtree of c must have infix reading β0 ,
where β0 is a prefix of β. Suppose β = β0 β1 ; then β1 is the
infix reading of the part of the tree above and to the right of
c. That is, DecT(αβ) must be as shown in the left-hand side of
(4.7) below.
DecT(αβ)
z
}|
{
β1
(4.7)
z
DecT(αkβ)
}|
{
c
α0
k
c
β0
α0
β1
β0
CRYSTALS AND TREES
15
The tree on the right-hand side of (4.7) has infix reading αkβ.
(Note that in the right-hand tree, the subtree β0 attaches to
the subtree β1 in the same place as c attaches to the subtree
β1 in the left-hand tree.) It is a decreasing tree since k is the
maximum symbol and the relative order of c, α0 , β0 , and β1 is
correct since DecT(αβ) is a decreasing tree. Hence the righthand tree in (4.7) is DecT(αkβ).
Since DecT(α′ β ′ ) has the same shape as DecT(αβ) and |α| =
|α′ |, the |α|-th symbol of α′ β ′ is also a left child and so DecT(α′ β ′ )
has the same shape as the left-hand tree in (4.7). Since k is also
the maximum symbol in α′ kβ ′ , the decreasing tree DecT(α′ kβ ′ )
has the same shape as the right-hand tree in (4.7). Thus DecT(αkβ)
and DecT(α′ kβ ′ ) have the same shape.
(b) c is a right child node or the root node. The infix reading of
DecT(αβ) is αβ, so the left subtree of c must have infix reading
α1 , where α1 c is a suffix of α, and the right subtree of c must
have infix reading β. Suppose α = α0 α1 c; then α0 is the infix
reading of the part of the tree above and to the left of c. That
is, DecT(αβ) must be as shown in the left-hand side of (4.8)
below.
DecT(αβ)
z
}|
{
α0
(4.8)
DecT(αkβ)
}|
{
k
α0
c
α1
z
β
c
β
α1
The tree on the right-hand side of (4.8) has infix reading αkβ. It
is a decreasing tree since k is the maximum symbol and the relative order of c, α0 , α1 , and β is correct since DecT(αβ) is a decreasing tree. Hence the right-hand tree in (4.8) is DecT(αkβ).
Since DecT(α′ β ′ ) has the same shape as DecT(αβ) and |α| =
|α′ |, the |α|-th symbol of α′ β ′ is also a right child or is also
the root node and so DecT(α′ β ′ ) has the same shape as the
left-hand tree in (4.8). Since k is also the maximum symbol in
α′ kβ ′ , the decreasing tree DecT(α′ kβ ′ ) has the same shape as
the right-hand tree in (4.8). Thus DecT(αkβ) and DecT(α′ kβ ′ )
have the same shape.
(2) Since k is greater than every symbol in αβ, the increasing tree
IncT(αkβ) is obtained from IncT(αβ) by adding k as a leaf node.
With the same notation and similar reasoning as in case (1), one
gets the following equivalents of (4.7) (when c is a left child node)
16
ALAN J. CAIN AND ANTÓNIO MALHEIRO
and (4.8) (when c is a right child node or the root node):
z
IncT(αβ)
}|
{
z
β1
IncT(αkβ)
}|
{
c
α0
β1
c
α0
β0
β0
k
z
IncT(αβ)
}|
{
α0
z
IncT(αkβ)
}|
{
α0
c
α1
c
β
α1
β
k
In each case, the right-hand tree is an increasing tree with infix read
αkβ, by the fact that k is the maximum symbol and the relative
order of c, α0 , α1 , and β is correct since IncT(αβ) is an increasing
tree. Hence in both cases the right-hand tree is IncT(αkβ). Thus,
by similar reasoning to case (1), IncT(αkβ) and IncT(α′ kβ ′ ) have
the same shape.
The second technical lemma is simply the dual of the first:
Lemma 4.9. Let α, α′ , β, β ′ ∈ A∗ be such that |α| = |α′ |, |β| = |β ′ |, and
α1β and α′ 1β ′ are standard words.
(1) If the decreasing trees DecT(αβ) and DecT(α′ β ′ ) have the same
shape as each other, then the decreasing trees DecT(α1β) and DecT(α′ 1β ′ )
are of the same shape as each other.
(2) If the increasing trees IncT(αβ) and IncT(α′ β ′ ) have the same shape
as each other, then the increasing trees IncT(α1β) and IncT(α′ 1β ′ )
are of the same shape as each other.
Equipped with these notions of binary search trees and insertion algorithms, it is now possible to use them to define the sylvester and Baxter
monoids (approach 2 in the introduction).
4.2. Sylvester monoid. Define the sylvester P-symbol of u ∈ A∗ by Psylv (u) =
TR (u), and the sylvester Q-symbol of u ∈ A∗ by Qsylv (u) = TRecR (u). The
definition of the relation ≡sylv using right strict binary search trees and insertion is:
u ≡sylv v ⇐⇒ Psylv (u) = Psylv (v).
Using this definition, it can be shown that ≡sylv is a congruence on A∗ ,
which is known as the sylvester congruence. The factor monoid A∗ /≡sylv is
the sylvester monoid and is denoted sylv. The congruence ≡sylv naturally
CRYSTALS AND TREES
17
restricts to a congruence on A∗n , and the factor monoid A∗n / ≡sylv is the
sylvester monoid of rank n and is denoted sylvn .
The monoid sylv is presented by hA | Rsylv i, where
Rsylv = (cavb, acvb) : a ≤ b < c, v ∈ A∗ ;
the monoid sylvn is presented by hAn | Rsylv i, where the set of defining relations Rsylv is naturally restricted to A∗n × A∗n . Notice that sylv and sylvn are
multihomogeneous. (See [CGM17] for a general discussion of homogeneous
and multihomogeneous monoids.)
It is straightforward to see that the map u 7→ Psylv (u), Qsylv (u) is a
bijection between words in A∗ and pairs consisting of a right strict binary
search tree and a decreasing tree of the same shape; this is the sylvester analogue of the Robinson–Schensted correspondence. For instance, if Psylv (u)
and Qsylv (u) are the right-hand trees in (4.1) and (4.2) respectively, then
u = 5451761524.
For any word u ∈ A∗ , the set of words in the ≡sylv -class of u is the
sylvester class of u. Extend this terminology to right strict binary search
trees: for such a tree T , the set of words w ∈ A∗ such that Psylv (w) = T
is the sylvester class of T . The left-to-right postfix reading Post(T ) is an
element of the sylvester class of T : that is, Psylv (Post(T )) = T . (Note that
the infix reading In(T ) is not in general an element of the sylvester class
of T .) The following lemma follows immediately from the definition of the
left-to-right postfix reading:
Lemma 4.10. Let T be an unlabelled binary tree. Under the sylvester analogue of the Robinson–Schensted correspondence, the set of postfix readings
of right strict binary search trees of shape T is obtained by fixing a particular
sylvester Q-symbol of shape T and varying the sylvester P-symbol of over all
right strict binary search trees of shape T .
The following corollary of Proposition 4.7 does not seem to have been
stated before:
Corollary 4.11. For u, v ∈ A∗ ,
u ≡sylv v ⇐⇒ wt(u) = wt(v)
∧ Sh DecT(std(u)−1 ) = Sh DecT(std(v)−1 ) .
Proof. Suppose that u ≡sylv v. Then TR (u) = TR (v). Thus wt(u) =
wt(TR (u)) = wt(TR (v)) = wt(v) and, by Proposition 4.7,
Sh DecT(std(u)−1 ) = Sh(Qsylv (u)) = Sh(Psylv (u))
= Sh(Psylv (v)) = Sh(Qsylv (v)) = Sh DecT(std(v)−1 ) .
On the other hand, suppose wt(u) = wt(v) and Sh DecT(std(u)−1 ) =
Sh DecT(std(v)−1 ) . Then
Sh(Psylv (v)) = Sh(Qsylv (v)) = Sh DecT(std(v)−1 )
= Sh DecT(std(u)−1 ) = Sh(Qsylv (u)) = Sh(Psylv (u)).
Thus Psylv (u) and Psylv (v) have the same shape. Since wt(u) = wt(v), the
trees Psylv (u) and Psylv (v) have the same content. By Proposition 4.1, the
18
ALAN J. CAIN AND ANTÓNIO MALHEIRO
infix reading of a binary search tree is the unique weakly increasing word
with the same content, and so it follows that the infix readings of Psylv (u)
and Psylv (v) are equal. Since Psylv (u) and Psylv (v) have the same shape, they
are thus identical. Hence u ≡sylv v.
4.3. Baxter monoid. The definition of the relation ≡baxt uses pairs of twin
binary search trees. For any word u ∈ A∗ , the binary search trees TL (u)
and TR (u) form a pair of twin binary search
trees. Define the Baxter Q
symbol of u by Pbaxt (u) = TL (u), TR (u) , and the Baxter Q-symbol of u by
Qbaxt (u) = TRecL (u), TRecR (u) . Define ≡baxt by:
u ≡baxt v ⇐⇒ Pbaxt (u) = Pbaxt (v).
Using this definition, it follows that ≡baxt is a congruence on A∗ , which is
known as the Baxter congruence. The factor monoid A∗ /≡baxt is the Baxter
monoid and is denoted baxt. The congruence ≡baxt naturally restricts to a
congruence on A∗n , and the factor monoid A∗n / ≡baxt is the Baxter monoid
of rank n and is denoted baxtn .
The monoid baxt is presented by hA | Rbaxt i, where
Rbaxt = (cudavb, cuadvb) : a ≤ b < c ≤ d, u, v ∈ A∗
∪ { (budavc, buadvc) : a < b ≤ c < d, u, v ∈ A∗ };
the monoid baxtn is presented by hAn | Rbaxt i, where the set of defining
relations Rbaxt is naturally restricted to A∗n × A∗n . Note that baxt and baxtn
are multihomogeneous.
The map
u 7→ Pbaxt (u), Qbaxt (u) = TL (u), TR (u) , TRecL (u), TRecR (u)
is a bijection between words in A∗ and pairs consisting of a pair of twin
binary search tree and a pair made up of an increasing and a decreasing
tree of the same shape as the corresponding binary search trees; this is the
Baxter analogue of the Robinson–Schensted correspondence. For instance,
if Pbaxt (u) is the pair of twin binary search trees in (4.1), and Qbaxt (u) is the
pair of increasing and decreasing trees in (4.2), then u = 5451761524.
For example, if u = 5451761524, then std(u) = 64711092835, so (std(u))−1 =
47921013865. Hence TRecL (u) and IncT((std(l))−1 ) are both the left-hand
tree in (4.2), and TRecR (u) and DecT((std(u))−1 ) are both the right-hand
tree in (4.2).
For any word u ∈ A∗ , the set of words in the ≡baxt -class of u is the Baxter
class of u. Extend this terminology to pairs of twin binary search trees: for
such a pair (TL , TR ), the set of words w ∈ A∗ such that Pbaxt (w) = T is the
Baxter class of T .
There is a straightforward method for computing words in the Baxter
class of a pair of twin binary search trees (TL , TR ):
Method 4.12. Input: A pair of twin binary search trees (TL , TR ).
Output: A word in the Baxter class of (TL , TR ).
(1) Set (UL , UR ) to be (TL , TR ). (Throughout this computation, UL is a
forest of left strict binary search trees and UR is a right strict binary
search tree.)
CRYSTALS AND TREES
19
(2) If UL and UR are empty, halt.
(3) Given some (UL , UR ), choose and output some symbol a that labels
a root of some tree in the forest UL and a leaf of the tree UR .
(4) Delete the corresponding root vertex of UL and the corresponding
leaf vertex of UL .
(5) Go to step 2.
This is essentially [Gir12, Algorithm on p.133], except that the method
given here is non-deterministic in that there may be several choices for a in
step 3. As these choices vary, all words in the Baxter class of (TL , TR ) are
obtained. If, in step 3, one always chooses the leftmost possible a, call the
resulting word the left-consistent reading of (TL , TR ). (By [Gir12, Proposition 4.16], the left-consistent reading of (TL , TR ) is the lexicographically
smallest word in the corresponding Baxter class.) The following lemma
follows immediately from the definition of the left-consistent reading:
Lemma 4.13. Let (TL , TR ) be a pair of twin binary trees. Under the Baxter
analogue of the Robinson–Schensted correspondence, the set of left-consistent
readings of pairs of twin binary search trees of shape (TL , TR ) is obtained by
fixing a particular Baxter Q-symbol of shape (TL , TR ) and varying the Baxter
P-symbol over all pairs of twin binary search trees of shape (TL , TR ).
The following result is the analogy of Corollary 4.11 for the Baxter monoid:
Corollary 4.14. For u, v ∈ A∗ ,
u ≡baxt v ⇐⇒ wt(u) = wt(v)
∧ Sh DecT(std(u)−1 ) = Sh DecT(std(v)−1 )
∧ Sh IncT(std(u)−1 ) = Sh IncT(std(v)−1 ) .
Proof. Follow the reasoning in the proof of Corollary 4.11, but working additionally with TL (·), TRecL (·), and IncT(std(·)−1 ).
5. Quasi-crystals with abstract shapes
An abstract shape is a map σ : A∗ → S satisfying the following axioms:
S1 The map σ is invariant under standardization, in the sense that for
all u ∈ A∗ σ(u) = σ(std(u)).
S2 For all u, v ∈ A∗ and a ∈ A, if wt(u) = wt(v) and σ(u) = σ(v), then
σ(ua) = σ(va) and σ(au) = σ(av) for all w ∈ A∗ .
Explicit examples of shapes will be given in the next section. For the moment, note that by Proposition 3.6 and Corollary 3.10, axiom S1 is equivalent to saying that σ is invariant under the quasi-Kashiwara operators,
in the sense that for all u ∈ A∗ and all i ∈ N, if ëi (u) is defined, then
σ(u) = σ(ëi (u)), and if f¨i (u) is defined, σ(u) = σ(f¨i (u)). Thus, if axiom S1
is satisfied, every element of a given connected component of Γ(hypo) will
have the same image under σ.
For u, v ∈ A∗ , a quasi-crystal isomorphism θ : Γ(hypo, u) → Γ(hypo, v) is
shape-preserving (with respect to an abstract shape σ) if σ(w) = σ(θw) for
all w ∈ Γ(hypo, u).
20
ALAN J. CAIN AND ANTÓNIO MALHEIRO
Define a relation ∼σ on the free monoid A∗ as follows: u ∼σ v if and only
if there is a shape-preserving quasi-crystal isomorphism θ : Γ(hypo, u) →
Γ(hypo, v) such that θ(u) = v. That is, u ∼σ v if and only if u and v
have the same image under σ and are in the same position in isomorphic
connected components of Γ(hypo). Note that ∼σ ⊆ ∼.
Proposition 5.1. The relation ∼σ is a congruence on the free monoid A∗ .
Proof. It is clear from the definition that ∼σ is an equivalence relation; it
thus remains to prove that ∼σ is compatible with multiplication in A∗ .
Suppose u ∼σ v and w ∈ A∗ . Suppose w = w1 · · · wk (where wi ∈ A).
The aim is to proceed by induction on i and show there is a shape-preserving
quasi-crystal isomorphism θi : Γ(hypo, uw1 · · · wi ) → Γ(hypo, vw1 · · · wi ) such
that θ(uw1 · · · wi ) = vw1 · · · wi for all i ∈ {0, 1, . . . , k} (formally taking
w1 · · · wi to be ε when i = 0).
For i = 0, since u ∼σ v, there is by definition a shape-preserving quasicrystal isomorphisms θ0 : Γ(hypo, u) → Γ(hypo, v) such that θ(u) = v. This
is the basis of the induction.
So assume there is a shape-preserving quasi-crystal isomorphism θi :
Γ(hypo, uw1 · · · wi ) → Γ(hypo, vw1 · · · wi ) such that θi (uw1 · · · wi ) = vw1 · · · wi .
In particular, uw1 · · · wi ∼ vw1 · · · wi . Since ∼ is a congruence, uw1 · · · wi wi+1 ∼
vw1 · · · wi wi+1 , and so there is a quasi-crystal isomorphism (which is not yet
known to be shape-preserving) θi+1 : Γ(hypo, uw1 · · · wi+1 ) → Γ(hypo, vw1 · · · wi+1 )
such that θi+1 (uw1 · · · wi+1 ) = vw1 · · · wi+1 . Furthermore, since θi is shapepreserving, σ(uw1 · · · wi ) = σ(θ(uw1 · · · wi )) = σ(vw1 · · · wi ). Combine this
with uw1 · · · wi ∼ vw1 · · · wi and use axiom S2 to see that σ(uw1 · · · wi+1 ) =
σ(vw1 · · · wi+1 ). Thus, by axiom S1, every element in the connected components Γ(hypo, uw1 · · · wi+1 ) and Γ(hypo, vw1 · · · wi+1 ) have the same shape.
Hence θi+1 is shape-preserving.
Hence, by induction, θk : Γ(hypo, uw) → Γ(hypo, vw) is a shape-preserving
quasi-crystal isomorphism; therefore uw ∼σ vw.
Similar reasoning shows that wu ∼σ wv; hence ∼σ is compatible with
multiplication and is thus a congruence.
Thus every abstract shape σ gives rise to a monoid A∗ /∼σ .
6. Crystallizing the sylvester monoid
6.1. Definition by crystals. Let S be the set of unlabelled binary trees
and define σsylv (u) = Sh(DecT(std(u)−1 )) for all u ∈ A∗ .
Proposition 6.1. σsylv is an abstract shape.
Proof. By Proposition 3.6, the map σsylv is invariant under the quasi-Kashiwara
operators, and so satisfies axiom S1.
To see that σsylv satisfies S2, proceed as follows. Let u, v ∈ A∗ be such
that u ∼ v and σsylv (u) = σsylv (v), and let a ∈ A.
The first aim is to prove that σsylv (ua) = σsylv (va). When one computes
std(ua) and std(va), the symbol a is replaced by the same symbol b in both
words, since u and v have the same content. Let k = |u| + 1. In std(ua)
(viewed as a permutation), k is mapped to b. Thus, k is the b-th symbol, and
the maximum symbol, in both std(ua)−1 and std(va)−1 . That is, std(u)−1 =
CRYSTALS AND TREES
21
αβ, std(ua)−1 = αkβ, std(v)−1 = α′ β ′ and std(va)−1 = α′ kβ ′ for some
α, α′ , β, β ′ ∈ A∗ , such that |α| = |α′ |, |β| = |β ′ |, the symbol k is greater than
every symbol in α, α′ , β, β ′ , and αkβ and α′ kβ ′ are standard words. Furthermore, DecT(std(u)−1 ) = DecT(αβ) and DecT(std(v)−1 ) = DecT(α′ β ′ )
have the same shape, since σsylv (u) = σsylv (v). Thus, by Lemma 4.8(1),
DecT(std(ua)−1 ) = DecT(αkβ) and DecT(std(va)−1 ) = DecT(α′ kβ ′ ) have
the same shape. Hence σsylv (ua) = σsylv (va).
The next aim is to prove that σsylv (au) = σsylv (av). When one computes
std(au) and std(av), the symbol a is replaced by the same symbol b in
both words, since u and v have the same content. In std(au) (viewed as
a permutation), 1 is mapped to b. Thus, 1 is the b-th symbol, and the
minimum symbol, in both std(au)−1 and std(av)−1 . By similar reasoning
to the above, but using Lemma 4.9(1), one sees that DecT(std(au)−1 ) and
DecT(std(av)−1 ) have the same shape. Hence σsylv (au) = σsylv (av).
As a consequence of Proposition 6.1, one can define a relation ∼sylv on
the free monoid A∗ to be the relation ∼σ from Section 5 with the abstract
shape σ = σsylv . By Propositions 5.1 and 6.1, ∼sylv is a congruence.
When viewing the quasi-crystal graph Γ(hypo) in the context of σsylv preserving isomorphisms, denote it by Γ(sylv). Similarly, denote the connected component Γ(hypo, u) by Γ(sylv, u). (This is arguably an abuse of
notation, since Γ(hypo) and Γ(sylv) are the same graph, and the difference
is in the notion of isomorphism and the relation to which it gives rise.)
Proposition 6.2. The relations ∼sylv and ≡sylv coincide.
Proof. Let u, v ∈ A∗ . Suppose that u ∼sylv v. Then u ∼ v and hence
wt(u) = wt(v). Furthermore, σsylv (u) = σsylv (v), and so
Sh DecT(std(u)−1 ) = σsylv (u) = σsylv (v) = Sh DecT(std(v)−1 ) .
Hence, by Corollary 4.11, u ≡sylv v.
Now suppose that u ≡sylv v. Then u ≡hypo v and so u ∼ v and so there is
a quasi-crystal isomorphism θ : Γ(hypo, u) → Γ(hypo, v) with θ(u) = v. By
Corollary 4.11,
σsylv (u) = Sh DecT(std(u)−1 ) = Sh DecT(std(v)−1 ) = σsylv (v).
Thus, by S2, σsylv (θ(w)) = σsylv (w) for all w ∈ Γ(hypo, u). Hence θ is a
σsylv -preserving quasi-crystal isomorphism and so u ∼sylv v.
Our aim here has been to show that ≡sylv can be defined using quasicrystals; thus we have avoided using the prior knowledge that ≡sylv is a
congruence [HNT05]. If we had used that result, it would have been straightforward to recover the fact that σsylv is an abstract shape.
6.2. Characterizing highest-weight elements.
Proposition 6.3. Let T be an unlabelled binary tree. The set of words in
A∗ that are postfix readings of right strict binary search trees of shape T
form a single connected component of Γ(sylv).
Proof. Suppose u and v are postfix readings of two right strict binary search
trees of shape T . Then std(u) and std(v) are also postfix readings of two
22
z
ALAN J. CAIN AND ANTÓNIO MALHEIRO
1
1
2
2
Γ(sylv, 1221)
}|
{
z
2121
Γ(sylv, 2121)
}|
{
z
2
1
1
2
Γ(sylv, 1212)
}|
{
2
1
1
2
1
1
4
1
2
3
1
1
5
1
2
2
4
3
|
{z
1
3
3
4
1
3
}
2151
|
3141
4
2
1
2
3
1
2
1
3
1
3
{z
3
1
3
3231
}
1
1
2
4
3131
2
2
Legend
1
2141
3
4
1
1
2
3
1
3
2131
3
5
4
1
1
2
1
|
2
3
{z
1
4
3
1
4
3
1
2
5
Figure 4. Part of the connected components Γ(sylv, 1221)
and Γ(sylv, 2121), which are linked by a shape-preserving
isomorphism, and part of the connected component
Γ(sylv, 1212), which is only isomorphic to the others as a
labelled digraph.
binary search trees of shape T [HNT05, Lemma 11]. Since there is exactly
one right strict binary search of a given shape labelled by a standard word
[HNT05, Note 3], it follows that std(u) = std(v) and so u and v are in the
same connected component of Γ(sylv) by Corollary 3.10.
Lemma 6.4. Let u ∈ A∗ and i ∈ N. Then ëi (u) is defined if and only if
TR (u) contains at least one node i + 1, but does not contain a node i + 1 in
the right subtree of the topmost node i.
Proof. By definition, ëi (u) is defined if and only if every symbol i + 1 is to
the right of every symbol i in u. In terms of Algorithm 4.5, this is equivalent
to every symbol i + 1 being inserted before every symbol i, which in turn
is equivalent to no symbol i + 1 being inserted into the right subtree of the
topmost symbol i.
Proposition 6.5. Let u ∈ A∗ . Then u is a highest-weight word if and only
if TR (u) has the following property: if the right interval partition of TR (u)
has ℓ parts, then for a ∈ {1, . . . , ℓ} ⊆ A, the nodes in the a-th right interval
are labelled by a.
3
}
CRYSTALS AND TREES
23
Before proceeding to the proof, an example is helpful. The word u =
1121335432 is highest weight, and TR (u) is
2
1
1
1
3
2
3
3
,
4
5
which satisfies the condition in Proposition 6.5. On the other hand, the word
u = 5451761524 does not have highest weight, and TR (u) is the right-hand
tree in (4.1), which does not satisfy the condition.
Proof. Suppose u is a highest-weight word. Then ëi (u) is undefined for all
i. Let ℓ be the number of right intervals of TR (u). Since the infix reading of
TR (u) is a weakly increasing word that must increase immediately after the
rightmost nodes in the first ℓ − 1 right intervals, there are at least ℓ distinct
symbols among the labels in TR (u).
If i > 1 is a symbol in u, then i − 1 is also a symbol in u, for otherwise
ëi−1 (u) would be defined. Consequently, u contains the symbols 1, . . . , m
for some m ≥ ℓ. If m > 1, then, by Lemma 6.4, there is a symbol 2 in the
right subtree of the topmost symbol 1. Thus the symbols 1 must label all
nodes in the first right interval of TR (u). Proceeding inductively and using
Lemma 6.4, one sees that a must label all nodes in the a-th right interval.
Hence TR (u) has the form described in the statement.
Now suppose that TR (u) has the form described. It is immediate that
a + 1 is in the right subtree of the topmost node a for all a ∈ {1, . . . , ℓ},
and so ëi (u) is undefined for i ≤ ℓ. Since ℓ + 1 is the maximum symbol in
TR (u), it follows that ëi (u) is undefined for all i by Lemma 6.4, and so u
has highest weight.
Note that it would be possible to prove Proposition 6.5 by appealing to the
corresponding result for the hypoplactic monoid and quasi-ribbon tableaux
[CM17, Proposition 6(2)] to characterize highest weight words, but it seems
clearer to give a direct proof that reasons only about trees.
6.3. The sylvester version of the Robinson–Schensted correspondence. The following result clarifies the interaction of the quasi-crystal
structure and the sylvester version of the Robinson–Schensted correspondence:
Proposition 6.6. Let u, v ∈ A∗ . The words u and v lie in the same connected component of Γ(sylv) if and only if Qsylv (u) = Qsylv (v).
Proof. Suppose Qsylv (u) = Qsylv (v). Hence TRecR (u) = TRecR (v) and by
Proposition 4.7, DecT(std(u)−1 ) = DecT(std(u)−1 ). Thus, by Proposition 4.1,
std(u)−1 and std(v)−1 are the infix readings of the same labelled binary tree,
and so std(u) = std(v). Since u and std(u) lie in the same connected component of Γ(sylv) by Corollary 3.10, and similarly for v and std(v), it follows
that u and v are in the same connected component of Γ(sylv).
On the other hand, if u and v lie in the same connected component of
Γ(sylv), then std(u) = std(v) by Corollary 3.10 and so Qsylv (u) = Qsylv (v).
24
ALAN J. CAIN AND ANTÓNIO MALHEIRO
Thus the sylvester Q-symbol indexes connected components of the quasicrystal graph and the sylvester P-symbol locates a word within that component. Note that combining Proposition 6.6 and Lemma 4.10, one recovers
Proposition 6.3.
The Knuth ‘hook-length’ formula
Q shows that the number of descending
trees of the same shape as T is n!/ v∈T hv , where hv is the number of nodes
in the subtree rooted at v [Knu98, § 5.1.4, Exer. 20]. This gives a formula
for the number of sylvester Q-symbols of a given shape, and thus for the
size of sylvester classes corresponding to left strict binary search trees of a
given shape, and thus for the number of quasi-crystal components that are
mapped to a given one by a σsylv -preserving quasi-crystal isomorphism.
In the infinite-rank setting, there are infinitely many sylvester P-symbols
of a given shape. On restricting to the rank n case, the number of sylvester
P-symbols of a given shape can be calculated using the same formula as for
the hypoplactic monoid [CM17, Theorem 4]:
Proposition 6.7. Let T be a binary search tree for which the right interval
partition has ℓ parts. The number of right strict binary search trees of shape
T labelled by symbols from An is
(
n+|T |−ℓ
if ℓ ≤ n
n−ℓ
0
if ℓ > n.
Proof. Consider a highest-weight word u ∈ A∗n such that TR (u) has shape
T . By Proposition 6.5, u contains the symbols {1, . . . , ℓ}. Let α be the
shape of the quasi-ribbon tableau Phypo (u). Since u is also a highest-weight
word in Γ(sylv) and thus in Γ(hypo), it follows that the column reading v of
Phypo (u) is also a highest weight word in Γ(hypo) and so α = wt(v) = wt(u)
[CM17, Proposition 6]. In particular, ℓ(α) = ℓ and |α| = |v| = |u| = |T |.
Since Γ(sylv, u) = Γ(hypo, u) is isomorphic to Γ(hypo, v), they both contain the same number of words that lie in A∗n , and so the result follows by
subsituting the given values for ℓ(α) and |α| into [CM17, Theorem 4].
6.4. Counting factorizations. One interpretation of the Littlewood–Richardson
rule [Ful97, ch. 5] is that the Littlewood–Richardson coefficients cνλµ are the
number of different factorizations of an element of plac corresponding to a
Young tableau of shape ν into elements corresponding to a tableau of shape
λ and a tableau of shape µ. In particular, it shows that the number of such
factorizations is independent of the content of the tableau, and is rather
dependent only on the shape. The quasi-crystal structure yields a similar
result for hypo [CM17, Theorem 5], and a minimal modification to the proof
gives the similar result for the sylvester monoid:
Theorem 6.8. The number of distinct factorizations of an element of the
sylvester monoid corresponding to a right strict binary search tree of shape
T into elements that correspond to right strict binary search trees of shapes
U and V is dependent only of T , U , and V , and not on the content of the
element.
Proof. For the purposes of this proof, a word u ∈ A∗ is a postfix word if it
is the postfix reading of a binary search tree; a postfix word u has shape U ,
CRYSTALS AND TREES
25
where U is an unlabelled binary tree, if it is the postfix reading of a binary
search tree of shape U .
Let T , U , and V be unlabelled binary trees. Let w ∈ A∗ be a postfix
word of shape T . Let
w
SU,V
= (u, v) : u, v ∈ A∗n are postfix words,
u has shape U ,
v has shape V ,
w ≡sylv uv .
w
SU,V
So
is a complete list of factorizations of w into elements whose corresponding right strict binary search trees have shapes U and V . Let i ∈ N
w . Then w ≡
and suppose ëi (w) is defined. Pick some pair (u, v) ∈ SU,V
sylv uv
and so ëi (uv) is defined, and ëi (w) = ëi (uv). By [CM17, Lemma 2], either
ëi (w)
,
ëi (uv) = uëi (v), or ëi (uv) = ëi (u)v. In the former case, (u, ëi (v)) ∈ SU,V
ë (w)
i
, since ëi preserves being a postand in the latter case (ëi (u), v) ∈ SU,V
fix word and the shape of the corresponding right strict binary search tree
ëi (w)
w
by Proposition 6.3. So ëi induces an injective map from SU,V
to SU,V
.
ëi (w)
w
¨
to S . Hence
It follows that fi induces the inverse map from S
U,V
U,V
w | = |S f¨i (w) |. Since
Similarly, if f¨i (w) is defined, then |SU,V
=
U,V
all the postfix words whose corresponding binary search trees have shape T
lie in the same connected component of Γ(sylv) by Proposition 6.3, it follows
w | is dependent only on T , not on w.
that |SU,V
w |
|SU,V
ëi (w)
|.
|SU,V
6.5. Satisfying an identity. Another application of the quasi-crystal structure was to prove that the hypoplactic monoid satisfies a non-trivial identity
[CM17, Theorem 6]. Similarly, it is possible to use the quasi-crystal structure to prove that the sylvester monoid satisfies an identity.
Theorem 6.9. The sylvester monoid satisfies the identity xyxy = yxxy.
Proof. Let x, y ∈ A∗ . The proof that xyxy = yxxy proceeds by reverse
induction on the weight of xyxy.
The base case of the induction is when xyxy is highest-weight. Thus
ëi (xyxy) is undefined for all i ∈ N. So xyxy must contain a subsequence
(i+1)i for all i ∈ {1, . . . , ℓ−1}, where ℓ is the maximum symbol appearing in
x or y. The symbols i + 1 and i may each lie in x or y, but in any case, there
is a subsequence (i + 1)i in yxxy. Hence ëi (yxxy) is undefined for all i ∈ N
and so yxyx is also highest-weight. Clearly wt(xyxy) = wt(yxxy). Consider
applying Algorithm 4.5 to xyxy and yxxy. In both cases, the algorithm
initially computes TR (xy). At this point, every symbol in B = {1, . . . , ℓ}
appears in the tree and so the infix reading of TR (xy) contains all these
symbols. Thus any symbol b ∈ B that is subsequently inserted will be
placed either as the left child of a node b that is already present, or into the
right subtree of the topmost node b − 1 that is already present. Thus the
trees TR (xyxy) and TR (yxxy) are identical.
The remaining reasoning parallels the proof of [CM17, Theorem 6]. For
the induction step, suppose xyxy is not highest-weight, and that x′ y ′ x′ y ′ ≡sylv
y ′ x′ x′ y ′ for all x′ , y ′ ∈ A∗ such that x′ y ′ x′ y ′ has higher weight than xyxy.
26
ALAN J. CAIN AND ANTÓNIO MALHEIRO
Then ëi (xyxy) is defined for some i ∈ N. Neither x nor y contains a symbol
i, since otherwise there would be a subsequence (i + 1)i. So ǫ̈i (xyxy) =
ǫ̈i (yxxy) = |xyxy|i+1 = |yxxy|i+1 = 2|x|i+1 + 2|y|i+1 = 2ǫ̈i (x) + 2ǫ̈i (y),
and ǫ̈i (xyxy) applications of ëi change every symbol i + 1 in xyxy to i.
ǫ̈ (xyxy)
ǫ̈ (yxxy)
That is, ëi i
(xyxy) and ëi i
(yxxy) are both defined and are equal
ǫ̈ (x)
ǫ̈ (y)
′
′
′
′
′
′
′
′
to x y x y and y x x y respectively, where x′ = ëi i (x) and y ′ = ëi i (y).
Since x′ y ′ x′ y ′ has higher weight than xyxy, it follows by the induction hypothesis that x′ y ′ x′ y ′ ≡sylv y ′ x′ x′ y ′ . Hence
ǫ̈ (xyxy) ′ ′ ′ ′
ǫ̈ (xyxy) ′ ′ ′ ′
xyxy = f¨i i
(x y x y ) ≡sylv f¨i i
(y x x y ) = yxxy.
This completes the induction step and thus the proof.
For an alternative proof of Theorem 6.9 that does not require the quasicrystal structure, see [CM, Section 3.3].
7. Crystallizing the Baxter monoid
7.1. Definition by crystals. Let S be the set of pairs of twin binary trees
and define σbaxt (u) = Sh(IncT(std(u)−1 )), Sh(DecT(std(u)−1 )) for all u ∈
A∗ .
Proposition 7.1. σbaxt is an abstract shape.
Proof. By Proposition 3.6, the map σbaxt is invariant under the quasi-Kashiwara
operators, and so satisfies axiom S1.
To see that σbaxt satisfies S2, proceed as follows. Let u, v ∈ A∗ be such
that wt(u) = wt(v) and σbaxt (u) = σbaxt (v), and let a ∈ A.
Then Sh(IncT(std(u)−1 )) = Sh(IncT(std(v)−1 )) and Sh(DecT(std(u)−1 )) =
Sh(DecT(std(v)−1 )). By the reasoning in the proof of Proposition 6.1, Sh(DecT(std(ua)−1 )) =
Sh(DecT(std(va)−1 )) and Sh(DecT(std(au)−1 )) = Sh(DecT(std(av)−1 )). Symmetrical reasoning, but using part (2) of Lemmata 4.8 and 4.9 shows that
Sh(IncT(std(ua)−1 )) = Sh(IncT(std(va)−1 )) and Sh(IncT(std(au)−1 )) =
Sh(IncT(std(av)−1 )). Hence σbaxt (ua) = σbaxt (va) and σbaxt (au) = σbaxt (av).
Hence σbaxt satisfies S2.
As a consequence of Proposition 7.1, one can define a relation ∼baxt on
the free monoid A∗ to be the relation ∼σ from Section 5 with the abstract
shape σ = σbaxt . By Propositions 5.1 and 7.1, ∼baxt is a congruence.
When viewing the quasi-crystal graph Γ(hypo) in the context of σbaxt preserving isomorphisms, denote it by Γ(baxt). Similarly, denote the connected component Γ(hypo, u) by Γ(baxt, u). (This is arguably an abuse of
notation, since Γ(hypo) and Γ(baxt) are the same graph, and the difference
is in the notion of isomorphism and the relation to which it gives rise.)
Proposition 7.2. The relations ∼baxt and ≡baxt coincide.
Proof. Let u, v ∈ A∗ . Suppose that u ∼baxt v. Then u ∼ v and hence
wt(u) = wt(v). Furthermore, σbaxt (u) = σbaxt (v), and so
Sh(IncT(std(u)−1 )), Sh(DecT(std(u)−1 )) = σbaxt (u)
= σbaxt (v) = Sh(IncT(std(v)−1 )), Sh(DecT(std(v)−1 ))
CRYSTALS AND TREES
27
Hence, by Corollary 4.14, u ≡baxt v.
Now suppose that u ≡baxt v. Then u ≡hypo v and so u ∼ v and so there is
a quasi-crystal isomorphism θ : Γ(hypo, u) → Γ(hypo, v). By Corollary 4.14,
σbaxt (u) = Sh(IncT(std(u)−1 )), Sh(DecT(std(u)−1 ))
= Sh(IncT(std(v)−1 )), Sh(DecT(std(v)−1 )) = σbaxt (v).
Thus, by S2, σbaxt (θ(w)) = σbaxt (w) for all w ∈ Γ(hypo, u). Hence θ is a
σbaxt -preserving quasi-crystal isomorphism and so u ∼baxt v.
As was the case for ≡sylv , our aim here has been to show that ≡baxt
can be defined using quasi-crystals; thus we have avoided using the prior
knowledge that ≡baxt is a congruence [Gir12], from which, it would have
been straightforward to recover the fact that σbaxt is an abstract shape.
7.2. Characterizing highest weight elements.
Proposition 7.3. Let (TL , TR ) be a pair of twin binary trees. The set of
words in A∗ that are left-consistent readings of pairs of twin binary search
trees of shape (TL , TR ) form a single connected component of Γ(baxt).
Proof. Suppose u and v are left-consistent readings of two pairs of twin
binary seach trees of shape (TL , TR ). Then std(u) and std(v) are also leftconsistent readings of shape (TL , TR ) (by [HNT05, Lemma 11] and its dual).
Since there is exactly one pair of twin binary search trees of a given shape
labelled by a standard word (by [HNT05, Note 3] and its dual), it follows
that std(u) = std(v) and so u and v are in the same connected component
of Γ(baxt) by Corollary 3.10.
The following lemma is the dual of Lemma 6.4, and can be proved by a
dual argument.
Lemma 7.4. Let u ∈ A∗ and i ∈ N. Then ëi (u) is defined if and only if
TL (u) contains at least one node i + 1, but does not contain a node i in the
left subtree of the topmost node i + 1.
The following result follows from the argument in the proof of Proposition 6.5
plus a dual argument:
Proposition 7.5. Let u ∈ A∗ . Then u is a highest-weight word if and only
if Pbaxt (u) = (TL (u), TR (u)) has the following properties:
• if the left interval partition of TL (u) has k parts, then for a ∈
{1, . . . , k} ⊆ A, the nodes in the a-th left interval are labelled by
a;
• if the right interval partition of TR (u) has ℓ parts, then for a ∈
{1, . . . , ℓ} ⊆ A, the nodes in the a-th right interval are labelled by a.
7.3. The Baxter version of the Robinson–Schensted correspondence.
The following result is the parallel of Proposition 6.6 for the Baxter monoid:
Proposition 7.6. Let u, v ∈ A∗ . The words u and v lie in the same connected component of Γ(baxt) if and only if Qbaxt (u) = Qbaxt (v).
28
ALAN J. CAIN AND ANTÓNIO MALHEIRO
Like the case of the sylvester monoid, the Baxter Q-symbol indexes connected components of the quasi-crystal graph and the Baxter P-symbol locates a word within that component.
Unlike the sylvester monoid, there is no known neat formula for the size
of classes of words that represent the same element of the Baxter monoid.
However, the quasi-crystal structure does make it clear that the size of
these classes is dependent only on the shape of the corresponding pair of
twin binary search trees.
Remark 7.7. One might consider trying to use the quasi-crystal structure
to obtain a formula for the size of these classes as follows. By the quasicrystal structure, it suffices to consider the size of the class for a standard
element; that is, for a pair of twin binary search trees (TL , TR ) containing
each symbol in K = {1, . . . , k} exactly once. Each of these binary search
trees can be considered as a partial order on the set K; denote these partial
orders by ≤L and ≤R . Let ≤′L be the dual of ≤L . Now, any reading of
(TL , TR ) can be viewed as a linear order on K that extends ≤′L and ≤R . Since
K is finite, there are finitely many such linear orders, so we can intersect
them to obtain a partial order ⊏. So the readings of (TL , TR ) are in oneto-one correspondence with the linear orders extending ⊏. However, the
problem of the number of linear orders extending a given partial order is
#P -complete [BW91].
Of course, not all partial orders on K will arise as ⊏, so it is possible that
a more refined version of this argument, using the structure of the pair of
twin binary search trees to gain information about ⊏, might still work.
7.4. Counting factorizations. The following result is the parallel of Theorem 6.8
for the Baxter monoid:
Theorem 7.8. The number of distinct factorizations of an element of the
Baxter monoid corresponding to a pair of twin binary search trees of shape
(TL , TR ) into elements that correspond to pairs of twin binary search trees of
shapes (UL , UR ) and (VL , VR ) is dependent only of (TL , TR ), (UL , UR ), and
(VL , VR ), and not on the content of the element.
Proof. For the purposes of this proof, a word u ∈ A∗ is a left-consistent
word if it is the left-consistent reading of a pair of twin binary search tree; a
left-consistent word u has shape (UL , UR ), where (UL , UR ) is a pair of twin
binary trees, if it is the left-consistent reading of a pair of twin binary search
trees of shape (UL , UR ).
Let (TL , TR ), (UL , UR ), and (VL , VR ) be pairs of twin binary trees. Let
w ∈ A∗ be a postfix word such that T is the shape of TR (w). Let
w
S(U
= (u, v) : u, v ∈ A∗n are left-consistent words,
L ,UR ),(VL ,VR )
u has shape (UL , UR ),
v has shape (VL , VR ),
w ≡baxt uv .
Now following the reasoning in the proof of Theorem 6.8, but using Propositions 7.6
w
and 7.3 shows that |S(U
| is dependent only on (TL , TR ), not on
L ,UR ),(VL ,VR )
w.
CRYSTALS AND TREES
29
7.5. Satisfying an identity. The analogy of Theorem 6.9 for the Baxter
monoid is the following:
Theorem 7.9. The Baxter monoid satisfies the identity xyxyxy = xyyxxy.
Proof. Let x, y ∈ A∗ . The proof that xyxyxy = xyyxxy proceeds by reverse
induction on the weight of xyxyxy.
The base case of the induction is when xyxyxy is highest-weight. By reasoning parallel to the proof of Theorem 6.9, xyyxxy is also highest-weight.
Clearly wt(xyxyxy) = wt(xyyxxy). Again by reasoning parallel to the proof
of Theorem 6.9, the trees TR (xyxyxy) and TR (xyyxxy) are equal. By dual
reasoning, the trees TR (xyxyxy) and TR (xyyxxy) are equal.
The induction step is essentially identical to the proof of Theorem 6.9,
except that ǫ̈i (xyxyxy) = 3ǫ̈i (x) + 3ǫ̈i (y) = ǫ̈i (xyyxxy) applications of ëi
change every symbol i + 1 to i. The result follows.
For an alternative proof of Theorem 7.9 that does not require the quasicrystal structure, see [CM, Section 3.5].
8. Perspectives
8.1. The Robinson–Schensted type correspondences. The graphs Γ(hypo),
Γ(sylv) and Γ(baxt) are all the same object; the difference is in the notion of
‘isomorphic’ components. In terms of the hypoplactic, sylvester, and Baxter versions of the Robinson–Schensted correspondence, the Q-symbol of a
word identifies a connected component of the graph, and the P-symbol of
a word identifies its position within that connected component. Viewed in
this context, the sylvester and Baxter Q-symbols are overdetermined. The
‘underlying’ Q-symbol of a word u ∈ A∗ is std(u), which lies in the same
component as u. This corresponds with the notions of ‘compatibility with
the destandardization process’ and std-goodness developed by [Pri13].
8.2. Bases of combinatorial Hopf algebras. In the infinite-rank setting,
there is a very natural connection between the crystal graph Γ(plac) and free
Schur functions. The free Schur functions are indexed by standard Young
tableau and form a basis for the algebra of free symmetric functions FSym.
The crystal graph offers an equivalent way of defining them:
X
ST =
w,
w∈Γ(plac,T )
where Γ(plac, T ) denotes the connected component of Γ(plac) corresponding
to a standard Young tableau T .
Similarly, the free quasi-ribbon functions, which form a basis for FQSym
can be defined with reference to the quasi-crystal graph:
X
Fu =
w,
w∈Γ(hypo,u)
where u ∈ A∗ is a standard word; this follows from the fact that standard
words index connected components of Γ(hypo).
30
ALAN J. CAIN AND ANTÓNIO MALHEIRO
For the basis of the algebra of planar binary trees PBT given by [HNT05,
Eqs (16–17)] (see also [Hiv07, Theorem 27]), the definition using the quasicrystal graph becomes:
X
PT =
w,
w∈Γ(sylv,U )
Sh(U )=T
where T is a binary tree, U ranges over the set of decreasing trees and
Γ(sylv, U ) denotes the connected component of Γ(sylv) indexed by the decreasing tree U .
For the basis of the algebra Baxter given in [Gir12, Theorem 6.12], the
definition using the quasi-crystal graph is:
X
w,
P′ (TL ,TR ) =
w∈Γ(baxt,(UL ,UR ))
Sh(UL ,UR )=(TL ,TR )
where (TL , TR ) is a pair of twin binary trees, (UL , UR ) ranges over the set
of pairs made up of increasing and a decreasing tree with the same infix
reading and Γ(baxt, (UL , UR )) denotes the connected component of Γ(baxt)
indexed by the pair (UL , UR ).
8.3. Speculations. In the quasi-crystal structure for the hypoplactic monoid,
one constructs the graph using the quasi-Kashiwara operators and one defines the congruence using labelled directed graph isomorphisms of connected
components. If one replaces the quasi-Kashiwara operators with the original
Kashiwara operators, one obtains the plactic monoid. On the other hand, if
one requires that labelled directed graph isomorphisms be shape-preserving,
one obtains the sylvester and Baxter monoids. It is thus natural to ask
what monoids arise when one makes both replacements and considers shapepreserving isomorphisms on the graph arising from the original Kashiwara
operators (see Figure 5). (If one uses the abstract notion of shape defined
in Section 5 and the original Kashiware operators, axiom S1 should probably be replaced with ‘invariance under the Kashiwara operators’, which is
a natural analogue in light of the discussion following axioms S1 and S2.
Alternatively, one could seek a natural analogue of standardization for the
original Kashiwara operators.)
References
[BW91]
G. Brightwell & P. Winkler. ‘Counting linear extensions’. Order, 8, no. 3 (1991),
pp. 225–242. doi: 10.1007/BF00383444.
[CGM]
A. J. Cain, R. D. Gray, & A. Malheiro. ‘Crystal monoids & crystal bases:
rewriting systems and biautomatic structures for plactic monoids of types An ,
Bn , Cn , Dn , and G2 ’. arXiv: 1412.7040.
[CGM17] A. J. Cain, R. D. Gray, & A. Malheiro. ‘On finite complete rewriting systems,
finite derivation type, and automaticity for homogeneous monoids’. Inform. and
Comput., 255, no. 1 (2017), pp. 68–93. doi: 10.1016/j.ic.2017.05.003.
[CM]
A. J. Cain & A. Malheiro. ‘Identities in plactic, hypoplactic, sylvester, Baxter,
and related monoids’. arXiv: 1611.04151.
[CM17] A. J. Cain & A. Malheiro. ‘Crystallizing the hypoplactic monoid: from quasiKashiwara operators to the Robinson–Schensted–Knuth-type correspondence
for quasi-ribbon tableaux’. J. Algebraic Combin., 45, no. 2 (2017), pp. 475–524.
doi: 10.1007/s10801-016-0714-6.
Q
ra
wa
hi
as rs
K to
si- era
ua op
?
Is
om
or
ph
ism
s
31
ra
wa s
hi tor
as a
K p er
o
Sh
a
iso pem pr
o r es
ph er
ism vin
s g
CRYSTALS AND TREES
sylv
baxt
plac
hypo
Figure 5. (Quasi-)-Kashiwara operators, notions of isomorphisms, and monoids.
[Ful97]
W. Fulton. Young Tableaux: With Applications to Representation Theory and
Geometry. No. 35 in LMS Student Texts. Cambridge University Press, 1997.
[Gir12]
S. Giraudo. ‘Algebraic and combinatorial structures on pairs of twin binary
trees’. J. Algebra, 360 (2012), pp. 115–157. doi: 10.1016/j.jalgebra.2012.03.020.
[Gre06] J. Green. Polynomial Representations of GLn. No. 830 in Lecture Notes in Mathematics. Springer, 2nd edition, 2006. doi: 10.1007/3-540-46944-3.
[Hiv00] F. Hivert. ‘Hecke algebras, difference operators, and quasi-symmetric functions’. Advances in Mathematics, 155, no. 2 (2000), pp. 181–238. doi:
10.1006/aima.1999.1901.
[Hiv07] F. Hivert. ‘An Introduction to Combinatorial Hopf Algebras: Examples and
realizations’. In J.-P. Gazeau, J. Nešetřil, & B. Rovan, eds, Physics and Theoretical Computer Science: From Numbers and Languages to (Quantum) Cryptography, no. 7 in NATO Security through Science Series D: Information and
Communication Security. IOS Press, 2007.
[HNT05] F. Hivert, J.-C. Novelli, & J.-Y. Thibon. ‘The algebra of binary search
trees’. Theoret. Comput. Sci., 339, no. 1 (2005), pp. 129–165. doi:
10.1016/j.tcs.2005.01.012.
[Jed11] F. Jedrzejewski. ‘Plactic classification of modes’. In C. Agon, M. Andreatta,
G. Assayag, E. Amiot, J. Bresson, & J. Mandereau, eds, Mathematics and
Computation in Music, no. 6726 in Lecture Notes in Comput. Sci., pp. 350–353.
Springer, 2011. doi: 10.1007/978-3-642-21590-2 31.
[Kas90] M. Kashiwara. ‘Crystalizing the q-analogue of universal enveloping algebras’. Comm. Math. Phys., 133, no. 2 (1990), pp. 249–260. url:
http://projecteuclid.org/euclid.cmp/1104201397.
[Kas91] M. Kashiwara. ‘On crystal bases of the q-analogue of universal enveloping
algebras’. Duke Mathematical Journal, 63, no. 2 (1991), pp. 465–516. doi:
10.1215/S0012-7094-91-06321-0.
[KN94] M. Kashiwara & T. Nakashima. ‘Crystal graphs for representations of the qanalogue of classical Lie algebra’. J. Algebra, 165, no. 2 (1994), pp. 295–345.
doi: 10.1006/jabr.1994.1114.
[Knu98] D. E. Knuth. The Art of Computer Programming: Sorting and Searching, vol. 3.
Addison-Wesley, 2nd edition, 1998.
32
[KT97]
[KT99]
[Lot02]
[LS78]
[LS81]
[LS85]
[LS90]
[Mac08]
[Nov00]
[Pri13]
[Sch97]
[Shi05]
[vL01]
ALAN J. CAIN AND ANTÓNIO MALHEIRO
D. Krob & J.-Y. Thibon. ‘Noncommutative Symmetric Functions IV: Quantum
Linear Groups and Hecke Algebras at q = 0’. Journal of Algebraic Combinatorics, 6, no. 4 (1997), pp. 339–376. doi: 10.1023/A:1008673127310.
D. Krob & J.-Y. Thibon. ‘Noncommutative Symmetric Functions V: A Degenerate Version of uq (gln )’. Int. J. Algebra Comput., 9, no. 3–4 (1999), pp. 405–430.
doi: 10.1142/S0218196799000254.
M. Lothaire. Algebraic Combinatorics on Words. No. 90 in Encyclopedia of
Mathematics and its Applications. Cambridge University Press, 2002.
A. Lascoux & M.-P. Schützenberger. ‘Sur une conjecture de H. O. Foulkes’. C.
R. Acad. Sci. Paris Sér. A-B, 286, no. 7 (1978), pp. A323–A324.
A. Lascoux & M.-P. Schützenberger. ‘Le monoı̈de plaxique’. In Noncommutative structures in algebra and geometric combinatorics, no. 109 in
Quaderni de ”La Ricerca Scientifica”, pp. 129–156, Rome, 1981. CNR. url:
http://igm.univ-mlv.fr/~ berstel/Mps/Travaux/A/1981-1PlaxiqueNaples.pdf.
A. Lascoux & M.-P. Schützenberger. ‘Schubert polynomials and the LittlewoodRichardson rule’. Lett. Math. Phys., 10, no. 2-3 (1985), pp. 111–124. doi:
10.1007/BF00398147.
A. Lascoux & M.-P. Schützenberger. ‘Tableaux and noncommutative Schubert polynomials’. Funct. Anal. Its. Appl., 23, no. 3 (1990), pp. 223–225. doi:
10.1007/BF01079531.
I. Macdonald. Symmetric Functions and Hall Polynomials. Clarendon Press,
Oxford University Press, 2008.
J.-C. Novelli. ‘On the hypoplactic monoid’. Discrete Mathematics, 217, no. 1–3
(2000), pp. 315–336. doi: 10.1016/S0012-365X(99)00270-8.
J.-B. Priez. ‘A lattice of combinatorial Hopf algebras: Binary trees with multiplicities’. In Formal Power Series and Algebraic Combinatorics, Nancy, 2013.
The Association. Discrete Mathematics & Theoretical Computer Science. url:
http://www.dmtcs.org/pdfpapers/dmAS0196.pdf.
M.-P. Schützenberger. ‘Pour le monoı̈de plaxique’. Mathématiques et sciences
humaines, 140 (1997). url: http://msh.revues.org/2764.
M.
Shimozono.
‘Crystals
for
dummies’,
2005.
Notes.
url:
http://www.aimath.org/WWN/kostka/crysdumb.pdf.
M. A. van Leeuwen. ‘The Littlewood-Richardson Rule, and Related Combinatorics’. In J. R. Stembridge, J.-Y. Thibon, & M. A. van Leeuwen, eds, Interaction of Combinatorics and Representation Theory, pp. 95–145, Tokyo, 2001.
Mathematical Society of Japan. doi: 10.2969/msjmemoirs/01101C030.
Centro de Matemática e Aplicações, Faculdade de Ciências e Tecnologia,
Universidade Nova de Lisboa, 2829–516 Caparica, Portugal
E-mail address: [email protected]
Centro de Matemática e Aplicações and Departamento de Matemática, Faculdade de Ciências e Tecnologia, Universidade Nova de Lisboa, 2829–516 Caparica, Portugal
E-mail address: [email protected]
| 4math.GR
|
Robust Detection in Leak-Prone Population
Protocols
Dan Alistarh15? , Bartłomiej Dudek2 , Adrian Kosowski3?? , David
Soloveichik4? ? ? , and Przemysław Uznański1
arXiv:1706.09937v1 [cs.DS] 29 Jun 2017
1
ETH Zürich, Switzerland
University of Wrocław, Poland
Inria Paris and IRIF, Université Paris Diderot, France
4
University of Texas, Austin, TX, USA
5
IST Austria
2
3
{dan.alistarh,przemyslaw.uznanski}@inf.ethz.ch,[email protected],
[email protected], [email protected].
Abstract. In contrast to electronic computation, chemical computation
is noisy and susceptible to a variety of sources of error, which has prevented
the construction of robust complex systems. To be effective, chemical
algorithms must be designed with an appropriate error model in mind.
Here we consider the model of chemical reaction networks that preserve
molecular count (population protocols), and ask whether computation can
be made robust to a natural model of unintended “leak” reactions. Our
definition of leak is motivated by both the particular spurious behavior
seen when implementing chemical reaction networks with DNA strand
displacement cascades, as well as the unavoidable side reactions in any
implementation due to the basic laws of chemistry. We develop a new
“Robust Detection” algorithm for the problem of fast (logarithmic time)
single molecule detection, and prove that it is robust to this general model
of leaks. Besides potential applications in single molecule detection, the
error-correction ideas developed here might enable a new class of robustby-design chemical algorithms. Our analysis is based on a non-standard
hybrid argument, combining ideas from discrete analysis of population
protocols with classic Markov chain techniques.
1
Introduction
A major challenge in designing autonomous molecular systems is to achieve a
sufficient degree of error tolerance despite the error-prone nature of the chemical
substrate. While considerable effort has focused on making the chemistry itself
more robust, here we look at the possibility of developing chemical algorithms that
?
??
???
Supported by an SNF Ambizione Fellowship.
Supported by Inria project GANG, ANR project DESCARTES, and NCN grant
2015/17/B/ST6/01897.
Supported by NSF grants CCF-1618895 and CCF-1652824.
are inherently resilient to the types of error encountered. Before designing robust
chemical algorithms, we must decide on a good error model that is relevant
to the systems we care about. In this paper we focus on a very simple and
general error model that is motivated both by basic laws of chemistry as well as
by implementation artifacts in strand displacement constructions for chemical
reaction networks. We begin by listing the types of errors we aim to capture.
Leaks due to Law of Catalysis. A fundamental law of chemical kinetics is
that for every catalyzed reaction, there is an uncatalyzed one that occurs at a
(often much) slower rate. By a catalytic reaction, we mean a reaction that involves
some species X but does not change its amount; this species is called a catalyst
of that reaction. For example, the reaction X + Y → X + Z is catalytic, and
species X is the catalyst since its count remains unchanged by the firing of this
reaction. By the law of catalysis, reaction X + Y → X + Z must be accompanied
by a (slower) leak reaction Y → Z. (A more general formulation of the law of
catalysis is that if any sequence of reactions does not change the net count of
X, then there is a pathway that has the same effect on all the other species,
but can occur in the absence of X (possibly much slower). Thus for example, if
X + Y → W and W → X + Z are two reactions, then there must also be a leak
reaction Y → Z. Formally defining catalytic cycles and catalysts is non-trivial
and is beyond the scope of this paper [1].)
Leaks due to Law of Reversibility. Another fundamental law of chemical
kinetics is that any reaction occurs also in the reverse direction at some (possibly
much slower) rate. In other words, reaction X + Y → Z + W must be accompanied
by Z + W → X + Y . (The degree of reaction reversibility is related to the freeenergy use, such that irreversible reactions would require “infinite” free energy.)
Leaks due to Spurious Activation in Strand-Displacement Cascades.
Arbitrary chemical reaction networks can in principle be implemented with DNA
strand displacement cascades [2,3]. Implementations based on strand displacement
also suffer from the problem of leaks [4]. The implementation of a reaction like
X + Y → Z + W consists of “fuel” complexes present in excess, that hold Z and
W sequestered. A cascaded reaction of the fuel complex with X and Y results in
the release of active Z and W . Leaks in this case consist of Z and W becoming
spuriously activated, even in the absence of active X or Y .
Importantly, for a catalytic reaction such as X + Y → X + Z, it is possible
to design a strand displacement implementation that does not leak the catalyst
X. This implementation would release the same exact molecule of X as was
consumed to initiate the process, as in the catalytic system described in [5].
Since fuels do not hold X sequestered, X cannot be produced in the leak process
(although Z can).
Modeling Reactions and Leaks. Note that in all cases above, we can guarantee
that a species does not leak if it is exclusively a catalyst in every reaction it
occurs in. This allows us some handle on the leak. In particular, we will ensure
that the species we are trying to detect (called D below) will be a catalyst in
every reaction that involves it. Otherwise, there might be a leak pathway to
generating D itself—which is fundamentally irrecoverable.
We express the implementations below in the population protocol formalism [6].
That is, we consider a system with n molecules (aka nodes), which interact
uniformly at random in a series of discrete steps. A population protocol is given
as a set of reactions (aka transition rules) of the form
A + B → C + D.
Note that unlike general reaction networks, population protocols conserve total
molecular count since molecules never combine or split. For this reason, compared
to general chemical reaction networks, this model is easier to analyze.
Given the set of reactions defining a protocol, we partition the species into
catalytic states, which never change count as a consequence of any reaction, and
non-catalytic, otherwise. Crucially, we model leaks as spurious reactions which
can consume and create arbitrary non-catalytic species. More formally, a leak is
a reaction of the type
S → S0,
where S and S 0 denote arbitrary non-catalytic species. In the following, we do not
make any assumptions on the way in which these leak transitions are chosen (i.e.,
they could in theory be chosen adversarially), but we assume an upper bound on
the rate at which leaks may occur in the system, controlled by a parameter β.
Leak-Robust Detection. A computationally simple task which already illustrates the difficulty of information processing in such an error-prone system is
single molecule detection. Consider a solution of n molecules, in which a single
molecule D may or may not be present. Intuitively, the goal is to generate largescale (in the order of n) change in the system, depending on whether or not D is
present or absent. Our time complexity measure is parallel time, defined as the
number of pairwise interactions, divided by n. This measure of time naturally
captures the parallelism of the system, where each molecule can participate in
a constant number of interactions per unit time. Subject to leaks, our goal is
to design the chemical interaction rules (formalized as a population protocol)
to satisfy the following behavior. If D is present then it is detected fast, in
logarithmic parallel time, and that the output is probabilistically “stable” in
the sense that sampled at a random future time the system is in the “detected
configuration” with high probability. By contrast, if D is absent, then the system
sampled at a random future time should be in the “undetected configuration”
with high probability. This basic task has several variations, for instance signal
amplification or approximate counting of D.
We first develop some intuition about this problem, by considering some
strawman approaches.
A first trivial attempt would be to have neutral molecules become “detectors”
(state T ) as soon as they encounter D, that is,
D + N → D + T.
This approach suffers from two fatal flaws. First, it is slow, in that detection
takes linear parallel time. Second, it has no way from recovering from leaks of
the type N → T .
A second attempt could try to implement an epidemic-style detection of D,
that is:
D + N →D + T
T + N → T + T.
This approach is fast, i.e. converges in logarithmic parallel time in case D is
present. However, if D is not present, the algorithm converges to a false positive
state: a leak of the type N → T brings the system to an all-T state, despite the
absence of D. One could try to add a “neutralization” pathway by having T turn
back to N after a constant number of interactions, but a careful analysis shows
that this approach also fails to recover from leaks of the type N → T .
Thus, it is not clear whether leak-resistant detection is possible in population
protocols (or more generally chemical reaction networks). There has been considerable work in the algorithmic community on diffusion based models, e.g. [7].
However, such results do not seem to apply to this setting, since leak models have
not been considered previously, and none of the known techniques are robust to
leaks. In particular, it appears that techniques for deterministic computation in
population protocols do not carry over in the presence of leaks. More generally,
this seems to create an unfortunate gap between the algorithmic community,
which designs and analyzes population protocols in leak-free models, and more
practically-minded research, which needs to address such implementation issues.
Contribution. In this paper, we take a step towards bridging this gap. We
provide a general algorithmic model of leaks, and apply it to the detection
problem. Specifically, our immediate goal is to elucidate the question of whether
efficient, leak-robust detection is possible.
We prove that the answer is yes. We present a new algorithm, called RobustDetect, which guarantees the following. Assume that the rate at which leaks
occur is upper bounded by β/n 1/n, and that we return the output mapping
(detect/non-detect) of a randomly chosen molecule after O(log n) parallel time.
Then the probability of a false negative is at most 1/e + o(1), and the probability
of a false positive is at most β. (Note that as the total molecular count n increases,
the chance that a particular interaction involves D decreases linearly with n. Thus
the leak rate must also decrease linearly with n, or else the leaks will dominate.
Alternatively, we can view some fixed leak rate as establishing an upper bound
on the molecular count n, see below. )
Algorithm Description. We now sketch the intuition behind the algorithm and
its properties, leaving the formal treatment to Sections 4 and 5. Fix a parameter
s ≥ 1, to be defined later. We define a set of “detecting” species X1 , . . . , Xs ,
arranged in consecutive levels. Whenever a molecule meets D, it moves to the
highest “alert” level, X1 . Since leaks might produce this species as well, we decay
it gracefully across s levels. More precisely, whenever a molecule at level Xi
meets another molecule at level Xj , both molecules move to state Xmin(i,j)+1 .
A molecule which would move beyond level Xs following a reaction becomes
neutral, i.e. moves to species N . Nodes in state Xi with i < s turn N into Xi+1 ,
whereas molecules in state Xs also become neutral when interacting with N .
Analysis. Intuitively, the algorithm’s dynamics for the case where a single
molecule is in state D are as follows. The counts of molecules in state Xi tend
to increase exponentially with the alert level i, up to levels ≈ log n, when the
count becomes a constant fraction of n. However, once level log n is reached,
these counts decrease doubly exponentially. Thus, it suffices to set s = log n to
obtain that a fraction of at least (1 − 1/e) molecules are in one of the alert states
Xi in case D is present. It is not hard to prove that leaks cannot meaningfully
affect the convergence behavior in this case.
The other interesting case is when D is not present, but leaks may occur,
leading to possible false positives. Intuitively, we can model this case as one where
states X1 at the highest alert level simply are created at a lower rate β/n 1/n.
A careful analysis of this setting yields that the probability of a false positive (D
detected, but not present) in this case is at most β, corresponding to the leak
rate parameter.
Our analysis technique works by characterizing the stationary behavior of the
Markov chain corresponding to the algorithm, and the convergence properties
(mixing time) of this chain. For technical reasons, the analysis uses a non-standard
hybrid argument, combining ideas from discrete analysis of population protocols
with classic Markov chain techniques. The argument proves that the algorithm
always stabilizes to the correct output in logarithmic parallel time.
The analysis further highlights a few interesting properties of the algorithm.
First, if the detectable species D is present in a higher count k > 1, then the
algorithm effectively skips the first log k levels, and thus requires log(n/k) +
O(log log n) states. Second, it is not necessary to know the exact value of log n,
as the counts of species past this threshold decrease doubly exponentially.
Alternative Formulations. An alternative view of this protocol is as solving
the following related amplification problem: we are given a signal of strength
(rate) φ, and the algorithm’s behavior should reflect whether this strength is
below or above some threshold. The detection problem requires us to differentiate
thresholds set at β/n and 1/n, for constant β 1, but our analysis applies to
more general rates.
Above, we have assumed that the leak rate decreases linearly with n, to
separate from the case where a single instance of D is present. However, it is
also reasonable to consider that the leak rate is fixed, say, upper bounded by a
constant λ. In this case, the analysis works as long as the number of molecules n
satisfies λ 1/n.
Self-stabilization. Our algorithm is self-stabilizing in the sense that if the count
of D changes due to some external reason, the output quickly adapts (within
logarithmic parallel time). This is particularly interesting if the algorithm is used
in the context of a “control module” for a cell detecting D and the amount of
D changes over time. Note that strawman solutions considered above cannot be
“untriggered” once D has been detected, and thus cannot adapt to a changing
input.
2
Related Work
There is much work on attempting to decrease error in the underlying chemical
substrate. A famous example includes kinetic proofreading [8]. In the context
of DNA strand displacement systems in particular, leak reduction has been a
prevailing topic [9]. Despite the importance of handling leaks, there are few
examples of non-trivial algorithms, where leaks are handled through computation
embedded in chemistry. One algorithm that appears to be able to handle errors is
approximate majority [10], originally analyzed in a model where a fraction of the
nodes are Byzantine, in that they can change their reported state in an adversarial
way. Potentially due to its robustness properties, the approximate majority
algorithm appears to be widely used in biological regulatory networks [11], and
it was also one of the first chemical reaction network algorithms implemented
with strand displacement cascades [4].
Our algorithm can be viewed as a timed, self-stabilizing version of rumor
spreading. For analysis of simple rumor-spreading, see [12]. Other work include
fault-tolerant rumor spreading [13], push-pull models [7] and self-stabilizing
broadcasting [14]. A rumor-spreading formulation of the molecule detection
problem is also considered in recent work [15], which relies on a different source
amplification mechanism based on oscillator dynamics. This protocol [15] is
self-stabilizing in a weaker (probabilistic) sense compared to the algorithms from
this paper and does not provide leak robustness guarantees.
3
3.1
Preliminaries
Population Protocols with Leaks
Population Protocols. We start from a standard population protocol model,
where n molecules (nodes) interact uniformly at random in a series of discrete
steps. In our formulation, in each step, a coin is flipped to decide whether the
current interaction is a regular reaction or a leak reaction. In the former case,
two molecules are picked uniformly at random, and interact according to the
rules of the protocol. In the latter case, a leak reaction occurs (see below).
A population protocol is given as a set of reactions (transition rules) of the
form
A + B → C + D,
(where some of A, B, C, D might be the same). We (arbitrarily) match the first
reactant (A) with the first product (C), and the second reactant (B) with the
second product (D), and think of A as changing state to C, and B as changing
state to D. If the two molecules picked to interact do not have a corresponding
interaction rule, then they don’t change state and we call this a null interaction.
Population protocols are a special case of the stochastic chemical reaction networks
kinetic model (e.g., [16](A.4)).
Catalytic and Non-Catalytic Species. Given a set of reactions, we define the
set of catalytic species as the set of states which never change as a consequence of
any reaction. That is, for every reaction, the species is present in the same count
both in the input and the output of the reaction. For example, in the reactions
A+C →B+C
A+B →A+D
we call C catalytic. Note that A acts as a catalyst in the second reaction, but its
count is changed by the first reaction, thus it is not overall catalytic. All species
whose count is modified by some reaction are called non-catalytic. Note that it is
possible that a species is never created, but disappears as a consequence of an
interaction. For example, in the reaction
L + L → A + B,
L is such as species. We define such species as non-catalytic, since their creation
is possible by the law of reversibility, and thus they can leak.
An Algorithmic Model of Leaks. A leak is a reaction of the type
S → S0
where S and S 0 are arbitrary non-catalytic species produced by the algorithm.
Note that the input and output species of a leak may be the same (although in
that case the reaction is trivial). In the following, we make no assumptions on
the way in which the input and output of a leak reaction are chosen—we assume
that they are chosen adversarially. Instead, we assume an absolute bound on the
probability of a leak.
We assume that each reaction is either a leak reaction or a normal reaction,
which follows the algorithm. We formalize this as follows.
Definition 1. Given an algorithm, defined by a set of reactions, the set of
catalysts is the set of species whose count does not change as a consequence of any
reaction. A leak is a spurious reaction, which changes an arbitrary non-catalytic
species to an arbitrary non-catalytic species. The leak rate β/n is the probability
that any given interaction is a leak reaction.
3.2
The Detection Problem
In the following, we consider the following detection task: we are given a distinct
species D, whose presence or absence must be detected by the algorithm, in
the presence of leaks. More precisely, if the species D is present, then the
algorithm should stabilize to a state in which molecules map to output value
“detect”. Otherwise, if D is not present, then the algorithm should stabilize to
a state in which molecules map to output value “non-detect”. To observe the
algorithm’s output, we sample a molecule at random, and return its output
mapping. (Alternatively, to boost accuracy, we can take a number of samples, and
return the majority output mapping.) We require that species D are catalytic.
4
The Robust-Detect Algorithm
Description. As given in the problem statement, we assume that there exists
a distinguished species D, which is to be detected, and which never changes
state. Our algorithm implements a chain of detection species X1 , . . . , Xs , for
some parameter s, each of which maps to output “detect”, but with decreasing
“confidence”. Further, we have a neutral species N , which maps to output “nondetect”. We assume that the parameter s = dlog ne, and that initially all molecules
are in state N . We specify the transitions below, and provide the intuition behind
them.
Algorithm 1
D + Xi → D + X1 ,
D + N → D + X1
∀i ∈ {2, . . . , s}
Xs + Xs → N + N
Xs + N → N + N
Xi + Xj → Xmin(i,j)+1 + Xmin(i,j)+1 ,
Xi + N → Xi+1 + Xi+1 ,
∀i, j ∈ {1, 2, . . . , s − 1}
∀i ∈ {1, 2, . . . , s − 1}
The intuition behind the algorithm is as follows. The “detecting” species X1 , . . . , Xs
are arranged in consecutive levels. Whenever a molecule meets D, it moves to
the highest “alert” level, X1 . Since leaks might produce this species as well, we
decay it gracefully across s levels. After going through these levels, a molecule
moves to neutral state N , in case it is not brought back either by meeting D, or
some molecule at a lower alert level. For this, whenever two of these species Xi
and Xj meet, they both move to level min(i, j) + 1. This reaction has the double
purpose of both decaying the alert level of the molecule at the lower level, and
of bringing back the molecule with the higher alert level. Further, whenever a
molecule at level Xi meets a neutral molecule N , it advances its level by 1. At
the same time, neutral molecules are turned into detector molecules whenever
meeting some molecule at an alert level smaller than s.
Intuitive Dynamics. Roughly, the chain of alert levels have the property that,
for the first ∼ log n levels, the count roughly doubles with level index. At the
same time, past this point, counts exhibit a steep (doubly exponential) drop, so
that a small constant fraction of molecules are always neutral. The presence of
D acts like a trigger, which maintains the chain in “active” state. The analysis in
the next section makes this intuition precise. These dynamics are illustrated in
Figure 1.
5
Analysis
Overview. We divide the analysis of the detection algorithm into two parts.
First, we derive stationary probabilities of the underlying Markov chain of
transitions of particles, by solving recursively the equations following from the
0.10
0.05
0
5
10
15
20
25
level i
Level CDF
dlog ne = 14
1.0
0.4
= 0.0
1
0.6
= 0.1
t
0.8
1D
prese
n
0.15
= 0.0
1
t
0.20
= 0.1
dlog ne = 14
0.25
(b)
probability that a random
molecule is one of D, X1 , . . . , Xi
Number of molecules in level
1D
prese
n
probability that
a random molecule is Xi
(a)
0.2
0
5
10
15
20
25
level i
Fig. 1. Steady state probabilities of the Robust-Detect algorithm for n = 104 molecules.
Three conditions are evaluated: (blue) 1 molecule of D is present and no leak (leak
parameter β = 0); (orange, red) no D is present but with worst-case leak for falsepositives (leak reactions Xi → X1 and N → X1 ) (orange: leak parameter β = 0.01, red:
leak parameter β = 0.1). (a) The probabilities of each level i. (b) The cumulative
probabilities of levels ≤ i, capturing the probability that a random molecule is in a
“detect” state. Note that it is enough to set the number of levels s = 14 = dlog ne to
have both false positive and false negative error probabilities small, although for smaller
leak rates (β = 0.01) increasing s beyond log n can help better distinguish true and
false positives. Numerical probabilities are computed using equations (1) and (2).
underlying dynamics. Later, we derive optimal bounds on the mixing time of
this Markov chain—that is we show that probability distribution of states at
every time t ≥ cn log n (for some constant c) is almost the same as the stationary
distribution.
Simplified Algorithm. For the purpose of analysis, let us consider a following
rephrasing of the detection algorithm: molecule states are D, X1 , X2 , X3 , . . ., and
interactions are as follows:
Algorithm 2
D + Xi → D + X1 ,
Xi + Xj → Xmin(i,j)+1 + Xmin(i,j)+1 .
This algorithm uses infinite number of states, thus it is useful only for purposes of
theoretical analysis. However, it captures the behavior of the original algorithm in
the following way: if in Algorithm 2 all states Xs+1 , Xs+2 , . . . are collapsed to N ,
the transitions are equivalent to Algorithm 1. However, formulation of Algorithm 2
is oblivious to parameter s, thus captures simultaneously the dynamics of all
possible instances of Algorithm 1.
5.1
Stationary Analysis
Let us consider an initial state when k ≥ 0 instances of state D are present, with
special attention given to k = 0 and k = 1. Those molecules do not change their
state.
We can imagine tracking a particular molecule through its state transitions,
such that its state can be expressed as a Markov chain. In the following, we will
focus on analyzing the stationary distribution of this Markov chain.
For any i ∈ {1, 2, . . . , s}, we let p?i be the stationary probability that a molecule
chosen uniformly at random is in state Xi . We let p?0 = nk be the (stationary)
probability that the molecule is in the state D. Let p?≤i = p?0 + . . . + p?i be the
probability that a molecule is in any of the states D, X1 , . . . , Xi .
Let us now analyze these stationary probabilities.
Stable State with No Leaks. We first analyze the simplified case where no
leaks occur. A molecule u is in one of states D, X1 , . . . , Xi at time t, in two cases:
– It was in state D, X1 , . . . , Xi at time t − 1, and did not get selected for a
reaction, which occurs with probability 1 − 2/n.
– It got selected for a reaction with element u0 , and either u or u0 was in one
of states D, X1 , . . . , Xi−1 .
Hence, by stationarity, we get that
2
2
?
?
p≤i = p≤i 1 −
+ · (1 − (1 − p?≤(i−1) )2 ).
n
n
From this we get that
1 − p?≤i = (1 − p?≤(i−1) )2 ,
which solves to
2i
k
p?≤i = 1 − 1 −
.
n
(1)
This gives us following estimates: if k ≥ 1, then p?i ≈ 2i−1 nk for i ≤ log(n/k).
j
Additionally, for i = log(n/k) + 1 + j, p?i ≈ e−2 . Thus, for i ≥ log(n/k) +
Θ(log log n) in all practicalities p?i ≈ 0.
To analyze the probability of detection when k = 1, we sum probabilities for
all i from 0 to s = dlog ne
Pr[detect] =
s
X
n
1
1
p?i = p?≤s ≥ 1 − 1 −
≥1− .
n
e
i=0
Probability of False Positives with Leaks. A useful side effect of the previous
analysis is that we also get probability bounds for detection in the case where D
is not present, i.e. false positives. We model this case as follows. Assume that
there exists an upper bound λ on the probability that a certain reaction is a leak.
Examining the structure of the algorithm, we note that the worst-case adversarial
application of leaks would be if this probability is entirely concentrated into leaks
which produce species X1 .
To preserve molecular count, we assume the following simplified leak model,
which is equivalent to the general one, but easier to deal with in the confines of
our algorithm.
Each reaction is a leak with probability λ = β/n, where β 1 is a small
constant. If a reaction is a leak, it selects a molecule at random, and transforms it
into an arbitrary state. In this case, we will assume adversarially that all leaked
molecules are transformed into state X1 . Notice that the assumption that β 1
is required to separate this setting from the case where D is present in the system,
where the probability of producing state X1 is 2/n.
We continue with calculations of p?0 , p?1 , . . . for the above formulation. Note
that the recurrence relation for p?≤i , i ≥ 1 is changed as follows:
– If at that round there was no leak, the transition probabilities are as previously.
This happens with probability 1 − nβ .
– If there was a leak, then the molecule either is selected as a leaked molecule
(this happens with probability n1 · nβ ) or it was not selected as a leaked
?
molecule, but it was already in the proper state (probability nβ · n−1
n p≤i ).
The recursive formulation gives
2
1
n−1 ?
2
β
β
p?≤i = p?≤i 1 −
+
+
+
p≤i
.
1 − (1 − p?≤(i−1) )2
1−
n
n
n
n
n
n
Which is equivalent to
1 − p?≤i
1 − nβ
(1 − p?≤(i−1) )2
=
β
1 − 2n
leading to (using estimate (1 − nβ )/(1 −
p?≤i
β
2n )
≈ (1 −
β
2n ))
1+2+...+2i−1
2i −1
β
β
≈1− 1−
=1− 1−
.
2n
2n
(2)
This gives us following estimates: p?i ≈ 2i−2 nβ for i ≤ log(2n/β). Additionally,
j
for i = log(2n/β) + 1 + j, p?i ≈ e−2 . Thus, for i ≥ log(2n/β) + Θ(log log n) in
all practicalities p?i ≈ 0.
This immediately implies that
2n
s
X
β
1
?
?
Pr[detect] =
pi = p≤s ≤ 1 − 1 −
= 1 − β ≈ β,
2n
e
i=0
which means that the probability that a randomly chosen molecule is in detect
state when chosen uniformly at random is at most β.
Probability of False Negatives with Leaks. Under the same leak model, it is
easy to notice that the “best” adversarial strategy for our algorithm in case D is
present is to concentrate all leaks to create the neutral species N (or X∞ in case
of Algorithm 2). It is easy to see that this just decreases the total probability of
detect states by the leak probability λ = β/n. More formally, we compute once
again stationary probabilities. The recurrent relation is
p?≤i
=
p?≤i
2
1−
n
2
β
β n−1 ?
?
2
+ · (1 − (1 − p≤(i−1) ) )
1−
+ ·
p .
n
n
n
n ≤i
β
β
Using estimate (1 − nβ )/(1 − 2n
) ≈ (1 − 2n
) we reach
β
(1 − (1 − p?≤(i−1) )2 ).
p?≤i = 1 −
2n
Thus we have for the first log(n/k) levels the dampening factor of (1 − β/(2n))
per level (compared to the leakless case). It can be easily shown by induction
that
i
2i
2i
β
k
k
1−
1−
≤ p?≤i ≤ 1 −
.
2n
n
n
The estimates for p?i follow from the leakless case, after taking into the account
the composed dampening factor:
log n
s
X
β
1
log n
1
· 1−
=1− −O
β .
Pr[detect] =
p?i = p?≤s ≥ 1 −
e
2n
e
n
i=0
Finally, we summarize the results in this section as follows:
Theorem 1. Assuming leak rate β/n for β 1, Robust-Detect guarantees the
following.
– The probability of a false positive is at most β.
– The probability of a false negative is at most 1/e + O(β · (log n)/n).
Notice that these probabilities can be boosted by standard sampling techniques.
5.2
Convergence Analysis
We now proceed with an analysis of the convergence speed of the previously
described protocols. To avoid separate analysis for each of the aforementioned
cases (no leaks, false positives, false negatives) and to be independent from all
possible initializations of the algorithm, we first start with showing that, under
no leaks and with no D present, all states X1 , . . . , Xc are quickly killed.
In this section, it is more convenient to use t to refer to the total number of
interactions, rather than parallel time. To convert to parallel time, one needs to
divide by n, the number of molecules.
Lemma 1. Assume arbitrary (adversarial) initial state in t = 0 and evolution
with no leaks (β = 0) and no D is present. For any c(n) ≥ 1, there is t =
O(n · (c(n) + log n)) such that with probability 1 − 1/nΘ(1) (with high probability)
there is no molecule in any of the states X1 , X2 , . . . , Xc(n) after t interactions.
Proof. We assign a potential to each molecule, based on the state it is currently in:
Φ(Xi ) = 3−i . We also define a global potential Φt as sum of all molecular potentials
after t interactions. Observe, that when two molecules interact, following rule
Xi + Xj → Xmin(i,j)+1 + Xmin(i,j)+1 , then there is:
Φ(Xmin(i,j)+1 ) + Φ(Xmin(i,j)+1 ) ≤ 2/3 · (Φ(Xi ) + Φ(Xj )) ,
which can be interpreted that each interacting molecule loses at least 1/3 of its
potential. Since each molecule participates in an interaction with probability n2
in each round, the following bound holds:
X
2
1
Φt ,
E[Φt+1 − Φt |Φt ] ≥ ·
Pr(v interacts in round t) · Φt (v) =
3
3n
v
2
Φt .
E[Φt+1 |Φt ] ≤ 1 −
3n
Substituting Φ0 ≤ n and fixing t ≥ 32 n ln(n · 3c(n) · nΘ(1) ) = O(n(c(n) + log n +
Θ(log n))) we have
t
c(n)
Θ(1)
2
1
E[Φt ] ≤ 1 −
· n ≤ e− ln(n·3 ·n ) · n = 3−c(n) · Θ(1) .
3n
n
By Markov’s inequality, this means that there is no molecule in any of the states
X1 , X2 , . . . , Xc with probability at least 1 − n−Θ(1) , that is with high probability.
t
u
We mention one additional useful property of Algorithm 2, that its actions on
population are decomposable with respect to levels. That is, define levelt (u) = i
if molecule u at time t is in state Xi , and levelt (u) = 0 if it is in state D.
Observation 1 Let {u1 , u2 , . . . , un }, {v1 , v2 , . . . , vn }, {w1 , w2 , . . . , wn } be 3 disjoint populations each on n molecules, following evolution defined by Algorithm 2.
Moreover, let their evolutions be coupled: at each time t, in each population the
corresponding molecules interact (i.e., the interaction is ui + uj , vi + vj , wi + wj
in the three populations for some i, j).
If ∀i level0 (ui ) = min(level0 (vi ), level0 (wi )), then at any time t > 0 :
∀i levelt (ui ) = min(levelt (vi ), levelt (wi )).
This observation can be naturally generalized to more than 3 populations. As
shown below, the observation implies that to analyze detection under noisy start,
we can decouple starting noise from detected particle and analyze evolution
under those two separately. Denote by pi (t) and p≤i (t) the probability for a randomly picked molecule after t interactions to be in the state Xi or D, X1 , . . . , Xi
respectively.
Theorem 2. Fix arbitrary leak model (i.e. no leaks, false-positives, false-negatives)
and arbitrary concentration of D. For any c ≥ 1, and t = Ω(n · (c + log n)), there
is
p?≤c − p≤c (t) ≤ 1/nΘ(1) ,
where p? is the stationary probability distribution of the identical process.
Proof. First, for simplicity we collapse all states Xc+1 , Xc+2 , . . . into N , since it
has no effect on p≤c distributions. Consider a population of size n, under no leaks,
no D, evolution. By Lemma 1, in τ = O(n · (c + log n)) steps it reaches all-N
state, regardless of initial configuration, with high probability. Thus evolution
of any population {ui }, under no leaks, with D present, is a coupling (as in
Observation 1) of following evolutions:
– initial configuration of population {ui }, with each D replaced with N ;
– for every timestep ti such that D interacted with Xi or N creating X1 , we
couple a population with corresponding molecule set to X1 and every other
molecule set to N , shifted in time so its evolution starts at time ti .
Observe, that evolution of population of both types will reach all-N state in
τ steps, with high probability. Thus, conditioned on this high probability, the
configuration at any t ≥ τ is the result of coupling of all-N (result of evolution
of first type population) with possibly several configurations of the second type,
where at each timestep t0 ∈ [t − τ, t] such population was created independently
with some probability only depending on n and k. However, the coupling we just
described is invariant from the choice of t, as long as t ≥ τ . Thus, P
for any t1 , t2 ≥ τ ,
t
there is |p≤c (t1 ) − p≤c (t2 )| ≤ 1/nΘ(1) . Since p?≤c = limt→∞ 1t i=1 p≤c (i), the
claimed bound follows.
To take into account errors, we say that whenever there is a leak changing
state of molecule v to some state S at time t, we change state of v at that time
in all existing populations to N , and create new population where v has state S,
and all other molecules are in N state. The same reasoning as in the error-less
case follows, since switching molecules to N state it only speeds up convergence of
populations to all-N state and since populations created due to leaks are created
at each step with the same probability depending only on n and error model. t
u
6
Simulation Results
We simulated the Robust-Detect algorithm (Algorithm 1) using a modified version
of the CRNSimulatorSSA Mathematica package [17]. Figure 2 shows the shape
of typical trajectories when there is one molecule in state D (k = 1), compared
with no molecules in state D (k = 0) but with the worst-case leak for falsepositives. Note that D is quickly detected if present, and if absent the system
exhibits random perturbations that are quickly extinguished and are clearly
distinguishable from the true positive case.
Fig. 2. Typical time-evolution of the Robust-Detect algorithm (Algorithm 1). Three
colors correspond to the three conditions described in Fig. 1: (blue) 1 molecule of D is
present and no leak (leak parameter β = 0); (orange, red) no D is present but with
worst-case false-positive leak Xi → X1 and N → X1 (orange: leak parameter β = 0.01,
red: leak parameter β = 0.1). All Xi states map to output value “detect”, and thus we
plot the sum of all their counts. (a) s = 14 layers, β = 0.1 (red). (b) s = 17 layers,
β = 0.01 (orange). See Fig. 1 for the corresponding steady state probabilities. Note
that with smaller leak (β = 0.01), it is possible to better distinguish true positives and
false positives by increasing the number of layers (from 14 to 17). In all cases there
are n = 104 molecules, and the initial configuration is all molecules in neutral state N .
Parallel time (number of interactions divided by n) corresponds to the natural model
of time where each molecule can interact with a constant number of other molecules
per unit time.
7
Conclusions
We have considered the problem of modeling and withstanding leaks in chemical
reaction networks, expressed as population protocols. We have presented an
arguably simple algorithm which is probabilistically correct under assumptions
on the leak rate, and converges quickly to the correct answer.
Beyond the specific example of robust detection, we hope that our results
motivate more systematic modeling of leaks, and further work on algorithmic
techniques to withstand them. As such errors appear to spring from the basic
laws of chemistry, their explicit treatment appears to be necessary. The authors
found it surprising that many of the algorithmic techniques developed in the
context of deterministically correct population protocols might not carry over to
implementations, due to their inherent non-robustness to leaks.
In future work, we plan to perform an exhaustive examination of which of
the current algorithmic techniques could be rendered leak-robust, and whether
known algorithms can be modified to withstand leaks via new techniques. Another
interesting avenue for future work is lower bounds on the set of computability
or complexity of fundamental predicates in the leak model. Finally, we would
like to examine whether our robust detection algorithm can be implemented in
strand displacement systems.
Acknowledgments. We thank Lucas Boczkowski and Luca Cardelli for helpful
comments on the manuscript.
References
1. M. Gopalkrishnan, “Catalysis in reaction networks,” Bulletin of mathematical biology,
vol. 73, no. 12, pp. 2962–2982, 2011.
2. D. Soloveichik, G. Seelig, and E. Winfree, “DNA as a universal substrate for chemical
kinetics,” Proceedings of the National Academy of Sciences, vol. 107, no. 12, pp.
5393–5398, 2010.
3. L. Cardelli, “Two-domain DNA strand displacement,” Mathematical Structures in
Computer Science, vol. 23, no. 02, pp. 247–271, 2013.
4. Y.-J. Chen, N. Dalchau, N. Srinivas, A. Phillips, L. Cardelli, D. Soloveichik, and
G. Seelig, “Programmable chemical controllers made from DNA,” Nature Nanotechnology, vol. 8, no. 10, pp. 755–762, 2013.
5. D. Y. Zhang, A. J. Turberfield, B. Yurke, and E. Winfree, “Engineering entropydriven reactions and networks catalyzed by DNA,” Science, vol. 318, no. 5853, pp.
1121–1125, 2007.
6. D. Angluin, J. Aspnes, Z. Diamadi, M. Fischer, and R. Peralta, “Computation in
networks of passively mobile finite-state sensors,” Distributed Computing, vol. 18,
pp. 235–253, 2006, preliminary version appeared in PODC 2004.
7. R. M. Karp, C. Schindelhauer, S. Shenker, and B. Vöcking, “Randomized rumor
spreading,” in 41st Annual Symposium on Foundations of Computer Science, FOCS
2000. IEEE Computer Society, 2000, pp. 565–574.
8. J. J. Hopfield, “Kinetic proofreading: a new mechanism for reducing errors in biosynthetic processes requiring high specificity,” Proceedings of the National Academy of
Sciences, vol. 71, no. 10, pp. 4135–4139, 1974.
9. C. Thachuk, E. Winfree, and D. Soloveichik, “Leakless DNA strand displacement
systems,” in DNA Computing and Molecular Programming. Springer, 2015, pp.
133–153.
10. D. Angluin, J. Aspnes, and D. Eisenstat, “A simple population protocol for fast
robust approximate majority,” Distributed Computing, vol. 21, no. 2, pp. 87–102,
2008.
11. L. Cardelli, “Morphisms of reaction networks that couple structure to function,”
BMC Systems Biology, vol. 8, no. 1, p. 84, 2014.
12. B. Pittel, “On spreading a rumor,” SIAM Journal on Applied Mathematics, vol. 47,
no. 1, pp. 213–223, 1987.
13. B. Doerr, C. Doerr, S. Moran, and S. Moran, “Simple and optimal randomized
fault-tolerant rumor spreading,” Distributed Computing, vol. 29, no. 2, pp. 89–104,
2016.
14. L. Boczkowski, A. Korman, and E. Natale, “Minimizing message size in stochastic
communication patterns: Fast self-stabilizing protocols with 3 bits,” in Proceedings
of the Twenty-Eighth Annual ACM-SIAM Symposium on Discrete Algorithms,
SODA 2017, pp. 2540–2559.
15. B. Dudek and A. Kosowski, “Spreading a confirmed rumor: A case for oscillatory
dynamics,” CoRR, vol. abs/1705.09798, 2017.
16. D. Soloveichik, “Robust stochastic chemical reaction networks and bounded tauleaping,” Journal of Computational Biology, vol. 16, no. 3, pp. 501–522, 2009.
17. http://users.ece.utexas.edu/~soloveichik/crnsimulator.html.
| 8cs.DS
|
1
LOOP Descriptor: Local Optimal Oriented Pattern
arXiv:1710.09317v2 [cs.CV] 2 Feb 2018
Tapabrata Chakraborti, Brendan McCane, Steven Mills and Umapada Pal
Abstract—This letter introduces the LOOP binary descriptor
(local optimal oriented pattern) that encodes rotation invariance
into the main formulation itself. This makes any post processing
stage for rotation invariance redundant and improves on both
accuracy and time complexity. We consider fine-grained lepidoptera (moth/butterfly) species recognition as the representative
problem since it involves repetition of localized patterns and
textures that may be exploited for discrimination. We evaluate
the performance of LOOP against its predecessors as well as
few other popular descriptors. Besides experiments on standard
benchmarks, we also introduce a new small image dataset on NZ
Lepidoptera. Loop performs as well or better on all datasets
evaluated compared to previous binary descriptors. The new
dataset and demo code of the proposed method are to be
made available through the lead author’s academic webpage and
GitHub.
Keywords—Local Binary Patterns, Collaborative Representation,
Local Optimal Oriented Pattern (LOOP), Fine-grained Recognition.
I. I NTRODUCTION
Local binary descriptors have been shown to be effective
encoders of repeated local patterns for robust discrimination
in several visual recognition tasks [1]. He and Wang [2], in
their seminal paper on the subject, introduced the concept of
textons. These encode localized textures/patterns in an image
into binary words, and the frequency histogram of these words
describes the image. The first popular implementation was
local binary pattern (LBP) by Ojala et al [3].
Since then, many interesting modifications and improvements of these descriptors have been developed. A few of
these are modified census transform (MCT) [4], local gradient
pattern (LGP) [5], local directional pattern (LDP) [6], uniform
local binary pattern (ULBP) [17], etc. LBP encodes the local
intensity variation in the 3 × 3 neighborhood of each image
pixel into a binary word, the decimal equivalent of which then
acts as a representative feature encapsulating the pattern of
local intensity variation in that neighborhood. The histogram
of the LBP values over all pixels of an image serves as the
descriptor for that image. MCT [4] is a modified version
of LBP, where instead of taking the central pixel intensity
of the neighborhood as the threshold, the average intensity
over the entire neighborhood is treated as the threshold while
forming the binary word. LGP [5] considers the gradient of
neighboring pixel intensities with respect to the central pixel
intensity to generate the binary word. Both MCT and LGP
Tapabrata Chakraborti, Brendan McCane and Steven Mills are with the
Department of Computer Science, University of Otago, NZ.
e-mail: tapabrata;mccane;steven at cs.otago.ac.nz.
Umapada Pal is with the Computer Vision and Pattern Recognition Unit at
the Indian Statistical Institute.
e-mail: umapada at isical.ac.in.
were demonstrated to be more robust to noise than the original
LBP operator [5],[6]. LDP was proposed by Jabid et al. [6] as
a more noise invariant alternative to LBP. It incorporates the
outputs of the 3 × 3 Kirsch masks in 8 directions, for each
pixel, into a binary word and its corresponding decimal value
is the desired LDP value, corresponding to that pixel.
An inherent shortcoming of both LBP and LDP, as well as
many of the other related descriptors like MCT, LGP, etc., is
that each assigns an ad hoc bias while sequencing the generated
binary digits to form the binary word. This is because the
choice of the initial pixel of this circular sequence of bits
determines the binary-to-decimal conversion weight assigned
to each of the binary digits, as explained later in the paper. This
adds an inherent rotation variance. A few rotation invariant
versions have also been suggested in existing literature, like
LDPri [16] for LDP. But these have their own set of limitations
and the present work presents an alternative rotation invariant
scheme.
The main contributions of this paper are:
1) LOOP descriptor: We encapsulate rotation invariance
into the main formulation of local binary descriptor,
thus overcoming a drawback of most existing descriptors of the genre. In the process we decrease post
processing time complexity and increase accuracy of
classification. The demo code for LOOP will be made
available through the lead author’s GitHub account.
2) NZ Lepidoptera dataset: We present a new finegrained image dataset of NZ Lepidoptera. Small finegrained datasets can have significant real life applications, like native/endemic species recognition and biodiversity analysis. It is difficult to effectively train deep
networks on such small specialized datasets, and hence
fine-tuned approaches like the present one still have
practical significance. The dataset will soon be available
at the lead author’s academic webpage.
Note that though we have shown the efficacy of the proposed
method on butterfly and moth species recognition, the methodology may be applied to other similar applications exhibiting
repeated local patterns/textures.
II.
P REVIOUS D ESCRIPTORS
A brief review of LBP and LDP is presented here, since
the proposed LOOP descriptor is an improvement designed on
them.
A. Local Binary Pattern (LBP)
LBP [3] is a popular descriptor which captures the local
intensity variation patterns of an image and has good discrimination characteristics.
2
Let ic be the intensity of an image I at pixel (xc , yc ) and
in (n = 0, . . . , 7) be the intensity of a pixel in the 3 × 3
neighborhood of (xc , yc ) excluding the center pixel ic .
Then the LBP value for the pixel (xc , yc ) is given by
LBP (xc , yc ) =
7
X
s(in − ic ).2n
(1)
n=0
where
s(x) =
if x ≥ 0
otherwise
1
0
(2)
A major disadvantage of LBP is the arbitrary sequence of
binarization weights. Depending on the chosen starting pixel
of the sequence of binary weights (2n , n = 0, . . . , 7), the 8
neighbors of the output 3 × 3 grid are allocated subsequent
weightage n sequentially. There is no clear logic behind
the proper assignment of weight and the result obtained is
susceptible to rotation variance. The same pattern rotated
between images of the same class or even within different
parts of the same image will generate a different binary word,
thereby confounding the classification process. In fact, this bias
has persisted over to other related descriptors as well, like LDP,
MCT, LGP, etc.
B. Local Directional Pattern (LDP)
LDP is an improved local pattern descriptor which incorporates a directional component by using Kirsch compass kernels.
It was shown to be less susceptible to noise than the traditional
LBP operator [6].
Let ic be the intensity of an image I at pixel (xc , yc ) and
in , n = 0, 1, . . . , 7 be the intensity of a pixel in the 3 × 3
neighborhood of (xc , yc ) excluding the center pixel ic . 3 × 3
Kirsch edge detectors centered at (xc , yc ) in eight possible
directions are given in Fig. 1.
The 8 responses of the Kirsch masks are mn , n = 0, . . . , 7
corresponding to pixels with intensity in , n = 0, . . . , 7 and
let mk be the k th highest Kirsch activation. Then all the
neighboring pixels having Kirsch response higher than mk is
assigned 1, and others 0.
But the empirically assigned value of k is ad hoc [6]. This
fixes the possible number of ones to k −1 and number of zeros
to (n + 1 − (k − 1) = n − k + 2 where n is as defined above by
the neighborhood pixel number. Hence the possible number of
n+1
binary words is reduced from 2(n+1) to Ck−1
.
Then the LDP value for the pixel (xc , yc ) is given by
LDPk (xc , yc ) =
7
X
s(mn − mk ).2n
as in previous formulations. Thus it assigns an emperical rule
to the starting point of the binary word construction. However,
it suffers from the self-imposed restriction of always having a
leading 1, which immediately reduces the number of available
combinations in the binary word by half. The problem of fixed
number of 1s and 0s also persists from the original LDP,
depending on the value of the threshold k.
III. P ROPOSED M ETHODOLOGY
The proposed Local Optimal Oriented Pattern (LOOP) is
described here in details.
A. Local Optimal Oriented Pattern (LOOP)
As discussed earlier, the major disadvantage of LBP and
LDP is the arbitrary sequence of binarization weights that adds
dependancy to orientation. LDP also suffers from the empirical
assignment of value to the threshold variable, which puts an
ad hoc restriction on the number of bits allowed to be 1, thus
reducing the number of possible words, as discussed before.
LOOP presents a non-linear amalgamation of LBP and LDP
that overcomes these drawbacks while preserving the strengths
of each.
Let ic be the intensity of an image I at pixel (xc , yc ) and
in (n = 0, 1, . . . , 7) be the intensity of a pixel in the 3 × 3
neighborhood of (xc , yc ) excluding the center pixel ic . The 8
Kirsch masks, as used in LDP previously, are oriented in the
direction of these 8 neighboring pixels in (n = 0, 1, . . . , 7)
thus giving a measure of the strength of intensity variation in
those directions, respectively.
This leads us to propose the incorporation of this information by assigning the binarization weight to each neighboring
pixel corresponding to the strength of Kirsch output in the
direction of that pixel. The underlying rationale behind this
approach is that the Kirsch mask output in a particular direction
provides an indication of the probability of occurrence of an
edge in that direction. Since the LBP indicates the intensity
variation over the neighboring pixels in the same directions, the
value of the Kirsch output is employed to assign the decimalto-binary weightage.
As discussed earlier, the 8 responses of the Kirsch masks
are mn corresponding to pixels with intensity in , n = 0, . . . , 7
Each of these pixels are assigned an exponential wn (a digit
between 0 and 7) according to the rank of the magnitude of
mn among the 8 Kirsch mask outputs.
Then the LOOP value for the pixel (xc , yc ) is given by
(3)
LOOP (xc , yc ) =
n=0
where
s(x) =
1
0
7
X
s(in − ic ).2wn
(5)
n=0
if x ≥ 0
otherwise
where
(4)
Rotation Invariant LDP (LDP ri ): A rotation invariant
version is introduced in [16]. Here the neighbor pixel with
highest Kirsch mask output is assigned the highest order in
the binary word, and then the other bits are taken sequentially
s(x) =
1
0
if x ≥ 0
otherwise
(6)
Thus the LOOP descriptor encodes rotation invariance into
the main formulation. Moreover, the proposed LOOP algorithm also negates the empirical assignment of the value of
the parameter k in the traditional LDP method (eqn. 3).
3
(a) Admiral
(b) Blue
(c) Copper
(d) Ringlet
(e) Erebidae
(f) Geometridae
(g) Hepialidae
(h) Noctuidae
Fig. 2: Sample images from each class of the new NZ Lepidoptera dataset. It has 640 images currently with 80 images per
class. (a)-(d) are butterfly species & (e)-(h) are moth species.
IV.
Fig. 1: Numerical Example to calculate LBP, LDP and LOOP.
B. Scale and Rotation Invariance
A multi-scaled amalgamated histogram is constructed to
achieve scale-independence. This is done by forming a spatial
gaussian pyramid and then concatenating the histograms of
LOOP values obtained at each scale to form the final histogram
which acts as the descriptor for the image.
Figure 1 illustrates the rotation invariance property of LOOP
descriptor, compared to the lack thereof in LBP and LDP. As
demonstrated in Figure 1, binary words are formed according
to LBP rule and the weights are assigned according to the LDP
mask activations.
Tie break in weight assignment: Referring to Figure 1,
the one with more differing nearest neighbour is assigned
higher weight. Eg. For the two -2155 for pattern 2, the nearest
neighbours for one are -2155 and -275 (difference is 1880)
while nearest neighbours for the other are -2155 and -2035
(difference is 120). So the former is assigned higher weight
than latter (21 vs. 20 ).
E XPERIMENTS AND R ESULTS
A. Experimental Setup
Datasets. 3 image datasets of moth and butterfly species
have been used to showcase the performance of the methods.
Leeds butterfly dataset [7] was built in 2009 at the University
of Leeds, UK. It contains 832 images of 10 species of
butterflies with 55 to 100 images per catagory.
Ponce butterfly dataset [8] was built by the Ponce Group at
Beckman Institute, University of Illinois at Urbana-Champaign
in 2004. It has 7 butterfly types with a total of 619 images.
The smallest class contains 42 images and the largest class has
111 images.
NZ Lepidoptera dataset is a new benchmark built during
this work at the Department of Computer Science, University
of Otago, NZ in collaboration with the CVPR Unit, Indian
Statistical Institute. It has 8 classes of NZ butterflies and
moths, 4 categories each. Currently it has 640 images with 80
images per class, subject to expansion in near future. Images
of NZ moths have been compiled from the publicly available
database of NZ Landcare Research. Sample images of the NZ
Lepidoptera dataset are presented in Figure 2.
Classifiers. The collaborative representation classifier
(CRC) [9] has been shown recently to be effective in handling
small fine-grained datasets where the differences in objects
between classes is subtle compared to randomized significant
background variation within and between classes [10]. CRC
represents the test image as an optimal weighted average of
training images of all classes and the subsequent residual per
class is used to calculate the predicted category. We adopt a
recent Probabilistic formulation of CRC called ProCRC [11].
For comparison with a popular off-the-shelf classifier, a
support vector machine (SVM) [12] with a χ2 kernel is
chosen with settings as in the ProCRC paper [11] for fair
comparison. Multiclass categorization is performed with the
binary SVM classifier in a one-versus-all fashion in turns.
4
TABLE I: Classification Accuracy (%)
Leeds
(a) Sample image
(b) LOOP output
Fig. 3: Standard test image (rice.png) and the LOOP output.
Descriptors. We compare the performance of several local
binary pattern encoders like LBP, MCT, LDP, LDPri , LGP
with the proposed LOOP descriptor. Among these, LOOP is
built influenced by LBP and LDP, while MCT and LGP are
chosen as relevant modifications of these methods. LDPri is
a rotation invariant modification of LDP. We also compare
with three popular modern binary descriptors: BRIEF (Binary Robust Independent Elementary Features) [13], BRISK
(Binary Robust Invariant Scalable Keypoints) [14], and ORB
(Oriented FAST and rotated BRIEF) [15]. BRIEF was the
first of these and presents a simple configuration similar to
LBP, without rotation invariance. ORB descriptor is rotation
invariant and uses an optimal sampling pair. BRISK has both
of these attributes and also has the additional characteristic of
a more advanced hand-crafted sampling pattern composed of
concentric rings.
B. Results and Discussion
For each dataset, experiments are conducted with 5 fold
cross validation and percentage classification accuracies are
presented in Table 1 with the highest accuracy in each column
highlighted in bold. The LOOP output on a standard test image
’rice.png’ is presented in Figure 3 as an illustrative example.
It is observed that LOOP performs much better than LBP
and LDP, the two descriptors on which it is based. Performance
of LOOP is also better than the competing LBP variants: MCT,
LGP and LDPri . LOOP successfully outperforms BRIEF,
but has comparable results with BRISK and ORB, both of
which are rotation invariant. However, LOOP has a simpler
formulation and lower execution time than ORB and BRISK.
Using the standard tic-toc functionalities of MATLAB, it is
observed that ORB and BRISK have 21% and 27% higher
computation time respectively than LOOP.
Among the competing classifiers listed in Table1, only
LDPri , BRISK and ORB are rotation invariant. LOOP yields
comparable results to BRISK and ORB, but has lower run
time due to simpler formulation. LOOP yields only marginal
improvement in performance compared to LDPri , but has
similar complexity in formulation and comparable run time.
We determine whether the increase in average accuracy of
LOOP over LDPri is statistically significant.
Ponce
NZ
SVM
ProCRC
SVM
ProCRC
SVM
ProCRC
LBP
62.1
64.7
68.4
70.5
55.6
59.3
MCT
63.9
65.8
69.9
72.3
58.7
61.1
LDP
66.6
68.5
71.7
74.1
60.9
64.5
LDPri
69.2
72.5
75.1
77.9
64.8
68.3
LGP
69.4
72.9
75.0
77.6
64.2
68.6
BRIEF
65.5
67.1
70.4
73.6
59.9
63.0
BRISK
69.8
73.5
77.7
79.3
66.2
69.5
ORB
71.0
73.8
78.4
79.9
65.8
70.1
LOOP
71.5
74.4
78.3
80.4
66.0
70.6
Sign Binomial Test. For each descriptor, we have 2 classifiers and 3 datasets, hence 6 combinations per descriptor. Also
there are 5 fold cross-validation per combination. So for each
descriptor we have 30 sets of accuracy results. Assuming the
null hypothesis to be that the two competing methods (LOOP
and LDPri ) are equally good, then there is 50% chance of each
beating the other.
It is observed that of the 30 experimental runs, LOOP
outperforms LDPri 22 times. The one-tail P value at 5% level
of significance is 0.0081. Now using Bonferroni correction, at
5% level of significance, the corrected α for the 6 combinations
(2 classifiers and 3 datasets per descriptor) is 0.05/6 = 0.0083.
Since the calculated chance is 0.0081 (less than the corrected
α), we can reject the null hypothesis and conclude that LOOP
has a statistically significant better performance than LDPri .
Also as explained earlier, LDPri has the constraint of always
having a leading 1 thus halving the number of possible words,
along with the restriction of having a fixed number of 1s and
0s in the binary word. LOOP is free from these limitations
design.
V.
C ONCLUSION
A novel binary local pattern descriptor, LOOP, which overcomes some disadvantages of its predecessors LBP and LDP, is
presented. It is tested on Lepidoptera species recognition with
encouraging initial results that warrant further exploration. It
outperforms the descriptors on which it is based, along with
a few other variants. It has comparable results with popular
binary descriptors like BRISK and ORB, but gains in time
complexity.
A new benchmark dataset of NZ Lepidoptera images is also
introduced. The proposed method may be used in similar finegrained applications with small datasets, where fine-tuning a
deep network is difficult.
The present letter has only presented experiments and results
on one representative problem, that of Lepidoptera classification. But LOOP is a generalized binary descriptor and may be
used in further research for other small fine-grained datasets
where it is difficult to train deep convolutional networks.
5
R EFERENCES
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
D. Huang, C. Shen, M. Ardabilian, Y. Wang, and L. Chen, “Local Binary
Patterns and Its Application to Facial Image Analysis: A Survey,” IEEE
Trans. on Systems, Man and Cybernetics- Part C, vol. 41, no. 6, pp. 765781, Nov 2011.
DC. He and L. Wang, “Texture Unit, Texture Spectrum, And Texture
Analysis,” IEEE Trans. on Geoscience and Remote Sensing, vol. 28, pp.
509 - 512, 1990.
T. Ojala, M. Pietikinen, and D. Harwood, “Performance evaluation of
texture measures with classification based on Kullback discrimination of
distributions,” In Proc. ICPR, 1994.
B. Frobaand and A. Ernst, “Face Detection with the Modified Census
Transform,” In Proc. IEEE Conf. on Face and Gesture Recognition, pp.
9196, May 2004.
B. Jun and D. Kim, “Robust face detection using local gradient patterns
and evidence accumulation,” Pattern Recognition, vol. 45, no. 9, pp.
3304-3316, Sep. 2012.
T. Jabid, M. H. Kabir, and O. S. Chae, “Gender Classification using
Local Directional Pattern (LDP),” In Proc. ICPR, 2010.
J. Wang, K. Markert, and M. Everingham, “Learning Models for Object
Recognition from Natural Language Descriptions,” In Proc. BMVC,
2009.
S. Lazebnik, C. Schmid, and J. Ponce, “Semi-Local Affine Parts for
Object Recognition,” In Proc. BMVC, 2004.
L. Zhang, M. Yang, and X. Feng, “Sparse Representation or Collaborative
Representation: Which Helps Face Recognition?,” In Proc. ICCV, 2011.
T. Chakraborti, B. McCane, S. Mills, and U. Pal, “Collaborative
representation basedfine-grained species recognition,” In Proc. IVCNZ,
2016.
S. Cai, L. Zhang, W. Zuo and X. Feng, “A Probabilistic Collaborative
Representation based Approach for Pattern Classification,” In Proc.
CVPR, 2016.
C. Cortes and V. Vapnik, “A support vector networks,” Machine
Learning, vol. 20, no. 3, pp. 273297, Sept 1995.
M. Calonder, V. Lepetit, C. Strecha, and P. Fua, Brief: Binary robust
independent elementary features, In Proc. ECCV, 2010.
S. Leutenegger, M. Chli, and R. Y. Siegwart, BRISK: Binary robust
invariant scalable keypoints, In Proc. ICCV, 2011.
E. Rublee, V. Rabaud, K. Konolige, and G. Bradski, ORB: an efficient
alternative to SIFT or SURF, In Proc. ICCV, 2011.
T. Jabid, M. H. Kabir, and O. S. Chae, “Local directional pattern
(LDP)A robust image descriptor for object recognition,” Proc. IEEE
Conf. on Advanced Video and Signal Based Surveillance (AVSS), 2010.
Y. Ming, G. Wang, C. Fan, “Uniform Local Binary Pattern Based
Texture-Edge Feature for 3D Human Behavior Recognition”, Plos One,
vol. 10, no. 5, pp: 1-15, 2015.
| 1cs.CV
|
A Line-Point Unified Solution to Relative Camera Pose
Estimation*
Ashraf Qadir, Jeremiah Neubert
Abstract— In this work we present a unified method of
relative camera pose estimation from points and lines
correspondences. Given a set of 2D points and lines
correspondences in three views, of which two are known, a
method has been developed for estimating the camera pose of
the third view. Novelty of this algorithm is to combine both
points and lines correspondences in the camera pose estimation
which enables us to compute relative camera pose with a small
number of feature correspondences. Our central idea is to
exploit the tri-linear relationship between three views and
generate a set of linear equations from the points and lines
correspondences in the three views. The desired solution to the
system of equations are expressed as a linear combination of
the singular vectors and the coefficients are computed by
solving a small set of quadratic equations generated by
imposing orthonormality constraints for general camera
motion. The advantages of the proposed method are
demonstrated by experimenting on publicly available data set.
Results show the robustness and efficiency of the method in
relative camera pose estimation for both small and large
camera motion with a small set of points and line features.
I. INTRODUCTION
Accurate and efficient camera pose estimation is
essential to a number of applications such as structure from
motion, robot navigation and simultaneous localization and
mapping. One key step is to estimate the relative position
and orientation of the views using the geometric constraints
arising from a 3D object and its corresponding 2D images in
multiple views. Most existing methods use images of scene
features such as corner points [1, 2, 3] or lines [4, 5] with
points being the primary feature of interest. However, almost
all of these methods use either points or lines but do not
handle both. Using either lines or points does not allow one
to take full advantages of the information available in the
images. While natural scenes are rich in feature points, manmade environment such as office interiors or city landscapes
can consists of texture less planar surfaces where there are
few point features that can be reliably detected. On the other
hand, such environments consists of large number of visible
lines. A system that can handle both lines and points, either
independently or combined has an advantage to be useful in
a wide range of applications. While visual SLAM systems
mostly deal with sequential images with small camera
motions, structure from motion methods often require to deal
with an unordered set of images with large camera motions.
In this work we propose an algorithm that gives a unified
linear approach that can deal with a mixture of points and
lines for relative pose estimation using three views. Given a
set of lines and points correspondences between views, the
algorithm exploits the tri-linearity relationship between the
views to generate a set of constraint equations from the
points and lines correspondences. The desired solution to the
system of equations are expressed as a linear combination of
the singular vectors and the coefficients are computed by
solving a small set of quadratic equations generated by
imposing orthonormality constraints for general camera
motion.
The proposed method considers the image sequence is
generated from a monocular camera and equation
representations used are adopted to three views. However,
the method can be easily extended to stereo systems with
immediate advantage that only two view are required. It also
generates an added set of constraint equations when using
points and line correspondences in all 4 views.
In theory given two known camera pose, it is possible to
reconstruct 3D lines and points from the corresponding
images and then estimate the third camera pose from the 3D
to 2D points and lines correspondences. However, the
quality of the camera pose estimation is critically dependent
on the accuracy of the reconstructed 3D points and lines. It
is especially critical with narrow base line and forward
motion [4].
The major contribution of this paper is the development
of a unified framework for relative camera pose estimation
combining points and lines. The algorithm is robust,
efficient and can handle a wide range of line and point
combination. The intended use of the algorithm is with
robust hypothesize-and-test frameworks but is also suitable
for scenarios with a large number of points and lines. Using
both points and line features mean the proposed method is
equally applicable in natural as well as manmade urban and
indoor environments.
II.
*Research supported partially by ND EPSCoR through NSF grants
#EPS-0814442, #IIA-1355466 and the University of North Dakota
Ashraf Qadir is with RFA Engineering, Minnesota, USA (phone: 1-701213-8094; e-mail: aqadir@ rfamec.com).
Jeremiah Neubert is with Mechanical Engineering Department,
University of North Dakota, Grand Forks, North Dakota.(email:
[email protected])
LITERATURE REVIEW
While line based structure from motion has received
considerable attention, most of the existing systems that are
robust and computationally efficient are points based
systems. A primary reason is that points provide stronger
constraints than lines as shown in [6]. For instance, the 5
points minimal solution exists [1] for two view
reconstruction for a calibrated camera. However, lines do
not put such constraints on the camera pose and at least 3
images are required for camera placement derivation. As a
result, a common approach is to use multifocal tensors [6, 7]
for structure from motions from lines. The disadvantage is
that a large number of lines in three views are required
making it less desirable in a hypothesis-and-test framework
for motion estimation. Yet there are situations where using
lines are beneficial or required for a successful structure
from motion system.
Important breakthrough in solving minimal problems for
points [1, 2] led to the development of efficient point based
camera pose estimation and reconstruction methods. A few
examples are visual odometry methods [8, 9, 10] or urban
reconstruction [11, 12] that uses the 5 points algorithm
inside a RANSAC framework for relative camera pose
estimation. These methods rely on the assumption that many
points can be reliably detected and localized. However,
stability of 5 points algorithm decreases for forward motions
as well as large depth of the features. For robotics
application, often additional sensors (IMU, odometer etc.)
and constraints are used along with vision for relative pose
estimation [13].
One of the earlier methods for line based 3D structure
and camera motion estimation is described in [14]. In order
to obtain an estimation of the scene structure and the motion
of the camera together, an objective function is defined that
minimizes the re-projection error in the image plane. The
method is robust to variable end points across views.
However, they use iterative nonlinear minimization which
requires initialization. The system also requires at least 6
line correspondences over three views. A similar method is
presented in [15] for urban environment.
A detail investigation on 3D line representation,
triangulation and bundle adjustment is performed in [16]
where the camera pose estimation was performed using
trifocal tensor. An efficient algorithm for relative camera
pose estimation using two stereo image pairs (four images)
is presented in [4]. The algorithm uses 2 or 3 line
correspondences in four images of two stereo image pairs to
estimate the relative camera pose between two stereo frames
in a RANSAC framework. That paper contains elements of
ideas contained in the present paper which extends the idea
to monocular camera with a combine points and line based
relative pose estimation as well as large camera motions.
Focusing on indoor and/or urban application, one
category of line based SFM from lines algorithms proposed
in [17, 18] based on the assumptions such as existence of
three orthogonal directions or primitive configurations such
as parallel and orthogonal lines. In [18], three dominant
directions are computed for camera pose recovery. All lines
in an image are used to find the dominant direction and the
method fails if a dominant direction cannot be detected. On
the other hand, the method presented in [5] requires a
primitive configuration of two parallel lines and an
orthogonal line to estimate the relative camera pose in a
RANSAC framework. As a result the method is not suitable
for environment with few or no primitive configurations
among lines.
In literature, there are a few methods present that
combines line and points together for camera pose recovery
or structure from motion. Once such method is presented in
[6] that involves generating a set of linear equations from the
point and line correspondences in three views and solve for
the trifocal tensor. Camera poses for the second and third
views are then computed from the estimated trifocal tensor.
A linear solution followed by an iterative refinement was
presented. The method presented in this work is
foreshadowed by the aforementioned work and extends to a
robust and efficient relative camera pose estimation.
III. OUR APPROACH TO RELATIVE POSE
ESTIMATION
Let the three cameras are denoted as P1 , P2 , and P3 .
Assuming that the intrinsic parameter matrix K is known,
the first two cameras can be parameterized as
P1 [I | 0] and P2 [R 0 | t 0 ] where R 0 is the known
rotation matrix and t 0 is the known translation vector of the
second camera relative to the first camera. Then the camera
matrix of the third camera relative to the first camera is
given by P3 [R | t] .
A. Problem formulation from 2D lines correspondences in
Three Views
Fig. 1: Geometry of 3D line formation. The back projected planes
joining three camera centers and the three corresponding lines intersect in a
3D line.
Let L be a 3D line and its projection on three image
planes are l 1 , l 2 and l 3 . The back projected planes through
the three camera centers containing the 3D line L and
imaged as l 1 , l 2 and l 3 are given by 1 P1T l 1 , 2 P2T l 2
and 3 P3T l 3 . Together the three equation can be stacked
in a 4x3 matrix form as
M
l1
0
R T0 l 2
t T0 l 2
R Tl 3
t Tl 3
(1)
Since the three back projected planes intersect at the 3D
line L , there are two independent 3D points X 1 and X 2 on
the line L that satisfy the point-plane incident relations
X Ti j 0 , for i 1,2 and j 1,2,3 . As a result the matrix
M is a rank 2 matrix with 1D null space and we can consider
that the first column m 1 is a linear combination of the last
two columns m1 am2 bm3 .
three images and the points are transformed as xˆ1j
xˆ 2j
G i j x 2i and xˆ 3j
Fi j x1i ,
H i j x 3i , then the transformed trifocal
tensor is written as
Similar to [], we perform a two-step QR decomposition of
the matrix M with householder rotation to obtain a matrix of
the form
Tˆi jk
(F
0
0
0
0
0 f 2 (R, t)
(2)
f 1 (R, t)
Where f1 ( R, t ) and f 2 ( R, t ) are two affine functions of
the parameters R, t . The matrix M is also of rank 2 since the
rank of a matrix is preserved by elementary matrix
operation. From the above matrix it readily follows that
f i ( R, t )
0; i 1,2
(3)
B. Problem formulation from Points Correspondences in
Three Views
xˆ 13Tˆk13
Fig 2. The correspondence
x1
x2
x3
arises from the image of a
X in space
Let X be a 3D point in space and its projection on three
image planes are x 1 , x 2 , and x 3 respectively. The trilinearity constraint between the three views gives rise to four
independent equations involving the three corresponding
points and the three camera matrices [19, 20]. Using tensor
notation, the four equations can be written as
x1k ( x 12 x 13Tk33
x 13Tk13
x 12Tk31
Tk11 )
x1k ( x 12 x 32Tk33
x 32Tk13
x 12Tk32
Tk12 )
0
(5)
x1k ( x 22 x 13Tk33
x 13Tk23
x 22Tk31
Tk21 )
0
(6)
x1k ( x 22 x 32 Tk33
x 32 Tk23
x 22Tk32
Tk22 )
(4)
0
0
(7)
where Ti jk aij b4k a 4j bik is the tri-focal tensor, and
ai , bi are the columns of the camera matrices
P2 a 1 a 2 a 3 a 4 and P3 b 1 b 2 b 3 b 4 . For
better numerical stability, the points are normalized
following the methods mentioned in [19]. If
F, G, and H represents three coordinate transformations in
(8)
xˆ 12Tˆk31
Tˆk11 )
(9)
0
xˆ 1k ( xˆ 12 xˆ 32 Tˆk33
xˆ 32 Tˆk13
xˆ 12 Tˆk32
Tˆk12 )
0
(10)
xˆ1k ( xˆ 22 xˆ 13Tˆk33
xˆ 13Tˆk23
xˆ 22 Tˆk31
Tˆk21 )
0
(11)
xˆ1k ( xˆ 22 xˆ 32 Tˆk33
xˆ 32 Tˆk23
xˆ 22 Tˆk32
Tˆk22 )
0
(12)
Similar to the equations for line correspondences, we rearrange the above equations to generate four independent
equations of the form
f j ( R, t )
0; j
1,2 ,3,4
(13)
C. Liner Solution
Every corresponding triplet of points x1 , x 2 , x 3
contributes four linearly independent equations. Similarly
every corresponding triplet of lines l 1 , l 2 , l 3 contributes two
linearly independent equations. The equations can be
stacked to construct a system of the form
Av
3D point
) G sj H tk T rst
Then the new set of equations become
xˆ1k ( xˆ 12 xˆ 13Tˆk33
M
1 r
s
b
( 14 )
where v [r1T r2T r3T t T ] and r1 , r2 , r3 are the columns of
the rotation matrix R while t is the translation vector. A is a
n 12 matrix, generated by arranging the coefficients from
the equations in (3) and (13). If the number of point triplets
is denoted as n1 and the number of line triplets is denoted
as n2 , then the number of rows is n 4 * n1 2 * n2 . Since
v is a 12 element vector, we need at least 12 constraint
equations generated from the lines or points
correspondences. However, such a system suffers from in
the presence of noise and the solution can be arbitrarily far
from a rigid motion since orthonormality constraints are
ignored.
D. Solution as a Linear Combination of Singular Vectors
with Orthonormality Constraints
Once the matrix A is constructed, the solution of the linear
system lies in the null space or the kernel of A. In noise free
cases the rank of the matrix A is one (because of scale
ambiguity) when the number of independent linear equations
n 12 generated from at least three points correspondences
or six line correspondences or a combination of points and
lines (assuming points and lines are all independent).
However, the matrix A can be rank deficient because of
noise or nearly parallel motion (principal axes are parallel).
In addition the solution is also required to satisfy the
geometric constraints among the elements of v . Similar to
[..] we express the desired solution as a linear combination
of the last k singular vectors of the matrix A, denoted as v i
r1
|
r2
|
1
r3
|
.......
v1
|
k
vk
|
|
t
(15)
The problem is then reduced to determining the
coefficients 1 , 2 ,....., k of the above linear combination,
subject to orthonormality constraints:
r1
2
r1T r2
1,
0,
r2
2
r2T r3
1,
0
r3
r3T r1
r1
r2
2
0, and r2
2
r3
2
Case N 3 : We consider the three singular vectors
corresponding to 3 smallest eigenvalues and the solution is
written as
x
v3
(20)
1v 1
2v2
For ease of computation we use the 6 constraint equations
in (16) and (17) and construct a linear system of 6 equations
of the form
A
2
1
0
(16)
(17)
Substituting ( R, t ) from equation (15) in equation (16)
and (17) we get 6 polynomials of degree two in the k
variables 1 , 2 ,....., k . This system of polynomial
equations will have no solution in the general noisy case and
we need to resort to a principled “least squares” approach to
extract the solution. The degree of the system can be reduced
by replacing the unit norm constraint by the equal norm
constraints:
2
using Singular Value Decomposition (SVD) and chose the
signs of so that the points have positive z coordinates.
0
(18)
Combining equation (17) and (18) now we have 5
polynomial constraint equations of degree 2 in the
variables 1 ,..., k . Note that this approach applies even
when the combined system is under constraint for example
when the number of constraint equations are 8 or 10 from a
limited number of points and/or line correspondences.
Case N 1 : The eigenvector corresponding to the lowest
eigenvector is considered as the solution. The rotation matrix
is constructed from the first nine element of the eigenvector
and the scale is fixed by setting the determinant as 1.
Case N 2 : Two singular vectors corresponding to the
lowest two eigenvalues are considered and the solution can
be written as x
1v 1
2 v 2 where we need to solve for
coefficients
and . Each of the five constraint equations
consists of three monomials 12 , ,
and we solve for
them using a linearization technique that was developed in
cryptography and used in [21, 22]. The five constraint
equations are arranged in a linear system of five equations of
the form
A
(19)
Where
is the matrix constructed from the constraint
equations and
[ 12 , ,
] . We solve this system
(21)
b
Now each of the constraint equations consists of 6
monomials
[ 12 , , 32 ,
,
,
],
is a 6x6
square matrix constructed from the constraint equations and
b [0,0,0,1,1,1]T . We follow the same procedure as before to
solve for i .
Case N 4 : We now have four unknown coefficients
i , i 1,....4 and the solution is written as
x
1v 1
2v2
3v 3
(22)
4v4
In theory, the 5 constraint equations should still suffice.
However, the linearization procedure treats all monomials in
the constraint equations as unknowns and there are not
enough constraints any more. We solve this problem by
using a relinearization method used in [4] to solve for 4
unknown coefficients. Initially the degree of the system is
reduced by dropping the scale. This is done by dividing each
equations in (22) by 4 to have a system of the form
x
1v 1
2 v2
3v 3
(23)
v4
Now each of the 5 constraint equations consists of 10
monomials. Discarding the superscripts from the
coefficients i ,
the
monomials
are
[ 12 , 22 , 1 2 , 1 , 2 , 1 3 , 2 3 , 3 ,1] . In order to solve for a
single variable, the constraint equations are rearranged to get
a matrix system of the form
. . . [1] [1] [2]
. . . [1] [1] [2]
. . . [1] [1] [2]
. . . [1] [1] [2]
. . . [1] [1] [2]
2
1
2
2
1 2
1
2
1
0
0
0
(24)
0
0
Where . represents some scalar value and [N ] denotes a
polynomial of degree N in the variable 3 . Once the matrix
is constructed the next step is to solve for the null vector.
The components of the null vector is obtained (up to a scale)
by omitting the corresponding column of the 5x6 coefficient
matrix, then taking the determinant of the resulting matrix
and multiply by
as appropriate. For example, the
A B
th
i component of the null vector can be computed as
ui
( 1) i 1 det( Fˆ )
(25)
Where F̂ is a 5x5 matrix constructed by deleting the
i column of the above 5x6 coefficient matrix. As a result,
the components of the null vector u [ 12 , 22 , 1 2 , 1, 2 ,1]
introduces a new set of polynomial equations of degree 4
( 12 , 22 , 1 2 ), degree 3( 1 , 2 ) and degree 2(1) in the
single variable 3 . The components of the null vector are not
independent and they satisfy three constraint equations
r
t
0
(27)
where r is the vector constructed from the elements of
the rotation matrix and t is the translation vector
th
(
1)
(
2)
(
(
1)
(
1)
(
(
2)
(
2)
(
1 2 ) (1)
2
1 ) (1)
2
2 ) (1)
(26)
A set of three polynomial equations of degree 6 in the
single variable
3 are constructed by substituting the
expressions for the components of the null vector from
equation (19) in the three constraint equations in (20).
Solving the nonlinear equation system (20) by linearization
technique might lead to an inconsistent result from
redundant equations and we approach it as an unconstrained
optimization problem. Denoting the three univariate
polynomials in equation (20) as f i ( 3 ); i 1,2,3 a cost
function F can be defined as a square sum of the three
3
polynomials as F
identified
f
i 1 i
by
derivative F
2
(
finding
3
f (
i 1 i
3)
. The minima of F can be
the
3) fi (
3)
roots
of
its
1st
0 . F is a 11th order
polynomial which can be easily solved by the eigenvalue
methods.
There can be up to 11 real roots of F ( 3 ) , all of which
need to be tested as a candidate solution. Once the candidate
for
3 are obtained, the corresponding candidates for
1,
2 are
obtained by substituting values of
3
in the
component expressions of u .
Once the coefficients i are computed with the method
described above, the rotation can now be obtained by
substituting i ; i 1....k in equation (14). Then the scale is
fixed by setting the determinant of the rotation matrix to 1.
Theoretically, the translation can also be found from
equation (14). However, experiments with real data
generated large translation error. Hence the translation is
estimated separately in the following method once the
rotation is estimated.
E. Translation Computation
In order to compute translation, we express the translation
vector in terms of the rotation parameters in a least-square
sense. We re-organize linear system of (23) in the form
We rearrange equation (24) in the form Ar
Bt and
then the least-square solution to the translation vector t can
be expressed as
t
( B T B) 1 B T Ar
(28)
F. Efficient Computation
In order to improve efficiency as well as to ensure
numerical stability, computations of critical components of
the algorithm are carefully designed. This way we ensure
that the number of arithmetic operations does not become
too large that might adversely affect the numerical behavior.
We pre-compute the symbolic interaction between
coefficients in the equation and then evaluate for values of
the variables.
As an example, the coefficients of the matrix A in
equation (14) are symbolically pre-computed and evaluated
with the set of points and lines correspondences. Similar
steps are taken for constructing systems of polynomials and
solve them to estimate rotation and translation. The symbolic
calculations are done using Matlab symbolic toolbox and the
scripts are available at request. For brevity, details of the
equations are skipped.
G. Correct Pose Selection
In order to find the correct camera, all the possible
camera pose candidates obtained by the method in previous
section need to be evaluated. We use a simple method to do
this. For the corresponding points in first and second view,
the 3D point is determined by back projecting the rays in
first and second view. The 3D point is then projected onto
the third view. A prescreening is performed by discarding
the camera poses where majority of the 3D points have
negative z coordinate in the 3rd camera before computing reprojection error. Similarly, lines back projected from the
first and second image intersects at a 3D line. The 3D line is
then projected onto the third camera and the line reprojection error is computed. The pose with the lowest reprojection error is selected as the final estimated pose.
IV.
EXPERIMENTAL EVALUATION
The proposed algorithm is evaluated for accuracy,
robustness and efficiency with both synthetic and real data.
We demonstrate the effectiveness of the proposed algorithm
by comparing results with state of the art line based pose
estimation algorithms.
A. Synthetic Experiments
Monte Carlo simulation with synthetic lines and points
were performed to quantify the performance of the proposed
algorithm. We evaluate the algorithm’s performance under
various noise level, different lines and points’ combination
as well as large and small camera motion. At each trial, a
predefined number of 3D line segments and 3D points are
generated by randomly placing segment end points as well
as the 3D points inside a cube (10 m length) that is centered
at the origin of the world coordinate system {W } . We
consider a virtual calibrated camera with image size
640x480 and focal length 800 pixels. The first view is
generated by placing the camera randomly at a distance of
20 meters from the origin. For large motion, the second and
third camera views are also generated randomly similar to
the first view. For small motion, the second and third camera
are generated by randomly selected camera positions that are
within 5 meters of the first camera position. The cameras
were then oriented so that they looked at the origin having
all the 3D points and the line segments in the field of views.
translation error as a function of noise level. The difference
between mean and median error mainly comes from the pose
estimation with large errors. The output of the algorithm is
shown in fig. 4. The top row shows the mean rotation and
translation error whereas the bottom rows shows the median
errors. The experiments are done with 50% points and 50%
lines where the numbers are varied from 4 to 8. It can be
seen that the algorithm’s performance is very stable when
number of points and lines used are 5 or above. Again we
consider large camera motion similar to experiment 1.
The 3D points and line segments are projected on each
view to generate the 2d points and line segments. Zero mean,
Gaussian noise of varying standard deviations is added to the
coordinates of the 2D points and line segments in each of the
three views. All the experiments are based on 500 trials.
In the first experiment, we evaluate the effectiveness of
the algorithm when used in a hypothesize-and-test
framework such as RANSAC. It is important to select a
small set of features for hypothesis generation in a RANSAC
scheme, we used a combination of 3 or 4 lines and 3 or 4
points in this experiment. Similar to [1, 4], the lower quartile
of the error distributions are used since finding a fraction of
good hypothesis is more important than to get consistent
results. Performance of the algorithm is shown in Fig. 3. The
left image shows the camera pose as well as the lines and
points. The relative distance between cameras could be as
much as 40 meters. It can be seen that rotational and
translational error increases with noise level. However with
a moderate noise level, all combination of lines and points
generate reasonable pose estimation. It can be observed that
the pose estimation is more accurate when the number of
points used is 4. This is explained by the fact that the
number of constraint equations is larger in case of points
than lines as well as how noise is added to the coordinates of
the points and lines. For each line, random noise is added to
both end points of the 2D lines resulting in a larger noise
effect compared to 2D points. However, in moderate noise,
the estimated camera poses are sufficiently accurate for
further optimization.
a. Experiment setup
b. Rotation Error
c. Translation Error
Fig. 3. Rotation and translation error with large camera motion as a
function of noise level. Experiments are conducted with lines n L =3 or 4
and points n p =3 or 4.
In the second experiment, the robustness of the algorithm
is evaluated by computing mean and median rotation and
Fig 4. Mean and Median rotation and translation error as a function of noise
level. Number of lines and points used ranges between 4 and 8. In each
experiment, 50% lines and 50% points are used.
Efficiency of the method was evaluated on a laptop
computer with an Intel Core i5-4340M 2.9 GHz CPU. Figure
5 shows the computational time requirement in millisecond
vs the number of features. Number of lines and points are
varied between 4 and 45. For 4 lines and 4 points the
average runtime is 37 milliseconds and for 5 lines and 5
points, average runtime is 41 milliseconds. Results show that
a combination of 4 points and 4 lines or 5points and 5 lines
are suitable for use in RANSAC framework while being
very stable.
Fig. 5: Computational time requirement vs number of lines and points.
All experiments are conducted with 50% lines and 50% points, averaged
over 500 runs. The computational cost grows almost linearly with
number of features used.
B. Real Images
In order to evaluate the algorithm’s performance in real
world situations with various magnitude and direction of
motions, we apply the algorithm on the VGG Multiview
dataset.[22]. The dataset contains indoor and outdoor image
sequences with extracted 2D lines and points, their
reconstructed 3D positions and camera projection matrices.
The Corridor sequence contains indoor images with small
forward camera motion whereas the other image sequences
are outdoor images of buildings with various camera
motions. In order to evaluate the accuracy and robustness of
our algorithm we consider three images from each sequence
and consider known camera poses for the first two views.
Then we apply the algorithm to estimate the camera pose of
the third view relative to the first view. The error is
represented as the ratio of the estimated relative pose and
ground truth relative pose between first and third view.
Fig. 6 shows the performance of the algorithm in boxplot
representation, where each column depicts the distribution of
errors for 100 runs of the algorithm for a number of points
and lines. From left to right, the number of 2D lines and
points combination considered are 2 lines + 2 points, 2
lines+ 3 points, 2 lines+4 points, 3 lines+3 points, 4 lines + 2
points, 4 lines + 4 points, 6 lines + 6 points, and 10 lines +
10 points. At each run, a predefined set of corresponding 2D
lines and points in three views are randomly selected from
the dataset. From the VGG Multiview dataset, the image
sequences considered are UL= University Library, MC-I =
Merton-College-I,
MC-II=Merton-College-II,
MCIII=Merton-College-III, WC=Wadham-College, C=Corridor.
UL
W
C
C
Fig. 6. Rotation and translation error as a function of
number of lines and points. The boxplot result is
generated by running the algorithm 100 times with
randomly selecting n lines and points. From left to right,
the combinations of lines and points used are 2 lines + 2
points, 2 lines+ 3 points, 2 lines+4 points, 3 lines+3
points, 4 lines + 2 points, 4 lines + 4 points, 6 lines + 6
points, and 10 lines + 10 points.
It can be seen that the algorithm is very robust when using
a combination of 3 lines + 3 points or above for both small
and large camera motions. Figure 7 shows example results
from the VGG dataset. In each row, the left and middle
images are overlaid with the corresponding points and lines
used for pose estimation of the third view relative to the first
view. A set of 3D lines are then projected onto the third view
using the estimated camera pose.
U
L
M
C-I
M
C
I
M
CII
M
CIII
M
C
I
I
M
C
I
I
I
W
C
[11] M. Pollefeys, D. Nistér, J. M. Frahm, A. Akbarzadeh, P. Mordohai, B.
C
O
Clipp, C. Engels, D. Gallup, S. J. Kim, P. Merrell, C. Salmi, “
Detailed real-time urban 3d reconstruction from video”, International
Journal of Computer Vision, Vol. 78(2-3), pp:143-67,2008
[12] Y. Furukawa, B. Curless, S. M. Seitz, R. Szeliski, ”Manhattan-world
Fig. 7: Example results from the VGG dataset. The left and the middle
images are overlaid with two corresponding lines and four corresponding
points that are used for relative pose estimation in the third view. The
right image shows re-projection of a subset of 3D line segments from the
dataset using estimated camera poses. Camera poses are estimated using
4 lines and 2 points and colors are chosen randomly for both points and
lines in each image.
V. CONCLUSION AND DISCUSSION
[13] B. Li, L. Heng, G. H. Lee, M. Pollefeys, “A 4-point algorithm for
relative pose estimation of a calibrated camera with a known relative
rotation angle” Intelligent Robots and Systems (IROS), 2013
IEEE/RSJ International Conference on, pp. 1595-1601, Nov 2013
[14] C. J. Taylor, D. J. Kriegman, “Structure and Motion from Line
Segments in Multiple Images” IEEE Transactions on Pattern Analysis
and Machine Intelligence, vol. 17(11), pp.1021-1032, 1995
[15] G. Schindler, P. Krishnamurthy, F. Dellaert, “Line-based structure
In this work, we have presented the development of a
unified relative camera pose estimation method that
combines points and lines. Given a set of point and line
correspondences in three views, the third camera view is
estimated relative to the first view when the first two views
are known. One of our primary contribution is a robust and
efficient algorithm that is suitable for both small and large
camera motions. We have demonstrated that the system,
though not fully optimized, perform robustly and suitable for
use in a RANSAC framework in both indoor and outdoor
environment.
While no nonlinear optimization was performed in any of
the experiments, it will conceivably improve the pose
estimation. Future work include applying the algorithm in
real time SFM pipeline and simultaneous localization and
mapping frameworks.
from motion for urban environments”, 3D Data Processing,
Visualization, and Transmission, Third International Symposium
on, pp. 846-853, June 2006
[16] A.
Bartoli, P. Sturm, “Structure-from-motion using lines:
Representation, triangulation, and bundle adjustment”, Computer
vision and image understanding, vol. 100(3), pp:416-41, Dec, 2005
[17] T. Werner, A. Zisserman “New techniques for automated architectural
reconstruction from photographs”, Computer Vision—ECCV , pp:8089, 2002
[18] J. Košecká, W. Zhang, “Extraction, matching, and pose recovery
based on dominant rectangular structures”, Computer Vision and
Image Understanding. vol. 100(3), pp::274-93, Dec 2005
[19] R. Hartley, A. Zisserman. “Multiple view geometry in computer
vision”. Cambridge university press; 2003.
[20] A. Shashua. “Projective structure from uncalibrated images: Structure
from motion and recognition”, IEEE Transactions on Pattern Analysis
and Machine Intelligence. vol. 16(8), pp:778-90, Aug, 1994
[21] V. Lepetit, F. Moreno-Noguer, P. Fua, “Epnp: An accurate o (n)
solution to the pnp problem” International journal of computer vision.
vol. 81(2), pp:155-66, Feb, 2009
[22] A. Ansar and K. Daniilidis. “Linear pose estimation from points or
REFERENCES
[1] D. Nistér, “An efficient solution to the five-point relative pose
problem.” IEEE transactions on pattern analysis and machine
intelligence, vol. 26(6), pp.756-770, 2004
[2] D. Nistér, and H. Stewénius, “A minimal solution to the generalised 3point pose problem.” Journal of Mathematical Imaging
Vision, vol. 27(1), pp.67-79, Jan, 2007
stereo” Computer Vision and Pattern Recognition, 2009. IEEE
Conference on, pp. 1422-1429, Jun 2009
and
[3] H. Li and R. Hartley, “Five-point motion estimation made easy,”
Pattern Recognition, 2006. ICPR 2006. 18th International Conference
on, vol. 1. IEEE, pp. 630–633, 2006
[4] M. Chandraker, J. Lim, D. Kriegman, “Moving in stereo: Efficient
structure and motion using lines”, Computer Vision, 2009 IEEE 12th
International Conference on, pp. 1741-1748. Sep, 2009
[5] A. Elqursh, A. Elgammal, “Line-based relative pose estimation.”
Computer Vision and Pattern Recognition (CVPR), 2011 IEEE
Conference on, pp. 3049-3056, June, 2011
[6] R. I. Hartley, “Lines and points in three views and the trifocal tensor”.
International Journal of Computer Vision. Vol. 22(2), pp:125-40,
Mar, 1997
[7] A. Shashua, L. Wolf, “On the structure and properties of the
quadrifocal tensor”, Computer Vision-ECCV pp.710-724, 2000
[8] D. Scaramuzza, F. Fraundorfer, R. Siegwart, “Real-time monocular
visual odometry for on-road vehicles with 1-point ransac.” In Robotics
and Automation, 2009. ICRA'09. IEEE International Conference on
pp. 4293-4299, May, 2009
[9] D. Nistér, O. Naroditsky, J. Bergen, “Visual odometry”, CVPR, Vol.
1, pp. 652-659, 2004
[10] A. Agrawal, K. Konolige, “Real-time localization in outdoor
environments using stereo vision and inexpensive gps”, Pattern
Recognition, 2006. ICPR 2006. 18th International Conference on,
Vol. 3, pp. 1063-1068, Aug 2006
lines”,
IEEE Transactions on Pattern Analysis and Machine
Intelligence, vol. 25(5), pp:578–589, 2003
| 1cs.CV
|
Control of Large Swarms via Random Finite Set Theory
arXiv:1801.07314v3 [cs.SY] 27 Feb 2018
Bryce Doerr1 and Richard Linares2
Abstract— Controlling large swarms of robotic agents has
many challenges including, but not limited to, computational
complexity due to the number of agents, uncertainty in the
functionality of each agent in the swarm, and uncertainty in the
swarm’s configuration. This work generalizes the swarm state
using Random Finite Set (RFS) theory and solves the control
problem using model predictive control which naturally handles
the challenges. This work uses information divergence to define
the distance between swarm RFS and a desired distribution.
A stochastic optimal control problem is formulated using a
modified L22 distance. Simulation results show that swarm
densities converge to a target destination, and the RFS control
formulation can vary in the number of target destinations.
I. INTRODUCTION
Control of large robotic groups or swarms is currently
an area of great interest in controls research. These robotic
groups are usually tiny robots with limited actuators that do
specific tasks in the formation. For example, the group can
use its combined effort to grasp or move in the environment
[1]. Specifically, in NASA applications, swarm control of
space systems including satellites and rovers can be used
for exploration of asteroids and other celestial bodies of
interest [2]. In military applications, swarms of UAVs can
be used for border patrol, search and rescue, surveillance,
communications relaying, and mapping of hostile territory
[3]. In reviewing these applications, swarm control of large
groups of swarming agents are required.
Changing how the swarm is modeled can alleviate computational complexity for different control methods. Previously,
the swarm/potential model using the random finite set (RFS)
formalism was used to describe the temporal evolution of
the probabilistic description of the robot swarm to promote
decentralized coordination [6]. By using a measure-value
recursion of the RFS formalism for the agents of the swarm,
the dynamics of the swarm in time can be determined with
computational efficiency.
Several control techniques have been implemented on
swarms previously. In centralized control, one agent in the
swarm is the designated controller and manages the execution of other agents in the swarm [7]. Thus, it oversees the
other agents’ system processes. This type of control is the
easiest to implement on a swarm of robots, and it solves
the problem of controlling the swarm to a desired state.
Unfortunately, this type of control suffers from two main
1 Bryce Doerr, Graduate Student, Department of Aerospace Engineering
and Mechanics, University of Minnesota, Minneapolis, MN 55455, USA
[email protected]
2 Richard Linares, Assistant Professor, Department of Aerospace Engineering and Mechanics, University of Minnesota, Minneapolis, MN 55455,
USA [email protected]
problems. As the number of agents of the swarm increases,
the computational workload becomes more difficult. This
is especially difficult when the swarm robots are low cost
robotic agents. Another issue that arises is that centralized
control is not robust against failures of other agents within
the swarm [8]. Since swarm robots on a level of a thousand
agents are low cost units, communication, actuation, and
sensing are performed with less reliability. Thus, centralized
control may not be a viable option for these systems.
Probabilistic swarm guidance has also been used to enable
the swarm to converge to target distributions [9]. Probabilistic swarm guidance controls a swarm density distribution
through distributed control, so that each agent determines
its own trajectory while the swarm converges to the target
distribution. Distributed control is defined as reformulating a
control problem as a set of interdependent subproblems and
solving these subproblems [5]. Probabilistic swarm guidance
solves issues that involve controlling an increasing number
of individual agents which is computationally complex by
controlling the swarm density distribution of the agents.
Using optimal transport, convergence is achieved faster than
a homogeneous Markov chain approach and the cost function
is minimized [9], [10], [11]. Another method is the use of the
inhomogeneous Markov chain [12], [13]. In this approach,
the agents move in a decentralized fashion which is like
the homogeneous Markov chain, but the algorithm allows
communication with other agents to reach and settle at the
target destination. These algorithms are robust to external
disturbances since the algorithms themselves use the swarm
density distribution.
Velocity field generation for swarm control is a nonoptimal decentralized control method for swarms that synthesizes smooth velocity fields as a function of time and position
[14]. With a designated target distribution, the swarm follows
the velocity field using the heat equation to convergence by
using local agent position information to estimate its local
density. The advantage of this method is that the agents
facilitate collision avoidance and move in a smoother manner
than previous Markov chain approaches [10]. Unfortunately,
the use of the heat equation diffuses the agents in a locally
uniform manner to the target density. Therefore, this is a
non-optimal method of controlling the swarm to a target
distribution.
Model predictive control (MPC) has been heavily studied
for nonlinear systems and for applications including spacecraft maneuvering and attitude control [15], [16], [17]. Decentralized MPC has also been studied for thousands of low
cost spacecraft with limited capabilities [18]. MPC computes
the control input by solving a finite horizon problem. This
allows it to optimize the current time step while taking the
future time steps into account.
One last decentralized approach to controlling swarms
is using sequential convex programming [19]. Sequential
convex programming uses multiple iterations to maintain
accuracy of the convex approximations of the constraints
which create more efficient trajectories. Uncertainties in the
trajectories are accounted for when the algorithm is tuned.
Thus, this provides the robustness for the swarm while
pushing the agents to the designated targets while using
this algorithm in combination with MPC in real time. Morgan also used sequential convex programming to do target
assignment (mapping of agents to targets) and trajectory
generation for varying swarm sizes through time [20]. This
method is viable for swarms, but through RFS control, target
assignment is not necessary.
II. R ANDOM F INITE S ET C ONTROL P ROBLEM
F ORMULATION
(1)
The control
sequence is also defined by a RFS in the form
Ut = u1,t , u2,t , · · · , uNtotal(t) ,t and a RFS probability
density given by
p(U = {u1 , u2 , · · · , un })
= p(|U |= n)p({u1 , u2 , · · · , un } | |U | = n),
where b(x), ps (ζ), and f (x|ζ) are the agents’ birth, survival
probability, and target motion models respectively and ζ is
the previous state [22]. The bar on v̄t (x) denotes that the
PHD is conditioned on measurements Zt−1 . For the time
update, the equation is given by
vt (x) = (1 − pd (x))v̄t (x)
X
pd (x)g(zt |x)v̄t (x)
R
,
+
c(z)
+
pd (ζ)g(zt |ζ)v̄t (ζ)dζ
z ∈Z
(4)
t
The control problem for swarming agents is set up by
using RFS theory [21], [22]. This theory addresses the
decentralized formulation for each agent in the formation.
Each agent has the challenge of estimating its local formation
configuration and designing a decentralized control policy
to achieve some local configuration. It is assumed that
each agent within the swarm is identical and using unique
identifiers on each agent is unnecessary. Using RFS theory,
the number of agents and their states is determined from
measurements. The number of agents in the field is denoted
by Ntotal (t) and
may be varying as agents enter and leave the
field. Xt = x1,t , x2,t , · · · , xNtotal(t) ,t denotes a realization
of the RFS distribution for agents. At any time, t, the RFS
probability density function can be written as
p(X = {x1 , x2 , · · · , xn })
= p(|X|= n)p({x1 , x2 , · · · , xn } | |X| = n).
the cardinality of the agent set that is Poisson distributed
[22]. It is noted that the assumptions made by the PHD filter
are strong assumption for swarming robotics. However, this
is a good starting point for initial proof-of-concept study.
The PHD recursion for a general intensity function, vt (x),
is given by
Z
v̄t (x) = b(x) + ps (ζ)f (x|ζ)v(ζ)dζ,
(3)
(2)
since the number of the agents in the field to be controlled
is also varying. The Bayesian multi-target state estimation
problem generally has no closed-form solution, but when
the form p(|X|= n) and p({x1 , x2 , ..., xn } | |X| = n) are
assumed, approximate solutions can be determined by the
Probability Hypothesis Density (PHD) filter. Using the PHD
filter assumption, the RFS is represented by its first moment (intensity function ν(x, t)) [21]. ν(x, t) represents the
probability of finding an agent in a region of state space.
The PHD is defined over the single agent’s state space and
has the property that the expected number of agents over a
region is simply the integral of the PHD over that region.
Propagation of the PHD can be determined if the agents are
assumed to be independent and identically distributed with
where pd (x), g(zt |x), and c(z) are the probability of detection, likelihood function, and clutter model of the sensor
respectively [22]. By using this recursion, the swarm probabilistic description can be updated. Unfortunately, Eqs. (3)
and (4) do not contain a closed-form solution and the numerical integration suffers from higher computational time as the
states increase. Fortunately, a closed-form solution exists if
it is assumed that the survival and detection probabilities are
state independent, and the intensities of the birth and spawn
RFSs are Gaussian mixtures of the form
ν(x, t) = f (x) =
ν̄(x, t) = g(x) =
Nf
X
(5)
i=1
(i)
wf N x; mif , Pfi ,
Ng
X
wg(i) N x; mig , Pgi ,
(6)
i=1
where w are the weights and N x; mi , P i is the probability
density function of a multivariate Gaussian distribution with
a mean and covariance respectively. Note that closed form
solutions using Gaussian mixtures exist for cases without
PNf (i)
the state independent assumption. Additionally, i=1
wf =
PNg (i)
= N̄total (t) where N̄total (t) is the
Ntotal (t) and
i=1 wg
desired number of agents. The intensity function ν(x, t) is
in terms of the swarm state while ν̄(x, t) is in terms of the
desired state. The swarm intensity function can be propagated
through updates on the mean and covariance as given by
mif,k+1 = Ak mif,k + Bk uif,k ,
(7)
i
i
Pf,k+1
= Ak Pf,k
ATk + Qk ,
(8)
where Qk is process noise. Then given the Gaussian mixture
intensities assumption, a control variable is calculated for
each component uif,k . Additionally, each Gaussian mixture
component can represent many agents since the intensity
function integrates to the total number of agents. Note
that although linear dynamics is used, the RFS-based cost
function is nonquadratic.
Each individual swarm agent will run a local PHD observer to estimate the state of the swarm by modeling
PSfrag replacements
the swarm as a distribution. Thus, using RFS PSfrag
theory, replacements
it is
assumed that the individual swarm agents form an intensity
function that is a Gaussian mixture density in which the
y
means and covariances of the Gaussian mixture are propx
agated and controlled. An optimal control problem is set
(a) Cauchy-Schwarz Divergence
up that tracks a desired swarm formation by minimizing its
control effort in the following objective function
Z T
J(u) =
E {u⊺ Ru} + D(ν(x, t), ν̄(x, t))dt, (9) PSfrag replacements
2.5
400
2
200
1.5
0
1
-200
0.5
60
-400
60
50
40
50
40
40
40
30
20
0
30
20
20
20
y
10
0
0
10
0
x
(b) L22 Distance
1500
1000
500
0
-500
0
where ν̄(x, t) is the desired formation and u is the control
effort for the Gaussian mixture swarm. Both ν(x, t) and
ν̄(x, t) are defined over the complete state space which
include position and velocity parameters. D(ν(x, t), ν̄(x, t))
is the distance between these Gaussian mixtures which has
several closed-form solutions.
III. D ISTRIBUTIONAL D ISTANCE BASED -C OST
A. Cauchy-Schwarz Divergence
The Cauchy-Schwarz divergence is based on the CauchySchwarz inequality for inner products, and it is defined for
two random vectors with probability densities f and g given
by
hf, gi
,
(10)
DCS (f, g) = ln
kf kkgk
where h·, ·i is the L22 inner product over the densities. The
argument of the logarithm is non-negative since probability
densities are non-negative, and it does not exceed one
by the Cauchy-Schwarz inequality. The Cauchy-Schwarz
divergence can be interpreted as an approximation to the
Kullback-Leibler divergence but has a closed-form expression for Gaussian mixtures [24]. This is useful for calculating
the distance between two-point processes represented by intensity functions. The Cauchy-Schwarz divergence between
two Poisson point processes with Gaussian mixture intensities is shown in Eq. (10) when Eqs. (5) and (6) are substituted
inside. Note that in the control formulation used, only ν(x, t)
is assumed to depend on the control u. Therefore, the term
that depends only on ν̄(x, t) is omitted from the objective
function since ν̄(x, t) does not depend on u.
Figure 1(a) shows the surface plot using the CauchySchwarz divergence for 4 densities in the swarm at an initial
time instance. Each initial state has hills while the destination
states have valleys. Thus, each density will repel each other
while moving towards the final state through time. If the
initial state is too large compared to the destination states,
it will take longer for the 4 densities to converge to the
destination values. Also, the repelling effect due to the hills
are relatively small. Thus, the Cauchy-Schwarz divergence
may not be the fastest converging solution for the objective
function minimization.
B. L22 Distance
Alternatively, the distance between two Poisson point processes with Gaussian mixture intensities can be determined
60
40
20
y
0
0
10
x
20
30
40
50
(c) L22 + Quadratic Term
Fig. 1: Surface Plots of the RFS divergences.
by using the L22 distance between the probability densities.
The L22 is given by
Z
2
DL22 (f, g) = (f − g) dx.
(11)
Figure 1(b) shows surface plot using the L22 distance for
a 4-density swarm at an initial time instance. Each initial
state has more defined hills compared to the Cauchy-Schwarz
divergence. Thus, the initial states have a stronger repel
effect from each other. Also, the destination states have
large valleys that create a large attraction effect for each
initial state to move to. Thus, the optimization solution will
be faster in the L22 distance case. Unfortunately, the L22
distance suffers from a similar issue to the Cauchy-Schwarz
divergence. If the initial conditions increase farther away
from the destination states, the optimization may take much
longer due to a flat surface away from the destination states.
C. L22 Distance with Quadratic Term
The issue of convergence remains for the L22 distance when
the initial states are far larger than the destination states. To
achieve faster convergence, an additional term is added to the
L22 distance to shape the gradient descent through a quadratic
term as given by
DL22 mod (f, g) = DL22 (f, g)
Ng Nf
X
X
(12)
(i)
wg(j) wf N (mjg ; mif , Pgi + Pfj ).
− ln
j=1 i=1
Note that this term is referred as quadratic, although it
may be more appropriate to call it quadratic-like. Figure
1(c) shows the surface plot using the modified L22 distance
for a 4-density swarm in Eq. (12). Compared to the L22
distance, the initial and destination states provide the hills
and valleys necessary to obtain convergence. However, as
the initial states move outwards, the surface map decreases
in a quadratic fashion instead of staying flat. Thus, the
optimization can occur at any point to reach convergence
in a timely manner. To solve the minimization problem for
the objective function, traditional LQR based solutions are
not applicable since the terms inside the objective functions
are nonquadratic [23]. The minimization of the objective
function in discrete time is
min
uk ,k =1,...,T
+
J(u) =
T
X
E{u⊺k Ruk }
k=1
Nf Nf
X
X
j
i
wf,k wf,k N (mjf,k ; mif,k , Pf,k
+ Pf,k
)
Ng Ng
X
X
j
i
wg,k wg,k N (mjg,k ; mig,k , Pg,k
+ Pg,k
)
(j)
(i)
(j)
(i)
j =1 i =1
+
j =1 i =1
−2
Ng Nf
X
X
(j)
V. R ESULTS
(i)
j
i
wg,k wf,k N (mjg,k ; mif,k , Pg,k
+ Pf,k
)
j =1 i =1
− ln
which measures the distance of the agents to the means of
the densities. As the swarm evolved, this distance determined
which units belong to a given component. Agents were
controlled according to their placement in each density
through an open loop method using the MPC control input
obtained for each density. Although an open loop method
was used, feedback control can be used if the state estimates
are determined from the PHD filter.
The great advantage of using MPC is that the algorithm
can handle nonlinearities in the objective function. Unfortunately, numerically solving an optimization at each timestep is not particularly fast, so it can be difficult to run
optimization problems in real-time to control swarm robots
towards a target state.
Ng Nf
X
X
j=1 i=1
(j) (i)
j
i
wg,k wf,k N (mjg,k ; mif,k , Pg,k
+ Pf,k
) ,
(13)
Subject to : mif,k+1 = Ak mif,k + Bk uif,k ,
(14)
N
where uk = [(u1f,k )T , · · · , (uf,kf )⊺ ]⊺ is the collection of all
control variables.
IV. R ECEDING H ORIZON DYNAMIC P ROGRAMMING
In the simulation experiment, a 4-density swarm on a 2D plane was initialized in a square grid where the densities
1, 2, 3, and 4 were defined counterclockwise starting on
the first quadrant. With the 4-density swarm, four different
test cases were implemented to bring the densities to the
target trajectories and to test the control theory involved
from the control formulation. The first test case shows the
limitations of using only the L22 distance with varying initial
conditions in a square grid and four target destinations
located at (±1,±1). The last three cases used the L22 distance
with a quadratic term and varying target destinations. In
Case 2, four target destinations are located at (±1,±1).
For Case 3, three target destinations are located at (±1,1)
and (-1,-1). Lastly in Case 4, five target destinations are
located at (±1,±1) and (0,0). The receding horizon dynamic
programming was used to determine the control effort and
trajectories for the simulation, but other methods including
iterative LQR can be used too [23].
m2,y
m3,y
m4,y
m1,y
m2,x
m3,x
m4,x
m1,x
An optimal solution based on minimizing the objective
uses MPC or receding horizon control. This uses dynamic
programming techniques to obtain an optimal control solution u from the objective function and has been implemented
on swarm control previously [20].
Conceptually, at a time t, the knowledge of the system
model is used to derive a sequence u(t|t), u(t + 1|t), u(t +
4
4
2|t), · · · , u(t + Tp |t) where Tp is the finite horizon from
2
2
the current state x̄(k) [26]. With the input sequence, the
0
0
0
0.1
0.2
0.3
0.4
0
0.1
0.2
0.3
0.4
state is moved forward in time by the control horizon,
Tc ;
PSfrag replacements
Time (s)
Time (s)
0
4
usually one time-step. Then the same strategy is repeated
-2
2
-4
0
for time t + 1. Tp can be chosen to be either small or large.
0
0.1
0.2
0.3
0.4
0
0.1
0.2
0.3
0.4
As Tp increases, the degrees of freedom in the optimization
Time (s)
Time (s)
0
0
increases which can slow down the algorithm considerably
-2
-2
-4
-4
even though more of the future reference trajectory would
0
0.1
0.2
0.3
0.4
0
0.1
0.2
0.3
0.4
Time (s)
Time (s)
be useful to bring the output closer to the reference. With
4
0
2
-2
a smaller Tp , the computation time will be faster, but the
0
-4
0
0.1
0.2
0.3
0.4
0
0.1
0.2
0.3
0.4
optimization may be more suboptimal. Thus, the swarm may
Time (s)
Time (s)
not converge to the desired configuration.
For the RFS control formulation, a u that controls the
swarm densities through their statistics (mean and covari- Fig. 2: The position responses of density statistics (means)
ance) was found by minimizing the objective as given by are shown for Case 1 and Case 2.2 Black-dashed and redEq. (13) and (14). This was done using MATLAB’s fminunc dotted lines are the responses for L2 distance with (±3,±3))
The bluesolver and the Quasi-Newton algorithm [27]. Then, the and (±1.5,±1.5)) initial conditions respectively.
2
solid
lines
are
the
responses
for
the
L
+Quadratic
term with
2
agents in the swarm were initialized to the closest density
initial
conditions
of
(±3,±3)).
The
green-dotted
lines
are the
using the Mahalanobis distance given by
target
destinations
for
both
cases.
q
DM (x, mif,k ) = (x − mif,k )⊺ P −1 (x − mif,k ), (15)
A. Case 1: L22
5
time = 0.00 s
5
time = 0.05 s
0.08
y
y
y
y
y
y
y
y
For Case 1, four swarm densities are controlled to move
0.07
0
0
towards the target destinations at initial conditions farther
0.06
away (square grid at (±3,±3)) and closer to (square grid
-5
-5
0.05
at (±1.5,±1.5)) the target destinations as shown by mean
-5
0
5
-5
0
5
x
x
PSfrag
replacements
responses given by the black-dashed and red-dotted
lines
in
0.04
time = 0.10 s
time = 0.40 s
5
5
Figure 2 respectively. From the trajectory snapshots given
0.03
by Figure 3(a), initial conditions that are far from the target
0
0
destinations do not have a converging control solution. From
0.02
the surface visualization in Figure 1(b), the general plane
0.01
-5
-5
is flat in areas away from the target destinations and states
-5
0
5
-5
0
5
x
x
of the density. Therefore, optimization using MPC is more
difficult in these flat areas and may not converge to a solution.
Fig. 4: Case 2: Trajectories of the swarm with individual
If the densities are initialized much closer to the target
agents (red dots) and four target destinations (black x’s).
destinations as shown in Figure 3(b), the flatness in the
general plane is minimal, and the optimization step in MPC
converges to a solution. By using the L22 distance, converging C. Case 3: L2 +Quadratic Term, Three Target Destinations
2
control solutions can only be found for initial conditions and
Case
3
illustrates
the effect of three target destinations
target destinations that are close.
on the final trajectories of the four swarm densities. From
Figure 5, it can be visually shown where the swarm densities
time = 0.00 s
time = 0.40 s
5
5
are located relative to the target destinations at each time
0.08
step. The trajectories for density 1 and density 3 reach their
0.07
0
0
target, but densities 2 and 4 reach the third target with
0.06
approximately 0.30 of steady-state error. The results obtained
-5
-5
follow directly from the RFS control theory using the L22 plus
Sfrag replacements
0.05
-5
0
5
-5
0
5
x (a) Initial conditions of (±3,±3) x
quadratic distance term. By using this L22 with a quadratic
0.04
time = 0.00 s
time = 0.40 s
5
5
term in the objective function, the individual densities will
0.03
attract towards the target destinations while repulsing away
0
0
from each other. This can be shown in the surface map shown
0.02
in Figure 1(c) where the hills are areas of repulsion and
0.01
-5
-5
valleys are areas of attraction. Thus, for Case 3, densities 2
-5
0
5
-5
0
5
time = 0.10 s
x (b) Initial conditions of (±1.5,±1.5) x
and 4 are attracted to the same target, but they stay away
from each other which causes both to share the same target
Fig. 3: Case 1: Trajectories of the swarm with individual
at a distance away.
agents (red dots) and four target destinations (black x’s).
5
time = 0.00 s
5
time = 0.05 s
0.12
B. Case 2: L22 +Quadratic Term, Four Target Destinations
y
y
y
y
0
0
0.1
In Case 2, four swarm densities move towards the four
target destinations given by the blue-solid lines (mean re0.08
-5
-5
sponses) in Figure 2. Figure 4 shows the trajectory snapshots
-5
0
5
-5
0
5
x
x
and final states of each of the swarm densitiesPSfrag
duringreplacements
the
0.06
time = 0.10 s
time = 0.40 s
5
5
simulation. The target destinations are plotted as black x’s.
0.04
The red dots are the individual swarm agents that form
0
0
the Gaussian mixture densities. From the figure, all four
0.02
densities converge to the target destinations in approximately
-5
-5
0.17 seconds and approximately 0.03 of steady-state error
-5
0
5
-5
0
5
x
x
between the densities’ position to the target destinations.
In comparison to Case 1, Figure 2 shows that for small Fig. 5: Case 3: Trajectories for the swarm with individual
distances between the initial state and the target destination, agents (red dots) and three target destinations (black x’s).
the L22 distance is sufficient for state convergence, but as the
2
distance increases, the L2 distance diverges away. By adding D. Case 4: L2 +Quadratic term, Five Target Destinations
2
the quadratic term to L22 , the optimization step can directly
determine the minimum for the control solution shown in
Figure 1(c). Therefore, the target destinations attract the
swarm densities at distances that fail for only L22 distance
given by Figure 2.
For Case 4, four swarm densities are attracted to five
target destinations. The trajectory snapshots of the densities
are visually shown relative to the five target destinations
in Figure 6. All densities have a steady state error of
approximately 0.17 which follow the theory as expected.
Since the swarm densities are far from each other, the effects
of repulsion are minimal. Also, the densities are attracted
to the four target destinations that make up a square, but
they are also attracted to the target destination at the origin.
This is due to the minimization of the objective function
that has a L22 and a quadratic term where the individual
densities will attract towards the target destinations. Since
there is an additional target destination of attraction at the
origin, all four swarm densities are affected by the origin
as they are moving towards the 4 square target destinations.
Thus, compared to Case 2 with only 4 target destinations,
the densities in this case will have steady state error due to
attraction of the target destination that has no density.
time = 0.00 s
5
5
time = 0.05 s
0.09
y
y
0.08
0
0
0.07
0.06
-5
-5
Sfrag replacements
5
0
-5
-5
5
x
time = 0.10 s
5
0
5
0.05
x
time = 0.40 s
0.04
y
y
0.03
0
0
0.02
0.01
-5
-5
0
5
-5
-5
x
0
5
x
Fig. 6: Case 4: Trajectories for the swarm with individual
agents (red dots) and five target destinations (black x’s).
VI. C ONCLUSIONS
The objective of this paper was to formulate a method to
control swarming formations by MPC and RFS theory. By
setting up the problem using information divergence to define
the distance between the swarm RFS and a desired distribution, an optimal control problem was found that tracked a
linear system with a nonquadratic objective function using
MPC that approximates the optimal control solution that
minimizes the objective function.
ACKNOWLEDGMENT
The authors wish to acknowledge support by the National
Aeronautics and Space Administration under Contract Number NNX16CP45P issued through the NASA STTR Program
by the Jet Propulsion Laboratory (JPL) and led by Amir
Rahmani at JPL. The authors wish to acknowledge useful
conversations related to satellite technologies with Chuck
Hisamoto, Vaughn Weirens, and Suneel Sheikh of ASTER
Labs, Inc.
R EFERENCES
[1] Kube, C. Ronald, and Hong Zhang. “Collective robotics: From social
insects to robots.” Adaptive behavior 2.2 (1993): 189-218.
[2] Vassev, Emil, Mike Hinchey, and Joey Paquet. “Towards an ASSL
specification model for NASA swarm-based exploration missions.”
Proceedings of the 2008 ACM symposium on Applied computing.
ACM, 2008.
[3] Ryan, Allison, et al. “An overview of emerging results in cooperative
UAV control.” Decision and Control, 2004. CDC. 43rd IEEE Conference on. Vol. 1. IEEE, 2004.
[4] Rubenstein, Michael, Alejandro Cornejo, and Radhika Nagpal.
“Programmable self-assembly in a thousand-robot swarm.” Science
345.6198 (2014): 795-799.
[5] Lunze, Jan. Feedback control of large-scale systems. London: PrenticeHall, (1992): 9-11.
[6] Pace, M., M. Birattari, and M. Dorigo. “The swarm/potential model:
Modeling robotics swarms with measure-valued recursions associated
to random finite sets.” IEEE Transactions on Robotics, page submitted
(2013).
[7] Centralized Control. Ian Sommerville , 2008. Web. 05 May 2017.
[8] Mondada, Francesco, et al. “The cooperation of swarm-bots: Physical
interactions in collective robotics.” IEEE Robotics & Automation Magazine 12.2 (2005): 21-28.
[9] Bandyopadhyay, Saptarshi, Soon-Jo Chung, and Fred Y. Hadaegh.
“Probabilistic swarm guidance using optimal transport.” Control Applications (CCA), 2014 IEEE Conference on. IEEE, 2014.
[10] Acikmese, Behet, and David S. Bayard. “A markov chain approach to
probabilistic swarm guidance.” American Control Conference (ACC),
2012. IEEE, 2012.
[11] Chattopadhyay, Ishanu, and Asok Ray. “Supervised self-organization
of homogeneous swarms using ergodic projections of markov chains.”
IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics) 39.6 (2009): 1505-1515.
[12] Bandyopadhyay, Saptarshi, Soon-Jo Chung, and Fred Y. Hadaegh. “Inhomogeneous Markov chain approach to probabilistic swarm guidance
algorithm.” 5th Int. Conf. Spacecraft Formation Flying Missions and
Technologies. 2013.
[13] Hadaegh, Fred Y., Soon-Jo Chung, and Harish M. Manohara. “On
development of 100-gram-class spacecraft for swarm applications.”
IEEE Systems Journal 10.2 (2016): 673-684.
[14] Eren, Utku, and Behet Açıkmeşe. “Velocity field generation for density
control of swarms using heat equation and smoothing kernels.” IFACPapersOnLine 50.1 (2017): 9405-9411.
[15] Camacho, Eduardo F., and Carlos Bordons Alba. Model predictive
control. Springer Science & Business Media, 2013.
[16] Di Cairano, S., H. Park, and I. Kolmanovsky. “Model predictive
control approach for guidance of spacecraft rendezvous and proximity
maneuvering.” International Journal of Robust and Nonlinear Control
22.12 (2012): 1398-1427.
[17] Manikonda, V., et al. “A model predictive control-based approach for
spacecraft formation keeping and attitude control.” American Control
Conference, 1999. Proceedings of the 1999. Vol. 6. IEEE, 1999.
[18] Morgan, Daniel, Soon-Jo Chung, and Fred Y. Hadaegh. “Decentralized
model predictive control of swarms of spacecraft using sequential
convex programming.” Advances in the Astronautical Sciences 148
(2013): 1-20.
[19] Morgan, Daniel, Soon-Jo Chung, and Fred Hadaegh. “Spacecraft
swarm guidance using a sequence of decentralized convex optimizations.” (2012): Art-No.
[20] Morgan, Daniel, Soon-Jo Chung, and Fred Y. Hadaegh. “Swarm assignment and trajectory optimization using variable-swarm, distributed
auction assignment and model predictive control.” AIAA guidance,
navigation, and control conference. 2015.
[21] Mahler, Ronald PS. “Multitarget Bayes filtering via first-order multitarget moments.” IEEE Transactions on Aerospace and Electronic
systems 39.4 (2003): 1152-1178.
[22] Vo, B-N., and W-K. Ma. “The Gaussian mixture probability hypothesis
density filter.” IEEE Transactions on signal processing 54.11 (2006):
4091-4104.
[23] Todorov, Emanuel, and Weiwei Li. “A generalized iterative LQG
method for locally-optimal feedback control of constrained nonlinear
stochastic systems.” American Control Conference, 2005. Proceedings
of the 2005. IEEE, 2005.
[24] Hoang, Hung Gia, et al. “The Cauchy Schwarz divergence for Poisson point processes.” IEEE Transactions on Information Theory 61.8
(2015): 4475-4485.
[25] Inaba, Masayuki, and Peter I. Corke. “Linear-quadratic control and
smoothing.” Robotics Research: The 16th International Symposium
ISRR. Switzerland: Springer, 2016. 43-47. Print.
[26] Findeisen, Rolf, and Frank Allgwer. “An introduction to nonlinear
model predictive control.” 21st Benelux Meeting on Systems and
Control. Vol. 11. Eindhoven, The Netherlands: Technische Universiteit
Eindhoven Veldhoven, 2002.
[27] Find minimum of unconstrained multivariable function - MATLAB
fminunc, MathWorks. Web.
| 3cs.SY
|
1
Mitigating Asymmetric Nonlinear Weight
Update Effects in Hardware Neural Network
based on Analog Resistive Synapse
Chih-Cheng Chang, Pin-Chun Chen, Teyuh Chou, I-Ting Wang, Boris Hudec, Che-Chia Chang,
Chia-Ming Tsai, Tian-Sheuan Chang, Senior Member, IEEE, and Tuo-Hung Hou,
Senior Member, IEEE
Abstract—Asymmetric nonlinear weight update is considered
as one of the major obstacles for realizing hardware neural
networks based on analog resistive synapses because it
significantly compromises the online training capability. This
paper provides new solutions to this critical issue through
co-optimization with the hardware-applicable deep-learning
algorithms. New insights on engineering activation functions and
a threshold weight update scheme effectively suppress the
undesirable training noise induced by inaccurate weight update.
We successfully trained a two-layer perceptron network online
and improved the classification accuracy of MNIST handwritten
digit dataset to 87.8/94.8% by using 6-bit/8-bit analog synapses,
respectively, with extremely high asymmetric nonlinearity.
Index Terms—Neuromorphic computing, RRAM, synapse,
multilayer perceptron
D
I. INTRODUCTION
EEP learning on software-defined neural networks has
greatly advanced many aspects of artificial intelligence,
such as image recognition, speech recognition, natural
language understanding, and decision making, by using a
general-purpose learning procedure on conventional
computing hardware, such as central processing units or
graphic processing units [1]. Recently, the limitations of the
conventional computing hardware on the form factor, cost, and
power consumption of deep learning systems have promoted
active research on hardware neural networks (HNNs) that are
inspired by the physical structure of high-density, low-power,
and distributed biological neural networks [2-3]. The HNNs
employing crossbar resistive random access memory
(RRAM)-based synaptic arrays are particularly interesting
because they may significantly improve the learning efficiency
by fully parallelizing the vector-matrix multiplication
(weighted sum) and outer-product weight update through the
This work was supported by the Ministry of Science and Technology of
Taiwan under grant: MOST 105-2119-M-009-009 and Research of Excellence
program MOST 106-2633-E-009-001, and Taiwan Semiconductor
Manufacturing Company. T.-H. Hou acknowledges support by NCTU-UCB
I-RiCE program, under grant MOST 106-2911-I-009-301.
The authors are with Department of Electronics Engineering and Institute of
Electronics, National Chiao Tung University, Hsinchu, Taiwan
(email:[email protected])
distributed weight storage [4]. Impressive RRAM-based HNN
prototypes have been demonstrated including speech and
electroencephalography signal recognition [5], pattern
recognition [6], dot product engine [7], nature image
processing [8], and face classification [9]. However, most
implementations were limited to primitive algorithms such as
single-layer perceptron or sparse coding, and many of them
relied on offline training. By contrast, our brain processes
information intelligently and in real time through a
considerably more complicated deep neural networks with a
layered structure [10]. The software approach also resorts to
extremely deep neural networks to improve recognition
accuracy that surpasses humans [11]. Designing a deep HNN
with online training capability is a nontrivial task because of the
challenges of non-ideal properties of RRAM synapses [4,
12–16]. In particular, the asymmetric nonlinear weight update
of RRAM synapses had been identified as an outstanding issue
that significantly degraded the learning accuracy [15–17].
Several approaches have been proposed aiming at improving
the asymmetric nonlinear weight update of RRAM synapses,
for example by careful material/device engineering [18], by
implementing an additional compensational circuit [19], or by
optimizing the operating pulse waveform [20]. However, to
further increase the immunity against this undesirable nonlinear
effect, co-optimization considering both the device
characteristics and the HNN algorithms suitable for future deep
learning would be essential.
This paper focused on the HNN design employing a
Ta/HfO2/Al-doped TiO2/TiN synaptic device deposited by
atomic layer deposition (ALD). The device exhibited
bidirectional analog weight change that is suitable for the
ultimate high-density neural network by using one
two-terminal RRAM cell per synapse [4]. By contrast, more
than one RRAM cell per synapse would be required by using
binary RRAM [21] or unidirectional analog RRAM [15].
Furthermore, we presented a hardware-applicable multilayer
perceptron (MLP) algorithm that accelerated supervised online
training through massive parallelism. The MLP algorithm is the
foundation of numerous deep-learning neural networks such as
convolutional neural network (CNN). We reported that
engineering the nonlinear activation function and employing a
cutoff threshold on weight update were effective for mitigating
the effects of asymmetric nonlinear weight update. The
2
Fig. 2. Non-ideal properties of RRAM-based synapse including bounded weight
with finite precision and asymmetric nonlinear weight update. The potentiation
and depression curves with different degrees of asymmetric nonlinearity (ANL)
are also plotted.
Fig. 1. (a) DC switching curve of the Ta/HfO2/Al-doped TiO2/TiN RRAM
devices. Inset shows the cross-sectional TEM image of the Ta/HfO2/Al-doped
TiO2/TiN RRAM device with a HfO2/TiO2 bilayer thickness of 12 nm. (b)
Conductance modulation of multiple devices during AC potentiation and
depression operations. Potentiation and depression pulses were 3 V for 0.01 ms
and -3 V for 70 ms, respectively. The conductance value were readout at -1.5 V.
The box-and-whisker plot indicating the device-to-device variation and the
fitting curve using Eq. (1)-(3) are also sketched. (c) Reproducible conductance
modulation during repeated potentiation and depression cycles with a total of
3,000 training pulses. The curve in the first and final 500 cycles are zoomed in.
classification accuracy of MNIST handwritten digit dataset was
improved to 94.8% by considering an extremely high
asymmetric nonlinearity in weight update. We also discussed
the effects of reduced precision in weights and neurons. The
results suggest the HNN algorithm, if designed properly, could
be highly robust against the non-ideal asymmetric nonlinearity
of weight update.
II. FULLY ANALOG RRAM SYNAPSE
The synaptic device we considered in this study was
implemented using a non-filamentary Ta/HfO2/Al-doped
TiO2/TiN RRAM. The device resembled similar analog
synaptic properties as those in the previously reported
Ta/TaOx/TiO2/Ti RRAM [20, 22–23]. Figure 1(a) shows the
DC switching characteristics of the device with positive set and
negative reset voltages. The total film thickness of the new
HfO2/TiO2 oxide bilayers, as shown in the transmission
electron microscope image (Inset in Fig. 1(a)), was further
scaled down to approximately 12 nm by using ALD compared
with the previous 40-nm-thick TaOx/TiO2 deposited by
sputtering. The conformal ALD growth and oxide thickness
scaling are highly desirable in future 3D crossbar synaptic
arrays with extremely high-density and high-connectivity [20].
In 3D crossbar synaptic arrays, which resemble the 3D neural
network in the human brain [24], the oxide bilayers must be
conformally deposited on the vertical sidewalls of deep
trenches or vias. The pitch size of the trenches or vias and thus
the ultimate packing density are determined by the thickness of
the oxide bilayers. More detailed discussions on the device
fabrication process and RRAM switching characteristics will
be presented elsewhere. Figure 1(b) shows the synaptic
plasticity determined by measuring the device conductance (G)
at -1.5 V. The G value changed from a G minimum (Gmin) to a
G maximum (Gmax) in a range of approximately 4 S within one
cycle of synaptic operation, which comprised 50 consecutive
potentiation pulses (P-pulses) at 3 V for 0.01 ms, followed by
another 50 consecutive depression pulses (D-pulses) at -3 V
for 70 ms. Here the increment and decrement of the device
conductance are referred to as potentiation and depression,
respectively. The device-to-device variation was considerably
small because of the uniform ALD growth. The multiple analog
states were precisely trained according to P/D-pulse inputs
provided by external neural circuits. The bidirectional analog
weight modulation with minimal conductance fluctuation
obtained in this device is inherently different from the
conventional filamentary synapses [25]. Furthermore, Fig. 1(c)
shows reproducible synaptic plasticity during repeated
potentiation and depression cycles with a total of 3,000 training
pulses.
Although this device demonstrated promising analog
synaptic responses, several non-ideal properties compared with
the weight presented by floating-point numbers in
software-defined neural networks are noticeable, as
summarized in Fig. 2. Their influence should be carefully
evaluated when implementing HNNs. First, the precision of
3
synaptic weight is limited because the total plasticity range
Gmax/Gmin is finite and precise control on the weight value is
governed by the integer number of training pulses. The discrete
weight values in this device are bounded between Gmax and Gmin
with approximately a 5 to 6-bit precision. Second, the weight
change is not linearly proportional to the P/D-pulse inputs. The
change of device conductance is more dramatic during the first
few P/D-pulses and saturates as the number of pulses increases,
which is referred to as nonlinear weight update. More
importantly, the potentiation and depression characteristics are
not symmetric. The weight change induced by the P-pulses is
much stronger at the low-conductance states while that induced
by the D-pulses is much stronger at the high-conductance states,
which is referred to as asymmetric weight update. This
asymmetric nonlinear weight update creates a hysteresis loop
[4] in Fig. 2 when using the number of P/D-pulses to update
weights, and results in a considerable error. For example, the
weight value undergoing 25 P-pulses is not identical as that
undergoing 50 P-pulses plus 25 D-pulses. The asymmetric
nonlinear relation between discrete G values and the
cumulative number of P/D-pulses (nP/nD) is fitted
mathematically by:
GP (nP ) Gmin A
nP
nP ek
GD (nD ) Gmax A
( N nD )
( N nD ) ek
A (Gmax Gmin ) (1
ek
)
N
(1)
(2)
(3)
where GP and GD are functions describing conductance of
potentiation and depression curves, respectively. N is the
maximum allowed number of input pulses, which defines the
weight precision of RRAM devices, and k is a fitting parameter.
Furthermore, this non-ideal effect could be quantitatively
defined using an asymmetric nonlinearity factor (ANL) as:
N
N
ANL GP ( ) GD ( ) / (Gmax Gmin )
2
2
(4)
The ANL value is between zero to one. Completely symmetric
potentiation and depression curves yield an ANL value of zero
while the device exhibits an ANL value of 0.7 in Fig. 1(b).
III. MULTILAYER PERCEPTRON ALGORITHM FOR HNNS
Here we present a modified two-layer perceptron
algorithm for implementing HNNs. The network includes one
input neuron layer where neurons are numbered from 1 to i, one
hidden neuron layer where neurons are numbered from 1 to j,
and one output neuron layer where neurons are numbered from
1 to k, as illustrated in Fig. 3(a). Two adjacent neuron layers are
referred to as pre-neuron and post-neuron layers along the
forward propagation direction, and are fully connected using
RRAM synapses. The fully connected network could be
implemented using a crossbar RRAM array, as illustrated in Fig.
Fig. 3. (a) Illustration of a two-layer fully-connected perceptron network. (b)
Implementation of weighted sum and outer-product weight update by using the
crossbar RRAM configuration. Only the first fully-connected layer was
illustrated. (c) Illustrations of sigmoid and ReLU activation functions with and
without parallel shifts.
3(b). The initial conductance values of RRAM synapses are
randomized in the range between Gmin and Gmax. The weighted
sum is performed by summing all current components on the
same column, and the post neurons in different columns
parallel convert analog current signal to digital voltage signal
by using appropriate integrated circuits, such as a leaky
integrate-and-fire neuron circuit. Note that time-consuming
readout of conductance values of individual devices is not
necessary when performing this weighted sum calculation by
using the crossbar configuration. However, the synaptic weight
values could be both positive and negative float-point numbers
in conventional multilayer perceptron algorithms while the
discrete conductance values in RRAM synapses are only
positive. Therefore, the effect of signed weight values are
emulated by subtracting a reference current flowing through a
separate dummy column. The weighted sum values of the
hidden (Hj) and output (Ok) neurons are given by:
I
H j f a ( S Hj ) f a b j X i Wij
i 1
I
f a b j X i Gij Gr
i 1
(5)
4
(13)
J
Ok f a ( SkO ) f a bk H j W jk
j 1
I
f a bk H j G jk Gr
i 1
(6)
where fa is a nonlinear activation function, and Xi is the value of
input neurons. S Hj and S kO are the weighted sum before
activation. bj(bk), Wij(Wjk), and Gij(Gjk) are bias value, signed
weight value, and device conductance value in the first (second)
fully-connected layer. Gr is a constant reference conductance in
the dummy column given by:
Gr (Gmax Gmin ) / 2
(7)
The nonlinear activation functions evaluated in this work
include the popular sigmoid function and rectified linear unit
(ReLU) function [26]. Fig. 3(c) shows the illustration of both
functions. The sigmoid function is given by:
f a ( x)
1
1 e ( x s )
(8)
The ReLU function is given by:
f a ( x) max 0, ( x s )
(9)
where s is the parallel shifted value of the functions, which is
set to zero by default. The supervised training is performed
using a gradient-descent backpropagation method. The
)
backpropagation values of the output ( OkBP ) and hidden ( H BP
j
neurons are given by:
OkBP Tk Ok f a' S kO
(10)
K
H BP
OkBPW jk f a' S Hj
j
(11)
k 1
where Tk is the target value of each output neuron, and f a' is the
derivative of the activation function. The conductance values of
RRAM synapses are updated according to multiplications of
forward-propagation
values
of
pre-neurons
and
backward-propagation values of post-neurons, which is
mathematically equivalent to perform outer product of two
vectors. The multiplication could be realized through the
careful overlapping of the pre- and post-neuron waveforms by
using the crossbar configuration [16, 27], and the effective
pulse number nij (njk) applied on RRAM synapses in the first
(second) fully-connected layer is given by:
nij round ( X i H BP
j )
n jk round ( H j OkBP )
(12)
where is the learning rate that is a scaling factor to regulate
training. Rounding is performed considering only an integer of
the effective pulse number. Positive and negative n induce
potentiation and depression of the device conductance,
respectively. This outer-product weight update is highly
efficient because all synapses in the array are updated
simultaneously. More importantly, the update is solely based
on the values of pre-neuron and post-neuron. No prior state of
RRAM synapses needs be stored or readout before updating. If
the weight change Wij(Wjk) of RRAM synapses is linearly
proportional to nij (njk), Eqs. (12) and (13) would guarantee
correct weight changes that follow the ideal gradient-descent
backpropagation method. However, practical RRAM synapses
possess nonlinear weight update behaviors as discussed in Eqs.
(1) and (2). The new conductance value of the updated RRAM
synapse, Gnew, should be considered using Eqs. (1) and (2) as:
G new GP (nPold n) for n 0
(14)
GD (nDold n) for n 0
where nPold and nDold are equivalent cumulative number of
P/D-pulses before update. Furthermore, rounding is performed
on all forward- and backward-propagation values, i.e. Xi, Hj, Ok,
, and OkBP , because the precision of practical neuron
H BP
j
circuits is finite. For the mathematically unbounded activations,
such as ReLU, an upper bound of neuron values was set to
facilitate rounding in hardware. This upper bound was
estimated in advance in this work to prevent clipping of high
positive neuron values, but it potentially could be optimized
empirically based on the classification results. This
hardware-applicable algorithm could be extended to more
sophisticate deep neural networks, such as CNN, with more
than one hidden layer.
To evaluate the non-ideal properties of the RRAM analog
synapse on the hardware-applicable MLP algorithm, we
performed simulations by using MNIST handwritten digit
dataset. The network size was fixed at 784×300×10. The bias
values were set to zero for simplicity in this work. Each training
epoch contained 60,000 training images, which were 28×28
pixels large with 256 gray levels. The weight update could be
performed after every training image by using stochastic
gradient descent or after a minibatch of training images by
using minibatch gradient descent. We verified that the
convergence errors of training did not yield significant
differences between these two methods. Therefore, the results
reported in this paper were obtained using minibatch gradient
descent to accelerate the simulation. The typical batch size was
set to 10 to 50. A constant learning rate was used throughout
training, but the value was optimized individually for different
hardware setting conditions to improve training convergence.
The classification accuracy was tested after training by using
another independent 10,000 testing images.
5
TABLE I
AVERAGE SPARSITY AND CLASSIFICATION ACCURACY BY USING SIGMOID AND
RELU ACTIVATIONS WITH DIFFERENT SHIFTED VALUES S AND ANL.
Fig. 4. Effects of various activation functions. (a) Dependence of asymmetric
nonlinearity on the classification accuracy of MNIST handwritten digit dataset.
(b) Convergence of MSE during training for both sigmoid and ReLU activation
functions when linear and asymmetric nonlinear weight updates are assumed.
IV. MITIGATING ASYMMETRIC NONLINEAR WEIGHT UPDATE
BY ENGINEERING ACTIVATIONS
We first verified the effects of asymmetric nonlinear
weight update on MLP training, which was reported to be one
of the major limitations of electronic synaptic devices [15–17].
Both neuron and weight precision were set to 8 bits. The
nonlinear but symmetric weight update did not significantly
affect the classification accuracy (not shown), which is
consistent with previous studies [15–17]. By contrast, Fig. 4(a)
shows that the asymmetric nonlinear weight update
substantially degraded the accuracy by using the sigmoid
activation function and a practical ANL factor larger than 0.4.
Figure 4(b) shows the convergence of mean square error (MSE)
during training. The asymmetric nonlinear weight update
resulted in a slower convergence rate and premature saturation
of MSE compared with the linear one because the intended
amount of weight change on the device is not linearly
proportional to the effective pulse number according to Eqs
(1)-(2). The amount of weight change depends on the current
weight state and is not equal when receiving an identical
number of P-pulses and D-pulses. Ideally corrections on Eqs
(1)-(2) would be possible in the neuron circuits to compensate
the effects of asymmetric nonlinear weight update if the current
weight state is known. However, this would involve reading the
weight (conductance) value of every individual RRAM
synapse and significantly comprise the advantage of hardware
acceleration through the crossbar configuration.
By contrast, we found that this undesirable effects could be
effectively mitigated by using the ReLU activation. Both the
classification accuracy in Fig. 4(a) and the convergence of
training MSE in Fig. 4(b) were significantly improved when
with a large ANL value of 0.8. Although both the ReLU and
sigmoid functions rectify negative neuron values, the ReLU
function produces a hard saturation effect where all negative
values and their corresponding derivatives are forced to be
or Ok and
exactly zero, i.e. both the values of Hj and H BP
j
OkBP are zero. We referred those neurons with zero/non-zero
values as inactive/active neurons. According to Eqs. (12)-(13),
inactive neurons prevent those weights connected to them from
updating (n=0). We define an average sparsity of the hidden
layer as the fraction of zeros in Hj [28]. For ANL=0.8, the
ReLU function greatly increased the average sparsity to 49%
compared with 0.12% when using the sigmoid function. The
quasi-hard saturation effect could be employed by using shifted
sigmoid activation functions, which also showed excellent
immunity against asymmetric nonlinear update in Fig. 4(a).
Table 1 list the average sparsity and classification accuracy by
using the sigmoid and ReLU activation functions with different
s values for both linear (ANL=0) and highly nonlinear
(ANL=0.8) weight update. The average sparsity increased with
the increase of positive s, and the classification accuracy also
increased with s before a turning point. For ANL=0, both
sigmoid and ReLU activations achieved high accuracy of over
97%. For ANL=0.8, the highest classification accuracy of
6
Fig. 5. Dependence of average sparsity on the classification accuracy for
different activation functions and asymmetric nonlinearity.
91.5% was obtained by using s = 3.5 for sigmoid while the
highest classification accuracy of 85.4 % was obtained by using
s = 0.2 for ReLU. Figure 5 illustrates that the optimized degrees
of sparsity depended strongly on the choice of activation
functions and the ANL values. The advantages of sparsity has
been widely recognized in the field of deep neural networks
[28]. The mathematical reasons behind had been explained by
improving
information
disentangling,
variable-size
representation,
linear
separability,
and
distributed
representation [28, 29]. In our MLP simulation, these factors
were not determinant because the classification accuracy was
relatively independent of the sparsity for the ideal linear update
case. Both sigmoid and ReLU activations could achieve similar
accuracy with different sparsity levels (~35% for sigmoid and
~65% for ReLU). We attribute the improvement by increasing
the average sparsity for the highly nonlinear update case to the
suppression of inaccurate weight update. The frequent
switching of the signs of the small n values during training
would induce substantial update error and unnecessary
perturbation toward convergence because of the asymmetric
nonlinear weight update. Their effects could be considered as
noise during weight training. The hard or quasi-hard saturation
effect introduced by ReLU or shifting promotes convergence
by suppressing noise induced by small n while favoring large
n that carries the true information. However, excessive
sparsity deteriorated the performance of forward propagation in
both training and testing for an equal number of neurons
because it reduces the effective capacity of the model [28].
More information was discarded in forward propagation by
using high s values. Therefore, there exists an optimized s
values in Fig. 5. Furthermore, for a high ANL, sigmoid
outperformed ReLU on the classification accuracy when
shifting is employed, as shown in Fig. 4 and Table 1. ReLU is
widely adopted in deep neural networks because its constant
derivative overcomes the vanishing gradient problem [26, 28].
However, the vanishing gradient problem is not a concern for
the simple MLP demonstrated here. The major difference
between ReLU and a shifted sigmoid is that ReLU is
unbounded with a constant derivative for high positive values.
Therefore, n is not equal to zero according to Eqs. (10)-(13)
Fig. 6. Effects of threshold weight update scheme with both the sigmoid and
ReLU activation functions in the presence of high asymmetric nonlinearity
(ANL=0.8). The improvement of using the shifted sigmoid (with the optimized
s=3.5) and the threshold update are cumulative while the additional benefit of
using shifted ReLU (with the optimized s=0.2) is less pronounced.
even if the output neuron values exceed the highest possible
target values. An upper bounded activation function with zero
derivation for high positive values, such as shifted sigmoid,
could suppress this unnecessary update when the nonideal
training noise is significant. Our preliminary data (not shown)
suggest that using a clipped ReLU activation by setting a
suitable upper bound value could achieve a comparable
accuracy as that using the shifted sigmoid in Fig. 4.
V. THRESHOLD WEIGHT UPDATE SCHEME
Inspired by the result that high sparsity is only required
during weight update but not forward propagation, we propose
a threshold weight update scheme where high degree of
sparsity is only introduced to suppress the noise of weight
update by modifying Eqs. (12)-(13) as:
nij round fth ( X i ) fth ( H BP
j )
(15)
n jk round fth ( H j ) fth (OkBP )
(16)
The cutoff threshold function fth with a threshold value th is
given by:
fth ( x) x for x th
(17)
0 for x th
The optimized th value is considered as an empirical parameter
similar to the learning rate for different datasets. The same th
value was applied for all Xi, Hj, H BP
, and OkBP in this work for
j
simplicity, but potentially different th values could be
optimized for each term separately. Figure 6 shows the
comparison of classification accuracy by using different th and
s values. For ANL=0.8, the highest classification accuracy of
94.8 % was obtained by using th = 0.99 for sigmoid (s=3.5),
which is close to the ideal 97% accuracy with ANL=0. The
highest classification accuracy of 92.9 % was obtained by using
th = 0.6 for ReLU (s=0). Because of the cutoff threshold,
7
Fig. 7. Evolution of weight distributions in the first and second fully-connected
layers for both ANL=0 and 0.8. The shifted sigmoid (s=3.5) and threshold
weight update (th=0.99) were used.
extremely high sparsity could be induced in the
backpropagation path (sparsity over 97% for the sigmoid case)
to suppress update noise without compromising the accuracy of
forward propagation. The effects of shifted activations and
threshold weight update were cumulative especially for the
sigmoid activation, which without any modification produces
an inherently less sparse network and is thus more prone to
noise induced by erroneous weight update. Independently
tuning sparsity during the forward propagation by using the
shifted activation and sparsity during weight update by using
the cutoff threshold function achieved higher classification
accuracy than that by tuning one of them alone. Figure 7 shows
the evolution of weight distributions in the first and second
fully-connected layers for the optimized networks with ANL=0
(97% accuracy) and ANL=0.8 (94.8% accuracy). The final
weight distributions were surprisingly different for these two
cases with a comparable accuracy. This result highlights that
the modified MLP algorithm, if designed properly, is
self-adaptive to high asymmetric nonlinearity of weight update.
VI. WEIGHT AND NEURON PRECISION
The effects of reducing weight and neuron precision were
further investigated in the presence of high asymmetric
nonlinearity. A lower weight/neuron precision relaxes the
design specifications on the synaptic devices/peripheral circuits
for the ease of hardware implementation. Figure 8a shows that
the requirement on the weight precision was higher for high
ANL values. The asymmetric nonlinearity compromised the
effective bit precision because of the irregular weight update.
The future development of synaptic devices with a higher
weight precision, for example by increasing the dynamic range
of the conductance change to accommodate more analog states,
should be continuously pursued. However, an acceptable
accuracy of 87.8% was obtained by using a 6-bit weight
precision, which is potentially obtainable by using practical
analog RRAM-based synapses such as the one shown in Fig. 1.
Figure 8b shows that the requirement on the neuron precision
was also higher for high ANL values because the activations
and threshold update implemented in the algorithm were
Fig. 8. Dependence of the classification accuracy on the (a) weight and (b)
neuron precision. The shifted sigmoid (s=3.5) and threshold weight update
(th=0.99) were used.
sensitive to the neuron precision. A 6-bit neuron precision
could be a good tradeoff between accuracy and hardware
complexity.
VII. CONCLUSION
Asymmetric nonlinear weight update is an outstanding
issue of online training for analog RRAM-based HNNs.
Previous studies mainly focused on improving this non-ideal
property from various perspectives of device optimization.
Through careful co-optimization between device properties and
hardware-applicable deep-learning algorithms, this paper
proposed an alternative solution to this problem. By
engineering the activation function and adopting the threshold
weight update scheme, the modified MLP algorithm
suppressed update noise induced by inaccurate small weight
changes and is self-adaptive to asymmetric nonlinearity. Both
modifications could be implemented in neuron circuits. We
successfully trained a two-layer perceptron network online in
the presence of extremely high asymmetric nonlinearity
(ANL=0.8) and improves classification accuracy of MNIST
handwritten digit dataset to 94.8% by using 8-bit weight and
neuron precisions. The tradeoff between accuracy and
8
hardware complexity was attainable using a reduced precision
to 6 bits. The results suggest a promising direction for
designing a robust HNN against asymmetric nonlinear weight
update in practical synaptic devices.
REFERENCES
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
Y. Lecun, Y. Bengio, and G. Hinton, "Deep learning," Nature, vol. 521,
no. 7553, pp. 436-444, 2015.
D. Kuzum, S. Yu, and H.-S. P. Wong, "Synaptic electronics: materials,
devices and applications," Nanotechnology, vol. 24, 382001, 2013.
S. B. Eryilmaz, D. Kuzum, S. Yu, and H.-S. P. Wong, "Device and system
level design considerations for analog-non-volatile- memory based
neuromorphic architectures," in Proc. IEEE Int. Electron Devices
Meeting (IEDM), 2015, pp. 64–67.
S. Yu, P.-Y. Chen, Y. Cao, L. Xia, Y. Wang, and H. Wu, "Scaling-up
resistive synaptic arrays for neuro-inspired architecture: challenges and
prospect," in Proc. IEEE Int. Electron Devices Meeting (IEDM), 2015, pp.
451–454.
S. Park, M. Chu, J. Kim, J. Noh, M. Jeon, B. H. Lee, H. Hwang, B. Lee,
and B.-G. Lee, "Electronic system with memristive synapses for pattern
recognition," Sci. Rep., vol. 5, 10123, 2015.
M. Prezioso, F. Merrikh-Bayat, B. D. Hoskins, G. C. Adam, K. K.
Likharev, and D. B. Strukov, "Training and operation of an integrated
neuromorphic network based on metal-oxide memristors," Nature, vol.
521, no. 7550, pp. 61–64, 2015.
M. Hu, J. P. Strachan, Z. Li, E. M. Grafals, N. Davila, C. Graves, S. Lam,
N. Ge, J. Yang, and R. S. Williams, "Dot-product engine for
neuromorphic computing: Programming 1T1M crossbar to accelerate
matrix-vector multiplication," in Proc. Design Automation Conf. (DAC),
2016, pp. 1–6.
P. M. Sheridan, F. Cai, C. Du, W. Ma, Z. Zhang, and W. D. Lu, "Sparse
coding
with
memristor
networks,"
Nature
Nanotech.,
doi:10.1038/nnano.2017.83, 2017.
P. Yao, H. Wu, B. Gao, S. B. Eryilmaz, X. Huang, W. Zhang, Q. Zhang, N.
Deng, L. Shi, H.-S. P. Wong, and H. Qian, "Face classification using
electronic synapses," Nature Communi., vol. 8, 15199, 2017.
S. Shipp, "Structure and function of the cerebral cortex," Current Biology,
vol. 17, no. 12, pp. R443–449, 2007.
K. He, X. Zhang, S. Ren, and J. Sun, "Deep residual learning for image
recognition," in Proc. Computer Vision and Pattern Recognition (CVPR),
2016, pp. 770–778.
M. Hu, H. Li, Y. Chen, Q. Wu, G. S. Rose, and R. W. Linderman,
"Memristor crossbar-based neuromorphic computing system: a case
study," IEEE Trans. Neural Networks and Learning Systems, vol. 25, no.
10, pp. 1864–1878, 2014.
P.-Y. Chen, B. Lin, I-T. Wang, T.-H. Hou, J. Ye, S. Vrudhula, J.-S. Seo, Y.
Cao, S. Yu, "Mitigating effects of non-ideal synaptic device
characteristics for on-chip learning," in Proc. Computer-Aided Design
(ICCAD), 2015, pp. 194–199.
P.-Y. Chen, D. Kadetotad, Z. Xu, A. Mohanty, B. Lin, J. Ye, S. Vrudhula,
J.-S. Seo, Y. Cao, and S. Yu, "Technology-design co-optimization of
resistive cross-point array for accelerating learning algorithms on chip," in
Proc. Design, Automation & Test in Europe Conference & Exhibition
(DATE), 2015, pp. 854–859.
[15] G. W. Burr, R. M. Shelby, C. Nolfo, J. W. Jang, R. S. Shenoy,
P. Narayanan, K. Virwani, E. U. Giacometti, B. Kurdi, and H.
Hwang, "Experimental demonstration and tolerancing of a large-scale
neural network (165,000 synapses), using phase-change memory as the
synaptic weight element," in Proc. IEEE Int. Electron Devices Meeting
(IEDM), 2014, pp. 697-700.
[16] S. Agarwal, S. J. Plimpton, D. R. Hughart, A. H. Hsia, I. Richter, J. A Cox,
C. D. James, M. J. Marinella, "Resistive memory device requirements for
a neural algorithm accelerator," in Proc. Int. Joint Conf. Neural Networks
(IJCNN), 2016, pp. 929-938.
[17] J.-W. Jang, S. Park, G. W. Burr, H. Hwang, Y.-H. Jeong, "Memristor
crossbar-based neuromorphic computing system: a case study," IEEE
Electron Device Lett., vol. 36, no. 5, pp. 457–459, 2015.
[18] J. Woo, K. Moon, J. Song, S. Lee, M. Kwak, J. Park, and H. Hwang,
"Improved synaptic behavior under identical pulses using AlOx/HfO2
bilayer RRAM array for neuromorphic systems," IEEE Electron Device
Lett., vol. 37, no. 8, pp. 994–997, 2016.
[19] D. Lee, J. Park, K. Moon, J. Jang, S. Park, M. Chu, J. Kim, J. Noh, M. Jeon,
B. H. Lee, B. Lee, B.-G. Lee, and H. Hwang, "Oxide based nanoscale
analog synapse device for neural signal recognition system," in Proc.
IEEE Int. Electron Devices Meeting (IEDM), 2015, pp. 91–94.
[20] I.-T. Wang, C.-C. Chang, L.-W. Chiu, T. Chou, and T.-H. Hou, "3D
Ta/TaOx/TiO2/Ti synaptic array and linearity tuning of weight update for
hardware neural network applications," Nanotechnology, vol. 27, 365204,
2016.
[21] D. Garbin, O. Bichler, E. Vianello, Q. Rafhay, C. Gamrat, L.Perniola, G.
Ghibaudo, B. DeSalvo, "Variability-tolerant convolutional neural
network for pattern recognition applications based on OxRAM synapses,"
in Proc. IEEE Int. Electron Devices Meeting (IEDM), 2014, pp. 661–664.
[22] I.-T. Wang, Y.-C. Lin, Y.-F. Wang, C.-W. Hsu, and T.-H. Hou, "3D
synaptic architecture with ultralow sub-10 fJ energy per spike for
neuromorphic computation," in Proc. IEEE Int. Electron Devices Meeting
(IEDM), 2014, pp. 665-668.
[23] Y.-F. Wang, Y.-C. Lin, I-T. Wang, T.-P. Lin, and T.-H. Hou,
"Characterization and modeling of nonfilamentary Ta/TaOx/TiO2/Ti
analog synaptic device," Sci. Rep., vol. 5, 10150, 2015.
[24] V. J. Wedeen, D. L. Rosene, R. Wang, G. Dai, F. Mortazavi, P. Hagmann,
J. H. Kaas, W.-Y. I. Tseng, "The geometric structure of the brain fiber
pathways," Science, vol. 335, 1628–1634, 2012.
[25] S. Yu, B. Gao, Z. Fang, H. Yu, J. Kang, and H.-S. P. Wong, "A
neuromorphic visual system using RRAM synaptic devices with sub-pJ
energy and tolerance to variability: experimental characterization and
large-scale modeling," in Proc. IEEE Int. Electron Devices Meeting
(IEDM), 2012, pp. 239–242.
[26] V. Nair and G. E. Hinton, "Rectified linear units improve restricted
boltzmann machines," in Proc. Int. Conf. Machine Learning (ICML),
2010, pp. 807–814.
[27] D. Kadetotad, Z. Xu, A. Mohanty, P.-Y. Chen, B. Lin, J. Ye, S. Vrudhula,
S. Yu, Y. Cao, and J.-S. Seo, "Parallel architecture with resistive
crosspoint array for dictionary learning acceleration," IEEE J. Emerging
Sel. Topics Circuits Syst., vol. 5, no. 2, pp. 194–204, 2015.
[28] X. Glorot, A. Bordes, and Y. Bengio, "Deep sparse rectifier neural
networks," in Proc. Conf. Artificial Intelligence and Statistics (AISTATS),
2011, pp. 315–323.
[29] Y. Bengio, "Learning deep architectures for AI," Foundations and Trends
in Machine Learning, vol. 2, no. 1, pp. 1–127, 2009.
| 9cs.NE
|
A Lottery Model for Center-type Problems With Outliers ∗
David G. Harris†
Thomas Pensyl‡
Aravind Srinivasan§
Khoa Trinh¶
arXiv:1710.00287v1 [cs.DS] 1 Oct 2017
Abstract
In this paper, we give tight approximation algorithms for the k-center and matroid center problems with outliers. Unfairness arises naturally in this setting: certain clients could always be considered as outliers. To address this issue, we introduce a lottery model in which each client j is allowed
to submit a parameter pj ∈ [0, 1] and we look for a random solution that covers every client j with
probability at least pj . Our techniques include a randomized rounding procedure to round a point
inside a matroid intersection polytope to a basis plus at most one extra item such that all marginal
probabilities are preserved and such that a certain linear function of the variables does not decrease
in the process with probability one.
1
Introduction
The classic k-center and Knapsack Center problems are known to be approximable to within factors
of 2 and 3 respectively [7]. These results are best possible unless P=NP [7, 8]. In these problems, we
are given a metric graph G and want to find a subset S of vertices of G subject to either a cardinality
constraint or a knapsack constraint such that the maximum distance from any vertex to the nearest vertex
in S is as small as possible. We shall refer to vertices in G as clients. Vertices in S are also called centers.
It is not difficult to see that a few outliers (i.e., very distant clients) may result in a very large optimal
radius in the center-type problems. This issue was raised by Charikar et. al. [4], who proposed a robust
model in which we are given a parameter t and only need to serve t out of given n clients (i.e. n − t
outliers may be ignored in the solution). Here we consider three robust center-type problems: the Robust
k-Center (RkCenter) problem, the Robust Knapsack Center (RKnapCenter) problem, and the Robust
Matroid Center (RMatCenter) problem.
Formally, an instance I of the RkCenter problem consists of a set V of vertices, a metric distance
d on V , an integer k, and an integer t. Let n = |V | denote the number of vertices (clients). The goal is
to choose a set S ⊆ V of centers (facilities) such that (i) |S| ≤ k, (ii) there is a set of covered vertices
(clients) C ⊆ V of size at least t, and (iii) the objective function
R := max min d(i, j)
j∈C i∈S
is minimized.
In the RKnapCenter problem, each vertex i ∈ VPhas a weight wi ∈ [0, 1], and the cardinality
constraint (i) is replaced by the knapsack constraint:
i∈S wi ≤ 1. Similarly, in the RMatCenter
problem, the constraint (i) is replaced by a matroid constraint: S must be an independent set of a given
matroid M. Here we assume that we have access to the rank oracle of M.
∗
Research supported in part by NSF Awards CNS-1010789 and CCF-1422569, and by a research award from Adobe,
Inc. A preliminary version of this work appeared in the Proc. International Workshop on Approximation Algorithms for
Combinatorial Optimization Problems (APPROX), 2017.
†
Department of Computer Science, University of Maryland, College Park, MD 20742.
Email:
[email protected]
‡
Department of Computer Science, University of Maryland, College Park, MD 20742. Email: [email protected]
§
Department of Computer Science and Institute for Advanced Computer Studies, University of Maryland, College Park,
MD 20742. Email: [email protected]
¶
Department of Computer Science, University of Maryland, College Park, MD 20742. Email: [email protected]
1
In [4], Charikar et. al. introduced a greedy algorithm for the RkCenter problem that achieves
an approximation ratio of 3. Recently, Chakrabarty et. al. [3] give a 2-approximation algorithm for
this problem. Since the k-center problem is a special case of the RkCenter problem, this ratio is best
possible unless P=NP.
The RKnapCenter problem was first studied by Chen et. al. [5], which showed that one can achieve
an approximation ratio of 3 if allowed to slightly violate the knapsack constraint by a factor of (1 + ).
It is still unknown whether there exists a true approximation algorithm for this problem. The current
inapproximability bound is still 3 due to the hardness of the Knapsack Center problem. The current best
approximation guarantee for the RMatCenter problem is 7 by Chen et. al. [5]. This problem has a
hardness of (3 − ) via a reduction from the k-supplier problem.
From a practical viewpoint, unfairness arises inevitably in the robust model: some clients might
always be considered as outliers and hence not covered within the guaranteed radius. To address this
issue, we introduce a lottery model for these problems. The idea is to randomly pick a solution from a
public list such that each client j ∈ V is guaranteed to be covered with probability at least pj , where
pj ∈ [0, 1] is the success rate requested by j. This is also motivated by the fact that different clients may
have different tolerances of getting connected to their closest facility. One possible way of determining
the pj values is to let each client j pay for the chance of being served. Note that the robust model is a
special case when pj = 0 for all j ∈ V . Similarly, when all pj ’s are equal to 1, it becomes the standard
model where all clients must be connected.
The lottery model might also be useful in the context of clustering. Recall that clustering is a
fundamental task in unsupervised machine learning. Basically, we want to partition a set of data points
to clusters in such a way that the points in the same cluster are “similar” to each other. The k-center
clustering is one of the popular approaches to this task. (See also k-means clustering [1,9] and k-median
clustering [2, 12, 15].) Naturally, the robust model can be applied to get rid of some “bad data points”
or “noise” and hence improve the overall quality of all clusters. Here over-fitting may occur when
this model misclassifies some good points as outliers. The lottery model offers the flexibility to decide
whether a point should be included in the solution via the pj values.
In this paper, we introduce new approximation algorithms for these robust center problems under
the lottery model. (Note that this model has been used recently for the Knapsack Center problems [6],
although the techniques and problems in that paper are different from ours.) We also propose improved
approximation algorithms for the RkCenter problem and the RMatCenter problem.
1.1
The Lottery Model
In this section, we formally define our lottery model for the above-mentioned problems. First, the Fair
Robust k-Center (FRkCenter) problem is formulated as follows. Besides the parameters V, d, k and t,
each vertex j ∈ V has a “target” probability pj ∈ [0, 1]. We are interested in the minimum radius R for
which there exists a distribution D on subsets of V such that a set S drawn from D satisfies the following
constraints:
Coverage constraint: |C| ≥ t with probability one, where C is the set of all clients in V that are within
radius R from some center S,
Fairness constraint: Pr[j ∈ C] ≥ pj for all j ∈ V , where C is as in the coverage constraint,
Cardinality constraint: |S| ≤ k with probability one.
Here we aim for a polynomial-time, randomized algorithm that can sample from D. Note that the
RkCenter is a special of this variant in which all pj ’s are set to be zero.
The Fair Robust Knapsack Center (FRKnapCenter) problem and Fair Robust Matroid Center
(FRMatCenter) problem are defined similarly except that we replace the cardinality constraint by a
knapsack constraint and a matroid constraint, respectively. More formally, in the FRKnapCenter problem, each vertex i has a weight wi ∈ [0, 1], and we require the total weight of centers in S to be at most
2
1 with probability one. Similarly, in the FRMatCenter problem, we are given a matroid M and we
require the solution S to be an independent set of M with probability one.
1.2
Our contributions and techniques
First of all, we give tight approximation algorithms for the RkCenter and RMatCenter problems.
Theorem 1.1. There exist a 2-approximation algorithm for the RkCenter problem 1 and a 3-approximation
algorithm for the RMatCenter problem.
Our main results for the lottery model are summarized in the following theorems.
Theorem 1.2. For any given constant > 0 and any instance I = (V, d, k, t, p~) of the FRkCenter
problem, there is a randomized polynomial-time algorithm A which can compute a random solution S
such that
• |S| ≤ k with probability one,
• |C| ≥ (1 − )t, where C is the set of all clients within radius 2R from some center in S and R is
the optimal radius,
• Pr[j ∈ C] ≥ (1 − )pj for all j ∈ V .
Theorem 1.3. For any > 0 and any instance I = (V, d, w, B, t, p~) of the FRKnapCenter problem,
there is a randomized polynomial-time algorithm A which can return random solution S such that
P
•
i∈S wi ≤ (1 + )B with probability one,
• |C| ≥ t, where C is the set of vertices within distance 3R of S.
• Pr[j ∈ C] ≥ pj for all j ∈ V .
Finally, the FRMatCenter can be reduced to (randomly) rounding a point in a matroid intersection
polytope. We design a randomized rounding algorithm which can output a pseudo solution, which
consists of a basis plus one extra center. By using a preprocessing step and a configuration LP, we
can satisfy the matroid constraint exactly (respectively, knapsack constraint) while slightly violating the
coverage and fairness constraints in the FRMatCenter (respectively, FRKnapCenter) problem. We
believe these techniques could be useful in other facility-location problems (e.g., the matroid median
problem [10, 14]) as well.
Theorem 1.4. For any given constant γ > 0 and any instance I = (V, d, M, t, p~) of the FRMatCenter
(respectively, FRKnapCenter) problem, there is a randomized polynomial-time algorithm A which can
return a random solution S such that
• S is a basis of M with probability one, (respectively, w(S) ≤ B with probability one)
• |C| ≥ t − γ 2 n, where C is the set of vertices within distance 3R of S.
• there exists a set T ⊆ V of size at least (1 − γ)n, which is deterministic, such that Pr[j ∈ C] ≥
pj − γ for all j ∈ T .
1
A 2-approximation algorithm has also been found independently by Chakrabarty et. al. [3], and in a private discussion
between Marek Cygan and Samir Khuller. Our algorithm here is different from the algorithm in [3].
3
1.3
Organization
The rest of this paper is organized as follows. In Section 2, we review some basic properties of matroids
and discuss a filtering algorithm which is used in later algorithms. Then we develop approximation
algorithms for the FRkCenter, FRKnapCenter, and FRMatCenter problems in the next three sections.
2
2.1
Preliminaries
Matroid polytopes
We first review
a few basic facts about matroid polytopes. For any vector z and set S, we let z(S) denote
P
the sum i∈S zi . Let M be any matroid on the ground set Ω and rM be its rank function. The matroid
base polytope of M is defined by
PM := x ∈ RΩ : x(S) ≤ rM (S) ∀S ⊆ Ω; x(Ω) = rM (Ω); xi ≥ 0 ∀i ∈ Ω .
Definition 2.1. Suppose Ax ≤ b is a valid inequality of PM . A face D of PM (corresponding to this
valid inequality) is the set D := {x ∈ PM : Ax = b} .
The following theorem gives a characterization for any face of PM (See, e.g., [11, 13]).
Theorem 2.2. Let D be any face of PM . Then it can be characterized by
D = x ∈ RΩ : x(S) = rM (S) ∀S ∈ L; xi = 0 ∀i ∈ J; x ∈ PM ,
where J ⊆ Ω and L is a chain family of sets: L1 ⊂ L2 ⊂ . . . ⊂ Lm . Moreover, it is sufficient to choose
L as any maximal chain L1 ⊂ L2 ⊂ . . . ⊂ Lm such that x(Li ) = rM (Li ) for all i = 1, 2, . . . , m.
Proposition 2.3. Let x ∈ PM be any point and I be the set of all tight constraints of PM on x. Suppose
D is the face with respect to I. Then one can compute a chain family L for D as in Theorem 2.2 in
polynomial time.
0 (S) = r (S) − x(S) for S ⊆ Ω
Proof. Recall that rM is a submodular function. Thus the function rM
M
is also submodular. It is well-known that submodular minimization can be done in polynomial time. We
0 (S) : S ⊆ Ω}. If there are multiple solutions, we
solve the following optimization problem: min {rM
let S0 be any solution of minimal size. (This can be done easily, say, by trying to drop each item from
the current solution and resolving the program.) We add S0 to our chain. Then we find some minimal
0 (S ) = 0, add S to our chain, and repeat the process.
superset S1 of S0 such that rM
1
1
Corollary 2.4. Let D be any face of PM . Then it can be characterized by
D = x ∈ RΩ : x(S) = bS ∀S ∈ O; xi = 0 ∀i ∈ J; x ∈ PM ,
where J ⊆ Ω and O is a family of pairwise disjoint sets: O1 , O2 , . . . , Om , and bO1 , . . . , bOm are some
integer constants.
Proof. By Theorem 2.2, we have
D = x ∈ RΩ : x(S) = rM (S) ∀S ∈ L; xi = 0 ∀i ∈ J; x ∈ PM ,
where J ⊆ Ω and L is the chain: L1 ⊂ L2 ⊂ . . . ⊂ Lm . Now let us define O1 := L1 , O2 := L2 \
L1 , O3 := L3 \L2 , . . . , Om := Lm \Lm−1 , and bO1 := rM (L1 ), bO2 := rM (L2 )−rM (L1 ), . . . , bOm :=
rM (Lm ) − rM (Lm−1 ). It is not difficult to verify that
D = x ∈ RΩ : x(S) = bS ∀S ∈ O; xi = 0 ∀i ∈ J; x ∈ PM .
4
2.2
Filtering algorithm
All algorithms in this paper are based on rounding an LP solution. In general, for each vertex i ∈ V , we
have a variable yi ∈ [0, 1] which represents the probability that we want to pick i in our solution. (In
the standard model, yi is the “extent” that i is opened.) In addition, for each pair of i, j ∈ V , we have a
variable xij ∈ [0, 1] which represents the probability that j is connected to i.
Note that in all center-type problems, the optimal radius R is always the distance between two
vertices. Therefore, we can always “guess” the value of R in O(log n) by a binary search on the sorted
entries of the matrix d. WLOG, we may assume that we know the correct value of R. For any j ∈ V ,
we define
P Bj := {i ∈ V : d(i, j) ≤ R} and we let Fj := {i ∈ V : d(i, j) ≤ R ∧ xij > 0} and
sj :=
i∈Bj xij . We shall refer to Fj as a cluster with cluster center j. Depending on a specific
problem, we may have different constraints on xij ’s and yi ’s. In general, the following constraints are
valid in most of the problems here:
XX
xij ≥ t,
(1)
j∈V i∈Bj
X
i∈Bj
xij ≤ 1,
xij ≤ yi ,
yi , xij ≥ 0,
For the fair variants, we may also require that
X
xij ≥ pj ,
i∈Bj
∀j ∈ V,
∀i, j ∈ V,
(2)
(3)
∀i, j ∈ V.
(4)
∀j ∈ V.
(5)
Constraint (1) says that at least t vertices should be covered. Constraint (2) ensures that each vertex
is only connected to at most one center. Constraint (3) means vertex j can only connect to center i if
it is open. Constraint (5) says that the total probability of j being connected should be at least pj . By
constraints (2) and (3), we have y(Fj ) ≤ 1.
The first step of all algorithms in this paper is to use the following filtering algorithm to obtain a
maximal collection of disjoint clusters. The algorithm will return the set V 0 of cluster centers of the
chosen clusters. In the process, we also keep track of the number cj of other clusters removed by Fj for
each j ∈ V 0 .
Algorithm 1 RF ILTERING (x, y)
1:
2:
3:
4:
5:
6:
7:
3
V0 ←∅
for each unmarked cluster Fj in decreasing order of sj do
V 0 ← V 0 ∪ {j}
Set all unmarked clusters Fk (including Fj itself) s.t. Fk ∩ Fj 6= ∅ as marked.
Let cj be the number of marked clusters in this step.
~c ← (cj : j ∈ V 0 )
return (V 0 , ~c)
The k-center problems with outliers
In this section, we first give a simple 2-approximation algorithm for the RkCenter problem. Then, we
give an approximation algorithm for the FRkCenter problem, proving Theorem 1.2.
5
3.1
The robust k-center problem
Suppose I = (V, d, k, t) is an instance the RkCenter problem with the optimal radius R. Consider the
polytope PRkCenter containing points (x, y) satisfying constraints (1)–(4), and the cardinality constraint:
X
yi ≤ k.
(6)
i∈V
Since R is the optimal radius, it is not difficult to check that PRkCenter 6= ∅. Let us pick any fractional
solution (x, y) ∈ PRkCenter . The next step is to round (x, y) into an integral solution using the following
simple algorithm.
Algorithm 2 R K C ENTER ROUND (x, y)
(V 0 , ~c) ← RF ILTERING (x, y) .
2: S ← the top k vertices i ∈ V 0 with highest value of ci .
3: return S
1:
Analysis. By construction, the algorithm returns a set S of k open centers. Note that, for
Peach i ∈ S,
ci is the number of distinct clients within radius 2R from i. Thus, it suffices to show that i∈S ci ≥ t.
By inequality (2), we have sj ≤ 1 for all j ∈ V 0 . Thus,
X
X
ci si ≥
si ≥ t,
i∈V 0
i∈V
where the first inequality is due to the greedy choice of vertices in V 0 and the second inequality follows
from (1). Now recall that the clusters whose centers in V 0 are pairwise disjoint. By constraint (6),
X
X
X
si ≤
y(Fi ) ≤
yi ≤ k.
i∈V 0
It follows by the choice of S that
3.2
P
i∈S ci
i∈V 0
i∈V
≥ t. This concludes the first part of Theorem 1.1.
The fair robust k-center problem
Assume I = (V, d, k, t, p~) be an instance of the FRkCenter problem
with the optimal radius R. Fix
any > 0. If k ≤ 2/, then we can generate all possible O n1/ solutions and then solve an LP to
obtain the corresponding marginal probabilities. So the problem can be solved easily in this case. We
will assume that k ≥ 2/ for the rest of this section. Consider the polytope PFRkCenter containing points
(x, y) satisfying constraints (1)–(4), the fairness constraint (5), and the cardinality constraint (6). We
now show that PFRkCenter is actually a valid relaxation polytope.
Proposition 3.1. The polytope PFRkCenter is non-empty.
Proof. It suffices to point out a solution (x, y) ∈ PFRkCenter . Since R is the optimal radius, there exists
a distribution D satisfying the coverage, fairness, and cardinality constraints. Suppose S is sampled
from D and C is the set of all clients in V that are within radius R from some
P center S. We now set
yi := Pr[i ∈ S] for all i ∈ V . Since |S| ≤ k with probability one, we have i∈V yi = E[|S|] ≤ k, and
hence constraint (6) is valid.
We construct the assignment variable x as follows. For each j ∈ V , set zj := 0. Then for each
i ∈ Bj , set xij := min{yi , 1 − zj } and update zj := zj + xij . It is not hard to see that inequalities (2)
and (3) hold by this construction. Now let us fix any j ∈ V . By fairness guarantee of D and the union
bound, we have
X
pj ≤ Pr[j ∈ C] ≤
yi .
i∈Bj
6
Thus, by construction of x, we have
X
i∈Bj
xij ≥ Pr[j ∈ C] ≥ pj ,
and hence inequality (5) is satisfied. Finally, we have
X
XX
E[|C|] =
Pr[j ∈ C] ≤
xij .
j∈V
j∈V i∈Bj
Since |C| ≥ t with probability one, E[|C|] ≥ t, implying that inequality (1) holds.
Our algorithm is as follows.
Algorithm 3 FR K C ENTER ROUND (, x, y)
(V 0 , ~c) ← RF ILTERING (x, y) .
2: Form vector y 0 by setting
1:
yj0 ← (1 − )
3:
4:
5:
6:
7:
X
xij
i∈Fj
for each j ∈ V 0 .
while at least three entries of y 0 are in the range (0, 1) do
0
Let δ ∈ RV , δ 6= 0 be such that δi = 0 ∀i ∈ V 0 : yi0 ∈ {0, 1}, δ(V 0 ) = 0, and ~c · δ = 0.
Choose scaling factors a, b > 0 such that
0
0
• y 0 + aδ ∈ [0, 1]V and y 0 − bδ ∈ [0, 1]V
• there is at least one new entry of y 0 + aδ which is equal to zero or one
• there is at least one new entry of y 0 − bδ which is equal to zero or one
b
With probability a+b
, update y 0 ← y 0 + aδ; else, update y 0 ← y 0 − bδ.
return S = {i ∈ V : yi0 > 0}.
Analysis. First, note that one can find such a vector δ in line 4 as the system of δ(V 0 ) = 0 and
~c · δ = 0 consists of two constraints and at least three variables (and hence is underdetermined.) Thus,
the algorithm rounds at least one fractional variable per iteration, and terminates after O(n) rounds. Let
S denote the (random) solution returned by FR K C ENTER ROUND and C be the set of all clients within
radius 3R from some center in S. Theorem 1.2 can be verified by the following propositions.
Proposition 3.2. |S| ≤ k with probability one.
Proof. By definition of y 0 at line 2 of FR K C ENTER ROUND, we have
X X
X X
y 0 (V 0 ) = (1 − )
xij ≤ (1 − )
yi ≤ (1 − )k ≤ k − 2,
j∈V 0 i∈Fj
j∈V 0 i∈Fj
since k ≥ 2/. Note that the sum y 0 (V 0 ) is never changed in the while loop (lines 4 . . . 7) because
δ(V 0 ) = 0. Then the final vector y 0 contains at most two fractional values at the end of the while loop.
By rounding these two values to one, the size of S is indeed at most k.
Proposition 3.3. |C| ≥ (1 − )t with probability one.
Proof. At the beginning of the while loop, we have
X
X
X
cj sj ≥ (1 − )
sj ≥ (1 − )t.
~c · y 0 =
cj yj0 (Fj ) = (1 − )
j∈V 0
j∈V 0
7
j∈V
Again, the quantity ~c · y 0 is unchanged in the while loop because ~c · δ = 0 implies that ~c · (y 0 + aδ) = ~c · y 0
0
and ~c · (y 0 − bδ) = ~c · y 0 in each iteration. Note that if y 0 ∈ {0, 1}V , then ~c · y 0 is the number of clients
within radius 2R from some center i such that yi0 = 1. Basically, we round the two remaining fractional
values of y 0 to one in line 8; and hence, the dot product should be still at least (1 − )t.
Proposition 3.4. Pr[j ∈ C] ≥ (1 − )pj for all j ∈ V .
Proof. Fix any j ∈ V . The algorithm RF ILTERING guarantees that there exists k ∈ V 0 such that
Fj ∩ Fk 6= ∅ and sk ≥ sj . Now we claim that E[yk0 ] = yk0 . This is because the expected value of yk0 does
not change after any single iteration:
E[yk0 ] = (yk0 + aδ)
b
a
+ (yk0 − bδ)
= yk0 .
a+b
a+b
Then
Pr[k ∈ S] = Pr[yk0 > 0] ≥ Pr[yk0 = 1] = E[yk0 ] = yk0 = (1 − )sk .
Therefore,
Pr[j ∈ C] ≥ Pr[k ∈ S] ≥ (1 − )sk ≥ (1 − )sj ≥ (1 − )pj ,
by constraint (5).
4
The Knapsack Center problems with outliers
We study the RKnapCenter and FRKnapCenter problems in this section. Recall that in these problems, each vertex has a weight and we want to make sure that the total weight of the chosen centers does
not exceed 1. We first give a 3-approximation algorithm for the RKnapCenter problem that slightly
violates the knapsack constraint. Although this is not better than the known result by [5], both our algorithm and analysis here are more natural and simpler. It serves as a starting point for the next results.
For the FRKnapCenter, we show that it is possible to satisfy the knapsack constraint exactly with small
violations in the coverage and fairness constraints.
4.1
The robust knapsack center problem
Suppose I = (V, d, w, t) is an instance the RKnapCenter problem with the optimal radius R. Consider the polytope PRKnapCenter containing points (x, y) satisfying constraints (1)–(4), and the knapsack
constraint:
X
wi yi ≤ 1.
(7)
i∈V
Again, it is not difficult to check that PRKnapCenter 6= ∅. Let us pick any fractional solution (x, y) ∈
PRKnapCenter . Our pseudo-approximation algorithm to round (x, y) is as follows.
Algorithm 4 RK NAP C ENTER ROUND (x, y)
(V 0 , ~c) ← RF ILTERING (x, y) .
2: For each i ∈ V 0 , let vi = arg minj∈Fi {wj } be the vertex with smallest weight in Fi
n
o
P
P
0
3: Let P 0 := z ∈ [0, 1]V :
c
z
≥
t
∧
w
z
≤
1
0
0
i
i
v
i
i
i∈V
i∈V
1:
Compute an extreme point Y of P 0
5: return S = {vi : i ∈ V, Yi > 0}
4:
8
Analysis. We first claim that P 0 6= ∅ which implies that the extreme point Y of P 0 (in line 4) does exist.
To see this, we claim that the vector s lies in P 0 . For, we have:
X
X
ci si ≥
si ≥ t.
i∈V 0
i∈V
Also,
X
i∈V
0
wvi si =
X
i∈V
0
wvi
X
j∈Fi
xji ≤
X
i∈V
0
wv i
X
j∈Fi
yj ≤
XX
0
i∈V j∈Fi
wj yj ≤
X
i∈V
wi yi ≤ 1.
All the inequalities follow from LP constraints and definitions of si , ci , and vi .
Proposition 4.1. The solution S returned by RK NAP C ENTER ROUND satisfies w(S) ≤ 1 + 2wmax and
|C| ≥ t, where C is the set of vertices within distance 3R of S and wmax is the maximum weight of any
vertex in V .
Proof. First, observe that any extreme point of P 0 has at most two fractional values. (In the worst case,
0
an
i = 1,
P extreme point z is
Pfully determined by |V | − 2 tight constraints of the form zi = 0 or∗ z∗∗
c
z
=
t,
and
w
z
=
1.)
By
construction
of
S,
there
are
at
most
two
vertices
i
,
i
such
i∈V 0 i i
i∈V 0 vi i
that Yi∗ , Yi∗∗ are fractional. Thus,
X
wvi Yi + wi∗ + wi∗∗ ≤ 1 + 2wmax .
w(S) =
i∈S\{i∗ ,i∗∗ }
Recall that each i ∈ V 0 has ci clients within distance 2R (and each client is counted only one time).
By the triangle inequality, these clients are within distance 3R from vi . Thus, S covers at least
X
X
ci Yi + ci∗ + ci∗∗ ≥
ci Yi ≥ t,
i∈S\{i∗ ,i∗∗ }
i∈S
clients within radius 3R.
4.2
The fair robust knapsack center problem
In this section, we will first consider a simple algorithm that only violates the knapsack constraint by
two times the maximum weight of any vertex. Then using a configuration polytope to “condition” on
the set of “big” vertices, we show that it is possible to either violate the budget by (1 + ) or to preserve
the knapsack constraint while slightly violating the coverage and fairness constraints.
4.2.1
Basic algorithm
Suppose I = (V, d, w, t, p~) is an instance the FRKnapCenter problem with the optimal radius R.
Consider the polytope PFRKnapCenter containing points (x, y) satisfying constraints (1)–(4), the fairness
constraint (5), and the knapsack constraint (7). The proof that PFRKnapCenter 6= ∅ is very similar to that
of Proposition 3.1 and is omitted here.
The following algorithm is a randomized version of RK NAP C ENTER ROUND.
Note that P 0 6= ∅, and so the decomposition at line 4 is well-defined (see the analysis in Section 4.1).
Proposition 4.2. The algorithm BASIC FRK NAP C ENTER ROUND returns a random solution S such that
w(S) ≤ 1 + 2wmax , |C| ≥ t, and Pr[j ∈ C] ≥ pj for all j ∈ V , where C is the set of vertices within
distance 3R of S and wmax is the maximum weight of any vertex in V .
Proof. By similar arguments to the proof of Proposition 4.1, we have w(S) ≤ 1 + 2wmax and |C| ≥ t.
To obtain the fairness guarantee, observe that vi ∈ S with probability at least zi = si . For any j ∈ V ,
let k ∈ V 0 be the vertex that removed j in the filtering step. We have
Pr[j ∈ C] ≥ Pr[vk ∈ S] ≥ sk ≥ sj ≥ pj ,
where the penultimate inequality is due to our greedy choice of k in RF ILTERING.
9
Algorithm 5 BASIC FRK NAP C ENTER ROUND (x, y)
(V 0 , ~c) ← RF ILTERING (x, y) .
2: For each i ∈ V 0 let vi := arg minj∈Fi {wj } be the vertex with smallest weight in Fi
n
o
P
P
0
3: Let P 0 := z ∈ [0, 1]V :
c
z
≥
t
∧
w
z
≤
1
i∈V 0 i i
i∈V 0 vi i
1:
4:
Decompose the vector s as a convex combination of extreme points z (1) , . . . , z (n+1) of P 0 :
s = p1 z (1) + . . . + pn+1 z (n+1) ,
P
where ` p` = 1 and p` ≥ 0 for all ` ∈ [n + 1].
5: Randomly choose Y ← z (`) with probability p` .
6: return S = {vi : i ∈ V, Yi > 0}
4.2.2
An algorithm slightly violating the budget constraint
Fix a small parameter > 0. A vertex i is said to be big iff wi > . Let U denote the collection of all
possible sets of big vertices. Since a solution can contain at most 1/ big vertices, we have |U| ≤ nO(1/) .
Consider the configuration polytope Pconfig1 containing points (x, y, q) with the following constraints:
P
qU = 1
PU ∈U U
∀j ∈ V, U ∈ U
i∈Bj xij ≤ qU
P
P
U
∀j ∈ V
U ∈U i∈Bj xij ≥ pj
U
U
≤ yi
∀i, j ∈ V, U ∈ U
x
Pij
U
w y ≤ qU B
∀U ∈ U
Pi∈V Pi i
U
x ≥ qU t
j∈V i∈Bj ij
U
yi = 1
∀U ∈ U, i ∈ U
U
yi = 0
∀U ∈ U, i ∈ V \ U, wi > 1/
U U
xij , yi , qU ≥ 0
∀i, j ∈ V, U ∈ U
We first claim that Pconfig1 is a valid relaxation polytope for the problem.
Proposition 4.3. The polytope Pconfig1 is non-empty.
Proof. Fix any optimal distribution D. Suppose S is sampled from D. For any U ∈ U, let
P E(U ) be the
event that U ⊆ S and S \U contains no big vertex, and let qU = Pr[E(U )]. It is clear that U ∈U qU = 1.
U
Let xU
ij = Pr[E(U ) ∧ j is connected to i] and let yi = Pr[E(U ) ∧ i ∈ S].
Now observe that
X
X
qU = Pr[E(U )] =
Pr[j is connected to i ∧ E(U )] =
xU
ij .
i∈Bj
i∈Bj
Similarly,
pj ≤ Pr[j is connected] =
=
X X
U ∈U i∈Bj
X
U ∈U
Pr[j is connected ∧ E(U )]
Pr[j is connected to i ∧ E(U )] =
X X
xU
ij .
U ∈U i∈Bj
U
Note that xU
ij /qU and yi /qU are the probabilities that j is connected to i and i ∈ S conditioned on
E(U ), respectively. Since the number of connected clients is at least t with probability one, we have
XX
XX
t ≥ E[# connected clients|E(U )] =
Pr[j is connected to i|E(U )] =
xU
ij /qU .
j∈V i∈Bj
j∈V i∈Bj
10
Similarly, w(S) ≤ 1 with probability one and so
X
1 ≥ E[w(S)|U ] =
wi (yiU /qU ).
i∈V
The other constraints can be verified easily. We conclude that (x, y, q) ∈ Pconfig1 .
We use the following Algorithm 6 to round any (x, y, q) ∈ Pconfig1 :
Algorithm 6 FRK NAP C ENTER ROUND 1 (x, y, q)
Randomly pick a set U ∈ U with probability qU
0
U
Let x0ij ← xU
ij /qU and yi ← min{yi /qU , 1}
3: return S = BASIC RFK NAP C ENTER ROUND (x0 , y 0 )
1:
2:
We are now ready to prove Theorem 1.3.
Proof of Theorem 1.3. We will show that solution S returned by FRK NAP C ENTER ROUND 1 satisfies
the requirements of Theorem 1.3. Let E(U ) denote the event that it select U ∈ U. Note that (x0 , y 0 )
satisfies the following constraints:
XX
x0ij ≥ t,
j∈V i∈Bj
X
i∈Bj
X
x0ij ≤ 1,
x0ij =
i∈Bj
X
i∈V
∀j ∈ V,
X
xij /qU ,
i∈Bj
x0ij ≤ yi0 ,
wi yi0
∀j ∈ V,
∀i, j ∈ V,
≤ 1.
Moreover, yi0 = 1 for all i ∈ U , and yi0 = 0 for all big vertices i ∈ V \ U . Thus, the two extra
fractional vertices opened by BASIC FRK NAP C ENTER ROUND have weight at most . By Proposition
4.2, we have w(S) ≤ 1 + 2. Moreover, conditioned on U , we have
X
X
x0ij =
xij /qU .
Pr[j ∈ C|E(U )] ≥
i∈Bj
i∈Bj
Thus, by definition of Pconfig1 and our construction of S, we get
Pr[j ∈ C] =
4.2.3
X
U ∈U
Pr[j ∈ C|E(U )] Pr[E(U )] ≥
X X
U ∈U i∈Bj
xij ≥ pj .
An algorithm that satisfies the knapsack constraint exactly
Let > 0 a small parameter to be determined. Let U denote the collection of all possible vertex sets of
at most d1/e; note that |U| ≤ nO(1/) . Suppose R is the optimal radius to our instance. Given a set
U ∈ U, we say that vertex j ∈ V is blue if d(j, U ) ≤ 3R; otherwise, vertex i is red. For any i ∈ V , let
RBall(i, U, R) denote the set of red vertices within radius 3R from i:
RBall(i, U, R) := {j ∈ V : d(i, j) ≤ 3R ∧ d(j, U ) > 3R}
11
Consider the configuration polytope Pconfig2 containing points (x, y, q) with the following constraints:
P
qU = 1
PU ∈U U
∀j ∈ V, U ∈ U
i∈B xij ≤ qU
P jP
U
∀j ∈ V
U ∈U i∈Bj xij ≥ pj
U
U
≤ yi
∀i, j ∈ V, U ∈ U
x
Pij
U
wi y ≤ qU B
∀U ∈ U
Pi∈V P i
U
j∈V
i∈Bj xij ≥ qU t
U
yi = 1
∀U ∈ U, i ∈ U
U
yi = 0
∀U ∈ U, i ∈ V \ U, |RBall(i, U, R)| ≥ n
U U
xij , yi , qU ≥ 0
∀i, j ∈ V, U ∈ U
We first claim that Pconfig2 is a valid relaxation polytope for the problem.
Proposition 4.4. The polytope Pconfig2 is non-empty.
Proof. Suppose S is a solution drawn from the optimal distribution D. Consider the following randomized procedure to generate a random subset US of S:
Algorithm 7 Generate random US :
Set US ← ∅.
while there exists ∈ S such that |RBall(i, US , R)| ≥ n do
3:
Select the vertex i of smallest index such that |RBall(i, US , R)| ≥ n
4:
Update US ← US ∪ {i}
5:
Mark all vertices within radius 3R of i as blue
1:
2:
Note that for all i ∈ S \ US , we have |RBall(i, US , R)| < n by the condition of the while-loop.
Moreover, we claim that |US | ≤ d1/e, so that US ∈ U. For, suppose |US | > 1/; for each i ∈ US ,
there are at least n red vertices turned into blue by i in the procedure. This implies that there are more
than (1/) × n = n vertices, which is a contradiction.
Now for any U ∈ U, we set qU := Pr[US = U ]. Let xU
= Pr[US = U ∧ j is connected to i], and
ij
P
finally let yiU = Pr[US = U ∧ i ∈ S]. Then it is clear that U ∈U qU = 1. Using similar arguments as
in the proof of Proposition 4.3, we have the following inequalities:
X
xU
ij ≤ qU , ∀j ∈ V, U ∈ U
i∈Bj
X X
U ∈U i∈Bj
X
i∈V
xU
ij ≥ pj ,
wi yiU ≤ qU ,
XX
j∈V i∈Bj
∀j ∈ V
∀U ∈ U
xU
ij ≥ qU t.
As mentioned before, if |RBall(i, US , R)| ≥ n then i ∈
/ S. Therefore,
yiU = Pr[US = U ∧ i ∈ S] = 0,
∀U ∈ U, i ∈ V \ U, |RBall(i, U, R)| ≥ n.
The other constraints can be verified easily. We conclude that (x, y, q) ∈ Pconfig2 .
12
Algorithm 8 FRK NAP C ENTER ROUND 2 (x, y, q)
1:
2:
3:
4:
5:
Randomly pick a set U ∈ U with probability qU
0
U
Let x0ij ← xU
ij /qU and yi ← min{yi /qU , 1}
0
S ← BASIC RFK NAP C ENTER ROUND (x0 , y 0 )
Let i1 , i2 be vertices in S 0 \ U having largest weights.
return S = S 0 \ {i1 , i2 }
Next, let us pick any (x, y, q) ∈ Pconfig2 and use the following algorithm to round it.
2
Analysis. Let us fix any γ > 0 and set := γ2 . Also, let E(U ) denote the event that U ∈ U is picked in
the algorithm. Again, observe that (x0 , y 0 ) satisfies the following inequalities:
XX
j∈V i∈Bj
X
i∈Bj
X
x0ij ≥ t,
x0ij ≤ 1,
x0ij =
i∈Bj
X
i∈V
∀j ∈ V,
X
xij /qU ,
i∈Bj
x0ij ≤ yi0 ,
wi yi0
∀j ∈ V,
∀i, j ∈ V,
≤ 1.
Recall that the algorithm BASIC FRK NAP C ENTER ROUND will return a solution S 0 consisting of a
set S 00 with w(S 00 ) ≤ 1 plus (at most) two extra “fractional” centers i∗ and i∗∗ . Since yi0∗∗ , yi0∗ are
fractional, we have that i∗ , i∗∗ ∈
/ U . Thus, by removing the two centers having highest weights in
S 0 \ U , we ensure that w(S) ≤ 1 with probability one.
Now we shall prove the coverage guarantee. By Proposition 4.2, S 0 covers at least t vertices within
radius 3R. If a vertex is blue, it can always be connected to some center in U and so it is not affected by
the removal of i1 , i2 . Because each of i1 and i2 can cover at most n other red vertices, we have
|C| ≥ t − 2n = 1 − γ 2 n.
For any j ∈ V , let Xj be the indicator random variable for the event that d(j, S 0 ) ≤ 3R but
d(j,
P S \ {i1 , i2 }) > 3R. We say that j is a bad vertex iff E[Xj ] ≥ γ, other, vertex j is good. Note that
j∈V Xj ≤ 2n with probability one. Thus, there can be at most 2n/γ bad vertices. Letting T denote
the set of good vertices, we have
|T | ≥ n − 2n/γ = (1 − γ)n.
By Proposition 4.2, Pr[j is covered by S 0 ] ≥ pj . For any j ∈ T , we have
Pr[j ∈ C] = Pr[j is covered by S 0 ∧ Xj = 0]
= Pr[j is covered by S 0 ] − Pr[j is covered by S 0 ∧ Xj = 1]
≥ Pr[j is covered by S 0 ] − Pr[Xj = 1] ≥ pj − γ.
This concludes the first part of Theorem 1.4 for the FRKnapCenter problem.
5
The Matroid Center problems with outliers
In this section, we will first give a tight 3-approximation algorithm for the RMatCenter problem, improving upon the 7-approximation algorithm by Chen et. al. [5]. Then we study the FRMatCenter
problem and prove the second part of Theorem 1.4.
13
5.1
The robust matroid center problem
Suppose I = (V, d, M, t) is an instance of the RMatCenter problem with optimal radius R. Let rM
denote the rank function of M. Consider the polytope PRMatCenter containing points (x, y) satisfying
constraints (1)–(4), and the matroid rank constraints:
y(U ) ≤ rM (U ),
∀U ⊆ V.
(8)
Since R is the optimal radius, it is not difficult to check that PRMatCenter 6= ∅. Let us pick any
fractional solution (x, y) ∈ PRMatCenter . The next step is to round (x, y) into an integral solution. Our
3-approximation algorithm is as follows.
Algorithm 9 RM AT C ENTER ROUND (x, y)
(V 0 , ~c) ←
RF ILTERING (x, y) .
0
2: Let P := z ∈ [0, 1]V : z(U ) ≤ rM (U ) ∀U ⊆ V ∧ z(Fi ) ≤ 1 ∀i ∈ V 0
3: Find a basic solution Y ∈ P 0 which maximizes the linear function f : [0, 1]V → R defined as
X X
f (z) :=
cj
zi for z ∈ [0, 1]V .
1:
j∈V 0
4:
i∈Fj
return S = {i ∈ V : Yi = 1}.
Analysis. Again, by construction, the clusters Fi are pairwise disjoint for i ∈ V 0 . Note that P 0 is the
matroid intersection polytope between M and another partition matroid polytope saying that at most
one item per set Fi for i ∈ V 0 can be chosen. Moreover, y ∈ P 0 implies that P 0 6= ∅. Thus, P 0 has
integral extreme points and optimizing over P 0 can be done in polynomial time. Note that the solution
S is feasible as it satisfies the matroid constraint. The correctness of RM AT C ENTER ROUND follows
immediately by the following two propositions.
Proposition 5.1. There are at least f (Y ) vertices in V that are at distance at most 3R from some open
center in S.
Proof. Recall that S is the set of vertices i ∈ V such that Yi = 1. Moreover, since Y (Fj ) ≤ 1, there can
be at most one open center in Fj (i.e., |S ∩ Fj | ≤ 1) for each j ∈ V 0 . For any j ∈ V 0 ,
• if Y (Fj ) = 0, then there is no open center in Fj and its contribution in f (Y ) is zero,
• if Y (Fj ) = 1, then we open some center i ∈ Fj and the contribution of j to f (Y ) is equal
to cj . (Recall that cj is the number of clusters Fk with Fj ∩ Fk 6= ∅.) By triangle inequality,
d(k, i) ≤ d(k, j) + d(j, i) ≤ 2R + R = 3R.
Proposition 5.2. We have f (Y ) ≥ t.
Proof. For each j ∈ V 0 and i ∈ Fj , define yi0 := xij (this is well-defined as all clusters Fj for j ∈ V 0
are pairwise disjoint). Also, set yi0 := 0 for other vertices i not belonging to any marked cluster. Then,
by greedy choice and constraint (1), we have
X
X
X
f (y 0 ) =
cj y 0 (Fj ) =
cj sj ≥
sj ≥ t.
j∈V 0
j∈V 0
By the choice of Y , we have f (Y ) ≥ f (y 0 ) ≥ t.
This analysis proves the second part of Theorem 1.1.
14
j∈V
5.2
The fair robust matroid center problem
In this section, we consider the FRMatCenter problem. It is not difficult to modify and randomize
algorithm RMC ENTER ROUND so that it would return a random solution satisfying both the fairness
guarantee and matroid constraint, and preserving the coverage constraint in expectation. This can be
done by randomly picking Y inside P 0 . However, if we want to obtain some concrete guarantee on
the coverage constraint, we may have to (slightly) violate either the matroid constraint or the fairness
guarantee. We leave it as an open question whether there exists a true approximation algorithm for this
problem.
We will start with a pseudo-approximation algorithm which always returns a basis of M plus at most
one extra center. Our algorithm is quite involved. We first carefully round a fractional solution inside
a matroid intersection polytope into a (random) point with a special property: the unrounded variables
form a single path connecting some clusters and tight matroid rank constraints. Next, rounding this
point will ensure that all but one cluster has an open center. Then opening one extra center is sufficient
to cover at least t clients.
Finally, using a similar preprocessing step similar to the one in Section 4.2.3, we correct the solution
by removing the extra center without affecting the fairness and coverage guarantees by too much. This
algorithm concludes Theorem 1.4.
5.2.1
A pseudo-approximation algorithm
Suppose I = (V, d, M, t, p~) is an instance the robust matroid center problem with the optimal radius
R. Let rM denote the rank function of M and PM be the matroid base polytope of M. Consider the
polytope PFRMatCenter containing points (x, y) satisfying constraints (1)–(4), the fairness constraint (5),
and the matroid constraints (8). We note that PFRMatCenter is a valid relaxation.
The main algorithm is summarized in Algorithm 10, which can round any vertex point (x, y) ∈
PFRMatCenter . Basically, we will round y iteratively. In each round, we construct a (multi)-bipartite
graph where vertices on the left side are the disjoint sets O1 , O2 , . . . in Corollary 2.4. Vertices on the
right side are corresponding to the disjoint sets F1 , F2 , . . . returned by RF ILTERING. Now each edge of
the bipartite graph, connecting Oi and Fj , represents some unrounded variable yv ∈ (0, 1) where v ∈ Oi
and v ∈ Fj . See Figure 1.
L
v
Oi
R
yv 2 (0, 1)
v
Fj
Figure 1: Construction of the multi-bipartite graph H = (L, R, EH ) in the main algorithm.
Then we carefully pick a cycle (path) on this graph and round variables on the edges of this cycle
(path). This is done by subroutines ROUND C YCLE, ROUND S INGLE PATH, and ROUND T WO PATHS.
See Figures 2, 3, and 4. Basically, these procedures will first choose a direction ~r which alternatively
15
increases and decreases the variables on the cycle (path) so that (i) all tight matroid constraints are
preserved and (ii) the number of (fractionally) covered clients is also preserved. Now we randomly
move y along ~r or −~r to ensure that all the marginal probabilities are preserved.
Finally, all the remaining, fractional variables will form one path on the bipartite graph. We round
these variables by the procedure ROUND F INAL PATH which exploits the integrality of any face of a
matroid intersection polytope. Then, to cover at least t clients, we may need to open one extra facility.
Algorithm 10 P SEUDO FRMC ENTER ROUND (x, y)
1:
2:
3:
4:
5:
6:
7:
8:
9:
(V 0 , ~c) ← RF ILTERING (x, y) and let F ← {Fj : j ∈ V 0 }
Set yi0 ← xij for all j ∈ V 0S
, i ∈ Fj
Set yi0 ← 0 for all i ∈ V \ j∈V 0 Fj
while y 0 still contains some fractional values do
Note that y 0 ∈ PM . Compute the disjoint sets O1 , . . . , Ot and constants bO1 , . . . , bOt as in
Corollary 2.4. S
S
Let O0 ← V \ ti=1 Oi and F0 ← V \ j∈V 0 Fj
Construct a multi-bipartite graph H = (L, R, EH ) where
• each vertex i ∈ L, where L = {0, . . . , t}, is corresponding to the set Oi
• each vertex j ∈ R, where R = {0} ∪ {k : Fk ∈ F}, is corresponding to the set Fj
• for each vertex v ∈ V such that yv ∈ (0, 1): if v belongs to some set Oi and Fj , add an
edge e with label v connecting i ∈ L and j ∈ R.
Check the following cases (in order):
• Case 1: H contains a cycle. Let ~v = (v1 , v2 , . . . , v2` ) be the sequence of edge labels on
this cycle. Update y 0 ← ROUND C YCLE(y 0 , ~v ) and go to line 4.
• Case 2: H contains a maximal path with one endpoint in L and the other in R. Let
~v = (v1 , v2 , . . . , v2`+1 ) be the sequence of edge labels on this path. Update y 0 ←
ROUND S INGLE PATH(y 0 , ~v ) and go to line 4.
• Case 3: There are at least 2 distinct maximal paths (not necessarily disjoint) having both
endpoints in R. Let ~v1 , ~v2 be the sequences of edge labels on these two paths. Update
y 0 ← ROUND T WO PATHS(y 0 , ~v1 , ~v2 , ~c) and go to line 4.
• The remaining case: all edges in H form a single path with both endpoints in
R. Let (v1 , v2 , . . . , v2` ) be the sequence of edge labels on this path. Let Y ←
ROUND F INAL PATH(y 0 , ~v ) and exit the loop.
return S = {i ∈ V : Yi = 1}.
Proposition 5.3. The polytyope PFRMatCenter is non-empty.
Proof. The proof is similar to Proposition 3.1, and is omitted.
We now explain how to implement the rounding steps in this algorithm. These are all based on the
subroutine ROUND S INGLE P OINT, which moves the vector y as far as possible along the direction of
vector ~r until it a new face of the polytope begins tight. We define it formally as Algorithm 11; note that
it can be implemented efficiently by solving an appropriate LP.
Algorithm 11 ROUND S INGLE P OINT (y, ~r)
Select δ ∗ ≥ 0 to be maximal such that the vector z + δ ∗ r is in PM .
2: return (y + δ ∗~
r, δ ∗ ).
1:
16
Algorithm 12 ROUND C YCLE (y 0 , ~v )
Initialize ~r = ~0, then set rvj = (−1)j for j = 1, 2, . . . , |~v |
2: (y1 , δ1 ) ←ROUND S INGLE P OINT(y 0 , ~
r)
3: return y1
1:
+
yv1
1
yv2
1
1
yv6
yv3
+
yv4
1
1
yv5
+
1
Figure 2: The left part shows a cycle. The right part shows how the variables on the cycle are being
changed by ROUND C YCLE.
Algorithm 13 ROUND S INGLE PATH (y 0 , ~v )
Initialize ~r = ~0, then set rvj = (−1)j+1 for j = 1, 2, . . . , |~v |
2: (y1 , δ1 ) ←ROUND S INGLE P OINT(y 0 , ~
r)
3: return y1
1:
+
yv1
1
yv2
1
yv3
+
yv4
1
+
yv5
O0
1
1
O0
Figure 3: The left part shows a single path. The right part shows how the variables on the path are being
changed by ROUND S INGLE PATH.
17
Algorithm 14 ROUND T WO PATHS (y 0 , ~v , ~v 0 , ~c)
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
WLOG, suppose j1 , j2 ∈ R are endpoints of v1 , v2` of the path ~v respectively and cj1 ≥ cj2
0 of the path ~
WLOG, suppose j10 , j20 ∈ R are endpoints of v10 , v2`
v 0 respectively and cj10 ≥ cj20
0
∆1 ← cj1 − cj2 ; ∆2 ← cj10 − cj20 ; ~r ← ~0
V1+ ← {v1 , v3 , . . . , v2`−1 }; V1− ← {v2 , v4 , . . . , v2` }
0 }; V − ← {v 0 , v 0 , . . . , v 0
V2+ ← {v20 , v40 , . . . , v2`
0
1 3
2
2`0 −1 }
+
for each v ∈ V1 : rv ← rv + 1; for each v ∈ V1− : rv ← rv − 1
for each v ∈ V2+ : rv ← rv + ∆1 /∆2 ; for each v ∈ V2− : rv ← rv − ∆1 /∆2
(y1 , δ1 ) ←ROUND S INGLE P OINT(y 0 , ~r)
(y2 , δ2 ) ←ROUND S INGLE P OINT(y 0 , −~r)
With probability δ1 /(δ1 + δ2 ): return y2
With remaining probability δ2 /(δ1 + δ2 ): return y1
j1
yv1
yv2
y
+
j1
2
1
v10
+
1
+
+
1
1
j10
yv3 , yv20
2
j2
1
j20
+
+
1
1
2
2
+
yv30
2
j10
1
yv4
j1
1
1
j2
2
2
2
1
2
j10
1
2
2
+
2
1
2
j2
yv40
1
1
2
j20
2
1
j20
2
Figure 4: The left part shows an example of two distinct maximal paths chosen in Case 3. The black
edge is common in both paths. The middle and right parts are two possibilities of rounding y. With
probability δ1 /(δ1 + δ2 ), the strategy in the right part is adopted. Otherwise, the strategy in the middle
part is chosen.
Algorithm 15 ROUND F INAL PATH (y, ~v )
1: P1 ← z ∈ [0, 1]V : z(U ) ≤ rM (U ) ∀U ⊆ V ∧ z(Oi ) = bOi ∀i ∈ L \ {0} ∧ zi = 0 ∀i : yi = 0
2: P2 ← {z ∈ [0, 1]V : z(Fj ) = y(Fj ) ∀j ∈ V 0 \ J ∧ z(Fj ) ≤ 1 ∀j ∈ J}, where J ⊆ R is the set of
vertices in R on the path ~v .
3: Pick an arbitrary extreme point ŷ of P 0 = P1 ∩ P2
4: for each j ∈ R and j is on the path ~
v : if ŷ(Fj ) = 0, pick an arbitrary u ∈ Fj and set ŷu ← 1.
5: return ŷ
18
5.2.2
Analysis of P SEUDO FRMC ENTER ROUND
Proposition 5.4. In all but the last iteration, the while-loop (lines 4 to 8) of P SEUDO FRMC ENTER ROUND preserves the following invariant: if y 0 lies in the face D of PM (w.r.t all tight matroid rank
constraints) at the beginning of the iteration, then y 0 ∈ D at the end of this iteration.
Proof. Observe that y 0 ∈ PM at the beginning of the first iteration due to the definition of y 0 . Fix any
iteration. Let y 00 be the updated y 0 at the end of the iteration. By Corollary 2.4, it suffices to show that
y 00 ∈ {x ∈ Rn : x(S) = bS ∀S ∈ O; xi = 0 ∀i ∈ J; x ∈ PM } ,
where J ⊆ V is the set of all vertices i with yi0 = 0. Note that y 00 is the output of one of the three subroutines ROUND C YCLE, ROUND S INGLE PATH, and ROUND T WO PATHS. Since we only round floating
variables strictly greater than zero, we have yi00 = 0 for all i ∈ J. Also, the procedure ROUND S INGLE P OINT guarantees that y 00 ∈ PM .
• When calling the procedure ROUND C YCLE, observe that each vertex j ∈ L on the cycle is adjacent to exactly two edges. By construction, we always increase the variable on one edge and
decrease the variable on the other edge at the same rate. See Figure 2. Therefore, y 00 (Oj ) =
y 0 (Oj ) = bOj for all j ∈ O.
• When calling the procedure ROUND S INGLE PATH, recall that our path is maximal and has one
endpoint in L and the other in R. We claim that the left endpoint of this path should be corresponding to the set O0 . Otherwise, suppose it is some set Oj with j > 0. We have the tight
constraint y 0 (Oj ) = bOj ∈ Z+ . Then the degree of the vertex j must be at least 2 as there must be
at least two fractional variables in this set. This contradicts to the fact that our path is maximal.
See Figure 3. By the same argument as before, we have y 00 (Oj ) = y 0 (Oj ) = bOj for all j ∈ O.
• In the procedure ROUND T WO PATHS, we round the variables on two paths which have both endpoints in R. Thus, any vertex j should be adjacent to either 2 or 4 edges. Again, by construction,
the net change in y 0 (Oj ) is equal to zero. See Figure 4.
Finally, the claim follows by induction.
Proposition 5.5. P SEUDO FRMC ENTER ROUND terminates in polynomial time.
Proof. Note that, in each iteration, each floating variable yv0 ∈ (0, 1) corresponds to exactly one edge in
the bipartite graph. This is because, by construction, the sets O0 , . . . , Ot form a partition of V and the
sets in F and F0 also form a partition of V . Thus, as long as there are fractional values in y 0 , our graph
will have some cycle or path.
Now we will show that the while-loop (lines 4 to 8) terminates after O(|V |) iterations. For any set
S, let χ(S) denote the characteristic vector of S. That is, χ(v) = 1 for v ∈ S and χ(v) = 0 otherwise.
Let us fix any iteration and let T = {χ(S) : S ⊆ V ∧ y 0 (S) = rM (S)} be the set of all tight
constraints. In this iteration, we will move y 0 along some direction ~r as far as possible (by procedure
ROUND S INGLE P OINT). It means that the new point y 00 = y 0 + δ ∗~r will either have at least one more
rounded variable or hit a new tight constraint y 00 (S0 ) = rM (S0 ) (while y 0 (S0 ) < rM (S0 )) for some
S0 ⊆ V . Indeed χ(S0 ) is linearly independent of all vectors in T .
Proposition 5.4 says that all the tight constraints are preserved in the rounding process. Therefore,
in the next iteration, we either have at least one more rounded variable or the rank of T is increased by
at least 1. This implies the algorithm terminates after at most |V | iterations.
Proposition 5.6. In all iterations, the while-loop (lines 4 to 8) of P SEUDO FRMC ENTER ROUND satisfies the invariant that y 0 (Fj ) ≤ 1 for all Fj ∈ F.
19
Proof. By constraints 2 and 3, this property is true at the beginning of the first iteration. By a very
similar argument as in the proof of Proposition 5.4, this is also true during all but the last iteration. (Note
that if j is an endpoint of a path, then j must be adjacent to exactly one fractional value yv0 , which could
be rounded to one, while other variables {yv0 0 : v 0 ∈ Fj , v 0 6= v} are already rounded to zero as our path
is maximal.) Finally, it is not hard to check that procedure ROUND F INAL PATH also does not violate this
invariant.
Proposition 5.7. P SEUDO FRMC ENTER ROUND returns a solution S which is some independent set of
M plus (at most) one extra vertex in V .
Proof. Let us focus on the procedure ROUND F INAL PATH. Recall that the polytope P 0 in ROUND F I NAL PATH is the intersection of the following two polytopes:
P1 = z ∈ [0, 1]V : z(U ) ≤ rM (U ) ∀U ⊆ V ∧ z(Oi ) = bOi ∀i ∈ L \ {0} ∧ zi = 0 ∀i : yi = 0 ,
and
P2 = {z ∈ [0, 1]V : z(Fj ) = y(Fj ) ∀j ∈ V 0 \ J ∧ z(Fj ) ≤ 1 ∀j ∈ J},
where J ⊆ R is the set of vertices in R on the path ~v .
By construction, P1 is the face of the matroid base polytope PM corresponding to all tight constraints of y. It is well-known that P1 itself is also a matroid base polytope. By Propositions 5.4 and
5.6, we have y ∈ P1 and y ∈ P2 . Thus, y ∈ P which implies that P 6= ∅. Moreover, P2 is a partition
matroid polytope. (Observe that z(Fj ) = y(Fj ) ∈ {0, 1} ∀j ∈ V 0 \ J since all fractional variables are
on the path ~v .) Therefore, P = P1 ∩ P2 has integral extreme points and the point ŷ chosen in line 3 is
integral.
Finally, recall that ~v = (v1 , v2 , . . . , v2` ) is a simple path with both endpoints in R. The constraints
of P1 and integrality of bOi ’s ensure that ŷv1 + ŷv2 = 1, ŷv3 + ŷv4 = 1, . . . , ŷv2`−1 + ŷv2` = 1. In other
words, every vertex i ∈ L on the path will be “matched” with exactly one vertex in R. Thus, there can
be at most one vertex j ∈ R on the path such that ŷ(Fj ) = 0 in line 4. Opening u ∈ Fj adds one extra
facility to our solution.
Recall that C is the (random) set of all clients within radius 3R from some center in S, where R is
the optimal radius. The following two propositions will conclude our analysis.
Proposition 5.8. |C| ≥ t with probability one.
P
P
Proof. Let f denote the function defined in Algorithm RMC ENTER ROUND (i.e., f (z) = j∈V 0 i∈Fi zi
for any z ∈ [0, 1]V .) Using a similar argument as in the proof of Proposition 5.1, one can easily verify
that there are at least f (Y ) vertices in V that are within radius 3R from some open center in S. Next, it
suffices to show that f (Y ) ≥ t.
By definition of y 0 in lines 2 and 3, we have f (y 0 ) ≥ t (see the proof of Proposition 5.2.) We now
claim that f (y 0 ) is not decreasing after each iteration of the rounding scheme. We check the following
cases:
• Case y 0 is rounded by ROUND C YCLE: observe that y 0 (Fj ) is preserved for all j ∈ R since j is
adjacent to two edges and we increase/decrease the corresponding variables by the same amount.
Thus, f (y 0 ) is unchanged.
• Case y 0 is rounded by ROUND S INGLE PATH: if j ∈ R is not the endpoint of the path then j is
adjacent to two edges on the path and y 0 (Fj ) is unchanged. If j is the endpoint, then we increase
the variable on the adjacent edge; and hence, y 0 (Fj ) will increase. See Figure 3.
• Case y 0 is rounded by ROUND T WO PATHS: again, for any j ∈ R \ {j1 , j2 , j10 , j20 }, the value of
y 0 (Fj ) remains unchanged in the process. We now verify the change in f caused by the four
20
endpoints j1 , j2 , j10 , and j20 . Suppose y1 is returned, the contribution of these points in f (y1 ) is
cj1 y1 (Fj1 ) + cj2 y1 (Fj2 ) + cj10 y1 (Fj10 ) + cj20 y1 (Fj20 )
∆1
∆1
0
0
0
0
0
0
= cj1 (y (Fj1 ) + δ1 ) + cj2 (y (Fj2 ) − δ1 ) + cj1 y(Fj1 ) − δ1
+ cj2 y(Fj2 ) + δ1
∆2
∆2
∆1
= cj1 y 0 (Fj1 ) + cj2 y 0 (Fj2 ) + cj10 y 0 (Fj10 ) + cj20 y 0 (Fj20 ) + δ1 (cj1 − cj2 ) + δ1
(c 0 − cj10 )
∆ 2 j2
= cj1 y 0 (Fj1 ) + cj2 y 0 (Fj2 ) + cj10 y 0 (Fj10 ) + cj20 y 0 (Fj20 ).
Hence, f (y1 ) = f (y 0 ). Similarly, one can verify that f (y2 ) = f (y 0 ).
• Case y 0 is rounded by ROUND F INAL PATH: we have shown in the proof of Proposition 5.7 that
y 0 (Fj ) = 1 for all j ∈ J where J is the set of vertices in R on the path ~v . This fact and the other
constraints of P2 ensure that y 0 (Fj ) is not decreasing for all j ∈ V 0 .
Proposition 5.9. Pr[j ∈ C] ≥ pj for all j ∈ V .
Proof. Let y 0 be the vector defined as in lines 2 and 3 of P SEUDO FRMC ENTER ROUND. It suffices to
show that, for all j ∈ V 0 , Pr[Y (Fj ) = 1] ≥ y 0 (Fj ). (Note that y 0 (Fj ) ≥ pj by constraint (5).) This
is because, for any vertex k ∈ V \ V 0P
, the algorithmPRF ILTERING guarantees that there exists j ∈ V 0
0
such that Fk ∩ Fj 6= ∅, and y (Fj ) = i∈Bj xij ≥ i∈V :d(i,k)≤R xik = y 0 (Fk ). Notice that the event
Y (Fj ) = 1 means there is some open center Fj and the distance from k to this center should be at most
3R. Thus,
Pr[k ∈ C] ≥ Pr[Y (Fj ) = 1] ≥ y 0 (Fj ) ≥ y 0 (Fk ) ≥ pk ,
by constraint (5).
Fix any j ∈ V 0 . Recall that Y is obtained by rounding y 0 and, by Proposition 5.6 and the proof of
5.7, we have Y (Fj ) ∈ {0, 1} and Pr[Y (Fj ) = 1] = E[Y (Fj )]. We now show that the expected value of
y 0 (Fj ) does not decrease after each iteration of the while-loop.
• Case y 0 is rounded by ROUND C YCLE: y 0 (Fj ) is unchanged as before.
• Case y 0 is rounded by ROUND S INGLE PATH: if j is not the endpoint of ~v then y 0 (Fj ) is unchanged.
Otherwise, y 0 (Fj ) is increase by some δ1 > 0 with probability one.
• Case y 0 is rounded by ROUND T WO PATHS: again, if j ∈
/ {j1 , j2 , j10 , j20 } then y 0 (Fj ) is unchanged.
Now suppose j = j1 . With probability δ1 /(δ1 + δ2 ), y 0 (Fj1 ) is increase by δ2 , and, with the
remaining probability, it is decreased by δ1 . Thus, the expected change in y 0 (Fj1 ) is
δ1
δ2
(δ2 ) +
(−δ1 ) = 0.
δ1 + δ2
δ1 + δ2
Similarly, one can verify that the expected values of y 0 (Fj2 ), y 0 (Fj10 ), and y 0 (Fj20 ) remain the same.
• Case y 0 is rounded by ROUND F INAL PATH: we have showed in the proof of Proposition 5.7 that if
j is on the path ~v , then Y (Fj ) = 1. Otherwise, the constraints of P2 ensure that Y (Fj ) = y 0 (Fj ).
So far we have proved the following theorem.
Theorem 5.10. P SEUDO FRMC ENTER ROUND will return a random solution S such that
• S is the union of some basis of M with (at most) one extra vertex,
• |C| ≥ t with probability one,
• Pr[j ∈ C] ≥ pj for all j ∈ V .
21
5.2.3
An algorithm satisfying the matroid constraint exactly
Using a similar technique as in Section 4.2.3, we will develop an approximation algorithm for the FRMatCenter problem which always returns a feasible solution. Let > 0 a small parameter to be
determined. Let U denote the collection of all possible sets of verticies with size at most d1/e such that
U is an independent set of M. Again, we have |U| ≤ nO(1/) . Suppose R is the optimal radius to our
instance. For any i ∈ V , recall that RBall(i, U, R) is the set of red vertices within radius 3R from i.
Consider the configuration polytope Pconfig3 containing points (x, y, q) with the following constraints:
P
qU = 1
PU ∈U U
x ≤ qU
∀j ∈ V, U ∈ U
Pi∈Bj Pij
U
∀j ∈ V
U ∈U
i∈Bj xij ≥ pj
U
U
≤ yi
∀i, j ∈ V, U ∈ U
x
Pij
U
y ≤ qU rM (W )
∀U ∈ U, W ⊆ V
Pi∈W Pi
U
j∈V
i∈Bj xij ≥ qU t
yiU = 1
∀U ∈ U, i ∈ U
U
yi = 0
∀U ∈ U, i ∈ V \ U, |RBall(i, U, R)| ≥ n
U U
xij , yi , qU ≥ 0
∀i, j ∈ V, U ∈ U
We first claim that Pconfig3 is a valid relaxation for the problem.
Proposition 5.11. The polytope Pconfig3 is non-empty.
Proof. Suppose S is a solution drawn from the optimal distribution D. We compute a subset US of S
using a similar procedure as in the proof of Proposition 4.4. Recall that |RBall(i, US , R)| < n for all
i ∈ S \ US and |US | ≤ d1/e. Since US ⊆ S, US is also an independent set of M, implying that
US ∈ U.
Now for any U ∈ U, we set qU := Pr[US = U ]. Let xU
ij be probability of the joint event: US = U
U be the probability of the joint event: U = U and i ∈ S. Then it
and j is connected
to
i.
Finally,
let
y
S
i
P
is clear that U ∈U qU = 1. Using similar arguments to the proofs of Propositions 4.4 and 4.3, we have
the following inequalities:
X
xU
∀j ∈ V, U ∈ U
(9)
ij ≤ qU ,
i∈Bj
X X
U ∈U i∈Bj
XX
j∈V i∈Bj
xU
ij ≥ pj ,
∀j ∈ V
(10)
xU
ij ≥ qU t,
(11)
yiU = 0 ∀U ∈ U, i ∈ V \ U, |RBall(i, U, R)| ≥ n.
(12)
Recall that yiU /qU is the Pr[i ∈ S | U = US ]. Since S is independent with probability one, we have
|S ∩ W | ≤ rM (W ) for all W ⊆ V . Therefore,
X
X
rM (W ) ≥ E[|S ∩ W | | U = US ] =
Pr[i ∈ S | U = US ] =
yiU /qU ,
i∈W
i∈W
for all W ⊆ V .
The other constraints can be verified easily. We conclude that (x, y, q) ∈ Pconfig3 .
Next, let us pick any (x, y, q) ∈ Pconfig3 and use the following algorithm to round it.
22
Algorithm 16 FRMC ENTER ROUND (x, y, q)
1:
2:
3:
4:
5:
Randomly pick a set U ∈ U with probability qU
0
U
Let x0ij ← xU
ij /qU and yi ← min{yi /qU , 1}
0
S ← P SEUDO FRMC ENTER ROUND (x0 , y 0 )
Let i∗ be the “extra” vertex in S 0 .
return S = S 0 \ {i}
Analysis. We are now ready to prove the second part of Theorem 1.4. Let us fix any γ > 0 and set
:= γ 2 . Also, let E(U ) be the event that Algorithm 16 selects U ∈ U. Note that (x0 , y 0 ) satisfies the
following inequalities:
XX
x0ij ≥ t,
j∈V i∈Bj
X
i∈Bj
X
x0ij ≤ 1,
x0ij =
i∈Bj
∀j ∈ V,
X
xij /qU ,
i∈Bj
∀j ∈ V,
x0ij ≤ yi0 , ∀i, j ∈ V,
X
yi0 ≤ rM (W ), ∀W ⊆ V.
i∈W
Moreover, every i ∈ U satisfies yi0 = 1; every i ∈ V \ U with RBall(i, U, R) ≥ n satisfies yi0 = 0.
Recall that the algorithm P SEUDO FRMC ENTER ROUND will return a solution S 0 , which is the union
of a basis of M with an extra center i∗ . The vertex i∗ has yi∗ ∈ (0, 1), and in particular i∗ ∈
/ U . By
removing i∗ from S 0 , we ensure that the resulting set is a basis of M with probability one.
Now we shall prove the coverage guarantee. By Theorem 5.10, S 0 covers at least t vertices within
radius 3R. Any blue vertex can be connected to some center in U and hence is not affected by the
removal of i∗ . Because i∗ covers at most n other red vertices, we have
|C| ≥ t − n = 1 − γ 2 n.
For any j ∈ V , let Xj be the indicator random variable for the event that d(j, S 0 ) ≤ 3R but
0
∗
d(j,
P S \ {i }) > 3R. We say that j is a bad vertex iff E[Xj ] ≥ γ, otherwise, j is good. Again,
j∈V Xj ≤ n with probability one. Thus, there can be at most n/γ bad vertices. Letting T be the set
of good vertices, we have
|T | ≥ n − n/γ = (1 − γ)n.
By Theorem 5.10, Pr[j is covered by S 0 ] ≥ pj . So, for any j ∈ T , we have
Pr[j ∈ C] ≥ Pr[j is covered by S 0 ] − Pr[Xj = 1] ≥ pj − γ.
23
References
[1] Sara Ahmadian, Ashkan Norouzi-Fard, Ola Svensson, and Justin Ward. Better guarantees for
k-means and euclidean k-median by primal-dual algorithms. CoRR, abs/1612.07925, 2016.
[2] Jaroslaw Byrka, Thomas Pensyl, Bartosz Rybicki, Aravind Srinivasan, and Khoa Trinh. An improved approximation for k-median, and positive correlation in budgeted optimization. In Proceedings of the Annual ACM-SIAM Symposium on Discrete Algorithms, (SODA), pages 737–756,
2015.
[3] Deeparnab Chakrabarty, Prachi Goyal, and Ravishankar Krishnaswamy. The non-uniform k-center
problem. In 43rd International Colloquium on Automata, Languages, and Programming (ICALP
2016), volume 55, pages 67:1–67:15, 2016.
[4] Moses Charikar, Samir Khuller, David M. Mount, and Giri Narasimhan. Algorithms for facility
location problems with outliers. In Proceedings of the Twelfth Annual ACM-SIAM Symposium
on Discrete Algorithms, SODA ’01, pages 642–651, Philadelphia, PA, USA, 2001. Society for
Industrial and Applied Mathematics.
[5] Danny Z. Chen, Jian Li, Hongyu Liang, and Haitao Wang. Matroid and knapsack center problems. Integer Programming and Combinatorial Optimization: 16th International Conference,
IPCO 2013, Valparaı́so, Chile, March 18-20, 2013. Proceedings, pages 110–122, 2013.
[6] David G. Harris, Thomas Pensyl, Aravind Srinivasan, and Khoa Trinh. Symmetric randomized
dependent rounding. CoRR, abs/1709.06995, 2017.
[7] Dorit S. Hochbaum and David B. Shmoys. A unified approach to approximation algorithms for
bottleneck problems. J. ACM, 33(3):533–550, May 1986.
[8] Wen-Lian Hsu and George L. Nemhauser. Easy and hard bottleneck location problems. Discrete
Applied Mathematics, 1(3):209 – 215, 1979.
[9] Tapas Kanungo, David M. Mount, Nathan S. Netanyahu, Christine D. Piatko, Ruth Silverman, and
Angela Y. Wu. A local search approximation algorithm for k-means clustering. Comput. Geom.,
28(2-3):89–112, 2004.
[10] Ravishankar Krishnaswamy, Amit Kumar, Viswanath Nagarajan, Yogish Sabharwal, and Barna
Saha. The matroid median problem. In Proceedings of the annual ACM-SIAM Symposium on
Discrete Algorithms (SODA), pages 1117–1130. SIAM, 2011.
[11] Lap-Chi Lau, R. Ravi, and Mohit Singh. Iterative Methods in Combinatorial Optimization. Cambridge University Press, New York, NY, USA, 1st edition, 2011.
[12] Shi Li and Ola Svensson. Approximating k-median via pseudo-approximation. In STOC, pages
901–910, 2013.
[13] Alexander Schrijver. Combinatorial optimization: polyhedra and efficiency, volume 24. Springer
Science & Business Media, 2003.
[14] Chaitanya Swamy. Improved approximation algorithms for matroid and knapsack median problems and applications. In APPROX/RANDOM 2014, volume 28, pages 403–418, 2014.
[15] David P. Williamson and David B. Shmoys. The Design of Approximation Algorithms. Cambridge
University Press, 2011.
24
| 8cs.DS
|
arXiv:1304.5144v2 [math.GR] 20 Jun 2016
Locally normal subgroups
of totally disconnected groups.
Part I: General theory
Pierre-Emmanuel Caprace∗1 , Colin D. Reid†2 , and George A. Willis‡2
1
Université catholique de Louvain, IRMP, Chemin du Cyclotron 2, bte L7.01.02,
1348 Louvain-la-Neuve, Belgique
2
School of Mathematical and Physical Sciences, University of Newcastle,
Callaghan, NSW 2308, Australia
April 22, 2015
Abstract
Let G be a totally disconnected, locally compact group. A closed subgroup of
G is locally normal if its normaliser is open in G. We begin an investigation
of the structure of the family of closed locally normal subgroups of G. Modulo
commensurability, this family forms a modular lattice LN (G), called the structure
lattice of G. We show that G admits a canonical maximal quotient H for which the
quasi-centre and the abelian locally normal subgroups are trivial. In this situation
LN (H) has a canonical subset called the centraliser lattice, forming a Boolean
algebra whose elements correspond to centralisers of locally normal subgroups. If
H is second-countable and acts faithfully on its centraliser lattice, we show that the
topology of H is determined by its algebraic structure (and thus invariant by every
abstract group automomorphism), and also that the action on the Stone space of
the centraliser lattice is universal for a class of actions on profinite spaces. Most of
the material is developed in the more general framework of Hecke pairs.
Contents
1 Introduction
3
∗
F.R.S.-FNRS research associate, supported in part by the ERC (grant #278469)
Supported in part by ARC Discovery Project DP120100996
‡
Supported in part by ARC Discovery Project DP0984342
†
1
2 Locally normal subgroups
2.1 Definition and first properties . . . . . . . . . . . . . . . . . . . . . . . . .
2.2 Local equivalence and the structure lattice . . . . . . . . . . . . . . . . . .
3 Quasi-centralisers and their
3.1 Motivation . . . . . . . .
3.2 The quasi-centre . . . . .
3.3 Quasi-centralisers . . . . .
3.4 C-stability . . . . . . . . .
3.5 Local C-stability . . . . .
stability properties
. . . . . . . . . . . . .
. . . . . . . . . . . . .
. . . . . . . . . . . . .
. . . . . . . . . . . . .
. . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
8
8
9
13
13
14
16
19
23
4 Direct decomposition of locally normal subgroups
26
4.1 Local decomposition lattices . . . . . . . . . . . . . . . . . . . . . . . . . . 26
4.2 Locally finitely decomposable t.d.l.c. groups . . . . . . . . . . . . . . . . . 31
5 The
5.1
5.2
5.3
5.4
centraliser lattice and weakly decomposable actions
The centraliser lattice . . . . . . . . . . . . . . . . . . . . .
Weakly decomposable actions . . . . . . . . . . . . . . . . .
Faithful weakly decomposable groups . . . . . . . . . . . . .
Topologically free actions . . . . . . . . . . . . . . . . . . .
6 Some radicals of locally compact groups
6.1 The quasi-hypercentre . . . . . . . . . . . . . . .
6.2 Relation with the quasi-hypercentre of subgroups
6.3 Other radicals . . . . . . . . . . . . . . . . . . . .
6.4 Canonical completions of Hecke pairs . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
33
33
39
42
47
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
48
48
49
53
60
7 Restrictions on the group topology
62
7.1 The structure lattice and refinements of the group topology . . . . . . . . 62
7.2 A group topology derived from the centraliser lattice . . . . . . . . . . . . 63
7.3 Subgroups abstractly generated by compact locally normal subgroups . . 66
Bibliography
69
Index
71
2
1
Introduction
The aim of the present article is to establish foundations for a study of general nondiscrete totally disconnected locally compact (t.d.l.c.) groups in terms of their local
structure. We define a canonical lattice in terms of this local structure that, in the case
of a p-adic Lie group, is equivalent to the lattice of ideals in the Lie algebra. Inspired by
earlier work of J. Wilson [27] on just-infinite groups and by Barnea–Ershov–Weigel [2]
on abstract commensurators of profinite groups, we develop properties of this lattice.
This article, which concerns general t.d.l.c. groups, is the first in a series. The second
paper in the series [8] uses the tools developed here to study the more special class of
compactly generated t.d.l.c. groups with few normal subgroups and, in particular, those
that are topologically simple.
Most of the definitions and results in this first part apply more generally to a Hecke
pair (G, U ), that is a group G with a commensurated subgroup U , in a way that depends
only on the (abstract) commensurability class of U as a subgroup of G. Although our
principal interest is in t.d.l.c. groups, we will state and prove results in this more general
framework in situations where it is unnecessary to use the topology of the group. For
example, in the case G = U , most of the results we obtain will be valid for residually
finite groups in general, with profinite groups as a special case.
1.1
Background
Since the solution of Hilbert’s fifth problem by Gleason, Montgomery–Zippin and Yamabe, connected locally compact groups are known to have a far-reaching structure theory,
ensuring notably that they are all projective limits of connected Lie groups. This deep
result extends the powerful machinery of Lie theory to all connected locally compact
groups, thereby yielding decomposition theorems and classification results.
In contrast, the solution of Hilbert’s fifth problem yields no information about totally
disconnected locally compact groups, that is, those in which the connected component
of each element is a singleton. Understanding this case would go a long way towards an
understanding of general locally compact groups because each general group is an extension of a totally disconnected group by a connected one. While a rather rich and deep
theory was developed for some special classes of t.d.l.c. groups (profinite groups, reductive groups over local fields, p-adic analytic groups, . . . ), for a long time the only general
result known was van Dantzig’s theorem, which asserts that compact open subgroups
exist and form a basis of identity neighbourhoods. Since the class of t.d.l.c. groups
includes all discrete groups, it seemed during this time that no meaningful structure
theory could be developed beyond the connected or almost connected case.
Three developments over the last twenty years have changed the situation. First, new
general results were obtained with the aid of the scale function, introduced by Willis
3
in [26]. The scale function on the t.d.l.c. group G measures the effect on compact
open subgroups of conjugation by elements of G; it takes positive-integer values and is
non-trivial only when G is not discrete. The scale function and associated ideas elicit
significant new features of general non-discrete t.d.l.c. groups such as tidy subgroups,
contraction groups, parabolic subgroups, etc. and have resolved several open questions.
A complementary approach was initiated by Barnea–Ershov–Weigel [2], based on the
following observation: since any two compact open subgroups of a t.d.l.c. group G are
commensurate, G admits a natural homomorphism to the group of abstract commensurators of its compact open subgroups. By definition, abstract commensurators can be
recovered from any compact open subgroup and are thus purely local objects depending
only on arbitrarily small identity neighbourhoods of G. This observation thus produces
a local approach to t.d.l.c. groups.
The third development is that techniques for isolating the non-discrete case from the
discrete one have been found. At the global level, Caprace–Monod [7] identified simple
pieces of a t.d.l.c. group G under the hypothesis that G has no discrete quotients while,
at the local level, Burger–Mozes [6] introduced, in analogy with the kernel of the adjoint
representation of a Lie group, the quasi-centre of G, which contains in particular any
discrete normal subgroup of G. The former of the two motivates the focus of our next
paper [8] on simple groups, while the latter was important in the approach introduced
in [2] and is developed further in Sections 3 and 6 below as part of the foundation for a
general theory of the local structure of t.d.l.c. groups.
1.2
Locally normal subgroups and their centralisers
The central concept in the general local theory is that of a locally normal subgroup
of a t.d.l.c. group, that is, a subgroup whose normaliser has finite index in an open
subgroup. A major part of [2] focuses on t.d.l.c. groups in which every open compact
subgroup is (hereditarily) just infinite; this corresponds exactly to the condition that
every non-trivial compact locally normal subgroup is open. Away from this case, the
arrangement of (closed) locally normal subgroups becomes more complex.
The intersection of two locally normal subgroups is locally normal. Moreover, if H and
K are closed locally normal subgroups of G, then they admit respective compact open
subgroups H ′ < H and K ′ < K which normalise each other, and such that the product
H ′ K ′ , which is also a compact subgroup, is locally normal. It is thus natural to consider
the collection of all compact locally normal subgroups modulo the equivalence relation
given by local equivalence: two subgroups H and K are locally equivalent if their
intersection H ∩K is open in both H and K. In this way, the collection of all equivalence
classes of compact locally normal subgroups of G, which is denoted by LN (G), endowed
with the partial ordering induced by the inclusion of subgroups, is easily seen to form a
modular lattice. We call it the structure lattice of G. There are two elements that we
4
regard as trivial (assuming G is non-discrete): the minimum element, which is the class
of the trivial subgroup, and the maximum element, which is the class of compact open
subgroups. The symbols 0 and ∞ will be used to denote the minimum and maximum
elements respectively of any bounded lattice, including LN (G). A non-trivial subset of
a bounded lattice is one that contains elements other than 0 and ∞.
Its structure lattice is a local invariant of a t.d.l.c. group G, since it is determined by any
open compact subgroup, which admits a natural action of the topological automorphism
group of G. In case G is a p-adic Lie group, the structure lattice LN (G) is canonically isomorphic to the lattice of ideals in the Qp -Lie algebra of G (see Proposition 2.5
below). Hence the structure lattice may be viewed as a group theoretic analogue for
arbitrary t.d.l.c. groups of the lattice of ideals in the p-adic case. Although it is difficult to determine the structure lattice of a particular t.d.l.c. group G explicitly, the
action of G on LN (G) provides an important tool for elucidating the overall structure
of G. Moreover, under the right conditions we obtain two natural G-invariant subsets
of LN (G) that are Boolean algebras when equipped with the induced order and, when
non-trivial, constitute valuable additional tools. The first is the local decomposition
lattice LD(G), which consists of all direct factors of open compact subgroups of G up
to equivalence. The second is the centraliser lattice LC(G): an element of LC(G) is
given by the equivalence class of the centraliser CU (K) for some open compact subgroup
U of G and closed normal subgroup K of U . In general, the structure lattice depends on
the topology of the open compact subgroups of G, but the centraliser and decomposition
lattices depend only on the Hecke pair structure (G, U ), where U is any open compact
subgroup of G.
Following Burger–Mozes [6], we define the quasi-centre QZ(G) of a locally compact
group G to be the collection of all elements whose centraliser is open (see Lemma 3.1
below for its basic properties).
Theorem I. Let G be a t.d.l.c. group.
(i) (See Theorem 4.5) Suppose that G has trivial quasi-centre. Then LD(G) is a
Boolean algebra.
(ii) (See Theorem 5.2) Suppose in addition that G has no non-trivial abelian locally
normal subgroups. Then LC(G) is a Boolean algebra and LD(G) is a subalgebra of
LC(G).
This opens up a new perspective: using the Stone representation theorem, a group action
on a Boolean algebra A can equivalently be regarded as a group action by homeomorphisms on a topological space S(A), called the Stone space of A, that is profinite, in
other words, compact and totally disconnected. An important property of the G-actions
on the Stone spaces of LD(G) and LC(G) is encapsulated in the following definition.
Let G be a locally compact group and X be a profinite space on which G acts faithfully
5
by homeomorphisms. The G-action is called weakly decomposable if for every nonempty clopen subset α ( X, the stabiliser of α in G is open, and the pointwise stabiliser
of α is non-trivial. Say that G is faithful weakly decomposable if it admits a faithful
weakly decomposable action on some profinite space.
Theorem II (See Theorem 5.18). Let G be a t.d.l.c. group such that QZ(G) = 1.
(i) Suppose that the only abelian locally normal subgroup of G is trivial, and suppose
that G acts faithfully on LC(G). Then the induced action of G on S(LC(G)) is a
faithful weakly decomposable action.
(ii) Suppose X is a profinite space with a faithful weakly decomposable action of G.
Then the only abelian locally normal subgroup of G is trivial and there exists a
G-equivariant continuous surjective map from S(LC(G)) to X.
In particular, if G is faithful weakly decomposable, then S(LC(G)) is the unique maximal
compact G-space for which the action is faithful and weakly decomposable.
1.3
Radical theory
The hypotheses of Theorem I are satisfied by all compactly generated, topologically
simple t.d.l.c. groups: for (i) this is Theorem 4.8 of [2], and for (ii) this will be proved
in [8, Theorem A]. We will show, by developing a suitable ‘radical theory’, that an
arbitrary t.d.l.c. group has a canonical maximal quotient which satisfies the hypotheses
of Theorem I(i) or (ii).
Theorem III. Let G be a t.d.l.c. group.
Then G has closed characteristic subgroups QZ∞ (G) ≤ R[A] (G) ≤ G with the following
properties:
(i) (See Theorem 6.2) QZ∞ (G) is the unique smallest closed normal subgroup N of G
such that G/N has trivial quasi-centre.
(ii) (See Theorem 6.11) R[A] (G) is the unique smallest closed normal subgroup M of
G such that G/M has trivial quasi-centre and no non-trivial abelian locally normal
subgroups.
In case G is a p-adic Lie group, R[A] (G) is the unique largest closed normal subgroup
whose Lie algebra is the soluble radical of Lie(G); see Proposition 6.18 below.
Both QZ∞ (G) and R[A] (G) are well behaved with respect to open subgroups, in the sense
that QZ∞ (H) = QZ∞ (G)∩H and R[A] (H) = R[A] (G)∩H for any open subgroup H of G
6
(see Propositions 6.3 and 6.15). Thus if the canonical quotients provided by Theorem III
turn out to be trivial, we have additional local information about the structure of G.
For instance it is easily seen that if R[A] (G) = G, then the open compact subgroups of
G are Fitting-regular in the sense of [22].
1.4
Group topologies
We now specialise to the case of first-countable t.d.l.c. groups and give some results on
the relationship between the commensurated compact subgroups and the compact locally
normal subgroups. For instance (Lemma 7.4), a closed subgroup K of a first-countable
profinite group U is commensurated by U if and only if U normalises an open subgroup
of K. Hence, any compact commensurated subgroup of a first-countable t.d.l.c. group
G is commensurate with a locally normal subgroup, whose commensurability class is
thus a G-fixed point in LN (G). As a consequence, the commensurated compact locally
normal subgroups have an interpretation in terms of group topologies.
Theorem IV (See § 7.1). Let G be a first-countable t.d.l.c. group. Then there is a
natural bijection between the elements of LN (G) that are fixed by the action of G, and
refinements of the topology of G that are locally compact and compatible with the group
structure.
When G is faithful weakly decomposable, we obtain a topological rigidity result of a
similar kind to those for semisimple Lie groups over real and p-adic fields given in [17]
and [19].
Theorem V (See Theorem 7.6). Let G be a second-countable t.d.l.c. group with trivial
quasi-centre, and suppose that G is faithful weakly decomposable. Then the topology of
G is the coarsest group topology on G for which it is a Baire space. In particular, it is
preserved by every automorphism of G as an abstract group. Consequently, the topology
of G is the unique σ-compact locally compact group topology on G, and the set of all
locally compact group topologies of G is in natural bijection with the set of fixed points
of G acting on LN (G) by conjugation.
Acknowledgements
We express our gratitude to John Wilson for numerous comments which helped in improving the presentation of this paper. We thank Corina Ciobotaru, David Hume,
Thierry Stulemeijer and Phillip Wesolek for their careful reading of an earlier draft of
the manuscript.
7
2
Locally normal subgroups
2.1
Definition and first properties
Let G be a t.d.l.c. group, which is assumed to be Hausdorff by convention. Then van
Dantzig’s theorem [13, Theorem II.7.7] asserts that the open compact subgroups of G
form a base of identity neighbourhoods. Denote this set by B(G). Any two elements of
B(G) are commensurate, as an elementary topological argument shows. Hence, letting
U be any open compact subgroup of G, the pair (G, U ) satisfies the following condition.
A Hecke pair (of groups) (G, U ) is a group G together with a commensurated subgroup U of G, that is, U ∩ gU g −1 has finite index in U for all g ∈ G.
Most of our arguments will be made in terms of general Hecke pairs, as the results then
apply directly to residually finite groups and not just to their profinite completions.
When specialising to the case that U is an open compact subgroup of the t.d.l.c. group
G, we will say that (G, U ) is a t.d.l.c. Hecke pair.
Given a Hecke pair (G, U ), a locally normal subgroup of (G, U ) is a subgroup K of G
such that the normaliser NU (K) has finite index in U . A subgroup K of G is bounded
if K ∩ U is commensurate with K. When (G, U ) is t.d.l.c., subgroups that are bounded
and closed are precisely the compact ones.
The most elementary properties of locally normal subgroups collected in the following
will frequently be used without comment.
Lemma 2.1. Let (G, U ) be a Hecke pair and K, L be subgroups of G.
(i) If K and L are locally normal, so is K ∩ L.
(ii) If (G, U ) is t.d.l.c. and K is compact and locally normal, so is every open subgroup
of K.
(iii) If K is locally normal, so is the centraliser CV (K) for any V ≤ G commensurate
with U .
(iv) If K is bounded, then there is V ≤ G commensurate with U that contains K. If
K is also locally normal, then V may be chosen to normalise K. When (G, U ) is
t.d.l.c. and K is compact and locally normal, then we can take V ∈ B(G).
Proof. (i) is clear since K ∩ L is normalised by NU (K) ∩ NU (L).
(ii) An open subgroup of K may be written as K ∩ U for some U ∈ B(G). Since any
element of B(G) is locally normal, (ii) follows from part (i).
8
(iii) Let W = U ∩ V . If K is locally normal, then CV (K) is a subgroup of V , so that
NW (K) ≤ NW (CV (K)) and we have
|U : NU (CV (K))| ≤ |U : W ||W : NW (CV (K))| ≤ |U : W ||W : NW (K)| < ∞;
hence CV (K) is locally normal.
T
(iv) If K is bounded, then W = k∈K kU k−1 is commensurate with U and normalised
by K, so that V = KW is a subgroup commensurate with U and containing K. If K
locally normal, then W = NU (K) is commensurate with U and V = KW is a subgroup
commensurate with U and containing K as a normal subgroup. When (G, U ) is t.d.l.c.
and K is compact and locally normal, then W open and compact with either definition,
and so V is an open compact subgroup of G.
Notice that the converse to Lemma 2.1(ii) does not hold: namely, the fact that K is
compact and contains a locally normal subgroup as an open subgroup does not generally
imply that K itself is locally normal, since otherwise all finite subgroups of G would be
locally normal and hence contained in the quasi-centre QZ(G).
2.2
Local equivalence and the structure lattice
Let (G, U ) be a Hecke pair and let H be a subgroup of G. Write [H] for the class of all
subgroups K of G such that K ∩ U is commensurate with H ∩ U and say that H and K
are locally equivalent if [H] = [K]. Although H might not be bounded, and such will
be the case in later arguments, it is the bounded subgroups of [H] that are of interest.
There is a natural partial ordering on local equivalence classes, given by [H] ≥ [K] if
H ∩ K ∩ U has finite index in K ∩ U . Note that, when (G, U ) is t.d.l.c. and H is closed,
the closed subgroups in [H] are those that have an open subgroup in common with H.
A subgroup H in G is locally equivalent to the trivial subgroup if and only if H ∩ U
is finite; in the t.d.l.c. case, this occurs if and only if H is discrete. We shall later be
interested in subgroups H of t.d.l.c. groups G such that H ∩ U = 1 for all U ∈ B(G).
Since every compact (and hence every finite) subgroup of G is contained in some open
compact subgroup, the latter property holds if and only if H is discrete and torsion-free.
For a given a Hecke pair (G, U ), let LN (G, U ) denote the set of local equivalence classes
of subgroups of G that have locally normal representatives. Unless otherwise stated,
when choosing a representative of α ∈ LN (G, U ), we only take representatives that are
bounded and locally normal. Then the partially ordered set LN (G, U ) is a lattice. The
greatest lower bound and least upper bound of α, β ∈ LN (G, U ) are respectively
α ∧ β = [H ∩ K] and α ∨ β = [HK],
where representatives H ∈ α and K ∈ β are chosen such that H, K E V for some V ≤ U
of finite index. That [H ∩ K] and [HK] are independent of how H, K and V are chosen
9
and that they are respectively the greatest lower bound and least upper bound of α and
β is easily verified. We call (LN (G, U ), ≤) the abstract structure lattice of (G, U ).
If (G, U ) is a t.d.l.c. Hecke pair, write LN (G) for the set of elements of LN (G, U ) that
have compact locally normal representatives. Then LN (G) is a sublattice of LN (G, U )
that does not depend on the choice of U so long as it is compact and open. We define
LN (G) with the induced operations to be the structure lattice of G.
It is clear that LN (G, U ) is determined entirely by the group structure of U , so effectively
LN (G, U ) = LN (U, U ). We retain the notation LN (G, U ) in order to distinguish the
abstract (Hecke pair) structure lattice from the topological structure lattice LN (G),
which is more useful in the study of t.d.l.c. groups.
Example 2.2.
(a) Let T be a regular tree of finite degree d, with d ≥ 3, and let G = Aut(T ). Then
G is a t.d.l.c. group, with topology generated by the pointwise stabilisers of finite
subtrees of T . Let F be a finite subtree of T , let U be the pointwise stabiliser of F ,
and let T1 , . . . , Tn be the connected components of T − F . Then U decomposes as a
direct product
U = K1 × K2 × · · · × Kn
where Ki is the pointwise stabiliser in G of all vertices outside Ti . Evidently Ki
is compact and normal in U , so it is locally normal in G, and since each tree Ti
is infinite and G induces an infinite group of automorphisms on it, each group Ki
is infinite. Thus the given decomposition of U provides 2n distinct elements of the
structure lattice of G. By varying the finite tree F , we see that LN (G) is infinite in
this case.
(b) Let G = Qnp for n ≥ 2. Then the every subgroup of G is locally normal, since G is
abelian. The compact subgroups of G are all isomorphic to Zm
p where 0 ≤ m ≤ n. It
is easily seen that two compact subgroups of G are commensurate if and only if they
span the same subspace of G regarded as a Qp -vector space, and conversely every
Qp -subspace of G is spanned by a compact subgroup. Consequently, the structure
lattice of G is naturally isomorphic to the lattice of Qp -subspaces of G. In particular,
LN (G) is uncountable.
(c) At the other extreme, let G be any t.d.l.c. group with a hereditarily just infinite
open subgroup U , for instance G = Qp or G = PSLn (Qp ). Then given a closed
locally normal subgroup K, the intersection K ∩ U is normal in an open subgroup
of U , and hence K is either finite or open. Hence LN (G) is the two-element lattice.
When G is a t.d.l.c. group, the lattice structure of LN (G) is preserved by the topological automorphism group Aut(G) of G. More generally, the structure of LN (G, U ) is
preserved by those automorphisms of G that preserve the commensurability class of U .
10
In particular, G acts on LN (G) (when defined) and LN (G, U ) by conjugation. If U is
infinite, there are two ‘trivial’ elements of LN (G, U ) and of LN (G), namely the class of
the trivial group, denoted by 0, and the class of the subgroups containing a subgroup of
U of finite index, denoted by ∞.
The use of the term structure lattice is borrowed from J. Wilson ([27]), whose ideas
developed in the context of just infinite groups were an important source of inspiration
for the present work. The structure lattice is a purely ‘local’ invariant of a t.d.l.c. group,
that is, it is completely determined by any open compact subgroup. It is also built out
of lattices of normal subgroups of open compact subgroups in a natural sense. Similar
observations apply more generally to the abstract structure lattice.
Lemma 2.3. Let (G, U ) be a Hecke pair.
(i) Let X be a finite subset of LN (G, U ), and let L be the sublattice generated by X.
Then there is a subgroup V of U of finite index and a choice of representatives Y for
the elements of X, so that every element of Y is a normal subgroup of V , and L is
isomorphic to the lattice of subgroups of V generated by Y modulo commensuration.
If L ⊆ LN (G), then we can take V ∈ B(G).
(ii) Let I be a lattice identity that is satisfied by the normal subgroup lattice of V
(modulo commensuration), for every finite index subgroup V of U . Then I is
satisfied by LN (G, U ). In particular, LN (G, U ) is a modular lattice.
T
Proof. (i) Let X = {[K1 ], . . . , [Kn ]}, let V =
NU (Ki ), and let Li = V ∩ Ki for
i ∈ {1, . . . , n}. Then Y = {L1 , . . . , Ln } consists of normal subgroups of V , and the
meet and join operations in L can evidently be given by [Ki ] ∧ [Kj ] = [Li ∩ Lj ] and
[Ki ] ∨ [Kj ] = [Li Lj ] = [hLi , Lj i]. Thus the map L 7→ [L] gives a surjective map from the
lattice generated by Y to L, and two subgroups have the same image if and only if they
are commensurate. If (G, U ) is t.d.l.c. and L ⊆ LN (G), then we can choose K1 , . . . , Kn
to be closed, so V is closed, and hence V ∈ B(G).
(ii) This is immediate from (i), recalling that the lattice of normal subgroups in any
group is modular (see [14, Theorem 8.3]).
If G is a t.d.l.c. group and U is an open compact subgroup, we see that LN (G) is
a sublattice of LN (G, U ). (In particular, LN (G) inherits the modular property from
LN (G, U ).) In general, elements of LN (G) may have non-closed representatives, since
a finite index subgroup of a profinite group is not necessarily closed. However, we can
recognise the elements of LN (G) inside LN (G, U ) by the existence of closed representatives (without reference to normalisers).
Proposition 2.4. Let (G, U ) be a t.d.l.c. Hecke pair and let α ∈ LN (G, U ). Then
α ∈ LN (G), that is, α has a compact locally normal representative, if and only if there
is closed subgroup H of G such that [H] = α.
11
Proof. If α ∈ LN (G), then certainly α has a closed representative. It remains to show the
converse. Suppose H is a closed subgroup of G such that [H] = α. Since α ∈ LN (G, U ),
there is some bounded locally normal subgroup K of G such that [K] = α. We see that
H ∩ U is closed and [H ∩ U ] = α, so we may assume H ≤ U . Now H ∩ K has finite index
in K, so the image of K in U/H is a finite set of cosets, say k1 H, . . . , kn H. Since H is
closed, there is a sufficiently small open normal subgroup V of U such that the cosets
k1 H, . . . , kn H remain pairwise distinct modulo V . It then follows that
H ∩ K ∩ V = K ∩ V.
Note that K ∩ V is locally normal and [K ∩ V ] = α. So by replacing K by K ∩ V , we
may assume K ≤ H. Since H is a closed subgroup of G, we in fact have K ≤ H. Now
K has finite index in K, since K is commensurate with H, so [K] = α; moreover, the
normaliser of K in G is a closed subgroup containing NG (K), so the normaliser of K is
open. Hence K is the desired compact locally normal representative of α.
We have seen that the structure lattice LN (G) of a t.d.l.c. group G is a modular lattice
that can be recovered from any identity neighbourhood in G. In the special case where
G is a p-adic Lie group, the structure lattice can be described in terms of the Qp -Lie
algebra of G, as follows.
Proposition 2.5. Let G be a p-adic Lie group. Then the structure lattice LN (G) is
G-equivariantly isomorphic to the lattice of ideals in the Qp -Lie algebra Lie(G).
Indeed, that proposition readily follows from the following known properties of the Lie
correspondence in p-adic Lie groups.
Proposition 2.6. Let G be a p-adic Lie group, and g = Lie(G) denote its Qp -Lie
algebra.
(i) For any two closed subgroups H1 , H2 ≤ G, we have [H1 ] = [H2 ] if and only if
Lie(H1 ) = Lie(H2 ).
(ii) If H ≤ G is closed with NG (H) open, then Lie(H) is an ideal.
(iii) If h is an ideal of g, then there exists a compact locally normal subgroup H with
Lie(H) = h.
Proof. We first recall that every closed subgroup of G is a Lie group, whose Lie algebra
is a subalgebra of g, see [5, Ch. III, §8, Th. 2].
We next observe that every vector subspace of g is closed: this is easily seen by induction on dimension, using the fact that every proper open subgroup of Qp is compact.
Therefore, every Lie subalgebra of g is the Lie algebra of a closed subgroup of G by [5,
Ch. III, §7, Th. 2].
12
This being recorded, assertion (i) now follows from [5, Ch. III, §7, Th. 2], while assertions
(ii) and (iii) follow from [5, Ch. III, §7, Prop. 2].
3
Quasi-centralisers and their stability properties
3.1
Motivation
Let (G, U ) be a Hecke pair. Our principal aim in the present section and Sections 4 and
5 is to obtain two subsets of the structure lattice of (G, U ), namely the local decomposition lattice LD(G, U ) and the centraliser lattice LC(G, U ) that are Boolean
algebras, provided that (G, U ) satisfies some unavoidable hypotheses. The local decomposition lattice will have as representatives the direct factors of finite index subgroups of
U ; more generally, one can define the local decomposition lattice LD(G, U ; K) relative
to a locally normal subgroup K, with representatives given by those direct factors of
finite index subgroups of K that are locally normal in G. The centraliser lattice will
have as representatives the centralisers of all locally normal subgroups. Here are some
observations that will lead us to the required hypotheses.
1. Given a direct factor K of a group U , observe that there may be many different
subgroups L such that K × L = U : consider for instance the direct factorisations
of U = Z2p as Zp × Zp . However, if the centre of U is trivial, then K has a unique
complement, namely L = CU (K).
2. A priori, the map
LN (G, U ) → LN (G, U ); [K] 7→ [CG (K)]
is not well-defined: given a locally normal subgroup K and a finite index subgroup
V of U , it could be the case that CG (K ∩ V ) lies in a local equivalence class
that is strictly above the local equivalence class of CG (K). So instead, we will
define a subgroup QCG ([K]) that is the union of all centralisers of representatives
of [K], and then define [K]⊥ to be [QCG ([K])]. We will see that under suitable
conditions, every representative of [K] actually has the same centraliser, so the
map [K] 7→ [CG (K)] is in fact well-defined.
3. There is a natural approach to obtaining a Boolean algebra from a bounded lattice
L, provided that L is pseudocomplemented. A pseudocomplementation is a
function ⊥ : L → L such that:
(∗) For all α, β ∈ L, we have α ∧ β = 0 if and only if β ≤ α⊥ .
If a pseudocomplementation exists, it is necessarily unique. Moreover, the set
A = {α⊥ | α ∈ L} forms a Boolean algebra, with the restriction of ⊥ serving as
the complementation map on A (in particular, ⊥3 = ⊥). We will obtain hypotheses
13
such that LN (G, U ) is pseudocomplemented, and the pseudocomplementation will
be given by [K] 7→ [CG (K)].
These observations lead naturally to the definitions of the quasi-centre (see 3.2), quasicentralisers (see 3.3) and C-stability (see 3.4). The purpose of the present section (Section 3) is to perform the technical groundwork needed to verify that the definitions have
the properties we desire of them.
Since the lattices we are describing are ‘local’ invariants of the Hecke pair (G, U ), for the
proofs, there will be no loss of generality in assuming G = U , and we may freely replace
G or U by a finite index subgroup when it is convenient to do so.
3.2
The quasi-centre
As observed in §3.1, to ensure unique complementation in the local decomposition lattice of the Hecke pair (G, U ), we need the finite index subgroups of U (or at least all
sufficiently small finite index subgroups of U ) to have trivial centre. When (G, U ) is
a t.d.l.c. Hecke pair, this becomes the question as to whether the quasi-centre of G,
defined as the set of all elements whose centraliser is open, is discrete. The notion of
quasi-centre is due to Burger–Mozes [6]; its basic properties are as follows.
Lemma 3.1. Let G be a locally compact group. Then:
(i) QZ(G) is a (not necessarily closed) characteristic subgroup of G.
(ii) QZ(O) ≤ QZ(G) for any open subgroup O ≤ G.
(iii) For any closed normal subgroup N of G, the image of QZ(G) under the projection
to G/N is contained in QZ(G/N ).
(iv) QZ(G) contains every discrete subgroup of G whose normaliser in G is open. In
particular, all discrete normal subgroups and all finite locally normal subgroups are
contained in QZ(G).
(v) For any discrete normal subgroup N of G, we have QZ(G)/N = QZ(G/N ). In
particular, if QZ(G) is discrete, then QZ(G/QZ(G)) = 1.
Proof. (i) For all g, h ∈ G, we have CG (gh) ≥ CG (g) ∩ CG (h). This implies that gh ∈
QZ(G) as soon as g, h ∈ QZ(G). Part (i) follows.
(ii) Let O ≤ G be open. Since CG (g) ≥ CO (g) for any g ∈ G, so that CG (g) is open as
soon as CO (g) is so. Part (ii) follows.
(iii) Clear since the projection G → G/N is an open map.
14
(iv) Let H be a discrete subgroup of G such that NG (H) is open and let V be an open
compact subgroup of NG (H). Consider h ∈ H. Since the V -conjugacy class of h is
compact, thus finite, h is centralised by an open subgroup of V and so h ∈ QZ(G) as
claimed.
(v) Let N ≤ G be a discrete normal subgroup. We have N ≤ QZ(G) by (iv) and
QZ(G)/N ≤ QZ(G/N ) by (iii). Let now g ∈ G such that gN ∈ QZ(G/N ). Thus there
is U ∈ B(G) such that [g, U ] ⊂ N . Since N is discrete, there exists also V ∈ B(G)
with V ∩ N = 1. Since the commutator map x 7→ [g, x] is continuous, we may, upon
replacing U by a sufficiently small open subgroup, assume that [g, U ] ⊂ V . We deduce
that [g, U ] = 1, whence g ∈ QZ(G). This confirms that QZ(G)/N ≥ QZ(G/N ), and the
required assertion follows.
Given a general Hecke pair (G, U ), we define the quasi-centre of G with respect
to U , denoted by QZ(G, U ), to be the set of all elements of G that have finitely many
U -conjugates. When (G, U ) is a t.d.l.c. pair, this is equal to the quasi-centre of G as
defined previously and we will adopt the notation QZ(G) for the quasi-centre of G with
respect to U in the general case as well when there is no ambiguity about the choice
of U . Subgroups H of G are understood to have the Hecke pair structure (H, H ∩ U ),
so QZ(H) is the set of all elements of H that have finitely many (H ∩ U )-conjugates.
Note that QZ(H) is a normal subgroup of H, and agrees with the usual definition of
the quasi-centre for topological groups in the case that (G, U ) is t.d.l.c. and H is closed.
When H is bounded with respect to (G, U ), observe that QZ(H) is just the union of all
finite conjugacy classes of H.
Note that a version of Lemma 3.1(iv) also holds for Hecke pairs (G, U ):
Lemma 3.2. Let (G, U ) be a Hecke pair and let H ≤ G. Suppose that there is a finite
index subgroup V of U such that V ≤ NG (H) and H ∩ V = 1. Then H ≤ QZ(G).
Proof. Let x ∈ H. Then there is a finite index subgroup W of V such that xW x−1 ≤ V ,
since V is commensurated by G. Given w ∈ W , we have [x, w] ∈ H, since H is normalised
by W , and also [x, w] ∈ xW x−1 W ⊆ V . Thus [x, w] ∈ H ∩ V , so [x, w] = 1. Hence
x ∈ CG (W ) ≤ QZ(G).
The quasi-centre has a useful characterisation in the case that it is locally equivalent to
the trivial group.
Corollary 3.3. Let (G, U ) be a Hecke pair. Suppose that QZ(G) ∈ [1], that is, there
exists a finite index subgroup V of U such that QZ(G) ∩ V = 1. Then QZ(G) is the
unique largest locally normal representative of [1]. In particular, if QZ(G) = 1, then the
trivial group is the only locally normal representative of [1] in G.
Proof. Follows immediately from Lemma 3.2.
15
Observe that if V is a finite index subgroup of U , then QZ(V ) = QZ(G) ∩ V . So in
taking the local approach, to prove results about Hecke pairs (G, U ) with QZ(G) ∈ [1],
we are usually free to assume that QZ(G) = 1.
3.3
Quasi-centralisers
Generalising the quasi-centre, we introduce a notion of the centraliser of a subgroup that
depends only on its local equivalence class.
Definition 3.4. Let (G, U ) be a Hecke pair, let H and K be subgroups of G and let C
be the class of subgroups of K ∩ U of finite index. The quasi-centraliser of K in H is
given by
[
QCH (K) :=
CH (L).
L∈C
Clearly the quasi-centraliser QCH (K) contains the centraliser CH (K). Moreover, the
quasi-centraliser is a subgroup of G, since CH (L1 )CH (L2 ) ⊆ CH (L1 ∩ L2 ). Similarly we
have QCH (K) = QCH (L) for any finite index subgroup L of K ∩ U , so that QCH (K)
depends only on the local equivalence class α = [K] and we can write QCH (α) without
ambiguity. Some caution is required in saying that two subgroups ‘quasi-commute’, as
the relation A is contained in the quasi-centraliser of B is not symmetric. We see that
QZ(H) = QCH (H).
We emphasise that the quasi-centraliser of a subgroup
Q need not be closed, even in the
context of`profinite groups. For example, if G = N S3 , then QCG (G) is the dense
subgroup N S3 . However, it is the case that QCG (H) = QCG (H) in this context.
Lemma 3.5. Let (G, U ) be a Hecke pair, such that G is a topological group and U is
open in G. Let H be a subgroup of G. Then
QCG (H) = QCG (H).
Proof. It may be assumed that H ≤ U and hence, since U is an open subgroup and
therefore closed, that H ≤ U . Hence, to compute the quasi-centralisers, we may work
with finite index subgroups of H and H.
Let K ≤ H have finite index. Then K∩H has finite index in H and CG (K∩H) ≥ CG (K).
Hence QCG (H) ≥ QCG (H). On the other hand, if L has finite index in H, then L has
finite index in H and CG (L) = CG (L). Hence QCG (H) ≤ QCG (H)
In the context of t.d.l.c. groups, the definition of the quasi-centraliser can be reformulated
as follows.
16
Lemma 3.6. Let (G, U ) be a t.d.l.c. Hecke pair and H, K be subgroups of G. Then we
have
[
QCH (K) :=
CH (K ∩ V ).
V ∈B(G)
Proof. In view of Lemma 3.5, we have QCH (K) = QCH (K). Moreover, for any V ∈
B(G) we have K∩V = K ∩ V because V is open, which yields CH (K∩V ) = CH (K ∩ V ) =
CH (K ∩ V ). We may therefore assume that K is closed.
S
For any
S V ∈ B(G), we have CH (K ∩ V ) ≥ CH (K ∩ U ∩ V ), so that V ∈B(G) CH (K ∩
V ) = V ∈B(G) CH (K ∩ U ∩ V ). Given a finite index subgroup L ≤ K ∩ U , we have
CH (L) = CH (L). Moreover, since K is closed and since L is an open subgroup of K ∩ U
there exists V ∈ B(G) with K ∩ U ∩ V = L. Denoting by C be the class of subgroups of
K ∩ U of finite index, we infer that
[
[
QCH (K) =
CH (L) =
CH (K ∩ U ∩ V )
L∈C
V ∈B(G)
as required.
The following observation is borrowed from [2, Prop. 2.6].
Lemma 3.7. Let G be a group, let U and V be subgroups of G, and let W ≤ U ∩ V .
Suppose θ : U → V is an (abstract) isomorphism such that θ(x) = x for all x ∈ W , and
suppose also that CG (W ∩ g−1 W g) = 1 for all g ∈ U . Then U = V and θ = idU .
Proof. Let g ∈ U and let x ∈ W ∩ g −1 W g. Then
gxg −1 = θ(gxg −1 ) = θ(g)θ(x)θ(g −1 ) = θ(g)xθ(g−1 ),
so g−1 θ(g) centralises x. This shows that g−1 θ(g) ∈ CG (W ∩ g−1 W g) = 1. Thus
g = θ(g).
In the following lemmas, we obtain some consequences of the hypothesis that a given
subgroup H of G has trivial quasi-centre.
Lemma 3.8. Let (G, U ) be a Hecke pair and H be a subgroup of G such that QZ(H) = 1.
Then:
(i) QCG (H) ∩ NG (H) = CG (H);
(ii) Every finite locally normal subgroup of G has trivial intersection with H.
17
Proof. For part (i), let t ∈ QCG (H) ∩ NG (H). Then conjugation by t induces an automorphism, θt , of H that fixes pointwise some finite index subgroup, K, of H ∩ U . Since
K is commensurated by H and QZ(H, H ∩ U ) = 1, we have CH (K ∩ g−1 Kg) = 1 for
all g ∈ H and so, by Lemma 3.7, θt = idH , that is, t ∈ CG (H). The reverse inclusion is
clear.
For part (ii), let K be a finite locally normal subgroup of G. Then there is a subgroup
V of G that is commensurate to U such that K ≤ V ≤ NG (K). Let H ′ = H ∩ V and
let K ′ = H ∩ K. Then K ′ is finite and normal in H ′ , and therefore contained in the
quasi-centre of H ′ . However, H ′ has trivial quasi-centre, and so K ′ is trivial.
Lemma 3.9. Let (G, U ) be a Hecke pair and let H and K be subgroups of G such that
QZ(H) = 1. Suppose that |H ∩ K|, |H : NH (K)| and |K : NK (H)| are all finite. Then
NK (H) = CK (H).
Proof. Set A = NH (K) and B = NK (H), so that A and B normalise each other. We
have A ∩ B ≤ H ∩ K, so that A ∩ B is finite by hypothesis. Since A ∩ B is normal in A,
we infer that A ∩ B is centralised by a finite index subgroup of A, hence by a finite index
subgroup of H since |H : A| is finite by hypothesis. Therefore A ∩ B ≤ QZ(H) = 1.
Since A and B normalise each other, it follows that they commute. Hence B centralises
a finite index subgroup of H, which implies that B ≤ QCG (H). Since B also normalises
H, we deduce that B ≤ CG (H) by Lemma 3.8(i).
Lemma 3.10. Let (G, U ) be a Hecke pair and H be a subgroup of G such that NU (H)
has finite index in U .
(i) Suppose (G, U ) is t.d.l.c. (as defined in §2.1 above), H is closed and QZ(H) is
discrete and torsion-free. Then QCG (H) is closed and H ∩ QCG (H) is discrete
and torsion-free. In particular, QCG (H) cannot be dense in G unless H is discrete
and torsion-free.
(ii) Let V be a subgroup of G that is commensurate with U . If QZ(H) ∩ V = 1, then
QCG (H) is locally equivalent to CG (H ∩ V ). In particular, we have
QCG (QCG (H)) = QCG (CG (H ∩ V )) = QCG (CV (H ∩ V )).
Proof. (i) Let V be any open compact subgroup of G. Then QZ(H ∩ V ) = 1 because
QZ(H) is discrete and torsion-free and QZ(H ∩V ) ≤ QZ(H). Therefore, by Lemma 3.8(i)
and because QCG (H ∩ V ) = QCG (H),
QCG (H) ∩ NG (H ∩ V ) = CG (H ∩ V ).
(1)
In particular, QCG (H) ∩ (H ∩ V ) ≤ Z(H ∩ V ) = 1. Hence H ∩ QCG (H) intersects
trivially any open compact subgroup of G and is therefore a discrete and torsion-free.
18
That QCG (H) is closed follows from (1) by noting that NG (H ∩ V ) is open because it
contains NU (H) ∩ V and that CG (H ∩ V ) is closed.
(ii) Let W = NV (H). Since QZ(H) ∩ V = 1, it follows that QZ(H ∩ V ) = 1, and
Lemma 3.8(i) implies that
QCG (H ∩ V ) ∩ W ≤ QCG (H ∩ V ) ∩ NG (H ∩ V ) = CG (H ∩ V ).
On the other hand, the inclusion CG (H ∩ V ) ∩ W ≤ QCG (H ∩ V ) ∩ W holds because
CG (H ∩V ) ≤ QCG (H ∩V ). Since QCG (H ∩V ) = QCG (H), we infer that QCG (H)∩W =
CG (H ∩ V ) ∩ W and hence that [QCG (H)] = [CG (H ∩ V )].
3.4
C-stability
Recall that one of our aims is to use the map K 7→ QCG (K) to define a pseudocomplementation on LN (G, U ), or at least on a subset of LN (G, U ). Observe that any
pseudocomplementation ⊥ on a lattice L satisfies the following identity:
2
α⊥ ∧ α⊥ = 0 ∀α ∈ L.
(2)
The following property applied to locally normal subgroups, closely related to the identity
(2), will prove to be the key criterion for our purposes.
Definition 3.11. Let (G, U ) be a Hecke pair and let H be a subgroup of G. Say H is
C-stable in G if
R(H) := QCG (H) ∩ QCG (CG (H)) ∈ [1].
We shall mostly be interested in the case that H is contained in a subgroup commensurate
with U . The following shows that, in that case, C-stability could have equivalently been
defined to mean that QCV (HCG (H)) = 1 for some finite index subgroup V of U . In the
t.d.l.c. case, this is equivalent to requiring that QCG (HCG (H)) be discrete.
Lemma 3.12. Let (G, U ) be a Hecke pair and let H be a bounded subgroup of G. Then
R(H) = QCG (H) ∩ QCG (CG (H)) = QCG (HCG (H)).
Proof. Let x ∈ QCG (HCG (H)), so x centralises some finite index subgroup K of
HCG (H) ∩ U . Then H ∩ K has finite index in H ∩ U and CG (H) ∩ K has finite index in
CG (H)∩U and x centralises both subgroups. Hence x ∈ QCG (H) and x ∈ QCG (CG (H)).
Conversely, consider x ∈ QCG (H)∩QCG (CG (H)) and let V be a subgroup commensurate
to U such that H ≤ V . Then there are subgroups L of finite index in H ∩ U and M of
finite index in CU ∩V (H) such that
x ∈ CG (L) ∩ CG (M ).
19
Hence x centralises the group LM . Since L has finite index in H and M has finite
index in CV (H), we thus have that x centralises a finite index subgroup of HCV (H).
Finally, the condition that H ≤ V implies that HCV (H) = HCG (H) ∩ V , and the latter
group is commensurate with HCG (H) ∩ U . Hence x centralises a finite index subgroup
of HCG (H) ∩ U as required.
We also note the role of the quasi-centre of (G, U ) as a whole in C-stability. Given any
subgroup H of G, then
QCG (H) ∩ QCG (CG (H)) ≥ QZ(G),
and equality holds in the case H = G. So if QZ(G) ∈ [1], then G is C-stable in itself,
and moreover, every subgroup of G commensurate to U is C-stable in G. If instead
QZ(G) 6∈ [1], then G does not have any C-stable subgroups. So in proving properties of
C-stable subgroups of a Hecke pair (G, U ), we may assume QZ(G) ∈ [1].
In particular, the following is an immediate consequence of Corollary 3.3:
Corollary 3.13. Let (G, U ) be a Hecke pair and let H be a C-stable locally normal
subgroup of G. Then
R(H) = QCG (H) ∩ QCG (CG (H)) = QZ(G).
The next lemma records the most basic properties of C-stable subgroups.
Lemma 3.14. Let (G, U ) be a Hecke pair and let H be a C-stable subgroup of G.
(i) Let K be a subgroup of H that is commensurate with H ∩ U . Then K is C-stable
in G.
(ii) QZ(H) ∈ [1]; indeed, QZ(H) ≤ QCG (H) ∩ QCG (CG (H)).
(iii) If H ∩ U is finite, then H ≤ QCG (H) ∩ QCG (CG (H)).
Proof. Let V be a subgroup of G that is commensurate with U .
(i) Since K is commensurate with H ∩ U , we have QCG (K) = QCG (H). Since K ≤ H,
we have CG (K) ≥ CG (H), so that QCG (CG (K)) ≤ QCG (CG (H)). Therefore
QCG (K) ∩ QCG (CG (K)) ≤ QCG (H) ∩ QCG (CG (H)) ∈ [1].
(ii) We have QZ(H) = QCH (H) ≤ QCG (H). Moreover, any g ∈ QZ(H) belongs to H
and therefore commutes with CG (H), so that QZ(H) ≤ CG (CG (H)) ≤ QCG (CG (H)).
Therefore
QZ(H) ≤ QCG (H) ∩ QCG (CG (H)) ∈ [1].
(iii) If H ∩ U is finite, then QZ(H) = H, so H ≤ QCG (H). Clearly also H is contained
in QCG (CG (H)).
20
For each integer n > 0, we define inductively
n
Cn+1
H (K) = CH (CH (K))
n
QCn+1
H (α) = QCH (QCH (α)),
and
where H and K are subgroups of G and α is either a subgroup, or a local equivalence class
of subgroups. For convenience we also adopt the convention C0H (K) = K = QC0H (K).
The following is elementary, and will be used frequently without comment.
Lemma 3.15. Let G be a group and let H be a subgroup of G. Then
C2G (H) ≥ H
and
C3G (H) = CG (H).
The following lemma gives a useful relationship between n-th centralisers and n-th quasicentralisers in the context of C-stable subgroups.
Lemma 3.16. Let (G, U ) be a Hecke pair, let H be a subgroup of G. As before, we set
R(H) = QCG (H)∩QCG (CG (H)). Let n ≥ 0 and let V be a subgroup of G commensurate
with U .
(i) If L is a subgroup of G that contains a finite index subgroup of U , then R(Cn+1
L (H)) ≤
R(CnL (H)). In particular, if H is C-stable, then so is CnL (H).
(ii) Suppose that H is C-stable and locally normal in G. Let V be a subgroup of G
commensurate with U such that QZ(G) ∩ V = 1. Then we have
QCV (QCnG (H)) = Cn+1
V (H ∩ V ).
In particular, QCV (H) = CV (H ∩ V ).
′
Proof. (i) Fix n ≥ 0, let H ′ = CnL (H) and let K = Cn+1
L (H) = CL (H ). On the one hand
H ′ ∩L is locally equivalent to H ′ and contained in CG (K), so QCG (CG (K)) ≤ QCG (H ′ );
on the other hand we also have QCG (K) = QCG (CG (H ′ )). Hence
QCG (K) ∩ QCG (CG (K)) ≤ QCG (H ′ ) ∩ QCG (CG (H ′ )),
n
in other words, R(Cn+1
L (H)) ≤ R(CL (H)). Thus if H is C-stable, then so are the groups
n
CL (H) for n ≥ 0.
(ii) Without loss of generality H ≤ V , since neither side of the desired equation is
affected by replacing H by H ∩ V , and H ∩ V retains the properties of being C-stable
and locally normal in G. Thus QZ(H) = 1 by Lemma 3.14(ii) and Corollary 3.13.
First suppose that n = 0: we must show that QCV (H) commutes with H. Consider
x ∈ QCV (H), so that x ∈ CV (K) for some finite index subgroup K of H ∩ U . Then
K has finite index in H and, replacing K by its core in H, we may assume that K is
normal in H.
21
Moreover, K is C-stable by Lemma 3.14(i), and it follows from part (i) and Lemma 3.14(ii)
that QZ(CV (K)) = 1.
Now CV (K) ∩ H = 1 since QZ(H) = 1. Also, H normalises CV (K) since H ≤ V and
since H normalises K. Furthermore H is locally normal, so that H is normalised by some
finite index subgroup W of U ∩ V such that W is normal in V . Hence H is normalised
by CW (K). Thus
[H, CW (K)] ≤ H ∩ CW (K) ≤ H ∩ CV (K) = 1,
so that H centralises CW (K). As CW (K) is a subgroup of CV (K) of finite index, we
see that H centralises CV (K) by Lemma 3.8(i), and in particular H centralises x. This
concludes the proof for n = 0.
For the general case, note first that CnV (H) is C-stable by part (i), has trivial quasi-centre
by Lemma 3.14(ii) and is also locally normal by Lemma 2.1(iii). Therefore, by repeated
applications of Lemma 3.10(ii), recalling that H ≤ V , we have the equation
n
n
QCn+1
V (H) = QCV (QCV (H)) = QCV (CV (H)).
The desired conclusion follows, since by the base case n = 0, we have
QCV (CnV (H)) = CV (CnV (H)).
We are now able to define a partial map α 7→ [QCG (α)] on the abstract structure lattice.
Corollary 3.17. Let (G, U ) be a Hecke pair and let K be a C-stable locally normal
subgroup of G. Suppose QZ(G) ∩ U = 1. Then for every finite index subgroup L of
K ∩ U , we have
QCU (K) = CU (L).
In particular we have [CG (K ∩ U )] = [CU (K ∩ U )] = [QCG (K)], so that the local equivalence class [CU (K)] depends only on [K].
Proof. By Lemma 3.16(ii), we have QCU (K) = CU (K ∩ U ).
Let now L be a finite index subgroup of K ∩ U . By definition of the quasi-centraliser,
we have QCG (K) = QCG (K ∩ U ) = QCG (L) ≥ CG (L). Since L ≤ K ∩ U , we also
have CG (L) ≥ CG (K ∩ U ). Intersecting with U and combining this with the observation
above, we obtain
QCU (K) = QCU (L) ≥ CU (L) ≥ CU (K ∩ U ) = QCU (K).
Therefore all those subgroups coincide, which yields the desired assertion.
22
3.5
Local C-stability
The previous section motivates the following definition.
Definition 3.18. A Hecke pair (G, U ) is called locally C-stable if all locally normal
subgroups of G are C-stable in G.
We can now establish the following theorem, which characterises locally C-stable Hecke
pairs under the assumption that U is residually finite; in particular, we obtain a characterisation of those t.d.l.c. Hecke pairs that are locally C-stable. We also obtain several
stability properties of quasi-centralisers; to avoid unnecessary complications, we state
these under the assumption that G has trivial quasi-centre. In view of Lemma 3.1(v)
this hypothesis is harmless: indeed if G is locally C-stable, then QZ(G) is discrete, so
that G/QZ(G) is locally C-stable and has trivial quasi-centre.
Theorem 3.19. Let (G, U ) be a Hecke pair such that U is residually finite.
Then G is locally C-stable if and only if QZ(G) ∈ [1] and every bounded abelian locally
normal subgroup of G is contained in QZ(G) (see §2.1 for the definition of the term
bounded).
Moreover, if G is locally C-stable and QZ(G) = 1, then every locally normal subgroup H
of G has the following properties:
(i) QZ(H) = 1.
(ii) QCG (H) = CG (H). In particular CG (H) depends only on [H].
(iii) Given a locally normal subgroup K of G, then H and K commute if and only if
H ∩ K = 1.
The proof requires some properties of subnormal subgroups of residually finite groups,
given in the following lemma.
Lemma 3.20. Let G be a group and let K be a normal subgroup of G.
(i) If M is an abelian minimal normal subgroup of K, then the normal closure of M
in G is abelian.
(ii) If M is a finite perfect normal subgroup of K and G is residually finite, then
|G : CG (M )| is finite.
Proof. (i) Let g ∈ G. Since M is a minimal normal subgroup of K, then either M =
gM g−1 , in which case [M, gM g−1 ] = 1 since M is abelian, or M ∩ gM g−1 = 1, in which
case [M, gM g−1 ] ≤ M ∩ gM g −1 = 1 since M and gM g−1 are normal subgroups of K. In
23
either case, we see that M commutes with gM g−1 , so every G-conjugate of M commutes
with every G-conjugate of M . Hence the normal closure of M in G is abelian.
(ii) Since M is finite and G is residually finite, there is a normal subgroup H of G
of finite index such that the quotient map G → G/H restricts to an injective map
M → G/H. In particular, M ∩ H = 1. Now consider [H, K]: this is a subgroup of
H ∩ K that is normal in G. Thus [H, K] and M are subgroups of G that normalise each
other and have trivial intersection, so [H, K, M ] = 1, where [H, K, M ] = [[H, K], M ]. In
particular, [H, M, M ] = 1, since M is a subgroup of K; equivalently [M, H, M ] = 1. By
the Three Subgroups Lemma (see [12, Theorem 10.3.5]), we have [M, M, H] = 1. But
M is perfect, so M = [M, M ], and hence [M, H] = [M, M, H] = 1, demonstrating that
H ≤ CG (M ).
Proof of Theorem 3.19. Suppose that G has an abelian locally normal subgroup L such
that L 6≤ QZ(G). Then L 6∈ [1] by Lemma 3.2, and moreover
L ≤ QCG (L) ∩ QCG (CG (L)),
so L is not C-stable in G.
From now on, we may suppose that QZ(G) ∈ [1] and every non-trivial abelian bounded
locally normal subgroup A of G is contained in QZ(G); in particular, every such A is
finite. If (G, U ) is t.d.l.c., it is enough to consider compact locally normal subgroups,
since the closure of an abelian locally normal subgroup is an abelian locally normal
subgroup.
We now claim that for every bounded locally normal subgroup L of G, then QZ(L) ≤
QZ(G). If L is finite, then L ≤ [1] by the fact that U is residually finite, so L ≤ QZ(G)
by Lemma 3.2. Given a bounded locally normal subgroup L of G, then the quasi-centre
of L is characteristic in L, so the quasi-centre is itself bounded locally normal. Let V be
a finite index subgroup of U such that QZ(G) ∩ V = 1. To show QZ(L) ≤ QZ(G), it is
enough to show that QZ(L) is finite, and hence it is enough to show that QZ(L ∩ V ) is
finite. Thus we may assume that L is an infinite subgroup of V such that L = QZ(L).
By replacing V with NV (L), we may assume L E V .
Since G has no infinite abelian bounded locally normal subgroup, the centre of L is
finite. By ([1], Proposition 3 and Theorem 2), it follows that L is the union of its finite
normal subgroups, so there exists a minimal finite normal subgroup M of L. Then M
is characteristically simple, so either M is abelian or M is perfect. If M is abelian,
then the normal closure N of M in V is a bounded locally normal subgroup of G, and
N is abelian by Lemma 3.20(i), so N = 1; if M is perfect, then |V : CV (M )| is finite
by Lemma 3.20(ii), so M ≤ QZ(G) = 1. In either case, we have a contradiction to
the assumption that M is non-trivial. This contradiction proves the claim that every
bounded locally normal subgroup L of G has quasi-centre contained in QZ(G).
24
We next claim that for any bounded locally normal subgroup L of G, we have
QCG (LCG (L)) ≤ QZ(G).
This implies that L is C-stable in view of Lemma 3.12.
Let V be a finite index subgroup of U such that QZ(G) ∩ V = 1 and V ≤ NG (L), and
let W be a subgroup
S of G commensurate to U such that L E W . Then L ∩ V has finite
index in L, so L = ni=1 (L ∩ V )li for some finite set of elements {l1 , . . . , ln } of L. Hence
(LCG (L)) ∩ W = LCW (L) =
n
[
(L ∩ V )CW (L)li =
m
n [
[
(L ∩ V )CV (L)li wj ,
i=1 j=1
i=1
where {w1 , . . . , wm } is a finite subset of W . Consequently QCG (LCG (L)) = QCG (K),
where K = (L ∩ V )CV (L).
Notice that CV (K) ≤ M ∩ CV (M ), where M = CV (L). In particular, CV (K) is abelian,
and moreover it is bounded and locally normal in G since CV (K) E V . It must thus
be contained in QZ(G) by hypothesis. Hence CV (K) ≤ QZ(G) ∩ V = 1. On the other
hand QZ(K) = 1 by the first part of the proof above. Therefore, by Lemma 3.8(i) we
have QCV (K) = CV (K) = 1. In particular, the group generated by QCG (K) and V is a
semidirect product QCG (K) ⋊ V . Given g ∈ QCG (K), there is a finite index subgroup
Y of V such that Y gY g−1 is contained in Y . We see that ygy −1 g−1 ∈ QCG (K) ∩ V = 1
for all y ∈ Y , so g ∈ CG (Y ) ≤ QZ(G). Thus QCG (K) ≤ QZ(G), as desired, and the
claim is proved.
It remains to prove that every locally normal subgroup H satisfies the assertions (i), (ii)
and (iii), under the assumption that QZ(G) = 1.
Let V = NU (H). Then H ∩ V is bounded locally normal, so
QZ(H) ∩ V = QZ(V ) ≤ QZ(G).
Hence QZ(H) is normalised by V and has trivial intersection with V , so QZ(H) ≤ QZ(G)
by Lemma 3.2, which proves (i).
Let H̃ = QCG (H). Since QZ(H) = 1, Lemma 3.8(i) ensures that
QCG (H) ∩ NG (H) = CG (H),
in other words NH̃ (H) = CG (H). Now NH̃ (H) contains a finite index subgroup L say
of H̃ ∩ U , so
H ≤ CG (L) ≤ QCG (H̃).
Moreover the definition H̃ = QCG (H) implies that H normalises H̃, and also that H̃
is locally normal. Then QZ(H̃) = 1 by (i), and we infer from Lemma 3.8(i) that H
centralises H̃. Hence H̃ = CG (H) and (ii) holds.
25
Let K be a locally normal subgroup of G. If K ≤ CG (H), then certainly H ∩ K = 1 by
property (i). Conversely, suppose H ∩K = 1. Then there is a finite index subgroup W of
U such that H ∩ W normalises K ∩ W , and hence H ∩ W ≤ CG (K ∩ W ) by Lemma 3.9.
Hence H ∩ W ≤ QCG (K), but then by property (ii) we have H ∩ W ≤ CG (K). Hence
by the same argument K ≤ QCG (H) = CG (H), proving (iii).
4
4.1
Direct decomposition of locally normal subgroups
Local decomposition lattices
Let (G, U ) be a Hecke pair and let H be a bounded C-stable locally normal subgroup of
G. In this section, we will obtain Boolean algebras corresponding to the decomposition
of finite index subgroups of H into direct factors.
Lemma 4.1. Let (G, U ) be a Hecke pair, let H be a bounded locally normal subgroup
of G with trivial quasi-centre, let K be a subgroup of H of finite index and let L be a
direct factor of K. Then L is C-stable in (H, H). Moreover CK (L) is the unique direct
complement of L in K and L = C2K (L).
If H is C-stable in G, then L is also C-stable in G.
Proof. Let K = L × M . Then LCH (L) is a finite index subgroup of H, so
QCH (LCH (L)) = QZ(H) = 1.
Thus L is C-stable as a subgroup of H by Lemma 3.12. In particular, L has trivial centre;
we see that CK (L) = M , and similarly CK (M ) = L. In addition LCG (L) contains K
and also contains CG (K), so KCG (K) ≤ LCG (L).
If H is C-stable in G, then K is C-stable in G by Lemma 3.14(i), and thus by Lemma 3.12,
QCV (LCG (L)) ≤ QCV (KCG (K)) = 1
for every subgroup V of G commensurate with U . In this case L is C-stable in G.
Lemma 4.2. Let G be a group such that Z(G) = 1, let K be a direct factor of G, and
let H be a subgroup of G such that H = C2G (H). Then H = (H ∩ K) × CH (K).
Proof. We have G = K × CG (K); let π1 and π2 be the projections onto K and CG (K)
respectively associated with this decomposition. Then H ≤ π1 (H) × π2 (H). Moreover
π1 (H) and π2 (H) are centralised by CG (H), so contained in H. Thus H = π1 (H) ×
π2 (H), which implies π1 (H) = H ∩ K and π2 (H) = CH (K).
26
Definition 4.3. Given a group H and a subgroup K of H, say K is an almost direct
factor of H if there is a finite index subgroup L of H such that K is a direct factor of
L.
Let H be a bounded locally normal subgroup of G, such that the quasi-centre of H
is finite. Define the local decomposition lattice LD(G, U ; H) of G at H to be the
subset of LN (G, U ) consisting of elements [K] where K is locally normal in G and K is
an almost direct factor of H.
In the case that G is a t.d.l.c. group and U is an open compact subgroup of G, we will
write LD(G; H) to mean LD(G, U ; H).
Lemma 4.4. Let (G, U ) be a Hecke pair, let H be a bounded locally normal subgroup
of G, and let K be a locally normal subgroup of G that is commensurate with H. Then
LD(G, U ; H) = LD(G, U ; K) as subsets of LN (G, U ).
Proof. Without loss of generality, we may assume that K is a finite index subgroup of
H.
It is clear from the definition that LD(G, U ; K) ⊆ LD(G, U ; H). Conversely, suppose
the finite index subgroup L of H decomposes as L1 × L2 with L1 and L2 locally normal
in G. Then Li ∩ K is locally normal, and has finite index in Li for i ∈ {1, 2}. Hence
(L1 ∩ K) × (L2 ∩ K) has finite index in K, so [L1 ] = [L1 ∩ K] and [L2 ] = [L2 ∩ K] are
contained in LD(G, U ; K).
Consequently, it makes sense to write LD(G, U ; α) where α is any element of LN (G, U )
that has a representative with trivial quasi-centre, and this lattice admits an action
of Gα by conjugation. The case α = [U ] is of particular interest, and we will define
LD(G, U ) := LD(G, U ; [U ]).
As the name suggests, the local decomposition lattice is a sublattice of LN (G, U ), but
more is true:
Theorem 4.5. Let (G, U ) be a Hecke pair and let α ∈ LN (G, U ) have a bounded locally
normal representative H with trivial quasi-centre. Then LD(G, U ; α) is a sublattice of
LN (G, U ), and internally it is a Boolean algebra (relative to the maximum α), with
complementation map ⊥α : β 7→ [QCH (β)]. If (G, U ) is t.d.l.c. and α has a closed
representative, then LD(G; α) is a sublattice of LN (G).
We will prove the following separately, as it will be reused later.
Lemma 4.6. Let M be a meet-semilattice with a least element 0. Suppose ⊥ : M 7→ M
is an order-reversing involution such that for all α, β ∈ M, α ∧ β = 0 if and only if β ≤
α⊥ . Then M is a Boolean algebra, with minimum 0 and maximum 0⊥ , complementation
given by ⊥ and join operation given by α ∨ β = (α⊥ ∧ β ⊥ )⊥ .
27
Proof. Let γ ∈ M. The existence of an order-reversing involution ensures that M is a
lattice, with maximum and join as given. For all α ∈ M we have α∨α⊥ = (α⊥ ∧α)⊥ = 0⊥
and clearly α ∧ α⊥ = 0, so α⊥ is a complement of α relative to the maximum element
0⊥ .
It remains to show that M is distributive. We first show
α ∧ β ≤ γ ⇔ α ≤ β ⊥ ∨ γ,
(3)
for all α, β, γ ∈ M. Using the properties of ⊥ in the hypothesis, we have α ∧ β ≤ γ if
and only if α ∧ β ∧ γ ⊥ = 0, while α ≤ β ⊥ ∨ γ if and only if δ = 0 where δ = α ∧ (β ⊥ ∨ γ)⊥ .
Now δ = α ∧ β ∧ γ ⊥ , so (3) is proved.
Now let α, β, γ, δ ∈ M. By (3) we have
(α ∨ δ) ∧ β ≤ γ ⇔ α ∨ δ ≤ β ⊥ ∨ γ ⇔ (α ≤ β ⊥ ∨ γ and δ ≤ β ⊥ ∨ γ).
Using (3) again we have α ≤ β ⊥ ∨ γ ⇔ α ∧ β ≤ γ and δ ≤ β ⊥ ∨ γ ⇔ δ ∧ β ≤ γ. Finally,
the conjunction of the inequalities α ∧ β ≤ γ and δ ∧ β ≤ γ is given by the inequality
(α ∧ β) ∨ (δ ∧ β) ≤ γ, proving the distributive law.
Proof of Theorem 4.5. Fix a bounded locally normal representative H of α, such that H
has trivial quasi-centre. Then every finite index subgroup of H has trivial quasi-centre.
Let K and L be almost direct factors of H such that K and L are locally normal in G.
Then KCH (K) and LCH (L) both have finite index in H, and K ∩CH (K) = L∩CH (L) =
1. We may replace K with C2H (K) and so assume K = C2H (K), since this substitution
does not affect the commensurability class of K or its status as an almost direct factor
of H that is locally normal in G. Let M = LCH (L) and let K2 = K ∩ M . Then
K2 = CM (CH (K)). Since K is C-stable in H by Lemma 4.1, in fact CH (K) = CH (K2 )
by Corollary 3.17; in turn, CH (K) has the same centraliser in H as CM (K) for the same
reason. So in fact K2 = C2M (K2 ). It follows by Lemma 4.2 that K2 ∩ L is a direct factor
of K2 . In turn, K2 is a direct factor of K2 CH (K), which has finite index in H. Thus
K2 ∩ L is an almost direct factor of H.
By a similar argument, there is a finite index subgroup R of CH (K) that factorises as
(R ∩ L) × CR (L), so there is a finite index subgroup of H that factorises as
K2 × (R ∩ L) × CR (L).
The first two factors are contained in K2 L, while the last factor centralises K2 L; additionally K2 L ∩ CR (L) ≤ QZ(H) = 1. It follows that K2 L is an almost direct factor
of H. Finally, note that K2 and L are both bounded and locally normal, and this
property is inherited by both their intersection and their product. We conclude that
[K2 ∩ L], [K2 L] ∈ LD(G, U ; α), so LD(G, U ; α) is a sublattice of LN (G, U ).
28
By Lemma 3.10, [QCH (K)] has a representative CH (K); this also represents an element
of LD(G, U ; α), so ⊥α preserves LD(G, U ; α). One has [K] ∧ β = 0 for β ∈ LD(G, U ; α)
whenever β ≤ [K]⊥α , since QZ(K) = 1, and conversely [K]∧β = 0 implies by Lemma 3.9
that some representative of β centralises a representative of [K], so in this case β ≤
[K]⊥α . It is also clear that ⊥α is an order-reversing involution on LD(G, U ; α). Thus
LD(G, U ; α) is a Boolean algebra by Lemma 4.6.
Now suppose (G, U ) is t.d.l.c. and that α has a closed representative. Then by Proposition 2.4, α has a compact locally normal representative H2 , and moreover from the
proof, H2 can be taken to be the closure of a finite index subgroup of H. We see that
H ∩ H2 has trivial quasi-centre. Since H ∩ H2 is a dense subgroup of H2 of finite index,
we have QCG (H2 ) = QCG (H ∩ H2 ). In particular, H2 has trivial quasi-centre. We now
follow the same argument as before, with H2 in place of H, and see that every element
of LN (G; α) has a representative of the form L = C2H2 (K), where K is locally normal.
Since centralisers are closed, it follows that L is compact and locally normal in G, so
LN (G; α) ⊆ LN (G); since LN (G; α) is a sublattice of LN (G, U ), it is also a sublattice
of LN (G).
Proof of Theorem I (i). Since G has trivial quasi-centre, the quasi-centraliser QCG (U )
of any open compact subgroup of G is trivial. In particular, U is C-stable in G, so by
Theorem 4.5, the lattice LD(G; [U ]) = LD(G) is a Boolean algebra.
In the case of a topological group G and a subgroup H, there naturally arises the
question of how direct factors of H relate to direct factors of H. Put another way, one
can consider the quasi-factors of a topological group, meaning the direct factors of its
dense subgroups. If H is compact and centreless, there is good control of this situation.
As a result, we obtain a relationship between decomposition lattices of t.d.l.c. groups
and their dense subgroups.
Lemma 4.7. Let G be a compact topological group such that Z(G) = 1 and let D be a
dense subgroup of G. Let K be a direct factor of D. Then
G = CG (K) × C2G (K),
and K = C2G (K) ∩ D.
In particular, the map K 7→ C2G (K) defines an injective homomorphism from LD to LG ,
where LD is the lattice of direct factors of D and LG is the lattice of direct factors of G.
Proof. Since centralisers are closed, we have CG (D) = 1.
Let D = K × L. Then Z(L) ≤ CG (D) = 1, so CD (L) = K. Similarly, CD (K) = L.
Thus D is contained in the set E = CG (K)CG (L). We see that both CG (K) and CG (L)
are closed in G, hence compact, so E is compact, hence closed. Since E is also dense,
it follows that E = G. Each of the subgroups CG (K) and CG (L) is normalised by
29
D, and hence by G, since the normaliser of a closed subgroup is closed. Moreover,
CG (K) ∩ CG (L) = CG (D) = 1. Hence
G = CG (K) × CG (L).
We see that Z(CG (K)) ≤ Z(G) = 1, so in fact CG (L) = C2G (K). It follows that
C2G (K) ∩ D = CD (L) = K.
The remaining conclusions are clear.
Proposition 4.8. Let G be a t.d.l.c. group such that QZ(G) = 1 and let D be a dense
subgroup of G. Let U be an open compact subgroup of G, and let R be a commensurated
subgroup of D that is also a dense subgroup of U . Let H be a bounded locally normal
subgroup of the Hecke pair (D, R). Then QZ(D) = 1; moreover the following assertions
hold.
(i) If H has trivial quasi-centre, then H has trivial quasi-centre. Moreover the map
2 (K)] defines an injective order-preserving map from LD(D, R; H) to
[K] 7→ [CH
LD(G, U, H).
(ii) If H has trivial quasi-centre, then LD(G, U ; H) is embedded in LD(D, R; H) as
a sublattice in a natural way. If H is closed in G, then this embedding is an
isomorphism.
Proof. We have QCG (U ) ≤ QZ(G) = 1, so QCG (R) = 1 by Lemma 3.5. In particular,
every finite index subgroup of R has trivial centraliser in D. Thus QZ(D) = 1.
Since H is locally normal in (D, R), without loss of generality we may assume H ≤ R.
(i) We have QCG (H) = QCG (H) by Lemma 3.5. Suppose QZ(H) = 1. Then QZ(H) = 1.
Let K be an almost direct factor of H that is locally normal in (D, R). Then NG (K)
contains a finite index subgroup of R, so NG (K) contains a finite index subgroup of
R = U ; thus K is compact and locally normal in (G, U ). We see from Lemma 4.7
that the map φ : [K] 7→ [C2H (K)] defines an order-preserving map from LD(D, R; H)
2 (K), where M is any finite index
to LD(G, U, H ). Moreover, we have K = M ∩ CH
subgroup of H of which K is a direct factor, so φ is injective.
(ii) If K is an almost direct factor of H that is locally normal in (G, U ), then it is also
locally normal in (D, R), since R ≤ U . We can therefore regard LD(G, U ; H) as a subset
of LD(D, R; H), consisting of those elements α ∈ LD(D, R; H) such that there exists a
representative K of α such that NG (K) contains a finite index subgroup of U . Moreover,
LD(G, U ; H) forms a sublattice of LD(D, R; H), since the meet and join operations for
LD(G, U ; H) and LD(D, R; H) are essentially the same (in both cases, the meet is the
30
class of the intersection of representatives, and the join is the class of the product of
suitably-chosen representatives).
If H is closed in G, then every almost direct factor K of H has finite index in a closed
almost direct factor C2H (K), and we see that C2H (K) is locally normal in (G, U ) if and
only if it is locally normal in (D, R), in other words LD(D, R; H) = LD(G, U ; H).
It is clear that the relation “A is a direct factor of B” is transitive. A similar transitivity
applies to local decomposition lattices:
Lemma 4.9. Let (G, U ) be a Hecke pair and let α, β ∈ LN (G, U ) have bounded locally
normal representatives with trivial quasi-centre. Then LD(G, U ; α) ⊆ LD(G, U ; β) if
and only if α ∈ LD(G, U ; β).
Proof. Suppose α ∈ LD(G, U ; β). Then there is a representative H of α and a representative K of β such that H is a direct factor of K. It follows that all almost direct
factors of H are also almost direct factors of K.
4.2
Locally finitely decomposable t.d.l.c. groups
The following are natural properties to consider in the context of local decomposition
lattices:
Definition 4.10. Let H be a group. Say H is finitely decomposable if
H = H1 × · · · × Hn
for some n ≥ 1, such that each Hi is indecomposable, i.e. all of whose direct product
decompositions are trivial. Given a t.d.l.c. group G, say G is locally finitely decomposable if every open compact subgroup of G is finitely decomposable.
Proposition 4.11. Let G be a group such that Z(G) = 1.
(i) Suppose G = G1 × · · · × Gn for some n ≥ 1, such that each Gi is indecomposable.
Then {G1 , . . . , Gn } is precisely the set of indecomposable direct factors of G and
every direct factor of G is a product of some subset of {G1 , . . . , Gn }. In particular,
Aut(G) acts on {G1 , . . . , Gn } and there are exactly 2n direct factors of G.
(ii) Suppose that G is compact and not finitely decomposable. Then G is topologically
isomorphic to an infinite Cartesian product of closed subgroups. In particular, G
has uncountably many closed direct factors up to commensurability.
Proof. For any direct factor H of G, then CG (H) is the complement of H and thus also
a direct factor, so H = CG (CG (H)). Suppose G = G1 × · · · × Gn for some n ≥ 1, such
31
that each Gi is indecomposable. Let K be a direct factor of G. Then
Q G = KCG (K),
and K ∩ CG (K) ≤ Z(G) = 1. Let I = {i | Gi 6≤ CG (K)}. Then K ≤ i∈I Gi , since K is
centralised by the remaining factors. Given i ∈ I, then K ∩ Gi is a direct factor of both
K and Gi , Q
by Lemma 4.2, so K = Gi by the fact that K and Gi are indecomposable.
Thus K = i∈I Gi .
Suppose G is compact and not finitely decomposable. We shall define a family {Gi,j }1≤j≤i
of non-trivial subgroups as follows, so that for all n, G = Gn,1 × · · · × Gn,n and Gn,n is
not finitely decomposable.
Set G1,1 = G. Suppose we have G = Gn,1 ×· · ·×Gn,n for some n. Choose some non-trivial
direct decomposition Gn,n = Hn × Kn , ensuring that Kn is not finitely decomposable
(this is always possible, as by induction Gn,n is not finitely decomposable). Now set
Gn+1,n = Hn , Gn+1,n+1 = Kn and Gn+1,j = Gn,j for 1 ≤ j < n.
T
Let K = ∞
j=1 Gj,j . We claim now that
G∼
=K×
∞
Y
Gj+1,j .
j=1
The decomposition G = Gn,1 × · · · × Gn+1,n+1 gives a continuous epimorphism
G →
Qn
Q
n
G
=
G
with
kernel
G
,
and
we
note
from
the
construction
that
n+1,j
n+1,j
n+1,n+1
j=1
Qj=1
n
G
.
These
epimorphisms
are
compatible
with
each
other
as
n
varies
so
that
j=1 j+1,j
Q
they generate an inverse system, with a limit homomorphism φ : L
G → ∞
j=1 Gj+1,j .
∞
G
We see that ker φ is precisely K. In addition φ(CG (K)) contains
j=1 j+1,j , since
Q
Gj+1,j centralises Gn+1,n+1 for all 1 ≤ j ≤ n, so φ(CG (K)) = ∞
G
j=1 j+1,j . In particular
G = KCG (K): this implies K ∩ CG (K) ≤ Z(G) = 1, so in fact G = K × CG (K). The
restriction of φ to CG (K) is consequently bijective, Q
giving the desired isomorphism. We
now obtain uncountably many direct factors Q
GJ = j∈J Gj+1,jQby letting J range over
all subsets of N. For two such direct factors j∈J1 Gj+1,j and j∈J2 Gj+1,j to be commensurate, the symmetric difference of the sets J1 and J2 must be finite. This defines
an equivalence relation on the power set of N whose equivalence classes are countable,
so there must be uncountably many equivalence classes.
Corollary 4.12. Let G be a t.d.l.c. group with QZ(G) = 1.
(i) G is locally finitely decomposable if and only if G has a basis of identity neighbourhoods consisting of open compact subgroups that are finitely decomposable.
(ii) Suppose G is first-countable. Then G is locally finitely decomposable if and only if
LD(G) is countable.
Proof. (i) The necessity of the condition is obvious. Conversely, if G is not locally
finitely decomposable, then G has some open compact subgroup V which is not finitely
decomposable. Since G has trivial quasi-centre, the group V must be center-free, and is
32
thus an infinite Cartesian product of closed subgroups by Proposition 4.11. Now, any
open subgroup W of V contains infinitely many direct factors of V ; those factors are
then also direct factors of W , and hence W cannot be finitely decomposable.
(ii) Suppose that G is locally finitely decomposable. Since LD(G) = LD(U ) for any
open compact subgroup U , we may assume that G is compact. In this case G has only
countably many open subgroups, each with only finitely many direct factors, and thus
LD(G) is countable.
Conversely, suppose LD(G) is countable. Then given an open compact subgroup U of
G, the direct factors of U fall into countably many commensurability classes. Thus U is
finitely decomposable by Proposition 4.11.
Proposition 4.13. Let G be a t.d.l.c. group and let H be a C-stable compact locally
normal subgroup that is locally finitely decomposable. Then LD(G; H) = LD(H); indeed
LD(G; H) contains LD(K) for every almost direct factor K of H.
Proof. It is clear that LD(G; H) ⊆ LD(H). Let L be an open subgroup of H, and let K
be a direct factor of L. Then NG (L) is open in G by Lemma 2.1(ii), and it preserves the
set X of indecomposable direct factors of L; since X is finite, there is an open subgroup
R of NG (L) that normalises every indecomposable direct factor of L. In turn K is a
product of indecomposable direct factors of L, so it is normalised by R. Hence K is
locally normal, so represents an element of LD(G; H). Hence LD(G; H) = LD(H), and
LD(H) contains LD(H; K) by Lemma 4.9. In turn, K is locally finitely decomposable,
because given an open subgroup M of K, the direct factors of M are also direct factors
of M CH (K), which is an open subgroup of H. Hence LD(H; K) = LD(K).
5
5.1
The centraliser lattice and weakly decomposable actions
The centraliser lattice
It would be useful to combine all the local decomposition lattices of (G, U ) (which are
not necessarily individually preserved by the action of G) into a single Boolean algebra
admitting a G-action. There are two difficulties here: one is that the property of having
trivial quasi-centre may not be well-behaved under intersections, and the other is that
we need a unified notion of complementation. To avoid the first difficulty and overcome
the second, we restrict our attention to locally C-stable Hecke pairs (G, U ) in the sense
of §3.5. Recall from Theorem 3.19 that under the assumption that U is residually finite,
these are precisely the Hecke pairs with trivial quasi-centre and no non-trivial abelian
locally normal subgroups. The resulting lattice will not be a sublattice of LN (G, U ) in
general (because the join operation is different), but it arises from the structure lattice
in a natural way.
33
Definition 5.1. Let (G, U ) be a locally C-stable Hecke pair. Define the map
⊥ : LN (G, U ) → LN (G, U ) : α 7→ [QCG (α)].
This is well-defined by Theorem 3.19, since there exists a finite index subgroup V of U
such that QZ(V ) = 1, and we have
[QCG (α)] = [QCV (α)] = [CG (K)]
for every locally normal representative K of α contained in V . The centraliser lattice LC(G, U ) is defined to be the set {α⊥ | α ∈ LN (G, U )} together with the map
⊥ restricted to LC(G, U ), the partial order inherited from LN (G, U ) and the binary
operations ∧c and ∨c given by:
α ∧c β = α ∧ β
α ∨c β = (α⊥ ∧ β ⊥ )⊥ .
In general we will write ∨ instead of ∨c in contexts where it is clear that we are working
inside the centraliser lattice.
Given a t.d.l.c. group G, we define LC(G) := LC(G, U ), where U is some open compact
subgroup of G. (The choice of U here is inconsequential.)
Theorem 5.2. Let (G, U ) be a locally C-stable Hecke pair, such that U is residually
finite.
(i) The poset LC(G, U ) is a Boolean algebra and ⊥2 : LN (G, U ) → LC(G, U ) is an
idempotent surjective lattice homomorphism. If G is t.d.l.c., then
LC(G) = {α⊥ | α ∈ LN (G)} ⊆ LN (G).
2
(ii) Fix α ∈ LN (G, U ) and let θ be the map on LN (G, U ) given by β 7→ β ⊥ ∧ α.
Then θ is an idempotent order-preserving map, and θ(LN (G, U )) = θ(LC(G, U )) ⊇
LD(G, U ; α). Moreover, θ −1 (LD(G, U ; α)) ∩ LC(G, U ) is a subalgebra of LC(G, U ).
We need the following basic group theoretic facts.
Lemma 5.3. Let G be a group such that every normal subgroup has trivial centre, in
other words that G has no non-trivial abelian normal subgroups. Then the following
equations are satisfied for any normal subgroups A and B of G.
(i) C2G (A ∩ B) = C2G (A) ∩ C2G (B);
(ii) CA (B) = CA (A ∩ B);
(iii) CA (CG (B)) = CA (CA (B)).
34
Proof. (i) Let K = C2G (A) and let L = C2G (B). It is clear that
CG (CG (A ∩ B)) ≤ CG (CG (A)CG (B)) = K ∩ L,
so it suffices to show that K ∩ L centralises CG (A∩ B). For this it is enough to show that
M = CG (A∩ B)∩ K ∩ L = 1, since all subgroups under consideration are normal. We see
that M ∩ A ∩ B ≤ Z(A ∩ B) = 1, so M ∩ A centralises B, in other words M ∩ A ≤ CG (B).
Then M ∩ A is centralised by L, so M ∩ A is central in M , so M ∩ A = 1, and hence
M ≤ CG (A). Since K centralises CG (A) and contains M , we see that M is central in
CG (A) and hence trivial, as required.
(ii) Notice that CA (A ∩ B) is normal in G and has trivial intersection with A ∩ B
(since Z(A ∩ B) = 1). Therefore [CA (A ∩ B), B] ⊂ CA (A ∩ B) ∩ B = 1, whence
CA (A ∩ B) ≤ CA (B). The reverse inclusion is clear.
(iii) This a special case of (ii), using the normal subgroups A and CG (B) instead of A
and B.
Proof of Theorem 5.2. Since we are concerned only with properties of a local nature, we
are free to replace G and U with a finite index subgroup V of U (which can be assumed
to be open in the t.d.l.c. case) such that QZ(V ) = 1. In other words, we may assume
that G has trivial quasi-centre.
(i) Let A be a bounded locally normal subgroup of G, let α = [A], and let V be a
subgroup of G commensurate with U such that A E V . By Lemma 3.16, CnV (A) is then
3
2
n
a representative of α⊥ . We deduce α⊥ ≥ α and α⊥ = α⊥ from Lemma 3.15. In
particular, ⊥ acts as an involution on LC(G, U ), so ⊥2 (LN (G, U )) = LC(G, U ). In the
2
2
t.d.l.c. case, we note that CU (CV (A)) is a closed representative of α⊥ , so α⊥ ∈ LN (G),
and since ⊥3 = ⊥, it follows that ⊥(LN (G)) = ⊥(LN (G, U )) =: LC(G) and that LC(G)
is contained in LN (G).
It is clear that ⊥ is order-reversing (so ⊥2 is order-preserving). Let α, β ∈ LN (G, U )
and choose representatives A and B of α, β respectively that are normal in a subgroup
V that is commensurate with U . Then we have CV (AB) = CV (A) ∩ CV (B), so that
(α ∨ β)⊥ = α⊥ ∧ β ⊥ .
(4)
This shows that LC(G, U ) is a meet-sublattice of LN (G, U ). Since the only abelian
normal subgroup of V is the trivial one, it also follows from Lemma 5.3(i) that ⊥2 is
compatible with the meet operation.
By Lemma 3.9, α ∧ β = 0 if and only if B centralises A, which amounts to saying that
β ≤ α⊥ . Hence LC(G, U ) is a Boolean algebra by Lemma 4.6.
We see that ⊥2 is compatible with the join operations in passing from LN (G, U ) to
LC(G, U ) as follows, using (4):
2
2
3
2
3
α⊥ ∨c β ⊥ = (α⊥ ∧ β ⊥ )⊥ = (α⊥ ∧ β ⊥ )⊥ = (α ∨ β)⊥ .
35
Hence ⊥2 is a lattice homomorphism.
(ii) It is clear that θ(LN (G, U )) = θ(LC(G, U )), since ⊥2 is a projection of LN (G, U )
to LC(G, U ). It is also clear that θ is order-preserving. Given β ∈ LN (G, U ) we have
2
4
2
2
2
θ 2 (β) = (β ⊥ ∧ α)⊥ ∧ α = β ⊥ ∧ α⊥ ∧ α = β ⊥ ∧ α,
so θ 2 = θ.
Let A be a bounded locally normal representative of α and let K be a direct factor of
A. Choose V commensurate with U such that both A and K are normal in U . Then
2
K = C2A (K), so K = CA (CV (K)) by Lemma 5.3(iii). In other words, [K] = β ⊥ ∧ α
2
where β = β ⊥ is the element [C2V (K)] = [QC2U (K)] of LC(G, U ). Thus θ(LC(G, U ))
contains LD(G, U ; α).
Let L = θ −1 (LD(G, U ; α)) ∩ LC(G, U ). Then L is closed under meets in LC(G, U ),
by the fact that ⊥2 is compatible with the meet operation. Furthermore L is closed
under complementation in LC(G, U ), since the complement of β ∧ α in LD(G, U ; α) is
(β ∧ α)⊥ ∧ α by Theorem 4.5. Using Lemma 5.3(ii) we see that
(β ∧ α)⊥ ∧ α = [CA (A ∩ B)] = [CA (B)] = β ⊥ ∧ α,
where A and B are representatives of α and β respectively that normalise each other.
So L is a subalgebra of LC(G, U ).
Theorem I (ii) now follows from Theorem 5.2 together with Theorem 3.19.
Let (G, U ) be a locally C-stable Hecke pair such that QZ(G) = 1. Then each element α
of the centraliser lattice has a canonical ‘global’ representative, namely QCG (α⊥ ). This
subgroup is the unique largest locally normal representative of α. If (G, U ) is t.d.l.c.,
then QCG (α⊥ ) is closed, but not necessarily compact.
We can consider the set L = {QCG (α) | α ∈ LN (G, U )}, ordered by inclusion, as
the global centraliser lattice of (G, U ). As a Boolean algebra, it is G-equivariantly
isomorphic to LC(G, U ). The elements of L also have a more direct characterisation
using Theorem 3.19, as follows:
Proposition 5.4. Let (G, U ) be a locally C-stable Hecke pair such that U is residually
finite and QZ(G) = 1. Then the set of subgroups L := {QCG (α) | α ∈ LN (G, U )} enjoys
the following properties:
(i) If H ∈ L, then H = C2G (H) and H is locally normal. Indeed H ∩ R = C2R (H),
where R is any locally normal subgroup of G.
(ii) Let K be a subgroup of G such that CG (K) is locally normal. Then CG (K) ∈ L.
If K is locally normal, in fact CG (K) = QCG ([K]) and C2G (K) = QCG ([K]⊥ ).
36
Proof. Let α ∈ LN (G, U ), let H = QCG (α). Since α is an element of LN (G, U ), then
|U : Uα | is finite; in turn Uα normalises H, since H is defined as a subgroup of G in
terms of α. Hence |U : NU (H)| is finite, so that H is locally normal.
Let L = QCG (α⊥ ). Then L = CG (H) by Theorem 3.19. The same argument shows that
H = CG (L), so H = C2G (H). Now let M = CG (H), and suppose R is a locally normal
subgroup of G. Then R ∩ M is locally normal and hence centreless, since G is locally
C-stable and QZ(G) = 1. Then CR (R ∩ M ) ≥ H ∩ R, since M centralises H. Moreover,
we have
CR (R ∩ M ) ∩ M = CR (R ∩ M ) ∩ (R ∩ M ) ≤ Z(R ∩ M ) = 1.
Thus CR (R ∩ M ) and M are locally normal subgroups that intersect trivially, so they
commute by Theorem 3.19(iii), so that CR (R ∩ M ) = CR (M ), and hence CR (R ∩ M ) ≤
C2G (H) = H. This proves (i).
Now let K be a locally normal subgroup of G. Then CG (K) = QCG ([K]) by Theorem 3.19, so in particular CG (K) ∈ L. Setting L = CG (K), the same argument shows
that CG (L) = QCG (α⊥ ). More generally, if K is a subgroup of G such that CG (K) is
locally normal, then CG (K) = C3G (K), and in turn C3G (K) is the centraliser of a locally
normal subgroup C2G (K) of G, so CG (K) ∈ L. This completes the proof of (ii).
As a consequence of Proposition 5.4 and Lemma 4.2, we obtain information about the
decomposability of centralisers in G.
Corollary 5.5. Let (G, U ) be a locally C-stable Hecke pair such that U is residually
finite and QZ(G) = 1. Let A and B be locally normal subgroups of G. Suppose that
B = K × L for some subgroups K and L. Then
CB (A) = CK (A) × CL (A).
Proof. Let H = CG (A) and let R = CB (A). Then R = C2B (H) by Proposition 5.4;
consequently R ≥ C2B (R), so in fact R = C2B (R). Thus by Lemma 4.2, R ∩ K and R ∩ L
are complementary direct factors of R, giving the required factorisation.
Similar to Proposition 4.8, there is a relationship between the centraliser lattice of a
t.d.l.c. group and that of its dense subgroups, viewed as Hecke pairs.
Proposition 5.6. Let G be a locally C-stable t.d.l.c. group such that QZ(G) = 1 and let
D be a dense subgroup of G. Let U be an open compact subgroup of G, and let R be a
commensurated subgroup of D that is also a dense subgroup of U .
(i) The Hecke pair (D, R) is locally C-stable, with QZ(D) = 1.
(ii) Let L be the global centraliser lattice of (G, U ) and let L′ be the global centraliser
lattice of (D, R). Then there is an injective D-equivariant order-preserving map
from L′ to L given by K 7→ C2G (K).
37
Proof. Given a locally normal subgroup L of D, then L and CG (L) are both closed
and locally normal in G; in particular, CG (L) ∈ L by Proposition 5.4. Assuming L is
non-trivial, then since L cannot be finite or abelian, neither can L. In addition, we have
QZ(D) = 1 by Proposition 4.8. This proves (i) in view of Theorem 3.19.
Given K ∈ L′ , then CG (K) ∈ L as above, hence C2G (K) ∈ L. Thus θ : K 7→ C2G (K) is a
map from L′ to L. Clearly θ is D-equivariant and order-preserving. Since K = C2D (K),
we see that K = θ(K) ∩ D as follows:
K ≤ θ(K) ∩ D = C2G (K) ∩ D ≤ CG (CD (K)) ∩ D = C2D (K) = K.
Thus θ is injective. This proves (ii).
Example 5.7. In general, given a t.d.l.c. group G, the centraliser lattice of a dense
subgroup of G does not determine the centraliser lattice of G. Let F be a non-abelian
free group (with the discrete topology), and consider F as a Hecke pair (F, F ). It is clear
that F is locally C-stable, so we can define the centraliser lattice LC(F, F ). Since CF (x)
is cyclic for all x ∈ F r {1}, in fact LC(F, F ) = {0, ∞}. Likewise LC(F̂p ) = {0, ∞},
where F̂p is the pro-p completion of F . However, there are many pro-p groups P with
dense free subgroups such that LC(P ) is large. For instance, if P is a pro-p branch
group, then by results of J. Wilson, P acts faithfully on LC(P ), and P contains a dense
non-abelian free subgroup. (See [27].)
The global centraliser lattice is a natural frame of reference for considering quasi-factors
of a locally C-stable t.d.l.c. group that has trivial quasi-centre. In contrast to the compact
case considered in Lemma 4.7, a closed quasi-factor of a non-compact t.d.l.c. group G
need not be a direct factor of G. (See [7], Appendix II.)
Proposition 5.8. Let G be a locally C-stable t.d.l.c. group with QZ(G) = 1. Let K be a
closed normal subgroup of G, and suppose KCG (K) is dense in G. Let α = [K]. Then
there is a dense subgroup of G of the form QCG (α)× QCG (α⊥ ), where both direct factors
are closed and normal in G. In addition, the following are equivalent:
(i) G = K × CG (K);
(ii) G = QCG (α) × QCG (α⊥ );
(iii) α ∈ LD(G);
2
(iv) α⊥ ∈ LD(G).
Proof. We have QCG (α) = CG (K) and QCG (α⊥ ) = C2G (K) by Proposition 5.4; in
particular, both QCG (α) and QCG (α⊥ ) are closed. Since K is normal, both QCG (α)
and QCG (α⊥ ) are normal in G. Since Z(G) = 1, we see that both K and CG (K) have
trivial centre. This ensures that the subgroup generated by QCG (α) and QCG (α⊥ ) is a
direct product.
38
Let U be an open compact subgroup of G and let L be a direct factor of G. Then
2
L = C2G (L), so [L] = [L]⊥ . Moreover, the topology of G is generated as a product
topology by the topologies of L and CG (L). In particular
(L ∩ U ) × CU (L)
is open in G, hence has finite index in U . Thus L ∩ U is an almost direct factor of U , so
α = [L ∩ U ] ∈ LD(G). Thus (i) implies (iii) and (ii) implies (iv).
On the other hand, let L be a closed normal subgroup of G such that LCG (L) is dense
and [L] ∈ LD(G).
(L ∩ U ) × CU (L ∩ U )
is open in U , since it is a closed subgroup of finite index, and moreover CU (L ∩ U ) =
CU (L) by Theorem 3.19. Hence LCG (L) contains an open subgroup of G. Since LCG (L)
is also dense, we conclude that G = L × CG (L). So (iii) implies (i) and (iv) implies (ii).
It is clear that (i) implies (ii), since K ≤ QCG (α⊥ ) and CG (K) = QCG (α). Conversely
if (ii) holds, then K × CG (K) is a closed subgroup of G = QCG (α) × QCG (α⊥ ), since K
is closed in QCG (α⊥ ) and CG (K) = QCG (α). Since K × CG (K) is also dense, it follows
that (i) holds. This completes the proof that (i), (ii), (iii) and (iv) are equivalent.
Corollary 5.9. Let G be a locally C-stable t.d.l.c. group such that LC(G) = LD(G) and
such that QZ(G) = 1. Then every closed quasi-factor of G is a direct factor of G.
5.2
Weakly decomposable actions
Let A be a Boolean algebra. Then by the Stone representation theorem, A defines
a profinite space (that is, a compact totally disconnected space), the Stone space
S(A) of A, whose points are the ultrafilters of A, with clopen subsets corresponding to
elements of A. Conversely, given a profinite space X, the set A(X) of clopen subsets of X
forms a Boolean algebra. This correspondence produces a natural isomorphism between
Aut(A) and Aut(S(A)). In practice we will often find it convenient to abuse notation
and treat elements of A as subsets of S(A), identifying α ∈ A with
S(α) := {p ∈ S(A) | α ∈ p} ∈ A(S(A)).
Given α ∈ A and p ∈ S(A), the expressions ‘α ∈ p’ and ‘p ∈ α’ can therefore be taken
to be synonymous. A partition P of α ∈ A is a finite subset of A such that the join
of P is α and the meet of any two distinct elements of P is 0; a partition of A is just a
partition of ∞ in A.
Definition 5.10. Let G be a topological group acting on a set Ω. Say that the action
of G on Ω is smooth if every point stabiliser is open; this implies that for every open
compact subgroup U of G, the orbits of U on Ω are all finite. Note that if the action is
smooth, then G is orbit-equivalent to any of its dense subgroups.
39
The following fact shows the relevance of smooth actions for t.d.l.c. groups:
Lemma 5.11. Let G be a t.d.l.c. group, let X be a profinite space and A a Boolean
algebra.
(i) If G acts on X by homeomorphisms, and if the G-action is continuous with respect
to the topology of uniform convergence, then the corresponding G-action on the
Boolean algebra of clopen subsets of X is smooth.
(ii) If G acts on A by automorphisms, and if the G-action is smooth, then the corresponding G-action on the Stone space S(A) is continuous with respect to the
topology of uniform convergence. In particular, the action map (g, x) 7→ gx is a
continuous map from G × S(A) to S(A).
Proof. Let U ∈ B(G) and let X = X1 ∪ · · · ∪ Xn be a clopen partition of S(A).
If G acts on X by homeomorphisms, and if the G-action is continuous with respect to
the topology of uniform convergence, then for any net (gα ) converging to the identity in
G, for all i and each x ∈ Xi , we have gα (x) ∈ Xi for all
T sufficiently large α. Therefore
gα eventually belongs to the common stabiliser H = ni=1 GXi , which is thus an open
subgroup of G. It follows that H ∩ U is open in G, and so it therefore UXi = U ∩ GXi .
Thus (i) holds.
Assume now that G acts smoothly on A by automorphisms
and set X = S(A). Since
Tn
the action is smooth, the common stabiliser V = i=1 UXi is open in U . Thus any net
(gα ) converging to the identity in G eventually belongs to V and thus stabilises each Xi .
Thus (gα ) uniformly converges to the identity in the space of homeomorphisms of X.
In particular the action of G on S(A) is jointly continuous (see [16, Chapter 7]). This
proves (ii).
As before, let G be a group acting on a set Ω. Given X ⊆ Ω, the rigid stabiliser of X
is the subgroup
ristG (X) := {g ∈ G | gx = x for all x ∈ Ω r X}.
Similarly, if G acts on a Boolean algebra A, then given α ∈ A we define
ristG (α) := {g ∈ G | gβ = β for all β ≤ α⊥ }.
(Equivalently, one could define ristG (α) to be the rigid stabiliser of the subset of S(A)
corresponding to α.)
Notice that if A is a subalgebra of LC(G) or of LD(G), then QCG (α) ≤ ristG (α⊥ ) for
all α ∈ A.
40
Lemma 5.12. Let G be a t.d.l.c. group acting smoothly on a Boolean algebra A and let
υ ⊆ S(A). Then ristG (υ) is closed.
Proof. It suffices to show that Gp is closed for all p ∈ S(A). This is immediate from
Lemma 5.11.
Related to the notion of rigid stabilisers is that of (weakly, locally) decomposable
actions. The centraliser and local decomposition lattices account for all faithful actions
of this kind of a t.d.l.c. group (if any such actions exist).
Definition 5.13. Let (G, U ) be a Hecke pair such that G acts on a Boolean algebra
A, with kernel K. Say the action is weakly decomposable if, for every α ∈ A r {0},
the U -orbit of α is finite and ristG (α)/K is non-trivial. Note that the fact that U has
finite orbits ensures that U K/K is residually finite. Moreover, taking α = ∞, we see
that G/K must be non-trivial. Say the action is locally decomposable if it is weakly
decomposable and if, for all α ∈ A, the product ristU (α)ristU (α⊥ ) has finite index in U .
Given an action of G by homeomorphisms on a profinite space X, we say the action is
weakly or locally decomposable if the action of G on the algebra of clopen subsets of X
is weakly or locally decomposable respectively.
Say a Hecke pair (G, U ) is faithful weakly decomposable if QZ(G) = 1 and G admits
a faithful weakly decomposable action on a Boolean algebra. (We require that QZ(G) = 1
to avoid degenerate situations, such as Hecke pairs of the form (G, 1).)
A property closed related to the weakly decomposable property has been considered by
R. Möller and J. Vonk ([20]), referred to in [20] as property H. In particular, the authors
give a criterion for obtaining non-discrete topologically simple t.d.l.c. groups, of a kind
that we will study further in [8]. As can easily be seen from the definitions, Property H
and weakly decomposable are related as follows:
Proposition 5.14. Let G be a group of automorphisms of a thick locally finite tree T ,
let v be a vertex of T and let U = Gv . Then (G, U ) is a Hecke pair, and the action of G
on T has property H in the sense of [20] if and only if the action of G on the space of
ends of T is weakly decomposable.
Here are some basic properties of weakly decomposable actions. We state them for actions on profinite spaces; of course these can be translated via the Stone correspondence
into corresponding statements for actions on Boolean algebras.
Lemma 5.15. Let (G, U ) be a Hecke pair such that G acts by homeomorphisms on a
profinite space X. Assume that the action is weakly decomposable with kernel K. Then:
(i) X has no isolated point.
41
(ii) If QZ(G/K) = 1, then for every non-empty clopen set α ⊆ X, we have [ristG (α)] >
[K].
(iii) Suppose QZ(G/K) = 1. Then for any two clopen subsets α, β, we have α 6= β if
and only if [ristG (α)] 6= [ristG (β)].
Proof. There is no loss of generality in replacing (G, U ) by the Hecke pair (G/K, U K/K),
so we may assume that the action is faithful. Since the action is weakly decomposable,
it is clear that |X| > 1.
(i) If a point x ∈ X is isolated, its complement is a non-empty clopen set whose rigid
stabiliser is trivial, contradicting the assumption that the action is weakly decomposable.
(ii) Let α ⊆ X be a non-empty clopen subset. Then Gα contains a finite index subgroup V
of U , and contains ristG (α) as a normal subgroup; by assumption ristG (α) is non-trivial.
Since QZ(G) = 1, it follows from Lemma 3.2 that ristU (α) is infinite, so [ristU (α)] > 0.
(iii) Assume that α 6= β. Upon swapping α and β, we may assume that α 6≤ β. Then
α∧β ⊥ 6= 0 so that ristU (α∧β ⊥ ) is an infinite bounded locally normal subgroup contained
in ristU (α). On the other hand, ristU (α ∧ β ⊥ ) ∩ ristU (β) = 1. Hence [ristG (α)] 6=
[ristG (β)].
5.3
Faithful weakly decomposable groups
We now study the rigid stabilisers of the G-action on the centraliser and local decomposition lattices. Of particular interest is the case of faithful actions of G.
In general, if a group G acts on a Boolean algebra A with an invariant subalgebra A′ ,
then ristG (α) could be larger when defined in terms of the action on A′ than it is when
defined in terms of the action on A. However, this ambiguity does not arise in the case of
faithful actions on subalgebras of the centraliser lattice: the rigid stabilisers turn out to
be exactly the corresponding elements of the global centraliser lattice discussed earlier.
Proposition 5.16. Let (G, U ) be a locally C-stable Hecke pair and let A be a G-invariant
subalgebra of LC(G, U ) on which G acts faithfully. Then QZ(G) = 1. Moreover, the
following hold.
(i) For each α ∈ A, we have ristG (α) = QC2G (α) = CG (QCG (α)) = QCG (α⊥ ).
(ii) If G is a topological group and U is open in G, then the action of G on A is smooth.
(iii) For each α ∈ A, we have α = [ristG (α)].
(iv) The action of G on A is weakly decomposable, and if A is a subalgebra of LD(G, U )
the action is locally decomposable.
42
Proof. Observe that for any Hecke pair (G, U ), the quasi-centre of G fixes every element
of LN (G, U ) by conjugation. Thus to have a faithful action on a subalgebra of LC(G, U ),
we must have QZ(G) = 1. Note also that U has finite orbits on A.
(i) We see that QCG (α⊥ ) ≤ ristG (α) and QCG (α) ≤ ristG (α⊥ ). In particular, ristG (α)
centralises QCG (α), so ristG (α) ≤ CG (QCG (α)). By Proposition 5.4, in fact
CG (QCG (α)) = QC2G (α) = QCG (α⊥ ).
Hence ristG (α) = CG (QCG (α)) = QC2G (α). This proves (i).
(ii) For each α ∈ A, then ristG (α) is closed by (i), so Gα = NG (ristG (α)) is closed. Since
moreover |U : Uα | is finite, we conclude that Uα is open in U , and hence Gα is open in
G.
2
(iii) Follows from (i) and the fact that α = α⊥ , see Theorem 5.2(i).
(iv) It is clear from (iii) that the action is weakly decomposable. If A is a subalgebra of
LD(G, U ), note that for all α ∈ A there are representatives K of α and L of α⊥ such
that KL has finite index in U . Consequently ristG (α)ristG (α⊥ ) contains a finite index
subgroup of U .
Corollary 5.17. Let (G, U ) be a locally C-stable Hecke pair and let A be a G-invariant
subalgebra of LC(G, U ) on which G acts faithfully. Let K be a locally normal subgroup
2
of G, and let α = [K]⊥ . Suppose α ∈ A. Then the set of points in S(A) moved by K
forms an open dense subset of α.
Proof. Let υ be the set of points in S(A) moved by K. We have K ≤ ristG (α) by
Proposition 5.4 and Proposition 5.16(i), which ensures that υ ⊆ α. Suppose υ is not
dense in α. Then υ is contained in a proper clopen subset β of α. But then K ≤ ristG (β),
so [K] ≤ [ristG (β)] = β < α, a contradiction.
A Hecke pair (G, U ) with trivial quasi-centre that admits a faithful weakly decomposable
action is locally C-stable. Moreover, the faithful locally or weakly decomposable actions
are controlled by LD(G, U ) and LC(G, U ) respectively.
Theorem 5.18. Let (G, U ) be a Hecke pair.
(i) Suppose that G has a faithful weakly decomposable action on some Boolean algebra
A. Then the following are equivalent:
(a) QZ(G) = 1;
(b) The action of U is weakly decomposable, and every non-trivial normal subgroup
of G has non-trivial intersection with U ;
43
(c) G is locally C-stable, and A is G-equivariantly isomorphic to a subalgebra of
LC(G, U ) (indeed, the set of rigid stabilisers of A form a subalgebra of the
global centraliser lattice).
(ii) Every G-equivalence class of faithful locally decomposable actions of G (if there are
any) occurs as the action of G on a subalgebra of LD(G, U ).
We begin the proof of Theorem 5.18 with two lemmas.
Lemma 5.19. Let A, N be subgroups of a group G. Assume that N is normal, and
contains an element x ∈ N such that [A, xAx−1 ] = 1. Then N contains [A, A].
Proof. Let a, b, c ∈ A. Since a and b commute with xcx−1 , we have [a, b] = [a, bxcx−1 ].
Setting c = b−1 , we infer that [a, b] = [a, [b, x]], which belongs to N since N is normal.
Lemma 5.20. Let U be a residually finite group acting faithfully on an atomless Boolean
algebra A. Let K be a non-trivial normal subgroup of U and let α ∈ A r {0} be such
that gα ∧ α = 0 for some g ∈ K.
(i) If K is abelian, then ristU (α) is abelian.
(ii) If K is finite, then ristU (α) is finite-by-abelian (and hence centre-by-finite).
In either case, there exists β ∈ A such that 0 < β ≤ α and ristU (β) = 1, so the action
of U on A is not weakly decomposable.
Proof. We see that K ∩ ristU (α) and K ∩ ristU (gα) are conjugate subgroups of K, and
also that ristU (α)∩ristU (gα) = 1 since U acts faithfully on A. Moreover, by Lemma 5.19,
K ∩ ristU (α) contains R = [ristU (α), ristU (α)].
Suppose K is abelian. Then the conjugation action of K on its subgroups is trivial, so
K ∩ristU (α) = K ∩ristU (gα); since ristU (α)∩ristU (gα) = 1, it follows that K ∩ristU (α) =
ristK (α) = 1. Thus R = [ristU (α), ristU (α)] is trivial and ristU (α) is abelian.
Suppose instead that K is finite. In particular, R = [ristU (α), ristU (α)] is finite, so
ristU (α) has a finite index subgroup Z such that Z ∩ R = 1. We see that [ristU (α), Z] ≤
Z ∩ R = 1, so Z is central in ristU (α) and hence ristU (α) is centre-by-finite.
Since A is atomless, there is a strictly descending sequence (αi )i∈N of non-zero elements
of A such that α0 = α. If ristU (α) is finite, then there exist i, j ∈ N with i < j such
that ristU (αi ) = ristU (αj ), and thus ristU (β) is trivial, where β = αi ∧ α⊥
j . Thus we
may suppose that ristU (α) is infinite. Let Z be a central subgroup of ristU (α) of finite
index. Since U acts faithfully on A, there must be some β ∈ A r {0} and z ∈ Z such
that zβ ∧ β = 0; it follows that β ≤ α, since ristU (α) fixes α⊥ pointwise as a subspace
of S(A). Now ristU (β) and ristU (zβ) are subgroups of ristU (α) that are Z-conjugate,
so ristU (β) = ristU (zβ), and hence ristU (β) = 1 since zβ ∧ β = 0.
44
Proof of Theorem 5.18. (i) Let A be a Boolean algebra equipped with a faithful weakly
decomposable action of G. Then A is atomless by Lemma 5.15. We also see that U is
residually finite, since U has finite orbits on A.
Suppose QZ(G) = 1. Then given a non-trivial subgroup K of G such that |U : NU (K)| is
finite, then K ∩ U is infinite. In particular, ristU (α) is infinite for every α ∈ A r {0}, so
the action of V on A is weakly decomposable, for every subgroup V of G commensurate
with U . Hence (a) implies (b). Moreover, by Lemma 5.20, V has no non-trivial finite or
abelian normal subgroups. Hence G is locally C-stable, by Theorem 3.19.
Let α ∈ A and let β = α⊥ . Then ristG (α) and ristG (β) commute. Let x ∈ CG (ristG (β)).
We claim that x ∈ ristG (α), in other words, the set υ of points moved by x is a
subset of α. If not, then υ ∩ β contains some γ ∈ A, since υ is an open subset of
S(A). Indeed there is some δ ∈ A such that 0 < δ ≤ γ and xδ 6= δ. The fact that
the action is weakly decomposable ensures that ristG (δ) and ristG (xδ) are not locally
equivalent by Lemma 5.15(iii), so x does not commensurate ristG (δ), and in particular
x 6∈ QCG (ristG (δ)). Since ristG (δ) ≤ ristG (β), we have a contradiction. This shows that
ristG (α) = CG (ristG (α⊥ )).
(5)
By Proposition 5.4, ristG (α) is an element of the global centraliser lattice of G. We
therefore have a map
θ : A → LC(G, U ); θ(α) = [ristG (α)].
By Lemma 5.15, θ is injective, and (5) shows that θ is compatible with complementation;
θ is also compatible with meets. Hence θ is a homomorphism of Boolean algebras.
From the above argument, we conclude that (a) implies (b) and (c).
In the other direction, it is clear that (c) implies (a). Suppose instead that (b) holds. By
Lemma 5.20, U has no non-trivial virtually abelian normal subgroups. This implies that
the quasi-centre of U is trivial, in other words, QZ(G) ∩ U = 1. Since every non-trivial
normal subgroup of G intersects U non-trivially, we conclude that QZ(G) = 1. Hence
(b) implies (a). This completes the proof of (i).
(ii) If the action of G on A is locally decomposable, then ristU (α) × ristU (β) has finite
index in U , so ristU (α) is an almost direct factor of U for all α ∈ A. So in this case θ(A)
is contained in LD(G, U ).
Corollary 5.21. Let G be a t.d.l.c. group with QZ(G) = 1 and let U be an open compact
subgroup of G. Then every faithful weakly decomposable action of G (if there are any)
on a profinite space is continuous with respect to the topology of uniform convergence.
Proof. By Theorem 5.18, G is locally C-stable and by Proposition 5.16, the action of G on
LC(G) is smooth. The conclusion now follows from Theorem 5.18 and Lemma 5.11.
45
Proof of Theorem II. (i) This follows from Proposition 5.16.
(ii) By Theorem 5.18 (i), there is a G-equivariant embedding ρ of the algebra A of clopen
subsets of X into LC(G). This embedding ρ then corresponds, via the contravariant
functor provided by the Stone representation theorem, to a G-equivariant quotient map
ρ∗ from S(LC(G)) to X.
Another consequence of Theorem 5.18 is that if (G, U ) is locally C-stable and has a
faithful action on its centraliser lattice, then both properties are inherited by subgroups
with large normalisers.
Proposition 5.22. Let (G, U ) be a faithful weakly decomposable Hecke pair. Let H be
a locally normal subgroup of G. Then the Hecke pair (H, H ∩ U ) is locally C-stable and
H has a faithful weakly decomposable action on a principal ideal of LC(G, U ); this ideal
is H-equivariantly isomorphic to a subalgebra of LC(H, H ∩ U ).
Proof. The existence of a faithful weakly decomposable action ensures that U is residually finite. By Proposition 5.16 and Theorem 5.18, (G, U ) is locally C-stable and has
2
a faithful weakly decomposable action on A = LC(G, U ). Let α = [H]⊥ and let I be
the ideal of A generated by α. By Corollary 5.17, the set of points in S(A) moved
by H forms an open dense subset of α. Since G acts faithfully on A, it follows that
H acts faithfully on I. Since I is a principal ideal, it is a Boolean algebra in its own
right. We have QZ(H) := QZ(H, H ∩ U ) = 1 by Theorem 3.19. Given 0 < β ∈ I,
3
then ristG (β) = QCG (β ⊥ ). Since β ⊥ > α⊥ = α⊥ , we see that [ristG (β)] > [QCG (H)].
If ristH (β) = 1, then 0 = [ristG (β) ∩ H] = [ristG (β)] ∧ [H] = β ∧ [H]. Hence [H] ≤ β ⊥
3
so that α ≤ β ⊥ = β ⊥ . Since β ≤ α, we obtain β = 0, a contradiction. Thus the action
of H on I is weakly decomposable. Hence by Theorem 5.18, H is locally C-stable and
I is H-equivariantly isomorphic to a subalgebra of LC(H, H ∩ U ).
Corollary 5.23. Let (G, U ) be a faithful weakly decomposable Hecke pair and let H be
a subgroup of G. Suppose there is a sequence of subgroups
H = H1 ≤ H2 ≤ · · · ≤ Hn = G
such that
|Hi+1 ∩ U : NHi+1 (Hi ) ∩ U |
is finite for 1 ≤ i < n. Then either H = 1 or (H, H ∩ U ) is faithful weakly decomposable.
In particular, QZ(H) = 1.
Proof. Follows from Proposition 5.22 by induction.
46
5.4
Topologically free actions
For groups acting on topological spaces, there is a natural notion of actions that are in
some sense the opposite of weakly decomposable:
Definition 5.24. Let G be a group acting faithfully on a topological space X. The
action is locally faithful at x ∈ X if for every neighbourhood O of x, the pointwise
stabiliser of O in G is trivial. The action is topologically free if it is locally faithful at
every point in X.
Let (G, U ) be a Hecke pair and let G act by homeomorphisms on a profinite space X,
such that U has finite orbits on the set of clopen subsets of X. Clearly, if the action of
G on X is weakly decomposable, then it is not locally faithful at any point. In general,
the action of (G, U ) on X may be neither weakly decomposable nor locally faithful at
any point, even if the action is minimal. For example, given a (faithful, minimal) weakly
decomposable action of (G, U ) on X, then (G × Z/2Z, U ) has a (faithful, minimal) action
on the disjoint union of two copies of X that is not weakly decomposable, but has no
locally faithful point: this is achieved by letting G act identically on each copy of X and
Z/2Z swap the two copies of X. If G admits a non-inner automorphism φ of order 2,
then one could construct a similar example of an action of (H, U ), where this time
H = G ⋊ hφi has trivial quasi-centre with respect to U . As a concrete example, let T
be a regular rooted tree, let G be the derived subgroup of Aut(T ), and let X be the
set of ends of T . Then (G, G) is a Hecke pair; the G-action on X is faithful, minimal
and weakly decomposable, and the G-orbit of any clopen subset is finite. The outer
automorphism group Out(G) contains (and is in fact equal to) Aut(T )/G, which is an
infinite direct product of cyclic groups of order 2. In particular Out(G) contains many
involutions, which can be lifted to involutions in Aut(G).
However, with a stronger restriction on the group structure of U , we can ensure the
existence of locally faithful points. This property may be regarded as the antithesis of
the faithful weakly decomposable property. The following definition and proposition are
inspired by [21], which uses similar methods to comprehend the actions of free groups
on rooted trees.
Definition 5.25. Let U be group. Say U has the virtually normal intersection
property (VNIP) if for any two non-trivial subgroups K, L of U such that NU (K) and
NU (L) have finite index in U , the intersection K ∩ L is infinite.
If (G, U ) is locally C-stable, one sees that U has VNIP if and only if LC(G, U ) is trivial.
Proposition 5.26. Let (G, U ) be a Hecke pair, let X be a profinite space on which G
acts faithfully by homeomorphisms, let A be the Boolean algebra of clopen subsets of X,
and suppose that U has finite orbits on A. Suppose that QZ(G) = 1 and that U has
VNIP. Then the set of points in X at which G acts locally faithfully is non-empty and
closed. In particular, if the action of G on X is minimal, then it is topologically free.
47
Proof. Let Y be the set of points at which G acts locally faithfully. We see from the
definition that X r Y is open and hence Y is closed, since if G does not act locally
faithfully at some p ∈ X, then there is an open neighbourhood O of p such that the
pointwise stabiliser of O is non-trivial, and hence G does not act locally faithfully at any
point in O.
Suppose Y = ∅. Then for each p ∈ X, we can find αp ∈ A such that p ∈ αp and ristG (α⊥
p )
))
has
is non-trivial. Since U has finite orbits on A, the normaliser V = NU (ristG (α⊥
p
)
to
be
infinite.
Since
finite index in U ; the fact that QZ(G) = 1 then forces ristV (α⊥
pW
every point in X is covered by some αp , by compactness weThave ni=1 αpi = ∞ for some
p1 , . . . , pn ∈ X. Since U has VNIP, the intersection K = ni=1 ristU (α⊥
pi ) is non-trivial.
W
V
But ni=1 αpi = ∞, in other words ni=1 α⊥
=
0,
so
K
acts
trivially
on
A, contradicting
pi
the hypothesis that G acts faithfully on A. Hence Y must be non-empty.
Clearly Y is G-invariant, so if the action of G on X is minimal, then Y = X.
6
Some radicals of locally compact groups
6.1
The quasi-hypercentre
In previous sections, we found it useful to assume that the groups under consideration
have trivial quasi-centre. Of course, there is no reason in general for a t.d.l.c. group to
have this property; more interesting, however, is the question of which locally compact
groups G have a non-trivial Hausdorff quotient G/K such that QZ(G/K) = 1.
To address this issue, we need to introduce the following definitions.
Definition 6.1. Let G be a locally compact group. A subgroup K of G is quasihypercentral in G if, whenever N is a closed normal subgroup of G such that K 6≤ N ,
then KN/N has non-trivial intersection with QZ(G/N ).
The quasi-hypercentre of G, denoted by QZ∞ (G), is the intersection of all closed
normal subgroups N of G such that QZ(G/N ) = 1.
By Lemma 3.1, we have QZ(G) ≤ QZ∞ (G). If moreover QZ(G) is discrete, then
Lemma 3.1 ensures that QZ(G/QZ(G)) = 1 so that QZ(G) = QZ∞ (G) in this case.
The main properties of the quasi-hypercentre are collected in the following.
Theorem 6.2. Let G be a locally compact group. Then:
(i) QZ∞ (G) is a closed characteristic subgroup.
(ii) The quotient G/QZ∞ (G) has trivial quasi-centre. In particular QZ∞ (G) is the
unique minimal closed normal subgroup of G with this property.
48
(iii) QZ∞ (G) is the unique largest quasi-hypercentral subgroup of G.
Proof of Theorem 6.2. (i) is clear by the definition of the quasi-hypercentre.
(ii) Set H = QZ∞ (G) and let N be the set of all closed normal subgroups N of G such
that QZ(G/N ) = 1. Let x ∈ G, and suppose xH ∈ QZ(G/H). For any N ∈ N , the
group G/N is a quotient of G/H, so that xN ∈ QZ(G/N ) by Lemma
T 3.1(iii). Since
QZ(G/N ) = 1 by the definition of N , we infer that x ∈ N . Thus x ∈ N ∈N N = H. In
other words G/H has trivial quasi-centre, as desired.
(iii) Since QZ(G/H) = 1 by part (ii), it follows that every quasi-hypercentral subgroup
of G is contained in H. It remains to show that H itself is quasi-hypercentral. Let M
be a closed normal subgroup of G not containing H. Then N = H ∩ M does not contain
H. Therefore, by the definition of H, the quasi-centre of G/N is non-trivial and, hence,
there is some element x of G such that x 6∈ N and xN ∈ QZ(G/N ). Since QZ(G/H) = 1
by part (ii), in fact we have x ∈ H r N = H r M . Since N ≤ M , the group G/M
is a quotient of G/N , and we deduce from Lemma 3.1(iii) that x is an element of H
whose image in G/M is non-trivial and contained in QZ(G/M ). We conclude that H is
quasi-hypercentral in G, as desired.
Recall that the quasi-centre QZ(G) is characteristic but need not be closed. Theorem 6.2
ensures that if QZ(G) is non-trivial, then QZ∞ (G) is a non-trivial closed characteristic
subgroup (that could be the whole of G).
6.2
Relation with the quasi-hypercentre of subgroups
We now specialise to t.d.l.c. groups. The goal is to establish the following stability
properties of the quasi-hypercentre by passage to subgroups.
Proposition 6.3. Let G be a t.d.l.c. group and H be a subgroup of G.
(i) If H is closed and normal in G, then QZ∞ (H) ≥ QZ∞ (G) ∩ H.
(ii) If H is open in G, then QZ∞ (H) = QZ∞ (G) ∩ H.
The proof of that proposition requires some preparations and will be given at the end
of the subsection. We shall need additional definitions.
Recall that given two normal subgroups K ≥ L of an abstract group G, the centraliser
of K modulo L is defined by
CG (K/L) = {g ∈ G | [g, K] ⊂ L},
where we define [u, g] = ugu−1 g −1 . The centraliser CG (K/L) is a normal subgroup of
G, since it coincides with the preimage in G of the centraliser CG/L (K/L). We shall
49
need to quasify this notion as follows. (Recall that B(G) denotes the set of open compact
subgroups of G.)
Definition 6.4. Let G be a t.d.l.c. group and let K and L be closed subgroups. The
quasi-centraliser of K modulo L is defined by
QCG (K/L) = {g ∈ G | there is U ∈ B(G) such that [g, U ∩ K] ⊂ L.}
The following shows that QCG (K/L) is a normal subgroup of G provided that L is
normal in K and that K and L are both ‘quasi-normal’ in G:
Lemma 6.5. Let G be a t.d.l.c. group and let K and L be non-trivial subgroups of G
such that L E K. Assume that gKg −1 ∈ [K] and gLg −1 ∈ [L] for all g ∈ G. Then:
(i) QCG (K/L) = QCG (K ′ /L′ ) for all K ′ ∈ [K] and L′ ∈ [L].
(ii) QCG (K/L) is a non-trivial normal subgroup of G.
Proof. (i) It is clear from the definition that QCG (K/L) is invariant under replacing K
with a subgroup of G that is locally equivalent to K.
Let M = V ∩ L for some open compact subgroup V of G. Certainly QCG (K/M ) ⊆
QCG (K/L). On the other hand, given g ∈ QCG (K/L) and an open compact subgroup
U of G such that [u, g] ∈ L for all u ∈ U ∩ K, there is an open subgroup W of V such
that W gW g−1 is a subset of V . Thus for all w ∈ W ∩ K we have [w, g] ∈ V ∩ L = M ,
so QCG (K/M ) = QCG (K/L). Hence QCG (K/L) is invariant under replacing L with a
subgroup of G that is locally equivalent to L. This proves (i).
(ii) We see that L ≤ QCG (K/L), since [K, L] ≤ L, so QCG (K/L) is non-trivial.
Let g, h ∈ QCG (K/L) and let U and V be open compact subgroups of G such that
[u, g] ∈ L for all u ∈ U ∩ K and [v, h] ∈ L for all v ∈ V ∩ K. Then for u ∈ U ∩ V we have
[u, g−1 h] = ug −1 hu−1 h−1 g = g −1 [g, u][u, h]g ∈ g −1 Lg.
Since g−1 Lg ∈ [L] by hypothesis, we deduce from (i) that g −1 h ∈ QCG (K/g−1 Lg) =
QCG (K/L), so QCG (K/L) is a subgroup of G. By definition, we have
gQCG (K/L)g−1 = QCG (gKg−1 /gLg −1 )
for all g ∈ G. Therefore QCG (K/L) is normal in G by (i).
We note the following facts separately as they will be used again later.
50
Lemma 6.6. Let G be a t.d.l.c. group. Suppose that to every open compact subgroup
U we have assigned a closed normal subgroup R(U ), such that the following statements
hold:
(a) Given two open compact subgroups U and V of G, then R(U ) is commensurate with
R(V ). If V = gU g −1 then R(V ) = gR(U )g −1 .
(b) Given an open compact subgroup U of G, then the quotient U/R(U ) has trivial quasicentre.
Let U ∈ B(G) and set K = QCG (U/R(U )). Then:
(i) K = QCG (V /R(V )) for any V ∈ B(G).
(ii) K ∩ V = R(V ) for every V ∈ B(G).
(iii) K is a closed normal subgroup of G.
(iv) QZ(G/K) = 1.
Proof. (i) Condition (a) ensures that the set {R(U ) | U ∈ B(G)} lies in a single commensurability class of subgroups of G, so that R(U ) is a compact locally normal subgroup of
G such that [R(U )] does not depend on U . It follows from Lemma 6.5 that K is normal
in G and does not depend on the choice of U . Thus (i) holds.
(ii) By the definition of K, we see that K ∩ U consists of those elements of U whose
image in U/R(U ) is quasi-central. By condition (b), we thus have K ∩ U = R(U ). Hence
(ii) follows from (i).
(iii) We have already noted that K is normal in G as a consequence of Lemma 6.5. Since
K ∩ U = R(U ) is closed in U by (ii), it follows that K is closed in G.
(iv) Let x ∈ G such that xK ∈ QZ(G/K); then x centralises V K/K for some open
compact subgroup V of G, but this ensures in turn that x ∈ QCG (V /R(V )) = K by (ii),
so xK is trivial. Hence QZ(G/K) = 1.
Proof of Proposition 6.3. (i) We see that the quasi-hypercentre of a t.d.l.c. group is a
characteristic subgroup, so QZ∞ (H) is normal in G. Let K = CG (H/QZ∞ (H)) be the
centraliser of H modulo QZ∞ (H). Since H and QZ∞ (H) are closed normal subgroups
of G, so is K. Moreover K ∩ H = QZ∞ (H), since H/QZ∞ (H) has trivial quasi-centre.
Consequently the quotient G/K of G has a normal subgroup HK/K ∼
= H/QZ∞ (H)
whose quasi-centre is trivial by Theorem 6.2. Since
QZ(G/K) ∩ HK/K ≤ QZ(HK/K) = 1,
51
we infer that QZ(G/K) is a normal subgroup of G/K with trivial intersection with
HK/K. Therefore HK/K and QZ(G/K) commute. On the other hand the definition
of K implies that HK/K has trivial centraliser in G/K. It follows that QZ(G/K) = 1.
Therefore K ≥ QZ∞ (G), and thus QZ∞ (H) = K ∩ H ≥ QZ∞ (G) ∩ H.
(ii) Let Q = QZ∞ (G). Note that H/(Q ∩ H) ∼
= HQ/Q, and HQ/Q is an open subgroup
of G/Q; thus QZ(HQ/Q) ≤ QZ(G/Q) = 1, so H/(Q ∩ H) has trivial quasi-centre. By
Theorem 6.2, we conclude that Q ∩ H ≥ QZ∞ (H).
Combining this with part (i), we see that QZ∞ (O) = QZ∞ (G) ∩ O in the case that O is
an open normal subgroup of G.
Let now V ≤ U be compact open subgroups of G. Let O be the core of V in U , so that
O is an open normal subgroup of both V and U . By the preceding paragraph, we have
QZ∞ (U ) ∩ O = QZ∞ (O) = QZ∞ (V ) ∩ O.
In particular QZ∞ (U ) and QZ∞ (V ) are commensurate. This implies that for all U, V ∈
B(G), the groups QZ∞ (U ) and QZ∞ (V ) are commensurate.
Define R(U ) := QZ∞ (U ) for every profinite group U . We have just checked that condition (a) from Lemma 6.6 is satisfied, and so is condition (b) by Theorem 6.2. Hence,
given V ∈ B(G), the group K = QCG (V /QZ∞ (V )) is a closed normal subgroup of G that
does not depend on the choice of V . Moreover Lemma 6.6 ensures that QZ(G/K) = 1,
so K ≥ Q. Also K ∩ V = QZ∞ (V ) ≤ Q ∩ V , where the latter containment was observed
in the first paragraph above. It follows that Q contains an open subgroup of K, so
K/Q is a discrete normal subgroup of G/Q; since QZ(G/Q) = 1, this forces K = Q by
Lemma 3.1.
This shows that QZ∞ (G) = QCG (V /QZ∞ (V )) for any open compact subgroup V of G.
In particular, if H is any open subgroup of G, we may replace V by V ∩ H and deduce
that
QZ∞ (G) ∩ H = QCG (V /QZ∞ (V )) ∩ H = QCH (V /QZ∞ (V )) = QZ∞ (H).
This confirms (ii).
Remark 6.7. We have shown that the local equivalence class of the quasi-hypercentre
is unchanged by passing to an open subgroup. In particular for any open subgroup H of
G and any V ∈ B(H), the group QZ∞ (H) is locally equivalent to QZ∞ (V ), so we have
QZ∞ (G) = QCG (V /QZ∞ (V )) = QCG (H/QZ∞ (H)).
Remark 6.8. In general, the quasi-hypercentre of an arbitrary closed subgroup of a
t.d.l.c. group bears no relation to that of the group itself. For instance, any profinite
group H can be embedded as a closed subgroup of the Cartesian product G of its finite
continuous images, in which case QZ∞ (G) = G since QZ(G) is dense in G. On the other
52
hand H can be embedded in a semidirect product K⋊H (where K is an infinite Cartesian
product of copies of the free profinite group on 2 generators, for instance, and H permutes
the terms of this product) of such a form that QZ∞ (K ⋊ H) = QZ(K ⋊ H) = 1.
The following question is natural.
Question 6.9. If H is a closed normal subgroup of G contained in QZ∞ (G), does it
follow that H is quasi-hypercentral in G?
We can at least say that H is quasi-hypercentral in itself; in other words, given any
proper closed normal subgroup K of H, then QZ(H/K) is non-trivial.
6.3
Other radicals
We have seen that locally compact groups admit a unique largest quotient with trivial
quasi-centre. In a similar manner, for some classes of profinite group there is a unique
largest quotient that has no non-trivial compact locally normal subgroups in the given
class. In particular, a motivation for eliminating virtually abelian locally normal subgroups in this way is the centraliser lattice described in Section 5.1.
Definition 6.10. Let C be a class of profinite groups. Say C is stable if the following
conditions hold:
(a) C contains all abelian profinite groups, as well as all finite simple groups;
(b) Given a group U ∈ C and a closed normal subgroup K of U , then K ∈ C and
U/K ∈ C;
(c) Given a profinite group U that is a (possibly non-direct) product of finitely many
closed normal C-subgroups, then U ∈ C;
(d) C is stable under isomorphisms, i.e. it contains the whole isomorphism class of each
of its elements.
Given a profinite group U , write C(U ) for the closed subgroup generated by all normal
C-subgroups of U .
Given a stable class C, a t.d.l.c. group G is C-semisimple if QZ(G) = 1 and the only
locally normal subgroup of G belonging to C is the identity subgroup. For a profinite
group U , a subgroup K of U is called C-regular in U if for every closed normal subgroup
L of U not containing K, the image of K in U/L contains a non-trivial locally normal
C-subgroup of U/L. For a t.d.l.c. group G and closed subgroup H, say H is C-regular
in G if U ∩ H is C-regular in U for all open compact subgroups U of G.
53
Write [A] for the smallest stable class; note that by Fitting’s Theorem, all groups in this
class are virtually nilpotent. However, the class of [A]-regular t.d.l.c. groups is larger
than just the locally [A] groups; for example, all t.d.l.c. groups with virtually soluble open
compact subgroups are [A]-regular. In a recent paper ([25, Theorem 1.11]), P. Wesolek
has shown that every second-countable [A]-regular t.d.l.c. group can be constructed
from profinite groups and countable discrete groups by a (possibly transfinite) sequence
of extensions and countable increasing unions. On the other hand, we shall see in
Proposition 6.17 below that a t.d.l.c. group is [A]-semisimple if and only if it is locally
C-stable with trivial quasi-centre.
Another example of a stable class (within the class of profinite groups) is the class of
profinite groups that are topologically generated by pronilpotent normal subgroups and
quasisimple subnormal subgroups; for this class C, then C(U ) is the generalised proFitting subgroup of U . The regular and semisimple classes associated to this class are
discussed in [22].
Here is the main theorem for this section.
Theorem 6.11. Let G be a t.d.l.c. group and let C be a stable class of topological groups.
Then G has a uniquely defined closed normal subgroup RC (G), the C-regular radical
of G, that is characterised by either of the following properties:
(i) RC (G) is the unique largest locally normal subgroup of G that is C-regular in G;
(ii) G/RC (G) is C-semisimple, and given any closed normal subgroup N of G such that
G/N is C-semisimple, then N ≥ RC (G).
The proof will be given at the end of the section. The following classical fact is well
known.
Lemma 6.12 (Dietzmann’s Lemma). Let G be a group and Σ be a finite subset consisting
of torsion elements. If Σ is invariant under conjugation, then hΣi is finite.
Proof. See [11] or [18, p. 154].
For the remainder of this section, C is a stable class of profinite groups. We need to
check that the local and global definitions of C-regularity are not in conflict; this will be
ensured by part (v) of the following lemma.
Lemma 6.13. Let U be a profinite group.
(i) Suppose that C(U ) = 1. Then U is C-semisimple.
54
(ii) Let N be the set of all closed normal subgroups N of U such that U/N is Csemisimple, and define K = R(U ) to be the intersection of N . Then U/K is
C-semisimple.
(iii) Let N be a non-trivial closed normal subgroup of U such that N does not contain
any non-trivial closed normal C-subgroups of U . Then U/CU (N ) is a non-trivial
C-semisimple image of U .
(iv) Let V ≥ K ≥ L be closed normal subgroups of U and suppose K is C-regular in U .
Then L is C-regular in U and K is C-regular in V .
(v) Let K be a closed normal subgroup of U , let V be an open subgroup of U and let
L ≤ K be a closed normal subgroup of U that has finite index in K. Then the
following are equivalent: (a) K is C-regular in U ; (b) L is C-regular in U ; (c)
K ∩ V is C-regular in V .
Proof. (i) Let H be a non-trivial finite locally normal subgroup of U . Then by Dietzmann’s Lemma, H is contained in a finite non-trivial normal subgroup K of U ; there is
then a characteristic subgroup M of K that is a direct power of a finite simple group, so
that M is a normal C-subgroup of U . Thus the condition C(U ) = 1 ensures that every
non-trivial locally normal subgroup of U is infinite.
Let H be an infinite locally normal C-subgroup of U , and let V be a closed normal
subgroup of U of finite index that is contained in NU (H). By intersecting with V , we
may assume H ≤ V . Let K be the smallest closed normal subgroup of U containing
H. Then K is topologically generated by the set of conjugates of H in U , which is a
finite set of normal subgroups of K (since K ≤ V ), so K is a normal C-subgroup of U ,
contradicting the assumption that C(U ) = 1.
It remains to show that QZ(U ) = 1. If not, then there is some open normal subgroup
V of U such that CU (V ) > 1. Then CU (V ) is normal in U , hence infinite, so that
Z(V ) = CU (V ) ∩ V is infinite as well. But then Z(V ) is a non-trivial abelian normal
subgroup of U , so 1 < Z(V ) ≤ C(U ), a contradiction.
(ii) We may assume K = 1. By (i) it suffices to show that C(U ) = 1. Let H be a
non-trivial closed normal subgroup of U such that H ∈ C. Then there is some N ∈ N
such that H 6≤ N . But then the image of H in U/N is a non-trivial normal C-subgroup,
contradicting the hypothesis that U/N is C-semisimple.
(iii) Let K = CU (N ). We must have Z(N ) = 1, since otherwise Z(N ) would be a
non-trivial normal C-subgroup of U ; thus K ∩ N = 1. As a consequence, U/K has
a normal subgroup N K/K isomorphic to N , and in particular U/K 6= 1; moreover,
CU/K (N K/K) = 1. Let H/K be a non-trivial normal C-subgroup of U/K. Then
H/K ∩ N K/K is non-trivial since N K/K has trivial centraliser. Since H > K, we then
have L > 1 where L = H ∩ N . Now L is isomorphic to (H ∩ N K)/K, so L is a normal
55
C-subgroup of U contained in N , a contradiction. Thus no such subgroup H/K of U/K
exists, so C(U/K) = 1; by part (i), it follows that U/K is C-semisimple.
(iv) If L is not C-regular in U , then there is some closed normal subgroup M of U
not containing L such that LM/M does not contain any non-trivial closed normal Csubgroups of U/M . It follows by part (iii) that there is a closed normal subgroup R of
U containing M but not L such that U/R is C-semisimple. But then every C-regular
closed normal subgroup of U must be contained in R, so K is not C-regular in U . This
confirms that L is C-regular in U as soon as K is so.
Let N be a closed normal C-subgroup of V that does not contain K, and let R be the
core of N in U . Then KR/R contains a non-trivial locally normal C-subgroup S/R of
U/R. Since S > R, there is some u ∈ U such that S 6≤ uN u−1 . Since S ≤ KR E U , we
infer that u−1 Su ≤ KREV . Thus u−1 SuN/N is a non-trivial locally normal C-subgroup
of V /N that is contained in KN/N . This confirms that K is C-regular in V .
(v) We have already seen in (iv) that (a) implies (b). In the other direction, suppose
that (b) holds and let M be a closed normal subgroup of U that does not contain K.
If KM/M is finite, then it certainly contains a non-trivial normal C-subgroup of U/M .
If KM/M is infinite, then LM/M is also infinite and so contains a non-trivial normal
C-subgroup of U/M . Thus (a) and (b) are equivalent.
Given the equivalence of (a) and (b), from now on we may assume K ≤ V . Suppose K
is C-regular in one of U and V . Then K is C-regular in the core W of V in U by (iv). Let
Y be either U or V . Given a closed normal C-subgroup R of Y not containing K, then
R∩W is a normal C-subgroup of W not containing K. It follows that K(R∩W )/(R∩W )
contains a non-trivial locally normal C-subgroup S/(R ∩ W ) of W/(R ∩ W ), so KR/R
contains the non-trivial locally normal C-subgroup SR/R of W/R; note that SR/R is
also locally normal in Y /R. Thus K is C-regular in Y , so (a) and (c) are equivalent.
Proof of Theorem 6.11. Assume for the moment that G is profinite.
Let N be the set of all closed normal subgroups of G such that G/N is C-semisimple,
and let K be the intersection of N . Then K satisfies property (ii) of the theorem by
Lemma 6.13 (ii). It is clear from the construction that K contains every subgroup of G
that is C-regular in G.
Let N be a closed normal subgroup of G not containing K. If KN/N is abelian, then it
is a locally normal C-subgroup of G/N . If KN/N is non-abelian, let L = CG (KN/N ) be
the centraliser of KN modulo N . Note that L does not contain K, so by the definition
of K, the quotient G/L cannot be C-semisimple. It then follows by Lemma 6.13 (iii) that
KN/N contains a non-trivial closed normal C-subgroup of G/N . Thus K is C-regular
in G.
We have proved the theorem in the profinite case. Now let G be an arbitrary t.d.l.c.
group. We have already obtained a C-regular radical for the compact subgroups of G.
56
Moreover, the C-regular radicals of the open compact subgroups satisfy the hypotheses
of Lemma 6.6: this is clear from Lemma 6.13 (v) and the fact that the C-regular radical
is the unique largest C-regular normal subgroup. Let U be an open compact subgroup of
G and define R = RC (G) := QCG (U/RC (U )). Then by Lemma 6.6, R is a closed normal
subgroup of G that does not depend on the choice of U , such that QZ(G/R) = 1 and
R ∩ U = RC (U ).
The open compact subgroups of G/R are of the form U R/R for U an open compact
subgroup of G; U R/R is isomorphic to U/U ∩ R = U/RC (U ), so U R/R is C-semisimple.
Hence G/R is C-semisimple; in particular RC (G/R) = 1.
Let N be a closed normal subgroup of G such that G/N is C-semisimple. Then by the
profinite case of the theorem, we have N ≥ RC (U ) for every open compact subgroup
U of G; this ensures that RN/N is a discrete normal subgroup of G/N . But G/N has
trivial quasi-centre, so all discrete normal subgroups are trivial. Thus N ≥ R. We have
now shown that property (ii) holds, which is clearly a characterisation of the C-regular
radical.
Given an open compact subgroup V of G, then the V ∩ R is C-regular in V by the
profinite case of the theorem. Hence R is C-regular in G. The fact that G/R is Csemisimple then ensures that R is the largest C-regular locally normal subgroup of G:
given K ≤ G such that K 6≤ R and K is locally normal in G, then KR/R is non-discrete
(since QZ(G/R) = 1), so KR/R ∩ U R/R is non-trivial for every open compact subgroup
U of G, but clearly KR/R does not contain any non-trivial locally normal C-subgroup
of U R/R.
Theorem III now follows immediately from Theorems 6.2 and 6.11.
Theorem 6.11 also has the following consequence:
Corollary 6.14. Let G be a t.d.l.c. group. Then at least one of the following holds:
(i) G is [A]-regular;
(ii) there is a closed characteristic normal subgroup R of G such that G/R is nondiscrete and [A]-semisimple.
In particular, consider a topologically characteristically simple t.d.l.c. group G. Then G
is either [A]-regular or [A]-semisimple.
As in the case of the quasi-hypercentre (see Proposition 6.3), the C-regular radical enjoys
stability properties with respect to certain closed subgroups.
Proposition 6.15. Let G be a t.d.l.c. group and let C be a stable class of profinite
groups.
57
(i) Let H be a closed normal subgroup of G. Then RC (H) ≥ RC (G) ∩ H.
(ii) Let H be an open subgroup of G. Then RC (H) = RC (G) ∩ H.
Proof. (i) Let U be an open compact subgroup of G. Then RC (U ) is C-regular in U , so
RC (U ) ∩ H is C-regular in U ∩ H by Lemma 6.13 (iv). Thus RC (U ) ∩ H ≤ RC (U ∩ H).
At the global level, we have QCH (G/(RC (U ) ∩ H)) = QCH (G/RC (U )) since H is normal
in G, so [g, h] ∈ H whenever g ∈ G and h ∈ H. Hence we have
RC (H) =
≥
=
=
QCH (G/RC (U ∩ H))
QCH (G/(RC (U ) ∩ H))
QCH (G/RC (U ))
RC (G) ∩ H.
(ii) We can obtain both RC (H) and RC (G) ∩ H as QCH (U/RC (U )), where U is an open
compact subgroup of G that is contained in H.
We finish this section by linking it with our previous discussion of locally C-stable t.d.l.c.
groups.
Note the following variant of Lemma 3.7, giving a condition under which a locally normal
subgroup has trivial centraliser in G.
Lemma 6.16. Let (G, U ) be a Hecke pair such that QZ(G) = 1 and let L be a locally
normal subgroup of G such that CU (L) is finite. Then:
(i) CG (L) = 1.
(ii) Let H and K be subgroups of G containing L, such that H ∩ K contains a finite
index subgroup V of U . Suppose θ : H → K is an isomorphism such that θ(x) = x
for all x ∈ L and such that θ(V ) is commensurate with V . Then H = K and
θ = idH .
Proof. Observe that CU (L) is finite and normalised by a finite index subgroup of U , so
in fact CU (L) ≤ QZ(U ) = 1.
We now prove (ii). We may assume V ≤ NG (L). Choose a finite index subgroup W of
V such that W θ(W ) is contained in V . Let g ∈ W and let x ∈ L. Then
gxg −1 = θ(gxg −1 ) = θ(g)θ(x)θ(g −1 ) = θ(g)xθ(g−1 ),
so g−1 θ(g) centralises x. This shows that g −1 θ(g) ∈ CV (L) = 1. Thus g = θ(g), that is,
θ fixes W pointwise.
We may now apply Lemma 3.7, since CG (W ∩ g−1 W g) ≤ QZ(G) = 1 for all g ∈ G. Thus
H = K and θ = idH . Part (ii) follows.
58
By applying (ii) to the automorphisms of G induced by conjugation by elements of
CG (L), we infer that CG (L) ≤ Z(G) = 1, thereby confirming (i).
Proposition 6.17. Let G be a t.d.l.c. group such that QZ(G) = 1. The following are
equivalent:
(i) G is locally C-stable;
(ii) For every compact locally normal subgroup H of G and open compact subgroup U
of G we have CG (HCU (H)) = 1;
(iii) G has no non-trivial abelian locally normal subgroups;
(iv) G is [A]-semisimple.
Proof. Suppose G is locally C-stable. Then (ii) follows from Lemma 6.16(i).
Suppose (ii) holds and let H be an abelian locally normal subgroup of G. Let U be any
open compact subgroup of G and let K = H ∩ U . Then K ≤ CG (KCU (K)) = 1, so H
is discrete. Since QZ(G) = 1, we conclude that H is trivial.
Suppose (iii) holds. If G is not [A]-semisimple, then there is an open compact subgroup
U of G such that U has a non-trivial normal [A]-subgroup H, by Lemma 6.13(i). But we
see by hypothesis that U has no non-trivial finite or abelian locally normal subgroups,
so by the construction of [A], there is no way to obtain a normal [A]-subgroup. Thus G
is [A]-semisimple.
Now suppose that G is [A]-semisimple. In particular, QZ(G) = 1 and there are no
non-trivial compact abelian locally normal subgroups. Then G is locally C-stable by
Theorem 3.19.
The following illustration of the above discussion also provides a justification for the
choice of terminology.
Proposition 6.18. Let G be a p-adic Lie group with Lie algebra g = Lie(G). Let also
R = R[A] (G) be the the [A]-regular radical of G.
Then Lie(R) coincides with the solvable radical r of g, and R is the largest closed normal
subgroup of G with Lie algebra r. It coincides with the kernel of the G action on g/r.
In particular R is locally soluble, and G is [A]-semisimple if and only if its Lie algebra
g is semisimple.
Proof. By Proposition 2.6, every ideal of g is the Lie algebra of a compact locally normal
subgroup of G. In particular G has a compact locally normal subgroup K whose Lie
algebra is the soluble radical r of g. By [5, Ch. III, §9, Cor. to Prop. 6], the group K
59
has a soluble open subgroup, say K0 . By Lemma 2.1 there is a compact open subgroup
U of G containing K0 as a normal subgroup.
By Theorem 6.11 and Proposition 6.15(ii), we have K0 = Lie(K0 ) ≤ R[A] (U ) = R ∩ U .
Since U/K0 is a compact p-adic Lie group whose Lie algebra g/r is a direct product of
simple (non-abelian) Lie algebras, it follows from Proposition 2.6 that U is virtually a
direct product of hereditarily just finite profinite groups that are not virtually abelian.
In particular U has trivial quasi-centre (see Proposition 5.1 in [2]) and no non-trivial
abelian locally normal subgroup. Therefore U is [A]-semisimple by Proposition 6.17, so
that K0 = Lie(K0 ) ≤ R[A] (U ) = R ∩ U . In particular K0 is an open subgroup of R that
is soluble, and we have r = Lie(K0 ) = Lie(R).
Any closed normal subgroup of G with Lie algebra r is locally solvable by [5, Ch. III,
§9, Cor. to Prop. 6], hence [A]-regular, and thus contained in R by Theorem 6.11. Thus
R is the largest such subgroup. It follows that R is the kernel of the G-action on g/r (a
short and complete argument showing this can be found in the proof of Lemma 6.4 in
[10]).
The remaining assertions are now clear.
The [A]-regular radical has been used recently to analyse the global structure of t.d.l.c.
group admitting an open subgroup which has a continuous faithful finite-dimensional
representation over a local field: by [9, Theorem R.1], the [A]-regular radical of such a
group is the largest closed normal subgroup that is locally soluble (i.e. that has a soluble
open subgroup).
6.4
Canonical completions of Hecke pairs
The radical theory developed so far in this section is specific to t.d.l.c. groups, as the
topology plays an important role. However, results about t.d.l.c. groups naturally have
consequences for Hecke pairs in general, due to the existence of a universal ‘completion’
for Hecke pairs that generalises the well-known profinite completion of a group. A
construction is given in [4] and a proof of universality in [23].
Theorem 6.19. Let (G, U ) be a Hecke pair, such that G is a topological group (possibly
discrete) and U is closed in G. Then there is an essentially unique group homomorphism
θ : G → ĜU , the localised profinite completion!of a Hecke pair of G at U , where
ĜU is a topological group, such that the following properties hold:
(i) The image of θ is dense, the restriction θ|U of θ to U is continuous, and θ(U ) is
an open profinite subgroup of ĜU .
(ii) Let R be a topological group and let φ : G → R be a group homomorphism. Suppose
60
that φ|U is continuous and φ(U ) is profinite. Then there is a unique continuous
homomorphism ψ : ĜU → R such that φ = ψθ.
By composing with the appropriate quotient map, we also have canonical homomorphisms with related universal properties for t.d.l.c. groups with trivial quasi-centre,
respectively locally C-stable t.d.l.c. groups with trivial quasi-centre, that follow immediately from the results we have so far.
Corollary 6.20. Let (G, U ) be a Hecke pair, such that G is a topological group (possibly
discrete) and U is closed in G. Let
θ1 : G → ĜU /QZ∞ (ĜU ) and θ2 : G → ĜU /R[A] (ĜU )
be the canonical homomorphisms obtained from Theorem 6.19, Theorem 6.2 and Theorem 6.11. Set G1 = ĜU /QZ∞ (ĜU ) and G2 = ĜU /R[A] (ĜU ). Then the homomorphisms
θi are characterised by the following properties:
(i) The image of θi is dense, the restriction θi |U of θi to U is continuous, and θi (U )
is an open profinite subgroup of Gi . Moreover, QZ(G1 ) = QZ(G2 ) = 1, and G2 is
locally C-stable.
(ii) Let R be a t.d.l.c. group and let φ : G → R be a group homomorphism. Suppose
that φ|U is continuous and φ(U ) is profinite. If QZ(R) = 1, then there is a unique
continuous homomorphism ψ1 : G1 → R such that φ = ψ1 θ1 . If in addition R is
locally C-stable, then there is a unique continuous homomorphism ψ2 : G2 → R
such that φ = ψ2 θ2 .
Let (G, U ) be a Hecke pair, equipped with the discrete topology, such that U is residually
finite. Two questions arise at this point:
Question 6.21. If (G, U ) has trivial quasi-centre, is θ1 : G → ĜU /QZ∞ (ĜU ) necessarily
injective? Equivalently, does (G, U ) necessarily admit a dense embedding into a t.d.l.c.
group with trivial quasi-centre?
Question 6.22. If (G, U ) is locally C-stable with trivial quasi-centre, is θ2 : G →
ĜU /R[A] (ĜU ) necessarily injective? Equivalently, does (G, U ) necessarily admit a dense
embedding into a locally C-stable t.d.l.c. group with trivial quasi-centre?
Both questions are of a local nature: if (G, U ) has trivial quasi-centre, then by Lemma 3.2,
every non-trivial normal subgroup of G has infinite intersection with U . Thus θi is injective for i ∈ {1, 2} if and only if θi |U is injective. The questions can thus be restated
as follows:
Question 6.23. Let U be a residually finite group with trivial quasi-centre. Can U be
densely embedded in a profinite group P , such that P has trivial quasi-centre?
Question 6.24. Let U be a residually finite group with no non-trivial virtually abelian
normal subgroups. Can U be densely embedded in a profinite group P , such that P has
no non-trivial virtually abelian normal subgroups?
61
7
Restrictions on the group topology
7.1
The structure lattice and refinements of the group topology
In this section, we investigate the structural significance of fixed points in the action of
a t.d.l.c. group G on its structure lattice. Of course, if G has a closed normal subgroup
K that is neither discrete nor open, then [K] is a non-trivial fixed point in LN (G).
Especially interesting, however, is the question of whether a compactly generated, topologically simple t.d.l.c. group can have non-trivial fixed points in LN (G). We shall return
to this question in the second paper [8], elaborating on the results of the present chapter. Most results we shall obtain here are for t.d.l.c. groups that are first-countable or
σ-compact; it is known (see [15]) that a compactly generated (or even σ-compactly generated) t.d.l.c. group without non-trivial compact normal subgroups is second-countable
(and hence also first-countable and σ-compact).
Our first goal is to prove Theorem IV, giving a general connection between the structure
lattice and group topologies for first-countable t.d.l.c. groups.
Definition 7.1. Let G be a t.d.l.c. group and let µ be a commensurability class (or
subset of a commensurability class) of compact subgroups of G. The localisation of
G at µ is the t.d.l.c. group H, unique up to isomorphism, that admits a continuous
embedding φ : H → G, such that for some (any) K ∈ µ, φ(H) = CommG (K) and
φ−1 (K) is an open compact subgroup of H. This group topology on H is called the
localised topology at µ; it is denoted by T(µ) .
Thus, as an abstract group, the localisation of G at µ coincides with CommG (K) so that
the localisation of G at µ is nothing but (CommG (K), T(µ) ).
Starting with a t.d.l.c. group G, one sees that the t.d.l.c. groups with the same group
structure as G but a finer topology are precisely the groups of the form (G, T(µ) ), where
µ is a G-invariant commensurability class of compact subgroups of G. If G is firstcountable, we can in fact classify these topological refinements of G using the structure
lattice: this is the purpose of Theorem IV.
First we recall a result of Caprace–Monod [7]. Recall that a t.d.l.c. group G is residually
discrete if the intersection of all open normal subgroups of G is trivial.
Proposition 7.2 ([7] Corollary 4.1). Let G be a compactly generated t.d.l.c. group. Then
the following are equivalent:
(i) G is residually discrete;
(ii) G has a base of neighbourhoods of the identity consisting of subsets that are invariant under conjugation;
62
(iii) G has a base of neighbourhoods of the identity consisting of open compact normal
subgroups.
Corollary 7.3. Let G be a residually discrete locally compact group and let x ∈ G. Then
the open subgroups of G normalised by x form a base of neighbourhoods of the identity.
Lemma 7.4. Let U be a first-countable profinite group and let K be a closed subgroup
of U . Then U commensurates K if and only if U normalises an open subgroup of K.
Proof. Certainly, if U normalises an open subgroup of K then U commensurates K.
Conversely, suppose U commensurates K. Let µ be the commensurability class of K.
Then U(µ) is residually discrete, so by Corollary 7.3
[
U = {NU (L) | L ≤o K}.
Now K is first-countable, so it has countably many open subgroups; thus we have expressed U as a union of countably many subgroups, each of which is closed for the given
topology on U . By the Baire Category Theorem, at least one of these, say NU (L), must
be open in U . Thus L has finitely many T
conjugates in U , and we have U = CommU (L)
since L is commensurate with U . Hence u∈U uLu−1 is a closed subgroup of L of finite
index and thus an open subgroup of K.
Proof of Theorem IV. Let T be the given topology of G and let T ′ be a refinement of
T compatible with the group structure. By van Dantzig’s theorem, under T ′ there is
a subgroup K of G that is open and compact; it follows that K is compact under T .
Moreover, the set of such subgroups forms a commensurability class µ that is invariant
under conjugation in G. In particular, given a subgroup U of G that is open and compact
under T , then by Lemma 7.4, U normalises a subgroup L ∈ µ. Thus µ corresponds to
an element α = [L] of LN (G); we have G = Gα since µ is G-invariant, and see that the
topology T ′ is exactly the topology T(α) .
Conversely, let α ∈ LN (G) that is fixed by G. Then G admits a locally compact
topological refinement by declaring those representatives of α that are closed in T to be
open, in other words, the group (G, T(α) ), which by definition has the same underlying
group structure as G.
Note that none of the group topologies in Theorem IV are σ-compact, except possibly
the original topology of G, since a closed subgroup of a profinite group is either open or
of uncountable index.
7.2
A group topology derived from the centraliser lattice
Let A be a Boolean algebra, and equip Aut(A) with the coarsest group topology such
that the stabiliser of α is open, for all α ∈ A. Then Aut(A) is a totally disconnected
63
group. Given a group G acting faithfully on A, write GA for the closure of G in Aut(A),
equipped with the subspace topology.
Of particular interest is the case in which an infinite commensurated subgroup U of G
(perhaps G itself) has finite orbits. In this case, we see that U is residually finite and U A
is profinite, so the group hG, U A i can be equipped with a t.d.l.c. topology that extends
the topology of U A .
Theorem 7.5. Let (G, U ) be a Hecke pair with trivial quasi-centre and let A be a Boolean
algebra on which G has a faithful, weakly decomposable action. Let E be the subgroup of
Aut(A) generated by G and U A . Then E commensurates U A , so that E can be equipped
with a t.d.l.c. group topology by extending the topology of U A (so that U A is embedded
as an open subgroup). Moreover, the following holds:
(i) As a t.d.l.c. group, E is locally C-stable and faithful weakly decomposable.
(ii) As a locally compact topological group, E is characterised as follows:
Given a locally compact group L and an injective homomorphism θ : G → L with
dense image such that θ(U ) is an open subgroup of L, then there exists a unique
continuous surjective homomorphism φ : L → E such that the composition φθ is
the embedding of G in E.
(iii) If |G : U | is countable, the following characterisation also applies:
Given a topological group L that is a Baire space and an injective homomorphism
θ : G → L such that L = hθ(G), θ(U )i, then there exists a unique continuous
homomorphism φ : L → E such that the composition φθ is the embedding of G in
E.
Proof. Let V = U A . Given g ∈ G, we see that gU g −1 contains a finite index subgroup of
U , which ensures that gV g−1 contains an open subgroup of V . Thus E can be equipped
with a t.d.l.c. group topology by extending the topology of U A .
Certainly E acts faithfully on A, and the action of E on A is weakly decomposable with
respect to V . Given x ∈ QZ(E), then x centralises a finite index subgroup of U , and
hence xristU (α)x−1 is commensurate with ristU (α), for all α ∈ A. This implies that x
acts trivially on A, so x = 1. Thus E is locally C-stable by Theorem 5.18.
Let us suppose we have a homomorphism θ : G → L as in either (ii) or (iii), with (G, U )
and A satisfying the necessary hypotheses. We will prove (ii) and (iii) together, except
where it is necessary to divide into cases.
Given α ∈ A, let Rα = θ(ristG (α)) and let Kα = CL (Rα⊥ ). Then Kα is a closed subgroup
of L, for all α ∈ A. Moreover, we have Rα ≤ Kα , but Kα ∩ Rα⊥ is trivial. This ensures
that Kα ∩ θ(G) = Rα ; in particular, given α, β ∈ A such that α 6= β, then Rα 6= Rβ since
the action of G is weakly decomposable, so Kα 6= Kβ . Now set Nα = NL (Kα ). Then Nα
64
is closed in L, since Kα is closed. Moreover, Nα ∩ θ(G) = Nθ(G) (Rα ) = θ(NG (ristG (α))).
We see that Nα ∩ θ(U ) has finite index in θ(U ), so Nα ∩ P has finite index in P , where
P = θ(U ).
In case (ii), it follows that Nα ∩ P is open in P , and hence Nα is open in L, since P
is open in L. In case (iii), Nα is a closed subgroup of L of countable index, and hence
open by the Baire category theorem. In either case, we conclude that Nα is open in L,
as is any finite intersection of subgroups Nα . Therefore the set {Kα | α ∈ A} ordered
by inclusion forms a Boolean algebra A′ on which L acts smoothly, so that θ(U ) is
orbit-equivalent to P on the finite subsets of A′ . The natural map Kα 7→ α induces
a continuous homomorphism ψ : L → Aut(A) such that ψθ is the given action of G
on A. In case (ii), ψ(P ) has a dense subgroup in common with U A , and both groups
are compact, so ψ(P ) = U A . In case (iii), UA is dense in ψ(P ), so ψ(P ) ≤ U A . In
either case, ψ(L) is generated by ψθ(G) and ψ(P ), so ψ(L) ≤ E, and hence ψ restricts
to a continuous homomorphism φ : L → E with the required properties (in particular,
φ is surjective in case (ii)). Moreover, in both cases the required properties specify φ
uniquely, since the condition on φθ ensures that φ is uniquely specified on a dense subset
of L.
We see from Theorem 7.5(i) that the embedding G → E is not dependent on the choice of
A, only on G and the commensurability class of U . Thus if (G, U ) is any faithful weakly
decomposable Hecke pair, we may define E to be the centraliser lattice completion
(LC-completion) of (G, U ), denoted ĜLC . For example, if G is a branch group, regarded
as a Hecke pair (G, G), then ĜLC is the congruence completion of G in the sense of
[3].
Theorem 7.5 is of interest even if (G, U ) is t.d.l.c., as it has implications for the topologies
that are compatible with the group structure of G. In this case, G = ĜLC as topological
groups. Notice that if G is second-countable, then |G : U | is countable.
In particular, we can prove a slightly more general form of Theorem V.
Corollary 7.6. Let G be a t.d.l.c. group that is faithful weakly decomposable. Let L be
a topological group and let θ : L → G be an abstract group isomorphism, such that one
of the following conditions holds:
(i) L is a t.d.l.c. group, and there exists an open compact subgroup U of L such that
θ(U ) is open and compact in G.
(ii) G is second-countable and the topology of L is a Baire space.
Then θ is continuous.
In particular, if G is second-countable, then the topology of G is the unique σ-compact
locally compact topology compatible with the group structure (in particular, it is preserved
65
by every automorphism of G as an abstract group), and the set of all locally compact
group topologies of G is in natural bijection with the set of fixed points of G acting on
LN (G) by conjugation.
Proof. By applying Theorem 7.5 to the group isomorphism θ −1 : G → L, we see that
there exists a continuous group homomorphism φ such that φθ −1 = idG . Clearly φ = θ.
In particular, if G is second-countable, then the topology of G is the coarsest group
topology on G that forms a Baire space, so in particular it is the coarsest locally compact
Hausdorff group topology on G. The remaining assertions follow from Theorem IV.
7.3
Subgroups abstractly generated by compact locally normal subgroups
Let G be a locally compact group. Recall that the group Aut(G) of homeomorphic automorphisms carries a natural Hausdorff group topology, called the Braconnier topology
(see Appendix I in [7]). Basic identity neighbourhoods in Aut(G) are provided by the
sets
A(K, U ) = α ∈ Aut(G) | ∀x ∈ K, α(x) ∈ U x and α−1 (x) ∈ U x ,
where K runs over compact subsets of G and U runs over identity neighbourhoods
of G. In particular, for each compact open subgroup U < G, the stabiliser {α ∈
Aut(G) | α(U ) = U } is an identity neighbourhood in Aut(G).
In what follows, automorphism groups of locally compact groups will always be implicitly
endowed with the Braconnier topology.
Lemma 7.7. Let G be a t.d.l.c. group, let I be a directed set, and let (Un )n∈I be a net
of open compact subgroups in G forming a base of identity neighbourhoods, such that
Ui < Uj whenever i > j. Let (αn )n∈I be a net in Aut(G) such that limn∈I αn (x) = x for
all x ∈ G.
If for each m ∈ I, there is some n0 ∈ I such that αn (Um ) = Um for all n ≥ n0 , then
(αn )n∈I converges to the identity with respect to the Braconnier topology.
Proof. Let K ⊂ G be a compact subset and U < G an identity neighbourhood.
Choose
S
Um ⊂ U . We can find a finite subset {x1 , . . . , xk } ⊂ K such that K ⊂ ki=1 Um xi . Since
(αn ) converges pointwise to the identity, so does the sequence (α−1
n ). Hence we may find
−1
n0 such that αn (xi ) ∈ Um xi and αn (xi ) ∈ Um xi for all n ≥ n0 .
Let now x ∈ K. Choose i such that Um x = Um xi . Thus we have
αn (x) ∈ αn (Um x) = αn (Um xi ) = Um αn (xi ) = Um xi = Um x.
66
Similarly, we get α−1
n (x) ∈ Um x. This shows that αn belongs to A(K, Um ) ⊂ A(K, U )
for all n ≥ n0 . Thus (αn ) converges to the identity with respect to the Braconnier
topology.
Definition 7.8. Let G be a t.d.l.c. group and let α ∈ LN (G). Define the compact
locally normal closure LN(G, α) of α in G to be the group generated by all compact
locally normal representatives of α.
We remark that LN(G, α) is a (not necessarily closed) subgroup of G with open normaliser, and if α is fixed by G then LN(G, α) is normal in G. If G is locally C-stable with
trivial quasi-centre and α ∈ LC(G), then it follows from Proposition 5.4 that LN(G, α)
is an open subgroup of QCG (α⊥ ), which in turn is a closed subgroup of Gα , and the
normaliser of LN(G, α) is exactly Gα .
Proposition 7.9. Let G be a t.d.l.c. group and let α be a fixed point of G in LN (G).
Let D = LN(G, α). Then the natural homomorphism G → Aut(D, T(α) ) induced by the
G-action by conjugation is continuous.
Proof. Since G commensurates the open compact subgroups of (D, T(α) ), it follows that
the homomorphism ϕ : G → Aut(D, T(α) ) induced by the conjugation action indeed takes
its values in the group of homeomorphic automorphisms of (D, T(α) ). Let (gn )n∈I ⊂ G
be a net converging to the identity in G.
Let us first show that ϕ(gn ) converges pointwise to the trivial automorphism. Let x ∈ D.
Let L1 , . . . , Lk beTcompact locally normal representatives of α such that x ∈ L1 L2 . . . Lk .
Then the group ki=1 NG (Li ) is open in G, and therefore contains gn for all sufficiently
large n. In particular, for such a large n, we have ϕ(gn ).x = gn xgn−1 ∈ L1 L2 . . . Lk . The
latter product being compact with respect to the localised topology T(α) , it follows that
the net of commutators (gn xgn−1 x−1 )n∈I is bounded in (D, T(α) ). Therefore, all we need
to show is that 1 is its only accumulation point.
Suppose for a contradiction that some subnet ([gf (n) , x])n∈I ′ converges to z 6= 1. Choose
a compact locally normal representative L of α which is sufficiently small so that z 6∈ L.
We have [gf (n) , x] ∈ Lz for all large n. In particular [gf (n) , x] ∈ Oz, where O is any open
compact subgroup of G containing L. Since z 6∈ L, we may choose O small enough so
that it does not contain z. Then O and Oz are disjoint so that [gf (n) , x] 6∈ O for all large
n. Since O is an open subgroup of G, we deduce that ([gf (n) , x])n∈I ′ does not converge
to the identity in G, which is absurd.
Let now L be a compact locally normal representative of α and let (Un )n∈I be as in
Lemma 7.7; we may assume Un ≤ NG (L) for all n ∈ I. Then the groups Lm = L ∩ Um
form a basis of identity neighbourhoods in (D, T(α) ). Clearly, for each m ≥ 0, we have
gn ∈ NG (Lm ) for all sufficiently large n. Lemma 7.7 now implies that ϕ(gn ) converges
to the identity with respect to the Braconnier topology.
67
Corollary 7.10. Let G be a t.d.l.c. group, and let α be a fixed point of G in LN (G).
Let D be an abstract normal subgroup of G generated by compact locally normal representatives of α, and let H = Γ.D where Γ is a discrete normal subgroup of G. Then
the natural homomorphism G → Aut(H, T(α) ) induced by the G-action by conjugation is
continuous.
Proof. Note that Γ is a discrete normal subgroup of (H, T(α) ), and we have H = Γ.D.
Note also that D is an open subgroup of LN(G, α) with respect to the localised topology
T(α) , so the Braconnier topology on Aut(D, T(α) ) is compatible with the Braconnier
topology on Aut(LN(G, α), T(α) ).
Let (gn ) be a net tending to the identity in G; let K ⊂ H be T(α) -compact and V ≤ H
be a T(α) -compact open subgroup. We need to show that [gn , K] ∪ [gn−1 , K] is contained
in V for all sufficiently large n. Since H = Γ.D, it suffices to deal with the case when
K = γC with γ ∈ Γ and C ⊂ D is T(α) -compact. For n large enough, we have [gn , γ] = 1
since Γ is contained in the quasi-center of G. Therefore [gn , γC] = γ[gn , C]γ −1 . Observe
that W = γ −1 V γ is a compact open subgroup of (H, T(α) ). By Proposition 7.9, we have
[gn , C] ⊂ W for all sufficiently large n. Hence [gn , γC] ⊂ V for all large n. The argument
with gn−1 instead of gn is similar.
In general, a compact subgroup of a t.d.l.c. group that is commensurate with a locally
normal subgroup need not itself be locally normal. For example, if G has trivial quasicentre, then the only finite locally normal subgroup is the trivial one, but G may still
have non-trivial finite subgroups. Proposition 7.9 leads to an interesting criterion for
compact subgroups to be locally normal.
Proposition 7.11. Let G be a t.d.l.c. group and let α ∈ LN (G). Let L be a compact
subgroup of G such that [L] = α. Then L is locally normal in G if and only if L ≤
LN(G, α).
Proof. If L is locally normal, then L ≤ LN(G, α) by definition.
Assume conversely that L ≤ LN(G, α). Since Gα is open, it follows that a subgroup
of G contained in Gα is locally normal in G if and only if it is locally normal in Gα .
Moreover, every locally normal representative of α is contained in Gα . Thus we have
LN(G, α) = LN(Gα , α), and there is no loss of generality in assuming that G = Gα .
Let D be the group LN(G, α) equipped with the localised topology T(α) . Note that
L is compact with respect to T(α) . Thus L is an open compact subgroup of D. In
particular the collection A(L, L) of automorphisms of D which stabilise L is an identity
neighbourhood with respect to the Braconnier topology on Aut(D). Since the natural
map G → Aut(D) is continuous by Proposition 7.9, there must be an open subgroup of
G that normalises L.
68
If U is an open compact subgroup of G that fixes α, there is a useful description of
LN(G, α) ∩ U .
Proposition 7.12. Let G be a t.d.l.c. group, let α ∈ LN (G) and let U be an open
compact subgroup of G such that U fixes α. Let L be a closed subgroup of U such that
[L] ≤ α. Then L ≤ LN(G, α) if and only if there is a closed normal subgroup M of U
such that L ≤ M and [M ] = α.
Proof. By Lemma 7.4, there is a normal subgroup N of U such that [N ] = α; by
definition, N ≤ LN(G, α).
Let L be a closed subgroup of U such that [L] ≤ α. Suppose L ≤ LN(G, α). By replacing
L with LN we may assume N ≤ L. By Proposition 7.11, L is locally normal in G. In
particular, L/N is a finite locally normal subgroup of U/N , so NU/N (L/N ) has finite
index in U/N , and hence CU/N (L/N ) has finite index in U/N . Now by Dietzmann’s
Lemma, it follows that the normal closure M/N of L/N in U/N is finite and hence
closed, so M is a closed normal subgroup of U such that [M ] = α.
Corollary 7.13. Let U be a profinite group and let L be a closed subgroup. Assume that
L is commensurated and locally normal in U . Then the abstract normal closure of L in
U is closed, and contains L as a finite index subgroup.
Proof. Let N be the abstract normal closure of L. Since α = [L] is fixed by G by
hypothesis, we have N ≤ LN(G, α). The previous proposition therefore implies that α
has a representative M that is normal in U and contains L. Thus M is closed, the index
of L in M is finite and we have L ≤ N ≤ M . Therefore N is closed and contains L as a
finite index subgroup.
References
[1] R. Baer, Finiteness properties of groups, Duke Math. J. 15 (1948), no. 4, 1021–1032.
[2] Y. Barnea, M. Ershov, T. Weigel, Abstract commensurators of profinite groups,
Trans. Amer. Math. Soc. 363 (2011), no. 10, 5381–5417.
[3] L. Bartholdi, O. Siegenthaler, P. Zalesskii, The congruence subgroup problem for
branch groups, Israel Journal of Mathematics 187 (2012), Issue 1, 419–450.
[4] V. V. Belyaev, Locally finite groups with a finite nonseparable subgroup (Russian)
Sibirsk. Mat. Zh. 34 (1993), no. 2, 23–41, 226, 233. Translation in Siberian Math.
J. 34 (1993), no. 2, 218–232 (1994)
69
[5] N. Bourbaki, Groupes et algèbres de Lie. Chapitre II: Algèbres de Lie libres.
Chapitre III: Groupes de Lie. Actualités Scientifiques et Industrielles, No. 1349.
Hermann, Paris, 1972. 320 pp.
[6] M. Burger, Sh. Mozes, Groups acting on trees: from local to global structure, Publ.
Math. IHÉS 92 (2000), 113–150.
[7] P-E. Caprace, N. Monod, Decomposing locally compact groups into simple pieces,
Math. Proc. Cambridge Philos. Soc. 150 Nr. 1 (2011), 97–128.
[8] P-E. Caprace, C. D. Reid and G. A. Willis, Locally normal subgroups of totally disconnected groups. Part II: Compactly generated simple groups, preprint,
http://arxiv.org/abs/arXiv:1401.3142.
[9] P-E. Caprace, T. Stulemeijer, Totally disconnected locally compact groups
with a linear open subgroup, Int. Math. Res. Not. IMRN (to appear),
http://arxiv.org/abs/1409.8184.
[10] R. Cluckers, Y. Cornulier, N. Louvet, R. Tessera, A. Valette, The Howe-Moore
property for real and p-adic groups, Math. Scand. 109 (2011), no. 2, 201–224.
[11] A. P. Dietzmann, On the criteria of non-simplicity of groups. C. R. (Dokl.) Acad.
Sci. URSS, n. Ser. 44 (1944), 89–91.
[12] M. Hall Jr., The theory of groups. Reprinting of the 1968 edition. Chelsea Publishing
Co., New York, 1976.
[13] E. Hewitt and K. A. Ross, Abstract harmonic analysis, Berlin-Göttingen-Heidelberg
1963.
[14] N. Jacobson, Basic algebra. I. Second edition. W. H. Freeman and Company, New
York, 1985.
[15] S. Kakutani, K. Kodaira, Über das Haarsche Mass in der lokal bikompakten Gruppe.
Proc. Imp. Acad. Tokyo 20, (1944). 444–450.
[16] J. L. Kelley, General topology. Reprint of the 1955 edition. Graduate Texts in
Mathematics, No. 27. Springer-Verlag, New York-Berlin, 1975.
[17] L. Kramer, The topology of a semisimple Lie group is essentially unique, Adv. Math.
228 (2011), no. 5, 2623–2633.
[18] A. G. Kurosh, The theory of groups. Translated from the Russian and edited by K.
A. Hirsch. Chelsea Publishing Co., New York, 1960.
[19] R. McCallum, Rigidity of the group topology for p-adic semisimple groups and
automorphism groups of trees, http://arxiv.org/abs/arXiv:1210.1405.
70
[20] R. G. Möller, J. Vonk, Normal subgroups of groups acting on trees and automorphism groups of graphs, J. Group Theory 15 (2012), no. 6, 831–850.
[21] V. Nekrashevych, Free subgroups in groups acting on rooted trees, Groups Geom.
Dyn. 4 (2010), no. 4, 847–862.
[22] C. D. Reid, The generalised Fitting subgroup of a profinite group, Communications
in Algebra 41 (2013), no. 1, 294–308.
[23] C. D. Reid, The profinite completion of a group localised at a subgroup,
http://arxiv.org/abs/arXiv:1208.3554.
[24] G. Schlichting, Operationen mit periodischen Stabilisatoren, Arch. Math. (Basel)
34 (1980), no. 2, 97–99.
[25] P. Wesolek, Elementary totally disconnected locally compact groups, Proc. London
Math. Soc. (to appear).
[26] G. A. Willis, The structure of totally disconnected, locally compact groups, Math.
Ann. 300 (1994), no. 2, 341–363.
[27] J. S. Wilson, On Just Infinite Abstract and Profinite Groups, ch. 5 of New Horizons
in Pro-p groups, editors M. du Sautoy, D. Segal, A. Shalev, Birkhäuser, 2000.
71
| 4math.GR
|
A Lambda-Calculus with letrec, case,
constructors and non-determinism
arXiv:cs/0011008v1 [cs.PL] 6 Nov 2000
Manfred Schmidt-Schauß and Michael Huber
Fachbereich Informatik
Johann Wolfgang Goethe-Universität
Postfach 11 19 32
D-60054 Frankfurt, Germany
E-mail: [email protected]
Abstract. A non-deterministic call-by-need lambda-calculus λndlr with
case, constructors, letrec and a (non-deterministic) erratic choice, based
on rewriting rules is investigated. A standard reduction is defined as
a variant of left-most outermost reduction. The semantics is defined by
contextual equivalence of expressions instead of using αβ(η)-equivalence.
It is shown that several program transformations are correct, for example
all (deterministic) rules of the calculus, and in addition the rules for
garbage collection, removing indirections and unique copy.
This shows that the combination of a context lemma and a metarewriting on reductions using complete sets of commuting (forking, resp.)
diagrams is a useful and successful method for providing a semantics of
a functional programming language and proving correctness of program
transformations.
1
Introduction
Functional programming languages are based on extended lambda calculi and
the corresponding rewrite semantics. There are several methods of giving these
languages a semantics and proving the correctness of program transformations:
– A denotational semantics uses a mathematical domain and a mapping from
expressions to their denotation. This defines an equivalence of expressions,
which can be used to define a notion of correctness of program transformations. This area is well-developed, but reaches its limits if non-deterministic
operations are possible in the language.
– An operational semantics defining the evaluation of expressions (the execution, resp.). Sometimes this is used with a kind of syntactic equality (e.g.
αβ(η)-equality in the lambda-calculus). It could also be complemented by a
behavioral equivalence, which can be used to define the a notion of correctness of program transformation.
– A contextual semantics is a kind of operational semantics as above enhanced
with an approximation relation based on a contextual preordering (see e.g.
[Smi92,MST96,Pit97]). An expression s has less information than an expression t, iff in all contexts C[], if C[s] gives some information (e.g. terminates),
then C[t] also gives some information (i.e. terminates). This notion is directly adapted to define a notion of correctness of program transformation.
Often it gives the intuitive correct notion of program equivalence, and hence
also of correct program transformations.
The advantage of the contextual semantics is that the number of equality relations is maximal and that the derived properties are independent of a specific
domain. The properties of the contextual preorder are comparable to the orderings in domains; for example it is possible to use fixed-point constructions
for recursion. The contextual semantics is superior to the more syntax-oriented
αβ(η)-equivalence, since contextual semantics permits considerably more program transformations.
An advantage of contextual semantics over the denotational approach becomes
obvious if non-determinism is on board and also sharing in the form of a (nonrecursive or recursive) let. It appears to be very hard to construct a useful domain for denotational semantics in the presence of non-determinism and higherorder functions, whereas it is possible to use the contextual equivalence for defining an intuitive correct semantics. This can then be used to prove correctness
of program transformations sometimes exploiting rewriting techniques. A slight
disadvantage of the contextual semantics (w.r.t. economy of proofs) is that it
depends on the available syntactic constructs, hence on the set of contexts, and
the defined standard reduction.
The prominent syntactic property of the lambda-calculus is confluence of reduction [Bar84]. In the framework of a contextual semantics for the lambda-calculus
(see e.g. [Abr90]), confluence is not thus important and is replaced by the correctness of program transformations. The really interesting propositions are:
– Every beta-reduction transforms a program P into an equivalent one P ′ ,
meaning that P and P ′ are contextually equivalent. This is the required
modification of confluence.
– (standardization) Whenever there is a reduction of an expression t to an
abstraction, then the standard reduction terminates, i.e. reduces t to an
abstraction.
These properties can be generalized to extended lambda-calculi, where confluence may be false (see e.g. [AK94]), but contextual equivalence can be easily
adapted.
Another approach is Rewriting Logic (see e.g. [Mes00]), which is a step in the
direction of providing a semantics for programming languages based on rewriting rules. This appears to work for deterministic languages based on rewriting
rules. However, the contextual semantics is our method of choice for the nondeterministic case.
In this paper we present the calculus λndlr that is rather close to a non-strict
functional core language. Reduction is like lazy call-by-need evaluation in functional programming languages. λndlr can be seen as a generalization of the calculus in [KSS98,Kut00] and thus of the calculi in [AFM+ 95,AF97,MOW98], which
2
treat sharing in the lambda calculus. It also is a generalisation of [MSC99] insofar as the language of expressions is not restricted to have only variables
as arguments in applications. The calculus λndlr is related to the calculus in
[SS00], where a similar language is investigated, but with the emphasis on an
IO-interface.
Another method for treating sharing are explicit substitutions [ACCL91], which
optimize resource usage of reductions by exploiting sharing, however, it is i)
based on αβ(η)-equivalence and ii) the reduction rules are in general not compatible with non-determinism, i.e. not with λndlr nor with the calculus in [Kut00];
in particular, the let-over-lambda-rules are incompatible with non-determinism.
Specific ingredients of λndlr are
– sharing by using letrec, which moreover allows recursive definitions.
– a non-deterministic (erratic) choice, which allows to choose between two
expressions.
– a modified beta-reduction that prevents an unwanted duplication of nondeterministic expressions.
The motivation to investigate non-determinism is to model interfaces of lazy
functional languages to the outside world, i.e. to model input/output. This is
done by a simulation of an IO-action by a nested choice-expression that represents the set of possible input values of the IO.
The paper proposes to investigate extended lambda-calculi by using operational
methods and a contextual semantics. The contextual semantics includes a measure for the number of non-deterministic steps. As a method for proving program
transformation to be correct we propose to use complete sets of reduction diagrams in combination with an appropriate context lemma.
The results are that for λndlr a rather large set of basic program transformations
is proved to be correct. The paper also demonstrates the power of the method,
since the reduction rules of λndlr are numerous and complex.
As a check-program for complete sets, a program “Jonah“ was implemented to
automatically test the complete reduction diagrams using a generate-and-test
scheme; Jonah can also be used to compute proposals for complete sets.
As a remaining open problem the paper can be seen as a recommendation to
start an investigation into adapting the Knuth-Bendix method to automatically
computing the reduction diagrams. However, the reduction diagrams for (llet) for
example show that it would be necessary to integrate a kind of meta-description
like the Kleene-∗.
In this paper we do not present all proofs, but give enough hints and evidence
of how the claims can be verified and that they are valid.
2
The calculus λndlr
The syntax of the language is as follows:
There is a set of type-names. For every type there are constructors c coming with
an arity ar(c). This partitions the set of all constructors into the constructors
3
belonging to different types. For a type A, |A| is defined to be the number of
constructors belonging to A. The constructors belonging to type A are indexed,
and cA,i denotes the ith constructor of type A.
E ::= V | C | (choice s t) | (caseA E Alt1 . . . Alt|A| ) | (E1 E2 )
| (λV.E) | (letrec V1 = E1 , . . . Vn = En in E)
Alt ::= (P at → E)
P at ::= c V1 . . . Var(c)
where E, Ei are expressions, A is a type, V, Vi are variables, and C is a constructor. The variables in a pattern P at must be different, and also new ones.
Moreover, in a caseA -expression, there is exactly one alternative with a pattern of the form (cA,i y1 . . . yn ) for every constructor cA,i . The constants caseA
and choice can only occur in a special syntactic construction. Thus expressions
where choice or caseA is applied to a wrong number of arguments are not
allowed.
The structure letrec obeys the following conditions: The variables in the bindings are all distinct. We also assume that the bindings in letrec are commutative, i.e. can be commuted without syntactic change. letrec is recursive: I.e.
the scope of xi in (letrec xi = Ei in E) is Ei , E. This allows to define closed,
open expressions and α-renamings. For simplicity we use the disjoint variable
convention. I.e., all bound variables in expressions are assumed to be disjoint.
The reduction rules are such that the bound variables in the result are also
made distinct by α-renaming. We also use the convention to omit parenthesis in
denoting expression: (s1 . . . sn ) denotes (. . . (s1 s2 ) . . . sn ).
We say that an expression of the form (c t1 . . . tn ) is a constructor application,
if n ≤ ar(c). A constructor application of the form (c x1 . . . xn ) is called a pure
constructor application. An expression of the form (c t1 . . . tar(c) ) is called a
saturated constructor application.
Definition 2.1. Let R, R− , be context classes defined as follows:
R− ::= [] | R− E | (caseA R− alts)
R ::= R− | (letrec x1 = E1 , . . . , xn = En in R− )
| (letrec x1 = R1− [·], x2 = R2− [x1 ], . . . , xj = Rj− [xj−1 ], . . . in R− [xj ])
where Rj− is a context of class R−
R is called a reduction context and R− is called a weak reduction context. For a
term t with t = R− [t0 ], we say R− is maximal, iff there is no larger weak reduction context with this property. For a term t with t = R[t0 ], we say the reduction
context R is maximal, iff it is either a maximal weak reduction context, or of
the form (letrec x1 = R1− [·], x2 = R2− [x1 ], . . . , xj = Rj− [xj−1 ], . . . in R− [xj ]),
4
where t = (letrec x1 = t1 , . . . in R− [xj ]), R1− [·] is maximal for t1 , and the
number j is maximal.
For example the maximal reduction context of (letrec x2 = λx.x, x1 =
x2 x1 in x1 ) is (letrec x2 = [], x1 = x2 x1 in x1 ), in contrast to the nonmaximal reduction context (letrec x2 = λx.x, x1 = x2 x1 in []).
The (call-by-need) reduction rules defined in 2.2 follow the principle of minimizing copying at the cost of perhaps following several indirections. This holds for
the copy rule (cpn) as well as (case). The technical reason is that this principle
assures well-behaved reduction diagrams.
Definition 2.2. The reduction rules are defined below in figure 2. If the context
is important, then we denote it as a label of the reduction or state it explicitly.
Note that for (case), the typical example is written down, where the position
of the case is left open. There are two variants, one where the caseis in the
in-expression, and one where the case-expression is in the right hand side of
a binding. An exceptional case, where perhaps a letrec-expression has to be
omitted, is the case of a constructor with zero arguments like (case N il . . .).
The union of ndl, ndr is called (nd). Reductions are denoted using an arrow with
llet
super and/or subscripts: e.g. −−→. Transitive closure is denoted by a +, reflexive
∗
transitive closure by a ∗. E.g. −→ is the reflexive, transitive closure of →.
As a short comment of the reduction rules:
– (lbeta) is a sharing version of beta-reduction
– (cpn) is a lazy version of the replacement done by usual beta-reduction,
where the copy may jump over several indirections.
– (case) is the generalized if for case analysis of values. To find the value to
be analyzed, it has to be virtually assembled by following the bindings.
– (llet), (lapp), (lcase) are used to adjust the let-environments
– (nd) is the non-deterministic (erratic) choice.
The next definition is intended to formalize the standard reduction. The idea is
to find the reduction that is outermost, in a reduction context and also necessary
for making progress in the evaluation.
Definition 2.3. Let t be an expression. Let R be the maximal reduction context such that t ≡ R[t′ ] for some t′ . The standard redex and the corresponding
st
standard reduction −−→ is defined by one of the following cases:
– t′ is a choice-expression: then use (ndr) or (ndl).
– R = (letrec x1 = t1 , . . . xn = tn in []), and t′ is a letrec-expression. Then
apply (llet) to R[t′ ].
– R = R0 [(· t′′ )] where R0 is a reduction context. If t′ is a letrec-expression,
then use (lapp) in context R0 ; If t′ is an abstraction, then use (lbeta) in
context R0 .
5
– R = R0 [caseA · alts].
If t′ is a letrec-expression, then use (lcase) in context R0 ;
If t′ is a saturated constructor application, then use (case) in context R0 , if
it is applicable.
– R = (letrec x1 = [], x2 = x1 , . . . , xj = xj−1 , . . . in R1− [xj ])] where R1− is a
weak reduction context.
If t′ is an abstraction, then use (cpn) as follows: (letrec x1 = t′ , x2 =
x1 , . . . , xj = xj−1 , . . . in R1− [xj ])] → (letrec x1 = t′ , x2 = x1 , . . . , xj =
xj−1 , . . . in R1− [t′ ])].
If t′ is a letrec-expression, then use (llet),(lcase), or (lapp) to flatten the
letrec-expression t′ into its superexpression.
If t′ is a constructor application, and (case) is applicable to a case-expression
in a reduction context, then apply this (case)-reduction.
– R = (letrec x1 = [], x2 = x1 , . . . , xj = xj−1 , xj+1 =
−
−
−
Rj+1
[xj ], . . . in R∞
[xk ])] where Ri− , R∞
are weak reduction contexts.
′
If t is an abstraction, then use (cpn) such that the result is: R =
−
−
(letrec x1 = t′ , x2 = x1 , . . . , xj = xj−1 , xj+1 = Rj+1
[t′ ], . . . in R∞
[xk ])].
′
If t is a letrec-expression, then use (llet),(lcase), or (lapp) to flatten the
letrec-expression t′ into its superexpression.
If t′ is a constructor application, and (case) is applicable to a case-expression
in a reduction context, then apply this (case)-reduction.
Lemma 2.4. For every term t: if t has a standard redex, then this redex is
unique. If the standard reduction is not an (nd), then the standard reduction is
also unique.
st
st
st
Definition 2.5. A standard reduction s −−→ s1 −−→ s2 . . . sn −−→ t has ndcount D, iff D is the number of (nd)-reductions in it.
Note that we use the notion standard reduction also for non-maximal reductions.
Definition 2.6. For a term t and an nd-count D, t⇓D holds if there is some
standard-reduction starting with t, and the reduction has nd-count D.
Note that a standard reduction for an nd-count D is in general not unique. Note
also that there may be expressions without a standard redex.
Definition 2.7.
define:
(contextual preorder and equivalence) Let s, t be terms. We
s ≤c t iff ∀C[].∀D : C[s]⇓D ⇒ (∃B.D ≤ B ∧ C[t]⇓B )
s ∼c t iff s ≤c t ∧ t ≤c s
Note that we permit contexts such that C[s] is an open term.
6
Proposition 2.8. ≤c is a preordering and ∼c is an equivalence relation.
s ≤c t implies that C[s] ≤c C[t] for all contexts C[.]. I.e., ≤c is a precongruence
on the set of expressions.
s ∼c t implies that C[s] ∼c C[t] for all contexts C[.]. I.e., ∼c is a congruence on
the set of expressions.
Note that there are terms t without a standard redex, i.e. the standard reduction
stops. The reasons could be classified as i) type-error like (caseA (λx.x) . . .), ii)
a kind of non-termination like (letrec x = x in x), iii) as a value or a kind of
normal form like (Cons True Nil) or λx.x.
The following lemma shows that it is sufficient to use reductions contexts for
checking contextual approximation.
Lemma 2.9. (Context Lemma) Let s, t be terms. If for all reduction contexts
R and all nd-counts D: R[s]⇓D ⇒ (∃B.D ≤ B ∧ R[t]⇓B ), then s ≤c t.
Proof. We prove the more general claim:
if for all i: si , ti satisfy the conditions of the lemma for reduction contexts,
then for all multicontexts C[·1 , . . . , ·m ]: C[s1 , . . . , sn ]⇓D ⇒ (∃B.D ≤
B ∧ C[t1 , . . . , tn ]⇓B .
Assume this is false. Then there is a multicontext C, an nd-count D, such that
C[s1 , . . . , sn ]⇓D , and for all B with D ≤ B: C[t1 , . . . , tn ]6⇓B .
We select a multicontext, C, terms si , ti , and an nd-count D, and a corresponding
reduction, such that the counterexample is minimal w.r.t. the following lexicographic ordering: i) the number of reduction steps of C[s1 , . . . , sn ], ii) the number
of holes of C[. . .].
The search for a standard redex is performed top-down. There are two cases:
1. The search for the reduction context inspects the term in a hole. Then we
can assume wlog that the first hole is inspected first. Hence C[·, t2 , . . . , tn ]
is a reduction context. Let C ′ := C[s1 , ·2 , . . . , ·n ]. Since C ′ [s2 , . . . , sn ] ≡
C[s1 , . . . , sn ], we can select the the same standard reduction for nd-count D.
Since the number of holes of C ′ is smaller than the number of holes in C, we
obtain some B ≥ D with C ′ [t2 , . . . , tn ]⇓B , which means C[s1 , t2 , . . . , tn ]⇓B .
Since C[·, t2 , . . . , tn ] is a reduction context, the preconditions of the lemma
imply that there is some B ′ ≥ B with C[t1 , t2 , . . . , tn ]⇓B ′ , a contradiction.
2. The search for the reduction context does not inspect any hole of C. Then
C[s1 , . . . , sn ] as well as C[t1 , . . . , tn ] can be reduced using the same standard
reduction, since the search for a standard redex takes place only in the outer
context C[. . .]. There are two cases for a reduction:
st
If the reduction C[s1 , . . . , sn ] −−→ s′ is not (nd), then this may result in
C ′ [...] with more holes, and the holes are filled with copies of si , ti . Then we
get a smaller counterexample since the number of reductions steps is smaller,
and since non-(nd) standard reductions are unique.
7
st
If the reduction C[s1 , . . . , sn ] −−→ s′ is an (ndl) (or (ndr), respectively), then
the reduction of s′ has nd-count D′ = D − 1. We make the corresponding
st,ndl
reduction: C[t1 , . . . , tn ] −−→ t′ . This is a smaller counterexample; hence we
get a contradiction also in this case.
Definition 2.10. A program transformation is a relation T between programs
(expressions). A program transformation T is called correct, iff for all expressions P, P ′ : P T P ′ implies P ∼c P ′ .
The reductions rules in definition 2.2 define corresponding program transformations if they are allowed in arbitrary contexts.
Definition 2.11. Let an internal reduction be a non-standard reduction that
takes place within a reduction context. Usually, this is denoted by the label i on
the reduction arrow.
We define complete sets of commuting and forking diagrams adapted from
[Kut99,Kut00]. In the following definition we use a notation for rewrite rules
(i,llet)
(st,a)
(st,a)
(i,llet)
on reduction sequences. For example −→ · −→ ❀ −→ · −→ , where a
is a reduction type. The · on the left hand side is like a joker, and the · on the
right hand side can be seen as an existentially quantified term.
Definition 2.12. Assume given a reduction type (red) and a set of (complementary) reduction types T , where the base calculus reduction types are contained in
T , as well as (red).
A complete set of commuting diagrams for a reduction (red) is a set of rewrite
rules on reduction sequences of the form
i,red
st,a1
st,ak
−−→ . −−→ . . . −−→
st,b1
❀
st,bm
i,ch
i,c1
−−→ . . . −−→ . −−→ . . . . . −−→,
i,red
st,∗
where ci ∈ T , such that for every reduction sequence s −−→ . −−→ t: Either
it can be transformed using one of the meta-reductions into another reduction
i,red
st,∗
sequence from s to t, such that at least −−→ can be replaced. Or . −−→ t can be
st,∗
st,+
prolonged into a longer standard reduction sequence . . . −−→ t −−→ t′ , such that
it can be replaced as above.
A complete set of forking diagrams for a reduction (red) is a set of rewrite rules
on reduction sequences of the form
st,a1
st,ak
i,red
←−− . . . ←−− . −−→
i,c1
❀
i,ch
st,b1
st,bm
−−→ . . . . . −−→ . ←−− . . . ←−−,
st,∗
i,red
where ci ∈ T , such that: Either every reduction sequence s ←−− . −−→ t can
be transformed into another reduction sequence between s and t, such that at
i,red
st,∗
least −−→ is replaced. Or s ←−− . can be prolonged into a reduction sequence
st,+
st,∗
s′ ←−− s ←−− . , such that it can be replaced as above.
We also use reductions not in the base calculus as internal reductions.
8
It is intended that the corresponding meta-rewriting on reduction sequences
terminates, which has to be proved for every such complete set. The complete
sets of commuting (forking) diagrams are not unique.
Note that in many cases, the forking diagrams can be derived from the commuting diagrams.
Lemma 2.13. For every reduction that is not a (llet) or (cp)-reduction, i.e.,
a ∈ {(nd), (lbeta), (lapp), (lcase), (case)}, there are no internal reductions. This
means, every internal a-reduction with a ∈ {(nd), (lbeta), (lapp), (lcase), (case)}
is a standard reduction.
Proof. By inspecting all the finitely many cases.
a
Proposition 2.14. If s −−→ t, where a ∈ {(lbeta), (lapp), (lcase), (case)}, then
s ∼c t.
I.e., all the program transformations defined by one of the reductions
{(lbeta),(lapp),(lcase),(case)} are correct.
a,[]
Proof. Let s′ −−→ t′ by a (a)-reduction on the surface with a ∈
{(lbeta), (lapp), (lcase), (case)}.
We show s′ ≤c t′ exploiting the context lemma. Let R be a reduction context.
i,a
st,a
Then R[s′ ] −→ R[t′ ] is not possible by Lemma 2.13. Then R[s′ ] −→ R[t′ ] by a
unique standard reduction, hence if there a reduction for R[s′ ] with nd-count D,
there is also one for R′ [t′ ] with nd-count D. The context lemma now shows that
s ′ ≤ c t′ .
To show t′ ≤c s′ using the context lemma is similar: If there a standard reduction
for R[t′ ] with nd-count D, there is also one for R[s′ ] with nd-count D. The context
lemma now shows that t′ ≤c s′ .
Hence we have shown s′ ∼c t′ . Since ∼c is a congruence, we have also that
C[s′ ] ∼c C[t′ ] for an arbitrary context C. Hence the proposition holds.
3
Correctness of the reduction (llet)
The union of the reductions (llet),(lapp),(lcase) is denoted as (lll). The reduction
lll+ means a reduction sequence consisting only of (lll)-reductions of length at
least 1. Accordingly lll∗ is defined as any number of (lll)-reductions. (i, llet)0∨1
means no reduction or 1 reduction (i, llet). In the following two lemmas, a stands
a
for an arbitrary reduction −−→.
Lemma 3.1. A complete set of commuting diagrams for (llet) is:
(i,llet)
(st,a)
(st,a)
(i,llet)
(i,llet)
(st,a)
(st,a)
(st,llet)
(i,llet)
(st,lll+ )
– −−→ · −−→ ❀ −−→ · −−→
– −−→ · −−→ ❀ −−→ · −−→
– −−→ · −−→
(st,lll+ )
❀
(i,llet)0∨1
−−→ · −−−→
9
Lemma 3.2. A complete set of forking diagrams for (llet) is:
(st,a)
(i,llet)
– ←−− · −−→
(st,llet)
(st,a)
(i,llet)
(i,llet)
(st,a)
– ←−− · ←−− · −−→
(st,lll+ )
(i,llet)
– ←−−−− · −→
(st,lll+ )
(i,llet)
– ←−−−− · −→
(st,a)
❀ −−→ · ←−−
❀
❀ ←−−
(i,llet
(st,lll+ )
−→ · ←−−−−
(st,lll+ )
❀ ←−−−−
(i,llet)
Proposition 3.3. If s −−→ t, then s ∼c t.
I.e. (llet) is a correct program transformation in any context.
Proof. First we assume that the reduction is on top level.
To use the context lemma, we have to show what happens in a reduction context.
I.e. assume that s ≡ R[s′ ] and s′ is the llet-redex.
Using the forking diagrams, it is possible to construct from a standard reduction
of s a standard reduction of t with the same nd-count. The context lemma then
shows that s ≤c t.
Using the commuting diagrams, it is possible to construct from a standard reduction of t a standard reduction of s with the same nd-count by shifting the
(i,llet)
−−→ to the right. The context lemma then shows that t ≤c s.
Together, this means s ∼c t.
Finally, the congruence property of ∼c implies that a (llet) can be applied everywhere in a term. ✷
4
Garbage Collection: ldel
Garbage collection in the calculus has two forms, a non-cyclic one, and the other
one that also collects cyclic references:
The noncyclic reduction (ldel) is defined as :
(ldel) (letrec x = s in t) → t if x does not occur in t
(ldel) (letrec x = s, E in t) → (letrec E in t) if x does not occur in t, E
The cyclic reduction (ldelcyc) consisting of (ldelcyc1), (ldelcyc2) is defined as :
(ldelcyc1) (letrec x1 = s1 , . . . , xm = sm in t) → (letrec xj = sj , . . . , xm = sm in t)
if xi for 1 ≤ i ≤ j − 1 does not occur in sj , . . . , sm , t and m > 1
(ldelcyc2) (letrec x1 = s1 , . . . , xm = sm in t) → t
if xi for 1 ≤ i ≤ m does not occur in t
Here we show the correctness of (ldel).
Lemma 4.1. A complete set of commuting diagrams for (ldel) is:
(ldel)
(st,a)
(st,a)
(ldel)
– −−→ · −−→ ❀ −−→ · −−→
10
(ldel)
– −−→ ❀
(st,lll+ )
(st,lll+ )
(ldel)
ldel
−−→ · −−→
– −−→ · −−→
(st,lll∗ )
❀
(ldel)
−−→ · −−→
As an example for computing commuting diagrams, we show one case: We write
\ instead of λ.
Example 4.2. We compute the overlap of an (ldel)-redex and a standard (lapp)redex. If the overlap is trivial, then it is not hard to see that the reductions
commute, including the property “standard“.
In the case of a proper overlap, the redex and the corresponding reduction is as
follows:
ldel
((letrec x = c in \y.y) d) −−→ (\y.y d)
On the other hand, if first the (lapp) rule is applied, then:
st,lapp
((letrec x = c in \y.y) d) −−→ (letrec x = c in (\y.y
ldel
−−→ (\y.y
d))
d)
(ldel)
This is covered by the rule −−→ ❀
(st,lll+ )
ldel
−−→ · −−→.
Lemma 4.3. A complete set of forking diagrams for (ldel) is:
(st,a)
(ldel)
– ←−− · −→
+
(st,lll )
❀
(ldel)
– ←−−−− · −→
(st,lll+ )
(ldel)
– ←−−−− · −→
(ldel)
❀
❀
(st,a)
−→ · ←−−
(ldel)
−→
(ldel)
(st,lll∗ )
−→ · ←−−
Lemma 4.4. There are no infinite lll-reductions
Proof. This can be shown by a natural-number valuation of expressions similar
as in [Kut00], which is strictly decreasing in every reduction step. ✷
(ldel)
Proposition 4.5. If s −−→ t, then s ∼c t.
Proof. (sketch)
Follows by induction on the length of reductions from the context lemma, and
since there are no infinite (lll)-reduction sequences. ✷
5
Copying variables
This section contains the reduction (lcv) which is like compressing references
used in letrecs. It can also be described as removing indirections.
(lcv) (letrec x = y, E in C[x]) → (letrec x = y, E in C[y])
(lcv) (letrec x1 = y, x2 = C[x1 ], E in t)
→ (letrec x1 = y, x2 = C[y], E in t)
11
Lemma 5.1. A complete set of commuting diagrams for (lcv) is:
(lcv)
(st,a)
(st,a)
(lcv)
– −→ · −→ ❀ −→ · −→
(lcv) (st,cpn)
(st,cpn) (lcv) (lcv)
– −→ · −→
❀
−→ · −→ · −→.
(lcv) (st,a)
(st,a)
– −→ · −→ ❀ −→ , where a ∈ {case, cpn, ndr, ndl}.
Lemma 5.2. A complete set of forking diagrams for (lcv) is:
(st,a)
(lcv)
(lcv)
(st,a)
– ←−− · −−→ ❀ −−→ · ←−−
(st,cpn)
(lcv)
(lcv)
(lcv)
(st,cpn)
– ←−− · −−→ ❀ −−→ · −−→ · ←−−
(st,a)
(lcv)
(st,a)
– ←−− · −−→ ❀ ←−− for a ∈ {cp, case, ndl, ndr}.
(lcv)
Proposition 5.3. If s −−→ t, then s ∼c t.
I.e., (lcv) is a correct program transformation in any context.
The proof uses the context lemma, and the complete set of commuting and
forking diagrams to meta-reduce reduction sequences.
6
Contextual equivalence of copy reductions
The required diagrams and the proof of correctness of non-standard copy reductions are complex. Only the complete set of commuting diagrams are presented.
For this rule we require a special class of contexts: surface contexts: Surface
contexts define expressions with holes not in the body of an abstraction.
Definition 6.1.
S ::= [] | (S E) | (E S) | (caseA S alts) | (caseA E . . . (p → S) . . .)
| (choice E S) | (choice S E)
| ((letrec . . . in S) | ((letrec . . . , xi = S, . . . in E)
where E stands for an expression. S is called surface context.
We consider the following atomic copy reductions:
(cp) (letrec x1 = s1 , . . . , xn = sn in C[x1 ])
→ (letrec x1 = s1 , . . . , xn = sn in C[s1 ])
where s1 is an abstraction
(cp) (letrec x1 = s1 , . . . , xn = sn in s)
→ (letrec x1 = s1 , . . . , xj = C[s1 ], . . . , xn = sn in s)
where s1 is an abstraction
and where sj ≡ C[x1 ]
We distinguish the (cp)-reduction into two subreductions: If the target occurrence of the variable is in a surface context, then (cpt), otherwise it is a (cpd).
Equivalently, it is a (cpd) iff the target variable is within an abstraction.
Thus ((letrec x = s, E in D[λz.C[x]])) → ((letrec x = s, E in D[λz.C[s]]))
is a reduction of type (cpd).
12
Lemma 6.2. A complete set of commuting diagrams for (cpt), (cpd) is:
(i,cpt)
(st,a)
(i,cpd)
(st,a)
(i,cpd)
(st,lbeta)
– −→ · −→ ❀
(i,cpt) (st,a)
– −→ · −→ ❀
(st,a)
({i,st},cpt)
(st,a)
(i,cpd)
−→ · −−→
(st,a)
−→ , where a ∈ {case, ndr, ndl}.
– −→ · −→ ❀ −→ · −→
(i,cpd) (st,cpn)
(st,cpn) (i,cpd) (i,cpd)
– −→ · −→
❀
−→ · −→ · −→
(i,cpd) (st,a)
(st,a)
– −→ · −→ ❀ −→ , where a ∈ {case, ndr, ndl}.
– −→ ·
−→
❀
(st,lbeta)
−→
·
({i,st},cpt)
−−→ .
This is sufficient to show that the (cp)-reductions retain contextual equivalence
by a meta-reduction on reduction sequences.
(cp)
Proposition 6.3. If s −−→ t, then s ∼c t.
I.e., (cp) is a correct program transformation in any context.
In summary, we can prove:
Theorem 6.4. All the reductions of the base calculus with the exception of
(ndr), (ndl) are correct program transformations
It is obvious that (ndr), (ndl) are not correct as program transformations,
since (choice T rue F alse) may reduce to T rue, but T rue is not equivalent
to (choice T rue F alse).
7
Unique Copy: Inlining
If a letrec-bound variable occurs only once, then it is possible to replace this
variable by the bound expression and to remove the binding:
(ucp) (letrec x = s, E in C[x]) → (letrec E in C[s]), where C[] is a surface
context, s arbitrary, x has exactly one occurrence in C[x] and no
occurrence in E nor in s.
(ucp) (letrec x = s in C[x]) → C[s], where C[] is a surface context, s arbitrary,
and x has exactly one occurrence in C[x] and no occurrence in s.
(ucp) (letrec x = s, y = C[x], E in t) → (letrec y = C[s], E in t), where C[]
is a surface context, s arbitrary, x has exactly one occurrence in C[x]
and no occurrence in E,s and t.
Note that if s is an abstraction, then the rule is a combination of (cp) and (ldel).
Lemma 7.1. A complete set of commuting diagrams for (ucp) is:
(ucp)
(st,a)
– −→ · −→ ❀
(ucp) (st,a)
– −→ · −→ ❀
(ucp)
(st,a)
– −→ · −→
❀
(st,a)
(ucp)
−→ · −→
(ldel)
−→ · −→ for a ∈ {case, ndr, ndl, cpn}
(st,a)
(st,a)
−→ for a ∈ {case, ndr, ndl}
13
(ucp)
(st,lll∗ )
(st,lll∗ )
(ucp)
– −→ · −→ ❀ −→ · −→
(ucp)
where the extreme case −→ ❀
(ucp)
(st,cpn) (ldel)
– −→ ❀
−→ · −→
(ucp)
−→ is excluded.
(ucp)
Proposition 7.2. If s −−→ t, then s ∼c t.
I.e., (ucp) is a correct program transformation in any context.
8
Conclusion
The rewriting based method of computing complete sets of commuting (resp.
forking) diagrams to prove program transformations to be correct is demonstrated to be successful. We are able to show that all deterministic reduction
rules in the rather complex lambda calculus λndlr and also some other rules are
correct. A general automatic method to compute diagrams by checking all nontrivial overlaps would be a valuable tool and deserves further research efforts.
References
Abr90.
Samson Abramsky. The lazy lambda calculus. In D. Turner, editor, Research
Topics in Functional Programming, pages 65–116. Addison-Wesley, 1990.
ACCL91. M. Abadi, L. Cardelli, P.-L. Curien, and J.-J Lévy. Explicit substitutions.
J. functional programming, 4(1):375–416, 1991.
AF97.
Z.M. Ariola and M Felleisen. The call-by-need lambda calculus. J. functional
programming, 7(3):265–301, 1997.
AFM+ 95. Z.M. Ariola, M. Felleisen, J. Maraist, M. Odersky, and P. Wadler. A callby-need lambda calculus. In Principles of programming languages, pages
233–246, San Francisco, California, 1995. ACM Press.
AK94.
Z.M. Ariola and Jan Willem Klop. Cyclic lambda graph rewriting. In LICS
94, pages 416–425. IEEE Press, 1994.
Bar84.
H.P. Barendregt. The Lambda Calculus. Its Syntax and Semantics. NorthHolland, Amsterdam, New York, 1984.
KSS98.
Arne Kutzner and Manfred Schmidt-Schauß. A nondeterministic call-byneed lambda calculus. In International Conference on Functional Programming 1998, pages 324–335. ACM Press, 1998.
Kut99.
Arne Kutzner. A non-deterministic call-by-need lambda-calculus with erratic choice: operational semantics, program transformations and applications. draft of thesis, 1999.
Kut00.
Arne Kutzner. Ein nichtdeterministischer call-by-need Lambda-Kalkül mit
erratic choice: Operationale Semantik, Programmtransformationen und Anwendungen. Dissertation, J.W.Goethe-Universität Frankfurt, 2000. in german.
Mes00.
José Meseguer. Rewriting Logic and Maude: concepts and applications. In
Leo Bachmair, editor, Proceedings RTA’2000, pages 1–26. Springer-Verlag,
2000.
MOW98. John Maraist, Martin Odersky, and Philip Wadler. The call-by-need lambda
calculus. J. of Functional programming, 8:275–317, 1998.
14
MSC99.
MST96.
Pit97.
Smi92.
SS00.
A.K.D. Moran, D. Sands, and M. Carlsson. Erratic fudgets: A semantic
theory for an embedded coordination language. In Coordination ’99, volume
1594 of Lecture Notes in Computer Science. Springer-Verlag, 1999.
Ian Mason, Scott F. Smith, and Carolyn L. Talcott. From operational semantics to domain theory. Information and Computation, 128:26–47, 1996.
Andrew D. Pitts. Operationally-based theories of program equivalence. In
Semantics and Logics of Computation. Cambridge University Press, 1997.
S.F. Smith. From operational to denotational semantics. In 7th International Conference MFPS, LNCS 598, pages 54–76. Springer-Verlag, 1992.
M. Schmidt-Schauß. FUNDIO: A lambda-calculus with a recursive let, case,
constructors, and an IO-interface, 2000. draft.
15
(lbeta) ((λx.s) t) → (letrec x = t in s)
(cpn) (letrec x1 = s1 , x2 = x1 , . . . , xj = xj−1 , xj+1 = sj+1 . . . in C[xj ])
→ (letrec x1 = s1 , x2 = x1 , . . . , xj = xj−1 , xj+1 = sj+1 . . . in C[s1 ])
where s1 is an abstraction
(cpn) (letrec x1 = s1 , x2 = x1 , . . . , xj = xj−1 , xj+1 = C[xj ], . . . in s)
→ (letrec x1 = s1 , x2 = x1 , . . . , xj = xj−1 , xj+1 = C[s1 ], . . . in s)
where s1 is an abstraction
(llet) (letrec x1 = s1 , . . . , xn = sn in (letrec y1 = t1 , . . . , ym = sm in r))
→ (letrec x1 = s1 , . . . , xn = sn , y1 = t1 , . . . , ym = sm in r)
(llet) (letrec x1 = s1 , . . . , xi =
(letrec y1 = t1 , . . . , ym = tm in si ), . . . , xn = sn in r)
→ (letrec x1 = s1 , . . . , xn = sn , y1 = t1 , . . . , ym = sm in r)
(lapp) ((letrec xi = ti in t) s) → (letrec xi = ti in (t s))
(lcase) (caseA (letrec E in t) alts) → (letrec E in caseA t alts)
(case) (caseA (cA,i t1 . . . tn ) . . . ((cA,i y1 . . . yn ) → t) . . .)
→ (letrec y1 = t1 . . . yn = tn in t)
letrec x1 = (cA,i t1 . . . tj1 ),
x2 = x1 tj1 +1 . . . tj2 ,
...
(case)
xm = xm−1 tjm−1 +1 . . . tjm ,
...
C[caseA (xm tjm +1 . . . tjm+1 ) . . . ((cA,i z1 . . . zn ) → t)])
−→
letrec x1 = (cA,i y1 . . . yj1 ), y1 = t1 , . . . yj1 = tj1 ,
x2 = x1 yj1 +1 . . . yj2 , yj1 +1 = tj1 +1 , . . . , yj2 = tj2 ,
...
xm = xm−1 yjm−1 +1 . . . yjm , yjm−1 +1 = tjm−1 +1 , . . . , yjm = tjm ,
...
C[(letrec yjm +1 = tjm +1 , . . . , yjm+1 = tjm+1 , z1 = y1 , . . . , zn = yn in t)]
where n = jm and the case-expression may be in a bound or in the in-expression
and where yi are fresh variables
(ndl)
(ndr)
(choice s t) → s
(choice s t) → t
Fig. 1. Reduction rules of λndlr
16
| 6cs.PL
|
1
A Linearly Relaxed Approximate Linear
Program for Markov Decision Processes
arXiv:1704.02544v1 [cs.SY] 9 Apr 2017
Chandrashekar Lakshminarayanan? , Shalabh Bhatnagar? , and Csaba Szepesvari†
Abstract
Approximate linear programming (ALP) and its variants have been widely applied to Markov
Decision Processes (MDPs) with a large number of states. A serious limitation of ALP is that it has an
intractable number of constraints, as a result of which constraint approximations are of interest. In this
paper, we define a linearly relaxed approximation linear program (LRALP) that has a tractable number
of constraints, obtained as positive linear combinations of the original constraints of the ALP. The main
contribution is a novel performance bound for LRALP.
Keywords:
Markov Decision Processes (MDPs), Approximate Linear Programming (ALP),
I. I NTRODUCTION
Markov decision processes (MDPs) have proved to be an indispensable model for sequential
decision making under uncertainty with applications in networking, traffic control, robotics, operations research, business, finance, artificial intelligence, health-care and more (see, e.g., [Whi93;
Rus96a; FS02; HY07; SB10; BR11; Put94; LL12; AA+15; BD17]). In this paper we adopt the
framework of discrete-time, discounted MDPs when a controller steers the stochastically evolving
state of a system while receiving rewards that depends on the states visited and actions chosen.
The goal is to choose the actions so as to maximize the return, defined as the total discounted
expected reward. A controller that uses past state information is called a policy. An optimal
policy is one that maximizes the value no matter where the process is started from [Put94].
In this paper we consider planning problems where the goal is to calculate actions of policies
?
Department of Computer Science and Automation, Indian Institute of Science, Bangalore 560012. E-mail: {chandrul,
shalabh}@csa.iisc.ernet.in
†
Department of Computing Science, University of Alberta, Edmonton, Alberta, Canada T6G 2E8. E-mail:
[email protected]
April 11, 2017
DRAFT
2
that give rise to high values and give new error bounds on the quality of solutions obtained by
solving linear programs of tractable size. To explain the contributions in more details, we start
by describing the computational challenges involved in planning.
The main objective of planning is to compute actions of an optimal policy while interacting
with an MDP model. In finite state-action MDPs, assuming access to individual transition
probabilities and rewards along transitions, various algorithms are available to perform this
computation in time and space that scales polynomially with the number of states and actions.
However, in most practical applications, the MDP is compactly represented and if it is not
infinite, the number of states scale exponentially with the size of the representation of the MDP.
If planners are allowed to perform some fixed amount of calculations for each state encountered,
it is possible to use sampling to make the per-state calculation-cost independent of the size of
the state space [Rus96b; Sze01; KMN02]. Nevertheless, the resulting methods are still quite
limited. In fact, various hardness results show that computing actions of (near-) optimal policies
is intractable in various senses and in various compactly represented MDPs [BT00]. Given these
negative results, it is customary to adopt the modest goal of efficiently computing actions of a
policy that is nearly as good as a policy chosen by a suitable (computationally unbounded, and
well-informed) oracle from a given restricted policy class. Here, within some restrictions (see
below), the policy class can be chosen by the user. The more flexibility the user is given in this
choice, the stronger a planning method is. The problem of planning with limited resources is also
one of the key problem in artificial intelligence (AI). The book of [KM12] gives a relatively fresh,
algorithm-centered summary of existing methods suitable for planning in MDPs. AI research tend
to focus on empirical results through the development of various benchmarks and little if any
effort is devoted to the theoretical understanding of the quality-effort tradeoff exhibited by the
that the various algorithms that are developed in this field.
A popular approach along these lines, which goes back to Schweitzer and Seidmann [SS85],
relies on considering linear approximations to the optimal value function: The idea is that,
similarly to linear regression, a fixed sequence of basis functions are combined linearly. The
user’s task is to use a priori knowledge of the MDP to choose the basis functions so that a good
approximation to the optimal value function will exist in the linear space spanned by the basis
functions. The idea then is to design some algorithm to find the coefficients of the basis functions
that gives a good approximation, while keeping computation cost in check. Finding a good
approximation is sufficient, since at the expense of an extra O(1/ε2 ) randomized computation,
April 11, 2017
DRAFT
3
a uniform O(ε)-approximation to the optimal value function can be used to calculate an action
of an O(ε)-optimal policy at any given state (e.g., follow the ideas in [Sze01; KMN02]; see
also Theorem 3.7 of Kallenberg [Kal17]). Since the number of coefficients can be much smaller
than the number of states, the algorithms that search for the coefficients have the potential to
run efficiently regardless of the number of states.
Following Schweitzer and Seidmann [SS85], most of the literature considers algorithms that
are obtained from restricting exact planning methods to search in the span of the fixed basis
functions when performing computations. In this paper we consider the so-called approximate
linear programming (ALP) approach, which was heavily studied during the last two decades,
e.g., [SP01; Gue+03; FV03; FV04; KH04; PZ09; DFM09; Tay+10; PP11; BFM12; AYBM14].
The basic idea here is to combine a linear program whose solution is the optimal value function
(and thus the number of optimization variables in it scales with the number of states) with a
linear constraint that restricts the optimization variables to lie in the subspace spanned by the
basis functions. As already noted by Schweitzer and Seidmann [SS85], the new LP can still
be kept feasible by just adding one special basis function, while by substituting the “value
function candidates” with their linear expansions, the number of optimization variables becomes
the number of basis functions. As shown by de Farias and Van Roy [FV03], the solution to the
resulting LP is within a constant factor of the best approximation to the optimal value function
within the span of the chosen bases. However, since the number of constraints in the LP is still
proportional to the number of states, it is not obvious whether a solution to the resulting LP can
be found in time independent of the number of states (other computations can be done in time
independent of the number of states, e.g., using sampling, at the price of a controlled increase
of the error, e.g., Theorem 6 of [PZ09]).
Most of the literature is thus devoted to designing methods to select a tractable subset of
the constraints while keeping the approximation guarantees, as well as keeping computations
tractable. Since a linear objective is optimized by a point on the boundary of the feasible region,
knowing the optimizer would be sufficient to eliminate all but as many constraints as the number
of optimization variables. The question is how to find a superset of these, or an approximating set,
without incurring much computational overhead. Schuurmans and Patrascu [SP01] and Guestrin
et al. [Gue+03] propose constraint generation in a setting where the MDP has additional structure
(i.e., factorized transition structure). This additional structure is then exploited in designing
constraint generation methods which are able to efficiently generate violated constraints. A more
April 11, 2017
DRAFT
4
general approach due to de Farias and Van Roy [FV04] is to choose a random subset of the
constraints by choosing states to be included at random from a distribution that reflects the
“importance” of states. While constraint generation can be powerful, it is not known how solution
quality degrades with the budget on the constraints generated (Guestrin et al. note that the number
of constraints generated can be at most exponential in a fundamental quantity, the induced
width of a so-called cost-network, which may be large and is in general hard to control). For
constraint sampling, de Farias and Van Roy [FV04] prove a bound on the suboptimality, but
this bound applies only in the unrealistic scenario when the constraints are sampled from an
idealized distribution, which is related to the stationary distribution of an optimal policy. While
it is possible to extend this result to any sampling distribution, the bound then scales with
the mismatch between the sampling and the idealized distributions, which, in general, will be
uncontrolled. Another weakness of the bound is related to that when constraints are dropped, the
linear program may become unbounded. To prevent this, de Farias and Van Roy [FV04] propose
imposing an extra constraint on the optimization variables. The bound they obtain, however,
scales with the worst approximation error over this constraint set. While in a specific example it
is shown that this error can be controlled, no general results are derived in this direction. Later
works, such as that of Desai, Farias, and Moallemi [DFM09] and Bhat, Farias, and Moallemi
[BFM12], repeat the analysis of de Farias and Van Roy [FV04] in combinations with other ideas.
However, no existing work that we know of addresses the above weaknesses of the result of
de Farias and Van Roy [FV04].
Another interesting approach is to consider the dual linear program, where the optimization
variables are measures over the state-action space and the feasible set is the set of discounted
state-occupation (DSO) measures of all possible policies. By adding an extra linear constraint on
the optimization variables, we arrive at an “approximate dual LP”. Feasibility of the resulting LP
can be ensured by adding basis functions that represent DSO measures of some select policies.
Abbasi-Yadkori, Bartlett, and Malek [AYBM14] considered this approach together and proposed
to use a randomized gradient method to minimize a penalized form of the linear objective to
approximately enforce the constraints. The algorithm computes the parameters of a measure over
the state-action space, from which a policy can be derived by normalization. The main result of
Abbasi-Yadkori, Bartlett, and Malek is a bound on the performance loss of this policy relative
April 11, 2017
DRAFT
5
to the performance of the best DSO measure in the feasible set of the approximate dual LP,1
while iteration cost to obtain an ε-competitive solution is O(1/ε4 ) provided that a number of
conditions hold. On the above complexity bound, the constants hidden are instance dependent,
but do not depend on the number of states or actions. The conditions under which the result is
proven are as follows: (i) the algorithm needs to be able to sample from distributions not too
dissimilar to the idealized distributions q1 , q2 , where q1 is a distribution over state-action pairs, q2
P
is a distribution over states and, e.g., q1 is defined by q1 (s, a) = kφ(s, a)k/ s0 ,a0 kφ(s0 , a0 )k with
φ(s, a) = (φ1 (s, a), . . . , φk (s, a))> and φ1 , . . . , φk being the chosen basis functions specifying the
linear constraints and (s, a) is a state-action pair; (ii) the Markov chains underlying all policies
in the MDP are uniformly fast mixing; (iii) for any state s0 ∈ S and index 1 ≤ i ≤ k, the
P
expression s,a φi (s, a)pa (s, s0 ) can be evaluated in O(1) time, where pa (s, s0 ) is the probability
of transitioning from state s to s0 provided action a is chosen. While the second assumption
limits the scope of MDPs that the result can be applied to, the other two assumptions limit the
choice of the basis functions. Among other things, it is unclear how feasibility can be ensured
while satisfying (i). Nevertheless, Abbasi-Yadkori, Bartlett, and Malek demonstrate promising
empirical results on a queuing problem.
Our main contribution is a new suboptimality bound for the case when the constraint system
is replaced with a smaller, linearly projected constraint system. We also propose a specific
way of adding the extra constraint to keep the resulting LP bounded. Rather than relying on
combinatorial arguments (such as those at the heart of de Farias and Van Roy [FV04]), our
argument uses previously unexploited geometric structure of the linear programs underlying
MDPs. As a result our bound avoids distribution-mismatch terms and we also remove the scaling
with worst approximation error. A specific outcome of our general result is the realization that
it is beneficial to select states so that the “feature vectors” of all states when scaled with a fixed
constant factor are included in the conic hull of the “feature vectors” underlying the selected
states. This suggests to choose the basis functions so that this property can be satisfied by
selecting only a few states. As we will argue, this property holds for several popular choices of
basis functions. A preliminary version of this paper without the theoretical analysis and without
the geometric arguments was published in a short conference communication [LB15].
1
The result shown is more general, allowing to use measures outside of the feasible set. However, for such measures the
performance bound degrades very rapidly and hence the greater generality does not seem to add much to the result.
April 11, 2017
DRAFT
6
II. BACKGROUND
The purpose of this section is to introduce the necessary background before we can present
the problem studied and the main results.
We shall consider finite state-action space, discounted total expected reward MDPs. We note
in passing that the assumption that number of states is finite is mainly made for convenience
and at the expense of a more technical presentation could be lifted. We will comment later
on the assumption concerning the number of actions. Let the set of states, or state space be
S = {1, 2, . . . , S} and let the set of actions be A = {1, 2, . . . , A}. For simplicity, we assume
that all actions are admissible in all states. Given a choice of an action a ∈ A in a state s ∈ S,
the controller incurs a reward (or gain) of ga (s) ∈ [0, 1] and the state moves to a next state
s0 ∈ S with probability pa (s, s0 ). A policy u is a mapping from states to actions.2 When a policy
is followed, the state sequence evolves as a Markov chain with transition probabilities given by
Pu matrix whose (s, s0 )th entry is Pu(s) (s, s0 ). Along the way the rewards generated from gu
.
defined by gu (s) = gu(s) (s). The value of following a policy from a starting state s is denoted
by Ju (s) and is defined as the expected total reward discounted reward. Thus,
∞
. X t t
α (Pu gu )(s) ,
Ju (s) =
t=0
where α ∈ (0, 1) is the so-called discount factor. We call Ju the value function of policy u. The
value function of a policy satisfies the fixed-point equation Ju = Tu Ju where the affine-linear
operator Tu is defined by Tu J = gu + αPu J. An optimal policy, is one that maximizes the
value simultaneously for all initial states. The optimal value function J ∗ is defined by J ∗ (s) =
maxu Ju (s) and is known to be the solution of the fixed-point equation J ∗ = T J ∗ where the
operator T is defined by (T J)(s) = maxu (Tu J)(s), s ∈ S, i.e., the maximization is componentwise. Optimal policies exist and in fact any policy u such that the equation Tu J ∗ = T J ∗ holds
is optimal (e.g., Corollary 3.3 of [Kal17]). A policy u is said to be greedy with respect to (w.r.t.)
J if Tu J = T J ∗ . Thus, any policy that is greedy w.r.t. J ∗ is optimal.
III. T HE L INEARLY R ELAXED ALP
In this section we introduce the computational model used and the “Linearly Relaxed Approximate Linear Program” a relaxation of the ALP.
2
For the scope of this paper, it suffices to restrict our attention to such policies as opposed to considering history dependent
policies. See Chapter 3, and specifically Corollary 3.3 of [Kal17].
April 11, 2017
DRAFT
7
As discussed in the introduction, we are interested in methods that compute a good approximation to the optimal value function. As noted earlier, at the expense of a modest additional
cost, knowing an O(ε) approximation to J ∗ at a few states suffices to compute actions of an
O(ε)-optimal policy. We will take a more general view, and we will consider calculating good
approximations to J ∗ with respect to a weighted 1-norm, where the weights c form a probability
distribution over S. Recall that the weighted 1-norm kJk1,c of a vector J ∈ RS is defined
P
S
as kJk1,c =
s c(s)|J(s)|. Note that here and in what follows we identify elements of R
(functions, mapping S = {1, . . . , S} to the reals) with elements of RS in the obvious way. This
P
allows us to write e.g. c> J, which denotes s c(s)J(s).
To introduce the optimization problem we study, first recall that the optimal value function
J ∗ is the solution of the fixed point equation T J ∗ = J ∗ . It follows from the definition of T
that J ∗ = maxu Tu J ∗ ≥ Tu J ∗ for any u, where ≥ is the componentwise partial ordering of
vectors (≤ is the reverse relation). With some abuse of notation, we also introduce Ta to denote
Tu where u(s) = a for any s ∈ S. It follows that J ∗ ≥ Ta J ∗ for any a ∈ A and also that
T = maxa Ta , where again the maximization is componentwise. We call a vector J that satisfies
J ≥ Ta J for any a ∈ A superharmonic. Note that this is a set of linear inequalities. By our
note on T and (Ta )a , these inequalities can also be written compactly as J ≥ T J. It is not
hard to show then that J ∗ is the smallest superharminoc function (i.e., for any J superharmonic,
.
J ≥ J ∗ ). It also follows that for any c ∈ RS++ = (0, ∞)S , the unique solution to the linear
program min{c> J : J ≥ T J} = min{c> J : J ≥ Ta J, a ∈ A} is J ∗ .
Now, let φ1 , . . . , φk : S → R be k basis functions. The Approximate Linear Program (ALP)
P
of Schweitzer and Seidmann [SS85] is obtained by adding the linear constraints J = ki=1 ri φi
to the above linear program. Eliminating J gives
X
X
X
min{
ri c> φi :
ri φi ≥ ga + α
ri Pa φi , a ∈ A, r = (ri ) ∈ Rk } .
i
i
i
As noted by Schweitzer and Seidmann [SS85], the linear program is feasible as long as 1,
defined as the vector with all components being identically equal to one, is in the span of
{φ1 , . . . , φk }. For the purpose of computations, it is assumed that the values c> φi , i = 1, . . . , k
and the values (Pa φi )(s) and ga (s) can be accessed in constant time. This assumption can be
relaxed to assuming that one can access ga (s) and φi (s) for any (s, a) in constant time, as well
as to that one can efficiently sample from c, from Pa (s, ·) for any (s, a) pair, but the details of
this are the beyond the scope of the present work. As shown by de Farias and Van Roy [FV03],
April 11, 2017
DRAFT
8
.
. P
if rALP denotes the solution to the above ALP then for JALP = i rALP (i)φi = ΦrALP it holds
that kJALP − J ∗ k1,c ≤
2ε
,
1−α
where ε = inf r kJ ∗ − Φrk∞ is the error of approximating the optimal
value with the span of the basis functions φ1 , . . . , φk and kJk∞ = maxs |J(s)| is the maximum
norm and Φ ∈ RS×k is the matrix formed by (φ1 , . . . , φk ). That the error of approximating J ∗
with JALP is O(ε) is significant: The user can focus on finding a good basis, leaving the search
for the “right” coefficients to a linear program solver.
While solving the ALP can be significantly cheaper than solving the LP underlying the MDP
and thus it can be advantageous for moderate-scale MDPs, the number of constraints in the
ALP is SA, hence the ALP is still intractable for huge-scale MDPs. To reduce the number of
constraints, we consider a relaxation of ALP where the constraints are replaced with positive
linear combinations of them. Recalling that the constraints took the form J ≥ ga + αPa J (with
J = Φr), choosing m to be target number of constraints, for 1 ≤ i ≤ m, the ith new constraint
P >
P >
is given by a wi,a
J ≥ a wi,a
(ga + αPa J), where the choice of m and that of the vectors
wi,a ∈ RS+ is left to the user. Note that this results in a linear program with k variables and m
constraints, which can be written as
min c> Φr
r∈Rk
s.t.
X
a
Wa> Φr ≥
X
Wa> (ga + αPa )Φr ,
(1)
a
where Wa = (w1,a , . . . , wm,a ) ∈ RS×m
. Note that the (i, j)th entry of the m × k constraint
+
P >
P >
matrix of the resulting LP is a wi,a φj −α a wi,a
Pa φj and assuming that (wi,a )a has p nonzero
elements, this can be calculated in O(p) time, making the total cost of obtaining the constraint
matrix to be O(mkp) regardless the value of S and A.
We will call the LP in (1) the linearly relaxed approximate linear program (LRALP). Any
LP obtained using any constraint selection/generation process can be represented by choosing
an appropriate binary-valued matrix W > = (W1> , . . . , WA> ) ∈ Rm×SA
. In particular, when the
+
constraints are selected in a random process as suggested by de Farias and Van Roy [FV04], the
matrix W would be a random, binary-valued matrix.
Note that the LRALP may be unbounded. Unboundedness could be avoided by adding an
April 11, 2017
DRAFT
9
extra constraint of the form r ∈ N to the LRALP, for a properly chosen polyhedron N ⊂ Rk .3
However, it seems to us that it is downright misleading to think that guaranteeing a bounded
solution will also lead to reasonable solutions. Thus we will stick to the above simple form,
forcing a discussion of how W should be chosen to get meaningful results.4
Further insight into the choice of W can be gained by considering the Lagrangians of the
ALP and LRALP. To write both LP’s in a similar form let us introduce E = (IS×S , . . . , IS×S )> ,
where IS×S is the S × S identity matrix. Further, let H : RS → RSA be the operator defined by
(HJ)> = ((T1 J)> , . . . , (TA J)> ) .
Note that H, which we call the linear Bellman operator, is a linear operator. Then, the ALP
can be written as
min{c> Φr | EΦr ≥ HΦr} ,
(ALP)
while LRALP takes the form
min{c> Φr | W > EΦr ≥ W > HΦr} .
(LRALP)
Hence, their Lagrangians are LALP (r, λ) = c> Φr + λ> (HΦr − EΦr) LLRALP (r, q) = c> Φr +
q > W > (HΦr − EΦr). Thus, we can view W q as a “linear approximation” to the dual variable
λ ∈ RSA
+ . This suggests that perhaps W should be chosen such that it approximates well the
optimal dual variable. If Φ spans RS , the optimal dual variable λ∗ is known to be the discounted
occupancy measure underlying the optimal policy (Theorem 3.18, [Kal17]), suggesting that the
role of W is very similar to the role of Φ excepts that the subspace spanned by the columns of
W should ideally be close to λ∗ .
3
In particular, to obtain their theoretical result, de Farias and Van Roy [FV04] need the assumption that the set N is bounded
and that it contains rALP . In fact, the error bound derived by de Farias and Van Roy depends on the worst error of approximating
J ∗ with Φr when r ranges over N . Hence, if N is unbounded, their bound is vacuous. In the context of a particular application,
de Farias and Van Roy [FV04] demonstrate that N can be chosen properly to control this term. However, no general construction
is presented to choose N .
4
The only question is whether there is some value in adding constraints beyond choosing W properly. Our position is that
the set N would most likely be chosen based on very little and general information; the useful knowledge is in choosing W ,
not in choosing some general set N . Since randomization does not guarantee bounded solutions, de Farias and Van Roy [FV03]
must use N : In their case, N incorporates all the knowledge that makes the LP bounded.
April 11, 2017
DRAFT
10
IV. M AIN R ESULTS
The purpose of this section is to present our main results. Let rLRA be a solution to the LRALP
given by (1) and let JLRA = ΦrLRA . When multiple solutions exist, we can choose any of them.
For the result, we assume that the LRALP is not unbounded, and hence a solution exist. In fact,
we will assume something much stronger. The discussion of why our assumptions are reasonable
and how to ensure that they hold is postponed to after the presentation of our results. Our main
results bounds the error kJ ∗ − JLRA k1,c .
The bound is given in terms of the approximation error of J ∗ with the basis functions Φ =
∗
∗
(φ1 , . . . , φk ), as well as the deviation between two functions, JALP
, JLRA
: S → R, which we
define next. In particular,
∗
JALP
(s) = min{r> φ(s) | Φr ≥ J ∗ , r ∈ Rk } ,
∗
JLRA
(s) = min{r> φ(s) | W > EΦr ≥ W > EJ ∗ , r ∈ Rk } ,
where s ∈ S. Recall that E : RS → RSA is defined so that (EJ)> = (J > , . . . , J > ), i.e., E
P
∗
stacks its argument A-fold. Hence, W > E = a Wa> . Our strong assumption is that JLRA
is
∗
finite-valued. Note that JALP
≥ J ∗ reflects the error due to using the basis functions (φj )j , and
∗
∗
the magnitude of the deviation JLRA
− JALP
reflects the error introduced due to the relaxed
constraint system.
Following de Farias and Van Roy [FV03; FV04], we will quantify the magnitude of the error
∗
∗
JLRA
− JALP
and also that of the error of approximating J ∗ with the subspace spanned by Φ, in
terms of a weighted maximum norm, kJk∞,ψ = maxs∈S |J(s)|/ψ(s), where ψ : S → R++ is a
positive-valued weighting function.5 As also stressed by de Farias and Van Roy, the appropriate
choice of ψ is crucial for MDPs with huge state-spaces: The problem is that if the range of
values of |J ∗ (s)| in different parts of the state space differ in orders of magnitude, we do not
expect to be able to control the error of approximating J ∗ uniformly over S. By choosing the
weighting function to reflect the magnitude of J ∗ , the weighted maximum norm is controlled as
soon as the relative errors are and this latter goal may be much easier to achieve than controlling
absolute errors.
5
As opposed to de Farias and Van Roy [FV03] and others, our definition uses division and not multiplication with the weights.
We choose this form for mathematical convenience: With this definition, nice duality results hold between weighted 1-norms
and weighted maximum norms.
April 11, 2017
DRAFT
11
Just like de Farias and Van Roy [FV03], we will also require that ψ is a stochastic Lyapunovfunction for the MDP. In particular, we require that the α-discounted stability coefficient
.
βψ = α max kPa ψk∞,ψ
a
(2)
is strictly less than one. This can be seen to imply that H : (RS , k · k∞,ψ ) → (RSA , k · k∞,ψ ) is
a contraction, where for J = (J1> , . . . , JA> )> ∈ RSA we let kJk∞,ψ = maxa kJa k∞,ψ . That H is
a contraction will play a crucial role in our results. Note that the condition βψ < 1 is closely
related to the condition that for any policy u, Pu ψ ≤ ψ, which can be viewed as a stability
condition on the MDP and which appeared in a slightly altered form in studying the stability
of MDPs with infinite state spaces [e.g., CM99]. Note also that one can always choose ψ = 1,
which gives β1 = α < 1. With this, we are ready to state our main result:
Theorem IV.1 (Error Bound for LRALP). Assume that c ∈ RS+ is such that 1> c = 1 and that
W ∈ RSA×m
is nonnegative valued. Let ψ ∈ RS+ be in the column span of Φ and assume that
+
the α-discounted stability coefficient of ψ is βψ < 1. Let ε = inf r∈Rk kJ ∗ − Φrk∞,ψ be the error
of approximation J ∗ using the basis functions in Φ. Then,
kJ ∗ − JLRA k1,c ≤
2c> ψ
∗
∗
(2.5ε + kJALP
− JLRA
k∞,ψ ) .
1 − βψ
Note that the result implicitly assumes that JLRA exists, because if JLRA does not exist then
∗
JLRA
is necessarily unbounded, making the last error term infinite. To ensure that ψ is in the
span of Φ, after choosing ψ, one can add ψ as one of the basis functions. Alternatively, the
bound can also be interpreted to hold for any ψ in the span of Φ with βψ < 1.
As noted earlier, de Farias and Van Roy [FV03] prove a similar error bound for JALP , the
solution of the ALP. In particular, their Theorem 3 states that under identical assumptions as in
our result, kJ ∗ − JALP k1,c ≤
2c> ψε
1−βψ
for ε defined as above (the result we cited previously is a
simplified form of this bound). The larger coefficient of ε is probably an artifact of our analysis.
Note that when W does not reduce the constraints, our bound is only a constant factor larger
∗
∗
than this previous result. The extra term kJALP
− JLRA
k∞,ψ can be seen as the price paid for
relaxing the constraints.
From linear programming theory, it follows that primal boundedness is equivalent to dual
feasibility. Since the dual of min{c> x : Ax ≥ b} is max{y > b : y ≥ 0, c = A> y}, we get
∗
that a necessary and sufficient condition for JLRA
to be finite-valued is that for any s ∈ S,
> >
φ(s) lies in the conic span, {U λ : λ ∈ RSA
+ }, of (the columns) of U = Φ E W . When W
April 11, 2017
DRAFT
12
is such that its constituents W1 , . . . , WA are all identical, the conic span of U is equal to the
conic span of Φ> W1 . It is particularly instructive to consider the case when the common matrix
Wa = (w1,a , . . . , wm,a ) “selects” the m states, i.e., when {w1,a , . . . , wm,a } = {es : s ∈ S0 } for
some S0 ⊂ S, |S0 | ≤ m, where es ∈ {0, 1}S are the s ∈ S vectors in the standard Euclidean
basis. In this case, the condition that φ(s) lies in the conic span of U is equivalent to that φ(s)
.
∗
lies in the conic span of φ(S0 ) = {φ(s0 ) : s0 ∈ S0 }. Thus, to ensure boundedness of JLRA
, the
chosen states should be selected to “conicly cover” all the vectors in φ(S) ⊂ Rk .6
The next theorem shows that magnitudes of the coefficients used in the conic cover control the
∗
∗
0
be the matrix of conic coefficients:
size of kJALP
− JLRA
k∞,ψ . For the theorem we let Λ ∈ RS×S
+
P
For any s ∈ S, φ(s) = s0 ∈S0 Λ(s, s0 )φ(s0 ). After the theorem we give constructions for creating
conic covers.
Theorem IV.2. Assume that W1 = · · · = WA , {w1,a , . . . , wm,a } = {es : s ∈ S0 } and that φ(S)
lies in the conic span of φ(S0 ) with conic coefficients given by Λ. Let ε = inf r kJ ∗ − Φrk∞,ψ .
Then,
∗
∗
∗
kJALP
− JLRA
k∞,ψ ≤ kJALP
− J ∗ k∞,ψ + (1 + kΛψkψ,∞ )ε .
Proof. Let r∗ be such that kJ ∗ − Φr∗ k∞,ψ = ε (this exists by continuity) and let δ = J ∗ − Φr∗ .
∗
Pick any s ∈ S and let rs = argmin{r> φ(s) : W > EΦr ≥ W > EJ ∗ , r ∈ Rk } so that JLRA
(s) =
∗
rs> φ(s). Note that by assumption, for any s0 ∈ S0 , JLRA
(s0 ) = rs> φ(s0 ) ≥ J ∗ (s0 ). Now, notice
∗
∗
∗
∗
that by definition, JLRA
≤ JALP
(the LP defining JLRA
is the relaxation of the LP defining JALP
).
Hence,
∗
∗
∗
∗
0 ≤ JALP
(s) − JLRA
(s) = JALP
(s) − J ∗ (s) + J ∗ (s) − JLRA
(s)
∗
and JLRA
(s) = rs> φ(s) = rs>
P
s0 ∈S0
Λ(s, s0 )φ(s0 ) =
P
s0 ∈S0
∗
Λ(s, s0 )JLRA
(s0 ) ≥
P
s0 ∈S0
Λ(s, s0 )J ∗ (s0 ).
Combining this with the previous inequality we get
∗
J ∗ (s) − J ∗ (s) J ∗ (s) −
J ∗ (s) − JLRA
(s)
≤ ALP
+
0 ≤ ALP
ψ(s)
ψ(s)
6
P
Λ(s, s0 )J ∗ (s0 )
.
ψ(s)
s0 ∈S0
The same implies that, under the same condition on W , boundedness of the LRALP holds if and only if
P
s
c(s)φ(s) is
in the conic span of φ(S0 ). Note that this is easy to fulfill if the support of c has a small cardinality by add all states in the
support of c to S0 .
April 11, 2017
DRAFT
13
Plugging in J ∗ (s) = φ(s)> r∗ +δ(s), using again that φ(s) =
P
s0 ∈S0
Λ(s, s0 )φ(s0 ), and also using
the triangle inequality after taking absolute values, we get
P
P
|J ∗ (s) − s0 ∈S0 Λ(s, s0 )J ∗ (s0 )|
Λ(s, s0 )|δ(s0 )|
|δ(s)|
0
≤
+ s ∈S0
ψ(s)
ψ(s)
ψ(s)
1 X
|δ(s0 )|
|δ(s)|
+
Λ(s, s0 )ψ(s0 )
≤
ψ(s)
ψ(s) s0 ∈S
ψ(s0 )
0
P
Λ(s, s0 )ψ(s0 )
|δ(s)|
0
≤
+ kδk∞,ψ s ∈S0
.
ψ(s)
ψ(s)
Combining this with the previous display and noting that kδk∞,ψ = ε finishes the proof.
Given φ : S → Rk , what is the minimum cardinality set S0 that conicly covers φ(S) and
how to find such a set? Further, how to keep the magnitude of kΛψk∞,ψ small? To control this
latter quantity it seems essential to make sure S0 contains states with high ψ-values. However, if
one is content with a bound that depends on kψk∞ , one can bound kΛψk∞,ψ by kψk∞ ζ where
P
ζ = maxs s0 ∈S0 Λ(s, s0 ), hence, the second term in the previous bound will be bounded by
(1 + kψk∞ ζ)ε.
Let us now return to the problem of finding conic covers. We will proceed by considering
some illustrative examples. As a start, consider the case when the basis functions are binary
valued. In this case, it is sufficient and necessary to choose one state for each binary vector
.
that appears in φ(S) ⊂ {0, 1}k . This gives that m0 = |S0 | ≤ 2k representative states will be
sufficient regardless of the cardinality of S. Further, in this case ζ = 1. For moderate to large
k (e.g., k 20), it will quickly become infeasible to keep 2k constraints. In this case we may
need to restrict what features are considered to guarantee the conic cover condition. Letting
Ai = {s ∈ S : φi (s) = 1}, if for a many pairs i 6= j, Ai and Aj do not overlap then N = |φ(S)|
can be much smaller than 2k . For example, in the commonly used state aggregation procedures
Ai ∩Aj = ∅ for any i 6= j, giving N = k. In the more interesting case of hierarchical aggregation
(when the sets {Ai } form a nested hierarchical partitioning of S), we have m0 ≤ D · k where
D is the depth of the hierarchy.
Another favourable example is the case of separable bases. In this case, the states are assumed
to be factored and the basis functions depend only on a few factors. Let us consider a simple
illustration. By abusing notation (redefining S), let S = S1 ×S2 , let there be k = 2 basis functions
and assume that φi (s) = hi (s) for some hi : Si → R, i = 1, 2. Assume further that 0 ∈ hi (Si )
for both i and specifically let si0 be such that hi (si0 ) = 0. In this case it is not hard to verify that
April 11, 2017
DRAFT
14
if Si0 is such that hi (Si ) is in the conic span of hi (Si0 ) then φ(S) is also in the conic span of
.
S0 = S1 × {s20 } ∪ {s10 } × S2 . The point is that |S0 | ≤ |S1 | + |S2 |, which is a tolerable increase of
growth. This example is not hard to generalize to more general, ANOVA-like basis expansions.
The moral is that as long as their limited order of interaction (which is usually necessary for
information theoretic reasons as well), the number of constraints may grow moderately with the
number of factors (dimensionality) of the state space.
In some cases, finding a conic cover with a small cardinality is not possible. This can already
happen in simple examples such as when S = {1, . . . , S} (as before) and φ(s) = (1, s, s2 )> .
In this case, the only choice is S0 = S. In examples similar to this one one possibility is to
quantize the range of φ, which may loose little on approximation quality, while it creates the
opportunity to construct a small cardinality conic cover.
Note that the bound of [FV04] and our main result can be seen as largely complementary.
Recall that de Farias and Van Roy consider adding an extra constraint r ∈ N , while they
propose to select all A constraints from the ALP corresponding to m states chosen at random
from some distribution µ. Then, with high probability, they show that, provided that rALP ∈ N ,
the extra price paid for relaxing the constraints of the ALP is O(ρεN k/m), where ρ = maxs
µ∗ (s)
,
µ(s)
µ∗ = (1 − α)c> (I − αPu∗ )−1 , u∗ is an optimal policy, and εN = supr∈N kJ ∗ − Φrk∞,ψ .7 The
bound is nontrivial when m ≥ ρεN k. In general, it may be hard to control ρ, or even εN while
ensuring that rALP ∈ N .
V. P ROOF OF T HEOREM IV.1
In this section we present the proof of the main result, Theorem IV.1. The proof uses
contraction-arguments. We will introduce a novel contraction operator, Γ̂ : RS → RS , that
captures the distortion introduced by the extra constraint in ALP and the relaxation in LRALP,
respectively. Then we relate the solution of LRALP to the fixed point of Γ̂.
∗
Note that for the proof it suffices to consider the case when JLRA
is finite-valued because
otherwise the bound is vacuous. Also, recall that it was assumed that ψ lies in the column space
of Φ, while βψ , the α-discounted stability of ψ w.r.t. the MDP (cf. (2)) is strictly below one.
We will let r0 ∈ Rk be such that ψ = Φr0 . We also assumed that the matrix W is nonnegative
P
valued, while c specifies a probability distribution over S: s c(s) = 1 and c ∈ RS+ .
7
The paper presents the results for µ = µ∗ giving ρ = 1, but the analysis easily extends to the general case.
April 11, 2017
DRAFT
15
The operator Γ̂ are defined as follows: For J ∈ RS , s ∈ S,
(Γ̂J)(s) = min{r> φ(s) : W > EΦr ≥ W > EHJ, r ∈ Rk } .
Note that (Γ̂J)(s) mimics the definition of ALP with c = es , except that the constraint J = Φr
is dropped.
Let us now recall some basic results from the theory of contraction maps. First, let us recall
the definition of contractions. Let k · k be a norm on RS and ρ > 0. We say that the map
B : RS → RS is (ρ, k · k)-Lipschitz if for any J, J 0 ∈ RS , kBJ − BJ 0 k ≤ ρkJ − J 0 k. We say
that B is a k · k-contraction with factor ρ if it is (ρ, k · k)-Lipschitz and ρ < 1. It is particularly
easy to check whether a map is a contraction map with respect to a weighted maximum norm
if it is known to be monotone. Here, B is said to be monotone if for any J ≤ J 0 , J, J 0 ∈ RS ,
BJ ≤ BJ 0 also holds, where ≤ is the componentwise partial order between vectors. We start
with the following characterization of monotone contractions with respect to weighted maximum
norms:
Lemma V.1. Let B : RS → RS , ψ : S → R++ , β ∈ (0, 1). The following are equivalent:
(i) B is a monotone contraction map with contraction factor β with respect to k · kψ,∞ .
(ii) For any J, J 0 ∈ RS , t ≥ 0, J ≤ J 0 + tψ implies that BJ ≤ BJ 0 + βtψ.
The proof, which essentially copies that of Lemma 3.1 of [Kal17], is given for completeness:
Proof. Introduce · to denote elementwise products: Thus, (ψ · J)(s) = ψ(s)J(s). We also let
ψ −1 (s) = 1/ψ(s) and we will use the shorthand k · k = k · k∞,ψ .
Let us first prove (i) ⇒ (ii). Thus, assume that B is a monotone contraction map with factor
β. Take any J, J 0 , t > 0, J ≤ J 0 + tψ. We have BJ = B(J + tψ) − BJ 0 + BJ 0 ≤ (ψ −1 · (B(J +
tψ) − BJ 0 )) · ψ + BJ 0 ≤ kB(J + tψ) − BJ 0 kψ + BJ 0 ≤ βtkJ − J 0 kψ + BJ 0 .
For the reverse direction, note that monotonicity follows by taking t = 0. Now, let ε = kJ−J 0 k.
Then, J ≤ J 0 + εψ and J 0 ≤ J + εψ. By monotonicity and the assumed property of B (using
t = ε ≥ 0), −βεψ ≤ BJ − BJ 0 ≤ βεψ, which implies that kBJ − BJ 0 k ≤ β.
Corollary V.2. If B is monotone and there exists some β ∈ [0, 1) such that for any J ∈ RS and
any t > 0,
B(J + tψ) ≤ BJ + βtψ
April 11, 2017
(3)
DRAFT
16
then B is a k · k∞,ψ contraction with factor β.
Proof. Let J, J 0 ∈ RS , t ≥ 0 and assume that J ≤ J 0 + tψ. By monotonicity BJ ≤ B(J 0 + tψ),
while by (3), B(J 0 + tψ) ≤ BJ 0 + βtψ. Hence, BJ ≤ BJ 0 + βtψ. This shows that (ii) of
Lemma V.1 holds. Hence, by this lemma, B is a contraction with factor β with respect to
k · k∞,ψ .
Let us now return to the proof of our main result. Recall that the goal is to bound kJ ∗ −
JLRA k1,c through relating this deviations from the fixed point of Γ̂, which was promised to be a
contraction. Let us thus now prove this. For this, it suffices to show that Γ̂ satisfies the conditions
of Corollary V.2. In fact, we will see this holds with β = βψ .
Proposition V.3. The operator Γ̂ satisfies the conditions of Corollary V.2 with β = βψ , and is
thus a k · k∞,ψ -contraction with coefficient βψ .
Proof. First, note that (as it is well known) H is monotone (all the Pa matrices in the definition
of H are nonnegative valued) and that it satisfies an inequality similar to (3): For any t ≥ 0,
J ∈ RS ,
H(J + tψ) ≤ HJ + βψ t Eψ .
(4)
This follows again because our assumption on ψ implies that for any a ∈ A, αPa ψ ≤ βψ ψ.
.
Let us now prove that Γ̂ is monotone. Given J ∈ RS , let F 0 (J) = { Φr : W > EΦr ≥
W > HJ, r ∈ Rk } . Choose any s ∈ S. Since J1 ≤ J2 , W is nonnegative valued and H is
monotone, we have W > HJ1 ≤ W > HJ2 . Hence, FJ2 ⊂ FJ1 and thus (Γ̂J1 )(s) ≤ (Γ̂J2 )(s).
Since s was arbitrary, monotonicity of Γ̂ follows.
Let us now turn to proving that (3) holds with β = βψ . By definition, for s ∈ S, t ≥ 0,
J ∈ RS , (Γ̂(J + tψ))(s) = min{r> φ(s) : W > EΦr ≥ W > H(J + tψ), r ∈ Rk }. By (4),
H(J + tψ) ≤ HJ + tβψ Eψ and hence W > H(J + tψ) ≤ W > (HJ + tβψ Eψ). Thus, (Γ̂(J +
tψ))(s) ≤ min{r> φ(s) : W > EΦr ≥ W > (HJ + tβψ Eψ), r ∈ Rk }.
To finish, we need the following elementary observation:
Claim V.4. Let A ∈ Ru×v , b ∈ Ru , d ∈ Rv and b0 = Ax0 for some x0 ∈ Rv . Then
min{d> x : Ax ≥ b + b0 , x ∈ Rv }
= min{d> y : Ay ≥ b, y ∈ Rv } + d> x0 .
April 11, 2017
DRAFT
17
Proof of Claim V.4. Set y = x − x0 .
Now, using Claim V.4 with A = W > EΦ, b = W > HJ, d = φ(s), b0 = tβψ W > Eψ and
x0 = tβψ r0 , thanks to Φr0 = ψ we have Ax0 = b0 . Hence the desired statement follows from
the claim.
Let us now return to bounding kJ ∗ − JLRA k1,c . For x ∈ R, let (x)− be the negative part
of x: (x)− = max(−x, 0). Then, |x| = x + 2(x)− . For a vector J ∈ RS , we will write (J)−
to denote the vector obtained by applying the negative part componentwise. We consider the
decomposition
kJLRA − J ∗ k1,c = c>(JLRA − J ∗ ) + 2c>(JLRA − J ∗ )− .
(5)
Let VLRA be the fixed point Γ̂. We know claim the following:
Claim V.5. We have JLRA ≥ VLRA , c> JALP ≥ c> JLRA .
Proof. The inequality c> JALP ≥ c> JLRA follows immediately from the definitions of JALP and
JLRA .
To prove the first part let s ∈ S, c = es and let rs be a solution to LRALP in (1). For s ∈ S,
let V0 (s) = mins0 ∈S rs>0 φ(s).
.
It suffices to show that V1 = Γ̂V0 ≤ V0 ≤ JLRA . Indeed, if this holds then Vn+1 = Γ̂Vn , n ≥ 1,
satisfies Vn+1 ≤ Vn and Vn → VLRA as n → ∞ since Γ̂ is a monotone contraction mapping.
Since rs>0 φ(s) ≥ rs> φ(s) also holds for any s, s0 ∈ S, we have V0 (s) = rs> φ(s). Also, since
JLRA (s) ≥ rs> φ(s), it follows that JLRA ≥ V0 . Now, fix some s ∈ S and define re0 s ,V0 be the
solution to the linear program defining (Γ̂V0 )(s). We need to show that V1 (s) = (Γ̂V0 )(s) =
(re0 s ,V0 )> φ(s) ≤ V0 (s). By the definition of re0 s ,V0 we know that (re0 s ,V0 )> φ(s) ≤ r> φ(s) holds
for any r ∈ Rk such that W > EΦr ≥ W > HV0 . Thus, it suffices to show that rs satisfies
W > EΦrs ≥ W > HV0 . By definition, rs satisfies W > EΦrs ≥ W > HΦrs . Hence, by the monotone
property of H and since W is nonnegative valued, it is sufficient if Φrs ≥ V0 . This however
follows from the definition of V0 .
Thanks to the previous claim, (JLRA − J ∗ )− ≤ (VLRA − J ∗ )− and c> JLRA ≤ c> JALP . Hence,
from (5) we get
kJLRA − J ∗ k1,c ≤ c>(JALP − J ∗ ) + 2c>(VLRA − J ∗ )− .
April 11, 2017
DRAFT
18
>
c ψ
ε, where recall that ε = inf r kJ ∗ −
By Theorem 3 of [FV03], the first term is bounded by 2 1−β
ψ
Φrk∞,ψ . Hence, it remains to bound the second term.
For this, note that for any J ∈ RS , (J)− ≤ |J| and also that kJk1,c ≤ c> ψkJk∞,ψ . Hence,
we switch to bounding kJ ∗ − VLRA k∞,ψ . A standard contraction argument gives
kJ ∗ − VLRA k∞,ψ = kJ ∗ − Γ̂J + Γ̂J ∗ − Γ̂VLRA k∞,ψ
≤ kJ ∗ − Γ̂J ∗ k∞,ψ + kΓ̂J ∗ − VLRA k∞,ψ
≤ kJ ∗ − Γ̂J ∗ k∞,ψ + βψ kΓ̂J ∗ − VLRA k∞,ψ .
Reordering and using another triangle inequality we get
kJ ∗ − VLRA k∞,ψ ≤
∗
∗
kJ ∗ − JALP
k∞,ψ + kJALP
− Γ̂J ∗ k∞,ψ
.
1 − βψ
∗
We bound the term kJ ∗ − JALP
k∞,ψ in the following lemma:
∗
Lemma V.6. We have kJ ∗ − JALP
k∞,ψ ≤ 2ε, where recall that ε = inf r∈Rk kJ ∗ − Φrk∞,ψ .
.
∗
∗
Proof. Let r∗ = argminr∈Rk kJ ∗ − Φrk∞,ψ . First, notice that JALP
≥ J ∗ . Hence, 0 ≤ JALP
− J ∗.
Now let r0 = r∗ + εr0 . Then, Φr0 = Φr∗ + εψ ≥ J ∗ , where the equality follows by the definition
of r0 and the inequality follows by the definition of ε. Hence, r0 is in the feasible set of the
∗
∗
∗
LP defining JALP
and thus JALP
≤ Φr0 . Thus, 0 ≤ JALP
− J ∗ ≤ Φr∗ − J ∗ + εψ. Dividing
componentwise by ψ, taking absolute value and then taking maximum of both sides gives the
result.
∗
The proof of the main result is finished by noting that Γ̂J ∗ = JLRA
and the chaining the
inequalities we derived.
VI. N UMERICAL I LLUSTRATION
In this section, we show via an example in the domain of controlled queues the consequences
of Theorem IV.2, which bounded the error when the constraints are chosen based on selecting
a set of representative states (further preliminary experimental results have been reported in
[LB15]).
Model: We ran the experiments in the context of a queuing model similar to the one in
Section 5.2 of [FV03]. We consider a (simple) small scale model so that we can compare
with the optimal policy. At the same time, we will use a small number of basis functions and
constraints, to “stress-test” the algorithm. The queuing system has a single queue with random
April 11, 2017
DRAFT
19
arrivals and departures. The state of the system is the queue length with the state space given by
S = {0, . . . , S − 1}, where S − 1 is the buffer size of the queue. The action set A = {1, . . . , A}
is related to the service rates. We let st denote the state at time t. The state at time t + 1
when action at ∈ A is chosen is given by st+1 = st + 1 with probability p, st+1 = st − 1 with
probability q(at ) and st+1 = st , with probability (1−p−q(at )). For states st = 0 and st = S −1,
the system dynamics is given by st+1 = st + 1 with probability p when st = 0 and st+1 = st − 1
with probability q(at ) when st = S − 1. The service rates satisfy 0 < q(1) ≤ . . . ≤ q(A) < 1
with q(A) > p so as to ensure ‘stabilizability’ of the queue. The reward associated with action
a ∈ A and state s ∈ S is given by ga (s) = −(s/N + q(a)3 ) (the idea here is to penalize higher
queue lengths and higher service rates).
Parameter Settings: We ran our experiments for S = 1000, A = 4 with q(1) = 0.2, q(2) =
0.4, q(3) = 0.6, q(4) = 0.8, p = 0.4 and α = 1 − S1 . The moderate size of S = 1000 enabled us
to compute the exact value of J ∗ (the most expensive part of the computation). We made use of
polynomial features in Φ (i.e., 1, s, . . . , sk−1 ) since they are known to work reasonably well for
this domain [FV03]. Note hat the conic span conditions will only be met with some lag, unless
all the constraints are selected. Hence, these features allow us to test the limits of the theory.
We chose k = 4, a low number, to counteract that the MDP is small scale.
Experimental Methodology: We compare two different sampling strategies (i) based on the
cone conditions, and (ii) based on constraint sampling. The two strategies are compared via
lookahead policies, wherein, the action at state s is obtained by computing the approximate
value functions of the next states and selecting the action that leads to the larger estimated
value. The details are as follows: Case (i): Except for the corner states i.e., s = 0 and s = 999,
each state 0 < s < S − 1 has two next states namely s0 = s − 1 and s0 = s + 1. We formulate two
separate LRALPs (or just one LRALP for s = 0 and s = S −1) for next states. When formulating
the LRALP for state s0 , we let c = es0 and choose the constraint corresponding to state s0 to
ensure the cone condition to be met for LRALP. We choose 5 more constraints corresponding to
states 1, 200, 400, 600, 800, 999 (uniformly spaced across the state space) and compute Jˆe 0 . The
s
number of constraints is kept very small as a way of emulating that in large-scale problems we
cannot expect a dense covering of the state-space when selecting the constraints. The lookahead
P
policy is formulated as uLRA (s) = argmina∈A ga (s) + s0 ∈S pa (s, s0 )Jˆes0 (s0 ). Case (ii): In a
manner similar to Case (i), we formulate two separate LRALPs for next states. However, as
opposed to the previous case, when formulating the LRALP for state s0 , we sample m = 6
April 11, 2017
DRAFT
20
states (defining the constraints) from a distribution dependent on s0 . We experimented with two
sampling distributions that lead to two the lookahead policies that we denote by uCS−ideal and
uCS , respectively. The sampling distribution that defines uCS−ideal is the sampling distribution
that minimizes the upper bound proved by de Farias and Van Roy [FV04]. In particular, the
−1
sampling distribution used at state s0 is cs0 = e>
s0 (1−α)(I −αPu∗ ) , with es denoting the standard
basis vector which is 1 in the sth co-ordinate and 0 in all the other co-ordinates. This sampling
distribution is used as a baseline; it is unrealistic to assume that one would be able to sample
from this distribution without access to the optimal policy u∗ , which is the quantity of ultimate
0
interest. As a more realistic approach, we also consider sampling from cs0 (s) = κ(1−α)(α)|s −s| ,
where κ > 0 is a normalization factor that ensures that cs0 is a distribution. Again, we sample
m = 6 states. This leads to the policy uCS .
0
4
−200
3
−400
−600
2
−800
1
0
200
400
600
J∗
JCS
JCS−ideal
JLRA
800
0
200
400
600
800 1,000
u∗
uCS
uCS−ideal
uLRA
Fig. 1. Results for a single-queue with polynomial features. On both figures the x axis represents the state space: the length
of the queue. The left-hand-side figure shows the value functions of the various policies computed, alongside with the optimal
value function (higher values are better), while the right-hand side subfigure shows the underlying policies. “CS” and “CS-ideal”
stand for constraint sampling, while LRA stands for choosing the constraints based on geometric principles proposed in the
paper. For further details, see the text.
The results are shown in Fig. 1. The right-hand-side figure shows the policies computed, while
the left-hand-side figure shows their value functions. Since constraint sampling (CS) produces
randomized results, we repeated the simulations 10 times. The results in all cases were quite
April 11, 2017
DRAFT
REFERENCES
21
close, hence we show the plot for a typical run. The plots show that the CS case with the ideal
sampler is slightly worse, which can be attributed to the fact that the in the case of ideal sampler,
the sampling distribution is concentrated near the start state s0 in comparison to the behaviour
0
of the distribution cs0 (s) = (1 − α)(α)|s −s| which distributes the mass more evenly. As can
be seen from the figure, choosing the constraints to (approximately) satisfy the constraint of
the theoretical results reliably produces better results: In fact, the value functions J ∗ and JLRA
are mostly on the top of each other. We expect that in larger domains, differences between
constraints chosen based on the principles discovered in this paper and choosing constraints in
more heuristic ways will lead to similar, or even larger differences. However, the study of this
is left for future work.
VII. C ONCLUSION
In this paper, we introduced and analyzed the linearly relaxed approximate linear program
(LRALP) whose constraints were obtained as positive linear combination of the original constraints of the ALP. The main novel contribution is a theoretical result which gives a geometrically
interpretable bound on the performance loss due to relaxing the constraint sets. Possibilities for
future work include extending the results to other forms of approximate linear programming
in MDPs (e.g., [DFM09]), exploring the idea of approximating dual variables and designing
algorithms that use the newly derived results to actively compute what constraints to select.
R EFERENCES
[AA+15]
M. Abu Alsheikh, D. T. Hoang, D. Niyato, H.-P. Tan, and S. Lin, “Markov decision
processes with applications in wireless sensor networks: A survey,” IEEE Comm.
Surveys & Tutorials, vol. 17, pp. 1239–1267, 2015.
[AYBM14]
Y. Abbasi-Yadkori, P. Bartlett, and A. Malek, “Linear programming for large-scale
Markov decision problems,” in ICML, 2014, pp. 496–504.
[BD17]
R. J. Boucherie and N. M. van Dijk, Eds., Markov decision processes in practice.
Springer, 2017, vol. 248.
[BFM12]
N. Bhat, V. Farias, and C. C. Moallemi, “Non-parametric approximate dynamic
programming via the kernel method,” in NIPS, 2012, pp. 386–394.
[BR11]
N. Bäuerle and U. Rieder, Markov decision processes with applications to finance.
Springer, 2011.
April 11, 2017
DRAFT
REFERENCES
[BT00]
22
V. D. Blondel and J. N. Tsitsiklis, “A survey of computational complexity results
in systems and control,” Automatica, vol. 36, pp. 1249–1274, 2000.
[CM99]
R.-R. Chen and S. P. Meyn, “Value iteration and optimization of multiclass queueing networks,” Queueing Systems, vol. 32, no. 1-3, pp. 65–97, 1999.
[DFM09]
V. V. Desai, V. F. Farias, and C. C. Moallemi, “A smoothed approximate linear
program,” in NIPS, 2009, pp. 459–467.
[FS02]
E. A. Feinberg and A. Shwartz, Handbook of Markov decision processes: Methods
and applications. Kluwer Academic Publishers, 2002.
[FV03]
D. P. de Farias and B. Van Roy, “The linear programming approach to approximate
dynamic programming,” Operations Research, vol. 51, pp. 850–865, 2003.
[FV04]
——, “On constraint sampling in the linear programming approach to approximate
dynamic programming,” Mathematics of Operations Research, vol. 29, pp. 462–
478, 2004.
[Gue+03]
C. Guestrin, D. Koller, R. Parr, and S. Venkataraman, “Efficient solution algorithms
for factored MDPs,” J. of Artificial Intelligence Research, vol. 19, pp. 399–468,
2003.
[HY07]
Q. Hu and W. Yue, Markov decision processes with their applications. Springer,
2007.
[Kal17]
L. Kallenberg. (2017). Markov decision processes: Lecture notes, [Online]. Available: https://goo.gl/yhvrph (visited on 03/16/2017).
[KH04]
B. Kveton and M. Hauskrecht, “Heuristic refinements of approximate linear programming for factored continuous-state Markov decision processes,” in ICAPS,
2004, pp. 306–314.
[KM12]
A. Kolobov and Mausam, “Planning with Markov decision processes: An AI
perspective,” Synthesis Lectures on Artificial Intelligence and Machine Learning,
2012.
[KMN02]
M. Kearns, Y. Mansour, and A. Y. Ng, “A sparse sampling algorithm for nearoptimal planning in large Markov decision processes,” Machine learning, vol. 49,
pp. 193–208, 2002.
[LB15]
C. Lakshminarayanan and S. Bhatnagar, “A generalized reduced linear program
for Markov Decision Processes,” in AAAI, 2015, pp. 2722–2728.
April 11, 2017
DRAFT
23
[LL12]
F. L. Lewis and D. Liu, Eds., Reinforcement learning and approximate dynamic
programming for feedback control. Wiley-IEEE Press, 2012.
[PP11]
J. Pazis and R. Parr, “Non-parametric approximate linear programming for MDPs,”
in AAAI, 2011.
[Put94]
M. L. Puterman, Markov decision processes: Discrete stochastic programming.
New York: John Wiley, 1994.
[PZ09]
M. Petrik and S. Zilberstein, “Constraint relaxation in approximate linear programs,” in ICML, 2009, pp. 809–816.
[Rus96a]
J. Rust, “Numerical dynamic programming in economics,” in Handbook of Computational Economics, vol. 1, Elsevier, North Holland, 1996, pp. 619–729.
[Rus96b]
——, “Using randomization to break the curse of dimensionality,” Econometrica,
vol. 65, pp. 487–516, 1996.
[SB10]
O. Sigaud and O. Buffet, Eds., Markov decision processes in artificial intelligence.
Wiley-ISTE, 2010.
[SP01]
D. Schuurmans and R. Patrascu, “Direct value-approximation for factored MDPs,”
in NIPS, 2001, pp. 1579–1586.
[SS85]
P. J. Schweitzer and A. Seidmann, “Generalized polynomial approximations in
Markovian decision processes,” Journal of Mathematical Analysis and Applications, vol. 110, pp. 568–582, 1985.
[Sze01]
Cs. Szepesvári, “Efficient approximate planning in continuous space Markovian
decision problems,” AI Communications, vol. 13, no. 3, pp. 163–176, 2001.
[Tay+10]
G. Taylor, M. Petrik, R. Parr, and S. Zilberstein, “Feature selection using regularization in approximate linear programs for Markov decision processes,” in ICML,
2010, pp. 871–878.
[Whi93]
D. J. White, “A survey of applications of markov decision processes,” The Journal
of the Operational Research Society, vol. 44, no. 11, pp. 1073–1096, 1993.
April 11, 2017
DRAFT
| 3cs.SY
|
A Systematic Aspect-Oriented Refactoring and Testing
Strategy, and its Application to JHotDraw
Arie van Deursen1,2 , Marius Marin2 , and Leon Moonen2,1
arXiv:cs/0503015v1 [cs.SE] 5 Mar 2005
1
2
Centrum voor Wiskunde en Informatica (CWI), The Netherlands
Software Evolution Research Lab, Delft Univ. of Technology, The Netherlands
[email protected], [email protected],
and [email protected]
Abstract. Aspect oriented programming aims at achieving better modularization for a system’s crosscutting concerns in order to improve its key quality
attributes, such as evolvability and reusability. Consequently, the adoption of
aspect-oriented techniques in existing (legacy) software systems is of interest to
remediate software aging. The refactoring of existing systems to employ aspectorientation will be considerably eased by a systematic approach that will ensure
a safe and consistent migration.
In this paper, we propose a refactoring and testing strategy that supports such
an approach and consider issues of behavior conservation and (incremental) integration of the aspect-oriented solution with the original system. The strategy is
applied to the JH OT D RAW open source project and illustrated on a group of selected concerns. Finally, we abstract from the case study and present a number of
generic refactorings which contribute to an incremental aspect-oriented refactoring process and associate particular types of crosscutting concerns to the model
and features of the employed aspect language. The contributions of this paper are
both in the area of supporting migration towards aspect-oriented solutions and
supporting the development of aspect languages that are better suited for such
migrations.
1 Introduction
Aspect-oriented software development is a programming paradigm that addresses crosscutting concerns: behavior of a software system that is hard to decompose and isolate in
existing paradigms (as object orientation) and requires its implementation to be spread
across many different modules. Aspect-oriented software development aims to overcome these limitations by capturing such crosscutting behavior in a new modularization
unit, the aspect, and offers (compile time) code generation facilities to weave aspect
code into the rest of the system. Claimed benefits include improved evolvability and
reusability of (parts of) the software system [12, 23].
Addressing the aforementioned modularization limitations and the resulting code
scattering and tangling does not only pay off in the development of new applications but
it will also have major benefits in existing software systems where these, and associated,
problems have become known as software aging [33] or software entropy [24–26].
The adoption of aspect-orientation in existing software requires refactoring: code
transformations that improve the internal structure of a system while preserving its external behavior. Existing work on aspect-introducing-refactorings has mainly focused
on presenting aspect-oriented solutions to typical crosscutting problems, especially in
the context of design patterns, and showing that this results in a better separation of
concerns [19, 23, 31, 32]. Also tool support for aspect extensions of refactorings, such
as method extraction, has been investigated [13].
We argue, however, that widespread adoption of aspect-oriented techniques in existing software systems is still hindered by a number of open issues:
– Lack of a systematic approach to refactor legacy code to employ aspect-oriented
solutions;
– Proper understanding of the testing challenges that rise from behavior preserving
migration towards a new or extended language, such as the development of an
aspect-oriented fault model and the definition of an explicit test adequacy criterion;
– Suitability of aspect languages. Analysis of crosscutting concerns that were identified in various object-oriented systems in earlier work ([29]) suggested that we
might encounter difficulties when trying to refactor those concerns into aspects: the
mechanisms offered by particular aspect languages, or the joint point model behind
the language were not always sufficient to capture all types of concerns that were
encountered;
– Availability of aspect-oriented and non-aspect-oriented implementations of the same
software system which can be used to show the evolution benefits of proposed solutions; and, more generally,
– An overall assessment of the benefits of aspect-oriented software development.
These issues are the principal motivation for the work described in this paper. To address
them, we propose a language and system independent refactoring and testing strategy
to adopt aspect-oriented solutions in legacy code. The strategy consists of a number
of systematic steps that guide the transformation, ensure conservation of observable
behavior and help one deal with the intricacies of aspect-oriented migrations.
We demonstrate suitability of the proposed strategy in a case study in which we migrate JH OT D RAW, a (relatively) large and well-designed open source Java application,
to AJ HOT D RAW, a corresponding aspect-oriented version which is based on A SPECT J,
an aspect language extending Java with crosscutting functionality.
Based on the difficulties that were encountered during our refactorings, we reflect
on the suitability of particular (features of) aspect languages for this type of work (i.e.
evolution of legacy systems as opposed to greenfield development). This provides designers of aspect languages with valuable insights into situations that require model or
feature extensions to address these concerns.
The remainder of the paper is structured as follows: In the next section we propose
an aspect-oriented refactoring and testing strategy together with its accompanying fault
model and test adequacy criterion. This is followed by a section that presents general
considerations about the case study. Next, a number of selected crosscutting concerns
are discussed on an individual basis, depicting the context in which they occur, the
original and refactored implementation together with the benefits and drawbacks. In
2
Section 5, a number of generic refactorings are abstracted from the case study and associations are made between types of crosscutting functionality and the aspect language
model and features. We conclude with a general discussion followed by an overview
of related work, summary of our contributions and present some directions for future
work.
2 The B ETTAR Refactoring Strategy
The refactoring strategy we propose is called B ETTAR: its objective is to obtain Better
Evolvability Through Tested Aspect Refactorings. We distinguish the following steps:
Identification of Crosscutting Concerns: Search for candidate aspects using “aspect
mining” techniques such as fan-in analysis or clone detection [6, 29, 36]. Assess the
scattering and tangling implications of the current non-aspect-oriented solution to
the crosscutting concerns identified.
Aspect Design: Identify how the concern could be implemented as an aspect. Assess
the pros and cons and compare these with the existing solution.
Refactoring Design: Devise a sequence of (small) steps, refactoring the object solution to the aspect solution. This may involve various traditional object-oriented
refactorings, in order to unplug the crosscutting concern from code implementing
other concerns, in addition to refactorings moving functionality to aspects.
Conduct trade-off analysis to determine whether the aspect benefits outweigh the
refactoring costs.
Test Suite Design: Conduct a baseline test on the existing implementation, and analyze the test adequacy of the current test suite with respect to the risks introduced
by the aspect-oriented solution as well as by the refactoring process itself. If necessary, create or extend the test suite – see the next section for further details on this
step.
Execute and Test: Carry out the refactorings, and verify that the behavior of the system is unaltered by means of the test suite.
In Section 4 we will apply these steps to various cross cutting concerns as occurring
in the open source JH OT D RAW system. As the interplay between refactoring and testing, in particular in an aspect-oriented setting, has received very little attention in the
literature so far, we start by elaborating the testing steps.
2.1 Ensuring Behavior Preservation while Refactoring to Aspects
Refactoring is the process of changing a software system in such a way that it improves
the code’s internal structure, without altering its external behavior [14]. In order to
ensure the latter constraint, most literature on refactoring assumes the presence of a test
suite that verifies the correct functional behavior of the system to be refactored. As long
as this test suite is executed before and after each refactoring, we can assume that we
will be warned as soon as one of our refactorings affects the correct behavior of the
system.
3
In practice, however, the creation of such a test suite is challenged by a number of
issues. These hold in the general (pure object-oriented) situation, as well as in a setting
where the refactoring includes the introduction of aspects. These issues are:
– To test effectively, testing should be based on a fault model. Such a fault model
guides our search for test cases that give the highest probability of finding typical
faults [4]. The adoption of aspects opens opportunities for different types of faults,
calling for an explicit aspect-oriented fault model.
– Systematic testing makes use of an explicit test adequacy criterion (see, e.g., [4]),
usually expressed as a coverage percentage to be achieved in some coverage model.
Refactoring changes the internal structure of the code. Since test adequacy is expressed in terms of code structures covered, a refactoring may very well affect
coverage negatively — a phenomenon referred to as the antiextensionality axiom
by Weyuker [38].
– In addition to new faults introduced by using aspects, the mechanics of actually
carrying out a refactoring may lead to a new fault. For example, when we moved
part of a method from a class to an aspect, we did not copy-paste all statements.
At other times, refactorings affect the public interface of classes, for example when
moving a public method. This implies that the test suite needs to be adapted as well,
causing an extra risk of letting errors pass [11].
This calls for a testing approach that is dedicated to refactorings involving the introduction of aspects. In this section, we provide such an approach. In order to do so,
we first propose an aspect-oriented fault model, as well as aspect-oriented test adequacy
criteria. Note that such a model and criteria can never be complete: we believe, however,
that our proposals represent an important first step.
The specific faults that can be made while refactoring depend on the actual refactoring applied. Therefore, we will not provide a general fault model for refactorings, but
will indicate typical faults and testing implications when discussing some of the individual refactorings that we have used. We will then also indicate whether the refactoring
may require changes to existing test cases.
The test strategy presented below is applicable to any development project making
use of aspects, and as such is independent of our case study on JH OT D RAW. In the
later sections we will discuss how we actually applied the proposed fault model and
adequacy criteria when while refactoring JH OT D RAW to AJ HOT D RAW.
2.2 An Aspect-Oriented Fault Model
A fault model identifies relationships and components of the system under test that are
most likely to have faults [4]. We distinguish faults for inter-type declarations, pointcuts, and advice.
Inter-type declarations are most error-prone (and most powerful) when used to create polymorphic functions. Therefore, our fault model for introductions is based on
Binder’s existing fault models for polymorphism and inheritance [4, p.501]. Our model
distinguishes the following faults that are specifically related to polymorphism in introductions and inter-type declarations:
4
– Wrong method name in introduction, leading to a missing or unanticipated method
override.
– Wrong class name in a member-introduction, leading to a method body in the
wrong place in the class hierarchy.
– Inconsistent parent declaration, resulting in a (sub)class that violates Liskov’s and
Wing’s behavioral notion of subtyping [27] and/or Meyer’s design-by-contract rules
for inheritance (such as require no more, ensure no less) [30].
– Inconsistent overridden method introduction, also resulting in a violation of behavioral subtyping.
– Omitted parent interface resulting in a method that was intended to implement an
interface method, but which now stands on its own.
Faults in pointcuts will have the effect that advice code is activated at the wrong
program execution points. Such faults include:
– Wrong primitive pointcut, using, for example, a call instead of an execution
construct.
– Errors in the conditional logic combining the individual pointcut conditions.
– Wrong type, method, field, or constructor pattern in pointcut. In particular, the use
of * as a pattern wildcard or in string matching easily leads to too many join points.
Furthermore, if the underlying classes are modified or extended, the wildcard may
become erroneous without the compiler being able to notice this.
Faults in advice will result in the wrong action at a certain point of execution. Such
faults include:
– Wrong advice specification (using before instead of after, using after with
the wrong argument, etc.).
– Wrong or missing proceed in around advice.
– Wrong or missing advice precedence.
– Advice code causing a method to break its class invariant or to fail to meet its
postcondition.
The fault model above states that aspect weaving should not conflict with class
invariants or method pre- and postconditions. The safe route to follow is that class resulting from weaving is a proper subtype of the original class. Put in terms of design by
contract, the class invariant of the resulting class cannot be weaker, its method preconditions cannot be stronger, and the postconditions cannot be weaker. Typical examples
are “harmless” aspects which add logging or tracing. In this situation, existing code
using the class need not be aware that new functionality has been woven into it. In
other words, the test suite for the original classes should pass on the classes extended
by introduction or advice as well, and doing so will help to find faults originating from
improper extensions.
An alternative route is that the aspect actually modifies the contract in a way that
conflicts with the inheritance rules from design-by-contract. This may include changes
in method pre- or postconditions, and may thus require weaving in additional code at
all affected call sites. A typical example is an aspect that adds security checks: this may
lead to additional exceptions which at some point should be handled in the original
5
application. Faults in this approach will not be restricted to the newly woven class,
but may be at any call site in the application. This setting is much harder to test and
immediate reuse of the test suite will not be possible.
Similar distinctions are made by Clifton and Leavens [7], who discuss the relation
between behavioral subtyping and aspect weaving, and distinguish observers from assistants. Rinard et al [34] classifies interactions between woven code and the original
code, recognizing augmentation, narrowing, replacement, and combinations between
them. The key concern of these authors is modular reasoning: in our setting it is modular testing, and reuse of test suites to woven classes.
2.3 Aspect-Oriented Test Adequacy
A test adequacy criterion prescribes the elements of the implementation under test that
need to be exercised by a test suite. The coverage achieved by a test suite is the percentage of elements actually exercised. In this section we formulate adequacy criteria
for aspects targeting the faults presented in the previous section.
Due to the mixed nature of an aspect definition, which can address both static and
dynamic crosscutting using pointcuts, intertype declarations, and advice, it is not so
easy to obtain a single criterion that allows us to make meaningful statements of the
form “we have tested 75% of this aspect”. Instead, we will define different criteria for
the various elements in an aspect definition.
Introducing a new method m in a class C is akin to directly adding the method
to C. Therefore, normal coverage goals such as statement or branch coverage apply.
However, as we have seen in the fault model, the most powerful and dangerous introductions are those where polymorphic methods are added. Therefore, adequacy criteria
explicitly based on exercising all possible polymorphic bindings are in place as well.
Rountev et al [35] include an up to date overview of criteria for polymorphic bindings.
They distinguish the all-receiver-classes criterion which requires exercising all possible
classes of the receiver object at a call site, and the all-target-methods criterion which
requires exercising all possible bindings between a call site and the methods that may
be invoked by that site.
The intertype declaration of a new supertype or interface for a given class changes
the inheritance hierarchy it belongs to. This, again, calls for adequacy criteria taking
polymorphic calls into account. Observe that these adequacy criteria take into account
all call sites within the rest of the application. Thus, polymorphic coverage goals are not
just a percentage of the aspect definition itself, but a percentage of how well affected
call sites are covered.
To deal with adequacy for pointcuts, we will say that a test case T exercises a
pointcut P if T activates advice at a join point captured by P . An adequate test suite for
a pointcut should maximize our chance of finding errors in the pointcut. We distinguish
primitive pointcuts and compound pointcuts built from conditional operators.
Primitive pointcut operators (such as call, cflow, and so on) can capture a multitude of join points. Which of these should we ensure we execute in order to maximize our chance of finding errors in the pointcut? In most cases we cannot answer this
question, so an arbitrary join point will do. For signature or type matching involving
6
wild cards, we can arrive at the equivalent of traditional boundary testing by ensuring we have one case where the asterisk matches the empty string, and one where its
match is non-empty. When matching types in a hierarchy, for example in a call(*
Class+.*) expression, the class named is a boundary. Using the one × one criterion
[4] insisting on one point on the boundary, and one just outside it, we would obtain one
test case for Class, and one for each immediate superclass (interface) above it in the
hierarchy.
Tests for pointcut expressions composed from multiple conditions should exercise
every relevant condition combination. In traditional testing, the most rigorous approach
is to test each true/false combination, leading to 2N test cases for an expression with
N conditions. Alternatively, the Each-Condition/All-Conditions criterion can be used
which leads to N + 1 test cases by insisting on one test case for each condition making
that condition true and all others false, in addition to one making all conditions true (for
and logic, replacing true and false for or logic) [4]. Pointcut logic, however, is different
from normal Boolean logic, for example in that certain operators (such as target) are
primarily meaningful in combinations with others. Moreover, there are typical idioms
for using pointcuts, such as a sequence of a general pointcut (such as all public calls)
conjuncted with several exceptions (each using the negation operator) for classes or
methods that are to be excluded from the pointcut.
Test adequacy for advice itself can again be based on branch or statement coverage.
It is most natural to compare advice with a method that is called at relevant join points.
Thus, to achieve branch coverage for the advice, we do not need to exercise all branches
at every join point: it suffices to find one join point at which we exercise all branches.
Furthermore, it is natural to insist that each join point at which the advice is activated
is exercised. Typically, a test suite achieving statement coverage for the full unwoven
application will get a far way in covering all join points. One may be tempted to think
that covering all captured join points also achieves adequate pointcut coverage. This,
however, is not the case, since the pointcut may be defined as a complex expression,
parts of which are used to prevent firing at a particular join point.
Last but not least, there may be (abstract) reusable aspects whose pointcuts do not
refer to particular (named) classes or methods they should be woven into. To test such
aspects, a stub application needs to be created, to which the aspect can be applied.
When creating these stubs, the test adequacy issues presented above can be used as a
guideline, ensuring for example that it is indeed possible to exercise all conditions in
the pointcut. As far as we can see, reusable aspects themselves provide no further test
adequacy constraints.
2.4 The B ETTAR Testing Strategy
The test strategy combining the fault model and various adequacy criteria consists of
three steps:
Responsibility-Based Testing (Black Box): Create or identify a functional test suite
for the concern at hand. Focus on answering the question whether the implementation of the concern does what it is supposed to do.
7
Risk-Based Testing (Grey Box): Use the fault model to refine the test suite so that
faults due to the refactoring process as well as the (aspect-oriented) target solution
are most likely to be captured.
Source-Based Test Adequacy Validation (White Box): Inspect the coverage of the
test suite developed so far (either by running it on instrumented code or by manual
analysis), and verify that relevant test adequacy criteria are indeed met. If not, return to the previous steps to create additional responsibility- or risk-based test cases
until the adequacy criteria are fullfilled.
If the refactorings do not affect the external interfaces of the classes under test, the
test suite can be applied to both the original and the refactored system. This has the
advantage that one can be certain that each new test also successfully passes on the
old implementation. If refactoring does require making adaptations to the test suite, the
following approaches are possible:
– Refactor the test suite so that it exercises more global functionality instead of invoking the modified methods directly, making it more robust to future implementation
changes but potentially making it harder to achieve the desired coverage;
– Apply an additional refactoring to the application offering for example an additional interface abstracting away from implementation differences between the
original and target solution;
– As a last resort, we could give up on our attempt to apply the test suite to the
original system, and apply new tests to the new system only, thereby losing them
as safeguard against behavior modification during refactoring. In our case study
we were never forced to do this, and could always refactor the test suite to permit
testing of both versions.
3
AJ HOT D RAW: An Open Source Aspect-Oriented Showcase
To experiment with the feasibility of adopting aspect-oriented solutions in existing software and demonstrate the strategy proposed earlier, we have created AJ HOT D RAW: an
aspect-oriented refactoring of JH OT D RAW,3 a relatively large and well-designed open
source Java application. In order to allow other researchers to benefit from our work and
to enable comparative software evolution research on a real-life aspect-oriented system,
we decided to release AJ HOT D RAW as an open source project.4
The next sections give a description of the case study and motivate the choice for
both the application and the refactoring language.
3.1 The JH OT D RAW Drawing Framework
JH OT D RAW is a (GUI-based) framework for drawing technical and structured 2D graphics. The application was originally developed as an exercise to show a good use of
3
4
jhotdraw.org, version 5.4b1
ajhotdraw.sourceforge.net Note for the reviewer: we are currently in the process of
cleaning up our refactored code, upgrading it to JHotDraw 6.0, and moving it to the sourceforge server (see the release plan on sourceforge). Our internal version is available upon
request.
8
object oriented design patterns in a Java implementation. The fact that JH OT D RAW is
considered a well-designed application makes it an ideal candidate for aspect-oriented
migration as it is unlikely that evolvability improvements can be made otherwise. The
version of JH OT D RAW analyzed consists of approximately 40,000 lines of code, 300
classes, and 2800 methods.
The JH OT D RAW editor comprises drawing tools, a set of user defined (geometrical,
image, text, etc.) figures, drawing views, and a collection of (tool and menu-associated)
commands. A number of additionally supported features include (re-)storing drawings
(from/)to storage devices, undo/redo activities for commands, and animation functions.
3.2 Evolving a JH OT D RAW Test Suite
The version of JH OT D RAW under study was shipped without a test suite. At the time
of writing, the most recent version (v6.0beta1) has a number of empty test classes,
automatically generated using a Java doclet. The intent is to fill them with test cases,
but so far these have not been made available.5
To safeguard our refactorings, we have developed our own functional test cases on
a by need basis. Since we were, initially, not familiar with the JH OT D RAW code, the
development of these test cases served as our program comprehension strategy ([10]):
We formulated hypotheses on JH OT D RAW’s implementation, expressed them as test
cases, and then attempted to refute them by running these test cases. The test suite that
was developed passes on the original, pure object oriented version of JH OT D RAW, as
well as on AJ HOT D RAW, the refactored, aspect-oriented version.
Our test suite is based on the JUnit framework [3]. Moreover, where needed, we
make use of Java 1.4 assertions to ensure that the alterations did not break invariants
or pre- and postconditions. Since we did wanted to minimize the number of changes to
the JH OT D RAW code, we injected these assertions by means of aspects. For ensuring
invariants, the aspect contains an inter-type declaration giving relevant classes a boolean
invariant method, as well as a pointcut ensuring this method is indeed checked
before and after each public method. Observe that the assertion aspects are independent
of our test suite, and can be woven into the production version of JH OT D RAW as well
in order to simplify debugging.
3.3
AJ HOT D RAW Organization
AJ HOT D RAW is organized into two parts: (1) the main project is the A SPECT J implementation of the system, where the identified crosscutting concerns are refactored to
aspects; (2) the test subproject (JHDTest) comprises all the test cases aimed at ensuring
equivalence between the the original Java solution and the refactored A SPECT J one.
The aspects are put in separate packages, one per concern. Changes to the original files
are restricted just to removing concerns that have been migrated to aspects.
5
We have agreed with the JH OT D RAW maintainers that our test suite will be integrated in their
project.
9
The tests suite can be compiled with and executed on the archived binary files (jar)
of any of the two solutions. Building and executing the test suite is automated using
A NT.6
4 Refactoring of Selected Concerns from JH OT D RAW
Previously, we have employed fan-in analysis for the identification of crosscutting concerns in JH OT D RAW [29]. This resulted in 10 types of concerns that were candidates
for refactoring into an aspect. In this section we discuss three of these concerns (persistence, contract enforcement, and undo) in considerable detail, covering the B ETTAR
steps aspect design, refactoring design, and test suite design.
A transparent, gradual process of refactoring is important for building confidence
in the aspect-oriented solution. Therefore, our refactorings aim at maintaining the conceptual integrity and stay close to the original design. An additional advantage of this
approach is that this preserves the understandability of the refactored system for the
original maintainers.
4.1 Refactoring The Persistence Concern
Aspect Design Drawings in JH OT D RAW are collections of figures that can optionally
be stored and recovered (write/read operations) by the application. The concern denoting this functionality, persistence, is defined by the Storable interface that declares
two methods, write(StorableOutput) and read(StorableInput). The entire hierarchy of
storable elements in a drawing comprises 94 interfaces and classes, of which 40 belong
to the Figure class hierarchy.
Because the persistence concern is already distinguished in the original design,
refactoring it to an aspect is fairly straightforward. The aspect can use introductions in
order to have the persistent elements of a drawing (e.g., figures) implement the Storable
interface. If not all variables comprising the state of the class are accessible through
public getters and setters, the aspect will need access to private members as well. The
A SPECT J way to achieve this is by declaring the aspect privileged.
The implementation of the Storable interface also implies an interesting enforcement constraint: “Objects that implement this interface and that are resurrected by
StorableInput have to provide a default constructor with no arguments.” This constraint
cannot be enforced by A SPECT J. A similar situation occurs when refactoring bean objects (see the “Bean Aspect” example in [2]) that must define no-argument constructors.
A concern related to persistence is serialization, which in JH OT D RAW is also implemented for the Figure hierarchy. According to Java API specification, classes requiring
special handling during de-serialization, such as a number of figures in JH OT D RAW,
must implement a special private method (readObject(ObjectInputStream)). A SPECT J
does not support introduction of private members into target classes. The visibility of
the inter-type declarations relates to the aspect and not to the target class. Although
6
ant.apache.org
10
already acknowledged as a shortcoming (see [2]) the language interpretation of visibility prevents a consistent refactoring of similar kinds (persistence and serialization) of
crosscutting concerns.
A summary of the various issues is provided in Figure 1.
Refactoring Design The refactoring itself is fairly straightforward, and just consists of
moving read and write method implementations to the persistence aspect. The complete
refactoring of the persistence concern can generally be described as Extract Interface
Implementation as discussed by [22].
Old situation
Aspect solution
Objects requiring persistence implement the Storable interface.
Implementation of Storable interface moved to aspect by means of introductions.
Code size
Remains the same
Benefits
All persistence related code in one aspect; classes oblivious of whether they
can be made persistent.
Risks
Encapsulation broken since persistence aspect requires privileged access.
A SPECT J issues Zero-argument constructor cannot be enforced; Private methods cannot be
introduced.
Fig. 1. Refactoring of the Persistence Concern
Test Suite Design Testing the persistence aspect is relatively simple. We nevertheless
discuss it in some detail, since the way of testing can be reused for other more complicated concerns that we will discuss next.
When refactoring persistence to an aspect we run a number of risks: The first is that
in our aspect, we accidentally introduce a read or write method body for a given figure
in the wrong class. The second is that we make an error when copy-pasting the body of
a method to an aspect. Last but not least, our removal of the persistence code from, e.g.,
figures may be incomplete.
In order to test persistence we proceed as follows: First, we create a top level
StorableTest class, which has a test method that (1) creates a Storable (typically a figure), (2) writes it to a stream, (3) reads it back into a different object, and (4) checks the
equivalence between the two. Next, the creation of the actual figure is deferred to subclasses of the StorableTest class using a virtual factory method. Thus, the test hierarchy
mimics the hierarchy of classes to be stored. Finally, our equivalence checking method
should be based on structure, not on object identity. Such a method is not included in the
JH OT D RAW implementation. We injected this method into the class hierarchy using an
aspect. Observe that a collection of static equivalence methods included in, for example the test class, would not work, since the equivalence method must be polymorphic
– which can be achieved by means of introductions in an aspect but not by means of
static methods.
This strategy implements Binder’s Polymorphic Server Test test design pattern [4]. It
can be used to verify that subclasses conform with superclass behavior, and that we are
11
setting up a correct polymorphic hierarchy. It requires exercising each superclass test
case to every possible subclass. In other words, we can reuse the write-read-compare
test case for every subclass of Storable.
4.2 Contract Enforcement in Commands
Aspect Design JH OT D RAW makes use of the Command design pattern in order to separate the user interface from the underlying model, and in order to support such features
as undoing and redoing user commands. Each command has to realize the Command interface, for which a default implementation is provided in the AbstractCommand class.
The key method is execute, which takes care of actually carrying out the command
(such as pasting text, inserting an image, etc.).
Each execute method should start with a consistency check verifying that the underlying “view” exists. Therefore, each concrete implementation of execute starts with
a call to the execute implementation in the superclass, which is always the one from the
AbstractCommand. This is illustrated in Figure 2.
public class AbstractCommand implements Command {
...
public void execute() {
if (view() == null) {
throw new JHotDrawRuntimeException(
"execute should NOT be getting called when view() == null");
} } }
public class PasteCommand extends AbtractCommand {
...
public void execute() {
super.execute();
...
} }
Fig. 2. Contract Enforcement using a super method idiom.
This is a typical example of what is called “contract enforcement” in the A SPECT J
manual [2]. We implemented it using a pointcut capturing all execute methods, putting
the check itself in the advice. Observe that mimicking the implementation where the
check is in a super method is not possible in A SPECT J: super methods cannot be accessed when advising a method. The resulting solution is shown in Figure 3.
The only surprise in this figure may be the within clause in the pointcut. It turns
out that anonymous subclasses of AbstractCommand do not implement the consistency
check. Such classes are used for simple commands such as printing, saving, and exiting
the application. Since A SPECT J does not provide a direct way to exclude anonymous
classes in a pointcut, we used the within operator to exclude executions occurring in
the context of the top level object creating the full user interface. One can also argue
that the anonymous classes should include the check (in which case the exclusion can
be omitted from the pointcut), but at present we focus on keeping the behavior as it was,
not on modifying it.
12
pointcut commandExecute(AbstractCommand aCommand) :
this(aCommand)
&& execution(void AbstractCommand.execute())
&& !within(*..DrawApplication.*);
before(AbstractCommand aCommand) : commandExecute(aCommand) {
if (aCommand.view() == null) {
throw new JHotDrawRuntimeException("...");
} }
Fig. 3. Enforcing the consistency check using before advice.
The main benefit of the aspect approach is that consistency checks cannot be forgotten. This is illustrated by the anonymous classes, but also by one non-anoymous
command,7 which does not extend the AbstractCommand default implementation. Consequently, it cannot reuse the consistency check using a supercall. Inspection of the
execute implementation, however, clearly shows that the code exits with a null pointer
exception in case the check fails. This suggests that the aspect that we are looking for
should implement the check not only for the AbstractCommand class, but for all implementations of the Command interface. Again, our current implementation does not yet
do this, but only injects the implementation in subclasses of AbstractCommand.
A summary of the main issues in the Contract Enforcement refactoring is provided
in Figure 4.
Old situation
Each concrete execute invokes its super execute in order to conduct certain
consistency checks.
Aspect solution The consistency check is implemented as advice, which is invoked before
each call to execute, as captured in a simple pointcut.
Code size
17 explicit consistency calls replaced by one pointcut; consistency check
itself moved from class to advice.
Benefits
Reliability: it becomes impossible to forget the consistency check. Omitted
checks can be fixed automatically thanks to the refactoring.
Risks
Check required that omissions are not on purpose.
A SPECT J issues No direct support to capture anonymous classes; Cannot refer to super methods in method advice.
Fig. 4. Refactoring Contract Enforcement for Commands.
Refactoring Design The restructuring can generally be described as an Advise Method
Overrides refactoring, as presented in Section 5.
Test Suite Design Simple as the pointcut in Figure 3 may be, it nevertheless illustrates
some of the issues involved in testing refactorings that make use of pointcuts.
7
Namely, the UndoableCommand.
13
First of all, adequate testing of the consistency check in the original (non-aspect)
JH OT D RAW version would typically correspond to branch coverage. This yields two
test cases for the top level execute method (one in which the consistency check passes,
and one in which it fails) in addition to one dedicated test for the execute implementation in each subclass. Since the super call can be resolved statically, even polymorphic
adequacy models will not add test cases to this.
It is interesting to observe that such a test suite would not capture the subtleties
involved in designing the aspect from Figure 3. For example, the test suite does not
exercise anonymous classes, nor execute methods occurring outside the scope of AbstractCommand.
The aspect-specific test adequacy criteria as discussed in Section 2.1, however, do
suggest creating the relevant additional test cases. Inspection of the pointcut leads to
the following tests:
– Since AbstractCommand occurs in a type match, we would like to test classes just
off this boundary as well, leading to a test case checking what happens for the
Command interface itself.
– Since the pointcut is a conditional expression, we also want to investigate what
happens if one of the conditions fails. This means that we want to verify that the
within clause does fire for anonymous classes.
Actually creating these test cases may, however, not be as easy as it seems. Testability is affected by controllability and observability, which are poor for anonymous
classes and join point execution.
In order to verify (observe) that our pointcut from Figure 3 does indeed capture
anonymous classes correctly, we created special advice used for testing purposes only,
which keeps track where a certain pointcut expression has fired. To do this, we first
refactored the aspect so that the individual conditions are in separate pointcuts, as shown
in Figure 5. The production aspect uses these pointcuts to perform the consistency check
at the right places. The testing aspect uses exactly the same pointcut definitions to weave
in code that keeps track of where (i.e. at which joinpoints) those pointcuts have fired.
This set of joinpoints is then used to verify intended behavior.
abstract aspect ContractEnforcementPointcut {
pointcut commandExecute(AbstractCommand aCommand) :
this(aCommand)
&& inExecuteMethod()
&& ! inAbstractClass()
pointcut inAbstractClass() :
within(*..DrawApplication.*);
pointcut inExecuteMethod() {
execution(void AbstractCommand.execute());
} }
Fig. 5. Separate pointcuts for each condition to improve aspect testability.
14
Concerning controllability, the instances of the anonymous classes are hard to access. They are normally activated via a mouse event, which must be mimicked in order to trigger the command’s execute method. We avoided the need for generating
mouse events by using an aspect: we intercept the constructors for anonymous command classes, and collect them in a set: after the full application has been built we can
apply the execute method to each command.
4.3 Refactoring the Undo Concern
Background and Current Approach Support for “undo” is a newly added feature in
the analyzed version of JH OT D RAW. As can be imagined, it is a concern that crosscuts
across many different classes. More than 30 elements of the JH OT D RAW framework,
comprising commands, tools and handles, have associated undo constructs to revert the
changes spawned by their underlying activities. The discussion here will focus on the
commands group, as it is the largest in terms of defined undo activities.
Some participants in JH OT D RAW’s undo implementation are shown in Figure 6:
– Each command is associated with one undo activity, whose method undo can be
invoked to revert the command.
– The undo activity is implemented in a nested class of the command, which is instantiated using a factory method called createUndoActivity.
– The primary abstraction in the undo activity is the list of affected figures: when the
command’s execute method is invoked, the relevant state of the affected figures is
stored in the undo activity.
– Undo activities are maintained on a stack by the undo manager.
DrawingEditor
UndoManager
*
<interface>
Command
execute()
createUndoActivity()
undoActivity
AbstractCommand
ConcreteCommand
<interface>
Undoable
undo()
*
affectedFigures
UndoableAdapter
has nested class
ConcreteUndoable
Fig. 6. Participants in JH OT D RAW’s undo implementation.
15
Figure
public class PasteCommand extends FigureTransferCommand {
public void execute() {
...
FigureSelection selection =
(FigureSelection)Clipboard.getClipboard().getContents();
if (selection != null) {
setUndoActivity(createUndoActivity());
... //core command logic and other undo setup
FigureEnumeration fe = insertFigures(...);
getUndoActivity().setAffectedFigures(fe);
...
}
}
}
Fig. 7. The original PasteCommand class.
Aspect Design The aspect solution to undo we propose consists of associating an undodedicated aspect to each undo-able command. The aspect implements the entire undo
functionality for the given command, while the associated class remains oblivious to its
secondary concern. By convention, for enforcing the relation with the command class,
each aspect will consistently be named by appending “UndoActivity” to the name of the
command class. In a successive step, the command’s nested UndoActivity class moves
to the aspect. The factory methods for the undo activities (createUndoActivity()) also
move to the aspect, from where they are introduced back into the associated command
classes using inter-type declarations.
The statements in the execute method that are responsible for setting up the undo
activity, are taken out of the execute method, and woven into it by means of advice.
In some cases the corresponding pointcut simply needs to capture all execute method
calls; in other cases the pointcut is more complex, depending on the way the undo code
is mixed with the regular code.
As an example, consider the paste command, whose execute method consists of
retrieving the selected figures from the clipboard, inserting them into the current view,
and clearing the clipboard. All this is done in a single method, using local variables and
if-then-else statements to deal with such situations as an empty clipboard. The undo
aspect will require the same conditional logic, and access to the same data in the same
order. The following aspect solutions are possible:
– If all getters are side effect free, an approach is to setup the undo activity in a simple
before advice. In JH OT D RAW, however, this is not the case, for example because
of figure enumerators that have an internal state.
– The alternative route is to intercept relevant getters, keep track of the data locally
in the advice as well, and inject advice after all data has been collected. This is
the approach we follow, but some of the pointcuts are somewhat artificial. Figure 8
illustrates such a pointcut in the undo aspect for the PasteCommand, which is also
shown in figure 7. The execute callClipboardgetContents() pointcut captures the
call that sets the reference to be checked by both the command’s core logic and the
undo functionality in the aspect.
– The last possibility is to refactor the long execute method into smaller steps using
non-private methods. The extra method calls can be intercepted allowing smooth
extension with setting up the undo activity, at the cost of creating a larger interface
and breaking encapsulation.
16
public aspect PasteCommandUndoActivity {
//store the Clipboard’s contents - common condition
FigureSelection selection;
pointcut execute_callClipboardgetContents() :
call(Object Clipboard.getContents()) && withincode(void PasteCommand.execute());
after() returning(Object select) : execute_callClipboardgetContents() {
selection = (FigureSelection)select;
}
...
pointcut executePasteCommand(PasteCommand cmd) :
this(cmd) && execution(void PasteCommand.execute());
}
// Execute undo setup
void after(PasteCommand cmd) : executePasteCommand(cmd) {
// the same condition as in the advised method
if(selection != null) {
cmd.setUndoActivity(cmd.createUndoActivity());
...
cmd.getUndoActivity().setAffectedFigures(...);
}
}
Fig. 8. The undo aspect for PasteCommand.
The resulting system differs in two ways from the original design. First, the original
design uses static nested classes to enforce a syntactical relation between the undo activity and its enclosing command class. Since the A SPECT J mechanisms do not allow
introduction of nested classes, the post-refactoring association will only be an indirect
one, based on naming conventions. This is a weaker connection than the one provided
by the original solution. A second difference is that the visibility of certain methods
has been altered, since A SPECT J cannot be used to introduce, for example, the required
factory method as protected.
Refactoring Design The complexity of the refactoring is determined by the complexity
of unplugging undo from the commands themselves. We distinguish different levels of
unpluggability:
1. The nested undo activity class of the command, and all its uses can be safely removed from the command. The fairly simple ChangeAttributeCommand class is an
example in this category.
2. The command’s core logic makes use of some of the data stored in the undo activity.
This is typically done for the list of affected figures. Since there is no real need for
this, we could easily refactor the core logic so that it does not refer to the undo
activity anymore.
3. The nested undo activity not only deals with undo, but also contains core logic
needed for the proper execution of the command. An example is the InsertImageCommand: its undo activity contains a method called insertImage which actually inserts the image (instead of undoing it). We consider this a design violation.
Our solution consists of applying traditional refactorings before starting with the
aspect refactoring, so that the command does not depend on the undo activity anymore.
17
4. The nested undo activity is not only used for this particular command, but also
for similar commands. This is the case for the PasteCommand. Our aspect refactoring will rename the undo activity, and hence requires a simple change to these
commands.
We anticipate that any non-trivial aspect refactoring will require similar objectoriented refactorings, before the crosscutting concern can be taken out of the available
system. A more detailed discussion of the undo concern refactoring, accompanied by
code snippets, is presented in [28].
Old situation
Each command’s execute sets up a corresponding undo activity, which is
implemented through a nested class.
Aspect solution One aspect per command, which contains the undo activity implementation,
and introduces the association into the command. Execute method intercepted to setup the proper undo activity state.
Code size
Remains the same.
Benefits
Strong tangling between commands and their undo activity eliminated;
commands are easier to understand.
Risks
Undo activity may require sophisticated pointcuts to intercept all relevant
state modifications of the command; Refactoring of commands needed in
order to unplug undo support from them.
A SPECT J issues No support for introducing nested classes. Visibility affected since protected
methods cannot be introduced. Modular reasoning affected by keeping track
of data set in the advised method.
Fig. 9. Refactoring Undo.
Test Suite Design In testing undo, we essentially combine the testing approaches of
the persistence and contract enforcement concerns discussed previously.
First of all, we create a reusable test suite at the Command level. This test can be
used for any command subclass, and ensures that each subclass complies with then
intended semantics. This test set takes care of:
– Setting up an appropriate JH OT D RAW application in which a concrete command
can be created. The actual command created is deferred to subclasses of the test
class.
– Bringing the application in a setting in which the execute can be carried out (for
example, many commands require that some figures in the drawing are selected),
and actually invoking it.
– Comparing the effects of the command execution with the intended behavior —
this step is specific to the actual command and deferred to subclasses. It usually
consists of comparing the modified selected figures with a set of figures actually
constructed in the test case.
– Invoking the undo method on the command’s undo-activity, and comparing that the
effects are indeed canceled. Again, this comparison typically involves the set of
affected figures.
18
Thus, the test case follows the template method design pattern, and defers the details of
certain steps to its subclasses.
To test the various pointcuts, the approach described for contract enforcement was
adopted, weaving in special advice that allowed us to observe which pointcut actually
fires.
5 Contributing to the Catalog of Refactorings
Several authors have proposed catalogs of aspect-oriented refactorings [22, 31, 32], in
the spirit of Fowler’s catalog of object-oriented refactorings [14]. We were able to
reuse several of these existing refactorings, such as Monteiro’s Encapsulate Implements
with Declare Parents, and Move Method from Class to Inter-type, or Laddad’s Extract
Method Calls refactoring which encapsulates calls to a method from multiple places
into an aspect.
In this section we add our contribution to these existing catalogs, casting some of
the experiences we obtained from building AJ HOT D RAW into generally reusable refactorings.
An open question is at what level of abstraction such refactorings should be defined.
Is introducing some design pattern considered a refactoring? It is, but Fowler’s book has
explicit refactorings described for just a few design patterns, not all. The reason for this
is, most likely, that the mechanics for introducing such a design pattern can hardly be
described in a reusable way, and for that reason the refactoring description would not
add much useful information to the pattern description. In this respect an interesting
approach is taken by Kerievsky [21], who explicitly addresses refactorings to patterns.
He focuses on a subset of the design patterns, namely those for which common coding
tricks are known that do not yet provide the benefits of using the full pattern, such as in
his Replace Hard-Coded Notifications with Observer refactoring.
A similar distinction holds for aspect refactorings. Introducing refactorings for each
of the prototypical concerns listed in, for example, the A SPECT J programming guide
[2] may not be particularly useful. But in some cases, the “old”, non-aspect solution can
be reasonably well described (for example an Observer implementation following the
guidelines from [15]), and it does make sense to describe how such an implementation
can be refactored into an aspect solution (such as the one from [19]).
If we look at the refactorings from Monteiro, these can be categorized as fairly
technical, elementary refactorings, such as introducing an inter-type declaration [31].
The refactorings from Laddad [22] are more of a mixed style, some being elementary,
others being closer to typical concerns from the A SPECT J manual. Below we try to
provide some building blocks for creating refactoring descriptions that give concrete
advice how certain concerns can be turned into aspects.
Move Role to Aspect Though not discussed in the previous section, several of our
refactorings involve the creation of an aspect-oriented implementation of a design pattern. As an example, JH OT D RAW contains several instantiations of the Observer pattern, which we essentially implemented according to the approach proposed by Hannemann and Kiczales [19].
19
The participants in this pattern can be an observer or a subject. The existing JH OTD RAW implementation does have a separate interface for the observer role, but not for
the subject role. We propose to refactor this and introduce a subject interface via an
aspect in order to: (a) make the two different roles explicit, and (b) remove the observer
pattern details from the primary concerns. Note that in some cases, one class can be
involved in multiple design patterns adopting different roles for them. For example, a
composite figure is a subject as well as an observer, listening to changes in its subfigures
while being listened to by, for example, drawings. The total number of methods implemented by such multi-role classes can be substantial, making them hard to understand;
a problem addressed by moving the roles to aspects.
Thus, Move Role to Aspect creates an interface for a particular role in a design
pattern, and superimposes this role on an existing class by means of an aspect.
Move Observer to Aspect A more high level refactoring is to move an observer implementation into an aspect. This is a compound refactoring, involving three elementary
steps: first, the Move Role to Aspect refactoring is applied twice, once for the subject
and once for the observer role. Subsequently, the calls made in subjects to notify the
observers of changes are captured into a pointcut and extracted into advice.
Override Method with Advice for Overlapping Roles Just like one class can fullfil
multiple roles from one or more different design patterns, one method can implement
features related to multiple roles. This is common in Java Swing design and also occurs
in one of our JH OT D RAW refactorings. This refactoring dealt with the CommandMenu,
which acts as both view and controller for the interactive drawing editor of the application. The method exhibiting the overlapping roles, checkEnabled(), enables/disables
menu items according to the status (executable/non-executable) of the command to be
activated when the item is selected. Although the method belongs to the interface of the
view component, allowing to set the view’s elements status, its implementation relies
on controller decisions.
The proposed refactoring places the method’s definition into the interface for the
role to which it belongs, in this case, the view role, making it accessible to the developer
of the GUI. Furthermore, the controller aspect uses an around advice to override the
default behavior of the method and to make it context(command)-aware.
Advise Method Overrides This refactoring aims at removing duplication arising from
statements that are common to (the start or end of) all method overrides of a given
(superclass) method. Such statements are replaced by advice to any refinement of the
superclass method. Examples in JH OT D RAW include the contract enforcement we discussed previously (the check at the beginning of each execute method), as well as a call
to the checkDamage method that is contained at the end of each execute method.
6 Discussion
What did we learn from refactoring JH OT D RAW to aspects and validating behavior
conservation by means of testing?
20
First of all, we once again learned that testing is actually needed for such refactorings. In several cases, we detected errors in our pointcuts, introductions, and copy-paste
activities thanks to our test suite. Although all of us will agree with this need for testing, it is alarming, to say the least, that neither the popular textbooks on aspect-oriented
programming (such as [23, 2, 16]) nor the existing work on aspect-oriented refactoring
[22, 32] provides any advice on how to approach aspect-oriented testing in a systematic
way.
Second, our fault model as well as our adequacy criteria illustrate how easy it is to
make errors during aspect-oriented programs, and how much needs to be done in order to have a reasonable chance of finding these errors using tests. Moreover, both the
observability (did this pointcut fire?) and the controllability (which inputs will cause a
pointcut to be exercised?) of aspect-oriented programs typically are problematic. Admittedly, at several points in time we were tempted to omit the testing since it seemed
too complicated to create a test suite capable of achieving the required coverage. Testing
tool support may very well help here: but this requires an adequacy model first, which
is what we proposed in the paper.
Concerning the refactorings themselves, our experiments illustrate that being oblvious to future extensions is not as easy as it may seem. For example, the undo concern
was added only in version 5.4 of JH OT D RAW. Could this have been implemented as a
separate aspect without modifying JH OT D RAW version 5.3? Our refactoring shows the
direction this would take. But for some commands, such as the paste command, artificial pointcuts are needed, which are very brittle if the underlying primary logic in the
command changes.
For most cases, assessing the benefits of an aspect-oriented refactoring turned out
to be a fairly subjective process that is hard to quantify. The aspect design step looks
for such solutions that would enhance the system’s evolvability; that is, to achieve a
better modularization for the, otherwise, scattered and tangled parts of a concern, and
to provide an implementation that better reflects the concern-based reasoning over the
system. It is not always apparent, however, in the context of a (relatively) large system as the analyzed case study, that the new, aspect solution surpasses the legacy one.
Although we argue to have improved the separation of concerns, for some more complex refactorings, e.g, undo, the downfalls of the aspect-oriented implementation make
it difficult to asses the improvements for the overall system, or even the gains in modular reasoning over the refactored crosscutting concern. Difficulties could also occur for
less demanding refactorings as for example, contract enforcement, depending on the
uniformity of the places where the contract needs to be enforced.
Last but not least, it is striking that almost every refactoring we experimented with
raised one or more issues concerning A SPECT J (such as visibility modifiers, nested
classes, or anymous classes). Some of these limitations are quite technical in nature,
and are likely to be resolved in future versions of A SPECT J. Also, other aspect-oriented
frameworks, such as AspectWerkz 28 , may offer solutions to some of the issues. Other
limitations are more fundamental (such as the constraint that a class should offer a
zero-argument constructor or the inability to access super methods), and call for a more
rigorous reconsideration of existing aspect-oriented models.
8
aspectwerkz.codehaus.org
21
7 Related Work
An important part of research into the area of refactoring to aspect-orientation has analyzed aspect solutions to a number of (sometimes complex) concerns that typically
crosscut the primary decomposition of a system [22, 19, 23]. The association between
the concern and its aspect solution is an important indication of how a specific language
model is intended to address types of crosscuttings. However, the specific implications
of applying the refactorings in the context of a large system, where deviations from the
examples used to describe the refactorings are very likely, are not considered. In this
paper we showed some of the difficulties that arise when these solutions are applied to
concerns in a large system.
A number of authors investigated the possibility of building catalogs of aspect refactorings. Monteiro and Fernandes [32] proposed a set of code transformations from Java
to A SPECT J specific modularization units, describing steps in a feature extraction process. The approach has followed the format used by Fowler [14] to describe objectoriented refactorings, and was further significantly extended [31]. The study emphasized the mechanics associated to code transformations as opposed to the relation with
typical crosscutting concerns [22, 2, 19]. A similar list is also proposed by Iwamoto and
Zhao [20], but the authors do not provide details about any of the specific refactorings. The attention tends to focus on potential conflicts between the aspect refactorings
and the traditional, object-oriented ones. This issue is also addressed by Hanenberg et
al [17], as well as Hannemann et al, who discuss the possibility of a refactoring approach based on a developer-tool dialog [18].
Specific techniques, like program slicing, are employed by Ettinger and Verbaere [13]
to extract tangled code into method and further into advice, as an extension of the
object-oriented refactoring to aspects.
Closely related to the work described in this paper, Coady et al investigate the benefits of aspect-oriented solutions for evolving operating system code and for better managing its variability [8, 5]. To that end they describe, for example, how the prefetching
concern can be separated from the page handling code in the FreeBSD kernel code [9].
Although their work aims at assessing the benefits of aspect-oriented software development, in contrast to the work presented in this paper, it has not led to a publicly available
aspect-oriented and non-aspect-oriented version of the same software system which can
be used for comparative experimental software evolution research by other researchers.
However, none of this refactoring work mentions a testing strategy that accompanies
the migration process. The attention given to testing in the context of aspect-orientation
is limited and not with concerns to refactoring. Few published test adequacy criteria
for aspect-oriented programming have been formulated: the only work we are aware of
is by Alexander et al, who propose a candidate model and raise a number of research
questions [1]. Ubayashi and Tamai [37] use model checking to verify object crosscutting properties in aspect-oriented programs. As a first attempt to define an approach
for testing aspect-oriented programs, Zhao [39] proposes a data-flow-based unit testing. The tests are oriented towards aspect and class modules that can potentially be
targeted by multiple aspects. Based on the modules’ accessibility three levels of testing
are considered, i.e., intra-module, inter-module, and intra-aspect or intra-class.
22
8 Concluding Remarks
Refactoring to aspect-orientation aims at improving the evolvability and reusability of a
system. Important issues to be considered in this context are (1) the adequacy of the aspect solutions discussed by a number of authors when applied to a large application, (2)
the assessment of the support for and improvements brought by refactoring to aspects,
and (3) the challenges of behavior conservation when migrating to aspect-supported
implementations.
This paper addresses these problems by proposing a refactoring and testing strategy to guide the migration process, and successively by applying it to an open source
Java system. The testing strategy aimed at ensuring migration consistency, introduces
an aspect-oriented fault model and adequacy criteria. Further, aspect and refactoring designs are analyzed for selected concerns in the system under investigation, which also
include new, complex examples of crosscuttings. The analysis consists of a proposed
aspect solution, associated validating tests, and a trade-off review of the pre- and postrefactoring implementations. The difficulties in assessing overall improvements due to
refactoring are turned into considerations about the suitability of the language features
and model for better supporting the types of identified crosscutting concerns. We believe that the development of aspect languages could benefit from catalogs that associate
types of crosscutting concerns to language mechanisms, and we provide further input
for such catalogs.
The paper’s main contributions are (1) an aspect-oriented fault model and adequacy
criteria; (2) a refactoring strategy that emphasizes testing and the use of aspect-oriented
solutions; (3) a detailed discussion of aspect refactorings and their testing implications,
as carried out on an existing system; and (4) the initiation of an open source project that
can be used to experiment with aspect-oriented testing and refactoring, and that can be
used to compare an object with an aspect solution.
The work described in this paper can be extended in various ways. First, we will
continue to experiment with AJ HOT D RAW and other case studies, in order to further
extend the fault model, adequacy criteria, and refactoring catalogs. Second we will use
the proposed models and the experience gained from these case studies to come up with
automated tool support for both testing and refactoring of aspect-oriented programs.
Last but not least, we will analyze the risks and benefits of the various aspect solutions,
and reflect on ways in which some of the limitations of the current solutions can be
resolved.
In order to put our work in a broader perspective, we would like to refer to Bray et
al who state: “assessment of aspect-oriented software development in general is still
arguably in its early days” [5]. We argue that one of the prerequisites for such an assessment is the availability of an aspect-oriented and non-aspect oriented version of the
same software system. Our work aims to create such versions for a publicly available
open source software system and thereby enables experimental comparative software
evolution research to asses the benefits of aspect-orientation.
Acknowledgments We would like to thank Magiel Bruntink (CWI), Hylke van Dijk
(TU Delft), Marco Lormans (TU Delft), and Tom Tourwé (CWI), for reading earlier
drafts of this paper.
23
Partial support was received from SENTERNovem, (Delft University of Technology, project MOOSE, ITEA 01002, and CWI, project IDEALS, hosted by the Embedded Systems Institute).
References
1. R. Alexander, J.M. Bieman, and A. A. Andrews. Towards the systematic testing of aspectoriented programs. Technical Report CS-4-105, Colorado State University, 2004.
2. The AspectJ Team. The AspectJ Programming Guide. Palo Alto Research Center, 2003.
Version 1.2.
3. K. Beck and E. Gamma. Test infected: Programmers love writing tests. Java Report, 3(7):51–
56, 1998. www.junit.org.
4. R. V. Binder. Testing Object-Oriented Systems: Models, Patterns, and Tools. AddisonWesley, 2000.
5. S. Bray, M. Yuen, Y. Coady, and M. E. Fiuczynski. Managing variability in systems: Oh
what a tangled OS we weave. In D. Beuche, K. Czarnecki, M. Mezini, C. Schwanninger,
and M. Voelter, editors, Managing Variabilities Consistently in Design and Code (MVCDC
OOPSLA04), 2004.
6. M. Bruntink, A. van Deursen, R. van Engelen, and T. Tourwé. An evaluation of clone detection techniques for identifying cross-cutting concerns. In Proc. of Int. Conf. on Software
Maintenance (ICSM 2004). IEEE CS, 2004.
7. C. Clifton and G. Leavens. Obliviousness, modular reasoning, and the behavioral subtyping
analogy. In Workshop on Software-Engineering Properties of Languages for Aspect Technologies (SPLAT’03), 2003. http://www.daimi.au.dk/˜eernst/splat03/.
8. Y. Coady and G. Kiczales. Back to the future: A retroactive study of aspect evolution in
operating system code. In Mehmet Akşit, editor, Proc. of 2nd Int. Conf. on Aspect-Oriented
Software Development (AOSD), pages 50–59. ACM Press, March 2003.
9. Y. Coady, G. Kiczales, M. Feeley, and G. Smolyn. Using AspectC to improve the modularity
of path-specific customization in operating system code. In Proc. of 8th European Softw.
Eng. Conf. (ESEC/FSE), pages 88–98. ACM Press, 2001.
10. A. van Deursen. Program comprehension risks and benefits in extreme programming. In
E. Burd, P. Aiken, and R. Koschke, editors, Proc. 8th Working Conf. on Reverse Engineering
(WCRE 2001), pages 176–185. IEEE CS Press, 2001.
11. A. van Deursen and L. Moonen. The video store revisited - thoughts on refactoring and
testing. In M. Marchesi and G. Succi, editors, Proc. of 3nd Int. Conf. on Extreme Programming and Flexible Processes in Software Engineering (XP2002), pages 71–76. University of
Cagliari, May 2002.
12. T. Elrad, M. Aksit, G. Kiczales, K. Lieberherr, and H. Ossher. Discussing aspects of AOP.
Communcations of the ACM, 44(10):33–38, 2001.
13. R. Ettinger and M. Verbaere. Untangling: A slice extraction refactoring. In Proc. of 3rd Int.
Conf. on Aspect-Oriented Software Development. ACM Press, 2004.
14. M. Fowler. Refactoring: Improving the Design of Existing Code. Addison Wesley, 1999.
15. E. Gamma, R. Helm, R. Johnson, and J. Vlissides. Design Patterns: Elements of Reusable
Object-Oriented Software. Addison-Wesley, 1994.
16. J. D. Gradecki and N. Lesiecki. Mastering AspectJ - Aspect Oriented Programmingin Java.
Wiley Publishing, Inc., 2003.
17. S. Hanenberg, C. Oberschulte, and R. Unland. Refactoring of aspect-oriented software. In
Proc. of Net.ObjectDays Conference, pages 19–35. Springer-Verlag, 2003.
24
18. J. Hannemann, T. Fritz, and G. C. Murphy. Refactoring to aspects: an interactive approach.
In Proc. of 2003 OOPSLA Workshop on Eclipse technology eXchange, pages 74–78. ACM
Press, 2003.
19. J. Hannemann and G. Kiczales. Design pattern implementation in Java and AspectJ. In
Proc. of 17th Annual ACM conf. on Object-Oriented Programming, Systems, Languages,
and Applications (OOPSLA), pages 161–173. ACM Press, 2002.
20. M. Iwamoto and J. Zhao. Refactoring aspect-oriented programs. In Proc. of 4th AOSD
Modeling With UML Workshop, UML’2003, San Francisco, California, USA, 2003.
21. J. Kerievsky. Refactoring to Patterns. Addison-Wesley, 2004.
22. R. Laddad. Aspect-oriented refactoring. www.theserverside.com, December 2003.
23. R. Laddad. AspectJ in Action - Practical Aspect Oriented Programming. Manning, 2003.
24. M. M. Lehman. On understanding laws, evolution and conservation in the large program life
cycle. Journal of Systems and Software, 1(3):213–221, 1980.
25. M. M. Lehman. Programs, life cycles and laws of software evolution. Proceedings of the
IEEE, 68(9):1060–1076, September 1980. Special Issue on Software Engineering.
26. M. M. Lehman, D. E. Perry, and J. F. Ramil. Implications of evolution metrics on software
maintenance. In T. M. Koshgoftaar and K. Bennett, editors, Proc. of Int. Conf. on Softw.
Maintenance (ICSM 1998), pages 208–217. IEEE CS Press, 1998.
27. B. Liskov and J. M. Wing. A behavioral notion of subtyping. ACM Trans. Program. Lang.
Syst., 16(6):1811–1841, 1994.
28. M. Marin. Refactoring JHotDraw’s Undo concern to AspectJ. In Proc. of First Workshop on
Aspect Reverse Engineering (WARE). Delft University of Technology, 2004.
29. M. Marin, A. van Deursen, and L. Moonen. Identifying aspects using fan-in analysis. In
Proc. of 11th Working Conf. on Reverse Engineering (WCRE2004). IEEE CS, 2004.
30. B. Meyer. Object-Oriented Software Construction. Prentice-Hall, 2nd edition, 1997.
31. M.P. Monteiro. Catalogue of refactorings for AspectJ. Technical Report UM-DI-GECSD200401, Universidade do Minho, 2004.
32. M.P. Monteiro and J.M. Fernandes. Object-to-aspect refactorings for feature extraction. In
Proc. of 3rd Int. Conf. on Aspect-Oriented Software Development. ACM Press, 2004. Industry paper.
33. D. L. Parnas. Software aging. In Proc. of 16th Int. Conf. on Softw. Eng. (ICSE), pages
279–287. IEEE CS Press, 1994.
34. M. C. Rinard, A. Salcianu, and S. Bugrara. A classification system and analysis for aspectoriented programs. In Proc. of 12th ACM SIGSOFT Int. Symp. on Foundations of Software
Engineering (FSE), pages 147–158. ACM Press, 2004.
35. A. Rountev, A. Milanova, and B. G. Ryder. Fragment class analysis for testing of polymorphism in Java software. IEEE Trans. Softw. Eng., 30(6):372–387, 2004.
36. P. Tonella and M. Ceccato. Migrating interface implementation to aspect oriented programming. In Proc. of Int. Conf. on Softw. Maintenance (ICSM 2004). IEEE CS Press, 2004.
37. N. Ubayashi and T. Tamai. Aspect-oriented programming with model checking. In Proc.
of 1st Int. Conf. on Aspect-oriented software development (AOSD), pages 148–154. ACM
Press, 2002.
38. E. Weyuker. The evaluation of program-based software test data adequacy criteria. Communications of the ACM, 31(6):668–675, June 1988.
39. J. Zhao. Data-flow-based unit testing of aspect-oriented programs. In Proc. of the 27th
Annual IEEE Int. Computer Software and Applications Conference (COMPSAC), pages 188–
197. IEEE CS, 2003.
25
| 2cs.AI
|
An Empirical Study of Online Packet Scheduling
Algorithms
Nourhan Sakr? and Cliff Stein??
arXiv:1603.07947v1 [cs.DS] 25 Mar 2016
Industrial Engineering and Operations Research, Columbia University, NY 10027
Abstract. This work studies online scheduling algorithms for buffer
management, develops new algorithms, and analyzes their performances.
Packets arrive at a release time r, with a non-negative weight w and an
integer deadline d. At each time step, at most one packet is scheduled.
The modified greedy (MG) algorithm is 1.618-competitive for the objective of maximizing the sum of weights of packets sent, assuming agreeable
deadlines. We analyze the empirical behavior of MG in a situation with
arbitrary deadlines and demonstrate that it is at a disadvantage when
frequently preferring maximum weight packets over early deadline ones.
We develop the MLP algorithm, which remedies this problem whilst
mimicking the behavior of the offline algorithm. Our comparative analysis shows that, although the competitive ratio of MLP is not as good
as that of MG, it performs better in practice. We validate this by simulating the behavior of both algorithms under a spectrum of simulated
parameter settings. Finally, we propose the design of three additional
algorithms, which may help in improving performance in practice.
1
Introduction
Efficient buffer management at a network router is a critical issue that motivates
the online packet scheduling problem. Kesselman et al. [13] introduce a buffer
management delay model and give algorithms to minimize end-to-end delay. We
adopt a similar model to analyze the empirical behavior of the modified greedy
(MG) algorithm introduced in [12], and propose new algorithms that do not have
as strong worst-case guarantees, yet perform better in our simulated settings.
Model. For simplicity, we investigate a network router with two nodes. Studying
a two node router is a first step towards understanding more complicated and
realistic models. In § 7 we briefly discuss possible model modifications. At each
integer time step, packets are buffered upon arrival at the source node, then at
most one packet is chosen from the buffer to be sent to the target node. A packet
(r,d,w) arrives at a release date r, has a non-negative weight w, and needs to
be sent by an integer deadline d. A packet not sent by d expires, and is dropped
from the buffer. The objective of a packet-scheduling algorithm A is to maximize
?
??
[email protected]
Supported in part by NSF grant CCF-1421161. [email protected]
its weighted throughput, ζA , defined as the total weight of packets sent by A. It
is easy to relate our model to an online version of the classical offline unit-job
scheduling problem where the input is a set of n unit-length jobs, each specified
by a similar triple (r,d,w) and the objective is to maximize weighted througput,
that is the total weight of jobs that are processed before their deadlines.
Parameters. We will typically be generating our input according to some type
of distribution. Let T denote the number of time steps during which the system
can generate arriving packets, and let λ denote an arrival rate. We choose values
for T and λ. Then at each integer time step t = 1, . . . , T , we first generate the
number of arriving packets according to a Poisson distribution with rate λ. For
each arriving packet, we set r = t and generate w from a uniform (integer)
distribution U (1, wmax ). To find d, we first generate τ , a time to expire, from
a uniform (integer) distribution U (0, dmax ), and set d = r + τ . We call this
Model 1. We also consider a bimodal distribution for τ with weights p and 1 − p,
respectively, for two distinct distributions centered on different means and call
this Model 2. Although a network may induce correlations between packets, we
use i.i.d. distributions as a first step in modeling the behavior of our algorithms.
In order to evaluate the performance of an online scheduling algorithm (A),
we use an offline algorithm (OFF) for comparison, which given all future arrivals
and packet characteristics, is able to statically find the optimal solution (e.g. using maximum-weight bipartite matching). Its solution gives the highest possible
throughput the system can achieve. The online algorithm is k-competitive if ζA
on any instance is at least 1/k of ζOF F on this instance. The smallest k for
which an algorithm is k-competitive is called the competitive ratio [3]. According to [11], k will be at most 2 for any algorithm that uses a static priority
policy. In this paper, we will simulate the online algorithm and evaluate the
ratio ζA /ζOF F . The average of these ratios across each batch of simulations will
be denoted by ρA , where A is the corresponding online algorithm.
Related Work. The literature is rich with works that acknowledge the importance of buffer management and present algorithms aiming at better router performance. Motivated by [13], [7] gives a randomized algorithm, RMIX, while [4]
e
-competitive against an adaptive-online adversary.
proves that it remains e−1
Many researchers attempt to design algorithms with improved competitive ratios. The best lower bound on the competitive ratio of deterministic algorithms is
the golden ratio φ [11, 5]. A simple greedy algorithm that schedules a maximumweight pending packet for an arbitrary deadline instance is 2-competitive [11, 13].
Chrobak et al. [8] introduce the first deterministic algorithm to have a competitive ratio strictly less than 2, namely 1.939. Li et al. [15] use the idea of
dummy packets in order to design the DP algorithm with competitive ratio
at most 1.854. Independently, [10] gives a 1.828-competitive algorithm. Further
research considers natural restrictions on packet deadlines with hopes of improving the competitive ratio. One type of restriction is the agreeable deadline
model considered in [12], i.e. deadlines are (weakly) increasing in their release
times. Motivated by a more general greedy algorithm, EDFα [7], that schedules
the earliest-deadline pending packet with weight at least 1/α of the maximumweight pending packet, [12] develop the MG algorithm which will be described
in § 2. In other models, researchers enforce the FIFO discipline using a model
where packets have no deadlines and the buffer is finite. One of the earliest such
algorithms is the FIFO preemptive model studied by [2]. Works such as [9, 14, 16]
adopt similar ideas. We do not consider the FIFO discipline in this paper.
Our Contribution. We observe that while MG is φ-competitive for the case
of agreeable deadlines, it may not be the best option to apply in practice. We
demonstrate the undesirable performance of MG under certain scenarios, e.g. frequently preferring maximum weight (late deadline) packets over early deadline
ones. Our proposed MLP algorithm remedies this drawback, as it outperforms
MG on most simulated instances. However, we are able to develop hard instances
to prove that MLP does not provide better worst-case guarantees, whereas on
those instances MG would produce the same results as an offline solution. Contrasting the advantages of MG and MLP motivates us to explore further algorithmic adjustments which may improve performance, at least in practice, as
supported by our preliminary analysis. Finally, we justify that a two-node model
with an infinite buffer is a sufficient model for our analysis. Moreover, extending
the model to multiple nodes or imposing a threshold on the capacity of the buffer
does not significantly alter the performance of the online algorithms.
2
Modified Greedy Algorithm (MG)
MG is a φ-competitive deterministic online algorithm for the agreeable deadline
model [12]. It focuses on two packets at each time step: the earliest deadline nondominated packet e (i.e. maximum weight among all earliest-deadline packets)
and the maximum weight non-dominated packet h (i.e. earliest deadline among
all maximum-weight packets in the buffer). Packet e is chosen if we ≥ wφh (φ ≈
1.618) and h is chosen otherwise. While [12] consider an agreeable deadline
model, we relax this assumption and explore MG in a more general setting.
MG analysis. Although MG has the best competitive ratio among deterministic
online algorithms, we believe that by better understanding MG, we can improve
on it in practice. Intuitively, if MG, at early stages, chooses packets with longer
deadlines (due to higher weights) over those with early deadlines, then as time
passes, many early packets expire while most of the heavy later-deadline packets
will have already been sent. Therefore, the algorithm may resort to choosing
packets with even smaller weights, thereby wasting an opportunity to send a
higher weight packet that has already expired. We, hence, explore the decisions
made by MG by observing its relative frequency of choosing h over e.
In order to consider a diverse set of instances, we set T to 200 and define
ranges [0.7,20], [1,20] and [1,40] for λ, wmax and dmax , respectively. Under the
assumptions of Model 1, we run a batch of 200 simulations each for 8000 sampled
parameter combinations. Given each parameter combination, we calculate the
relative frequency of choosing h over e and average the frequencies over λ to
obtain the empirical probability of P (wh > 1.618we ), denoted by ψ.
We suspect that when ψ is high, especially if h expires at later deadlines, MG
will be at a major disadvantage. Figure 1
plots ψ vs. wmax , where each curve corresponds to a fixed level for dmax . In general, ψ increases with wmax and dmax . The
decreasing curve slope implies that ψ is
more sensitive to lower values of wmax .
Further analysis shows that at any level
of wmax , MG will choose h over e at most
66% of the time. We also observe that regardless the average number of packets in
the buffer, if wmax is small (less than 3),
the event of interest occurs at most 40%
of the time.
Fig. 1: ψ vs. wmax , colored by dmax
(a) ρM G vs. λ
(b) ρM G vs. dmax
Fig. 2: Performance of MG under Scenario 1
From this probability analysis, we conclude that unless wmax or λ are small,
MG tends to choose packet h too frequently. To show that this property may ”fire
back”, we construct Scenario 1, forcing MG to favor later deadlines. We reuse
the data of the generated packets above and adjust the weight of each packet by
multiplying it by its deadline. We let MG run on the new data and plot ρM G
against different parameters. Figure 2 depicts lower ratios for the new dataset.
While ρM G originally increased with λ, it now decreases with λ and dmax . wmax
does not affect the performance much. A gradient boosted tree predictive model
(i.e. a sequence of decision tree models where the next model is built upon the
residuals of the previous one) shows that dmax is the most important factor
under Scenario 1, as it accounts for 30% of the variability in the model.
3
Mini LP Algorithm (MLP)
Description. In light of the previous analysis, we develop a new online algorithm that is more likely to send early deadline packets. The mini LP Algorithm
(MLP) runs a ”mini” assignment LP at each time step in order to find the optimal schedule for the current content of the buffer, assuming no more arrivals.
Assuming nt is the number of packets in the buffer at current time t, we search
for the packet with the latest deadline (dt,max ) and set a timeline from t̂= 0 to
dt,max − t. We then solve the following optimization problem, where wi is the
weight of packet i and xit̂ is 1 if packet i is sent at time t̂ and 0 otherwise:
min
X
wi xit̂
i,t̂
s.t.
X
xit̂ ≤ 1
t̂ = 0, . . . , dt,max − t
xit̂ ≤ 1
i = 1, . . . , nt
xit̂ ≥ 0
i = 1, . . . , nt
i
X
t̂
MLP then uses the optimal solution to send the packet that receives the first
assignment, i.e. the packet i for which xi0 =1, while the rest of the schedule is
ignored and recomputed in subsequent time steps.
3.1
Initial Analysis
Similar to the MG analysis, we compute ρM LP and are interested in its behavior
as the load varies. A way to measure load is to define the average number of
packets in the buffer as n̄, which is a byproduct of λ. We expect higher ρM LP
at low n̄, since the online algorithm would not have many packets to choose
from and hence, is more likely to choose the same packet as the offline algorithm
at each time step. However, we expect ρM LP to decrease as n̄ increases, since
the discrepancy between online and offline solutions increases. To test this, we
sample parameter combinations from T ∈ [100, 500], λ ∈ [0.7, 20] and wmax ,
dmax ∈ [1, 20] and run a batch of 1000 simulations per combination.
Figure 3 plots ρM LP vs. n̄ and interestingly shows a dip-shaped graph:
ρM LP starts at a very high value (≈ 1),
decreases as expected with increasing
n̄ until eventually it increases again,
thereby forming a dip, and finally converges to 1. Our claim is true at first,
when λ is relatively low, as the first
range for λ is quite sensitive (λ = 2.3
vs. 2.8 makes a difference). However,
when λ increases, the problem loses its
sensitivity. An explanation for such beFig. 3: ρM LP vs. n̄
havior may be that as n̄ increases (with
increasing λ), we are more likely to have multiple packets achieving maximum
weight, in which case both the online and offline algorithms are likely to choose
those packets and have less discrepancy between their choices, especially if the
weight or deadline ranges are not wide enough. We conclude that when the system is heavily or lightly loaded, both algorithms perform well. The dip happens
in the interesting area. Consequently, we will investigate how the dip moves and
what the effect of parameter choices will have on such graph.
3.2
Parameter Effect on MLP Behavior
Changing the parameters to generate different graphs did not change the structure of the dip-shaped graph that we have seen in Figure 3. Nonetheless, the
dip gets narrower/wider and shifts to the left/right, as parameters change. In
this section, we will only focus on a restricted range for the values of λ, namely
0.7 to 10. However, we believe that the restriction does not mask any interesting results, since MLP converges at higher values of λ, as we have seen before.
Therefore, a heavily loaded system is not significant for our analysis.
Arrival Rates. The graph inevitably depends on λ, as it directly affects n̄, i.e.
the x-axis. However, λ does not have a direct effect on the shape of the graph. By
tuning the range for λ, we are able to “zoom in” onto the dip area and monitor
the behavior more accurately where the system is neither lightly nor heavily
loaded. The range for such sensitive values is on average between 1.3 and 4.2.
Figure B.1.1a (in the appendix) zooms in on the dip where λ is most sensitive.
Weight Ranges. The range of the weights moves the dip to the right (left), as
it gets narrower (wider). Very narrow ranges (i.e. low values for wmax ) are the
most influential. As wmax increases, its impact decreases. In fact, this result
seems intuitive and one can see an example in Figure B.1.1b where the weight
range is designed to be very narrow (wmax is set at 2). Some experimentation
led us to the explanation of this phenomenon: When there are few options for
weights, both algorithms converge together. Let’s say weights are only 1 and 2,
then the higher the n̄, the more likely we will have packets of weight 2. In this
case both algorithms find the optimal choice to be the packet with higher weight
(we don’t have much choice here so it must be 2). Hence, both behave alike.
We note that it is not in particular the range of weights that has this effect but
rather the number of distinct weights available, i.e. choosing between weights 1
and 2 vs. 100 and 200, would depict the same behavior.
Time Period and Deadline Range. T and dmax have a combined effect. Figures B.1.1c and B.1.1d give two examples: Allowing a longer timeline T results
in a second but higher dip and slows down convergence, such that suddenly
higher values of λ become slightly more interesting. Meanwhile lower dmax values (combined with shorter T ’s) result in a graph with one sharp dip as well as
much faster convergence to 1.
3.3
Influence of maximum-weight packets
The motivation of MLP was mainly to remedy the drawback we observed for
MG when later deadline packets are preferred. Therefore, it is essential to verify
that MLP outperforms MG under Scenario 1. In fact, one-sided 99% confidence
intervals (CI) imply that ρM G is at most 91.98% while ρM LP is at most 96.28%.
The difference in performance between both algorithms increases with λ. Figure 4a shows the behavior of ρ against n̄ for both algorithms. While MLP is
not influenced by n̄ under this scenario, the performance of MG gets worse as
, denoted by ρ̂, is (1.0443,1.0552),
n̄ increases. A 99% two-sided CI for ρρMMLP
G
implying that MLP produces a total weight at least 4.43% more than that of
MG under this scenario. Better performance is observed with higher T or lower
dmax , but wmax does not seem to influence the algorithms’ performances.
(a) ρ vs. n̄ under Scenario 1
(b) ρ vs. λ under Scenario 2
Fig. 4: ρM LP (red) and ρM G (green)
4
Comparative Analysis
In this section, we contrast the behavior of MG and MLP under a spectrum of
parameter settings. We are interested in the behavior of the ratios against our
parameters and expect MLP to perform better in our simulations. The general
procedure for our simulations is based on sampling parameter combinations from
a predefined parameter space. we impose the following parameter range restrictions: T ∈ (50, 750), λ ∈ (0.5, 50), wmax ∈ (2, 50) and dmax ∈ (1, 50) (Scenario
2). For each combination, we run MG, MLP (5 times each) and the offline algorithm in order to obtain values for ρM G , ρM LP , as well as ρ̂ = ζζMMLP
. Detailed
G
steps for simulations are given in § A.1.
4.1
Ratio behavior w.r.t. model parameters
Comparing ρM LP and ρM G against values of λ implies that on average MLP
outperforms MG (Figure 4b). As λ increases, both algorithms perform better. A
99% one-sided CI for ρM G is (0,0.9734), implying that we are 99% confident that
ζM G is at most 97.34% of ζOF F , while the one-sided CI for ρM LP is (0,0.9926).
In Figure B.2.1, it is evident that MG produces a wider spread of the ratios. All
(a) under Scenario 2
(b) under Scenario 3
Fig. 5: ρ̂ vs. λ and colored by T
else constant, the performance of each algorithm improves with higher T , lower
dmax or higher n̄, whereas it is not influenced by the values of wmax .
Figure 5a plots ρ̂ vs. λ, colored by T . For very small λ’s, there is the possibility
that MLP and MG perform similarly; in some cases, MG outperforms MLP,
regardless of the value of T . However, for large λ, MLP tends to outperform
MG. A 99% two-sided CI for ρ̂ is (1.0188, 1.0216), implying that we are 99%
confident that ζM LP is at least 1.88% more than ζM G . However, both algorithms
have similar performance as the upper bound of the CI shows that ζM LP is at
most 2.16% more. Whether this is beneficial depends on the use case as well
as time constraints (see § 5). § A.2 presents a brief analysis where we construct
gradient booted tree predictive models on the ratios for inference purposes.
4.2
Changing the distribution of τ
So far, we have only considered uniform distributions, however, real inputs are
more complicated. Here we make one step towards modeling more realistic inputs
and consider a τ that follows a bimodal distribution of two distinct peaks (recall
Model 2); with probability p, τ is N (2, 0.52 ) and with probability 1 − p, τ is
N (8, 0.752 ). We restrict our parameters to the following ranges: T ∈ (100, 300),
Fig. 6: ρM LP (red) and ρM G (green) vs. λ under Scenario 3
λ ∈ (0.7, 6), wmax ∈ (2, 7) and p ∈ (0.75, 0.95) (Scenario 3). We choose a bimodal
distribution because these distributions are often hard for scheduling algorithms.
Indeed, we see that the results for Scenario 3 are slightly different.
While MG performs worse with increasing λ, ρM LP improves with λ and still
outperforms ρM G (Figure 6).The graph for ρM LP resembles a dip-shaped graph,
yet we find this dip to be entirely above the confidence interval of ρM G . All else
constant, neither algorithm is influenced greatly by any of the parameters T ,
dmax or p. Figure 5b plots ρ̂ vs. λ, where lighter points correspond to longer
T ’s. For very small λ, MLP and MG perform similarly. In some cases, MG
outperforms MLP, regardless of the value of T . However, for large λ, a 95% CI
shows that MLP outperforms MG by at least 2.80% and at most 3.30%.
5
Hard Instances
The previous analysis presents evidence
that MLP gives better competitive ratios than MG. An index plot (Figure 7)
of ρM G and ρM LP shows that, for the
same instances, MLP not only outperforms MG, but also gives a ratio of 1
for most of the instances that are hard
for MG. However, it would be incorrect
to conclude that MLP always has a better competitive ratio than MG. In fact,
we are able to create hard instances for
MLP where it performs worse than MG.
A small example is given in Table 1.
Fig. 7: Index plot of ρM LP (red) and
ρM G (green)
Packet(r,d,w)
(1, 1, w1 )
(1, 2, w2 )
(2, 2, w2 )
Throughput
MLP
MG
Offline
Assign to t = 1
Assign to t = 2
w1 + w2
Assign to t = 1
Assign to t = 2
2w2
Assign to t = 1
Assign to t = 2
2w2
Table 1: Hard Instance for MLP
For w2 > w1 , we can easily see that MLP is 2-competitive, while ρM G on
those instances is 1. However, such worst-case instances may be rare and our
results show that MLP performs better on a varied set of data. On the other
hand, MG, which bases its decisions on simple arithmetic, is simpler than MLP
which solves an LP at each time step. In our experiments, MLP was as much as
140 times slower than MG. In the next section, we consider some modifications
to take advantage of the strengths of both approaches.
6
Algorithm Modifications
In an attempt to find faster solutions, we introduce some possible algorithmic
modifications. Preliminary results show slight performance improvement when
these modifications are applied. However, more analysis is needed to verify the
results and choose the best parameters for improvement.
6.1
The Mix and Match Algorithm (MM)
Algorithm. The Mix and Match Algorithm (MM) combines both MG and MLP.
At each time step, MM chooses to either run MG or MLP-according to n̄. If n̄
is high, then by previous analysis, MG and MLP each converges to 1 (assuming
Model 1), and MM runs MG, as it is faster and has a competetive ratio that is
as good as that of MLP. If n̄ is low, MM runs MLP, as it is more accurate and
the running time is also small since n̄ is low. To distinguish between ”high” and
”low”, we define a threshold N̄ . Although MM suffers from the same limitations
as MLP, it might still be preferred due to its smaller computation time.
Simulation. We set T to 200 and define ranges [0.7,15], [1,30], [1,23] and [5,20] for
λ, wmax , dmax and N̄ , respectively. We compare ζM M under different values of
N̄ . We also average ζM M , use the same simulations to run MG and average ζM G .
We take the ratio of both averages and plot it against n̄ (Figure 8). Preliminary
results show that that for small n̄, the algorithm, at higher N̄ , does slightly
worse than at lower N̄ . However, the opposite is true for large n̄.
(a) N̄ = 5
(b) N̄ = 20
Fig. 8: Average ζM M over average ζM G vs. n̄
Future work. Further ideas are needed to set the optimal choice for N̄ . We may
want to look at the percentage of times the algorithm chose to run MG over
MLP in order to monitor time complexity. Another idea would be to take hard
instances of MG into consideration and explore how to derive conditions such
that the algorithm switches to MLP in such cases.
6.2
The Learning Modified Greedy Algorithm (LMG)
Algorithm. MG and MLP are memoryless [12], i.e. the assignment of packets at
time t uses no information about assignments at times t0 < t. We introduce a
non-memoryless modification to MG. Recall that MG compares we to wφh . In the
learning MG (LMG) algorithm, we try to make use of the past performance of
MG, in order to replace φ by a more suitable divisor. If f defines the frequency of
learning, then every f steps, we calculate the throughput at the current divisor,
φ∗ . Then we search for a divisor, φbetter , which lies in the vicinity of φ∗ , but
yields higher throughput using the previous data. Finally, we proceed with the
new divisor φ∗new , given by (αφ∗ + (1 − α)φbetter ) for some smoothing factor
α ∈ [0, 1]. The detailed procedure for LMG is given below:
Step 0. Set the frequency of learning, f , i.e. the time window needed to define
a learning epoch. Then run MG and use the following procedure every f steps
in order to replace the divisor, φ, by a sequence of better divisors as follows:
1. Generate a sequence of divisors φi ’s starting at the current divisor φ∗ and
having jumps of ±0.05, without going below 1 or above 2.5. For instance, if
φ∗ =1.62, we generate the following sequence: 1.02, 1.07, . . ., 1.57, 1.62, 1.67,
. . ., 2.47.
2. Start with the throughput associated with φ∗ and move left in our generated
sequence. At each φi , we calculate the throughput of MG on the previous
data. We keep moving to subsequent divisors as long as there is an increase
in throughput. Next, we do the same with divisors to the right. Given a left
endpoint and a right one, we choose the divisor associated with the higher
throughput and denote it by φbetter . Some toy examples are shown in Table 2.
For simplicity, we only observe the weighted throughput for four values of φ.
Thruput/φi 1.57 φ∗ =1.62
Case 1
2
4
Case 2
2
4
Case 3
6
4
Case 4
4
4
Case 5
6
4
1.67
4
2
5
4
3
1.72 φbetter
2
1.62
6
1.62
7
1.72
4
1.62
7
1.57
Table 2: Examples for choosing φbetter
3. The new divisor φ∗new is given by smoothing φ∗ with φbetter , i.e. for some
α ∈ [0, 1]
φ∗new = αφ∗ + (1 − α)φbetter
Simulation. We use the same parameter space as in § 6.1. We set f = max(0.1 ∗
30
) and for simplicity, α = 0.5. The choice for α in general must ensure
T, min(1,λ)
that the process of finding an optimal divisor φ does not generate a jumpy
sequence of divisors. Our analysis for 8000 sampled scenarios shows that LMG
outperforms MG 83.3% of the time. The range of the improvement is [-0.6%,
2.8%], implying that LMG brings as much as 2.8% increase in the ratio over MG.
Performance is worse when the sequence of φ∗new ’s is around 1.618, implying that
LMG is picking up on some noise and should not change the divisor.
Future work. One can avoid this noise by statistically testing and justifying the
significance of changing the divisor. In terms of time complexity, LMG is not
slower than MG, as it can be done while the regular process is running. Finally,
no direct conclusion is made about a threshold on the number of packets beyond
which LMG is particularly effective. Further analysis could yield such conclusion,
thereby indicating at which instances LMG should be used.
6.3
The Second Max Algorithm (SMMG)
Algorithm. Inspired by the dummy packet (DP) algorithm discussed in [15] for
cases of non-agreeable deadlines, we realize the importance of extending the
comparison to a pool of more than two packets. The key idea in SMMG is to
prevent the influence a single heavy packet may have on subsequent steps of
the algorithm. We try to find an early-deadline packet that is sufficiently large
compared to the heaviest packet. We set a value for p ∈ (0, 1) and the iterations
are as follows:
1. If MG chooses e, send e and STOP.
2. Else find the earliest second largest packet in the buffer, denoted by s.
3. If ds < dh and ws ≥ max(we , p ∗ wh ), send s. Else send h.
The intuition here is that sending packet e is always a good choice, so we need
no modification. However, we limit over-choosing packet h by finding the earliest
second-largest packet s. The concern is that keeping s in the buffer may bias
the choice of the packets. Hence, we send s, if its weight is significant enough,
in order to eliminate its influence and keep the possibility of sending h for a
subsequent iteration (as h expires after s). To evaluate that ws is significant
enough, we verify that it exceeds we (otherwise, we should have sent e), as well
as p ∗ wh , a proportion of wh . Note that for instance, if p = 0.95, it means that
SMMG is very conservative allowing the fewest modifications to MG.
Simulation. We use the same parameter space as in § 6.1 and try values for p
as follows: 0.65, 0.75, 0.85, 0.95. Figure 9 plots the improvement of SMMG over
MG (ρM G - ρSM M G ) vs. n̄, colored by p. As expected, the lower the value of p,
the bigger the deviation from MG. At very low n̄, we see that applying SMMG
is not useful, however, as n̄ increases, the improvement remains positive. At all
values of p, the improvement is at its highest when n̄ is between 8 and 12 packets.
Hence, SMMG is useful when n̄ is in the vicinity of the interval between 4 and
17. Whether this is a significant result, depends on the nature of our problem.
Even if p = 0.95, the minimum improvement within that interval is around 0.8%.
However, the maximum improvement is 1.5% (at p = 0.95).
Fig. 9: (ρM G - ρSM M G ) vs. n̄, colored by p
7
Model Discussion
The two-node model with no buffer limitations clearly does not capture all aspects of realistic network models. Thus, in this section, we will consider a multinode model and also the case of finite buffer capacity.
Multi-node model. In the previous analysis, we only considered a two-nodesystem, namely the source and the target nodes. In order to understand multinode systems, we consider first a three node system, that is a system of two
tandem queues and see how the throughput behaves.
Assume the arrival rate at node 1 is λ1 and that each packet has to be sent
to visit node 2 then reach node 3 by its deadline. Some packets will be lost at
node 1, as they expire before being sent. Node 2 will, hence, have less traffic. We
assume as before that we can send at most one packet per node at each time step.
Within this framework, we are interested to know whether this setup results in a
deterministic online algorithm of better performance. Our simulation shows that
node 2 either has the same throughput as node 1 or lower. After tracing packets,
it turns out that this is a fairly logical result because node 2 only receives packets
from node 1. The packet will either expire at stage 2 or go through to node 3.
So the throughput can only be at most the same as that of node 1.
The following minor adjustment slightly improves the performance at node 2:
Each arriving packet has a deadline to reach node 3, denoted by d. We introduce
a temporary deadline for that packet to reach node 2 by d − 1. This modification
guarantees that we only send packets to node 2 if after arriving at node 2 there
is at least one more time unit left to its expiration in order to give the packet a
chance to reach node 3. Here is a trivial example: A packet arrives with deadline
7, i.e. it should arrive at node 3 by 7. Before the adjustment it was possible for
this packet at time 7 to be still at node 1 and move to node 2, then be expired at
node 2 and get lost. After the adjustment, this packet will have a deadline of 6
for node 2. So if by time 6, the packet hasn’t been sent yet, it gets deleted from
the buffer of node 1 (one time step before its actual deadline). This adjustment
improved the throughput of node 2 to be almost equal to that of node 1 because
the arrival rate at node 2 is at most 1 (at most one packet is sent from node 1
at each time step). So node 2 is usually making a trivial decision of sending the
only packet it has in its buffer.
In conclusion, our model implicitly imposes a restriction on the maximum
possible throughput at internal nodes, hence, making the multi-node model,
where only one packet is sent at each time step, an uninteresting problem. In § 8,
we give, however, a few future directions for more interesting extensions.
Finiteness of Buffer. Throughout this paper, we chose not to put any restrictions
on the buffer capacity and therefore we
now look to verify whether the finiteness
of the buffer has a major effect on the
algorithm performance and its bounds.
We run a set of experiments where, given
λ ∈ [2, 100], we find the corresponding
Fig. 10: b/λ vs. λ
buffer size b, such that the probability of
exceeding b is one in a million. We create
an index plot for the ratio of b/λ (Figure 10) and conclude that imposing a buffer
size is unnecessary. Even when λ = 100, the buffer size needs to be about 1.53
as much, i.e. 153. Furthermore, for the interesting values of λ used throughout
this report, a buffer size of around 30 would be more than sufficient. We conclude that imposing a capacity limit on the buffer is not necessary as choosing
a reasonable buffer size should not affect the algorithm’s performance.
8
Conclusion
In this paper, we consider several old and new packet scheduling algorithms. By
analyzing the empirical behavior of MG, we observe that MG chooses packet
h over e too frequently. We therefore develop a new algorithm, MLP, which
mimics the offline algorithm and gives higher attention to early deadline packets.
We then show that on a wide variety of data, including uniform and bimodal
distributions, MLP is slower, but has a better empirical competitive ratio than
MG. (This is in contrast to the worst-case analysis where MG has a better
competitive ratio.)
We then propose three new algorithms that may offer an improvement in
empirical performance, as they combine features of both algorithms. MM, at each
time step, chooses between using MG or MLP in order to make a decision on the
packet to send. LMG learns from previous behavior to correct the divisor used in
MG, while SMMG is motivated by the idea of influential packets in extending the
comparison to a pool of three packets, namely e, h and s. The improvements for
these algorithms are small, yet encouraging for further analysis. Moreover, it is
important to consider extensions for the network model and run the algorithms
on one where induced correlations are captured by more realistic distributions
that are not i.i.d. Contrasting the behavior of any of the algorithms mentioned
in this paper on an actual router, rather than a simulated environment, would
also be important to consider.
Several interesting future directions remain. One important extension would
be a multi-node model. We showed how the straightforward extension does not
yield much insight but other extensions may be more interesting. For example,
one could have nodes that process at different rates; this would prevent the first
node from being an obvious bottleneck. Another possibility is to allow feedback,
that is, if a packet expires somewhere in the multi-node system, it could return
to the source to be resent. A final possibility is for each packet to have a vector
of deadlines, one per node, so that different nodes could be the bottleneck at
different times.
Acknowledgments
The authors would like to thank Dr. Shokri Z. Selim and Javid Ali.
References
[1] S. Albers and M. Schmidt. On the performance of greedy algorithms in packet
buffering. SIAM Journal on Computing (SICOMP), 35(2):278–304, 2005.
[2] N. Andelman, Y. Mansour, and A. Zhu. Competitive queuing polices for QoS
switches. In Proceedings of the 14th Annual ACM-SIAM Symposium on Discrete
Algorithms (SODA), pages 761–770, 2003.
[3] A. Borodin and R. El-Yaniv Online Computation and Competitive Analysis. Cambridge University Press, 1998
[4] M. Bieńkowski, M. Chrobak, and L. Jeż. Randomized algorithms for buffer management with 2-bounded delay. In Proceedings of the 6th Workshop on Approximation
and Online Algorithms (WAOA), pages 92–104, 2008.
[5] F. Y. L. Chin and S. P. Y. Fung Online Scheduling with Partial Job Values: Does
Timesharing or Randomization Help? Algorithmica, 37(3): 149–164, 2003
[6] F. Y. L. Chin and S. P. Y. Fung Improved competitive algorithms for online scheduling with partial job values. Theoretical computer science, 325(3): 467-478,2004.
[7] F. Y. L. Chin, M. Chrobak, S. P. Y. Fung, W. Jawor, J. Sgall, and T. Tichy. Online
competitive algorithms for maximizing weighted throughput of unit jobs. Journal
of Discrete Algorithms, 4(2):255–276, 2006.
[8] M. Chrobak, W. Jawor, J. Sgall, and T. Tichy Improved Online Algorithms for
Buffer Management in QoS Switches In Proceedings of 12th European Symposium
on Algorithms (ESA), pages 204–215, 2004.
[9] M. Englert and M. Westermann Lower and Upper Bounds on FIFO Buffer Management in QoS Switches. In Proceedings of 14th Annual European Symposium on
Algorithms (ESA), pages 352–363, 2006.
[10] M. Englert and M. Westermann Considering Suppressed Packets Improves Buffer
Management in QoS Switches. In Proceedings of 18th Annual ACM-SIAM Symposium on Discrete Algorithms (SODA), pages 209–218, 2007.
[11] B. Hajek. On the competitiveness of online scheduling of unit-length packets with
hard deadlines in slotted time. In Proceedings of 2001 Conference on Information
Sciences and Systems (CISS), pages 434–438, 2001.
[12] L. Jeż, F. Li, J. Sethuraman, and C. Stein. Online scheduling of packets with
agreeable deadlines. ACM Transactions on Algorithms (TALG), 9(1):5, 2012.
[13] A. Kesselman, Z. Lotker, Y. Mansour, B. Patt-Shamir, B. Schieber, and M. Sviridenko. Buffer overflow management in QoS switches. SIAM Journal on Computing
(SICOMP), 33(3):563–583, 2004.
[14] A. Kesselman, Y. Mansour, R. van Stee. Improved Competitive Guarantees for
QoS Buffering. Algorithmica, 43:63–80, 2005.
[15] F. Li, J. Sethuraman, and C. Stein. Better online buffer management. In Proceedings of the 18th Annual ACM-SIAM Symposium on Discrete Algorithms (SODA),
pages 199–208, 2007.
[16] Z. Lotker and B. Patt-Shamir Nearly Optimal FIFO Buffer Management for
DiffServ. In Proceedings of the 21st ACM Symposium on Principles of Distributed
Computing (PODC), pages 134–142, 2002.
A
Appendix
A.1
General Procedure for Simulations in § 4
Our simulations try to avoid any initial bias by starting with a nonempty system.
This is the role of the variable κ which ”warms up” the system; that is the simulation will take place for κ time steps before collecting any data. This warmup
period allows the content of the buffer to adjust to normal running conditions
for the system. The general simulation procedure is as follows:
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
Identify a parameter range for the parameters T , λ, wmax and dmax .
Randomly select a parameter combination within the parameter space.
Set κ = max(40, 0.4T ).
Generate inputs for T +κ time steps using the chosen parameter combination.
Find the optimal offline solution, for the last T steps only, i.e. sum of weights
of packets sent from t = κ + 1 to t = κ + T and let that be our ζOF F .
Run MG and record the weighted throughput for the last T steps only.
Repeat Step 6 five times and let our ζM G be the average weighted throughput
over the five runs.
Run MLP and record the weighted throughput for the last T steps only.
Repeat Step 8 five times and let our ζM LP be the average weighted throughput over the five runs.
.
Compute ρM G , ρM LP , as well as ρ̂ = ζζMMLP
G
Repeat Steps 2-10 500 times.
Note that the input generation in Step 3 has been described in § 1.
A.2
Predictive Model
We construct a gradient boosted tree (GBT) predictive model on the results for
inference purposes. The predictors were dmax , wmax , T , λ and whether MG or
MLP were used and the response was ρ of the respective algorithm. In terms of
variable importance (as measured by the model and stated as a percentage), λ
and T turn out to be the two most important parameters.
Feature
λ T dmax wmax MLP? MG?
Variable Importance 41% 15% 8% 3%
17% 16%
Table A1: Variable Importance for a GBT predictive model
Figure A.2.1a plots both model parameters for each of the algorithms, and
colors them by ρ. This allows us to view the behavior of the ratio in a multivariate
manner: We see here that there is a point where the ratios tend to level off,
approximately (as can be seen by the constancy of the colors) around λ > 20
and T > 200. We looked to see if this was perhaps a result of the average number
of packets in the buffer but saw slight correlation, as is evident in Figure A.2.1b.
(a) Colored
ρM LP (right)
by
ρM G (left)
and
(b) Colored
n̄M LP (right)
by
n̄M G (left)
Fig. A.2.1: T vs. λ
(a) acutal ρM LP
(b) predicted ρM LP
(c) acutal ρM G
(d) predicted ρM G
Fig. A.2.2: 3D locally smoothed plot of ρ vs. T and λ
and
Next, we perform predictions from the model for a much larger range of T and
λ (risking extrapolation issues). In Figure A.2.2b, ρM LP is well behaved for large
values of T and λ: there seems to be a dip but that may be due to randomness.
Furthermore, the ratios do not dip as sharply as for MG in Figure A.2.2d. This
leads us to investigate more the behavior of ρM G for larger values of λ, which
leads us to Scenario 4 where values for λ are extended to (0.5, 250) and the same
analysis is run for MG.
The results for the behavior of ρM G vs. λ shows an increasing performance
under Scenario 4. We are 99% confident that ζM G is at most 98.69% of ζOF F .
Plots against other parameters show that the performance of MG, all else constant, would get better with larger T or lower dmax and only slightly better for
larger wmax . A gradient boosted tree predictive model again shows that λ and
T are the two most important variables. As λ and T get larger, ρM G gets better.
B
Figures
B.1
Parameter Effect on MLP Behavior
(a) λ focuses on sensitive values
(b) wmax reflects narrow weight ranges
(c) longer T
(d) smaller dmax
Fig. B.1.1: ρM LP vs. n̄
B.2
ρM LP (red) and ρM G (green) vs. different parameters
(a) T
(b) wmax
(c) dmax
(d) n̄
Fig. B.2.1: ρM LP (red) and ρM G (green) vs. different parameters
| 8cs.DS
|
Distributed Computation of Large-scale Graph Problems
Hartmut Klauck∗
arXiv:1311.6209v6 [cs.DC] 16 Sep 2015
Abstract
Danupon Nanongkai†
Gopal Pandurangan‡
Peter Robinson§
lower bound for connectivity, ST verification and other
related problems. Our lower bounds develop and use
new bounds in random-partition communication complexity.
To complement our lower bounds, we also give algorithms for various fundamental graph problems, e.g.,
PageRank, MST, connectivity, ST verification, shortest paths, cuts, spanners, covering problems, densest
subgraph, subgraph isomorphism, finding triangles, etc.
We show that problems such as PageRank, MST, connectivity, and graph covering can be solved in Õ(n/k)
time (the notation Õ hides polylog(n) factors and an additive polylog(n) term); this shows that one can achieve
almost linear (in k) speedup, whereas for shortest
paths,
√
we present algorithms that run in Õ(n/ k) time (for
(1 + ǫ)-factor approximation) and in Õ(n/k) time (for
O(log n)-factor approximation) respectively.
Our results are a step towards understanding the
complexity of distributively solving large-scale graph
problems.
Motivated by the increasing need for fast distributed
processing of large-scale graphs such as the Web graph
and various social networks, we study a number of
fundamental graph problems in the message-passing
model, where we have k machines that jointly perform a
computation on an arbitrary n-node (typically, n ≫ k)
input graph. The graph is assumed to be randomly
partitioned among the k ≥ 2 machines (a common
implementation in many real world systems). The
communication is point-to-point, and the goal is to
minimize the time complexity, i.e., the number of
communication rounds, of solving various fundamental
graph problems.
We present lower bounds that quantify the fundamental time limitations of distributively solving graph
problems. We first show a lower bound of Ω(n/k)
rounds for computing a spanning tree (ST) of the input graph. This result also implies the same bound for
other fundamental problems such as computing a minimum spanning tree (MST), breadth-first tree (BFS),
and shortest paths tree (SPT). We also show an Ω(n/k 2 ) 1 Introduction
The emergence of “Big Data” over the last decade or so
∗ Division of Mathematical Sciences, Nanyang Technological
has led to new computing platforms for distributed proUniversity, Singapore 637371 & Centre for Quantum Technolo- cessing of large-scale data, exemplified by MapReduce
gies, Singapore 117543. E-mail: [email protected]. This work [12] and more recently systems such as Pregel [29] and
is funded by the Singapore Ministry of Education (partly through
Giraph[34]. In these platforms, the data — which is simthe Academic Research Fund Tier 3 MOE2012-T3-1-009) and by
ply too large to fit into a single machine — is distributed
the Singapore National Research Foundation.
† KTH Royal Institute of Technology, Sweden, and University
across a group of machines that are connected via a
of Vienna, Austria E-mail: [email protected]. Work done while communication network and the machines jointly proat ICERM, Brown University, USA, and Nanyang Technological cess the data in a distributed fashion. The focus of this
University, Singapore.
paper is on distributed processing of large-scale graphs
‡ Department of Computer Science, University of Houston,
Houston, TX 77204, USA. E-mail: [email protected]. which is increasingly becoming important with the rise
Work done while at Division of Mathematical Sciences, Nanyang of massive graphs such as the Web graph, social netTechnological University, Singapore 637371 & Department of works, biological networks, and other graph-structured
Computer Science, Brown University, Providence, RI 02912, USA. data and the consequent need for fast graph algorithms
Supported in part by the following research grants: Nanyang
on such large-scale graph data. Indeed, there has been
Technological University grant M58110000, Singapore Ministry
of Education (MOE) Academic Research Fund (AcRF) Tier 2 a recent proliferation of systems designed specifically
grant MOE2010-T2-2-082, Singapore MOE AcRF Tier 1 grant for large-scale graph processing, e.g., Pregel [29], GiMOE2012-T1-001-094, and a grant from the US-Israel Binational raph [34], GraphLab [26], and GPS [1]. MapReduce
Science Foundation (BSF).
(developed at Google [12]) has become a very success§ Department of Computer Science, National University of Sinful distributed computing platform for a wide variety
gapore. E-mail: [email protected]. Research supported
by the grant MOE2011-T2-2-042 “Fault-tolerant Communica- of large-scale computing applications and also has been
tion Complexity in Wireless Networks” from the Singapore MoE used for processing graphs [25]. However, as pointed out
AcRF-2.
by the developers of Pregel (which was also developed
at Google), MapReduce may sometimes be ill-suited for
implementing graph algorithms; it can lead to “suboptimal performance and usability issues” [29]. On the
other hand, they mention that graph algorithms seem
better suited to a message-passing distributed computing model [33, 28] and this is the main design principle
[29] behind Pregel (and other systems that followed it
such as Giraph [34] and GPS [1]). While there is a rich
theory for the message-passing distributed computing
model [33, 28], such a theory is still in its infancy for
distributed graph processing systems.
In this work, our goal is to investigate a theory for
large-scale graph computation based on a distributed
message-passing model. A fundamental issue that we
would like to investigate is the amount of “speedup”
possible in such a model vis-a-vis the number of machines used: more precisely, if we use k machines, does
the run time scale linearly (or even super-linearly) in k?
And what are the fundamental time bounds for various
graph problems?
1.1 Model We consider a network of k > 1 (distinct)
machines N = {p1 , . . . , pk } that are pairwise interconnected by bidirectional point-to-point communication
links — henceforth called the k-machine model1 . Each
machine executes an instance of a distributed algorithm
A. The computation advances in synchronous rounds
where, in each round, machines can exchange messages
over their communication links. Each link is assumed to
have a bandwidth of W , i.e., W bits can be transmitted
over the link in one round. (In stating our time bounds,
for convenience, we will assume that bandwidth W = 1;
in any case, it is easy to rewrite our upper bounds to
scale in terms of parameter W — cf. Theorem 4.1).
(Note that machines have no other means of communication and do not share any memory.) There is an
alternate — but equivalent — way to view our communication restriction: instead of putting bandwidth
restriction on the links, we can put a restriction on the
amount of information that each machine can communicate (i.e. send/receive) in a round. The results that
we obtain in the bandwidth-restricted model will also
apply to the latter model (cf. Section 4.2). Local computation within a machine is considered free, while communicating messages between the machines is the costly
operation2 .
1 Our results can also be generalized to work if the communication network is a sparse topology as well if one assumes an
underlying routing mechanism; details are omitted here.
2 This assumption is reasonable in the context of large-scale
data, e.g., it has been made in the context of theoretical analysis
of MapReduce, see e.g., [35] for a justification. Indeed, typically
We are interested in solving graph problems where
we are given an input graph G of n vertices (assume
that each vertex has a unique label) and m edges
from some input domain G. To avoid trivialities, we
will assume that n ≥ k (typically n ≫ k). Unless
otherwise stated, we will consider G to be undirected,
although all our results can be made to apply in
a straightforward fashion to directed graphs as well.
Initially, the entire graph G is not known by a single
machine, but rather partitioned among the k machines
in a “balanced” fashion, i.e., the nodes and/or edges
of G must be partitioned approximately evenly among
the machines. We will assume a vertex-partition model,
where vertices (and their incident edges) are partitioned
across machines. One type of partition that we will
assume throughout is the random (vertex) partition, i.e.,
vertices (and its incident edges) of the input graph are
assigned randomly to machines. (This is the typical
way that many real systems (e.g., Pregel) partition
the input graph among the machines; it is simple and
and easy to accomplish, e.g., via hashing 3 .) Our
upper bounds will also hold (with slight modifications)
without this assumption; only a “balanced” partition of
the input graph among the machines is needed. On
the other hand, our lower bounds apply even under
random partitioning, hence they apply to worst-case
partition as well. It can be shown that (cf. Lemma 4.1)
a random partition gives rises to an (approximately)
balanced partition.
Formally, in the random vertex partition (RVP)
model, each vertex of G is assigned independently and
randomly to one of the k machines. If a vertex v is
assigned to machine pi we call pi the home machine of
v. Note that when a vertex is assigned to a machine, all
its incident edges are assigned to that machine as well;
i.e., the home machine will know the labels of neighbors
of that vertex as well as the identity of the home
machines of the neighboring vertices. A convenient way
to implement the above assignment is via hashing: each
vertex (label) is hashed to one of the k machines. Hence,
if a machine knows a vertex label, it also knows where
it is hashed to.
Depending on the problem P, the vertices and/or
edges of G have labels chosen from a set of polynomial
(in n) size. Eventually, each machine pi (1 ≤ i ≤ k)
in practice, even assuming the links have a bandwidth of order of
gigabytes of data per second, the amount of data that have been
to be communicated can be in order of tera or peta bytes which
generally dominates the overall computation cost [35].
3 Partitioning based on the structure of the graph — with the
goal of minimizing the amount of communication between the
machines — is non-trivial; finding such a “good” partition itself
might be prohibitively expensive and can be problem dependent.
Some papers address this issue, see e.g., [40, 3, 39].
must (irrevocably) set a designated local output variable oi (which will depend on the set of vertices assigned to machine pi ) and the output configuration
o = ho1 , . . . , ok i must satisfy certain feasibility conditions w.r.t. problem P. For example, when considering the minimum spanning tree (MST) problem, each
oi corresponds to a set of edges (which will be a subset of edges incident on vertices mapped to machine)
pi and the edges in the union of the sets oi must form
an MST of the input graph G; in other words, each
machine pi will know all the MST edges incident on
vertices mapped to pi . (Note that this is a natural generalization of the analogous assumption in the standard
distributed message passing model, where each vertex
knows which of its incident edges belong to the MST
[33].) We say that algorithm A solves problem P if A
maps each G ∈ G to an output configuration that is feasible for P. The time complexity of A is the maximum
number of rounds until termination, over all graphs in
G. In stating our time bounds, for convenience, we will
assume that bandwidth W = 1; in any case, it is easy to
rewrite our upper bounds to scale in terms of parameter
W (cf. Theorem 4.1).
Notation. For any 0 ≤ ǫ ≤ 1, we say that a protocol
has ǫ-error if, for any input graph G, it outputs the
correct answer with probability at least 1 − ǫ, where
the probability is over the random partition and the
random bit strings used by the algorithm (in case it is
randomized).
For any n > 0 and function T (n), we say that an
algorithm A terminates in O(T (n)) rounds if, for any nnode graph G, A always terminate in O(T (n))) rounds,
regardless of the choice of the (random) input partition.
For any n and problem P on n node graphs, we let
the time complexity of solving P with ǫ error probability
in the k-machine model, denoted by Tǫk (P), be the
minimum T (n) such that there exists an ǫ-error protocol
that solves P and terminates in T (n) rounds. For any
0 ≤ ǫ ≤ 1, graph problem P and function T : Z+ → Z+ ,
we say that Tǫk (P) = O(T (n)) if there exists integer
n0 and c such that for all n ≥ n0 , Tǫk (P) ≤ cT (n).
Similarly, we say that Tǫk (P) = Ω(T (n)) if there exists
integer n0 and real c such that for all n ≥ n0 , Tǫk (P) ≥
cT (n). For our upper bounds, we will usually use
ǫ = 1/n, which will imply high probability algorithms,
i.e., succeeding with probability at least 1 − 1/n. In this
case, we will sometimes just omit ǫ and simply say the
time bound applies “with high probability”. We use ∆
to denote the maximum degree of any node in the input
graph, and D for the diameter of the input graph.
of distributed “rounds”, for solving various fundamental
graph problems. The time complexity not only captures
the (potential) speed up possible for a problem, but it
also implicitly captures the communication cost of the
algorithm as well, since links can transmit only a limited
amount of bits per round; equivalently, we can view our
model where instead of links, machines can send/receive
only a limited amount of bits per round. We develop
techniques to obtain non-trivial lower and upper bounds
on the time complexity of various graph problems.
Lower Bounds. Our lower bounds quantify the fundamental time limitations of distributively solving graph
problems. They apply essentially to distributed data
computations in all point-to-point communication models, since they apply even to a synchronous complete
network model where the graph is partitioned randomly
(unlike some previous results, e.g., [41], which apply
only under some worst-case partition).
We first give a tight lower bound on the complexity of computing a spanning tree (cf. Section 2). The
proof shows that Ω(n/k) rounds of communication are
needed even for unweighted and undirected graphs of
diameter 2, and even for sparse graphs. We give an
information theoretic argument for this result. This result also implies the same bound for other fundamental
problems such as computing a minimum spanning tree,
breadth-first tree, and shortest paths tree. This bound
shows that one cannot hope to obtain a run time that
scales (asymptotically) faster than 1/k. This result, in
conjunction with our upper bound of Õ(n/k) for computing a MST, shows that this lower bound is essentially
tight.
We then show an Ω(n/k 2 ) lower bound for connectivity, spanning tree (ST) verification and other related
verification problems (cf. Section 3). To analyze the
complexity of verification problems we give reductions
from problems in the 2-player communication complexity model using random partitions of the input variables.
As opposed to the standard fixed partition model here
all input bits are randomly assigned to Alice and Bob.
We give a tight lower bound for randomized protocols
for the well-studied disjointness problem in this setting.
In particular, we show a lower bound on the randomized
average partition communication complexity of the disjointness problem which might be of independent interest. Random partition communication complexity has
also been studied by Chakrabarti et al. [8], but their
results apply to the promise disjointness problem in the
multiparty number in hand model for a sufficiently large
number of players only. In our proof we apply the rectangle based arguments of Razborov [36], but we need
1.2 Our Results and Techniques Our main goal
to take care of several issues arising. A core ingrediis to investigate the time complexity, i.e., the number
ent of Razborov’s proof is the conditioning that turns
the input distribution into a product distribution. With
randomly assigned inputs we need to recover the necessary product properties by conditioning over badly
assigned input bits. Even when doing so the size of the
sets in the input are no longer exactly as in Razborov’s
proof. Furthermore, there is a large probability that the
set intersection is visible to a single player right from
the start, but with a small enough error probability the
communication still needs to be large.
pute exact shortest paths, this might take significantly
longer (e.g., using Bellman-Ford — cf. Section 4.3). For
graph covering problems such a Maximal Independent
Set (MIS) and (approximate) Minimum Vertex cover
(MVC), we show a bound of Õ(min(n/k, m/k 2 + ∆/k));
note that this implies a bound of Õ(n/k 2 ) for (constant)
bounded degree graphs, i.e., we can get a speed up that
scales superlinearly in k.
We finally note that our results also directly apply
to an alternate (but equivalent) model, where instead
Algorithms and Upper Bounds. We introduce techof having a restriction on the number of bits individual
niques to obtain fast graph algorithms in the k-machine
links can transmit in a round, we restrict the number
model (cf. Section 1.1). We first present a general reof bits a machine can send/receive (in total) per round
sult, called the Conversion Theorem (cf. Theorem 4.1)
(cf. Section 4.2).
that, given a graph problem P, shows how fast algorithms for solving P in the k-machine model can be
1.3 Related Work The theoretical study of (largedesigned by leveraging distributed algorithms for P in
scale) graph processing in distributed systems is relathe standard CONGEST message-passing distributed
tively recent. Several works have been devoted to decomputing model (see e.g., [33]). We note that fast
veloping MapReduce graph algorithms (e.g., see [25, 35]
distributed algorithms in the standard model do not diand the references therein). There also have been sevrectly imply fast algorithms in the k-machine model. To
eral recent theoretical papers analyzing MapReduce alachieve this, we consider distributed algorithms in an ingorithms in general, including Mapreduce graph algotermediate clique model (cf. Section 4.2) and then show
rithms see e.g., [24, 35, 19] and the references therein.
two ways — parts (a) and (b) respectively of the ConWe note that the flavor of theory developed for MapReversion Theorem — to efficiently convert algorithms in
duce is quite different compared to the distributed comthe clique model to the k-machine model. Part (b) applexity results of this paper. Minimizing communicaplies to converting distributed algorithms (in the clique
tion (which leads in turn to minimizing the number of
model) that only uses broadcast, while part (a) applies
communication rounds) is a key motivation in MapReto any algorithm. Part (a) will sometimes give better
duce algorithms (e.g., see [35]); however this is gentime bounds compared to part (b) and vice versa —
erally achieved by making sure that the data is made
this depends on the problem at hand and the type of
small enough quickly (in a small number of MapReduce
distributed algorithm considered, as well as on the graph
rounds) to fit into the memory of a single machine. An
parameters. (The latter can be especially useful in apexample of this idea is the filtering technique of [24] applications where we might have some information on the
plied to graph problems. The main idea behind filtering
graph parameters/topology as explained below.) Using
is to reduce the size of the input in a distributed fashion
this theorem, we design algorithms for various fundaso that the resulting, much smaller, problem instance
mental graph problems, e.g., PageRank, minimum spancan be solved on a single machine. Filtering allows for
ning tree (MST), connectivity, spanning tree (ST) vera tradeoff between the number of rounds and the availification, shortest paths, cuts, spanners, covering probable memory. Specifically, the work of [24] shows that
lems, densest subgraph, subgraph isomorphism, Trianfor graphs with at most n1+c edges and machines with
gle finding (cf. Table 1). We show that problems such as
memory at least n1+ε will require O(c/ǫ) (MapReduce)
PageRank, MST, and connectivity, graph covering etc.
rounds.
can be solved in Õ(n/k) time; this shows that one can
The work that is closest in spirit to ours is the reachieve almost linear (in k) speedup. For graph conneccent work of [41]. The above work considers a number
tivity, BFS tree construction, and ST verification, we
of basic statistical and graph problems in the messageshow Õ(min(n/k, m/k 2 + D∆/k)) bound — note that
passing model (where the data is distributed across a set
the second part of the above bound may be better in
of machines) and analyzes their communication comsome cases, e.g., if the graph is sparse (i.e., m = O(n))
plexity — which denotes the total number of bits exand D and ∆ are small (e.g., bounded by O(log n))
changed in all messages across the machines during a
2
— then we get a bound of Õ(n/k ). For single-source
computation. Their main result is that exact compushortest paths, another classic
√ and important problem, tation of many statistical and graph problems in the
we show a bound of Õ(n/ k) for a (1 + ǫ)-factor apdistributed setting is very expensive, and often one canproximation and a bound of Õ(n/k) for O(log n)-factor
not do better than simply having all machines send
approximation. We note that if one wants to com-
Problem
Upper Bound
Lower Bound
Minimum Spanning Tree (MST)
Connectivity, Spanning Tree Verification (Conn,ST)
Breadth First Search Tree (BFS)
Single-Source Shortest-Paths Distances (SSSP)
Single-Source Shortest-Paths Tree (SPT)
All-Pairs Shortest-Paths Distances (APSP)
PageRank with reset prob. γ (PageRank)
Graph Covering Problems (MIS, MVC)
Maximal Ind. Set on Hypergraphs (HMIS)
(2δ − 1)-Spanner (Spanner) (δ ∈ O(log n))
Densest Subgraph (DSGraph)
Triangle Verification (Tri)
Subgraph Isomorphism (SubIso) (d-vertex subgraph)
Õ(n/k)
Õ(min(n/k, m/k 2 + D⌈∆/k⌉))
2
Õ(min(n/k
+ D⌈∆/k⌉))
√ † + D, m/k
$
Õ(n/√k) , Õ(n/k)
Õ(n/ k)† , Õ(n/k)$
√
Õ(n n/k)# , Õ(n/k)$
Õ(n/γk)
Õ(min(n/k, m/k 2 + ∆/k))
Õ(n/k + k)
Õ(n/k)
Õ(n/k) (for (2 + ǫ)-approx.)
Õ(min(n∆2 /k 2 + ∆⌈∆/k⌉, n7/3 /k 2 + n4/3 /k))
Õ(n2+(d−2)/d /k 2 + n1+(d−2)/d /k)
Ω̃(n/k)∗
Ω̃(n/k 2 )
Ω̃(n/k)
†
(1 + ǫ)-approximation.
#
(2 + ǫ)-approximation.
$
O(log n)-approximation.
∗
Ω̃(n/k)∗
For any approx. ratio.
Table 1: Complexity bounds in the k-machine model for an n-node input graph with m edges, max degree ∆, and diameter
D. ǫ > 0 is any small constant. The notation Õ hides polylog(n) factors and an additive polylog(n) term. For clarity of
presentation, we assume a bandwidth of Θ(log n) bits.
their data to a centralized server. The graph problems considered are computing the degree of a vertex,
testing cycle-freeness, testing connectivity, computing
the number of connected components, testing bipartiteness, and testing triangle-freeness. The strong lower
bounds shown for these assume a worst-case distribution of the input (unlike ours, which assumes a random
distribution). They posit that in order to obtain protocols that are communication-efficient, one has to allow
approximation, or investigate the distribution or layout
of the data sets and leave these as open problems for
future. Our work, on the other hand, addresses time
(round) complexity (this is different from the notion of
round complexity defined in [41]) and shows that nontrivial speed up is possible for many graph problems.
As posited above, for some problems such as shortest
paths and densest subgraph etc., our model assumes
a random partition of the input graph and also allows
approximation to get good speedup, while for problems
such as MST we get good speedups for exact algorithms
as well. For spanning tree problems we show tight lower
bounds as well.
The k-machine model is closely related to the wellstudied (standard) message-passing CONGEST model
[33], in particular to the CONGEST clique model (cf.
Section 4). The main difference is that while many vertices of the input graph are mapped to the same machine in the k-machine model, in the standard model
each vertex corresponds to a dedicated machine. More
“local knowledge” is available per vertex (since it can
access for free information about other vertices in the
same machine) in the k-machine model compared to the
standard model. On the other hand, all nodes assigned
to a machine have to communicate through the links
incident on this machine, which can limit the bandwidth (unlike the standard model where each vertex has
a dedicated processor). These differences manifest in
the time complexity — certain problems have a faster
time complexity in one model compared to the other
(cf. Section 4). In particular, the fastest known distributed algorithm in the standard model for a given
problem, may not give rise to the fastest algorithm in
the k-machine model. Furthermore, the techniques for
showing the complexity bounds (both upper and lower)
in the k-machine model are different compared to the
standard model. The recently developed communication complexity techniques (see e.g, [37, 32, 15]) used to
prove lower bounds in the standard CONGEST model
are not applicable here.
2
Tight Lower Bounds for Spanning Tree
Computation
In this section, we show that any ε-error algorithm takes
Ω(n/k) rounds to compute a spanning tree (STC). Recall
that in the STC problem, for every edge uv, the home
machines of u and v must know whether uv is in the ST
or not.
Theorem 2.1. (Lower bound for STC) Every
public-coin ǫ-error randomized protocol on a k-machine
network that computes a spanning tree of an n-node
input graph has an expected round complexity of
More specifically, there exists a constant
Ω nk .
ǫ > 0 such that, for any k ≥ 3 and large enough
n,
Tǫk (STC), Tǫk (MST), Tǫk (BFS), Tǫk (SPT) ∈ Ω nk .
Observation 1. For any X, Y chosen as described
above, and all 1 ≤ i ≤ b such that Xi = Yi = 1 exactly
one of the edges uvi or vi w must be part of any spanning
tree, except for exactly one such i, for which both edges
Proof. We first show the theorem for Tǫk (STC) using are in the spanning tree. For all other i the one edge uvi
an information theoretic argument. Assume for a or vi w that is in the graph must also be in the spanning
contradiction that there exists a distributed algorithm tree.
in the k-machine model, denoted by R, that violates
Theorem 2.1. In other words, R solves STC correctly
Since all edges are adjacent to either u or w, and
with probability at least 1 − ǫ and always terminates in any spanning tree must have b + 1 edges, one of p1 or p2
δn
k rounds, for some δ ∈ o(1). We will show that the must output at most b/2 edges. Before the first round
information flow to at least one machine must be large. of communication the entropy H(Y |X) is 2b/3 by the
following calculation:
Graph Gb (X, Y ). Let b = n − 2. For any X, Y ⊆ [b],
X
we construct the following graph, denoted by Gb (X, Y ).
H(Y
|X)
=
Pr(X = x) · H(Y |X = x)
The vertices (and adjacent edges of Gb (X, Y )) will be
x
assigned to a random machine.
b
X
Gb (X, Y ) consists of b + 2 nodes, denoted by
b ℓ
= 3−b
2 · log 2ℓ
v1 , . . . , vb and u, w. For each 1 ≤ i ≤ b, we add edge uvi
ℓ
ℓ=0
to Gb (X, Y ) if i ∈ X, and we add edge vi w to Gb (X, Y )
b−1
X
b − 1 ℓ+1
if i ∈ Y .
= 3−b b
2
The random strings X, Y will be drawn from a disℓ
ℓ=0
tribution that will ensure that Gb (X, Y ) is connected.
= 2b/3.
Furthermore the graph will contain roughly 4b/3 edges
with high probability (2b/3 adjacent to u, w respectively), hence roughly b/3 edges must be removed to
Besides X the machine p1 also knows some vertices
obtain a spanning tree.
vi and their edges, giving it access to some bits of Y .
To produce the correct output in R the machine
It is easy to see via the Chernoff bound that with very
that receives u must know which edges uvi are included
high probability p1 knows at most (1 + ζ)b/k bits of
in the spanning tree, and similarly the machine that
Y for ζ = 0.01, lowering the conditional entropy of Y
receives w must know which edges vi w are included
given those bits to no less than 2b/3 − (1 + ζ)b/k. The
in the spanning tree. Since X, Y are encoded in the
event where p1 knows more cannot influence the entropy
graph, the machine p1 who receives u knows X via
2
by more than 2−ζ b/(3k) · b = o(1) (for b large enough).
the edges uvi at the start, but has initially limited
Hence the entropy of Y given the initial information
information about Y , unless it also receives w, which
of p1 , which we denote by a random variable A, is
happens with probability 1/k. With probability 1 − 1/k
H(Y |A) ≥ 2b/3 − (1 + ζ)b/k − o(1).
the information about Y initially held by p1 comes from
Assume that p1 outputs at most b/2 edges. This
the set of vertices vi held by p1 which is of size ≈ b/k
event or the corresponding event for p2 must happen
with high probability, giving p1 at most ≈ b/k bits of
with probability 1 − ǫ assuming failure probability ǫ.
information about Y . Hence all information needed to
Conditioned on the event that p1 outputs b/2 or fewer
decide which edges to not include must come into p1 via
edges we can estimate the entropy H(Y |A, T0 ) for the
communication with the other k − 1 machines. We show
random variable T0 containing the transcript of all
that this communication is Ω(b), if p1 outputs at most
messages to p1 . H(Y |A, T0 ) ≤ H(Y |X, E) where E is
b/2 edges as part of the spanning tree (and a symmetric
the random variables of edges in the output of p1 (we
argument holds for the machine p2 that holds w). Hence
use that X and E can be computed from A, T0 ). Given
the number of rounds is Ω(b/k) = Ω(n/k).
that X = x there are |x| possible edges for E. With
In order to give a clean condition on which edges
probability 1 − o(1) we have |y| < 2b/3 + ζb.
are necessary for a spanning tree we use the following
For at most b/2 edges in E there are at most
distribution on X, Y . X, Y (viewed as characteristic
b
b
vectors) are chosen uniformly from {0, 1} × {0, 1}
X b/2
b/2
under the condition that for every i ∈ [b] we have
≤b·
ℓ
b/6 + ζb
Xi + Yi ≥ 1. Hence there are exactly 3b possible values
ℓ<b/6+ζb
for X, Y . The following simple observation is crucial to
our proof.
possibilities for Y = y such that |y| ≤ 2b/3 + ζb. Hence
we can estimate the remaining entropy of Y as follows:
b/2
H(Y |X, E) ≤ Pr(|Y | < 2b/3 + ζb)(log
b/6 + ζb
+ log b) + o(1)
n
o( k2 log
n ) expected rounds, and no (ǫ, ǫ)-error protocol
n
that solves Conn correctly in o( k2 log
n ) rounds.
To prove Theorem 3.1, we introduce a new
model called random-partition (two-party) communication complexity and prove some lower bounds in this
≤ H(1/3 + 2ζ)b/2 + o(b)
model. This is done in Section 3.1. We then use these
for the binary entropy function H. For k ≥ 7 we lower bounds to show lower bounds for the k-machine
can simply use the upper bound b/2 for the above model in Sections 3.1.1 and 3.1.2.
quantity and conclude that I(T0 : Y |A) = H(Y |A) −
H(Y |A, T0 ) ≥ 2b/3 − (1 + ζ)b/k − o(1) − b/2 ≥ Ω(b), and 3.1 Random-Partition Communication Comhence p1 must have received messages of length Ω(b). plexity We first recall the standard communication
This happens with some probability γ for p1 and with complexity model, which we will call worst-partition
probability (1 − ǫ) − γ for p2 . Hence |T0 | + |T1 | ≥ Ω(b). model, to distinguish it from our random-partition
The above analysis is under the assumption that model. (For a comprehensive review of the subject, we
different machines hold u, w, which happens with proba- refer the reader to [22].)
In the worst-partition communication complexity
bility 1 − 1/k. Without this assumption the information
model,
there are two players called Alice and Bob.
flow must be at least (1 − 1/k) · Ω(b).
Each
player
receives a b-bit binary string, for some
For k < 7 we need to make a more careful analysis.
integer
b
≥
1. We denote the string received by
For instance, p1 actually gets only around 2b/(3k) bits
Alice
and
Bob
by x and y respectively. Together, they
of information from knowing b/k bits of Y , and the
both
want
to
compute
f (x, y) for a Boolean function
estimate on H(Y |A, E) needs to be made more precise.
b
b
f
:
{0,
1}
×
{0,
1}
→
{0, 1}. At the end of the
We skip the details.
This completes the proof of
n
k
process,
we
want
both
Alice
and Bob to know the value
Tǫ (STC) = Ω k .
of
f
(x,
y).
We
are
interested
in the number of bits
To
see
that
this
also
implies
n
k
k
k
exchanged
between
Alice
and
Bob
in order to compute
Tǫ (MST), Tǫ (BFS), Tǫ (SPT) ∈ Ω k , it is suffif
.
We
say
that
a
protocol
R
has
complexity t if it
cient to observe that any BFS tree (resp. MST, and
4
always
uses
at
most
t
bits
in
total
.
For any function
SPT, for any approximation ratio) is also a spanning
f
,
the
worst-partition
communication
complexity for
tree.
(f ),
computing f with (ǫ0 , ǫ1 )-error, denoted by Rǫcc−pub
0 ,ǫ1
is
the
minimum
t
such
that
there
is
an
(ǫ
,
ǫ
)-error
0
1
3 Lower Bounds for Verification Problems
protocol with complexity t. (Note that a protocol is
In this section, we show lower bounds for the spanning
(ǫ0 , ǫ1 )-error if it outputs 0 with probability at least
tree (ST) and connectivity (Conn) verification problems.
1 − ǫ0 when f (x, y) = 0 and outputs 1 with probability
An algorithm solves the Conn verification problem in
at least 1 − ǫ1 when f (x, y) = 1.)
our model if the machines output 1 if and only if the
The random-partition model is slightly different
input graph G is connected; the ST problem is defined
from the worst-partition model in that, instead of giving
similarly. We note that our lower bounds hold even
every bit of x to Alice and every bit of y to Bob, each
when we allow shared randomness, i.e. even when all
of these bits are sent to one of the players randomly.
machines can read the same random string. For a
To be precise, let xi be the ith bit of x and yi be the
problem P where a two-sided error is possible, we define th
i bit of y. For any pair of input strings (x, y), we
the time complexity of solving P with (ǫ0 , ǫ1 ) error
partition it by telling the value of each xi and yi (by
k
probabilities, denoted by Tǫ0 ,ǫ1 (P), be the minimum
sending a message of the form “xi = 0” or “xi = 1”)
T (n) such that there exists a protocol that solves
to a random player. As before, we say that a protocol
P, terminates in T (n) rounds, and errs on 0-input
R has complexity t if it always uses at most t bits in
with probability at most ǫ0 and errs on 1-input with
total, regardless of the input and its partition. We note
probability at most ǫ1 .
that the error probability of a protocol R is calculated
over all possible random choices made by the algorithm
Theorem 3.1. (ST verification and Conn) There
exists a constant ǫ > 0 such that,
for any kk ≥ 2 and
k
and Tǫ,ǫ (Conn) =
large enough n, Tǫ,0
(ST) = Ω̃ kn2
Ω̃ kn2 . In other words, there is no public-coin (ǫ, 0)error randomized protocol on a k-machine model that,
on any n-node input graph, solves ST correctly in
4 We
emphasize that we allow R to incur at most t bits of
communication regardless of the input and random choices made
by the protocol. We note a standard fact that one can also define
the complexity t to be the expected number of bits. The two
notions are equivalent up to a constant factor.
and all possible random partitions; e.g., it is possible
that an (ǫ0 , ǫ1 )-error protocol never answers a correct
answer for some input partition. Also note that, while
we pick the input partition randomly, the input pair
itself is picked adversarially. In other words, an (ǫ0 , ǫ1 )error protocol must, for any input (x, y), output 0 with
probability at least 1 − ǫ0 when f (x, y) = 0 and output
1 with probability at least 1 − ǫ1 when f (x, y) = 1,
where the probability is over all possible random strings
given to the protocol and the random partition. For
any function f , the random-partition communication
complexity for computing f with (ǫ0 , ǫ1 )-error, denoted
(f ), is the minimum t such that there is an
by Rǫrcc−pub
0 ,ǫ1
(ǫ0 , ǫ1 )-error protocol with complexity t.
The problems of our interest are equality and disjointness. In the rest of Section 3.1, we show that the
communication complexity of these problems are essentially the same in both worst-partition and randompartition models. The techniques used to prove these results are different between the two problems, and might
be of an independent interest.
3.1.1 ST Verification and Random-Partition
Communication Complexity of Eq. The equality
function, denoted by eq, is defined as eq(x, y) = 1
if x = y and eq(x, y) = 0 otherwise. Note that
this problem can be solved by the fingerprinting technique which makes a small error only when x 6= y, i.e.
cc−pub
Rǫ,0
(eq) = O(log b) (see, e.g. [22]). Interestingly, if
we “switch” the error side, the problem becomes hard:
cc−pub
R0,ǫ
(eq) = Ω(b). We show that this phenomenon remains true in the random-partition setting. Lemma 3.1
is proved in Appendix B.
Lemma 3.1. (Random-Partition Equality) For
rcc−pub
some ǫ > 0, R0,ǫ
(eq) = Ω(b). This lower bound
holds even when Alice knows x and Bob knows y.
Lower bound for ST verification. We now show a
lower bound of Ω̃(n/k 2 ) on (0, ǫ)-error algorithms for
ST verification. For the b-bit string inputs x and y of
the equality problem, we construct the following graph
G(x, y): The nodes are u0 , . . . , ub and v0 , . . . , vb . For
any i, there is an edge between u0 and ui if and only if
xi = 1, and there is an edge between v0 and vi if and
only if yi = 0. Additionally, there is always an edge
between uj and vj , for 0 ≤ j ≤ b. Observe that G(x, y)
is a spanning tree if and only if x = y. Also note that
G(x, y) has n = Θ(b) nodes.
Now assume that there is a (0, ǫ)-error algorithm R
in the k-machine model that finishes in õ(n/k 2 ) rounds.
Alice and Bob simulate R as follows. Let p1 , . . . , pk
be machines in the k-machine model, and assume that
k is even. First, Alice and Bob generate a random
partition of nodes in G(x, y) using the random partition
of input (x, y) and shared randomness. Using shared
randomness, they decide which machine the nodes u0
and v0 should belong to. Without loss of generality,
we can assume that u0 belongs to p1 . Moreover, with
probability 1 − 1/k, v0 is not in p1 , and we can assume
that v0 is in p2 without loss of generality. (If u0 and
v0 are in the same machine then Alice and Bob stop
the simulation and output 0 (i.e. x 6= y).) Alice will
simulate machines in PA = {p1 , p3 , p5 . . . , pk−1 } and
Bob will simulate machines in PB = {p2 , p4 , . . . , pk }.
This means that Alice (Bob respectively) can put and
get any information from PA (PB respectively) with no
cost. At this point, Alice assigns node u0 on p1 ; i.e., she
tells p1 whether ui has an edge to u0 or not, for all i (this
can be done since she knows x). Similarly, Bob assigns
node v0 on p2 . (Note that to do this we need Alice to
know x and Bob to know y in addition to the random
partition of x and y. We have the lower bound of this as
claimed in Lemma 3.1.) Next, they randomly put every
node in a random machine. For any i, if Alice gets xi ,
then she assigns ui in a random machine in PA , i.e., she
tells such a random machine whether ui has an edge to
u0 or not. Otherwise, Bob puts ui in a random machine
in PB in the same way. Since each xi and yi belongs to
Alice with probability 1/2, it can be seen that each ui
and vi will be assigned to a random machine. Similarly,
node vi is assigned to a random machine depending on
who gets yi . Note that both Alice and Bob know which
machine each node is assigned to since they use shared
randomness.
Now Alice and Bob simulate R where Alice simulates R on machines in PA and Bob simulates R on
machines in PB . To keep this simulation going, they
have to send messages to each other every time there is
a communication between machines in PA and PB . This
means that they have to communicate Õ(k 2 ) bits in each
round of R. Since R finishes in õ(n/k 2 ) rounds, Alice
and Bob have to communicate õ(n) = õ(b) bits. Once
they know whether G(x, y) is an ST or not, they can answer whether x = y or not. Since R is (0, ǫ)-error, Alice
and Bob’s error will be (0, ǫ + 1/k), where the extra 1/k
term is because they answer 0 when u0 and v0 are in the
same machine. For large enough k, this error is smaller
than the error in Lemma 3.1, contradicting Lemma 3.1.
This implies that such an algorithm R does not exist.
3.1.2 Conn and Random-Partition Communication Complexity of Disjointness The disjointness
function, denoted by disj, is defined as disj(x, y) = 1
if there is i such that xi = yi and disj(x, y) = 0 otherwise. This problem in the worst-partition model is
a fundamental problem in communication complexity,
having tons of application (e.g. [9]). Through a series of results (e.g. [2, 4, 7, 18, 36]), it is known that
cc−pub
Rǫ,ǫ
(disj) = Ω(b). By adapting the previous proof
of Razborov [36], we show that this lower bound remains true in the random-partition setting. Lemma 3.2
is proved in Appendix A.
machine each node is assigned to since they use shared
randomness.
Now Alice and Bob simulate R where Alice simulates R on machines in PA and Bob simulates R on
machines in PB . To keep this simulation going, they
have to send messages between each other every time
there is a communication between machines in PA and
PB . This means that they have to communicate Õ(k 2 )
Lemma 3.2. (Random-Partition Disjointness)
bits in each round of R. Since R finishes in õ(n/k 2 )
rcc−pub
For some ǫ > 0, Rǫ,ǫ
(disj) = Ω(b). This lower
rounds, Alice and Bob have to communicate õ(n) = õ(b)
bound holds even when Alice knows x and Bob knows y.
bits. Once they know whether G(x, y) is connected or
not, they can answer whether x and y are disjoint or
Lower bound for Conn verification. We now show
not. Since R is (ǫ, ǫ)-error, Alice and Bob’s error will
a lower bound of Ω̃(n/k 2 ) on (ǫ, ǫ)-error algorithms for
be (ǫ, ǫ+ 1/k), where the extra 1/k term is because they
Conn verification, for a small enough constant ǫ > 0.
answer 0 when u0 and v0 are in the same machine. For
For the b-bit string inputs x and y of the disjointness
large enough k, this error is smaller than the error in
problem, we construct the following graph G(x, y): The
Lemma 3.2, contradicting Lemma 3.2. This implies that
nodes are u0 , . . . , ub and v0 , . . . , vb . For any i, there
such an algorithm R does not exist.
is an edge between u0 and ui if and only if xi = 0,
and there is an edge between v0 and vi if and only if
4 Algorithms and Techniques
yi = 0. Additionally, there is always an edge between
uj and vj , for 0 ≤ j ≤ b. Observe that G(x, y) is 4.1 A Mapping Lemma We first prove a “mapconnected if and only if x and y are disjoint. Also ping” lemma for the random vertex partition model,
note that G(x, y) has n = Θ(b) nodes. Assume that which we will use in our Conversion theorem (cf. Thethere is an (ǫ, ǫ)-error algorithm R in the k-machine orem 4.1). Consider an input graph G = (V, E) that is
model that finishes in õ(n/k 2 ) rounds. Alice and Bob partitioned (according to the random vertex partition
simulate R as follows. Let p1 , . . . , pk be machines in model) among the k machines in N = {p1 , . . . , pk } of
the k-machine model, and assume that k is even. First, the network. Let |V | = n and |E| = m. We will assume
Alice and Bob generate a random partition of nodes (without loss of generality) throughout that n ≥ k. We
in G(x, y) using the random partition of input (x, y) say that a vertex v of G is mapped to a machine h of
and shared randomness. Using shared randomness, they N if v is assigned to N , i.e., h is the home machine of
decide which machine the nodes u0 and v0 should belong v (cf. Section 1.1). We say that an edge e = (u, v) of G
to. Without loss of generality, we can assume that u0 is mapped to a link (pi , pj ) of the k-machine network, if
belongs to p1 . Moreover, with probability 1 − 1/k, v0 is u is mapped to pi and v is mapped to vj or vice versa.
not in p1 , and we can assume that v0 is in p2 without loss The following Mapping Lemma gives a concentration
of generality. (If u0 and v0 are in the same machine then bound on the number of edges mapped to any link of
Alice and Bob stop the simulation and output 0.) Alice the network.
will simulate machines in PA = {p1 , p3 , p5 . . . , pk−1 } and
Bob will simulate machines in PB = {p2 , p4 , . . . , pk }.
This means that Alice (Bob respectively) can put and
get any information from PA (PB respectively) with no
cost. At this point, Alice assigns node u0 on p1 ; i.e.,
she tells p1 whether ui has an edge to u0 or not, for all
i (this can be done since she knows x). Similarly, Bob
assigns node v0 on p2 . Next, they randomly put every
node in a random machine. For any i, if Alice gets xi ,
then she assigns ui in a random machine in PA , i.e., she
tells such a random machine whether ui has an edge to
u0 or not. Otherwise, Bob puts ui in a random machine
in PB in the same way. Since each xi and yi belongs to
Alice with probability 1/2, it can be seen that each ui
and vi will be assigned to a random machine. Similarly,
node vi is assigned to a random machine depending on
who gets yi . Note that both Alice and Bob knows which
Lemma 4.1. (Mapping Lemma) Let an n-node, medge graph G be partitioned among the k machines
in N = {p1 , . . . , pk }, according to the random vertex
partition model (assume n ≥ k). Then with probability
at least 1 − 1/nα , where α > 1 is an arbitrary fixed
constant, the following bounds hold:
(1) The number of vertices of G mapped to any machine
is Õ(n/k).
(2) The number of edges of G mapped to any link of
the network is Õ(m/k 2 + ∆/k), where ∆ is the
maximum node degree of G.
Proof. (1) This follows easily from a direct Chernoff
bound application. Since each vertex of G is mapped
independently and uniformly to the set of k machines,
the expected number of vertices mapped to a machine
is n/k. The concentration follows from a standard mapped to link ℓ; the total number of edges mapped pto
X
⌉
Chernoff bound [14].
ℓ is bounded by Zℓp + Zℓq . We have, E[Zℓp |X p ] ≤ ⌈ k−1
q
q
p
X
q
(2) We first note that we cannot directly apply a and similarly E[Zℓ |X ] ≤ ⌈ k−1 ⌉. We next show that Zℓ
q
2
Chernoff bound to show concentration on the number and Zℓ are both bounded by Õ(m/k +∆/k) whp which
p
of edges mapped, as these are not independently dis- will complete the proof. We focus on Zℓ ; (the case of
q
p
Zℓ is similar). By proof of (a), X < Φ with probability
tributed. We show the bound in two steps:
at
least 1 − 1/n2α , where Φ = c log n(m/k + ∆) for some
1. (a) We first show a concentration bound on the
total degree of the vertices assigned to any machine; sufficiently large constant c > 0, depending on constant
2. (b) then we bound the number of edges assigned to α. Observe that an edge that has one endpoint mapped
to p is mapped to link ℓ independently. Hence we can
any link.
p
′
To show (a) we use Bernstein’s inequality [14]. Fix apply a standard Chernoff bound [14]: Pr(Zℓ < c Φ/k+
p
p
′
p
p
a machine p. Let random variable Xi be defined as 1) ≤ Pr(X > Φ) + Pr(Zℓ < c Φ/k + 1)|X < Φ) ≤
′
follows: Xip = d(vi ) (d(vi ) is the degree of vi ) if 1/n2α + 2−c Φ/k+1 ≤ 1/n2α + 2−3 log n+Ω(1) = O(1/n2α ),
p
′
vertex vi is assigned
Pn to pmachine p, otherwise Xi = for a sufficiently large constant c > 0.
p
Thus, with probability 1 − 1/n2α , the number of
0. Let X =
i=1 Xi denote the total degree of
the vertices assigned to machine p; in other words, it edges mapped to link ℓ is c′ Φ/k = Õ(m/k 2 + ∆/k).
is the total number of edges that have at least one Applying a union bound over all k(k − 1)/2 links, yields
endvertex assigned to P
machine p. WeP
have E[Xip ] = the result.
n
n
p
d(vi )/k and E[X p ] =
i=1 d(vi )/k =
i=1 E[Xi ] =
p
p 2
2m/k. Furthermore, V ar(Xi ) = E[(Xi ) ] − E[Xip ]2 = 4.2 The Conversion Theorem We now present a
2
1
2
general conversion theorem that enables us to leverage
( d(vk i ) )2 = (d(vki )) (1 − 1/k)
k (d(vi )) − P
Pn and hence
n
p
1
1
2
p
V ar(X ) P
= i=1 V ar(Xi ) = k (1 − k ) i=1 (d(vi )) ≤ results from the standard message-passing model [33].
n
1
1
1
1
Our conversion theorem allows us to use distributed
i=1 ∆d(vi ) = k (1 − k )∆m.
k (1 − k )
Using Bernstein’s inequality, we have (for some algorithms that leverage direct communication between
nodes, even when such an edge is not part of the
t > 0):
input graph. More specifically, we can translate any
t2
distributed algorithm that works in the following clique
Pr(X p > E[X p ] + t) ≤ e− 2V ar(X p )+(2/3)bt)
model to the k-machine model.
where b = max1≤i≤n |Xip − E[Xip ]|. Now, |Xip −
E[Xip ]| ≤ d(vi )(1 − 1/k) ≤ ∆(1 − 1/k) = b.
Let γ > 0 and let A be the event that X p >
2m/k + γ(2m/k + ∆). Hence, for any γ > 0 and letting
t = γ(2m/k + ∆), we have:
t2
t2
Pr(A) ≤ e− 2(1/k)(1−1/k)∆m+(2/3)∆(1−1/k)t ≤ e− Θ(m∆/k+∆t)
γ 2 (2m/k+∆)2
≤e
− Θ(m∆/k+∆2 )
≤ e−γ
2
m
Θ( ∆k
+ ∆k
m +1)
= O(1/n3α ),
√
if γ = Θ(α log n).
The above tail bound applies to a single machine
p; applying a union bound over all the k machines, we
have X p = Õ(m/k + ∆) whp for every machine p ∈ N .
We now show (b) which will complete the proof
of Part 2. Fix a link ℓ = (p, q) of the k-machine
network. Let X p (resp. X q ) be defined as before, i.e.,
the total number of edges of G with at least one endvertex assigned to p (resp. q). Note that each such
edge mapped to p has probability of 1/(k − 1) of being
mapped to link ℓ (independently of other edges mapped
to p). A similar statement holds for edges mapped to
q as well. Let r.v. Zℓp (resp. Zℓq ) denote the number
of edges, among those mapped to p (resp. q), that are
The Clique Model. Consider a complete n-node
network C and a spanning subgraph G of C determined
by a set of (possibly weighted) edges E(G). The nodes
of C execute a distributed algorithm and each node
u is aware of the edges that are incident to u in G.
Each node can send a message of at most W ≥ 1
bits over each incident link per round. For a graph
problem P , we are interested in distributed algorithms
that run on the network C and, given input graph G,
compute a feasible solution of P . In addition to time
complexity (the number of rounds in the worst case), we
are interested in the message complexity of an algorithm
in this model which is the number of messages (in the
worst case) sent over all links. Additionally, we are also
interested in communication degree complexity which is
the maximum number of messages sent or received by
any node in any round; i.e., it is the minimum integer
M ′ such that every node sends a message to at most
M ′ other nodes in each round. Note that we can
simulate any “classic” distributed algorithm running
on a network G of an arbitrary topology that uses
messages of O(log n) size in the clique model by simply
restricting the communication to edges in E(G) ⊂ E(C)
and by splitting messages into packets of size W . In
Let Gi be the graph whose node set is the same as the
input graph (as well as the clique model), and there is
an edge between nodes u and v if and only if a message
is sent between u and v in round i of the algorithm;
in other words, Gi captures the communications happening in round i. From Lemma 4.1(2), we know that
(w.h.p.) each communication link of N is mapped to
at most Õ(|E(Gi )|/k 2 + ∆i /k) edges of Gi , where ∆i
is the maximum degree of Gi . This means that each
machine needs to send at most Õ(|E(Gi )|/k 2 + ∆i /k)
Theorem 4.1. (Conversion Theorem) Suppose
messages over a specific communication link with high
that there is an ε-error algorithm AC that solves probability. In other words, the ith round of AC can be
problem P in time TC (n) ∈ Õ(n) in the clique model, simulated in Õ(|E(Gi )|/k 2 W +∆i /kW ) rounds, and, by
for any n-node input graph. Then there exists an taking a union bound, the same is true for all rounds in
ε-error algorithm A that solves P in the k-machine [1, TC (n)]. By summing up over all rounds of AC , we
model with bandwidth W satisfying the following time can conclude that the number rounds needed to simucomplexity bounds with high probability:
late AC is
(a) If AC uses point-to-point communication with mes
′
TC (n)
X
|E(Gi )|
∆i
M
∆
sage complexity M and communication
degree com
Õ
=
Õ
+
+
T
(n)
C
∆′
2W
2W
k
kW
k
kW
+
T
(n)⌈
⌉
plexity ∆′ , then A runs in Õ kM
C
2W
kW
i=1
time.
the equality is because of the following facts: (1)
(b) If AC is a broadcast algorithm with broadcast com- where
PTC (n)
B
|E(Gi )| = O(M ) since |E(Gi )| is at most two
plexity B, then A takes Õ( kW + TC (n)) time.
i=1
times the number of messages sent by all nodes in the
Proof. Consider any n-node input graph G with m ith round, and (2) ∆i ≤ ∆′ . This proves (a).
edges and suppose that nodes in G are assigned to the Proof of (b): We first slightly modify the previous
k machines of the network N according to the vertex simulation to simulate broadcast algorithms: Note that
partitioning process (cf. Section 1.1).
if AC is a broadcast algorithm, then for the ith round
We now describe how to obtain algorithm A for the (i ≥ 1) of algorithm AC , if a node u belonging to
k-machine model from the clique model algorithm AC : machine p1 sends messages to nodes v1 , . . . , vj (j ≥ 1)
Each machine locally simulates the execution of AC at belonging to machine p2 , we know that u sends the
each hosted vertex. First of all, we only need to consider same message to v1 , . . . , vj . Thus, when we simulate
inter-machine communication, since local computation this round AC , we will let machine p1 send only one
at each machine happens instantaneously at zero cost. message to p2 , instead of j messages. Then, machine
If algorithm AC requires a message to be sent from a p2 will pretend that this message is sent from u1 to
node u1 ∈ C hosted at machine p1 to some node u2 ∈ C all nodes belonging to p2 that have an edge to node u.
hosted at p2 , then p1 sends this message directly to p2 (We cannot specify the destination nodes v1 , . . . , vj in
via the links of the network N . (Recall that a machine this message as this might increase the length of the
p1 knows the hosting machines of all endpoints of all message significantly.)
edges (in G) that are incident to a node hosted at p1 .)
We now analyze this new simulation. We show that
Moreover, p1 adds a header containing the IDs of u1 and this simulation finishes in Õ( B + TC (n)) rounds. Let
k
u2 to ensure that p2 can correctly deliver the message Bi be the number of nodes that perform a broadcast
to the simulation of AC at u2 . Each message is split in round i of the run of AC in the clique model, and
into packets of size W , which means that sending all note that B = PTC (n) Bi . According to Lemma 4.1(a),
i=1
packages that correspond to such a message requires the number of nodes contributing to Bi broadcasts
⌈O(log n)/W ⌉ rounds. In the worst case (i.e. W = 1), that are assigned to a single machine is Õ(⌈Bi /k⌉)
this requires additional O(log n) rounds, which does not w.h.p.; in other words, w.h.p., each machine contains
change our complexity bounds. Thus, for the remainder ℓi = Õ(⌈Bi /k⌉) of the Bi nodes. Thus, for every i, we
of the proof, we assume that W is large enough such that instruct algorithm A to simulate these Bi broadcasts in
any message generated by AC can be sent in 1 round in the k-machine model in ⌈ℓi /W ⌉ rounds. Since AC takes
the k-machine model.
at most TC (n) rounds, we can take a union bound, and
B
Proof of (a): We will bound the number of messages sent it follows that algorithm A takes Õ( kW
+TC (n)) rounds
in each round through each link using Lemma 4.1(2). in the k-machine model.
this case, the time and message complexities remain
the same (up to log-factors) while the communication
degree complexity can be bounded by the maximum
degree of G. We say that an algorithm is a broadcast
algorithm if, in every round and for every node u, it
holds that u broadcasts the same message to other nodes
(or remains silent). We define the broadcast complexity
of an algorithm as the number of times nodes broadcast
messages.
It is easy to see that a simulation similar to the one
employed in the proof of Theorem 4.1 provides the same
complexity bounds, if we limit the total communication
of each machine (i.e. bits sent/received) to at most kW
bits per round, instead of restricting the bandwidth
of individual inter-machine links to W bits. To see
why this is true, observe that throughout the above
simulation, each machine is required to send/receive at
most kW bits in total per simulated round with high
probability.
4.3 Algorithms We now consider various important
graph problems in the k-machine model. For the sake
of readability, we assume a bandwidth of Θ(log n) bits,
i.e., parameter W = Θ(log n). Observe that the simple
solution of aggregating the entire information about the
input graph G at a single machine takes O(m/k) rounds;
thus we are only interested in algorithms that beat this
trivial upper bound. Our results are summarized in
Table 1.
Breadth-First Search Tree (BFS). To get an intuition for the different bounds obtained by applying either Theorem 4.1(a) or Theorem 4.1(b) to an algorithm
in the clique model, consider the problem of computing a breadth-first search (BFS) tree rooted at a fixed
source node. If we use Theorem 4.1(a) we get a bound of
Õ(m/k 2 + D⌈∆/k⌉) rounds. In contrast, recalling that
each node performs O(1) broadcasts, Theorem 4.1(b)
k
(BFS) ∈ Õ(n/k + D). We will leverage these
yields T1/n
bounds when considering graph connectivity and spanning tree verification below.
Minimum Spanning Tree (MST), Spanning Tree
Verification (ST) and Graph Connectivity (Conn).
An efficient algorithm for computing the MST of an
input graph was given by [16], which proceeds by
merging “MST-fragments” in parallel; initially each
vertex forms a fragment by itself. In each of the O(log n)
phases, each fragment computes the minimum outgoing
edge (pointing to another fragment) and tries to merge
with the respective fragment. Since any MST has
n − 1 edges, at most n − 1 edges need to be added in
total. This yields a total broadcast complexity of Õ(n)
and thus Theorem 4.1(b) readily implies the bound of
Õ(n/k). We can use an MST algorithm for verifying
graph connectivity which in turn can be used for ST. We
assign weight 1 to all edges of the input graph G and
then add an edge with infinite weight between any pair
of nodes u, v where (u, v) ∈
/ E(G), yielding a modified
graph G′ . Clearly, G is disconnected iff an MST of
G′ contains an edge with infinite weight. This yields
the first part of the upper bound for graph connectivity
stated in Table 1. We now describe how to verify
whether an edge set S is an ST, by employing a given
algorithm A for Conn. Note that, for ST verification,
each machine p initially knows the assumed status of
the edges incident to its nodes wrt. being part of the
ST, and eventually p has to output either yes or no.
First, we run A on the graph induced by S and then
we compute the size of S as follows: Each machine
locally adds 1 to its count for each edge (u, v) ∈ S,
if p is the home machine for vertices u, v. Otherwise,
if one of u or v reside on a different machine, then p
adds 1/2. Then, all machines exchange their counts
via broadcast, which takes 1 round (since each count is
at most n and W ∈ Θ(log n)) and determine the final
count by summing up over all received counts including
their own. Each machine outputs yes iff (1) the output
of the Conn algorithm A returned yes and (2) the final
count is n − 1. Thus we get the same bounds for ST
verification as for graph connectivity.
Recalling that we can compute a BFS in Õ(m/k 2 +
D⌈∆/k⌉) rounds, it is straightforward to see that the
same bound holds for Conn (and thus also ST verification): First, we run a leader election algorithm among
the√k machines. This can be done in O(1) rounds (and
Õ( k) messages) by using the algorithm of [23] (whp).
The designated leader machine then chooses an arbitrary node s as the source node and executes a BFS
algorithm. Once this algorithm has terminated, each
machine locally computes the number of its vertices that
are part of the BFS and then computes the total number
of vertices in the BFS by exchanging its count (similarly
to the ST verification above). The input graph is connected iff the BFS contains all vertices.
PageRank. The PageRank problem is to compute the
PageRank distribution of a given graph (may be directed or undirected). A distributed page rank algorithm was presented in [10], based on the distributed
random walk algorithm: Initially, each node generates
Θ(log n) random walk tokens. A node forwards each token with probability 1−δ and terminates the token with
probability δ (called the reset probability). Clearly, every token will take at most O(log n/δ) steps with high
probability before being terminated. From Lemma 2.2
of [11] we know that these steps can be implemented
in O(log2 n) rounds in the clique model, and since this
requires O(n log2 n/δ) messages to be sent in total, Theorem 4.1(a) yields that, for any δ > 0, there is a randomized algorithm for computing PageRank in the kn
k
).
machine model such that T1/n
(PageRank) ∈ Õ( δk
Computing a (2δ − 1)-Spanner. The algorithm of [5]
computes a (2δ − 1)-spanner, for some (δ ∈ O(log n)),
of the input graph G in δ 2 rounds (using messages of
O(log n) size) that has an expected number of δm1+1/δ
edges (cf. Theorem 5.1 in [5]). That is, each node needs
to broadcast δ 2 times, for a total broadcast complexity
of nδ 2 . Applying Theorem 4.1(b) yields a bound of
Õ(n/k) rounds in the k machine model.
Single-Source Shortest Paths (SSSP, SPT), and
All-Pairs Shortest Paths(APSP). We show that,
in the k-machine model, SSSP
√ and APSP can be
(1 + ǫ)-approximated√in Õ(n/ k) time, and (2 + ǫ)approximated in Õ(n n/k) time, respectively.
Recall that, for SSSP, we need to compute the distance between each node and a√designated source node.
1/4
Nanongkai [31] presented a Õ(
√ nD )-time algorithm
for SSSP, which implies a Õ( n)-time algorithm in the
clique model. We show that the ideas
√ in [31], along
with Theorem 4.1(b), leads to a Õ(n/ k)-time (1 + ǫ)approximation algorithm in the k-machine model. We
sketch the algorithm in [31] here5 . First, every node
broadcasts ρ edges incident to it of minimum weight
(breaking tie arbitrarily), for some parameter ρ which
will be fixed later. Using this information, every node
internally compute Õ(1) integral weight functions (without communication). For each of these weight functions, we compute a BFS tree of depth n/ρ from the
source node, treating an edge of weight w as a path
of length w. Using two techniques called light-weight
SSSP and shortest-path diameter reduction, the algorithm of [31] gives a (1 + ǫ)-approximation solution.
Observe that this algorithm uses broadcast communication. Its time complexity
√ is clearly TC (n) = Õ(ρ + n/ρ).
n, we have the running time of
(Thus,
by
setting
ρ
=
√
Õ( n) in the clique model.) Its broadcast complexity is B = Õ(nρ) since every node has to broadcast ρ
edges in the first step and the BFS tree algorithm has
O(n) broadcast complexity. Its message complexity is
M = Õ(n2 ρ) (the BFS tree algorithm has O(n2 ) message complexity since a message will be sent through
each edge once). By Theorem 4.1(b), we have that in
the k-machine model, the time√we need to solve SSSP
is Õ( nρ
+ n/ρ). Using ρ = k, we have the running
k + ρ√
√
√
time of Õ( k + n/ k) = Õ(n/ k) where the equality
is because k ≤ n.
In [31], a similar √
idea was also used to obtain a
(2 + ǫ)-approximation n-time algorithm for APSP on
the clique model. This algorithm is almost identical to
the above algorithm except that it creates BFS trees of
depth n/ρ from n/ρ centers instead of just the source.
By this modification, it can be shown that the running
time remains the same,
√ i.e. TC (n) = Õ(ρ + n/ρ).
n, we have the running time of
(Thus,
by
setting
ρ
=
√
Õ( n) in the clique model.) The broadcast complexity
5 The algorithm is in fact a simplification of the algorithm in
[31] since we only have to deal with the clique model
becomes B = Õ(nρ + n2 /ρ) since each BFS tree
algorithm has O(n) broadcast complexity. Its message
complexity becomes M = Õ(n2 ρ + n3 /ρ) since each
BFS tree algorithm has O(n2 ) message complexity. By
Theorem 4.1(b), we have that in the k-machine model,
2
/ρ
the time we need to solve APSP is Õ( nρ+n
+ρ+n/ρ).
k
√
√
Using ρ = n, we have the running time of Õ( n k n ).
Since the algorithm of [31] also constructs a shortest
path tree while computing the shortest path distances,
we get analogous bounds for the SPT problem, which
requires each machine to know which of its edges are
part of the shortest path tree to the designated source.
We can leverage the technique of computing a
(2δ − 1)-spanner in Õ(n/k) rounds that has an expected
number of Õ(m1+1/δ ) edges. We can simply collect
all edges at one designated machine p, which takes
time Õ(m1+1/δ /k) and then locally compute a (2δ −
1)-approximation for the shortest path problems at
machine p. In particular, for δ = Θ(log n), we have
a spanner of (expected) O(n) edges and thus we get a
O(log n)-approximation for the shortest path problems
in expected Õ(n/k) rounds.
For computing the exact APSP (resp. SSSP) problems, we can use a distributed algorithm by BellmanFord ([33, 28]). This algorithm takes S rounds, where
S is the shortest path diameter, and thus the broadcast
complexity is nS. By virtue of Theorem 4.1(b), we get
a round complexity of Õ(nS/k + S) in the k-machine
model.
Densest Subgraph. We show that Theorem 4.1
implies that the densest subgraph problem can be
approximated in Õ(min( km2 , nk )) time in the k-machine
model. In [38], Das Sarma et al. presented a (2 +
ǫ)-approximation algorithm for the densest subgraph
problem. The idea is very simple: In every round, they
compute the average degree of the network and delete
nodes of degree less than (1 + ǫ) times of the average
degree. This process generates several subgraphs of
the original network, and the algorithm output the
densest subgraph among the generated subgraphs. It
was shown in [38] that this algorithm produces a (2+ǫ)approximate solution. They also proved that this
algorithm stops after log1+ǫ n rounds, implying a time
complexity of TC (n) = O(log n) in the clique model.
The message complexity is M = Õ(m) since every node
has to announce to its neighbors in the input graph
that it is deleted at some point. For the same reason,
the broadcast complexity is B = Õ(n). Note that this
algorithm is broadcast. So, By Theorem 4.1, we have
that in the k-machine model, the time we need to solve
this problem is Õ( nk ) as desired.
Cut Sparsifier, Min Cut, Sparsest Cut, etc.. An
ǫ-cut-sparsification of a graph G is a graph G′ on the
same set of nodes such that every cut in G′ is within
(1 ± ǫ) of the corresponding cut in G. We show that
we can use the technique called refinement sampling
of Goel et al. [17], to compute an ǫ-sparsification of
Õ(n) edges in Õ(n/k) time in the k-machine model.
By aggregating this sparsification to a single machine,
we can approximately solve cut-related problems such
as a (1 ± ǫ)-approximate solution to the minimum cut
and sparsest cut problems. The main component of the
algorithm of Goel et al. is repetitively sparsify the graph
by keeping each edge with probability 2−ℓ for some ℓ,
and compute the connected components after every time
we sparsify the graph. By doing this process for Õ(1)
times, we can compute a probability z(e) on each edge e.
Goel et al. showed that we can use this probability z(e)
to (locally) sample edges and assign weights to them to
obtain an ǫ-sparsification. It is thus enough to be able
to compute the connected components quickly in the
k-machine model. This can be done by simply invoke
the MST algorithm, which takes Õ(n/k) rounds. We
have to runs this algorithm for Õ(1) times, so the total
running time is Õ(n/k).
Covering Problems on Graphs. We now describe
how to solve covering problems like maximal independent set (MIS) in our model. We first consider MIS
and related covering problems on simple graphs, and
then describe how to obtain an MIS on an input hypergraph. A well known distributed algorithm for computing a maximal independent set (MIS) is due to [27]: The
algorithm proceeds in phases and in each phase, every
active node v—initially every node—marks itself with
probability 1/2dv where dv is the degree of v. If v turns
out to be the only marked node in its neighborhood,
v enters the MIS, notifies all of its neighbors who no
longer participate (i.e. become inactive) in future phases
and terminates. When 2 neighboring nodes both mark
themselves in the same phase, the lower degree node
unmarks itself. Nodes that were not deactivated proceed to the next phase and so forth. It was shown in
[27] that this algorithm terminates in O(log n) rounds
with high probability. Since each node sends the same
messages to all neighbors, we can analyze the communication in terms of broadcasts, yielding a broadcast complexity of O(n log n) (whp). Applying Theorem 4.1(b)
yields a round complexity of Õ(n/k). Alternatively, for
bounded degree graphs, applying Theorem 4.1(a) gives
us a running time of Õ(m/k 2 + ∆/k), which is faster
when ∆ ≪ k. Considering the locality preserving reductions (cf. [21]) between MIS, maximal matching (MM),
minimal dominating set (MDS), and computing a 2approximation of the minimum vertex cover (MVC), we
k
k
k
k
get that T1/n
(MIS), T1/n
(MM), T1/n
(MVC), T1/n
(MDS)
are Õ(min(n/k, m/k 2 + ∆/k)).
We now describe how to obtain an Õ(n/k + k) time
algorithm for all graph covering problems directly in
the k-machine model, without resorting to Theorem√4.1;
note that this translates to Õ(n/k) when k ≤ Õ( n).
In particular, this yields an Õ(n/k + k) algorithm for
the problem of finding a maximal independent set in a
hypergraph (HMIS), which has been studied extensively
in the PRAM model of computation (cf. [20, 6]). Note
that, for the hypergraph setting, the input graph G is
a hypergraph and if some node u has home machine p1 ,
then p1 knows all hyperedges (and the corresponding
machines) that contain u. To the best of our knowledge,
there is no efficient distributed algorithm known for
HMIS. First assume that there is an ordering of the
machines with ids 1, . . . , k. Such an ordering can
be obtained by running the O(1)-time leader election
algorithm of [23]. The elected leader (machine) then
arbitrarily assigns unique ids to all the other machines.
Then, we sequentially process the nodes at each machine
and proceed in k phases. In the first phase, machine 1
locally determine the membership status of being in the
HMIS, for all of its nodes. Next, machine 1 computes an
arbitrary enumeration of its nodes and sends the status
(either 0 or 1) and node id of the first k nodes over its
k links (i.e. a single status is sent over exactly 1 link).
When a machine receives this message from machine 1,
it simply broadcasts this message to all machines in the
next round. Thus, after 2 rounds, all machines know
the status of the first k nodes of machine 1. Then,
machine 1 sends the status of the next k nodes and so
forth. By Lemma 4.1, there will be Õ(n/k) nodes with
high probability, and therefore every machine will know
the status of all the nodes of machine 1 after Õ(n/k 2 )
rounds. After machine 1 has completed sending out all
statuses, all other machines locally use this information
to compute the statuses of their nodes (if possible). For
example, if some node u at machine 1 is in the HMIS (i.e.
has status 1) and adjacent to some node v at machine
2, then machine 2 sets the status of v to 0. Then,
machine 2 locally computes the status of its remaining
undetermined nodes such that they are consistent with
the previously received statuses, and starts sending this
information to all other machines in the same fashion.
Repeating the same process for each of the k machines
yields a total running time of Õ(n/k + k).
Finding Triangles and Subgraphs. For the subgraph isomorphism problem SubIsod , we are given 2
input graphs: the usual n-node graph G and a dvertex graph H, for d ∈ O(1). We want to answer
the question whether H ⊆ G. A distributed algorithm
for the clique model that runs in O(n(d−2)/d ) rounds
was given by [13]. Since the total number of mes-
sages sent per round is O(n2 ), Theorem 4.1(a) gives
rise to an algorithm for the k-machine model that runs
in Õ(n2+(d−2)/d /k 2 + n1+(d−2)/d /k) rounds.
We use Tri to denote the restriction of SubIso3
to the case where H is a triangle. The following is
a simple algorithm for the clique model: Each node
locally collects its 2-neighborhood information, checks
for triangles and then either outputs yes or no. This
requires each node to send a message for each of its
at most ∆ neighbors. The total number of messages
sent is n∆2 , and the algorithm takes ∆ rounds to
send all messages. Applying Theorem 4.1(a), we get
a distributed algorithm in the k-machine model with
round complexity of Õ(n∆2 /k 2 + ∆⌈∆/k⌉), which is
better than the above bound when ∆ is sufficiently
k
small. Thus, we have T1/n
(Tri) ∈ Õ(min(n∆2 /k 2 +
∆⌈∆/k⌉, n7/3 /k 2 + n4/3 /k)).
5
Conclusion
We presented algorithms and lower bounds for distributed computation of several graph problems. Our
bounds are (almost) tight for problems such as computing a ST or a MST, while for other problems such as connectivity and shortest paths, there is a non-trivial gap
between upper and lower bounds. Understanding these
bounds and investigating the best possible can provide
insight into understanding the complexity of distributed
graph processing.
References
[1] Gps:
A
graph
processing
system,
http://infolab.stanford.edu/gps/. 1, 2
[2] L. Babai, P. Frankl, and J. Simon. Complexity classes
in communication complexity theory (preliminary version). In FOCS, pages 337–347, 1986. 9
[3] Nikhil Bansal, Kang-Won Lee, Viswanath Nagarajan,
and Murtaza Zafer. Minimum congestion mapping in
a cloud. In PODC, pages 267–276, 2011. 2
[4] Ziv Bar-Yossef, T. S. Jayram, Ravi Kumar, and
D. Sivakumar. An information statistics approach to
data stream and communication complexity. J. Comput. Syst. Sci., 68(4):702–732, 2004. Also in FOCS’02.
9
[5] Surender Baswana and Sandeep Sen. A simple and linear time randomized algorithm for computing sparse
spanners in weighted graphs. Random Struct. Algorithms, 30(4):532–563, 2007. 12, 13
[6] Paul Beame and Michael Luby. Parallel search for
maximal independence given minimal dependence. In
SODA, pages 212–218, 1990. 14
[7] Mark Braverman, Ankit Garg, Denis Pankratov, and
Omri Weinstein. From information to exact communication. In STOC, pages 151–160, 2013. 9
[8] Amit Chakrabarti, Graham Cormode, and Andrew
McGregor. Robust lower bounds for communication
and stream computation. Electronic Colloquium on
Computational Complexity (ECCC), 18:62, 2011. 3
[9] Arkadev Chattopadhyay and Toniann Pitassi. The
Story of Set Disjointness. SIGACT News, 41(3):59–85,
2010. 9
[10] Atish Das Sarma, Anisur Rahaman Molla, Gopal Pandurangan, and Eli Upfal. Fast distributed pagerank
computation. In ICDCN, pages 11–26, 2013. 12
[11] Atish Das Sarma, Danupon Nanongkai, and Gopal
Pandurangan. Fast distributed random walks. In
PODC, pages 161–170, 2009. 12
[12] Jeffrey Dean and Sanjay Ghemawat. Mapreduce:
Simplified data processing on large clusters. In OSDI,
pages 137–150, 2004. 1
[13] Danny Dolev, Christoph Lenzen, and Shir Peled. ”tri,
tri again”: Finding triangles and small subgraphs in
a distributed setting - (extended abstract). In DISC,
pages 195–209, 2012. 14
[14] Dubhashi and Panconesi. Concentration of Measure
for Analysis of Randomized Algorithms. Cambridge
University Press, 2012. 10
[15] Michael Elkin, Hartmut Klauck, Danupon Nanongkai,
and Gopal Pandurangan. Can quantum communication speed up distributed computation? In PODC,
2014. 5
[16] Robert G. Gallager, Pierre A. Humblet, and Philip M.
Spira. A distributed algorithm for minimum-weight
spanning trees. ACM Trans. Program. Lang. Syst.,
5(1):66–77, 1983. 12
[17] Ashish Goel, Michael Kapralov, and Sanjeev Khanna.
Graph sparsification via refinement sampling. CoRR,
abs/1004.4915, 2010. 14
[18] Bala Kalyanasundaram and Georg Schnitger. The
Probabilistic Communication Complexity of Set Intersection. SIAM J. Discrete Math., 5(4):545–557, 1992.
Also in CCC’87. 9
[19] Howard J. Karloff, Siddharth Suri, and Sergei Vassilvitskii. A model of computation for mapreduce. In
SODA, pages 938–948, 2010. 4
[20] Pierre Kelsen. On the parallel complexity of computing
a maximal independent set in a hypergraph. In STOC,
pages 339–350, 1992. 14
[21] Fabian Kuhn, Thomas Moscibroda, and Roger Wattenhofer. Local computation: Lower and upper bounds.
CoRR, abs/1011.5470, 2010. 14
[22] E. Kushilevitz and N. Nisan. Communication complexity. Cambridge University Press, New York, NY, USA,
1997. 7, 8, 20
[23] Shay Kutten, Gopal Pandurangan, David Peleg, Peter
Robinson, and Amitabh Trehan. Sublinear bounds for
randomized leader election. In ICDCN, pages 348–362,
2013. 12, 14
[24] Silvio Lattanzi, Benjamin Moseley, Siddharth Suri, and
Sergei Vassilvitskii. Filtering: a method for solving
graph problems in mapreduce. In SPAA, pages 85–94,
2011. 4
[25] J. Lin and C. Dyer. Data-Intensive Text Processing
with MapReduce. Morgan Claypool Publishers, 2010.
1, 4
[26] Yucheng Low, Joseph Gonzalez, Aapo Kyrola, Danny
Bickson, Carlos Guestrin, and Joseph M. Hellerstein.
Graphlab: A new framework for parallel machine
learning. In UAI, pages 340–349, 2010. 1
[27] Michael Luby. A simple parallel algorithm for the
maximal independent set problem. SIAM J. Comput.,
15(4):1036–1053, 1986. 14
[28] Nancy Lynch. Distributed Algorithms. Morgan Kaufmann, 1996. 2, 13
[29] Grzegorz Malewicz, Matthew H. Austern, Aart J. C.
Bik, James C. Dehnert, Ilan Horn, Naty Leiser, and
Grzegorz Czajkowski. Pregel: a system for large-scale
graph processing. In SIGMOD Conference, pages 135–
146, 2010. 1, 2
[30] M. Mitzenmacher and E. Upfal. Probability and Computing: Randomized Algorithms and Probabilistic Analysis. Cambridge University Press, 2005. 20
[31] Danupon Nanongkai. Distributed approximation algorithms for weighted shortest paths. In STOC, pages
565–573, 2014. 13
[32] Danupon Nanongkai, Atish Das Sarma, and Gopal
Pandurangan. A tight unconditional lower bound on
distributed randomwalk computation. In PODC, pages
257–266, 2011. 5
[33] David Peleg.
Distributed computing: a localitysensitive approach. Society for Industrial and Applied
Mathematics, Philadelphia, PA, USA, 2000. 2, 3, 4, 5,
10, 13
[34] The
Apache
Project.
Apache
Giraph,
http://giraph.apache.org/. 1, 2
[35] Anand Rajaraman and Jeff Ullman. Mining of Massive
Datasets. Cambridge University Press, October 2011.
2, 4
[36] Alexander A. Razborov.
On the Distributional
Complexity of Disjointness. Theor. Comput. Sci.,
106(2):385–390, 1992. Also in ICALP’90. 3, 9, 16
[37] Atish Das Sarma, Stephan Holzer, Liah Kor, Amos
Korman, Danupon Nanongkai, Gopal Pandurangan,
David Peleg, and Roger Wattenhofer. Distributed
verification and hardness of distributed approximation.
SIAM J. Comput., 41(5):1235–1265, 2012. 5
[38] Atish Das Sarma, Ashwin Lall, Danupon Nanongkai,
and Amitabh Trehan. Dense subgraphs on dynamic
networks. In Distributed Computing - 26th International Symposium, DISC 2012, Salvador, Brazil, October 16-18, 2012. Proceedings, pages 151–165, 2012.
13
[39] Isabelle Stanton. Streaming balanced graph partitioning for random graphs. CoRR, abs/1212.1121, 2012.
2
[40] Isabelle Stanton and Gabriel Kliot. Streaming graph
partitioning for large distributed graphs. In KDD,
pages 1222–1230, 2012. 2
[41] David P. Woodruff and Qin Zhang. When distributed
computation is communication expensive. In DISC,
pages 16–30, 2013. 3, 4, 5
A Proof of Lemma 3.2
Our proof adapts the proof of Razborov [36] from the
situation of a fixed partition to a random partition of
the inputs. In the following we try to keep our notation
close to that of Razborov’s. In particular, we use N
to replace b as elsewhere in this paper. We first recall
the standard definition of deterministic and randomized
protocols here.
Definition 1. The communication cost of a deterministic protocol P computing a function f : X ×Y → {0, 1}
on an input x, y is the number of bits exchanged between
Alice and Bob during P on x, y. The communication
cost of P is the maximum communication cost over all
inputs. The deterministic communication complexity of
f is the minimum communication cost (over all protocols) computing f .
A randomized protocol has to be correct with probability 2/3 (over the random choices used in the protocol) on every input. The randomized communication
complexity is then defined as above.
In the above definition, the inputs x, y are partitioned
in a fixed way. We are interested in the communication
cost over an average partition of the inputs to Alice and
Bob. In this situation the communication cost (for a
fixed input z) is the average over all partitions of the
bits of z to Alice and Bob. More precisely, for every
partition ρ of z there is a protocol Pρ computing f ,
and the communication cost on z is the average over
the communication costs of all the Pρ on z. Here the
distribution over partitions is the one in which every bit
zj is assigned uniformly random to Alice or Bob (but
not both).
Definition 2. A deterministic average partition protocol P consists of a protocol Pρ for every partition of
the inputs z = z1 , . . . , zℓ . Every protocol Pρ has to be
correct on all inputs partitioned according to ρ.
The average partition communication cost of a deterministic protocol P on an input z is the average number of bits exchanged by Pρ , where the distribution on ρ
is defined such that each zj is assigned as input to Alice or Bob uniformly at random. The average partition
communication cost of P is then the maximum over all
inputs.
The deterministic average partition communication
complexity of a function f is the minimum average
partition communication cost over all protocols for f .
For randomized protocols the following correctness
criteria must hold: for an input z the probability of P
giving the correct output must be at least 2/3, over both
the random choices inside the Pρ and the choice of Pρ
itself.
We note that in the definition above, the notion of
average partition is slightly different from the notion of
random-partition we define in Section 3.1 in that we are
interested in expected communication complexity for the
case of average partition, and worst-case communication
complexity for the case of random partition. Since a
protocol with worst-case communication complexity c
also gives an average communication complexity c, a
lower bound for the average partition model is also a
lower bound for the random partition model. In this
section, we will prove a lower bound for the average
partition model.
We will show a lower bound on the randomized average partition communication complexity of the Disjointness problem disj. This complexity measure is defined as a maximum over all inputs (of the average cost
of randomized protocols under a random partition). As
usual we can instead (as in the easy direction of the Yao
principle) bound the expected complexity of deterministic protocols over a hard distribution on inputs (and
over the random partitions).
Recall that the partitions ρ of {z1 , . . . , zℓ } are
chosen uniformly. Furthermore, inputs z will be chosen
from a distribution µ defined by Razborov. It is
important that inputs and their partition are chosen
independent of each other. Due to independence, when
analyzing the error of a protocol, we may also choose the
partition first, and then consider the error for random
inputs given a fixed partition.
This error is the expectation over inputs z (from µ) and
over partitions ρ of the input positions {1, . . . , 2N }. We
denote the expected error under a fixed partition ρ, i.e.,
the expected error of Pρ , by ǫρ . The probability (over
ρ) that ǫρ is larger than ǫ/δ is at most δ.
Besides the error also the communication cost on
input z is an expectation, this time over the choice of the
partition. Recall that we are dealing with deterministic
protocols Pρ . Denote by c the expected communication
over the choice of ρ and the choice of z, and by cρ the
expectation over z with ρ fixed.
In our main lemma we will show that for most
partitions ρ this is impossible unless c = Ω(N ).
We first define the input distribution used by
Razborov for disj on inputs of size N .
Definition 3. Set N = 4m−1, the size of the universe
of elements. To choose inputs according to Razborov’s
distribution one first chooses a frame. The frame
consists of a partition of {1, . . . , N } into 3 disjoint
subsets zx , zy , {i} such that |zx | = |zy | = 2m − 1.
For the distribution µ1 on 1-inputs to disj one then
chooses a subset x of size m of zx as an input to Alice,
and a subset of size m of zy as an input to Bob.
For the distribution µ0 on 0-inputs one chooses a
subset x of size m − 1 from zx and lets Alice’s input be
x ∪ {i}, and correspondingly for Bob.
The distribution µ is the mixture 3/4 · µ0 + 1/4 · µ1.
Note that µ is the uniform distribution on all inputs
for which both sets are of size m and the intersection
size is at most 1.
In our setting the partition ρ of input positions (not
Theorem A.1. The randomized average partition com- to be confused with the partitions from the definition of
munication complexity of the Disjointness problem disj µ) is chosen independent of the inputs (which are chosen
is Ω(N ).
from µ). We next observe that usually enough inputs
The theorem also holds, if Alice also gets to know xi , yi are distributed among Alice and Bob.
all of x and Bob gets to know all of y.
Lemma A.1. Let ρ be a uniformly random partition of
Proof of Theorem A.1 The inputs to Disjointness
are z = x1 , . . . , xN , y1 , . . . , yN . We first take a look
at a simple property that most partitions have, namely
that on most partitions many pairs xj , yj are distributed
among Alice and Bob. For simplicity we only care about
pairs where xj is with Alice and yj is with Bob.
After that we follow Razborov’s proof for the hardness of Disjointness under a fixed distribution µ, but we
have to be more flexible with parameters, and restrict
attention to a part of the distribution µ that is still hard
(for a given partition of the inputs). Furthermore we
have to analyze certain parts of rectangles separately to
keep the product properties in Razborov’s proof intact.
In the following δ will be a small all-purpose constant. Denote the expected error of the protocol by ǫ.
the input variables x1 , . . . , xN , y1 , . . . , yN . pThen with
probability 1 − δ we have that at least N/4 − N/δ pairs
xj , yj are distributed such that Alice holds xi and Bob
holds yi .
Proof: For each pair xi is on Alice’s side and yi on
Bob’s side with probability 1/4. Let C denote number
of i’s for which this is the case. The expectation of C is
N/4 and the variance is 3N/16. Hence, by Chebyshev’s
inequality
probability that |C − N/4| is larger than
√ √ the √
(1/ δ)( 3/4) N is at most δ.
We call partitions ρ under which more than n =
N/5 positions j are such that xj is with Alice and yj
with Bob good partitions. In the following we analyze
protocols under an arbitrary fixed good partition, showing that protocols with low communication cost will
have large error. Since the expected communication is
at least 1 − δ times the cost under the best good partition, a linear lower bound will follow for a random
partition. On the other hand, at most a δ fraction of all
partitions is allowed to have larger error than ǫ/δ.
So fix any good partition ρ. For notational simplicity assume that x1 , . . . , xn belong to Alice, and
y1 , . . . , yn to Bob. The remaining inputs are distributed
in any other way. We consider inputs drawn from the
distribution µ.
Note first, that the probability (under µ) that for
the ”intersection position” i is larger than n, i.e., that
both xi and yi might belong to Alice (or Bob) is 4/5, and
in this case we assume the protocol already knows the
answer without communication. So we have to assume
that ǫ < 1/20 to get a lower bound, since i is contained
in both x and y with probability 1/4, and hence with
error 1/20 the problem to be solved is trivial. If the error
is smaller than, say 1/50, then the protocol must give
the correct output for most inputs with the intersecting
position i smaller than n + 1.
Define µρ as the distribution µ restricted to i (as
part of the frame) being chosen from 1 to n. Note that
µ = 1/5 · µρ + 4/5 · σ for some distribution σ. Since
the protocols Pρ we consider have error ǫρ on µ (using
partition ρ), they can have error at most 5ǫρ on µρ . So
we restrict our attention to µρ .
µρ,1 is the distribution µ1 restricted to i being from
1 to n, and µρ,0 is the distribution µ0 under the same restriction. We will denote the random variables of inputs
chosen from µρ,1 by x0 , y0 (indicating intersection size
0), and the random variables of inputs chosen from µρ,0
by x1 , y1 . x, y is the pair of random variables drawn
from µ. Note that x is the random variable corresponding to the marginal of µ on the x-variables.
Furthermore we will denote by a, b etc. the random
variables x, y grouped into Alice’s and Bob’s inputs.
Note that the first n variables in x belong to Alice
and the first n variables in y belong to Bob, the other
variables are assigned in any way. Subscripts again
indicate drawing inputs from µρ,0 or µρ,1 .
We can now state the Main Lemma.
In other words, all rectangles R that are trying
to cover mostly 1-inputs to disj are either small or
corrupted by a constant fraction of 0-inputs. A low
error cover made up of rectangles hence needs to contain
many rectangles.
Now we show that the main lemma implies the
lower bound for disj. The randomized average partition communication complexity is lower bounded by
Eρ Eµ cρ,z (f ), where cρ,z is the communication cost on
input z under partition ρ. Since all good partitions
together have probability 1 − δ, we just need to show
that for any good partition ρ the average communication Eµ cρ,z ≥ Ω(N ).
We have Prρ,µ (cρ,z > δc) < δ. Hence, if we stop all
protocols Pρ once their communication on an input z
exceeds c/δ, we increase the error by at most δ. Hence
in the following we can assume that for each ρ and z we
have cρ,z ≤ c/δ, and in particular, that cρ ≤ c/δ.
The protocol Pρ has communication at most c/δ,
and corresponds to a partition of the communication
matrix according to ρ into 2c/δ rectangles with total
error ǫρ +δ under µ. Since µρ makes up one fifth of µ, the
error under µρ can be at most 5(ǫρ + δ). Furthermore,
the inputs in the support of µρ (i.e., the inputs with
intersection size at most 1 and intersections happening
at position n or less) are partitioned into 2c/δ rectangles
with error 10(ǫρ +δ). Since µρ puts weight 3/4 on inputs
x, y that are disjoint, there must be a single rectangle
R that covers (3/12) · 2−c/δ of 1-inputs and at the same
time has error at most 30(ǫ/δ + δ) (both under µρ ). The
Main Lemma now leads to a contradiction for ζN < c/δ
and 30(ǫ/δ + δ) being a suitably small constant.
We now proceed to show the main lemma. In
Razborov’s proof the rectangles R to be analyzed are
product sets with respect to the divide between x and y.
For us this is not the case, because Alice and Bob hold
both x- and y-inputs. However, Alice holds x1 . . . , xn
and Bob y1 , . . . , yn . Call all the remaining variables
(that are distributed arbitrarily) O.
In the statement of the Main Lemma α/(1 + α) is
(a lower bound on) the error of R under µρ . Since α is a
constant we will determine later it is enough to show a
constant lower bound on the error of any large rectangle
Lemma A.2. Let ρ be a good partition. x, y are chosen R. Denote the error of R under µρ by α′ .
from µρ (i.e., µ with i < n + 1), and partitioned into
The rectangle R can be partitioned into smaller
random variables as described above. Let R = Ā × B̄ rectangles by fixing the variables O to some value o.
denote a rectangle of size 2−ζN (under µρ ) in the We denote a resulting rectangle by Ro (note that Ro
communication matrix induced by disj and ρ, for some is indeed a rectangle). Let γo denote the average error
small constant ζ > 0.
of Ro under µρ , normalized to Ro (where o is chosen
Then
according to the size of Ro ). Clearly the expected γo is
α′ .
Pr ((a1 , a1 ) ∈ R) ≥ α · Pr ((a0 , a0 ) ∈ R) ,
Ro = Āo ×B̄ o corresponds to a rectangle in {0, 1}n ×
{0, 1}n by ignoring fixed input variables. R itself is a
for some constant α > 0.
rectangle in {0, 1}N × {0, 1}N . o fixes x, y on the last
N −n input positions, and if we define a as x∩{1, . . . , n}
and b = y ∩ {1, . . . , n}, then ma = |a| and mb = |b| are
fixed on Ro .
For every a, b with fixed ma , mb there are
N −n
N − n − m + ma
fa,b =
·
m − ma
mb
in the same situation as Razborov, except that the sets
a, b are not exactly of size n/4.
Denote by ma and mb the sizes of a, b respectively.
We consider the distribution ν which is defined as
follows.
Definition 4. n = 4m′ − 1. To choose inputs one
first chooses a frame. There are n input positions. The
frame consists of a partition of {1, . . . , n} into 3 disjoint
extensions o. And hence as many (disjoint) rectangles subsets za , zb , {i} such that |za | = |zb | = 2m′ − 1.
For the distribution ν1 on 1-inputs to disj one then
Ro in R. Later we will restrict our attention to o for
which ma , mb are very close to n/4. Recall that R has chooses a subset a of size ma of za as an input to Alice,
size µρ (R) ≥ 2−ζN by the assumptions in the Main and a subset of size mb of zb as an input to Bob.
For the distribution µ0 on 0-inputs one chooses a
Lemma.
To apply a modification of Razborov’s original subset a of size ma − 1 from za and lets Alice’s input be
argument for disj on {0, 1}n × {0, 1}n we want to find a ∪ {i}, and correspondingly for Bob.
The distribution ν is the mixture 3/4 · ν0 + 1/4 · ν1 .
an o such that Ro
1. is large (as a rectangle in {0, 1}n × {0, 1}n)
2. has small error
3. the size of x∩{1, . . . n} and y ∩{1, . . . , n} is roughly
n/4 for all inputs in Ro .
Note that ν can be obtained from µρ by dropping
the last N − n elements of the universe. Since µ is
invariant under permuting the last N − n or the first n
elements of the universe, for all x, y and a, b obtained
from them ν(a, b) = µρ (x, y) · fa,b . In our case all fa,b
are within 2O(κn) of each other. This means that our
Ro has size ν(Ro ) ≥ 2−2ζN −O(κN ) .
What is left to do is to prove Razborov’s Main
Lemma under the modification that the sets x, y have
fixed sizes that are slightly different from n/4. The argument is almost entirely the same, with some estimates
looser than in the original.
We now restate that Lemma. In this restatement
we re-use the notation a1 etc., which now refer to the
domain {0, 1}n × {0, 1}n and ν etc.
−1
The rectangles Ro with µρ (R0 ) < fa,b
· 2−2ζN contribute at most 2−2ζN to R and can thus be disregarded without changing the error much (even if the
small Ro contain only 1-inputs). We choose o now
by picking x, y from µρ restricted to R, and removing
x1 , . . . , xn , y1 , . . . , yn . This clearly chooses a Ro by size,
and hence the expected error is at most 2α′ , and the
probability that the error is larger than 2α′ /δ is at most
δ.
We now have to consider the size of sets a = Lemma A.3. a, b are chosen from ν (as above). Let
x ∩ {1, . . . , n} and b = y ∩ {1, . . . , n}. Note that o fixes R = Ā × B̄ denote a rectangle of size 2−ζ ′ , for some
some kx of the x-variables to 1 and some ky of the y- small constant ζ ′ > 0.
variables, and the size of a is m − kx , the size of b is
Then
m − ky .
Pr ((a1 , a1 ) ∈ R) ≥ α · Pr ((a0 , a0 ) ∈ R) ,
The following claim follows from a simple application of the Chernoff inequality.
for some constant α > 0.
Claim 1. Choose o by choosing x, y from µρ and disWe omit the details of this. As discussed above this
carding the variables outside O.
implies our Main Lemma, and hence the lower bound
2
Pr(| |a| − n/4 | > κn or | |b| − n/4 | > κn) ≤ 2−κ n/3 . for disj.
To see that the above proof also applies if Alice
2
knows
all of x and Bob knows all of y note that we
For a small enough κ (namely κ /3 · n ≪ ζN ) this
actually
analyze the rectangles R0 , in which the extra
means that when we choose o as above (i.e., x, y from R
inputs
are
fixed anyway.
according to µ and dropping the unneeded variables),
then with probability > 2/3 the size of a and b are
B Proof of Lemma 3.1
within ±κn of n/4.
Hence we can find o such that the above 3 criteria We will use the following lower bound which can be
are satisfied, and all that is left to do is to show that Ro derived using basic facts in communication complexity.
must have large error after all. By now we are almost We sketch its proof for completeness.
cc−pub
Theorem B.1. For some ǫ > 0, R0,ǫ
(eq) ≥ b.
gives the additional error of 4/b.) Moreover, the number
of communication bits that their√simulation requires
is
√
Proof. [Proof sketch] We need the following additional less than |Sb |+(b−|Sb|)/2+(1/2) 6b ln b+b/4− 6b ln b
√
notations. Let N 1 (f ) denote the nondeterminic com- which is less than b since |Sb | ≤ b/2+(1/2) 6b ln b. This
munication complexity of computing f (see, e.g., [22, implies that Rcc−pub (eq) < b as desired.
0,ǫ
Definition 2.3]). It is well-known that N 1 (eq) ≥ b
Note that in the reduction above Alice knows x and
(see, e.g. [22, Example 2.5]). Note further that Bob knows y. Thus, the claimed lower bound holds even
cc−pub
R0,ǫ
(f ) ≥ N 1 (f ) (see, e.g., [22, Proposition 3.7]6 ). when Alice and Bob have this information in addition
cc−pub
Thus, R0,ǫ
(eq) ≥ N 1 (eq) ≥ b as claimed.
to the random partition.
Proof. [Proof Lemma 3.1]
prove that if
√ We will now cc−pub
rcc−pub
R0,ǫ
(eq) < b/4 − 6b ln b, then R0,ǫ
(eq) <
b, thus contradicting Theorem B.1. This immedircc−pub
ately implies that R0,ǫ
(eq) = Ω(b), as claimed in
Lemma 3.1. Let A be a protocol in the random-partition
model which requires strictly less than b/4 bits of communication. Alice and Bob simulates A as follows.
First, they use shared randomness to pick a random
partition of x and y, denoted by P. Let SB be the set
of positions in x that
√ Bob receives in partition P. If
|SB | > b/2 + (1/2) 6b ln b, then they stop the simulation and answer 0 (representing “x 6= y”). Note that
this could cause an error when x = y; however, by
Chernoff’s bound (e.g. [30, Section 4.2.2]), this happens with probability at most 2/b. If Alice and Bob
do not stop the simulation, Alice sends to Bob bits in
x at positions
√ in SB . This causes a communication of
b/2 + (1/2) 6b ln b (note that Alice simply sends |SB |
bits in order and Bob can figure out which bit belongs
to which position since he also knows the partition P).
Next, Bob checks whether bits in x sent from Alice is
the same as his bits (in the same position). If not, he
immediately answers 0 (“x 6= y”); otherwise, we let
SA be the set of positions not in SB such that Alice
receives bits in y at positions
√ in SA . Bob checks if
|SA | > (b − |SB |)/2 + (1/2) 6b ln b. If this is the case,
then Alice and Bob stops the simulation and answer 0.
This again could cause an error but it happens with
probability at most 2/b by Chernoff’s bound. If this
does not happen, Bob sends bits in y at positions
√ in SA
to Alice. This causes at most (b−|SB |)/2+(1/2) 6b ln b
bits of communication. Then, Alice and √
Bob simulates
A, which will finish after less than b/4 − 6b ln b bits of
communication and produces an answer that is (ǫ, 0)error. Alice and Bob use this answer as their answer.
Combining with the previous errors, we have that Alice
and Bob’s answer is (ǫ+ 4/b, 0)-error. (This error comes
from the worst-case scenario where A could be always
correct for the partitions that Alice and Bob stops. This
6 We note that our notations are slightly different from [22].
It can be checked from [22, Definitions 3.1 and 3.3] that the
rcc−pub
definition of R0,ǫ
is the same as R1 , i.e. they correspond
to the case where we allow an error only when f (x, y) = 1.
| 8cs.DS
|
1
Spectral analysis for non-stationary audio
arXiv:1712.10252v1 [eess.AS] 29 Dec 2017
Adrien Meynard and Bruno Torrésani
Abstract
A new approach for the analysis of non-stationary signals is proposed, with a focus on audio applications. Following earlier contributions, non-stationarity is modeled via stationaritybreaking operators acting on Gaussian stationary random signals. The focus is here on time
warping and amplitude modulation, and an approximate maximum-likelihood approach based
on suitable approximations in the wavelet transform domain is developed. This papers provides theoretical analysis of the approximations, and describes and analyses a corresponding
estimation algorithm. The latter is tested and validated on synthetic as well as real audio
signal.
Index Terms
Non-stationary signals, deformation, wavelet analysis, local spectrum, Doppler effect
I. I NTRODUCTION
Non-stationarity is a key feature of acoustic signals, in particular audio signals. To
mention a few examples, a large part of information carried by musical and speech
signals is encoded by their non-stationary nature, as is the case for environment sounds
(think of car noises for example, where non-stationarity informs about speed variations),
and many animals (bats, dolphins,...) use non-stationary signals for localization and
communication. Beyond acoustics, amplitude and frequency modulation are of prime
importance in telecommunication.
A. Meynard and B. Torrésani are with Aix Marseille Univ, CNRS, Centrale Marseille, I2M, UMR 7373, Marseille,
France.
Part of this work was done when the two authors were at Centre de Recherches Mathématiques, UMI 3457, CNRS
and Université de Montréal, Canada.
Manuscript received ???, 2017; revised ???, 2017.
2
While stationarity can be given rigorous definitions, non-stationarity is a very wide
concept, as there are infinitely many ways to depart from stationarity. The theory of
random signals and processes (see [1], [2] and references therein) gives a clear meaning
to the notion of stationarity. In the context of time series analysis, Priestley [2], [3] was
one of the first to develop a systematic theory of non-stationary processes, introducing
the class of locally stationary processes and the notion of evolutionary spectrum. A
similar approach was followed in [4], who proposed a wavelet-based approach to
covariance estimation for locally stationary processes (see also [5]). An alternate theory
of locally stationary time series was developed by Dahlhaus [6] (see also [7] for a
corresponding stationarity test). In a different context, frequency modulated stationary
signal were considered in [8], [9], and time warping models were analyzed in [10]. In
several of these approaches, wavelet, time-frequency and similar representations happen
to play a key role for the characterization of non-stationarity.
In a deterministic setting, a popular non-stationarity model expresses the signal as
a sum of K sinusoidal components y(t) = ∑K
k=1 Ak (t) cos (2πφk (t)). This model has
been largely used in speech processing since early works by McAulay and Quatieri [11]
(see also [12] and references therein for more recent developments, and [13], [14] for
probabilistic approaches). The instantaneous frequencies φk′ of each mode give important information about the physical phenomenon. Under smoothness assumptions
on functions Ak and φk′ , techniques such as ridge/multiridge detection (see [15] and
references therein), synchrosqueezing or reassignment have been developed to extract
theses quantities from a single signal observation (see [16], [17] for recent accounts).
In sound processing, signals often possess an harmonic structure, which corresponds
to a special case of the above model where each instantaneous frequency φk′ is multiple
of a fundamental frequency φ0′ : φk′ (t) = λk φ0′ (t), λk ∈ R. If the amplitudes Ak are
such that Ak (t) = αk A0 (t), we can describe such signals as a stationary signal x (t) =
∑K
k=1 α k cos(2πλk t + ϕ k ) modified by time warping and amplitude modulation: y(t) =
A0 (t) x (φ0 (t)). A major limit of this model is that each component is purely sinusoidal
while audio signals often contain broadband information. However sounds originating
from physical phenomena can often be modeled as stationary signals which have been
deformed by a stationarity breaking operator (time warping, amplitude modulation,...).
For example sounds generated by a variable speed engine or any stationary sound
3
deformed by Doppler effect can be described as such. A stochastic time warping model
has been introduced in [18], [19], where wavelet-based approximation and estimation
techniques were developed. In [9], [20], an approximate maximum-likelihood approach
was proposed for the joint estimation of the time warping and power spectrum of the
underlying Gaussian stationary signal, exploiting similar approximations.
In this paper, we build on results of [9], [20] which we extend and improve in
several ways. We develop an approximate maximum likelihood method for estimating
jointly time warping and amplitude modulation from a single realization. While the
overall structure of the algorithm is similar, we formulate the problem as a continuous
parameter estimation problem, which avoids quantization effects present in [9], [20],
and allows computing a Cramér-Rao bound for assessing the precision of the estimate.
After completing the estimation, the inverse deformation can be applied to the input
signal, which yields an estimate for the power spectrum.
The plan of the paper is as follows. After giving some definitions and notations
in Section II, we detail in Section III the non-stationary signal models we consider,
and specify the assumptions made on the underlying stationary signal. We also analyze the effect of time warping and amplitude modulation in the wavelet domain,
which we exploit in designing the estimation procedure. We finally propose an alternate
estimation algorithm, and analyze the expected performances of the corresponding
estimator. Section IV is devoted to numerical results, on both synthetic signals and
real sounds. We also shortly describe in this section an extension published in [21]
involving simultaneously time warping and frequency modulation. More mathematical
developments are postponed to the Appendix.
II. N OTATIONS
AND BACKGROUND
A. Random signals, stationarity
Throughout this paper, we will work in the framework of the theory of random
signals. Signals of interest will be modeled as realizations of random processes1 t ∈
R → Xt ∈ C. In this paper, the random processes will be denoted by upper-case letters
1 Signals
of interest are real-valued, however we will also use complex-valued functions since we will use complex-
valued wavelet transforms later on.
4
while their realizations will be denoted by lower-case letters. The random processes
will be assumed to zero mean (E { Xt = 0} for all t) and be second order, i.e. they
have well defined covariance kernel E Xt X s . A particularly interesting class of such
stochastic processes is the class of second order (or weakly) stationary processes, for
∆
which CX (t − s) = E Xt X s is a function of t − s only. Under these assumptions, the
Wiener-Khinchin theorem states that the covariance kernel may be expressed as the
inverse Fourier transform of a non-negative measure dηX , which we will assume to be
continuous with respect to the Lebesgue measure: dηX (ν) = SX (ν)dν, for some nonnegative L1 function SX called the power spectrum. We then write
CX ( t ) =
Z
SX (ν)e2iπνt dν .
We refer to textbooks such as [1], [2] for a more complete mathematical account of the
theory, and to [20] for an extension to distribution theory setting.
B. Elementary operators
Our approach rests on non-stationary models obtained by deformations of stationary
random signals. We will mainly use as elementary operators the amplitude modulation
Aα , translation Tτ , dilation Ds , and frequency modulation Mν defined as follows:
Aα x (t) = αx (t) ,
Ds x (t) = q−s/2 x (q−s t) ,
Tτ x (t) = x (t − τ ) ,
Mν x (t) = e2iπνt x (t) .
where α, τ, s, ν ∈ R and q > 0 is a fixed number. The amplitude modulation commutes
with the other three operators, which satisfy the commutation rules
Tτ Ds = Ds Tq−s τ , Tτ Mν = e−2iπντ Mν Tτ , Mν Ds = Ds Mνqs .
C. Wavelet transform
Our analysis relies heavily on transforms such as the continuous wavelet transform
(and discretized versions). In particular, the wavelet transform of a signal X : t ∈ R → Xt
is defined as:
W X (s, τ ) = hX, ψsτ i , with ψsτ = Tτ Ds ψ .
(1)
where ψ is the analysis wavelet, i.e. a smooth function, with fast decay away from the
origin. It may be shown that for suitable choices of ψ the wavelet transform is invertible
5
(see [15]), we will not use that property here. Notice that when X is a realization of
a continuous time random process, it does not need to decay at infinity. However,
for a suitably smooth and localized wavelet ψ, the wavelet transform can still be well
defined (see [15], [20] for more details). In such a situation the wavelet transform of X
is a two-dimensional random field, which we analyze in the next section. Besides, in
this paper the analysis wavelet ψ is complex valued and belongs to the set H 2 (R ) =
ψ ∈ L2 (R ) : supp(ψ̂ ) ⊂ R + . In that framework, a useful property is that if X is a
real zero mean Gaussian random process then, W X is a complex zero mean circular
Gaussian random field.
Classical choices of wavelets in H 2 (R ) are (analytic) derivative of Gaussian ψk (which
has k vanishing moments), and the sharp wavelet ψ♯ (with infinitely many vanishing
moments) introduced in [21]. These can be defined in the positive Fourier domain by
ψ̂k (ν) = νk e−kν
2 /2ν2
0
,
δ ( ν,ν0 )
ψ̂♯ (ν) = ǫ δ(ν1,ν0 ) , ν > 0
(2)
and vanish on the negative Fourier half axis. Here ν0 is the mode of ψ̂. In the expression
of ψ̂♯ , ν1 is chosen so that ψ̂♯ (ν1 ) = ǫ (a prescribed numerical tolerance at cutoff
frequency ν1 ), and the divergence δ is defined by δ(a, b) = 12 ba + ba − 1.
III. J OINT
ESTIMATION OF TIME WARPING AND AMPLITUDE MODULATION
A. Model and approximations
Let us first describe the deformation model we will mainly be using in the following.
As said above, the non-stationary signals of interest are obtained as linear deformations
of stationary random signals. The deformations of interest here are amplitude modulations and time warpings. Amplitude modulations are pointwise multiplications by
smooth functions, defined as
Aa :
Aa x (t) = a(t) x (t) ,
(3)
where a ∈ C1 is a real valued function, such that
0 < c a ≤ a(t) ≤ Ca < ∞,
∀t ,
(4)
for some constants c a , Ca ∈ R ∗+ . Time warpings are compositions with smooth and
monotonic functions,
Dγ :
Dγ x ( t ) =
q
γ′ (t) x (γ(t))
(5)
6
where γ ∈ C2 is a strictly increasing smooth function, satisfying the control condition [20]
0 < cγ ≤ γ′ (t) ≤ Cγ < ∞,
∀t ,
(6)
for some constants cγ , Cγ ∈ R ∗+ .
Assume one is given a (unique) realization of a random signal of the form
Y = A a Dγ X
(7)
where X is a stationary zero mean real random process with (unknown) power spectrum
SX . The goal is to estimate the deformation functions a and γ from this realization of
Y, exploiting the assumed stationarity of X.
Remark 1: Clearly enough, the stationarity assumption is not sufficient to yield unambiguous estimates, since affine functions γ(t) = λt + µ do not break stationarity: for
any stationary X, Dγ X is stationary too. Therefore, the warping function γ can only be
estimated up to an affine function, as analyzed in [19] and [20]. Similarly, the amplitude
function a can only be estimated up to a constant factor.
A key ingredient of our approach is the smoothness of the deformation functions a
and γ, and their slow variations. This allows us to perform a local analysis using smooth
and localized test functions, on which the action of Aa and Dγ can be approximated
τ
g
fτ
by their so-called tangent operators A
a and Dγ (see [19], [9], [20], [22]). More precisely,
given a test function g located near t = τ (i.e. decaying fast enough as a function of
|t − τ |), Taylor expansions near t = τ yield
τ
g
fτ ∆
Aa g(t) ≈ A
a g(t) , with Aa = A a(τ ) ,
∆
fγτ g(t) , with D
fγτ =
Dγ g ( t ) ≈ D
Tτ D− logq(γ′ (τ )) T−γ(τ ) .
(8)
(9)
∆
fY (s, τ ) =
Therefore, the wavelet transform of Y will be approximated by WY (s, τ ) ≈ W
E
D
τD
τ X, T D ψ , i.e.
g
f
A
τ
s
a
γ
′
f
(10)
WY (s, τ ) = a(τ )W X s + logq (γ (τ )), γ(τ ) .
Here we have used the standard commutation rules of translation and dilation operators
given in Section II-B.
The result below provides a quantitative assessment of the quality of the approximation. There, we denote by k f k∞ = ess supt | f (t)| the essential absolute supremum of a
function f .
7
Theorem 1: Let X be a second order zero mean stationary random process, let Y be the
non-stationary process defined in (7). Let ψ be a smooth test function, localized in such
a way that |ψ(t)| ≤ 1/(1 + |t| β ) for some β > 2. Let WY denote the wavelet transform of
fY its approximation given in (10), and let ε = WY − W
fY denote the approximation
Y, W
error. Assume ψ and SX are such that
rZ
∞
(ρ) ∆
IX =
0
ξ 2ρ SX (ξ )dξ < ∞ , where ρ =
β−1
.
β+2
Then the approximation error ε is a second order, two-dimensional complex random
field, and
where
n
o
ρ
E |ε(s,τ )|2 ≤ Ca2 q3s K1kγ′′ k∞ + K2 qµs kγ′′ k∞ + K3 a′
βσX
√ ,
2( β − 2) c γ
p
Cγ βσX
,
K3 =
( β − 2) c a
(ρ)
K2 = I X
K1 =
µ=
π ρ q
2
∞
Cγ
2
4
,
3ρ
β−4
,
β+2
σX2 being the variance of X.
The proof, which is an extension of the one given in [20], is given in appendix A.
Remark 2: The assumption on β ensures that the parameters belong to the following
intervals: 1/4 < ρ < 1 and −1/2 < µ < 1. Therefore, the variance of the approximation
error tends to zero when the scales are small (i.e. s → −∞). Besides, the error is
inversely proportional to the speed of variations of γ′ and a. This is consistent with
the approximations of the deformation operators by their tangent operators made in
equations (8) and (9).
From now on, we will assume the above approximations are valid, and work on the
fY , which
approximate random fields. The problem is then to estimate jointly a, γ from W
is a zero mean random field with covariance
o
n
fY (s′ , τ ′ ) = C(s, s′ , τ, τ ′ )
fY (s, τ )W
E W
(11)
where
C(s, s′, τ, τ ′ ) = a(τ )a(τ ′ )q
s+s′
2
q
Z ∞
SX (ξ )
γ′ (τ )γ′ (τ ′ )
0
s′ ′ ′ 2iπξ (γ(τ )−γ(τ ′ ))
s ′
dξ.
×ψ̂ q γ (τ )ξ ψ̂ q γ (τ )ξ e
(12)
8
B. Estimation
1) Estimation procedure: Our goal is to estimate both deformation functions γ and a
fy of a realization y of Y, assuming the
from the approximated wavelet transform W
latter is a reliable approximation of the true wavelet transform. From now on, we
fY is a zero mean
additionally assume that X is a Gaussian random process. Therefore, W
circular Gaussian random field and its probability density function is characterized by
the covariance matrix. However, equation (12) shows that besides deformation functions
the covariance also depends on the power spectrum SX of the underlying stationary
signal X, which is unknown too. Therefore, the evaluation of the maximum likelihood
estimate for a and γ requires a guess for SX . This constraint naturally brings the
estimation strategy to an alternate algorithm. In [21], an estimate for the power spectrum
was obtained at each iteration by computing a Welch periodogram on a “stationarized”
signal Aã−1 Dγ̃−1Y, ã and γ̃ being the current estimates for the deformation functions a
and γ. We use here a simpler estimate, computed directly from the wavelet coefficients.
The two steps of the estimation algorithm are detailed below.
Remark 3: The alternate likelihood maximization strategy is reminiscent of the Expectation-Maximization (EM) algorithm, the power spectrum being the nuisance parameter.
However, while it would be nice to apply directly the EM paradigm (whose convergence
is proven) to our problem, the dimensionality of the latter (and the corresponding size
of covariance matrices) forces us to make additional simplifications that depart from the
EM scheme. Therefore we turn to a simpler approach with several dimension reduction
steps.
(a) Deformation estimation. Assume that the power spectrum SX is known (in fact, only
an estimate S˜X is known). Thus, we are able to write the likelihood corresponding to
the observations of the wavelet coefficients. Then the maximum likelihood estimator is
implemented to determine the unknown functions γ and a.
The wavelet transform (1) is computed on a regular time-scale grid Λ = s × τ, δs
being the scale sampling step and Fs the time sampling frequency. The sizes of s and τ
are respectively denoted by Ms and Nτ .
Considering the covariance expression (12) we want to estimate the vector of pa∆
fy (Λ ) denote the
rameters Θ = (θ1 , θ2 , θ3 ) = (a(τ )2 , logq (γ′ (τ )) , γ(τ )). Let Wy = W
9
discretized transform and let CW (Θ ) be the corresponding covariance matrix. The related log-likelihood is
1
1
L (Θ ) = − ln |det(CW (Θ ))| − CW (Θ )−1 Wy · Wy .
2
2
(13)
The matrix CW (Θ) is a matrix of size Ms Nτ × Ms Nτ , which is generally huge. For
instance, when the signal is a sound of 5 seconds sampled at frequency Fs = 44.1
kHz and the wavelet transform is computed on 8 scales, the matrix CW (Θ ) has about
3.1 trillion elements which makes it numerically intractable. In addition, due to the
redundancy of the wavelet transform, CW (Θ ) turns out to be singular, which makes
the evaluation of the likelihood impossible.
To overcome these issues, we use a block-diagonal regularization of the covariance
matrix, obtained by forcing to zeros entries corresponding to different time indices. In
other words, we disregard time correlations in the wavelet domain, which amounts to
fy (s, τn ) as independent circular Gaussian vectors
consider fixed time vector wy,τn = W
with zero mean and covariance matrix
C(Θn )ij = θn,1 C0 (θn,2 )ij , 1 ≤ i, j ≤ Ms ,
where
C0 (θn,2 )ij = q(si +s j )/2
Z ∞
0
SX (q−θn,2 ξ )ψ̂ (qsi ξ )ψ̂ (qs j ξ )dξ .
(14)
(15)
In this situation, the regularized log likelihood L r splits into a sum of independent
terms
L r (Θ ) = ∑ L (Θn ) ,
n
∆
where Θn = (θn,1 , θn,2 ) = (θ1 (n), θ2 (n)) corresponds to the amplitude and warping
parameters at fixed time τn = τ (n). Notice that in such a formalism, θn,3 = γ(τn )
does not appear any more in the covariance expression. Thus we are led to maximize
independently for each n
1
1
L (Θn ) = − ln |det(C(Θn ))| − C(Θn )−1 wy,τn · wy,τn .
2
2
(16)
For simplicity, the estimation procedure is done by an iterative algorithm (given in
more details in part III-B2), which rests on two main steps. On the one hand, the loglikelihood is maximized with respect to θn,2 using a gradient ascent method, for a fixed
10
value of θn,1 . On the other hand, for a fixed θn,2 , an estimate for θn,1 is directly obtained
which reads
θ̃n,1 =
1 −1
C (θn,2 )wy,τn · wy,τn ,
Ms 0
(17)
(b) Spectrum estimation. Assume the amplitude modulation and time-warping parameters θ1 and θ2 are known (in fact, only estimates θ̃1 and θ̃2 are known). For any n we
can compute the wavelet transform
1 f
Wy (s − θn,2 , τn ) = W x (s, γ (τn )) ,
(18)
1/2
θn,1
For fixed scale sm , w x,sm = W x (sm , γ(τ )) ∈ C Nτ is a zero mean random circular Gaussian
vector with time independent variance (as a realization of the wavelet transform of
a stationary process). Hence, the empirical variance is an unbiased estimator of the
variance. We then obtain
(
)
Z ∞
1
1
2
2
E
kwx,sm k =
SX (ξ )qsm ψ̂ (qsm ξ ) dξ
2
2
Nτ kψk2
k ψ k2 0
∆
= SX,ψ (q−sm ω0 ) ,
(19)
where ω0 is the central frequency of |ψ̂ |2 . SX,ψ is a band-pass filtered version of SX
centered around frequency νm = q−sm ω0 . Besides, the bandwidth of the filter is propor-
tional to the frequency νm . This motivates the introduction of the following estimator
S˜X of SX
S˜X (q−sm ω0 ) =
1
kwx,sm k2 .
Nτ kψk22
(20)
Finally, the estimate SX is extended to all ξ ∈ [0, Fs /2] by linear interpolation.
2) Algorithm: The estimation procedure is implemented in an iterative alternate op-
timization algorithm, whose pseudo-code is given as Algorithm 1. The initialization
requires an initial guess for the power spectrum SX of X. We use for this the spectrum
estimator (20) applied to the original observations Y.
(k)
After k iterations of the algorithm, estimates Θ̃n
(k)
and S˜X for Θn and SX are
(k)
available. Hence we can only evaluate the plug-in estimate C̃0
of C0 , obtained by
replacing the power spectrum with its estimate in the covariance matrix (15). This yields
an approximate expression L (k) for the log-likelihood, which is used in place of L
in (16) for maximum likelihood estimation. The influence of such approximations on
the performances of the algorithm are discussed in section III-C.
11
To assess the convergence of the algorithm, the relative update of the parameters is
chosen as stopping criterion:
(k)
( k − 1) 2
θ̃j − θ̃j
2
( k − 1) 2
θ̃j
2
< T , for j = 1 and 2 ,
(21)
where 0 < T < 1 is a given threshold.
Finally, after convergence of the algorithm to the estimated value Θ̃
(k)
, logq (γ′ ) and a2
are estimated for all time by cubic spline interpolation. Besides, γ is given by numerical
integration assuming that γ(0) = 0.
Algorithm 1 Joint spectrum and deformations estimation
Initialization: Compute an estimate S˜Y of the power spectrum of Y as an initial
(0)
guess S˜X for SX . Initialize the estimator of the squared amplitude modulation with
(0)
θ̃n,1 = 1, ∀n.
Compute the wavelet transform Wy of y.
k := 1
while criterion (21) is false and k ≤ kmax do
( k + 1)
• For each n, subsample wy,τn on scales s p , and estimate θ̃n,2
(k)
approximate log-likelihood L (k) θ̃n,1 , θn,2 in (16).
by maximizing the
( k + 1)
• For each n, estimate θ̃n,1
by maximizing the approximate log-likelihood
( k + 1)
L (k) θn,1 , θ̃n,1
with respect to θn,1 in (16). Or, in absence of noise, directly apply
equation (17) using the regularized covariance matrix given by (22).
• Construct the estimated wavelet transform W x of the underlying stationary signal
(k)
by interpolation from Wy and θ̃
with equation (18). Estimate the corresponding
( k + 1)
with (20).
power spectrum S˜X
• k := k + 1
end while
• Compute ã and γ̃ by interpolation from Θ̃
(k)
.
Remark 4: In order to control the variances of the estimators, and the computational
cost, two different discretizations of the scale axis are used for θ̃1 or θ̃2 . Indeed, the
computation of the log-likelihood involves the evaluation of the inverse covariance
matrix. In [20], a sufficient condition for invertibility was given in the presence of
12
noise. The major consequence induced by this condition is that when δs is close to
zero (i.e. the sampling period of scales is small), the covariance matrix could not be
numerically invertible. The scale discretization must then be sufficiently coarse to ensure
good conditioning for the matrix. While this condition can be reasonably fulfilled to
estimate θn,2 without impairing the performances of the estimator, it cannot be applied
to the estimation of θn,1 because of the influence of Ms on its Cramér-Rao bound (see
section III-C below). The choice we made is to maximize L (Θn ) for θn,2 with wy,τn
corresponding to a coarse sampling s p which is a subsampled version of the original
vector s, the scale sampling step and the size of s p being respectively pδs and ⌊ Ms /p⌋
for some p ∈ N ∗ . While L (Θn ) is maximized for θn,1 on the original fine sampling s,
a regularization of the covariance matrix has to be done to ensure its invertibility. The
regularized matrix is constructed by replacing covariance matrix C0 (θn,2 ) given by (15)
by its regularized version C0,r (θn,2 ), given by
C0,r (θn,2 ) = (1 − r )C0,r (θn,2 ) + rI ,
(22)
for some regularization parameter 0 ≤ r ≤ 1.
Remark 5: After convergence of the estimation algorithm, the estimated functions ã
and γ̃ allow constructing a “stationarized” signal
x̃ = Dγ̃−1 Aã−1 y .
x̃ is an estimation of the original underlying stationary signal x. Furthermore, the Welch
periodogram [23] may be computed from x̃ to obtain an estimator of SX whose bias is
not depending on frequency (unlike the estimator used within the iterative algorithm).
Remark 6: In order to accelerate the speed of the algorithm, the estimation can be done
only on a subsampled time. The main effect of this choice on the algorithm concerns
the final estimation of a and γ which is more sensitive to the interpolation operation.
In the following section, we analyze quantities that enable the evaluation of the
expected performances of the estimators, and their influence of the algorithm. The reader
who is not directly interested in the statistical background may skip these section and
jump directly to the numerical results in part IV.
13
C. Performances of the estimators and the algorithm
(a) Bias. For θn,1 , the estimator is unbiased when the actual values of θn,2 and SX are
n
o
(k)
(k)
known. In our case, the bias bn,1 (θn,1 ) = E θ̃n,1 − θn,1 is written as
θn,1
(k) (k) −1
(k)
C0 (θn,2 ) − I .
(23)
Trace C̃0 θ̃n,2
bn,1 (θn,1 ) =
Ms
(k)
As expected, the better the covariance matrix estimation, the lower the bias bn,1 .
For θn,2 , as we do not have a closed-form expression for the estimator we are not
able to give an expression of the bias. Nevertheless, if we assume that the two other
true variables are known, as a maximum likelihood estimator we make sure that θ̃n,2 is
asymptotically unbiased (i.e. θ̃n,2 → θn,2 when Ms → ∞).
Regarding SX , equation (19) shows that the estimator yields a smoothed, thus biased
version of the spectrum. Besides, proposition 1 below shows that the estimated spectrum
converges to this biased version when the deformation parameters converge to their
actual values.
Proposition 1: Let ψ ∈ H 2 (R ) be an analytic wavelet such that ψ̂ is bounded and
ψ̂(u) = Ou→∞ (u−η ) with η > 2. Let ϕ1 and ϕ2 be bounded functions defined on R +
by ϕ1 (u) = u ψ̂(u)
2
and ϕ2 (u) = u2 ψ̂ (u) . Assume SX is such that
JX =
(k)
Let SX
Z ∞
0
ξ −1 SX (ξ )dξ < ∞.
denote the estimation of the spectrum after k iterations of the algorithm. Let
(k)
bSX denote the bias defined for all m ∈ [[1, Ms ]] by
n
o
(k)
(k) − sm
˜
bSX (m) = E SX (q
ω0 ) − SX,ψ (q−sm ω0 ) .
(k)
Assume there exists a constant cθ1 > 0 such that θn,1 > cθ1 , ∀n, k. Then
J
(k)
(k)
(k)
≤ X 2 K1′ θ1 − θ̃1
bSX
+ K2′ θ̃2 − θ2
,
∞
∞
∞
k ψ k2
where
K1′ =
k ϕ1 k ∞
<∞,
c θ1
K2′ = ln(q) k ϕ1 k∞ + 2kψ̂′ k∞ k ϕ2 k∞ < ∞ .
The proof is given in appendix B.
(24)
14
Remark 7: If
(k)
θ1
→ θ1 and
(k)
θ2
n
(k)
→ θ2 as k → ∞, we have E S˜X (νm )
which is the expected property.
o
→ SX,ψ (νm )
k
Formula (24) enables the control of the bias of the spectrum at frequencies νm =
q−sm ω0 only. We can also notice the required property JX < ∞ forces SX to vanish at
zero frequency.
(b) Variance. The Cramér-Rao lower bound (CRLB) gives the minimum variance that
can be attained by unbiased estimators. The Slepian-Bangs formula (see [24]) directly
gives the following CRLB for component θn,i
(
CRLB(θn,i ) = 2 Trace
C(Θn )
−1 ∂C(Θn )
∂θn,i
2 )!−1
.
This bound gives information about the variance of the estimator at convergence of the
algorithm, i.e. when both SX and the other parameters are well estimated.
Applying this formula to θn,1 gives
2
n
2θn,1
2 o
≥ CRLB(θn,1 ) =
E θ̃n,1 − E θ̃n,1
.
Ms
This implies that the number of scales Ms of the wavelet transform must be large enough
to yield an estimator with sufficiently small variance.
For θn,2 , no closed-form expression is available for the CRLB. Therefore, the evaluation
of this bound and its comparison with the variance of the estimator θ̃n,2 can only be
based on numerical results, see section IV.
(c) Robustness to noise. Assume now the observations are corrupted by a random
2 (supposed to be known). The model becomes
Gaussian white noise W with variance σW
Y = A a Dγ X + W .
(25)
The estimator θ̃n,1 is not robust to noise. Indeed, if the maximum likelihood estimator
(k)
of model (7) in the presence of such white noise, a new term bn,1| (θn,1 ) must be added
W
to the bias expression (23), which becomes
1
(k) −1
(k)
(k)
Cwn ,
Trace C̃0 θ̃n,2
bn,1| (θn,1 ) =
W
Ms
R
2 q(si + s j )/2 ∞ ψ̂ (qsi ξ )ψ̂ (qs j ξ )dξ. In practice, this term can take large
where (Cwn )ij = σW
0
values, therefore noise has to be taken into account. To do so, the covariance matrix is
now written as
C(Θn)ij = q
si +s jZ
2
∞
0
(θn,2 SX (q−θn,1 ξ ) + σW2 )ψ̂ (qsi ξ )ψ̂ (qs j ξ )dξ ,
(26)
15
and the likelihood is modified accordingly. Formula (17) is no longer true and no closedform expression can be derived any more, the maximum likelihood estimate θ̃n,1 must
be computed by a numerical scheme (here we use a simple gradient ascent).
The estimator θ̃n,2 is very robust to noise. Indeed, equation (26) shows that the only
change in the covariance matrix formula is to replace the power spectrum SX by SZ =
σ2
W
. The additive constant term is not impairing the estimator as long as it is small
SX + θn,2
in comparison with the maximum values of SX .
Moreover, the estimator S˜X is modified because when computing
on scale sm , we compute:
wz,sm = w x,sm + ww∗ ,sm ,
1 f
Ww (sm − θ2 , τ ) is the wavelet transform of a
θ1/2
1
by a−1 . Thus a constant term σ̃W independent of
1 f
1/2 W y
θn,1
(s − θn,2 , τn )
where ww∗ ,sm =
white noise modulated
in amplitude
frequency is added to
the new spectrum estimator S˜Z , so that
2
E S˜Z = SX,ψ + σ̃W
where
2
2
σ̃W
= σW
1
Nτ
Nτ
∑
1
θ
n =1 n,1
.
D. Extension: estimation of other deformations
To describe other non-stationary behaviors of audio signals, other operators can be
investigated. For example, combination of time warping and frequency modulation can
be considered, as was done in [21], we shortly account for this case here for the sake
of completeness. Let α ∈ C2 be a smooth function, and set
Mα :
Mα x (t) = e2iπα(t) x (t) ,
(27)
The deformation model in [21] is of the form
Y = A a M α Dγ X .
(28)
To perform joint estimation of amplitude and frequency modulation and time warping
for each time, a suitable time-scale-frequency transform V is introduced, defined as
V X (s, ν, τ ) = hX, ψsντ i, with ψsντ = Tτ Mν Ds ψ. In that case, an approximation theorem
similar to 1 can be obtained from which the corresponding log-likelihood can be written.
At fixed time τ, the strategy of estimation is the same as before, but the parameter
space is of higher dimension, and the extra parameter θ3 = α′ (τ ) complicates the loglikelihood maximization. In particular, the choice of the discretization of the two scale
16
and frequency variables s and ν influences performances of the estimator, in particular
the Cramér-Rao bound.
IV. N UMERICAL
RESULTS
We now turn to numerical simulations and applications. A main ingredient is the
choice of the wavelet transform. Here we shall always use the sharp wavelet ψ♯ defined
in (2) and set the scale constant q to q = 2.
We will systematically compare our approach to simple estimators for amplitude
modulation and time warping, commonly used in applications, defined below.
•
Amplitude modulation: we use as baseline estimator of a(τn )2 the average energy
(B)
θ̃n,1 defined as follows:
(B)
θ̃n,1 =
1
kwy,τn k2 .
Ms
This amounts to replace the estimated covariance matrix in (17) by the identity
(B)
matrix. Notice that θ̃n,1 does not depend on the time warping estimator, and can
be computed directly on the observation.
•
(B)
Time warping: the baseline estimator θ̃n,2 is the scalogram scale center of mass
defined as follows:
(B)
θ̃n,2 = C0 +
1
kwy,τn k2
Ms
∑
s[m]|wy,τn [m]|2 .
m=1
(B)
C0 is chosen such that θ̃n,2 is a zero mean vector.
Numerical evaluation is performed on both synthetic signals and deformations and real
audio signals.
A. Synthetic signal
We first evaluate the performances of the algorithm on a synthetic signal. This allows
us to compare variance and bias with their theoretical values.
The simulated signal has length Nτ = 216 samples, sampled at Fs = 8 kHz (meaning the signal duration is t F = ( Nτ − 1)/Fs ≈ 8.2 s). The spectrum SX is written
(l )
(l )
(l )
(l )
if |ν − ν0 | < ∆ν /2
as SX = S1 + S2 where Sl (ν) = 1 + cos 2π (ν − ν0 )/∆ν
and vanishes elsewhere (for l ∈ {1, 2}). The amplitude modulation a is a sine wave
R
1 tF 2
a(t) = a0 (1 + a1 cos(2πt/T1 )), where a0 is chosen such that t−
F
0 a (t)dt = 1. The
17
3
2.5
2
1.5
1
0.5
0
0
1
2
3
4
5
6
7
8
5
6
7
8
Time (s)
1
0.5
0
-0.5
0
1
2
3
4
Time (s)
Fig. 1. Joint amplitude modulation/time warping estimation on a synthetic signal. Top: amplitude modulation
estimation (a1 = 0.4 and T1 = t F /3). Bottom: warping estimation (T2 = t F /2 and T3 = t F /2).
Estimation
Amplitude
Time
method
modulation
warping
Baseline
2.01 × 10−1
2.32 × 10−2
Algorithm 1
7.01 × 10−2
4.91 × 10−4
TABLE I
M EAN SQUARE ERROR OF THE ESTIMATION METHODS FOR BOTH DEFORMATIONS
time warping function γ is such that logq (γ′ (t)) = Γ + cos(2πt/T2 )e−t/T3 , where Γ is
R
1 tF ′
chosen such that t−
F
0 γ (t)dt = 1.
The estimation algorithm is implemented with Ms = 106 and p = 7. Results are
shown in Fig. 1 and compared with baseline estimations. For the sake of visibility,
the baseline estimator of the amplitude modulation (which is very oscillatory) is not
displayed, but numerical assessments are provided in Table I, which gives MSEs for the
different estimations. The proposed algorithm is clearly more precise than the baseline
algorithm, furthermore its precision is well accounted for by the Cramér-Rao bound: in
Fig. 1, the estimate is essentially contained within the 95 % confidence interval provided
by the CRLB (assuming gaussianity and unbiasedness).
The left hand side of Fig. 2 displays the estimated spectrum given by the algorithm
with formula (20). The agreement with the actual spectrum is very good, with a slight
enlargement effect due to to the filtering by |ψ̂|2 . The right hand side of Fig. 2 gives
18
10 0
2
1.8
10 -1
1.6
Relative update
Spectrum
1.4
1.2
1
0.8
0.6
10 -2
10
-3
10
-4
0.4
0.2
0
400
600
800
1000
1200
1400
1600
1
T =0.1 %
Time warping
Amplitude modulation
Spectrum
2
3
Frequency (Hz)
( 1)
Fig. 2. Left: Spectrum estimation (ν0
4
5
6
7
Iteration
( 1)
= 600 Hz, ∆ ν
( 2)
= 200 Hz, ν0
( 2)
= 1.2 kHz, ∆ ν
= 400 Hz): actual (dash-dot
blue line) and estimated (solid red line) spectra. Right: Stopping criterion evolution.
the evolution of the stopping criterion (21) with iterations. Numerical results show
that time warping estimation converges faster than amplitude modulation estimation.
Nevertheless, when fixing a stopping criterion to 0.1 % only 7 iterations are necessary
for the algorithm to converge.
B. Application to Doppler estimation
After studying the influence of the various parameters, let us now turn to a real life
audio example. The analyzed sound is a recording (from a fixed location) from a racing
car, moving with constant speed. The car engine sound is then deformed by the Doppler
effect, which results in time warping the sound emitted by the car. Besides, as the car is
moving, the closer the car to the microphone, the larger the amplitude of the recorded
sound. Thus, our model fits well this signal.
The wavelet transforms of the original signal and the two estimations of the underlying stationary signal are shown in Fig. 3. While the estimation of time warping
only corrects the displacement of wavelet coefficients in the time-scale domain, the joint
estimation of time warping and amplitude modulation also approximately corrects nonstationary variations of the amplitudes.
The physical relevance of the estimated time warping function can be verified. Indeed,
denote by V the (constant) speed of the car and by c the sound velocity. Fixing the time
origin to the time at which the car passes in front of the observer at distance d, the time
19
5513
2756
1378
689
345
172
0
0.5
1
1.5
Time (s)
2
0
0.5
1
1.5
2
Time (s)
1.15
1.1
1.05
1
0.95
0.9
0.85
0.2
0.4
0.6
0.8
1
Fig. 3. Doppler estimation. Top left: Scalogram of the original signal. Top right: Scalogram of the unwarped and
unmodulated signal. Bottom left: Scalogram of the unwarped signal. Bottom right: Estimated time warping compared
with the theoretical value given in (29).
warping function due to Doppler effect can be shown to be
c2
γ′ (t) = 2
c − V2
1− p
V 2t
d2 (c2 − V 2 ) + (cVt)2
!
(29)
.
We plot in Fig. 3 (bottom right) the estimation γ̃′ compared with its theoretical value
where d = 5 m and V = 54 m/s. Clearly the estimate is close to the corresponding
theoretical curve obtained with these data, which are therefore realistic values.
Nevertheless, a closer look at scalograms in Fig. 3 shows that the amplitude correction
is not perfect, due to the presence of noise, and the fact that the model is still too simple:
the amplitude modulation actually depends on frequency, which is not accounted for.
V. C ONCLUSIONS
We have discussed in this paper extensions of methods and algorithms described
earlier in [9], [22], [20], [21] for the joint estimation of deformation operator and power
spectrum for deformed stationary signals, a problem already addressed in [19] with
a different approach. The main improvements described in this paper concern the
following two points
1) the extension of the algorithm to the joint estimation of deformations including
amplitude modulation to the model and its estimation;
2) a statistical study of the estimators and of the performances of the algorithm.
20
The proposed approach was validated on numerical simulations and an application to
Doppler estimation.
The results presented here show that the proposed extensions result in a significant
improvement in terms of precision, and a better theoretical control. In particular, the continuous parameter estimation procedure avoids quantization effects that were present
in [20] where the parameter space was discrete and the estimation based on exhaustive
search. Regarding the approach of [19], its domain of validity seems to be limited to
small scales (i.e. high frequency) signals, which is not the case here.
Contrary to [19] our approach is based on (approximate) maximum likelihood estimation in the Gaussian framework. Because of our choice to disregard time correlations, the
estimates obtained here generally present spurious fluctuations, which can be smoothed
out by appropriate filtering. A natural extension of our approach would be to introduce
a smoothness prior that would avoid such filtering steps when necessary.
The code and datasets used to produce the numerical results of this paper are available
at the web site
https://cv.archives-ouvertes.fr/bruno-torresani
A PPENDIX A
P ROOF
OF THEOREM
1
√
To simplify notations, let Bγ denote the operator Dγ / γ′ . We split the approximation
error as
τ fτ
g
ε(s, τ ) = hAa Dγ X, ψsτ i − hA
a Dγ X, ψsτ i
q
(2)
(3)
(1)
′
= a(τ ) ε (s, τ ) + γ (τ ) ε (s, τ ) + ε (s, τ ) ,
where
(1)
p
q
γ′ − γ′ (τ ) Bγ X, ψsτ
q
p
′
′
γ − γ (τ ) ψsτ ,
= X, Bγ−1
D
E
(2)
τ
f
ε (s, τ ) =
Bγ − Bγ X, ψsτ
E
D
−1
τ
f
ψsτ ,
= X, Bγ−1 − Bγ
ε
(s, τ ) =
21
ε
(3)
E
−1
τ
g
− 1 Dγ X, ψsτ
(s, τ ) = Aa Aa
E
D
−1
τ
g
− 1 ψsτ .
= X, Dγ−1 Aa Aa
D
Furthermore, the triangle inequality gives:
r n
o
n
E |ε(s,τ)|
2
≤
Ca2
2
ε(1) (s,τ)
E
o
r
n
o
2
+ Cγ E ε(2) (s,τ)
!
r n
o 2
2
.
+ E ε(3) (s, τ )
(30)
Let us now determine an upper bound for each error term. To this end, they are
written as follows:
E
ε
(k)
(s, τ )
2
=
Z ∞
0
2
(k)
SX (ξ ) fˆsτ (ξ ) dξ ,
with k ∈ {1, 2, 3}.
√
Concerning the first error term, a Taylor expansion of γ′ around τ gives
Z q
q
(
1
)
fˆsτ (ξ ) =
γ′ (t) − γ′ (τ ) ψsτ (t)e−2iπγ(t)ξ dt
R
where Iψ =
write
R
≤
R
Z
R
γ′′
√ ′
2 γ
3s
∞
|t − τ | |ψsτ (t)|dt ≤ q 2
kγ′′ k∞
Iψ ,
√
2 cγ
|tψ(t)|dt. Furthermore, the localization assumption on ψ allows us to
Iψ ≤ 2
Z ∞
0
t
dt ≤ 2
1 + tβ
Z
1
0
tdt +
Z ∞
1
1
t β−1
dt
=
β
.
β−2
Finally, we can control the first error term as follows:
2
′′
2
βσX
3s/2 kγ k∞
(1)
.
≤ q
E ε (s, τ )
√
c γ 2( β − 2)
Concerning the second error term, we have
Z
′
(2)
ˆf sτ
e−2iπγ(t)ξ − e−2iπ (γ(τ )+(t−τ )γ (τ ))ξ ψsτ (t)dt
(ξ ) =
Z
R
′
1 − e−2iπ (γ(τ )+(t−τ )γ (τ )−γ(t))ξ |ψsτ (t)| dt
R
Z
π
≤
2 sin
ξ (t − τ )2 γ′′ (t∗ ) |ψsτ (t)| dt ,
2
R
≤
for some t∗ between t and τ. Besides, we have | sin(u)| ≤ |u| and | sin(u)| ≤ 1 so that:
Z
Z
π
(
2
)
′′
2s
2
s/2
ξ kγ k∞ q t |ψ(t)| dt +
fˆsτ (ξ ) ≤ 2q
| ψ(t)| dt .
R\ J
J2
22
where J = [− T, T ]. One can prove that the value of T minimizing the the right-hand
−1/( β+2)
. Therefore, we have:
side of the latter equation is T = π2 ξ kγ′′ k∞ q2s
4( β + 2)
(2)
fˆsτ (ξ ) ≤ q 2( β+2)
3( β − 1)
5β−2
π
2
′′
ξ kγ k∞
Finally, we can control the second error term as follows:
E
ε
(2)
(s, τ )
2
≤ q
5β−2
2( β +2)
β +2
β −1
.
β +2
4( β + 2)π ′′ β−1 (ρ)
kγ k∞
IX
3( β − 1) 2
!2
.
Concerning the third error term, we have
Z q
a
(
t
)
(
3
)
fˆsτ (ξ ) =
γ′ (t)
− 1 ψsτ (t)e−2iπγ(t)ξ dt
a
(
τ
)
R
Z
1 k a′ k
1
3s
k a′ k∞
∞
2
|t − τ | |ψsτ (t)|dt = q 2 Cγ2
Iψ ,
≤ Cγ
ca
R ca
Finally, we can control the third error term as follows:
!2
p
2
C
γ βσX
E ε(3) (s, τ )
≤ q3s/2
k a′ k∞
.
ca β − 2
To conclude the proof, the three errors terms in equation (30) are replaced by their
upper bounds to obtain the approximation error given in the theorem.
A PPENDIX B
P ROOF
OF PROPOSITION
1
(k)
Let w̃ x,sm ∈ C Nτ denote the estimation of w x,sm after k iterations of the algorithm.
(k)
(k)
1 f
q
Considering equation (18), we have w̃ x,s =
Wy sm − θ̃ , τ , thus
m
E
n
(k)
S˜X (q−sm ω0 )
o
2
(k)
θ̃1
1
=
E
Nτ kψk22
(k) 2
w̃ x,sm
.
(ξ )
To simplify notations, let us introduce some variables. We define sm = sm + logq (ξ )
and h( x ) = ϕ1 (qx ) = qx |ψ̂ (qx )|2 for x ∈ R.
By means of the covariance expression given in equation (12) we can write
Nτ
1
(
k
)
(
k
)
(k) 2
fy sm − θ̃ , τn
fy sm − θ̃ , τn W
E w̃x,sm =∑ (k)E W
n,2
n,2
n =1 θ̃n,1
Nτ
=
∑
θn,1
(k)
n =1 θ̃n,1
Z ∞
SX (ξ ) (ξ )
0
ξ
(k)
h sm + θn,2 − θ̃n,2 dξ
23
Z ∞
SX (ξ ) Nτ θn,1 (ξ )
=
ξ
0
∑
(k)
n =1 θ̃n,1
h sm
(k)
(k)
+ θn,2 − θ̃ dξ.
n,2
Let us now split the bias into two terms such that bSX (m) = g1 (m) + g2 (m), where g1
and g2 are defined as
Z
N −1 ∞SX (ξ ) Nτθn,1 (ξ )
(k)
−
1
+
θ
−
θ̃
h
s
g1 ( m ) = τ 2
n,2
m
n,2 dξ,
(k)
ξ
k ψ k2 0
θ̃
n =1 n,1
∑
N −1
g2 ( m ) = τ 2
k ψ k2
Z ∞
SX (ξ )
0
ξ
Nτ
∑
n =1
(k)
(ξ )
h sm + θn,2 − θ̃n,2
!
(ξ )
− h sm dξ.
Regarding the first term, we directly have
(k) Z
∞ S (ξ )
khk∞ 1 Nτ θn,1 − θ̃n,1
X
| g1 (m)| ≤
dξ
∑
2
(
k
)
ξ
kψk2 Nτ n=1
0
θ̃n,1
≤
khk∞
(k)
θ1 − θ̃1
2
k ψ k 2 c θ1
∞
JX .
Besides, we have khk∞ = k ϕ1 k∞ and the smoothness and decay assumptions on ψ̂ allow
us to write ϕ1 (u) = Ou→∞ (u1−2η ) → 0. Then ϕ1 is bounded and K1′ < ∞. This yields
u→∞
JX K1′
(k)
| g1 (m)| ≤
.
θ1 − θ̃1
2
∞
k ψ k2
(ξ )
Regarding the second term, a Taylor expansion of h around sm gives
Z
N −1 ∞SX(ξ ) Nτ (ξ )
(ξ )
(k)
+
θ
−
θ̃
| g2 (m)| ≤ τ 2
h
s
n,2
m
∑
n,2 − h s m dξ
ξ
k ψ k2 0
n =1
≤
≤
Furthermore, ∀ x ∈ R
kh′ k∞ 1
kψk22 Nτ
kh′ k
∞
kψk22
Nτ
∑
n =1
(k)
θn,2 − θ̃n,2
(k)
θ2 − θ̃2
∞
Z ∞
SX (ξ )
ξ
0
dξ
JX .
2
|h′( x )|ln (q)−1 = qxϕ1′ (qx ) ≤ qx ψ̂(qx ) + 2q2x ψ̂ (qx )ψ̂′(qx )
≤ k ϕ1 k∞ + 2kψ̂′ k∞ k ϕ2 k∞ = ln(q)−1 K2′ .
Besides, the decay assumption on ψ̂ gives | ϕ2 (u)| = Ou→∞ (u2−η ) → 0 because η > 2.
Then ϕ2 is bounded and K2′ < ∞. This yields
JX K2′
(k)
| g2 (m)| ≤
θ2 − θ̃2
2
k ψ k2
u→∞
∞
.
The proof is concluded by summing up the upper bounds of | g1 | and | g2 | to obtain the
(k)
upper bound of bSX . Notice that this bound does not depends on m.
24
R EFERENCES
[1] L. H. Koopmans, The spectral analysis of time series, 2nd ed., ser. Probability and mathematical statistics, Z. W.
Birnbaum and E. Lukacs, Eds. Elsevier, 1995, vol. 22.
[2] M. B. Priestley, Spectral analysis and time series, ser. Probability and mathematical statistics.
Academic Press,
1982, no. v. 1-2.
[3] ——, Non-linear and non-stationary time series analysis, 1988.
[4] S. Mallat, G. Papanicolaou, and Z. Zhang, “Adaptive covariance estimation of locally stationary processes,”
Ann. Statist., vol. 26, no. 1, pp. 1–47, 02 1998. [Online]. Available: http://dx.doi.org/10.1214/aos/1030563977
[5] G. P. Nason, R. von Sachs, and G. Kroisandt, “Wavelet processes and adaptive estimation of the evolutionary
wavelet spectrum,” Journal of the Royal Statistical Society. Series B (Statistical Methodology), vol. 62, no. 2, pp.
271–292, 2000. [Online]. Available: http://www.jstor.org/stable/3088859
[6] R. Dahlhaus, “A likelihood approximation for locally stationary processes,” Ann. Statist., vol. 28, no. 6, pp.
1762–1794, 12 2000. [Online]. Available: http://dx.doi.org/10.1214/aos/1015957480
[7] R. von Sachs and M. H. Neumann, “A wavelet-based test for stationarity,” Journal of Time Series Analysis,
vol. 21, no. 5, pp. 597–613, 2000. [Online]. Available: http://dx.doi.org/10.1111/1467-9892.00200
[8] S. Wisdom, L. Atlas, and J. Pitton, “Extending coherence time for analysis of modulated random processes,” in
2014 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), May 2014, pp. 340–344.
[9] H. Omer and B. Torrésani, “Estimation of frequency modulations on wideband signals; applications
to audio signal analysis,” in Proceedings of the 10th International Conference on Sampling Theory and
Applications (SampTA), G. Pfander, Ed.
Eurasip Open Library, 2013, pp. 29–32. [Online]. Available:
http://hal.archives-ouvertes.fr/hal-00822186
[10] S. T. Wisdom, “Improved statistical signal processing of nonstationary random processes using time-warping,”
Master’s thesis, School of Electrical Engineering, University of Washington, USA, March 2014.
[11] R. J. McAulay and T. F. Quatieri, “Speech analysis/synthesis based on a sinusoidal representation,” IEEE
Transactions on Acoustics, Speech, Signal Processing, vol. 34, pp. 744–754, August 1986.
[12] M. Lagrange, S. Marchand, and J. B. Rault, “Enhancing the tracking of partials for the sinusoidal modeling of
polyphonic sounds,” IEEE Transactions on Audio, Speech, and Language Processing, vol. 15, no. 5, pp. 1625–1634,
July 2007.
[13] P. Depalle, G. Garcı́a, and X. Rodet, “Tracking of partials for additive sound synthesis using hidden markov
models,” in Proceedings of the IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP-93),
vol. 1, April 1993, pp. 225–228.
[14] S. Godsill and M. Davy, “Bayesian harmonic models for musical pitch estimation and analysis,” in The IEEE
International Conference on Acoustics, Speech, and Signal Processing (ICASSP 02), 2002, pp. 1769–1772.
[15] R. Carmona, W. L. Hwang, and B. Torrésani, Practical time-frequency analysis: Gabor and Wavelet Transforms With
an Implementation in S, C. K. Chui, Ed. Academic Press, 1998.
[16] F. Auger, P. Flandrin, Y.-T. Lin, S. McLaughlin, S. Meignen, T. Oberlin, and H.-T. Wu, “Time-frequency
reassignment and synchrosqueezing: An overview,” IEEE Signal Processing Magazine, vol. 30, no. 6, pp. 32–41,
2013.
[17] H.-T. Wu, “Instantaneous frequency and wave shape function (i),” Applied and Computational Harmonic Analysis,
vol. 35, pp. 181–199, 2013.
25
[18] M. Clerc and S. Mallat, “The texture gradient equation for recovering shape from texture,” IEEE Transactions on
Pattern Analysis and Machine Intelligence, vol. 24, no. 4, pp. 536–549, 2002.
[19] ——, “Estimating deformations of stationary processes,” Ann. Statist., vol. 31, no. 6, pp. 1772–1821, 12 2003.
[Online]. Available: http://dx.doi.org/10.1214/aos/1074290327
[20] H. Omer and B. Torrésani, “Time-frequency and time-scale analysis of deformed stationary processes, with
application to non-stationary sound modeling,” Applied and Computational Harmonic Analysis, vol. 43, no. 1, pp.
1 – 22, 2017. [Online]. Available: https://hal.archives-ouvertes.fr/hal-01094835
[21] A. Meynard and B. Torrésani, “Spectral estimation for non-stationary signal classes,” in Sampling
Theory and Applications, ser. Proceedings of SampTA17, Tallinn, Estonia, Jul. 2017. [Online]. Available:
https://hal.archives-ouvertes.fr/hal-01475534
[22] H. Omer, “Modèles de déformation de processus stochastiques généralisés. application à l’estimation des non
stationnarités dans les signaux audio.” Ph.D. dissertation, Aix-Marseille Université, 2015.
[23] P. Welch, “The use of fast fourier transform for the estimation of power spectra: A method based on time
averaging over short, modified periodograms,” IEEE Transactions on Audio and Electroacoustics, vol. 15, no. 2, pp.
70–73, Jun 1967.
[24] P. Stoica and R. Moses, Introduction to Spectral Analysis. Prentice Hall, 1997.
| 10math.ST
|
Doubly Balanced Connected Graph Partitioning
arXiv:1607.06509v1 [math.CO] 21 Jul 2016
Saleh Soltan∗
Electrical Engineering
Columbia University
New York, NY
[email protected]
Mihalis Yannakakis†
Computer Science
Columbia University
New York, NY
[email protected]
Gil Zussman∗
Electrical Engineering
Columbia University
New York, NY
[email protected]
Abstract
We introduce and study the Doubly Balanced Connected graph Partitioning (DBCP) problem: Let G=(V, E) be a P
connected graph with a weight (supply/demand) function p:V →
{−1, +1} satisfying p(V )= j∈V p(j)=0. The objective is to partition G into (V1 , V2 ) such that
1 | |V2 |
G[V1 ] and G[V2 ] are connected, |p(V1 )|, |p(V2 )|≤cp , and max{ |V
|V2 | , |V1 | }≤cs , for some constants
cp and cs . When G is 2-connected, we show that a solution with cp =1 and cs =3 always exists
and can be found in polynomial time. Moreover, when G is 3-connected, we show that there
is always a ‘perfect’ solution (a partition with p(V1 )=p(V2 )=0 and |V1 |=|V2 |, if |V |≡0(mod 4)),
and it can be found in polynomial time. Our techniques can be extended, with similar results,
to the case in which the weights are arbitrary (not necessarily ±1), and to the case that p(V )6=0
and the excess supply/demand should be split evenly. They also apply to the problem of partitioning a graph with two types of nodes into two large connected subgraphs that preserve
approximately the proportion of the two types.
∗
Research partially supported by DTRA grant HDTRA1-13-1-0021 and CIAN NSF ERC under grant EEC0812072.
†
Research partially supported by NSF grant CCF-1320654.
1
Introduction
Power Grid Islanding is an effective method to mitigate cascading failures in power grids [17].
The challenge is to partition the network into smaller connected components, called islands, such
that each island can operate independently for a while. In order for an island to operate, it is
necessary that the power supply and demand at that island are almost equal.1 Equality of supply
and demand in an island, however, may not be sufficient for its independent operation. It is also
important that the infrastructure in that island has the physical capacity to safely transfer the
power from the supply nodes to the demand nodes. When the island is large enough compared to
the initial network, it is more likely that it has enough capacity. This problem has been studied in
the power systems community but almost all the algorithms provided in the literature are heuristic
methods that have been shown to be effective only by simulations [8, 14, 16, 17].
Motivated by this application, we formally introduce and study the Doubly Balanced Connected graph Partitioning (DBCP) problem: Let G=(V,
P E) be a connected graph with a weight
(supply/demand) function p:V →Z satisfying p(V )= j∈V p(j)=0. The objective is to partition V
|V1 | |V2 |
into (V1 , V2 ) such that G[V1 ] and G[V2 ] are connected, |p(V1 )|, |p(V2 )|≤cp , and max{ |V
,
}≤cs ,
2 | |V1 |
for some constants cp and cs . We also consider the case that p(V )6=0, in which the excess supply/demand should be split roughly evenly.
The problem calls for a partition into two connected subgraphs that simultaneously balances
two objectives, (1) the supply/demand within each part, and (2) the sizes of the parts. The
connected partitioning problem with only the size objective has been studied previously. In the
most well-known result, Lováz and Gyori [9,13] independently proved, using different methods, that
every k-connected graph can be partitioned into k arbitrarily sized connected subgraphs. However,
neither of the proofs is constructive, and there are no known polynomial-time algorithms to find
such a partition for k>3. For k=2, a linear time algorithm is provided in [18] and for k=3 an
O(|V |2 ) algorithm is provided in [20].2 The complexity of the problem with the size objective and
related optimization problems have been studied in [3, 5, 6] and there are various NP-hardness and
inapproximability results. Note that the size of the cut is not of any relevance here (so the extensive
literature on finding balanced partitions, not necessarily connected, that minimize the cut is not
relevant.)
The objective of balancing the supply/demand alone, when all p(i) are ±1, can also be seen as
an extension for the objective of balancing the size (which corresponds to p(i)=1). Our bi-objective
problem of balancing both supply/demand and size, can be seen also as an extension of the problem
of finding a partition that balances the size for two types of nodes simultaneously: Suppose the
nodes of a graph are partitioned into red and blue nodes. Find a partition of the graph into two
large connected subgraphs that splits approximately evenly both the red and the blue nodes.
We now summarize our results and techniques. Since the power grids are designed to withstand
a single failure (“N −1” standard) [1], and therefore 2-connected, our focus is mainly on the graphs
that are at least 2-connected. We first, in Section 3, study the connected partitioning problem
with only the supply/demand balancing objective, and show results that parallel the results for
balancing size alone, using similar techniques: The problem is NP-hard in general. For 2-connected
graphs and weights p(i)=±1, there is always a perfectly balanced partition and we can find it easily
using an st-numbering. For 3-connected graphs and weights p(i)=±1, there is a perfectly balanced
partition into three connected graphs, and we can find it using a nonseparating ear decomposition
1
If the supply and demand are not exactly equal but still relatively close, load shedding/generation curtailing can
be used in order for the island to operate.
2
For k=2, a much simpler approach than the one in [18] is to use the st-numbering [11] for 2-connected graphs.
1
of 3-connected graphs [4] and similar ideas as in [20].
The problem is more challenging when we deal with both balancing objectives, supply/demand
and size. This is the main focus and occupies the bulk of this paper. Our main results are existence
results and algorithms for 2- and 3-connected graphs. It is easy to observe that we cannot achieve
perfection in one objective (cp =0 or cs =1) without sacrificing completely the other objective. We
show that allowing the supply/demand of the parts to be off balance by at most the weight of one
node suffices to get a partition that is roughly balanced also with respect to size.
First, in Section 3.1, we study the case of 3-connected graphs since we use this later as the basis
of handling 2-connected graphs. We show that if ∀i, p(i)=±1, there is a partition that is perfectly
balanced with respect to both objectives, if |V |≡0(mod 4) (otherwise the sizes are slightly off for
parity reasons); for general p, the partition is perfect in both objectives up to the weight of a single
node. Furthermore, the partition can be constructed in polynomial time. Our approach uses the
convex embedding characterization of k-connectivity studied by Linial, Lováz, and Wigderson [12].
We need to adapt it for our purposes so that the convex embedding also has certain desired geometric
properties, and for this purpose we use the nonseparating ear decomposition of 3-connected graphs
of [4] to obtain a suitable embedding.
Then, in Section 3.2, we analyze the case of 2-connected graphs. We reduce it to two subcases:
either (1) there is a separation pair that splits the graph into components that are not very large,
or (2) we can perform a series of contractions to achieve a 3-connected graph whose edges represent
contracted subgraphs that are not too large. We provide a good partitioning algorithm for case (1),
and for case (2) we extend the algorithms for 3-connected graphs to handle also the complications
arising from edges representing contracted subgraphs. Finally, in Section 4, we briefly discuss the
problem of finding a connected partitioning of a graph with two types of nodes that splits roughly
evenly both types.
The body of the paper contains proofs and sketches for some of the results, and Appendices B, C,
and D contain the remaining proofs. Graph-theoretic background and definitions (e.g., the notions
of st-numbering, nonseparating ear decomposition, convex embedding of k-connected graphs) are
given in Appendix A.
2
Balancing the Supply/Demand Only
In this section, we study the single objective problem of finding a partition of the graph into
connected subgraphs that balances (approximately) the supply and demand in each part of the
partition, without any regard to the sizes of the parts. We can state the optimization problem
as follows, and will refer to it as the Balanced Connected Partition with Integer weights (BCPI)
problem.
Definition
1. Given a graph G=(V, E) with a weight (supply/demand) function p : V → Z satisP
fying j∈V p(j)=0. The BCPI problem is the problem of partitioning V into (V1 , V2 ) such that
1. V1 ∩V2 =∅ and V1 ∪V2 =V ,
2. G[V1 ] and G[V2 ] are connected,
P
3. |p(V1 )|+|p(V2 )| is minimized, where p(Vi )= j∈Vi p(j).
Clearly, the minimum possible value for |p(V1 )|+|p(V2 )| that we can hope for is 0, which occurs
iff p(V1 )=p(V2 )=0. It is easy to show that the problem of determining whether there exists such a
‘perfect’ partition (and hence the BCPI problem) is strongly NP-hard. The proof is very similar
2
to analogous results concerning the partition of a graph into two connected subgraphs with equal
sizes (or weights, when nodes have positive weights) [3, 6]
Proposition 1. (1) It is strongly NP-hard to determine whether there is a solution to the BCPI
problem with value 0, even when G is 2-connected.
(2) If G is not 2-connected, then this problem is NP-hard even when ∀i, p(i)=±1.
Although it is NP-hard to tell whether there is a solution satisfying p(V1 )=p(V2 )=0, even when
∀i, p(i)=±1, in this case, if the graph G is 2-connected there is always such a solution. For general
weights p, there is a solution such that |p(V1 )|, |p(V2 )|≤maxj∈V |p(j)|/2 and it can be found easily
in linear time using the st-numbering between two nodes (see the Appendix B).
Proposition 2. Let G be a 2-connected graph and u, v any two nodes in V such that p(u)p(v)>0.
(1) There is a solution such that u∈V1 , v∈V2 , and |p(V1 )|=|p(V2 )|≤maxj∈V |p(j)|/2.
(2) If ∀i, p(i)=±1, we can find a solution such that u∈V1 , v∈V2 , and p(V1 )=p(V2 )=0.
In both cases, the solution can be found in O(|E|) time.
Remark. The bound in Proposition 2 (1) is tight. A simple example is a cycle of length 4 like
(v1 , v2 , v3 , v4 ) with p(v1 )=−p, p(v2 )=−p/2, p(v3 )=p, and p(v4 )=p/2. It is easy to see that in this
example |p(V1 )|+|p(V2 )|=maxj∈V |p(j)|=p is the best that one can do.
2.1
Connected Partitioning into Many Parts
The BCPI problem can be extended to partitioning a graph
P into k=3 or more parts. Let G=(V, E)
be a graph with a weight function p : V → Z satisfying j∈V p(j)=0. The BCPIk problem is the
problem of partitioning G into (V1 , V2 , . . . , Vk ) such that for any 1≤i≤k, G[Vi ] is connected and
P
k
i=1 |p(Vi )| is minimized.
In the following proposition, we show that for k=3, if p(i)=±1, ∀i, then there is always a perfect
partition (i.e., with p(V1 )=p(V2 )=p(V3 )=0) and it can be found efficiently. For general p, we can
find a partition such that |p(V1 )|+|p(V2 )|+|p(V3 )|≤2 maxj∈V |p(j)|. The proof and the algorithm use
a similar approach as the algorithm in [20] for partitioning a 3-connected graph to three connected
parts with prescribed sizes, using the nonseparating ear decomposition of 3-connected graphs as
described in Subsection A.4. The proof is given in the Appendix B.
Proposition 3. Let G be a 3-connected graph and u, v, w three nodes in V such that p(u), p(v), p(w)>
0 or p(u), p(v), p(w)<0.
(1) There is a solution such that u∈V1 , v∈V2 , w∈V3 , and |p(V1 )|+|p(V2 )|+|p(V3 )|≤2 maxj∈V |p(j)|.
(2) If ∀i, p(i)=±1, then there is a solution such that u∈V1 , v∈V2 , w∈V3 , and |p(V1 )|=|p(V2 )|=
|p(V3 )|=0.
In both cases, the solution can be found in O(|E|) time.
3
Balancing Both Objectives
In this section, we formally define and study the Doubly Balanced Connected graph Partitioning
(DBCP) problem.
Definition 2.PGiven a graph G=(V, E) with a weight (supply/demand) function p : V → Z satisfying p(V )= j∈V p(j)=0 and constants cp ≥0, cs ≥1. The DBCP problem is the problem of
partitioning V into (V1 , V2 ) such that
1. V1 ∩V2 =∅ and V1 ∪V2 =V ,
3
2𝑠 + 1
−1
−1
−1
−1
+1
+1
+1
−1
−1
−1
+1
+1
+1
−1
−1
−1
+1
+1
+1
−1
−1
+1
+1
+1
−1
2𝑡 + 1
+1
2𝑡 + 1
Figure 1: Series-parallel graphs with 2s+1 paths of length 4t+2 used in Observations 1 and 2.
2. G[V1 ] and G[V2 ] are connected,
|V1 | |V2 |
3. |p(V1 )|, |p(V2 )|≤cp and max{ |V
,
}≤cs , where p(Vi )=
2 | |V1 |
P
j∈Vi
p(j).
Remark. Our techniques apply also to the case that p(V )6=0. In this case, the requirement 3
on p(V1 ) and p(V2 ) is |p(V1 )−p(V )/2|, |p(V2 )−p(V )/2|≤cp , i.e., the excess supply/demand is split
approximately evenly between the two parts.
We will concentrate on 2-connected and 3-connected graphs and show that we can construct
efficiently good partitions. For most of the section we will focus on the case that p(i)=±1, ∀i∈V .
This case contains all the essential ideas. All the techniques generalize to the case of arbitrary p,
and we will state the corresponding theorems.
We observed in Section 2 that if the graph is 2-connected and p(i)=±1, ∀i∈V then there is
always a connected partition that is perfect with respect to the weight objective, p(V1 )=p(V2 )=0,
i.e., (3) is satisfied with cp =0. We know also from [9, 13] that there is always a connected partition
that is perfect with respect to the size objective, |V1 |=|V2 |, i.e., condition 3 is satisfied with cs =
1. The following observations show that combining the two objectives makes the problem more
challenging. If we insist on cp =0, then cs cannot be bounded in general, (it will be Ω(|V |)), and
if we insist on cs =1, then cp cannot be bounded. The series-parallel graphs of Figure 1 provide
simple counterexamples.
Observation 1. If cp =0, then for any cs <|V |/2−1, there exist a 2-connected graph G such that
the DBCP problem does not have a solution even when ∀i, p(i)=±1.
Proof. In the graph depicted in Figure 1, set t=0.
Observation 2. If cs =1, then for any cp <|V |/6, there exist a 2-connected graph G such that the
DBCP problem does not have a solution even when ∀i, p(i)=±1.
Proof. In the graph depicted in Figure 1, set s=1.
Thus, cp has to be at least 1 to have any hope for a bounded cs . We show in this section that
cp =1 suffices for all 2-connected graphs. We first treat 3-connected graphs.
3.1
3-Connected Graphs
Let G=(V, E) be a 3-connected graph. Assume for the most of this section that ∀i, p(i)=±1 and
p(V )=0 (we will state the results for general p at the end). We show that G has a partition that is
essentially perfect with respect to both objectives, i.e., with cp =0 and cs =1. We say “essentially”,
4
because p(V1 )=p(V2 )=0 and |V1 |=|V2 | imply that |V1 |=|V2 | are even, and hence V must be a
multiple of 4. If this is the case, then indeed we can find such a perfect partition. If |V |≡2(mod 4)
(|V | has to be even since p(V )=0), then we can find an ‘almost perfect’ partition, one in which
|p(V1 )|=|p(V2 )|=1 and |V1 |=|V2 | (or one in which p(V1 )=p(V2 )=0 and |V1 |=|V2 |+2).
We first treat the case that G contains a triangle (i.e., cycle of length 3). In the following Lemma,
we use the embedding for k-connected graphs introduced in [12] and as described in Subsection A.6,
to show that if G is 3-connected with a triangle and all weights are ±1, then the DBCP problem
has a perfect solution.
Lemma 1. If G is 3-connected with a triangle, ∀i, p(i)=±1, and |V |≡0(mod 4), then there exists a
solution to the DBCP problem with p(V1 )=p(V2 )=0 and |V1 |=|V2 |. If |V |≡2(mod 4), then there is a
solution with p(V1 )=p(V2 )=0 and |V1 |=|V2 |+2. Moreover, this partition can be found in polynomial
time.
Proof. Assume that |V |≡0(mod 4); the proof for the case |V |≡2(mod 4) is similar. In [12] as
described in Subsection A.6, it is proved that if G is a k-connected graph, then for every X⊂V
with |X|=k, G has a convex X-embedding in general position. Moreover, this embedding can be
found by solving a set of linear equations of size |V |. Now, assume v, u, w∈V form a triangle in G.
Set X={v, u, w}. Using the theorem, G has a convex X-embedding f :V →R2 in general position.
Consider a circle C around the triangle f (u), f (v), f (w) in R2 as shown in an example in Fig. 2.
Also consider a directed line L tangent to the circle C at point A. If we project the nodes of G
onto the line L, since the embedding is convex and also {u, v}, {u, w}, {w, v}∈E, the order of the
nodes’ projection gives an st-numbering between the first and the last node (notice that the first
and last nodes are always from the set X). For instance in Fig. 2, the order of projections give an
st-numbering between the nodes u and v in G. Hence, if we set V1 to be the |V |/2 nodes whose
projections come first and V2 are the |V |/2 nodes whose projections come last, then G[V1 ] and G[V2 ]
are both connected and |V1 |=|V2 |=|V |/2. The only thing that may not match is p(V1 ) and p(V2 ).
Notice that for each directed line tangent to the circle C, we can similarly get a partition such that
|V1 |=|V2 |=|V |/2. So all we need is a point D on the circle C such that if we partition based on the
directed line tangent to C at point D, then p(V1 )=p(V2 )=0. To find such a point, we move L from
being tangent at point A to point B (AB is a diameter of the circle C) and consider the resulting
partition. Notice that if at point A, p(V1 )>0, then at point B since V1 and V2 completely switch
places compared to the partition at point A, p(V1 )<0. Hence, as we move L from being tangent at
point A to point B and keep it tangent to the circle, in the resulting partitions, p(V1 ) goes from
some positive value to a non-positive value. Notice that the partition (V1 , V2 ) changes only if L
passes a point D on the circle such that at D, L is perpendicular to a line that connects f (i) to
f (j) for some i, j∈V . Now, since the embedding is in general position, there are exactly two points
on every line that connects two points f (i) and f (j), so V1 changes at most by one node leaving V1
and one node entering V1 at each step as we move L. Hence, p(V1 ) changes by either ±2 or 0 value
at each change. Now, since |V |≡0(mod 4), p(V1 ) has an even value in all the resulting partitions.
Therefore, as we move L from being tangent at point A to point B, there must be a point D such
that in the resulted partition p(V1 )=p(V2 )=0.
It is also easy to see that since V1 may change only when a line that passes through 2 nodes of
graph G is perpendicular to L, we can find D in at most O(|V |2 ) steps.
When G is a triangle-free 3-connected graph, however, the proof in Lemma 1 cannot be directly
used anymore. The reason is if for example {u, v}∈E
/ and we project the nodes of G onto the line
L, this time the order of the nodes projection does not give an st-numbering between the first and
the last node if for example u and w are the first and last node, since some of the middle nodes may
5
𝑉2
𝑉1
𝑢′
ℒ
𝑤′
𝑣′
𝐴
𝑤
𝒞
𝑢
𝑣
𝐵
Figure 2: Proof of Lemma 1.
only be connected to v. To prove a similar result for triangle-free 3-connected case, we first provide
the following two Lemmas. The main purpose of the following two Lemmas are to compensate for
the triangle-freeness of G in the proof of Lemma 1. The idea is to show that in every 3-connected
graph, there is a triple {u, w, v}∈V , such that {u, w}, {w, v}∈E and in every partition that we get
by the approach used in the proof of Lemma 1, if u and v are in Vi , so is a path between u and v.
Lemma 2. If G is 3-connected, then there exists a set {u, v, w}∈V and a partition of V into
(V10 , V20 ) such that:
1. V10 ∩V20 =∅ and V10 ∪V20 =V ,
2. G[V10 ] and G[V20 ] are connected,
3. {u, w}, {v, w}∈E,
4. w∈V10 , u, v∈V20 ,
5. |V20 |≤|V |/2.
Moreover, such a partition and {u, v, w} can be found in O(|E|) time.
Proof. Using the algorithm presented in [4], we can find a non-separating cycle C0 in G such that
every node in C0 has a neighbor in G\C0 . Now, we consider two cases:
(i) If |C0 |≤|V |/2+1, then select any three consecutive nodes (u, w, v) of C0 and set V20 =C0 \{w}
and V10 =V \V20 .
(ii) If |C0 |>|V |/2+1, since every node in C0 has a neighbor in G\C0 , there exists a node w∈V \C0
such that |N (w)∩C0 |≥2. Select two nodes u, v∈N (w)∩C0 . There exists a path P in C0
between u and v such that |P |<|V |/2−1. Set V20 ={u, v}∪P and V10 =V \V20 .
Lemma 3. Given a partition (V10 , V20 ) of a 3-connected graph G with properties described in
Lemma 2, G has a convex X-embedding in general position with mapping f :V →R2 such that:
1. X={u, w, v}, f (u)=(0, 0), f (v)=(1, 0), and f (w)=(0, 1),
2. Every node i in V10 is mapped to a point (f1 (i), f2 (i)) with f2 (i)≥1/2,
3. Every node i in V20 is mapped to a point (f1 (i), f2 (i)) with f1 (i)≥f2 (i) and f1 (i)+2f2 (i)≤1.
6
ℒ
𝐴
𝑤
ℒ3
(0,1)
ℒ2
ℒ1
𝒞
𝑉1 ′
(0.5,0.5)
(0,0.5)
𝑉2 ′
(0,0)
(1,0)
𝑣
𝑢
𝐵
Figure 3: Proof of Lemma 3 and Theorem 1.
Moreover, such an embedding can be found in in polynomial time.
Sketch of the proof. Set X={v, u, w}. Using [12], G has a convex X-embedding in R2 in general
position with mapping f :V →R2 such that f (u)=(0, 0), f (v)=(1, 0), and f (w)=(0, 1). In the Xembedding of the nodes, we have a freedom to set the elasticity coefficient vector ~c to anything that
we want (except a measure zero set of vectors). So for any edge {i, j}∈G[V1 ]∪G[V2 ], set cij =g; and
for any {i, j}∈E[V10 , V20 ], set cij =1. Since both G[V10 ] and G[V20 ] are connected, as we increase g,
nodes in V10 get closer to w and nodes in V20 get closer to the line uv (as g→∞, nodes in V10 get in
the same position as w and node in V20 get on the line uv). Hence, intuitively there exists a value
g, for which all the nodes in V1 are above line L1 and all the nodes in V20 are below the lines L2
and L3 as depicted in Fig. 3. In the appendix we give the detailed proof and show that a g with
polynomially many bits suffices.
Using Lemmas 2 and 3, we are now able to prove that for any 3-connected graph G such that
all the weights are ±1, the DBCP problem has a solution for cp =0 and cs =1. The idea of the proof
is similar to the proof of Lemma 1, however, we need to use Lemma 2 to find a desirable partition
(V10 , V20 ) and then use this partition to find an embedding with properties as described in Lemma 3.
By using this embedding, we can show that in every partition that we obtain by the approach in
the proof of Lemma 1, if u and v are in Vi , so is a path between u and v. This implies then that
G[V1 ] and G[V2 ] are connected. So we can use similar arguments as in the proof of Lemma 1 to
prove the following theorem (see the Appendix C for the proof details).
Theorem 1. If G is 3-connected, ∀i, p(i)=±1, and |V |≡0(mod 4), then there exists a solution to
the DBCP problem with p(V1 )=p(V2 )=0 and |V1 |=|V2 |. If |V |≡2(mod 4), then there is a solution
with p(V1 )=p(V2 )=0 and |V1 |=|V2 |+2. Moreover, this partition can be found in polynomial time.
It is easy to check for a 3-connected graph G, by using the same approach as in the proof of
Lemma 1 and Theorem 1, that even when the weights are arbitrary (not necessarily ±1) and also
p(V )6=0, we can still find a connected partition (V1 , V2 ) for G such that |p(V1 )−p(V )/2|, |p(V1 )−
p(V )/2|≤maxi∈V |p(i)| and |V1 |=|V2 |.
Corollary 1. If G is 3-connected, then the DBCP problem (with arbitrary p, and not necessarily
satisfying p(V )=0) has a solution for cp =maxi∈V |p(i)| and cs =1. Moreover, this solution can be
found in polynomial time.
7
3.2
2-Connected Graphs
We first define a pseudo-path between two nodes in a graph as below. The definition is inspired by
the definition of the st-numbering.
Definition 3. A pseudo-path between nodes u and v in G=(V, E), is a sequence of nodes v1 , . . . , vt
such that if v0 =u and vt+1 =v, then for any 1≤i≤t, vi has neighbors vj and vk such that j<i<k.
Using the pseudo-path notion, in the following lemma we show that if G is 2-connected and
has a separation pair such that none of the resulting components are too large, then the DBCP
problem always has a solution for some cp =cs =O(1). The idea used in the proof of this lemma is
one of the building blocks of the proof for the general 2-connected graph case.
Lemma 4. Given a 2-connected graph G and an integer q>1, if ∀i:p(i)=±1 and G has a separation pair {u, v}⊂V such that for every connected component Gi =(Vi , Ei ) of G[V \{u, v}], |Vi |<
(q−1)|V |/q, then the DBCP problem has a solution for cp =1, cs =q−1, and it can be found in
O(|E|) time.
Proof. Assume for simplicity that |V | is divisible by q. There is a separation pair {u, v}∈V such
that if G1 , . . . , Gk are the connected components of G\{u, v}, for any i, |Vi |<(q−1)|V |/q. Since G
is 2-connected, G1 , . . . , Gk can be presented by pseudo-paths P1 , . . . , Pk between u and v. Assume
P1 , . . . , Pk are in increasing order based on their lengths. There exists
two subsets of the pseudoP
paths S1 and S2 such that S1 ∩S2 =∅, S1 ∪S2 ={P1 , . . . , Pk } and Pj ∈Si |Pj |≥|V |/q−1 for i=1, 2.
The proof is very simple. If |Pk |<|V |/q, the greedy algorithm for the partition problem gives the
desired partition of {P1 , . . . , Pk }. If |Pk |≥|V |/q, since |Pk |<(q−1)|V |/q, S1 ={P1 , . . . , Pk−1 } and
S2 ={Pk } is the desired partition.
Now, if we put all the pseudo-paths in S1 back to back, they will form a longer pseudo-path Q1
between u and v. Similarly, we can form another pseudo-path Q2 from the paths in S2 . Without
loss of generality we can assume |Q1 |≥|Q2 |. From u, including u itself, we count |V |/q of the nodes
in Q1 towards v and put them in a set V 0 . Without loss of generality, we can assume p(V 0 )≥0.
If p(V 0 )=0, then (V 0 , V \V 0 ) is a good partition and we are done. Hence, assume p(V 0 )>0. We
keep V 0 fixed and make a new set V 00 by continuing to add nodes from Q1 to V 0 one by one until
we get to v. If p(V 00 ) hits 0 as we add nodes one by one, we stop and (V 00 , V \V 00 ) is a good
partition and we are done. So, assume V 00 =Q1 ∪{u} and p(V 00 )>0. Since |Q2 ∪{v}|≥|V |/q, |V 00 |≤
(q−1)|V |/q. If |V 00 |<(q−1)|V |/q, we add nodes from Q2 one by one toward u until either |V 00 |=0
or |V 00 |=(q−1)|V |/q. If we hit 0 first (i.e., p(V 00 )=0) and |V 00 |≤(q−1)|V |/q, define V1 =V 00 \{u},
then (V1 , V \V1 ) is a good partition. So assume |V 00 |=(q−1)|V |/q and p(V 00 )>0. Define V 000 =V \V 00 .
Since p(V 00 )>0 and |V 00 |=(q−1)|V |/q, then p(V 000 )<0 and |V 000 |=|V |/q. Also notice that V 000 ⊆Q2 .
We consider two cases. Either |p(V 0 )|≥|p(V 000 )| or |p(V 0 )|<|p(V 000 )|.
If |p(V 0 )|≥|p(V 000 )|, then if we start from u and pick nodes one by one from Q1 in order, we can
get a subset V10 of V 0 such that |p(V10 )|=|p(V 000 )|. Hence, if we define V1 =V10 ∪V 000 , then (V1 , V \V1 )
is a good partition.
If |p(V 0 )|<|p(V 000 )|, then we can build a new set V1 by adding nodes one by one from Q2 to V 0
until P (V1 )=0. It is easy to see that since |p(V 0 )|<|p(V 000 )|, then V1 \V 0 ⊆V 000 . Hence, (V1 , V \V1 ) is
a good partition.
Corollary 2. If G is a 2-connected series-parallel graph and ∀i:p(i)=±1, then the DBCP problem
has a solution for cp =1, cs =2, and the solution can be found in O(|E|) time.
The graph in Figure 1 with s=1 shows that these parameters are the best possible for series
parallel graphs: if cp =O(1) then cs must be at least 2.
8
𝑣1
𝑣3
𝑣4
𝑣5
𝑣2
𝐺1
𝑣3
𝑣6
𝐺2
𝐺3
𝑣7
𝑣7
𝐺
(a)
𝑣1
𝑣13
𝑣12
𝑣11
𝑣2
𝑣3
𝑣4
𝑣7
𝑣8
𝑣9
𝑣5
𝑣4
𝑣10
𝑣14 𝑣15
𝑣1
𝑣13
𝑣12
𝑣11
𝑣2
𝑣3
𝑣13
𝑣12
𝑣11
3
𝑣7
𝑣7
3
3
𝑣6
𝐺
𝑣1
𝑣5
𝑣6
𝑣14 𝑣15
𝑣5
𝑣6
𝑣14 𝑣15
𝐺∗
(b)
Figure 4: Lemma 5.
To generalize Lemma 4 to all 2-connected graphs, we need to define the contractible subgraph
and the contraction of a given graph as below.
Definition 4. We say an induced subgraph G1 of a 2-connected graph G is contractible, if there
is a separating pair {u, v}⊂V such that G1 =(V1 , E1 ) is a connected component of G[V \{u, v}].
Moreover, if we replace G1 by a weighted edge e0 with weight w(e0 )=|V1 | between the nodes u and v
in G to obtain a smaller graph G0 , we say G is contracted to G0 .
Remark. Notice that every contractible subgraph of a 2-connected graph G can also be represented
by a pseudo-path between its associated separating pair. We use this property in the proof of Theorem 2.
Using the notion of the graph contraction, in the following lemma, we show that to partition
a 2-connected graph, we can reduce it to one of two cases: either G can be considered as a graph
with a set of short pseudo-paths between two nodes, or it can be contracted into a 3-connected
graph as illustrated in Fig. 4.
Lemma 5. In every 2-connected graph G=(V, E), given an integer q≥3, one of the following cases
holds, and we can determine which in O(|E|) time:
1. There is a separation pair {u, v}⊂V such that if G1 , . . . , Gk are the connected components of
G[V \{u, v}], for all i, |Vi |<(q−1)|V |/q.
2. After a set of contractions, G can be transformed into a 3-connected graph G∗ =(V ∗ , E ∗ ) with
weighted edges representing contracted subgraphs such that for every e∗ ∈E ∗ , w(e∗ )<|V |/q.
Proof. If there is no separation pairs in G, then G is 3-connected and there is nothing left to
prove. So assume {u, v}⊂V is a separation pair and G1 , . . . , Gk are the connected components of
G[V \{u, v}]. If ∀i, |Vi |<(q−1)|V |/q, we are done. So let’s assume there is a connected component
9
Gj such that |Vj |≥(q−1)|V |/q. Then for every i6=j, Gi can be contracted and represented by an
edge of weight less than |V |/q between u and v. Now, we repeat the process by considering the
weight of the edges in the size of each connected component. An example for each case is shown in
Fig. 4 for q=3. We can find either a suitable separation pair as in case 1 or a suitable contracted
graph G∗ as in case 2 in linear time using the Hopcroft-Tarjan algorithm for finding the triconnected
components [10].
Using Lemma 5 for q=4, then Lemma 4, and the idea of the proof for Theorem 1, we can
prove that when G is 2-connected and all p(i)=±1, the DBCP problem has a solution for cp =1 and
cs =3. There are some subtleties in adapting Lemma 2 for this case to account for the fact that the
edges of G∗ are now weighted, and the partition (V10 , V20 ) has to take into account the edge weights.
We find a suitable convex embedding of the 3-connected graph G∗ and then embed the nodes of
the contracted pseudopaths appropriately along the segments corresponding to the weighted edges.
Some care is needed to carry out the argument of the 3-connected case, since as the line tangent to
the circle rotates, the order of the projections of many nodes may change at once, namely the nodes
on an edge perpendicular to the rotating line. The details of the proof are given in Appendix C.
We have:
Theorem 2. If G is 2-connected, ∀i, p(i)=±1, then the DBCP problem has a solution for cp =1
and cs =3. Moreover, this solution can be found in polynomial time.
Similar to Corollary 1, the approach used in the proof of Theorem 2, can also be used for the
case when the weights are arbitrary (not necessarily ±1) and p(V )6=0. It is easy to verify that
in this case, if G is 2-connected, the DBCP problem has a connected partition (V1 , V2 ) such that
|p(V1 )−p(V )/2|, |p(V2 )−p(V )/2|≤maxj∈V |p(j)| and |V1 |, |V2 |≥|V |/4.
Corollary 3. If G is 2-connected, then the DBCP problem (with general p and not necessarily
satisfying p(V )=0) has a solution for cp =maxj∈V |p(j)| and cs =3. Moreover, this solution can be
found in polynomial time.
4
Graphs with Two Types of Nodes
Assume G is a connected graph with nodes colored either red (R⊆V ) or blue (B⊆V ). Let |V |=n,
|R|=nr , and |B|=nb . If G is 3-connected, set p(i)=1 if i∈R and p(i)=−1 if i∈B. Corollary 1 implies
then that there is always a connected partition (V1 , V2 ) of V that splits both the blue and the red
nodes evenly (assuming nr and nb are both even), i.e., such that |V1 |=|V2 |, |R∩V1 |=|R∩V2 |, and
|B∩V1 |=|B∩V2 |. (If nr and/or nb are not even, then one side will contain one more red or blue
node.)
Corollary 4. Given a 3-connected graph G with nodes colored either red (R⊆V ) or blue (B⊆
V ). There is always a partition (V1 , V2 ) of V such that G[V1 ] and G[V2 ] are connected, |V1 |=|V2 |,
|R∩V1 |=|R∩V2 |, and |B∩V1 |=|B∩V2 | (assuming |R| and |B| are both even). Such a partition can
be computed in polynomial time.
If G is only 2-connected, we may not always get a perfect partition. Assume wlog that nr ≤nb .
If for every v∈R and u∈B, we set p(v)=1 and p(u)=−nr /nb , Corollary 3 implies that there is
always a connected partition (V1 , V2 ) of V such that both |(|R∩V1 |−nr /nb |B∩V1 |)|≤1 and |(|R∩
1 | |V2 |
V2 |−nr /nb |B∩V2 |)|≤1, and also max{ |V
|V2 | , |V1 | }≤3. Thus, the ratio of red to blue nodes in each side
Vi differs from the ratio nr /nb in the whole graph by O(1/n). Hence if the numbers of red and blue
nodes are ω(1), then the two types are presented in both sides of the partition in approximately
the same proportion as in the whole graph.
10
Corollary 5. Given a 2-connected graph G with nodes colored either red (R⊆V ) or blue (B⊆V ),
and assume wlog |R|≤|B|. We can always find in polynomial time a partition (V1 , V2 ) of V such
that G[V1 ] and G[V2 ] are connected, |V1 |, |V2 |≥|V |/4, and the ratio of red to blue nodes in each side
Vi differs from the ratio |R|/|B| in the whole graph by O(1/n).
5
Conclusion
In this paper, we introduced and studied the problem of partitioning a graph into two connected
subgraphs that satisfy simultaneouly two objectives: (1) they balance the supply and demand within
each side of the partition (or more generally, for the case of p(V )6=0, they split approximately equally
the excess supply/demand between the two sides), and (2) the two sides are large and have roughly
comparable size (they are both Ω(|V |)). We showed that for 2-connected graphs it is always possible
to achieve both objectives at the same time, and for 3-connected graphs there is a partition that is
essentially perfectly balanced in both objectives. Furthermore, these partitions can be computed
in polynomial time. This is a paradigmatic bi-objective balancing problem. We observed how it
can be easily used to find a connected partition of a graph with two types of nodes that is balanced
with respect to the sizes of both types. Overall, we believe that the novel techniques used in this
paper can be applied to partitioning heterogeneous networks in various contexts.
There are several interesting further directions that suggest themselves. First, extend the theory
and algorithms to find doubly balanced connected partitions to more than two parts. Second,
even considering only the supply/demand objective, does the analogue of the results of Lováz and
Gyori [9,13] for the connected k-way partitioning of k-connected graphs with respect to size (which
corresponds to p(i)=1) extend to the supply/demand case (p(i)=±1) for k>3? And is there a
polynomial algorithm that constructs such a partition? Finally, extend the results of Section 4 to
graphs with more than two types of nodes, that is, can we partition (under suitable conditions)
a graph with several types of nodes to two (or more) large connected subgraphs that preserve
approximately the diversity (the proportions of the types) of the whole population?
11
A
Definitions and Background
In this section, we provide a short overview of the definitions and tools used in our work. Most
of the graph theoretical terms used in this paper are relatively standard and borrowed from [2]
and [21]. All the graphs in this paper are loopless.
A.1
Terms from Graph Theory
Cutpoints and Subgraphs: A cutpoint of a connected graph G is a node whose deletion results
in a disconnected graph. Let X and Y be subsets of the nodes of a graph G. G[X] denotes the
subgraph of G induced by X. We denote by E[X, Y ] the set of edges of G with one end in X and
the other end in Y . The neighborhood of a node v is denoted N (v).
Connectivity: The connectivity of a graph G=(V, E) is the minimum size of a set S⊂V such that
G\S is not connected. A graph is k-connected if its connectivity is at least k.
A.2
st-numbering of a Graph
Given any edge {s, t} in a 2-connected graph G, an st-numbering for G is a numbering for the nodes
in G defined as follows [11]: the nodes of G are numbered from 1 to n so that s receives number 1,
node t receives number n, and every node except s and t is adjacent both to a lower-numbered and
to a higher-numbered node. It is shown in [7] that such a numbering can be found in O(|V |+|E|).
A.3
Series-Parallel Graphs
A Graph G is series-parallel, with terminals s and t, if it can be produced by a sequence of the
following operations:
1. Create a new graph, consisting of a single edge between s and t.
2. Given two series parallel graphs, X and Y with terminals sX , tX and sY , tY respectively,
form a new graph G=P (X, Y ) by identifying s=sX =sY and t=tX =tY . This is known as the
parallel composition of X and Y .
3. Given two series parallel graphs X and Y , with terminals sX , tX and sY , tY respectively, form
a new graph G=S(X, Y ) by identifying s=sX , tX =sY and t=tY . This is known as the series
composition of X and Y .
It is easy to see that a series-parallel graph is 2-connected if, and only if, the last operation is a
parallel composition.
A.4
Nonseparating Induced Cycles and Ear Decomposition
Let H be a subgraph of a graph G. An ear of H in G is a nontrivial path in G whose ends lie in H
but whose internal vertices do not. An ear decomposition of G is a decomposition G=P0 ∪· · ·∪Pk of
the edges of G such that P0 is a cycle and Pi for i≥1 is an ear of P0 ∪P1 ∪· · ·∪Pi−1 in G. It is known
that every 2-connected graph has an ear decomposition (and vice-versa), and such a decomposition
can be found in linear time.
A cycle C is a nonseparating induced cycle of G if G\C is connected and C has no chords. We
say a cycle C avoids a node u, if u∈C.
/
12
Theorem 3 (Tutte [19]). Given a 3-connected graph G(V, E) let {t, r} be any edge of G and let
u be any node of G, r6=u6=t. Then there is a nonseparating induced cycle of G through {t, r} and
avoiding u.
Notice that since G is 3-connected in the previous theorem, every node in C has a neighbor in
G\C. Cheriyan and Maheshwari showed that the cycle in Theorem 3 can be found in O(E) [4].
Moreover, they showed that any 3-connected graph G has a nonseparating ear decomposition
G=P0 ∪· · ·∪Pk defined as follows: Let Vi =V (P0 )∪V (P1 ) · · ·∪V (Pi ), let Gi =G[Vi ] and Gi =G[V \Vi ].
We say that G=P0 ∪P1 ∪· · ·∪Pk is an ear decomposition through edge {t, r} and avoiding vertex u
if the cycle P0 contains edge {t, r}, and the last ear of length greater than one, say Pm , has u as
its only internal vertex. An ear decomposition P0 ∪P1 · · ·∪Pk of graph G through edge {t, r} and
avoiding vertex u is a nonseparating ear decomposition if for all i, 0≤i<m, graph Gi is connected
and each internal vertex of ear Pi has a neighbor in Gi .
Theorem 4 (Cheriyan and Maheshwari [4]). Given an edge {t, r} and a vertex u of a 3-connected
graph G, a nonseparating induced cycle of G through {t, r} and avoiding u, and a nonseparating
ear decomposition can be found in time O(|V |+|E|).
A.5
Partitioning of Graphs to Connected Subgraphs
The following theorem is the main existing result in partitioning of graphs into connected subgraphs
and is proved independently by Lováz and Gyori [9, 13] by different methods.
Theorem 5 (Lováz and Gyori [9,13]). Let G=(V, E) be a k-connected graph. Let n=|V |, v1 , v2 , . . . , vk ∈
V and let n1 , n2 , . . . , nk be positive integers satisfying n1 +n2 +· · ·+nk =n. Then, there exists a partition of V into (V1 , V2 . . . , Vk ) satisfying vi ∈Vi , |Vi |=ni , and G[Vi ] is connected for i=1, 2, . . . , k.
Although the existence of such a partition has long been proved, there is no polynomial-time
algorithm to find such a partition for k>3. For k=2, it is easy to find such partition using stnumbering. For k=3, Wada and Kawaguchi [20] provided an O(n2 ) algorithm using the nonseparating ear decomposition of 3-connected graph.
A.6
Convex Embedding of Graphs
In this subsection, we provide a short overview of the beautiful work by Linial, Lováz, and Wigderson [12] on convex embedding of the k-connected graphs. Let Q={q1 ,P
q2 , . . . , qm } be P
a finite set
of points in Rd . The convex hull conv(Q) of Q is the set of all points m
λ
q
with
i=1 i i
i=1 λi =1.
The rank of Q is defined by rank(Q)=1+dim(conv(Q)). Q is in general position if rank(S)=d+1
for every (d+1)-subset S⊆Q. Let G be a graph and X⊂V . A convex X-embedding of G is any
mapping f :V →R|X|−1 such that for each v∈V \X, f (v)∈conv(f (N (v))). We say that the convex
embedding is in general position if the set f (V ) of the points is in general position.
Theorem 6 (Linial, Lováz, and Wigderson [12]). Let G be a graph on n vertices and 1<k<n.
Then the following two conditions are equivalent:
1. G is k-connected
2. For every X⊂V with |X|=k, G has a convex X-embedding in general position.
Notice that the special case of the Theorem for k=2 asserts the existence of an st-numbering
of a 2-connected graph. The proof of this theorem is inspired by physics. The embedding is
13
found by letting the edges of the graph behave like ideal springs and letting its vertices settle. A
formal summary of the proof (1→2) is as follows (for more details see [12]). For each vi ∈X, define
f (vi ) arbitrary in Rk−1 such that f (X) is in general position. Assign to every edge (u, v)∈E a
positive elasticity coefficient cuv and let c∈R|E| be the vector of coefficients. It is proved in [12]
that
vector c, an embedding f that minimizes the potential function P =
P for almost any coefficient
2 provides a convex X-embedding in general position (k.k is the Euclidean
c
kf
(u)−f
(v)k
{u,v}∈E uv
norm). Moreover, the embedding that minimizes P can be computed as follows,
f (v)=
1 X
cuv f (u) for all v∈V \X,
cv
u∈N (v)
P
in which cv = u∈N (v) cuv . Hence, the embedding can be found by solving a set of linear equations
in at most O(|V |3 ) time (or matrix multiplication time).
B
Missing Proofs from Section 2
Proof of Proposition 1. We use the proof of [3, Theorem 2] with a modest change. The reduction is
from the X3C problem [15], which is a variant of the Exact Cover by 3-sets and defined as follows:
Given a set X with |X|=3q and a family C of 3-element subsets of X such that |C|=3q and each
element of X appears in exactly 3 sets of C, decide whether C contains an exact cover for X. Given
an instance
set V =X∪C∪{a, b} and edge
S (X, C) of X3C, let G=(V, E) be the graph with the vertex
2 +q, p(C )=−1, and p(x )=−3q.
set E= 3q
[{C
x
|x
∈C
}∪{C
a}∪{C
b}].
Set
p(a)=2q,
p(b)=9q
j i i
j
j
j
j
i
j=1
It is easy to verify that C contains an exact cover for X if and only if the BCPI problem has a
solution such that p(V1 )=p(V2 )=0. This shows the first claim.
For the second claim, attach to nodes a, b, and the xi s, paths of length 2q, 9q 2 +q, and 3q,
respectively, and set the supply/demand values of a, b, the xi ’s and the new nodes equal to +1 (for
the paths for a and b) or −1 (for the xi ’s).
Proof of Proposition 2. Clearly, part (2) follows immediately from part (1) because in this case,
p(V1 ), p(V2 ) are integer and maxj∈V |p(j)|/2=1/2. To show part (1), pick two arbitrary nodes
u, v∈V with p(u)p(v)>0. Since we want to separate u and v, we can assume wlog that initially
{u, v}∈G. Since G is 2-connected, an st-numbering between nodes u and v as u=v1 , v2 , . . . , vn =v
(i)
(1)
can be found in O(|V |+|E|) [7]. Define V1 :={v1 , v2 , . . . , vi }. It is easy to see that p(V1 )=p(u)>0
(n−1)
(i∗ )
and p(V1
)=−p(v)<0. Hence, there must exist an index 1≤i∗ <n such that |p(V1 )|>0 and
(i∗ +1)
(i)
(i+1)
(i∗ )
(i∗ +1)
|p(V1
)|≤0. Since |p(V1 )−p(V1
)|=|p(i∗ +1)|, either |p(V1 )|≤|p(i∗ +1)|/2 or |p(V1
)|≤
(i∗ )
(i∗ +1)
∗
|p(i +1)|/2; Accordingly set V1 =V1
or V1 =V1
. Let V2 =V \V1 . Hence, (V1 , V2 ) is a solution
∗
with |p(V1 )|=|p(V2 )|≤|p(i +1)/2|≤maxj∈V |p(j)|/2. It is easy to see that i∗ can be found in O(|V |).
Proof of Proposition 3. Consider the case of general function p, and let pmax =maxj∈V |p(j)|. We
will show that we can find a solution such that u∈V1 , v∈V2 , w∈V3 with |p(V1 )|, |p(V2 )|≤pmax /2. Since
|p(V3 )|=|p(V1 )+p(V2 )| (recall p(V )=0), this implies that |p(V3 )|≤pmax , and hence |p(V1 )|+|p(V2 )|+
|p(V3 )|≤2pmax . Furthermore, if p(i)=±1 for all i∈V , hence pmax =1, then |p(V1 )|, |p(V2 )|≤pmax /2
implies that p(V1 )=p(V2 )=0, and therefore also p(V3 )=0. Thus, both claims will follow.
Assume u, v, w∈V and p(u), p(v), p(w)>0 (the case of negative p(u), p(v), p(w) is symmetric).
Since we want to separate u from v, we can assume without loss of generality that {u, v}∈E.
14
Using [4], there is a non-separating ear decomposition through the edge {u, v} and avoiding node w.
Ignore the ears that do not contain any internal nodes, and let Q0 ∪Q1 ∪· · ·∪Qr be the decomposition
consisting of the ears with nodes; we have w∈Qr . Let Vi =V (Q0 )∪V (Q1 ) · · ·∪V (Qi ), let Gi =G[Vi ]
and Gi =G[V \Vi ]. We distinguish two cases, depending on whether p(V0 )≤0 or p(V0 )>0.
(i) If p(V0 )≤0, then consider an st-numbering between u and v in V0 , say u=v1 , v2 , . . . , vs =v.
(i)
Define V0 ={v1 , v2 , . . . , vi }. Since p(u), p(v)>0 and p(V0 )≤0, there must exist indices 1≤i∗ ≤
(i∗ )
(i∗ +1)
(j ∗ +1)
(j ∗ )
j ∗ <s such that p(V0 )>0, p(V0
)≤0 and p(V0 \V0
)>0, p(V0 \V0 )≤0.
(i∗ )
(i∗ +1)
(i∗ )
(a) If i∗ =j ∗ , since p(V0 )+p(vi∗ +1 )+p(V0 \V0
)=p(V0 )<0, we have p(V0
|p(vi∗ +1 )|. Now, one of the following three cases happens:
(i∗ )
(i∗ +1)
)+p(V0 \V0
)≤
(i∗ +1)
- If p(V0 )≤|p(vi∗ +1 )|/2 and p(V0 \V0
)≤|p(vi∗ +1 )|/2, then it is easy to see that
(i∗ )
(i∗ +1)
V1 =V0 , V2 =V0 \V0
, and V3 =V \(V1 ∪V2 ) is a good partition.
(i∗ )
(i∗ +1)
(i∗ )
- If p(V0 )>|p(vi∗ +1 )|/2 and p(V0 \V0
)≤|p(vi∗ +1 )|/2, then p(V0 )+p(vi∗ +1 )=
(i∗ +1)
(i∗ +1)
(i∗ +1)
∗
p(V0
)≤|p(vi +1 )|/2. Hence, V1 =V0
, V2 =V0 \V0
, and V3 =V \V0 is a good
partition.
(i∗ )
(i∗ +1)
(i∗ +1)
- If p(V0 )≤|p(vi∗ +1 )|/2 and p(V0 \V0
)>|p(vi∗ +1 )|/2, then p(V0 \V0
)+p(vi∗ +1 )=
(i∗ )
(i∗ )
(i∗ )
p(V0 \V0 )≤|p(vi∗ +1 )|/2. Hence, V1 =V0 , V2 =V0 \V0 , and V3 =V \V0 is a good
partition.
(i∗ )
(i∗ +1)
(b) If i∗ <j ∗ , then either p(V0 )≤|p(vi∗ +1 )|/2 or |p(V0
)|≤|p(vi∗ +1 )|/2, accordingly set
(i∗ )
(i∗ +1)
(j ∗ +1)
(j ∗ )
V1 =V0 or V1 =V0
. Similarly, either p(V0 \V0
)≤|p(vj ∗ +1 )|/2 or |p(V0 \V0 )|≤
(j ∗ +1)
(j ∗ )
|p(vj ∗ +1 )|/2, so accordingly set V2 =V0 \V0
or V2 =V0 \V0 . Set V3 =V \(V1 ∪V2 ). It
is easy to check that (V1 , V2 , V3 ) is a good partition.
(ii) If p(V0 )>0, then since p(w)>0 and therefore p(Vr−1 )<0, there must exist an index 0≤j<r−1
such that p(Vj )>0 and p(Vj+1 )≤0. Consider an st-numbering between u and v in G[Vj ]
(i)
as u=v1 , v2 , . . . , vs =v and define Vj ={v1 , v2 , . . . , vi }. The ear Qj+1 is a path of new nodes
q1 , q2 , . . . , qt attached to two (distinct) nodes vx , vy of G[Vj ] through edges {vx , q1 }, {qt , vy }∈E;
assume wlog that 1≤x<y≤s. For simplicity, we will use below Qj+1 to denote also the
(i)
set {q1 , q2 , . . . , qt } of internal (new) nodes of the ear. Also define Qj+1 ={q1 , q2 , . . . , qi } and
(0)
Qj+1 =∅. One of the following cases must happen:
(i∗ )
(a) Suppose there is an index 1≤i∗ <(y−1) such that p(Vj
(i∗ +1)
)>0 and p(Vj
)≤0 or there is
(i∗ −1)
(i∗ )
an index
such that p(Vj \Vj
)>0 and p(Vj \Vj )≤0. Let’s assume there is
(i∗ )
(i∗ +1)
an index 1≤i∗ <(y−1), such that p(Vj )>0 and p(Vj
)≤0 (the other case is exactly
(i∗ )
(i∗ +1)
similar). Then either p(Vj )≤|p(vi∗ +1 )|/2 or |p(Vj
)|≤|p(vi∗ +1 )|/2, accordingly set
(i∗ )
(i∗ +1)
0
either V1 =Vj
or V1 =Vj
. Set V2 =Vj \V1 . One of the following cases happens:
x+1<i∗ <s
(i∗ )
(i∗ +1)
(i∗ +1)
- If V1 =Vj and p(V20 )≤0, then since p(Vj
)≤0, we have p(Vj \Vj
)>0. Hence,
0
0
0
p(V2 \{vi∗ +1 })>0. So, it is either |p(V2 )|≤|p(vi∗ +1 )|/2 or p(V2 \{vi∗ +1 })≤|p(vi∗ +1 )|/2.
Now if p(V20 \{vi∗ +1 })≤|p(vi∗ +1 )|/2, since also p(V1 )≤|p(vi∗ +1 )|/2, p(Vj )≤0 which
contradicts with the assumption. Therefore, |p(V20 )|≤|p(vi∗ +1 )|/2. Set V2 =V20 and
V3 =V \(V1 ∪V2 ). It is easy to check that (V1 , V2 , V3 ) is a good partition.
(i∗ )
- If V1 =Vj
and p(V20 )>0, then since p(Vj ∪Qj+1 )<0, there is an index 0<t∗ ≤t,
(t∗ )
(t∗ −1)
such that p(V20 ∪(Qj+1 \Qj+1 ))>0 and p(V20 ∪(Qj+1 \Qj+1 ))≤0. Hence, either p(V20 ∪
15
(t∗ )
(t∗ −1)
(Qj+1 \Qj+1 ))≤|p(qt∗ )|/2 or |p(V20 ∪(Qj+1 \Qj+1 ))|≤|p(qt∗ )|/2, accordingly set V2 =
(t∗ )
(t∗ −1)
V20 ∪(Qj+1 \Qj+1 ) or V2 =V20 ∪(Qj+1 \Qj+1 ). Set V3 =V \(V1 ∪V2 ). It is easy to see
that (V1 , V2 , V3 ) is a good partition.
(i∗ +1)
- If V1 =Vj
, then since p(V1 )≤0, we have p(V20 )>0. The rest is exactly like the
(i∗ )
previous case when V1 =Vj
and p(V20 )>0.
(i)
(i)
(b) Suppose that for every 1≤i<y, p(Vj )>0 and for every x<i<s, p(Vj \Vj )>0. Set
(y−1)
V10 =Vj
and V20 =Vj \V10 . Based on the assumption p(V10 ), p(V20 )>0. Since p(Vj+1 )≤
(i∗ )
(i∗ +1)
0, there are indices 0≤i∗ ≤j ∗ <t such that p(V10 ∪Qj+1 )>0, p(V10 ∪Qj+1 )≤0 and p(V20 ∪
(j ∗ +1)
(j ∗ )
(Qj+1 \Qj+1 ))>0, p(V20 ∪(Qj+1 \Qj+1 ))≤0. The rest of the proof is similar to case (i)
when p(V0 )≤0.
C
Missing Proofs from Section 3
Proof of Lemma 3. Set X={v, u, w}. Using [12], G has a convex X-embedding in R2 in general
position with mapping f :V →R2 such that f (u)=(0, 0), f (v)=(1, 0), and f (w)=(0, 1). In the Xembedding of the nodes, we have a freedom to set the elasticity coefficient vector ~c to anything that
we want (except a measure zero set of vectors). So for any edge {i, j}∈G[V10 ]∪G[V20 ], set cij =g; and
for any {i, j}∈E[V10 , V20 ], set cij =1. Assume L1 is the line y=0.5, L2 is the line x+2y=1, and L3 is
the line x=y.
First, we show that there exist a g for which all the nodes in V10 will be embedded above the line
L1 . To show
P this, from [12], we know the embedding is such that it minimizes the total potential
P (f, ~c)= {i,j}∈E cij kf (i)−f (j)k2 . Notice that we can independently minimize P on x-axis values
and y-axis values as below:
min P
f
= min Px +min Py
f1
f2
X
X
= min
cij (f1 (i)−f1 (j))2 +min
cij (f2 (i)−f2 (j))2
f1
f2
{i,j}∈E
{i,j}∈E
Now, notice that if we place all the nodes in V10 at point (0,1) and all the nodes in V20 on
the line uv, then Py ≤|E|. Hence, if f2 minimizes Py , then Py (f2 , c)≤|E|. Set g≥4|V |2 |E|. We
show that if f2 minimizes Py , then for all edges {i, j}∈G[V10 ]∪G[V20 ], (f2 (i)−f2 (j))2 ≤1/(4|V |2 ).
By contradiction, assume there is an edge {i, j}∈G[V10 ]∪G[V20 ] such that (f2 (i)−f2 (j))2 >1/(4|V |2 ).
Then, cij (f2 (i)−f2 (j))2 =g(f2 (i)−f2 (j))2 >|E|. Hence, Py (f2 , c)>|E| which contradicts with the
fact the f2 minimizes Py . Therefore, if g≥4|V |2 |E|, then for all {i, j}∈G[V10 ]∪G[V20 ], |f2 (i)−f2 (j)|≤
1/(2|V |). Now, since G[V10 ] is connected, all the nodes in V10 are connected to w with a path of
length (in number of hops) less than |V |−1. Hence, using the triangle inequality, for all i∈V10 :
|f2 (w)−f2 (i)|≤(|V |−1)/(2|V |)<1/2⇒|1−f2 (i)|<1/2,
which means that all the nodes in V10 are above L1 .
With the very same argument, if g≥t2 |V |2 |E|, then for all i∈V20 , f2 (i)<1/t.
Now, we want to prove that there is a g such that all the nodes in V20 will be embedded below
the lines L2 and L3 . Define n1 (i):=|N (i)∩V10 | and n2 (i):=|N (i)∩V20 |. From [12], we know the
16
P
P
embedding is such that for all i∈V \{u, v, w}, f (i)=1/ci j∈N (i) cij f (j), where cj = j∈N (i) cij f (j).
Since G[V20 ] is connected, for any i∈V20 there is a path i=v1 , v2 , . . . , vr =v in V20 . Using this ordering:
(
1
f1 (vj )≥ n2 (vj )g+n
gf1 (vj+1 )≥(1/|V |)f1 (vj+1 ), ∀j∈{1, . . . , r−1}
1 (vj )
f1 (vr )=f1 (v)=1
⇒∀i∈V20 \{u, v}: f1 (i)≥(1/|V |)r ≥(1/|V |)|V | .
On the other hand, from the previous part, if we set g≥|V |2|V |+2 |E|, then for all i∈V20 , f2 (i)≤
(1/|V |)|V | . Hence, for all i∈V20 , f2 (i)≤f1 (i), which means that all the nodes in V20 will be placed
below the line L3 .
With the very same idea, we show that there exist a g for which all the nodes in V20 will be placed
below the line L2 . Since G[V20 ] is connected, for P
any i∈V20 there is a path u=u1 , v2 , . . . , ut =i in V20 .
Notice that for all i∈V \{u, v, w}, 1−f1 (i)=1/ci j∈N (i) cij (1−f1 (j)). Hence, since ∀j∈V :f1 (j)≤1,
we have,
(
1
1−f1 (uj )≥ n2 (uj )g+n
g(1−f1 (ui−1 ))≥(1/|V |)(1−f1 (ui−1 )), ∀j∈{2, . . . , t}
1 (uj )
1−f1 (u)=1−f1 (u1 )=1
⇒∀i∈V20 \{u, v}: 1−f1 (i)≥(1/|V |)t ≥(1/|V |)|V | .
From the previous part, if we set g≥4|V |2|V |+2 |E|, then for all i∈V20 , f2 (i)≤1/2(1/|V |)|V | . Hence,
for i∈V20 , f1 (i)+2f2 (i)≤1, which means that all the nodes in V20 will be placed below the line L3 .
Therefore, if we set g≥4|V |2|V |+2 |E|, then we will get an embedding as depicted in Fig. 3. Note
that a polynomial number of bits suffices for g.
Notice that if ~c is a “good” vector, then so is ~c+~ in which ~ is a vector with very small
Euclidean norm. Hence, we can always find a “good” vector ~c which result in a X-embedding in
general position.
Proof of Theorem 1. Assume that |V |≡0(mod 4); the proof for the case |V |≡2(mod 4) is similar.
Using Lemma 2, we can find {u, v, w}∈V and a partition (V10 , V20 ) of V with properties described in
the Lemma. Set X={u, v, w}. Using Lemma 3, we can find a convex X-embedding of G in general
position with properties described in the Lemma as depicted in Fig. 3. The rest of the proof is
very similar to the proof of Lemma 1. We consider again a circle C around f (u), f (v), f (w) in R2
as shown in Fig. 3. Also consider a directed line L tangent to the circle C at point A. If we project
the nodes of G onto the line L, this time the order of the nodes projection gives an st-numbering
between the first and the last node only if u and v are the first and last node. However, if we set V1
to be the |V |/2 nodes whose projections come first and V2 are the |V |/2 nodes whose projections
come last, then G[V1 ] and G[V2 ] are both connected even when u and v are not the first and last
nodes. The reason lies on the special embedding that we considered here. Assume for example w
and v are the first and the last projected nodes, and V1 and V2 are set of the |V |/2 nodes which
projections come first and last, respectively. Two cases might happen:
(i) If u, w∈V1 and v∈V2 , then since {u, w}∈E, both G[V1 ] and G[V2 ] are connected because of
the properties of the embedding.
(ii) If w∈V1 and u, v∈V2 , since |V20 |≤|V |/2 and |V2 |=|V |/2, then either V2 =V20 or V2 ∩V10 6=∅. If
V2 =V20 , and hence V1 =V10 then there is nothing to prove. So assume there is a node z∈V2 ∩V10 .
From the properties of the embedding, the triangle {z, u, v} contains all the nodes of V20 .
17
Since {z, u, v}∈V2 , and V2 contains all the nodes that are on a same side of a halfplane, we
should also have V20 ⊂V2 . Now, from the properties of the embedding, it is easy to see that
every node in V2 has a path either to u or v. Since V20 ⊂V2 , there is also a path between u
and v. Thus, G[V2 ] is connected. From the properties of the embedding, G[V1 ] is connected
as before.
The rest of the proof is exactly the same as the proof of Lemma 1. We move L from being tangent
at point A to point B (AB is a diameter of the circle C) and consider the resulting partition. Notice
that if at point A, p(V1 )>0, then at point B since V1 and V2 completely switch places compared to
the partition at point A, p(V1 )<0. Hence, as we move L from being tangent at point A to point
B and keep it tangent to the circle, in the resulted partitions, p(V1 ) goes from some positive value
to a negative value. Notice that the partition (V1 , V2 ) changes only if L passes a point D on the
circle such that at D, L is perpendicular to a line that connects f (i) to f (j) for i, j∈V . Now, since
the embedding is in general position, there are exactly two points on every line that connects two
points f (i) and f (j), so V1 changes at most by one node leaving V1 and one node entering V1 .
Hence, p(V1 ) changes by either ±2 or 0 value at each change. Now, since |V |≡0(mod 4), p(V1 ) has
an even value in all the resulting partitions. Therefore, as we move L from being tangent at point
A to point B, there should be a point D such that in the resulted partition p(V1 )=p(V2 )=0.
Proof of Corollary 2. Every series-parallel graph G has a separation pair {u, v} such that every
connected component of G[V \{u, v}] has less that 2|V |/3 nodes, and furthermore, such a separation
pair can be found in linear time. To see this, consider the derivation tree T of the construction of
G. The root of T corresponds to G, the leaves correspond to the edges, and every internal node i
corresponds to a subgraph Gi =(Vi , Ei ) that is the series or parallel composition of the subgraphs
corresponding to its children. Starting at the root of T , walk down the tree following always
the edge to the child corresponding to a subgraph with the maximum number of nodes until the
number of nodes becomes ≤2|V |/3. Thus, we arrive at a node i of the tree such that |Vi |>2|V |/3
and |Vj |≤2|V |/3 for all children j of i. Let ui , vi be the terminals of Gi . Note that ui , vi separate all
the nodes of Gi from all the nodes that are not in Gi . Since |Vi |>2|V |/3, we have |V \Vi |<|V |/3. If
Gi is the parallel composition of the graphs corresponding to the children of i, then the separation
pair {ui , vi } has the desired property, i.e. all the components of G[V \{u, v}] have less than 2|V |/3
nodes.
Suppose Gi is the series composition of the graphs Gj , Gk corresponding to the children j, k of
i, and let w be the common terminal of Gj , Gk ; thus, Gi has terminals ui , w, and Gk has terminals
w, vi . Assume wlog that |Vj |≥|Vk |. Then |V |/3<|Vj |≤2|V |/3. The pair {ui , w} of terminals of Gj
separates all the nodes of Vj \{ui , w} from all the nodes of V \Vj , and both these sets have less than
2|V |/3 nodes. Thus, {ui , w} has the required property.
Proof of Theorem 2. Using Lemma 5 for q=4, we consider two cases:
(i) There is a separation pair {u, v}∈V such that if G1 , . . . , Gk are the connected components of
G\{u, v}, for any i, |Vi |<3|V |/4. In this case Lemma 4 for q=4 proves the theorem.
(ii) After a set of contractions, G can be transformed into a 3-connected graph G∗ =(V ∗ , E ∗ ) with
weighted edges such that for any edge e∗ ∈E ∗ , w(e∗ )<|V |/4. In this case the proof is similar to
the proof of Theorem 1. Notice that if G∗ contains a triangle then the proof is much simpler
18
as in the proof of Lemma 1 but here to avoid repetition, we use the approach in the proof of
Theorem 1 and prove the theorem once for all cases of G∗ .
Recall that for each edge in G∗ , its weight represents the number of nodes in the contracted
subgraph that it represents. So if e∗ ∈E ∗ represents an induced subgraph of G with l nodes,
then w(e∗ )=l; and if e∗ ∈E∩E ∗ then w(e∗ )=0. Using Lemma 2, we can find {u, v, w}∈V ∗
and a partition (V1∗ , V2∗ ) of V ∗ with properties described in the Lemma. Since in G∗ edges
have weights
that actually represent nodes in G, we want a partition such that we also have
P
|V2∗ |+ e∈G[V ∗ ] w(e)≤|V |/2. Notice that we can modify the proof of Lemma 2 to take into
2
P
account the weights for the edges and find a partition such that |V2∗ |+ e∈G[V ∗ ] w(e)≤|V |/2
2
as follows. Again, using the algorithm presented in [4], we can find a non-separating proper
cycle C0 in G∗ such that every node in C0 has a neighbor in G∗ \C0 . We consider three cases:
P
(a) If |C0 |+ e∈G[C0 ] w(e)≤|V |/2+1, then the proof is as in the proof of Lemma 2. Select
any three consecutive nodes (u, w, v) of C0 and set V2∗ =C0 \{w} and V1∗ =V ∗ \V2∗ .
P
(b) If |C0 |>|V ∗ |/2 and |C0 |+ e∈G[C0 ] w(e)>|V |/2+1, then as in the proof of Lemma 2,
since |C0 |>|V ∗ \C0 | and every node is C0 has a neighbor in V ∗ \C0 , there exist a node
w∈V ∗ \C0 such that |N (w)∩C0 |≥2. Select two
P nodes u, v∈N (w)∩C0 . There exists a
path P in C0 between u and v such that |P |+ e∈G[P ] w(e)<|V |/2−1. Set V2∗ ={u, v}∪P
and V1∗ =V ∗ \V2∗ .
P
(c) If |C0 |≤|V ∗ |/2 and |C0 |+ e∈G[C0 ] w(e)>|V |/2+1, then if there exist a node w∈V ∗ \C0
such that |N (w)∩C0 |≥2, the proof is as in the previous part. So assume for all w∈V ∗ \C0 ,
|N (w)∩C0 |≤1. Assume w1 , w2 , . . . , wt ∈V ∗ \C0 are all the nodes with |N (wi )∩C0 |=1. We
show that there is a 1≤i≤t, such that wi is not a cut-point of G∗ [V ∗ \C0 ]. Let T be a
spanning tree of G∗ [V ∗ \C0 ]. If there is a 1≤i≤t, such that wi is a leaf of T , then wi is not
a cut-point of G∗ [V ∗ \C0 ] and there is nothing left to prove. So assume none of wi s is a
leaf of T . Suppose 1≤p, q≤t are such that the path between wp and wq in T is the longest
between all pairs of wi s. We show that wp and wq cannot be cut-points of G∗ [V ∗ \C0 ].
Assume for example wp is a cut-point of G∗ [V ∗ \C0 ] and G∗q̄ is a connected component of
G∗ [V ∗ \C0 ]\{wp } such that wq ∈G
/ ∗q̄ . Since the path between wp and wq in T is the longest
between all pairs of wi s, ∀1≤i≤t:wi ∈G
/ ∗q̄ , otherwise we can find a longer path from wq
to some other wi . On the other hand, if ∀1≤i≤t:wi ∈G
/ q̄ , in G∗ \{wp } the cycle C0 is
∗
disconnected from Gq̄ which contradicts with the 3-connectedness of G∗ . Hence, there
should be at least a noncut-point w∈V ∗ \C0 with |N (w)∩C0 |=1. Since G∗ is 3-connected,
each node has degree at least 3. Hence, there are two nodes u, v∈N (w)∩(V ∗ \C0 ). Set
V1∗ =C0 ∪{w} and V2∗ =V ∗ \V1∗ . (V1∗ , V2∗ ) is a desirable partition for V ∗ .
Hence, we can find {u, v, w}∈V ∗ andPa partition (V1∗ , V2∗ ) of V ∗ with properties described in
the Lemma 2 as well as having |V2∗ |+ e∈G[V ∗ ] w(e)≤|V |/2. Set X={u, v, w}. Using Lemma 3,
2
G∗ has a convex X-embedding in general position like f ∗ :V ∗ →R2 as described in the lemma
and depicted in Fig. 3. Now, from this embedding, we get a convex X-embedding for G
like f :V →R2 as follows. For any i∈V ∩V ∗ , f (i)=f ∗ (i). For any edge {i, j}∈E ∗ such that
{i, j} represents an induced subgraph of G, we represent it by a pseudo-path between i and
j in G like P and place the nodes in P in order on random places on the line segment that
connects f (i) to f (j). If the edge {i, j}∈E ∗ is between a node in V1∗ and a node in V2∗ and
represents a pseudo-path P in G , we place the nodes in P in order on random places on the
segment that connects f (i) to f (j) but above the line L1 . Hence, by this process, we get a
convex X-embedding for G which is in general position (almost surely) except for the nodes
19
that are part of a pseudo-path (which we know have length less than |V |/4). Notice that if
V20 ⊂V contains all the nodes in V2∗ and all the nodes that are part of a pseudo-path between
the nodes of V2∗ (represented by weighted edges in G∗ ) in G, and V10 =V \V20 , then (V10 , V20 ) is
partition of G with all the properties of Lemma 2. Moreover, the X-embedding f of G that
we derived from f ∗ for G∗ , has all the properties of Lemma 3 for the partition (V10 , V20 ).
The rest of the proof is similar to the proof of Theorem 1. We consider again a circle C
around f (u), f (v), f (w) in R2 as shown in Fig. 3. Also consider a directed line L tangent to
the circle C at point A and project nodes of G onto the line L. With the same argument as
in the proof of Theorem 1, since the embedding f has the properties in Lemma 2 and 3, if we
set V1 to be the |V |/2 nodes whose projections come first and V2 are the |V |/2 nodes whose
projections come last, then G[V1 ] and G[V2 ] are both connected. If |p(V1 )|≤1, then (V1 , V2 ) is
a good partition and there is nothing left to prove. Otherwise, we move L from being tangent
at point A to point B (AB is a diameter of the circle C) and consider the resulting partition.
Notice that if at point A, p(V1 )>0, then at point B since V1 and V2 completely switch places
compared to the partition at point A, p(V1 )<0. Hence, as we move L from being tangent at
point A to point B and keep it tangent to the circle, in the resulting partitions, p(V1 ) goes
from some positive value to a negative value. Notice that the partition (V1 , V2 ) changes only
if L passes a point D on the circle such that at D, L is perpendicular to a line that connects
f (i) to f (j) for i, j∈V . Now, since the embedding is in general position except for few lines
∗ or {i, j}∈E∩E ∗ there are exactly two points on
that contain a pseudo-path, if {i, j}∈E∪E
/
the line that connects f (i) and f (j), so V1 changes at most by one node leaving V1 and one
node entering V1 . Thus, in this case p(V1 ) changes by at most 2. However, if {i, j}∈E ∗ , then
since there can be at most |V |/4−1 points on the line that connects f (i) to f (j), p(V1 ) may
change by |V |/4−1. So let’s consider that L is perpendicular to the line that connects f (i)
and f (j) at point D and p(V1 )>0 slightly before L passes point D and p(V1 )≤0 slightly after
L passes point D. We consider 2 cases:
∗ or {i, j}∈E∩E ∗ . In this case since V changes by at most 2,
(a) Suppose {i, j}∈E∪E
/
1
therefore, p(V1 ) slightly after L passes D is either 0 or −1. Hence, the partition (V1 , V2 )
that we get after L passes D is a good partition with |V1 |=|V2 |=|V |/2.
(b) Suppose {i, j}∈E ∗ . Let (Vb1 , Vb2 ) be the partition slightly before L passes D and let
(Va1 , Va2 ) be the partition slightly after. Assume P0 is the pseudo-path between i and j
and P =P0 ∪{i, j}. It is easy to see that (Vb1 \P, Vb2 \P )=(Va1 \P, Va2 \P ). If p(Vb1 \P )=0,
since |P0 |<|V |/4, then V1 =Vb1 \P and V2 =V \V1 is a good partition and there is nothing
left to prove. So either p(Vb1 \P )>0 or p(Vb1 \P )<0.
If p(Vb1 \P )>0, since p(Va1 )≤0 and Vb1 \P =Va1 \P , we can add a set of nodes from P ∩Va1 ,
say P 0 , to Va1 \P to get p((Va1 \P )∪P 0 )=0. Now since |P0 |<|V |/4, V1 =(Va1 \P )∪P 0 and
V2 =V \V1 is a good partition.
If p(Vb1 \P )<0, since p(Vb1 )>0, we can add a set of nodes from P ∩Vb1 , say P 0 , to Vb1 \P
to get p((Vb1 \P )∪P 0 )=0. Now since |P0 |<|V |/4, V1 =(Vb1 \P )∪P 0 and V2 =V \V1 is a good
partition.
20
D
Missing Proofs from Section 4
Proof of Corollary 4. Suppose without loss of generality that nr ≥nb and let nr −nb =2t and nr +nb =
n=2m. Set p(i)=1 for i∈R and p(i)=−1 for i∈B. Then p(V )=2t. From the equations, we have
nr =m+t and nb =m−t.
From Corollary 1 we can find a partition (V1 , V2 ) such that |V1 |=|V2 | and |p(V1 )−p(V )/2|, |p(V1 )−
p(V )/2|≤1. Let r1 =|R∩V1 | and b1 =|B∩V1 |. We have r1 +b1 =n/2=m and t−1≤r1 −b1 ≤t+1. Therefore, (m+t)/2−(1/2)≤r1 ≤(m+t)/2+(1/2). Since r1 is an integer and nr =m+t is even, it follows
that r1 =(m+t)/2=nr /2. Hence, b1 =(m−t)/2=nb /2. Therefore, V2 also contains nr /2 red nodes
and nb /2 blue nodes.
References
[1] D. Bienstock. Electrical Transmission System Cascades and Vulnerability: An Operations
Research Viewpoint, volume 22. SIAM, 2016.
[2] J. Bondy and U. Murty. Graph theory, 2nd printing, 2008.
[3] F. Chataigner, L. R. Salgado, and Y. Wakabayashi. Approximation and inapproximability
results on balanced connected partitions of graphs. DMTCS, 9(1):177–192, 2007.
[4] J. Cheriyan and S. Maheshwari. Finding nonseparating induced cycles and independent spanning trees in 3-connected graphs. J. Algorithms, 9(4):507–537, 1988.
[5] J. Chlebı́ková. Approximating the maximally balanced connected partition problem in graphs.
Inform. Process. Lett., 60(5):225–230, 1996.
[6] M. Dyer and A. Frieze. On the complexity of partitioning graphs into connected subgraphs.
Discrete Applied Math., 10:139–153, 1985.
[7] S. Even and R. E. Tarjan. Computing an st-numbering. Theoretical Computer Science,
2(3):339–344, 1976.
[8] N. Fan, D. Izraelevitz, F. Pan, P. M. Pardalos, and J. Wang. A mixed integer programming
approach for optimal power grid intentional islanding. Energy Systems, 3(1):77–93, 2012.
[9] E. Gyori. On division of graphs to connected subgraphs. In Combinatorics (Proc. Fifth
Hungarian Colloq., Keszthely, 1976), volume 1, pages 485–494, 1976.
[10] J. E. Hopcroft and R. E. Tarjan. Dividing a graph into triconnected components. SIAM J.
Comput., 2(3):135–158, 1973.
[11] A. Lempel, S. Even, and I. Cederbaum. An algorithm for planarity testing of graphs. In
Theory of graphs: International symposium, volume 67, pages 215–232. Gordon and Breach,
New York, 1967.
[12] N. Linial, L. Lovasz, and A. Wigderson. Rubber bands, convex embeddings and graph connectivity. Combinatorica, 8(1):91–102, 1988.
[13] L. Lovász. A homology theory for spanning tress of a graph. Acta Mathematica Hungarica,
30(3-4):241–251, 1977.
21
[14] S. Pahwa, M. Youssef, P. Schumm, C. Scoglio, and N. Schulz. Optimal intentional islanding
to enhance the robustness of power grid networks. Physica A, 392(17):3741–3754, 2013.
[15] C. H. Papadimitriou and M. Yannakakis. The complexity of restricted spanning tree problems.
JACM, 29(2):285–309, 1982.
[16] R. J. Sánchez-Garcı́a, M. Fennelly, S. Norris, N. Wright, G. Niblo, J. Brodzki, and J. W. Bialek.
Hierarchical spectral clustering of power grids. IEEE Trans. Power Syst., 29(5):2229–2237,
2014.
[17] K. Sun, Q. Zhao, D.-Z. Zheng, J. Ma, and Q. Lu. A two-phase method based on obdd for
searching for splitting strategies of large-scale power systems. In Proc. IEEE PowerCon’02,
2002.
[18] H. Suzuki, N. Takahashi, and T. Nishizeki. A linear algorithm for bipartition of biconnected
graphs. Inform. Process. Lett., 33(5):227–231, 1990.
[19] W. T. Tutte. How to draw a graph. Proc. London Math. Soc, 13(3):743–768, 1963.
[20] K. Wada and K. Kawaguchi. Efficient algorithms for tripartitioning triconnected graphs and
3-edge-connected graphs. In Graph-Theoretic Concepts in Computer Science, pages 132–143.
Springer, 1994.
[21] D. B. West et al. Introduction to graph theory, volume 2. Prentice hall Upper Saddle River,
2001.
22
| 8cs.DS
|
arXiv:cs/0703155v1 [cs.PL] 30 Mar 2007
Replace this file with prentcsmacro.sty for your meeting,
or with entcsmacro.sty for your meeting. Both can be
found at the ENTCS Macro Home Page.
Liveness of Heap Data for Functional Programs
Amey Karkare1 Uday Khedker Amitabha Sanyal
{karkare,uday,as}@cse.iitb.ac.in
Department of CSE, IIT Bombay
Mumbai, India
Abstract
Functional programming languages use garbage collection for heap memory management. Ideally, garbage collectors should
reclaim all objects that are dead at the time of garbage collection. An object is dead at an execution instant if it is not used in
future. Garbage collectors collect only those dead objects that are not reachable from any program variable. This is because
they are not able to distinguish between reachable objects that are dead and reachable objects that are live.
In this paper, we describe a static analysis to discover reachable dead objects in programs written in first-order, eager functional programming languages. The results of this technique can be used to make reachable dead objects unreachable, thereby
allowing garbage collectors to reclaim more dead objects.
Keywords: Compilers, Liveness, Garbage Collection, Memory Management, Data Flow Analysis, Context Free Grammars
1
Introduction
Garbage collection is an attractive alternative to manual memory management because it
frees the programmer from the responsibility of keeping track of object lifetimes. This
makes programs easier to implement, understand and maintain. Ideally, garbage collectors
should reclaim all objects that are dead at the time of garbage collection. An object is
dead at an execution instant if it is not used in future. Since garbage collectors are not able
to distinguish between reachable objects that are live and reachable objects that are dead,
they conservatively approximate the liveness of an object by its reachability from a set of
locations called root set (stack locations and registers containing program variables) [14].
As a consequence, many dead objects are left uncollected. This has been confirmed by
empirical studies for Haskell [19], Scheme [16] and Java [22,23,24].
Compile time analysis can help in distinguishing reachable objects that are live from
reachable objects that are dead. This is done by detecting unused references to objects.
If an object is dead at a program point, none of its references are used by the program
beyond that program point. If every unused reference is nullified, then the dead objects
may become unreachable and may be claimed by garbage collector.
1
Supported by Infosys Technologies Limited, Bangalore, under Infosys Fellowship Award.
c
2018 Published by Elsevier Science B. V.
Karkare, Khedker and Sanyal
(define (app list1 list2)
π1 :(if π2 :(null? π3 :list1)
π4 :list2
π5 :(cons π6 :(car π7 :list1)
π8 :(app π9 :(cdr π10 :list1)
π11 :list2))))
(let z ←(cons (cons 4 (cons 5 Nil))
(cons 6 Nil)) in
(let y ← (cons 3 Nil) in
π12 :(let w ← π13 :(app y z) in
π14 :(car (car (cdr w))))))
y
w
×
z
×
×
3
×
Nil
×
6
4
5
Nil
Nil
(b) Memory graph at π14 .
(b) Thick edges denote live links.
(b) Edges marked × can be nullified.
(a) Example program.
Fig. 1. Example Program and its Memory Graph.
Example 1.1 Figure 1(a) shows an example program. The label π of an expression e denotes the program point just before the evaluation of e. At a given program point, the heap
memory can be viewed as a (possibly unconnected) directed acyclic graph called memory
graph. The locations in the root set form the entry nodes for the memory graph. Figure 1(b) shows the memory graph at π14 . Each cons cell is an intermediate node in the
graph. Elements of basic data types and the 0-ary constructor Nil form leaf nodes of the
graph. They are assumed to be boxed, i.e. stored in separate heap cells and are accessed
through references. The edges in the graph are called links.
If we consider the execution of the program starting from π14 , the links in the memory
graph that are traversed are shown by thick arrows. These links are live at π14 . Links that
are not live can be nullified by the compiler by inserting suitable statements. If an object
becomes unreachable due to nullification, it can be collected by the garbage collector.
In the figure, the links that can be nullified are shown with a ×. Note that a link need
not be nullified if nullifying some other link makes it unreachable from the root set. If a
node becomes unreachable from the root set as a consequence of nullifying the links, it will
be collected during the next invocation of garbage collector.
✷
In this example, starting at π14 , there is only one execution path. In general, there could
be multiple execution paths starting from a program point π. The liveness information at π
is a combination of liveness information along every execution path starting at π.
In this paper, we describe a static analysis for programs written in first-order, eager
functional programming languages. The analysis discovers live references at every program
point, i.e. the references that may be used beyond the program point in any execution of
the program. We use context free grammars as a bounded representation for the set of live
references. The result of the analysis can be used by the compiler to decide whether a given
reference can be nullified at a given program point. Our analysis is context-sensitive yet
modular in that a function is analyzed only once.
The rest of the paper is organized as follows: Section 2 describes the language used to
explain our analysis along with the basic concepts and notations. The analysis in Section 3
captures the liveness information of a program as a set of equations. The method to solve
2
Karkare, Khedker and Sanyal
p ::= d1 . . . dn e1
— program
d ::= (define ( f v1 . . . vn ) e1 ) — function definition
e ::=
— expression
κ
— constant
|v
— variable
| Nil
— primitive constructor
— primitive constructor
| (cons e1 e2 )
| (car e1 )
— primitive selector
| (cdr e1 )
— primitive selector
| (pair? e1 )
— primitive tester
| (null? e1 )
— primitive tester
| (+ e1 e2 )
— generic primitive
| (if e1 e2 e3 )
— conditional
| (let v1 ← e2 in e3 ) — let binding
| ( f e1 . . . en )
— function application
Fig. 2. The syntax of our language
these equations is given in Section 4. Section 5 describes how the result of the analysis can
be used to nullify unused references. Finally, we compare our approach with related work
in Section 6 and conclude in Section 7.
2
Language, Concepts and Notations
The syntax of our language is described in Figure 2. The language has call-by-value semantics. The argument expressions are evaluated from left to right. We assume that variables
in the program are renamed so that the same name is not defined in two different scopes.
For notational convenience, the left link (corresponding to the car) of a cons cell is
denoted by 0 and the right link (corresponding to the cdr) is denoted by 1. We use e.0 to
denote the link corresponding to (car e) for an expression e (assuming e evaluates to a list)
and e.1 to denote the link corresponding to (cdr e). A composition of several cars and cdrs
is represented by a string α ∈ {0, 1}∗ . If an expression e evaluates to a cons cell then e.ε
corresponds to the reference to the cons cell.
For an expression e, let [e] denote the location in the root set holding the value of e.
Given a memory graph, the string e.α describes a path in the memory graph that starts
at [e]. We call the string e.α an access expression, the string α an access pattern, and
the path traced in the memory graph an access path. In Figure 1, the access expression
w.100 represents the access path from w to the node containing the value 4. Most often,
the memory graph being referred to is clear from the context, and therefore we shall use
access expressions to refer to access paths. When we use an access path to refer to a link in
the memory graph, it denotes the last link in the access path. Thus, w.100 denotes the link
incident on the node containing the value 4. If σ denotes a set of access patterns, then e.σ
is the set of access paths rooted at [e] and corresponding to σ. i.e.
e.σ = {e.α | α ∈ σ}
A link in a memory graph is live at a program point π if it is used in some path from π
3
Karkare, Khedker and Sanyal
to the program exit. An access path is defined to be live if its last link is live. In Example 1,
the set of live access paths at π14 is {w.ε, w.1, w.10, w.100, z.0, z.00}. Note that the access
paths z.0 and z.00 are live at π14 due to sharing. We do not discover the liveness of such
access paths directly. Instead, we assume that an optimizer using our analysis will use alias
analysis to discover liveness due to sharing.
The end result of our analysis is the annotation of every expression in the program
with a set of access paths rooted at program variables. We call this liveness environment,
denoted L . This information can be used to insert nullifying statements before expressions.
The symbols 0 and 1 extend the access patterns of a structure to describe the access
patterns of a larger structure. In some situations, we need to create access patterns of a
substructure from the access patterns of a larger structure. For this purpose, we extend our
alphabet of access patterns to include symbols 0̄ and 1̄. The following example motivates
the need for these symbols.
Example 2.1 Consider the expression at program point π1 in
π1 :(let w ← π2 :(cons x y) in π3 :· · ·)
Assuming Lπ3 = {w.α}, we would like to find out which reference of the list x and y are
live at π1 . Let x.α′ be live at π1 . Then, the two possible cases are:
•
If α = 1β or α = ε, no link in the structure rooted at x is used. We use ⊥ to denote the
access pattern describing such a situation. Thus, α′ = ⊥.
•
If α = 0β then the link represented by w.α that is x rooted and live at π1 can be represented by x.β. Thus, α′ = β.
This relation between α and α′ is expressed by α′ = 0̄α. 1̄ can be interpreted similarly. ✷
With the inclusion of 0̄, 1̄ and ⊥ in the alphabet for access patterns, an access pattern
does not directly describe a path in the memory graph. Hence we define a Canonical Access
Pattern as a string restricted to the alphabet {0, 1}. As a special case, ⊥ is also considered
as a canonical access pattern.
We define rules to reduce access patterns to their canonical forms. For access patterns
α1 and α2 :
α1 α′2 if α2 ≡ 0α′2
(1) α1 0̄α2 →
⊥
if α2 ≡ 1α′2 or α2 ≡ ε
α1 α′2 if α2 ≡ 1α′2
(2) α1 1̄α2 →
⊥
if α2 ≡ 0α′2 or α2 ≡ ε
(3) α1 ⊥α2 → ⊥
k
α → α′ denotes the reduction of α to α′ in k steps, and → denotes the reflexive and transitive
closure of →. The concatenation (·) of a set of access patterns σ1 with σ2 is defined as a
set containing concatenation of each element in σ1 with each element in σ2 , i.e.
⋆
σ1 · σ2 = {α1 α2 | α1 ∈ σ1 , α2 ∈ σ2 }
3
Computing Liveness Environments
Let σ be the set of access patterns specifying the liveness of the result of evaluating e. Let
L be the liveness environment after the evaluation of e. Then the liveness environment
4
Karkare, Khedker and Sanyal
before the computation of e is discovered by propagating σ backwards through the body of
e. This is achieved by defining an environment transformer for e, denoted X E .
Since e may contain applications of primitive operations and user defined functions,
we also need transfer functions that propagate σ from the result of the application to the
arguments. These functions are denoted by X P and X F . While X P is given directly based
on the semantics of the primitive, X F is inferred from the body of a function.
3.1
Computing X E
For an expression e at program point π, X E (e, σ, L ) computes liveness environment at π
where σ is the set of access patterns specifying the liveness of the result of evaluating e
and L is the liveness environment after the evaluation of e. Additionally, as a side effect,
the program point π is annotated with the value computed. However, we do not show this
explicitly to avoid clutter. The computation of X E (e, σ, L ) is as follows.
(4)
(5)
(6)
(7)
(8)
(9)
(10)
X E (κ, σ, L ) = L
X E (v, σ, L ) = L ∪ v.σ
X E ((P e1 e2 ), σ, L ) = let L ′ ← X E (e2 , X PP2 (σ), L ) in
X E (e1 , X PP1 (σ), L ′ )
where P is one of cons, +
X E ((P e1 ), σ, L ) = X E (e1 , X PP1 (σ), L )
where P is one of car, cdr, null?, pair?
X E ((if e1 e2 e3 ), σ, L ) = let L ′ ← X E (e3 , σ, L ) in
let L ′′ ← X E (e2 , σ, L ) in
X E (e1 , {ε}, L ′ ∪ L ′′)
X E ((let v1 ← e1 in e2 ), σ, L ) = let L ′ ← X E (e2 , σ, L ) in
X E (e1 , σ′ , L ′ − v1 .σ′ )
where σ′ = {α | v1 .α ∈ L ′ }
X E (( f e1 . . . en ), σ, L ) = let L1 ← X E (en , X F fn (σ), L ) in
..
.
let Ln−1 ← X E (e2 , X F f2 (σ), Ln−2 ) in
X E (e1 , X F f1 (σ), Ln−1 )
We explain the definition of X E for the if expression. Since the value of the conditional
expression e1 is boolean and this value is used, the liveness access pattern with respect to
which e1 is computed is {ε}. Further, since it is not possible to statically determine whether
e2 or e3 will be executed, the liveness environment with respect to which e1 is computed is
the union of the liveness environments arising out of e2 and e3 .
3.2
Computing X P and X F
If σ is the set of access patterns specifying the liveness of the result of evaluating (P e1 . . . en ),
where P is a primitive, then X PPi (σ) gives the set of access patterns specifying the liveness
of ei . We describe the transfer functions for the primitives in our language: car, cdr, cons,
null?, pair? and +. The 0-ary constructor Nil does not accept any argument and is ignored.
5
Karkare, Khedker and Sanyal
Assume that the live access pattern for the result of the expression (car e) is α. Then,
the link that is denoted by the path labeled α starting from location [(car e)] can also be
denoted by a path 0α starting from location [e]. We can extend the same reasoning for set
of access patterns (σ) of result, i.e. every pattern in the set is prefixed by 0 to give live
access pattern of e. Also, since the cell corresponding to e is used to find the value of car,
we need to add ε to the live access patterns of e. Reasoning about (cdr e) similarly, we have
1
(σ) = {ε} ∪ {0} · σ,
(11) X Pcar
1
X Pcdr
(σ) = {ε} ∪ {1} · σ
As seen in Example 2.1, an access pattern of α for result of cons translates to an access
pattern of 0̄α for its first argument, and 1̄α for its second argument. Since cons does not
read its arguments, the access patterns of the arguments do not contain ε.
1
(12) X Pcons
(σ) = {0̄} · σ,
2
X Pcons
(σ) = {1̄} · σ
Since the remaining primitives read only the value of the arguments, the set of live access
patterns of the arguments is {ε}.
1
(13) X Pnull? (σ) = {ε},
1
X Ppair?
(σ) = {ε},
X P+1 (σ) = {ε},
X P+2 (σ) = {ε}
We now consider the transfer function for a user defined function f . If σ is the set of
access patterns specifying the liveness of the result of evaluating ( f e1 . . . en ), then X F fi (σ)
gives the set of access patterns specifying the liveness of ei . Let f be defined as:
(define ( f v1 . . . vn ) π : e)
Assume that σ is the live access pattern for the result of f . Then, σ is also the live access
/ computes live access patterns for vi (1 ≤ i ≤ n) at π. Thus, the
pattern for e. X E (e, σ, 0)
transfer function for the ith argument of f is given by:
/
(14) X F fi (σ) = {α | vi .α ∈ X E (e, σ, 0)}
1≤i≤n
The following example illustrates our analysis.
Example 3.1 Consider the program in Figure 1. To compute the transfer functions for app,
/ in terms of a variable σ. Here e is
we compute the environment transformer X E (e, σ, 0)
the body of app. The value of the liveness environment at each point in the body of app is
shown in Appendix A. From the liveness information at π1 we get:
1
1
X F app
(σ) = {ε} ∪ {00̄} · σ ∪ {1} · X F app
({1̄} · σ)
2
2
X F app
(σ) = σ ∪ X F app
({1̄} · σ)
Let epgm represent the entire program being analyzed and σpgm be the set of access
patterns describing the liveness of the result. Then, the liveness environment at various
/ The liveness environments at
points in the epgm can be computed as X E (epgm , σpgm , 0).
π14 and π12 are as follows:
Lπ14 = { w.({ε, 1, 10} ∪ {100} · σpgm ) }
Lπ12 =
(
1
({ε, 1, 10} ∪ {100} · σpgm ),
y.X F app
2
z.X F app ({ε, 1, 10} ∪ {100} · σpgm )
)
✷
We assume that the entire result of the program is needed, i.e., σpgm is {0, 1}∗.
6
Karkare, Khedker and Sanyal
4
Solving the Equations for X F
In general, the equations defining the transfer functions X F will be recursive. To solve
such equations, we start by guessing that the solution will be of the form:
(15) X F fi (σ) = I fi ∪ D if · σ,
where I fi and D if are sets of strings over the alphabet {0, 1, 0̄, 1̄}. The intuition behind this
form of solution is as follows: The function f can use its argument locally and/or copy a
part of it to the return value being computed. I fi is the live access pattern of ith argument
due to local use in f . D if is a sort of selector that selects the liveness pattern corresponding
to the ith argument of f from σ, the liveness pattern of the return value.
If we substitute the guessed form of X F fi in the equations describing it and equate the
terms containing σ and the terms without σ, we get the equations for I fi and D if . This is
illustrated in the following example.
1
(σ) from Example 3.1:
Example 4.1 Consider the equation for X F app
1
1
X F app
(σ) = {ε} ∪ {00̄} · σ ∪ {1} · X F app
({1̄} · σ)
Decomposing both sides of the equation, and rearranging the RHS gives:
1
1
1
1
Iapp
∪ Dapp
· σ = {ε} ∪ {00̄} · σ ∪ {1} · (Iapp
∪ Dapp
· {1̄} · σ)
1
1
= {ε} ∪ {1} · Iapp
∪ {00̄} · σ ∪ {1} · Dapp
· {1̄} · σ
Separating the parts that are σ dependent and the parts that are σ independent, and equating
them separately, we get:
1
1
Iapp
= {ε} ∪ {1} · Iapp
1
1
· {1̄}σ
Dapp
· σ = {00̄} · σ ∪ {1} · Dapp
1
= ({00̄} ∪ {1} · Dapp
· {1̄}) · σ
As the equations hold for any general σ, we can simplify them to:
1
1
1
1
Iapp
= {ε} ∪ {1} · Iapp
and Dapp
= {00̄} ∪ {1} · Dapp
· {1̄}
2
(σ), we get:
Similarly, from the equation describing X F app
2
2
2
2
Iapp
= Iapp
and Dapp
= {ε} ∪ Dapp
· {1̄}
The liveness environment at π12 and π14 in terms Iapp and Dapp are:
Lπ14 = { w.{100} · σpgm }
1
1
y.(Iapp
∪ Dapp
· ({ε, 1, 10} ∪ {100} · σpgm )),
Lπ12 =
2
2
· ({ε, 1, 10} ∪ {100} · σpgm ))
∪ Dapp
z.(Iapp
Solving for Iapp and Dapp gives us the desired liveness environments at these program
points.
4.1
✷
Representing Liveness by Context Free Grammars
The values of I and D variables of a transfer function are sets of strings over the alphabet
{0, 1, 0̄, 1̄}. We use context free grammars (CFG) to describe these sets. The set of terminal
symbols of the CFG is {0, 1, 0̄, 1̄}. Non-terminals and associated rules are constructed as
illustrated in Examples 4.2 and 4.3.
7
Karkare, Khedker and Sanyal
Example 4.2 Consider the following constraint from Example 4.1:
1
1
Iapp
= {ε} ∪ {1} · Iapp
1 i and the productions with right hand sides directly derived
We add non-terminal hIapp
from the constraints:
1 i → ε | 1hI 1 i
hIapp
app
The productions generated from other constraints of Example 4.1 are:
1
1
hDapp
i → 00̄ | 1hDapp
i1̄
2 i → hI 2 i
hIapp
app
2
2
hDapp
i → ε | hDapp
i1̄
These productions describe the transfer functions of app.
✷
The liveness environment at each program point can be represented as a CFG with a
start symbol for every variable. To do so, the analysis starts with hSpgm i, the non-terminal
describing the liveness pattern of the result of the program, σpgm . The productions for
hSpgm i are:
hSpgm i → ε | 0hSpgm i | 1hSpgm i
Example 4.3 Let Sπv denote the non-terminal generating liveness access patterns associated
with a variable v at program point π. For the program of Figure 1:
hSπw14 i → ε | 1 | 10 | 100hSpgm i
2 i | hD 2
2
2
2
hSπz 12 i → hIapp
app i | hDapp i1 | hDapp i10 | hDapp i100hSpgm i
1
1
1
1 i | hD 1
hSπy12 i → hIapp
app i | hDapp i1 | hDapp i10 | hDapp i100hSpgm i
✷
The access patterns in the access paths used for nullification are in canonical form
but the access patterns described by the CFGs resulting out of our analysis are not. It is
not obvious how to check the membership of a canonical access pattern in such CFGs.
To solve this problem, we need equivalent CFGs such that if α belongs to an original
⋆
CFG and α → β, where β is in canonical form, then β belongs to the corresponding new
CFG. Directly converting the reduction rules (Equations (1, 2, 3)) into productions and
adding it to the grammar results in unrestricted grammar [11]. To simplify the problem,
we approximate original CFGs by non-deterministic finite automata (NFAs) and eliminate
0̄ and 1̄ from the NFAs.
4.2
Approximating CFGs using NFAs
The conversion of a CFG G to an approximate NFA N should be safe in that the language
accepted by N should be a superset of the language accepted by G. We use the algorithm
described by Mohri and Nederhof [18]. The algorithm transforms a CFG to a restricted
form called strongly regular CFG which can be converted easily to a finite automaton.
Example 4.4 We show the approximate NFAs for each of the non-terminals in Example 4.2 and Example 4.3.
1
0
hSpgm i:
start
1 i:
hIapp
1
8
start
Karkare, Khedker and Sanyal
1̄
1
1
hDapp
i:
start
0
1̄
0̄
2
hDapp
i:
1̄
0
hSπw14 i:
start
1
start
0
0̄
hSπz 12 i :
start
0
1
0
0
1
1̄
1
hSπy12 i:
0
0
start
1
0
1
0
0
1
2 i. This is because the least solution of the equaNote that there is no automaton for hIapp
2 i → hI 2 i is 0.
1
/ Also, the language accepted by the automaton for Dapp
tion hIapp
is
app
approximate as it does not ensure that there is an equal number of 1 and 1̄ in the strings
1
generated by rules for hDapp
i.
✷
4.3
Eliminating 0̄ and 1̄ from NFA
We now describe how to convert an NFA with transitions on symbols 0̄ and 1̄ to an equivalent NFA without any transitions on these symbols.
Input: An NFA N with underlying alphabet {0, 1, 0̄, 1̄} accepting a set of access patterns
Output: An NFA N with underlying alphabet {0, 1} accepting the equivalent set of canonical access patterns
Steps:
i←0
N0 ← Equivalent NFA of N without ε-moves [11]
do {
N′i+1 ← Ni
foreach state q in Ni such that q has an incoming edge from q′
with label 0̄ and outgoing edge to q′′ with label 0 {
/⋆ bypass 0̄0 using ε ⋆/
add an edge in N′i+1 from q′ to q′′ with label ε.
}
foreach state q in Ni such that q has an incoming edge from q′
with label 1̄ and outgoing edge to q′′ with label 1 {
/⋆ bypass 1̄1 using ε ⋆/
add an edge in N′i+1 from q′ to q′′ with label ε.
}
Ni+1 ← Equivalent NFA of N′i+1 without ε-moves
i ← i+1
} while (Ni 6= Ni−1 )
N ← Ni
delete all edges with label 0̄ or 1̄ in N.
The algorithm repeatedly introduces ε edges to bypass a pair of consecutive edges labeled
0̄0 or 1̄1. The process is continued till a fixed point is reached. When the fixed point is
reached, the resulting NFA contains the canonical access patterns corresponding to all the
9
Karkare, Khedker and Sanyal
access patterns in the original NFA. The access patterns not in canonical form are deleted
by removing edges labeled 0̄ and 1̄. Note that by our reduction rules if α is accepted by
⋆
N and α → ⊥, then ⊥ should be accepted by N, However, N returned by our algorithm
does not accept ⊥. This is not a problem because the access patterns which are tested for
membership against N do not include ⊥ as well.
Example 4.5 We show the elimination of 0̄ and 1̄ for the automata for hSπy12 i and hSπz 12 i.
The automaton for hSπw14 i remains unchanged as it does not contain transitions on 0̄ and 1̄.
The automata at the termination of the loop in the algorithm are:
1̄
1
hSπy12 i:
start
0
0̄
0
0
0
1
0
0
1
1̄
hSπz 12 i
:
start
0
0
0
1
0
1
Eliminating the edges labeled 0̄ and 1̄, and removing the dead states gives:
1
hSπy12 i:
start
0
0
0
0
hSπz 12 i :
start
0
1
0
0
1
1
The language accepted by these automata represent the live access paths corresponding to
y and z at π12 .
✷
We now prove the termination and correctness of our algorithm.
Termination
Termination of the algorithm follows from the fact that every iteration of do-while loop
adds new edges to the NFA, while old edges are not deleted. Since no new states are added
to NFA, only a fixed number of edges can be added before we reach a fix point.
Correctness
The sequence of obtaining N from N can be viewed as follows, with Nm denoting the NFA
at the termination of while loop:
N
deletion
of ε-edges
N0
addition
of ε-edges
N′1
deletion
of ε-edges
N1
addition
···
addition
of ε-edges
of ε-edges
deletion of
Nm
N
0̄, 1̄ edges
N′i
deletion
of ε-edges
Ni · · ·
deletion
of ε-edges
Nm
Then, the languages accepted by these NFAs have the following relation:
L(N) = L(N0 ) ⊆ L(N′1 ) = L(N1 ) ⊆ · · · ⊆ L(N′i ) = L(Ni ) ⊆ · · · = L(Nm )
L(N) ⊆ L(Nm )
We first prove that the addition of ε-edges in the while loop does not add any new
information, i.e. any access pattern accepted by the NFA after the addition of ε-edges is a
reduced version of some access pattern existing in the NFA before the addition of ε-edges.
10
Karkare, Khedker and Sanyal
Lemma 4.6 for i > 0, if α ∈ L(Ni ) then there exists α′ ∈ L(Ni−1 ) such that α′ → α.
⋆
Proof. As L(Ni ) = L(N′i ), we have α ∈ L(N′i ). Only difference between N′i and Ni−1 is that
N′i contains some extra ε-edges. Thus, any ε-edge free path in N′i is also in Ni−1 . Consider
a path p in N′i that accepts α. Assume the number of ε edges in p is k. The proof is by
induction on k.
(BASE) k = 0, i.e. p does not contains any ε-edge: As the path p is ε-edge free, it must be
⋆
present in Ni−1 . Thus, Ni−1 also accepts α. α → α.
(HYPOTHESIS) For any α ∈ L(Ni ) with accepting path p having less than k ε-edges there
⋆
exists α′ ∈ L(Ni−1 ) such that α′ → α.
(INDUCTION) p contains k ε-edges e1 , . . . , ek : Assume e1 connects states q′ and q′′ in N′i .
By construction, there exists a state q in N′i such that there is an edge e′1 from q′ to q with
label 0̄(1̄) and an edge e′′1 from q to q′′ with label 0(1) in N′i . Replace e1 by e′1 e′′1 in p to get a
1
new path p′′ in N′i . Let α′′ be the access pattern accepted by p′′ . Clearly, α′′ → α. Since p′′
has k − 1 ε-edges, α′′ is accepted by N′i along a path (p′′ ) that has less than k ε-edges. By
1
induction hypothesis, we have α′ ∈ L(Ni−1 ) such that α′ → α′′ . This along with α′′ → α
⋆
✷
gives α′ → α.
⋆
Corollary 4.7 for each α ∈ L(Nm ), there exists α′ ∈ L(N) such that α′ → α.
⋆
Proof. The proof is by induction on m, and using Lemma 4.6.
✷
The following lemma shows that the the language accepted by Nm is closed with respect
to reduction of access patterns.
Lemma 4.8 For α ∈ L(Nm ), if α → α′ and α′ 6= ⊥, then α′ ∈ L(Nm ).
⋆
k
Proof. Assume α → α′ . The Proof is by induction on k, number of steps in reduction.
0
(BASE) case k = 0 is trivial as α → α.
k−1
(HYPOTHESIS) Assume that for α ∈ L(Nm ), if α → α′ , then α′ ∈ L(Nm ).
k
k−1
1
(INDUCTION) α ∈ L(Nm ), α → α′ . There exists α′′ such that: α → α′′ → α′ . By
induction hypothesis, we have α′′ ∈ L(Nm ).
1
For α′′ → α′ to hold we must have α′′ = α1 0̄0α2 and α′ = α1 α2 , or α′′ = α1 1̄1α2 and
α′ = α1 α2 . Consider the case when α′′ = α1 0̄0α2 . Any path in Nm accepting α′′ must have
the following structure (The states shown separately may not necessarily be different):
start
q0
α1
q′
0̄
q
0
q′′
α2
qF
As Nm is the fixed point NFA for the iteration process described in the algorithm, adding
an ε-edge between states q′ and q′′ will not change the language accepted by Nm . But, the
access pattern accepted after adding an ε-edge is α1 α2 = α′ . Thus, α′ ∈ L(Nm ). The case
when α′′ = α1 1̄1α2 is identical.
✷
Corollary 4.9 For α ∈ L(N), if α → α′ and α′ 6= ⊥, then α′ ∈ L(Nm ).
⋆
Proof. L(N) ⊆ L(Nm ) ⇒ α ∈ L(Nm ). The proof follows from Lemma 4.8.
✷
The following theorem asserts the equivalence of N and N with respect to the equivalence of access patterns, i.e. every access pattern in N has an equivalent canonical access
11
Karkare, Khedker and Sanyal
pattern in N, and for every canonical access pattern in N, there exists an equivalent access
pattern in N.
Theorem 4.10 Let N be an NFA with underlying alphabet {0, 1, 0̄, 1̄}. Let NFA N be the
NFA with underlying alphabet {0, 1} returned by the algorithm. Then,
(i) if α ∈ L(N), β is a canonical access pattern such that α → β and β 6= ⊥, then β ∈ L(N).
⋆
(ii) if β ∈ L(N) then there exists an access pattern α ∈ L(N) such that α → β.
⋆
Proof.
(i) From Corollary 4.9: α ∈ L(N), α → β and β 6= ⊥ ⇒ β ∈ L(Nm ). As β is in canonical
form, the path accepting β in Nm consists of edges labeled 0 and 1 only. The same
path exists in N. Thus N also accepts β ⇒ β ∈ L(N).
⋆
(ii) L(N) ⊆ L(Nm ) ⇒ β ∈ L(Nm ). Using Corollary 4.7, there exists α ∈ L(N) such that
⋆
α → β.
✷
5
An Application of Liveness Analysis
The result of liveness analysis can be used to decide whether a given access path v.α can be
nullified at a given program point π. Let the link corresponding to v.α in the memory graph
be l. A naive approach is to nullify v.α if it does not belongs to the liveness environment
at π. However, the approach is not safe because of two reasons: (a) The link l may be
used beyond π through an alias, and may therefore be live. (b) a link l ′ in the access path
from the root variable v to l may have been created along one execution path but not along
another. Since the nullification of v.α requires the link l ′ to be dereferenced, a run time
exception may occur.
To solve the first problem, we need an alias analysis phase to detect sharing of links
among access paths. A link in the memory graph can be nullified at π if none of the access
paths sharing it are live at π . To solve the second problem, we need an availability analysis
phase. It detects whether all links in the access path have been created along all execution
paths reaching π. The results of these analysis are used to filter out those access paths
whose nullification may be unsafe. We do not address the descriptions of these analyses in
this paper.
6
Related Work
In this paper, we have described a static analysis for inferring dead references in first order
functional programs. We employ a context free grammar based abstraction for the heap.
This is in the spirit of the work by Jones and Muchnick [13] for functional programs. The
existing literature related to improving memory efficiency of programs can be categorized
as follows:
Compile time reuse. The method by Barth [2] detects memory cells with zero reference count and reallocates them for further use in the program. Jones and Le Metayer [15]
describe a sharing analysis based garbage collection for reusing of cells. Their analysis
incorporates liveness information: A cell is collected even when it is shared provided expressions sharing it do not need it for their evaluation.
12
Karkare, Khedker and Sanyal
Explicit reclamation. Shaham et. al. [25] use an automaton called heap safety automaton to model safety of inserting a free statement at a given program point. The analysis is
based on shape analysis [20,21] and is very precise. The disadvantage of the analysis is
that it is very inefficient and takes large time even for toy programs. Free-Me [7] combines
a lightweight pointer analysis with liveness information that detects when short-lived objects die and insert statements to free such objects. The analysis is simpler and cheaper as
the scope is limited. The analysis described by Inoue et. al. [12] detects the scope (function) out of which a cell becomes unreachable, and claims the cell using an explicit reclaim
procedure whenever the execution goes out of that scope. Like our method, the result of
their analysis is also represented using CFGs. The main difference between their work and
ours is that we detect and nullify dead links at any point of the program, while they detect
and collect objects that are unreachable at function boundaries. Cherem and Rugina [5]
use a shape analysis framework [8] to analyze a single heap cell at a time for deallocation.
However, multiple iterations of the analysis and the optimization steps are required, since
freeing a cell might result in opportunities for more deallocations.
Making dead objects unreachable. The most popular approach to make dead objects
unreachable is to identify live variables in the program to reduce the root set to only the
live reference variables [1]. The major drawback of this approach is that all heap objects
reachable from the live root variables are considered live, even if some of them may not be
used by the program. Escape analysis [3,4,6] based approaches discover objects escaping
a procedure (an escaping object being an object whose lifetimes outlives the procedure that
created it). All non-escaping objects are allocated on stack, thus becoming unreachable
whenever the creating procedure exits. In Region based garbage collection [9], a static
analysis called region inference [26] is used to identify regions that are storage for objects.
Normal memory blocks can be allocated at any point in time; they are always allocated
in a particular region and are deallocated at the end of that region’s lifetime. Approaches
based on escape analysis and region inference detect garbage only at the boundaries of
certain predefined areas of the program. In our previous work [17], we have used bounded
abstractions of access paths called access graphs to describe the liveness of memory links
in imperative programs and have used this information to nullify dead links.
A related work due to Heine and Lam [10] attempts to find potential memory leaks
in C/C++ programs by detecting the earliest point in a program when an object becomes
unreachable.
7
Conclusions
In this paper we presented a technique to compute liveness of heap data in functional programs. This information could be used to nullify links in heap memory to improve garbage
collection. We have abstracted the liveness information in the form of a CFG, which is then
converted to NFAs. This conversion implies some imprecision. We present a novel way to
simplify the NFAs so they directly describe paths in the heap. Unlike the method described
by Inoue et. al. [12], our simplification does not cause any imprecision.
In future, we intend to take this method to its logical conclusion by addressing the issue
of nullification. This would require us to perform alias analysis which we feel can be done
in a similar fashion. We also feel that with minor modification our method can be used for
dead code elimination and intend to extend our analysis to higher order languages.
13
Karkare, Khedker and Sanyal
References
[1] Agesen, O., D. Detlefs and J. E. Moss, Garbage collection and local variable type-precision and liveness in Java virtual
machines, in: PLDI ’98: Proceedings of the ACM SIGPLAN 1998 conference on Programming language design and
implementation (1998), pp. 269–279.
[2] Barth, J. M., Shifting garbage collection overhead to compile time, Commun. ACM 20 (1977), pp. 513–518.
[3] Blanchet, B., Escape analysis for object-oriented languages: application to Java, in: OOPSLA ’99: Proceedings of the
14th ACM SIGPLAN conference on Object-oriented programming, systems, languages, and applications (1999), pp.
20–34.
[4] Blanchet, B., Escape analysis for JavaT M : Theory and practice, ACM Transactions on Programming Languages and
Systems 25 (2003), pp. 713–775.
[5] Cherem, S. and R. Rugina, Compile-time deallocation of individual objects, in: ISMM ’06: Proceedings of the 2006
international symposium on Memory management (2006), pp. 138–149.
[6] Choi, J.-D., M. Gupta, M. Serrano, V. C. Sreedhar and S. Midkiff, Escape analysis for Java, in: OOPSLA ’99:
Proceedings of the 14th ACM SIGPLAN conference on Object-oriented programming, systems, languages, and
applications (1999), pp. 1–19.
[7] Guyer, S. Z., K. S. McKinley and D. Frampton, Free-me: a static analysis for automatic individual object
reclamation, in: PLDI ’06: Proceedings of the 2006 ACM SIGPLAN conference on Programming language design
and implementation (2006), pp. 364–375.
[8] Hackett, B. and R. Rugina, Region-based shape analysis with tracked locations, in: POPL ’05: Proceedings of the 32nd
ACM SIGPLAN-SIGACT symposium on Principles of programming languages (2005), pp. 310–323.
[9] Hallenberg, N., M. Elsman and M. Tofte, Combining region inference and garbage collection, in: PLDI ’02:
Proceedings of the ACM SIGPLAN 2002 Conference on Programming language design and implementation (2002),
pp. 141–152.
[10] Heine, D. L. and M. S. Lam, A practical flow-sensitive and context-sensitive c and c++ memory leak detector, in: PLDI
’03: Proceedings of the ACM SIGPLAN 2003 conference on Programming language design and implementation (2003),
pp. 168–181.
[11] Hopcroft, J. E. and J. D. Ullman, “Introduction To Automata Theory, Languages, And Computation,” Addison-Wesley
Longman Publishing Co., Inc., Boston, MA, USA, 1990.
[12] Inoue, K., H. Seki and H. Yagi, Analysis of functional programs to detect run-time garbage cells, ACM Trans. Program.
Lang. Syst. 10 (1988), pp. 555–578.
[13] Jones, N. D. and S. S. Muchnick, Flow analysis and optimization of lisp-like structures, in: POPL ’79: Proceedings of
the 6th ACM SIGACT-SIGPLAN symposium on Principles of programming languages (1979), pp. 244–256.
[14] Jones, R. and R. Lins, “Garbage collection: algorithms for automatic dynamic memory management,” John Wiley &
Sons, Inc., New York, NY, USA, 1996.
[15] Jones, S. B. and D. L. Metayer, Compile-time garbage collection by sharing analysis, in: FPCA ’89: Proceedings of the
fourth international conference on Functional programming languages and computer architecture (1989), pp. 54–74.
[16] Karkare, A., A. Sanyal and U. Khedker, Effectiveness of garbage collection in MIT/GNU scheme,
http://arxiv.org/abs/cs/0611093 (2006).
[17] Khedker, U., A. Sanyal and A. Karkare, Heap reference analysis using access graphs, Submitted to ACM Transactions
on Programming Languages and Systems, copy available at http://arxiv.org/abs/cs.PL/0608104 (2006).
[18] Mohri, M. and M.-J. Nederhof, Regular approximation of context-free grammars through transformation, in: J.-C.
Junqua and G. van Noord, editors, Robustness in Language and Speech Technology, Kluwer Academic Publishers,
Dordrecht, 2000 pp. 251–261.
[19] Röjemo, N. and C. Runciman, Lag, drag, void and use—heap profiling and space-efficient compilation revisited, in:
ICFP ’96: Proceedings of the first ACM SIGPLAN international conference on Functional programming (1996), pp.
34–41.
[20] Sagiv, M., T. Reps and R. Wilhelm, Parametric shape analysis via 3-valued logic, in: POPL ’99: Proceedings of the
26th ACM SIGPLAN-SIGACT symposium on Principles of programming languages (1999), pp. 105–118.
[21] Sagiv, M., T. Reps and R. Wilhelm, Parametric shape analysis via 3-valued logic, ACM Transactions on Programming
Languages and Systems 24 (2002), pp. 217–298.
[22] Shaham, R., E. K. Kolodner and M. Sagiv, On the effectiveness of gc in java, in: ISMM ’00: Proceedings of the 2nd
international symposium on Memory management (2000), pp. 12–17.
[23] Shaham, R., E. K. Kolodner and M. Sagiv, Heap profiling for space-efficient java, in: PLDI ’01: Proceedings of the
ACM SIGPLAN 2001 conference on Programming language design and implementation (2001), pp. 104–113.
[24] Shaham, R., E. K. Kolodner and M. Sagiv, Estimating the impact of heap liveness information on space consumption
in Java, in: ISMM ’02: Proceedings of the 3rd international symposium on Memory management (2002), pp. 64–75.
[25] Shaham, R., E. Yahav, E. K. Kolodner and M. Sagiv, Establishing local temporal heap safety properties with
applications to compile-time memory management, Sci. Comput. Program. 58 (2005), pp. 264–289.
[26] Tofte, M. and L. Birkedal, A region inference algorithm, ACM Transactions on Programming Languages and Systems
20 (1998), pp. 724–767.
14
A
Live Access Patterns
Liveness Environment
Liveness Environment
Point
for e at π
after e
at π
(π)
(σ)
(L )
π1
π4
15
π5
π7
π8
π9
)
(
1
list1.({ε} ∪ {00̄} · σ ∪ {1} · X F app
({1̄} · σ)),
1
list1.({ε} ∪ {00̄} · σ ∪ {1} · X F app
({1̄} · σ)),
)
(
1
list1.({ε} ∪ {00̄} · σ ∪ {1} · X F app
({1̄} · σ)),
)
1
list1.({ε} ∪ {00̄} · σ ∪ {1} · X F app
({1̄} · σ)),
)
(
1
list1.({ε} ∪ {00̄} · σ ∪ {1} · X F app
({1̄} · σ)),
)
(
1
list1.({ε} ∪ {00̄} · σ ∪ {1} · X F app
({1̄} · σ)),
)
(
1
list1.({ε} ∪ {00̄} · σ ∪ {1} · X F app
({1̄} · σ)),
)
(
1
list1.({ε} ∪ {00̄} · σ ∪ {1} · X F app
({1̄} · σ)),
0/
{ε}
(
{ε}
(
σ
(
1
list1.({ε} ∪ {00̄} · σ ∪ {1} · X F app
({1̄} · σ)),
2
list2.(σ ∪ X F app
({1̄} · σ))
2
list2.(σ ∪ X F app
({1̄} · σ))
2
list2.X F app
({1̄} · σ)
0/
σ
{0̄} · σ
(
{ε} ∪ {00̄} · σ
(
{1̄} · σ
X F app ({1̄} · σ)
1
1
list1.{1} · X F app
({1̄} · σ),
2
list2.X F app
({1̄} · σ)
1
list1.{1} · X F app
({1̄} · σ),
2
list2.X F app
({1̄} · σ)
2
list2.(σ ∪ X F app
({1̄} · σ))
2
list2.(σ ∪ X F app
({1̄} · σ))
2
list2.(σ ∪ X F app
({1̄} · σ))
2
list2.X F app
({1̄} · σ)
2
list2.X F app
({1̄} · σ)
2
list2.X F app
({1̄} · σ)
0/
(
1
list1.{1} · X F app
({1̄} · σ),
)
{list2.X F app ({1̄} · σ)}
(
1
list1.{1} · X F app
({1̄} · σ),
)
(
1
list1.{1} · X F app
({1̄} · σ),
2
π10
{1} · X F app ({1̄} · σ)
{list2.X F app ({1̄} · σ)}
π11
2
X F app
({1̄} · σ)
0/
2
2
list2.(σ ∪ X F app
({1̄} · σ))
2
2
list2.X F app
({1̄} · σ)
2
list2.X F app
({1̄} · σ)
2
list2.X F app
({1̄} · σ)
2
{list2.X F app
({1̄} · σ)}
)
)
)
)
)
)
Karkare, Khedker and Sanyal
π3
π6
1
list1.({ε} ∪ {00̄} · σ ∪ {1} · X F app
({1̄} · σ)),
σ
π2
(X E (e, σ, L ))
(
Computation of Liveness for app
Program
| 6cs.PL
|
Exact 3D seismic data reconstruction using Tubal-Alt-Min algorithm
Feng Qian1 , Quan Chen2 , Ming-Jun Su3 , Guang-Min Hu2 , Xiao-Yang Liu∗
1 Center of Information Geoscience, University of Electronic Science and Technology of China,
2 School of Rresources and Environment, University of Electronic Science and Technology of China,
3 PetroChina Research Institute of Petroleum Exploration and Development (RIPED)-Northwest, Lanzhou, China,
∗
Department of Electrical Engineering, Columbia University, NY, USA
arXiv:1704.02445v1 [cs.CE] 8 Apr 2017
SUMMARY
Data missing is an common issue in seismic data, and many
methods have been proposed to solve it. In this paper, we
present the low-tubal-rank tensor model and a novel tensor
completion algorithm to recover 3D seismic data. This is a
fast iterative algorithm, called Tubal-Alt-Min which completes
our 3D seismic data by exploiting the low-tubal-rank property
expressed as the product of two much smaller tensors. TubalAlt-Min alternates between estimating those two tensor using
least squares minimization. We evaluate its reconstruction performance both on synthetic seismic data and land data survey. The experimental results show that compared with the
tensor nuclear norm minimization algorithm, Tubal-Alt-Min
improves the reconstruction error by orders of magnitude.
INTRODUCTION
Seismic data quality is vital to various geophysical processing. However, due to the financial and physical constraints,
the real seismic survey data are usually incomplete. Seismic
data reconstruction is a complex problem, and a large number
of researchers have devoted themselves to the research of this
field. Consequently, many approaches have been proposed to
handle this problem. From the point of view of data organization, these methods can be divided into two categories.
The low dimensional based methods, such as transform-based
methods which utilize the properties of the seismic trace in an
auxiliary domain (Hennenfent et al., 2010). Liner prediction
theory use the predictability of the signal in the f-x or t-x domain (Naghizadeh and Sacchi, 2007). And methods which exploit the low-rank nature of seismic data embedded in Hankel
matrices (Oropeza and Sacchi, 2011).
Those low dimensional based methods usually ignores the spatial structure of the seismic traces. While the spatial structure coherence is very important for the seismic data completion, hence recent developments in high dimensional tensor
completion approaches exploit various tensor decomposition
model are widely used in seismic data reconstruction (Kreimer
et al., 2013; Ely et al., 2015a). Those high-order tensor decomposition approaches have became a trend for seismic data
completion, and the exist different definitions for tensor decomposition that lead to different tensor completion model,
i.e. the higher-order singular value decomposition (HOSVD)
(Kreimer and Sacchi, 2011), the tuker decomposition (Da Silva
and Herrmann, 2013), and tensor SVD (tSVD) decomposition
(Ely et al., 2015b).
In this paper, we focus on a new seismic data reconstruction
algorithm which based on low-tubal-rank tensor decomposition model possesses extremely high precision seismic data
recovery performance. Because of the high redundancy or coherence between one seismic trace to the others (Ely et al.,
2015a), we assume that the fulled sampled seismic volume has
low-tubal-rank property in the tSVD domain. Therefore, we
can solve the seismic tensor completion problem through an
alternating minimization algorithm for low-tubal-rank tensor
completion (Tubal-Alt-Min) (Liu et al., 2016a,b). We have
evaluated the performance of this approach on both synthetic
and field seismic data.
NOTATIONS
The data in seismic survey is a natural high-dimensional tensor, such as the 3D poststack seismic data which consists of
one time or frequency dimension and two spatial dimensions
corresponding to xline and inline directions.
Throughout the paper, we denote those 3D seismic tensor in
time domain by uppercase calligraphic letter, T ∈ Rm×n×k ,
and denote the frequency domain 3D seismic tensor by Te ∈
Rm×n×k correspondingly. Uppercase letter A ∈ Rm×n denotes
matrix, and lowercase boldface letter x ∈ Rn denotes vector.
Let [n] denotes the set {1, 2, . . . , n}. In addition, we introduce
an important tensor operator t-product (Kilmer et al., 2013).
t-product. The tensor-product T = X ∗ Y of X ∈ Rn1 ×n2 ×k
n2 ×n3 ×k is a tensor of size n × n × k, T (i, j, :) =
and
1
2
Pn2Y ∈ R
s=1 X (i, s, :) ∗ Y(s, j, :), for i ∈ [n1 ] and j ∈ [n3 ].
PROBLEM SETUP
From what we have stated above, we know that seismic data
comprises many traces that provide a spatio-temporal sampling of the reflected wavefield. However, caused by various
factors, such lost many important informations. Such as the
exist of reservoir, residential or any other obstacle in the seismic data acquisition areas will lead to under-sampled seismic
record. The missing traces will complicate certain data processing steps such as the prediction accuracy of underground
reservoirs. Hence, the completion step in seismic data processing is of grate significance.
In this paper, we explored the relationship between low-tubalrank and under-sampled rate firstly, and found that low-tubal
property is positively correlated with the sampling rate. Figure
1 shows the detail experiment result. Base on this work, we as-
Empirical CDF
100
SOLUTION
90
In the Tubal-Alt-Min algorithm, the target 3D seismic volume
Tb ∈ Rm×n×k can be decomposed as Tb = X ∗ Y, X ∈ Rm×r×k ,
Y ∈ Rr×n×k , and r is the target tubal-rank. With this decomposition, the problem (1) reduces to
80
70
CDF
60
50
Tb =
40
20% sampling rate
40% sampling rate
60% sampling rate
80% sampling rate
100% sampling rate
30
20
10
10
20
30
40
50
60
70
80
90
kPΩ (T ) − PΩ (X ∗ Y)k2F .
(2)
This cost function can be solved by the alternating minimization algorithm for low tubal rank tensor completion designed
by Liu. The main algorithm steps are showing as Alg 1.
0
0
argmin
X ∈Rm×r×k ,Y∈Rr×n×k
100
Top portion of singular value %
Figure 1: The empirical CDF of singular value.
Alg. 1 Tubal Alternating Minimization: AM(PΩ (T ), Ω, L, r)
Input: Observation set Ω ∈ [n] × [m] × [k] and the corresponding elements PΩ (T ) , number of iterations L, target tubal-rank
r.
1: X0 ← Initialize(PΩ (T ), Ω, r),
2: Xe0 ← fft(X0 , [], 3), TeΩ ← fft(TΩ , [], 3),
eΩ ← fft(PΩ , [], 3),
P
3: for l = 1 to L do
eΩ , Xel−1 , r ),
el ← LS Y( TeΩ , P
Y
eΩ , Y
el , r),
Xel ← LS X( TeΩ , P
4: end for
eL , [], 3),
5: X ← ifft( XeL , [], 3); Y ← ifft( Y
The key problem of Alg. 1 is the tensor least square minimization, which was solved by the providing methods in Liu’s
paper. The main ideal is to decompose (2) into n separate standard least squares minimization problem in the frequency domain. Then, we just need to solve a least square problem like
the following form each step:
b
x = argmin kb − A1 A2 xk2F .
(3)
x∈Rrk×1
PERFORMANCE EVALUATION
To evaluate the performance of the algorithm we adopted, two
commonly used evaluation criteria in seismic data completion
filed have been used for comparison - the reconstruction error
and the convergence speed.
• Reconstruction error: here we adopted the relative square error as a scale standard which defined as RSE =
kTb − T kF /kT kF .
Output: Pair of tensors ( X , Y ).
sume that the full sampled seismic data volume has low-tubalrank property and under-sampled traces will increase the tubal
rank. Therefore, the poststack 3D seismic data reconstruction can be tackled with tensor completion tools that using
the low tubal rank property. The statement above transform
to mathematical representation is that T ∈ Rm×n×k is a 3D
seismic data with tubal-rank equal to r. Then, by observing a
set Ω ⊂ Rm×n×k of T ’s elements, we get the under-sampled
seismic data TΩ . Then, our aim is to recover T from TΩ . The
tensor reconstruction problem can be formulated as following
• Convergence speed: we measure decreasing rate of the
RSE across iterations by linearly fitting the measured
RSEs.
In order to have an intuitive performance comparison, we compared with two other seismic data volume completion algorithm. The Parallel matrix factorization algorithm (PMF) (Gao
et al., 2015) and the tensor singular value decomposition based
algorithm, also called tensor nuclear norm algorithm (TNN)
(Ely et al., 2015b) for seismic reconstruction. We applied those
algorithm both on synthetic and real seismic data, the following subsections will demonstrate the detail performance comparison.
optimization function:
Tb = argmin rank(X ),
X ∈Rm×n×k
(1)
s.t. PΩ (X ) = PΩ (T ).
Here, PΩ (·) denote the projection of a tensor onto the observed
set Ω, TΩ = PΩ (T ). The Tubal-Alt-Min algorithm proposed
by Liu recently, which can complete the low tubal rank tensor with very high currency in several iterations is a perfect
approach to solve this problem.
Synthetic data example
We use a Ricker wavelet with central frequency of 40 Hz to
generate a simple 3D seismic model with two dipping planes.
The seismic data corresponds to a spatial tensor of size 64 ×
64 × 256, 256 time samples with the time sampling rate of 1
ms and 64 corresponding to inline and xline direction. Then,
through tSVD decomposition we get U, S, and V. According
to the decomposition result, we choose the first 2 tubes of S,
and make other tubes elements equate to zeros, form a new
b Then we get the tensor Tb = U ∗ Sb ∗ V which tubaltensor S.
rank is 2.
Reconstruction error
RSE in log-scale
100
PMF
TNN
Tuabl-Alt-Min
10-5
10-10
10-15
(a)
0
10
20
30
40
50
60
70
80
90
(b)
100
Sampling rate %
Trace number
0
20
40
60
0
Figure 2: Reconstruction error for the synthetic data set as a
function of the downsampling rate.
Time(ms)
50
Convergence speed
100
PMF
TNN
Tubal-Alt-Min
150
200
10-1
RSE in log-scale
100
250
10-2
(c)
Figure 4: (a) Full sampled synthetic data. (b) Under-sampled
measured data for the case when 40% of the traces in (a) was
removed. (c) The recovered traces of (b), it’s RSE< 1e − 4. (d)
The relative residual between (c) and (a), when sampling rate
equal to 40%, there almost no residual between original data
and reconstruction data.
10-3
10-4
(d)
0
10
20
30
40
50
60
70
80
Iterations
Figure 3: Convergence speed for the three algorithms on synthetic data.
Firstly, we apply Algorithm 1 to decompose the low-tuabl-rank
tensor of our data set at different sampling rate vary from 5%
to 95% and set the tubal-rank equate to 2. Using these decompositions, we generate our reconstructed data and compare the
error between the low-tubal-rank reconstruction and full sampled data. Applying two other algorithm we stated above to
reconstruct the data at the same condition. Because of all of
the three algorithms include random sampling operator, we
averaged those algorithms’ performance under 20 above experiments. Figure 2 shows the relative square error (RSE) of
the three algorithms. From these curves, we see that the performance of our algorithm is very outstanding. Even there are
only 30% sampling traces, it can also get a relative perfect reconstruction data. As the increase of sampling rate, the reconstruction error of Tubal-Alt-Min algorithm decrease rapidly.
Comparing with TNN, when the sampling rate over 50%, it
improves the recovery error by orders of magnitude. It’s recovery error also better than the other algorithm almost in the
whole sampling rate range.
Secondly, to evaluate the reconstruction error, we fixed the sa-
mpling rate to be 40% and set the RSE tolerance of the three algorithms equal to 1e-4. Then evaluate relative square error and
convergence speed for the three algorithms. Figure 3 shows
that Tubal-Alt-Min convergence at 9th iteration, TNN-ADMM
terminates at the 60-th iteration, PMF used 77-th iterations to
reach the preset threshold. The convergence speed of our algorithm is obviously better than other algorithms.
Then, we fixed the sampling rate at 40% to evaluate the completion performance on seismic data. The portion of reconstruction result shows in figure 4. From figure 4 (b), 4 (c) and 4 (d),
we observe that the missing traces are effectively recovered.
Filed data example
We also test the performance of the reconstruction method on
a land data set that was acquired to monitor a heavy oil field
in Anyue mountain, China. Our test data is a 25 × 200 × 600
tensor cut from the full-sampled area of the filed data. On this
basis, we randomly sample 50% traces in the seismic tensor,
then reconstruct it. To reconstruct the traces, we evaluated the
low-tubal-rank of the data manually. Determined by the tSVD,
we found it’s first 16 eigentubes’s l2 norm is much larger than
the rest in almost 40 times. Consequently, we set the parameter
of Alg. 1 r = 16. Figure 4 shows portion of the result cut along
inline direction. From the comparison of figure 5 (a), 5 (c) and
Columns
100
Columns
150
200
50
100
100
200
200
Time(ms)
Time(ms)
50
300
100
150
200
is that it can achieve very high recovery accuracy in several it-
300
400
400
500
500
(a)
600
600
(b)
Columns
100
150
200
50
100
100
200
200
Time(ms)
Time(ms)
50
(a)
Columns
300
150
200
300
400
400
500
500
600
100
(b)
600
(c)
(d)
Figure 6: The reconstruction result of filed data. (a) The
horizontal-slice of filed data from Anyue survey areas. (b)
Randomly sampled 40% trances from (a). (c) The data recovered by Tubal-Alt-Min algorithm, it’s RSE reached 1e-7. (d)
The data recovered by TNN whose RSE converged to 1e-3.
(c)
5 (d), the reconstruction performance is such amazing that almost no residual between 5 (a) and 5 (c) Compared with the
reconstruction result of TNN, it seems no difference. However,
the RSE gap between them almost have a few orders of magnitude. Figure 5 shows the difference between the specific details
of the seismic data traces restored by the two algorithms.
Figure 7: The reconstruction traces of filed data. (a) The
lateral-slices of filed data from Anyue survey areas. (b) Traces
recovered by TNN. (c) Traces recovered by our algorithm.
From the recovered traces in the blue rectangles, our algorithm
has an obvious performance advantages, which can recover
more details of the seismic traces.
CONCLUSION
In this article, we formulate the problem of poststack seismic
data reconstruction as an low-tubal-rank decomposition problem. The adopted method is based on the alternating minimization approach for low-tubal rank tensor completion. The
unknown low-tubal-rank tensor is parameterized as the product of two much smaller tensors with the low-tubal-rank property being automatically incorporated, and Tubal-Alt-Min alternates between estimating those two tensors using tensor least
squares minimization. The biggest advantage of our algorithm
erations. From the experimental result, it potently proved that
compared with contrast algorithms our algorithm improves the
recovery error by orders of magnitude with much better convergence speed for higher sampling rates.
ACKNOWLEDGMENTS
We would like to acknowledge financial support from the National Natural Science Foundation of China (Grant No.U1562218).
REFERENCES
Da Silva, C., and F. Herrmann, 2013, Hierarchical tucker tensor optimization-applications to 4d seismic data interpolation: Presented at the 75th EAGE Conference & Exhibition incorporating SPE EUROPEC 2013.
Ely, G., S. Aeron, N. Hao, and M. E. Kilmer, 2015a, 5d and 4d pre-stack seismic data completion using tensor nuclear norm (tnn),
in SEG Technical Program Expanded Abstracts 2013: Society of Exploration Geophysicists, 3639–3644.
Ely, G. T., S. Aeron, N. Hao, and M. E. Kilmer, 2015b, 5d seismic data completion and denoising using a novel class of tensor
decompositions: Geophysics, 80, V83–V95.
Gao, J., A. Stanton, and M. D. Sacchi, 2015, Parallel matrix factorization algorithm and its application to 5d seismic reconstruction
and denoising: Geophysics, 80, V173–V187.
Hennenfent, G., L. Fenelon, and F. J. Herrmann, 2010, Nonequispaced curvelet transform for seismic data reconstruction: A
sparsity-promoting approach: Geophysics, 75, WB203–WB210.
Kilmer, M. E., K. Braman, N. Hao, and R. C. Hoover, 2013, Third-order tensors as operators on matrices: A theoretical and
computational framework with applications in imaging: SIAM Journal on Matrix Analysis and Applications, 34, 148–172.
Kreimer, N., and M. D. Sacchi, 2011, A tensor higher-order singular value decomposition (hosvd) for pre-stack simultaneous
noise-reduction and interpolation: Seg Technical Program Expanded Abstracts, 3069–3074.
Kreimer, N., A. Stanton, and M. D. Sacchi, 2013, Tensor completion based on nuclear norm minimization for 5d seismic data
reconstruction: Geophysics, 78, V273–V284.
Liu, X.-Y., S. Aeron, V. Aggarwal, and X. Wang, 2016a, Low-tubal-rank tensor completion using alternating minimization: SPIE
Defense+ Security, International Society for Optics and Photonics, 984809–984809.
——–, 2016b, Low-tubal-rank tensor completion using alternating minimization: 9848, 984809.
Naghizadeh, M., and M. D. Sacchi, 2007, Multistep autoregressive reconstruction of seismic records: Geophysics, 72, V111–V118.
Oropeza, V., and M. Sacchi, 2011, Simultaneous seismic data denoising and reconstruction via multichannel singular spectrum
analysis: Geophysics: Geophysics, 76, V25–V32.
| 5cs.CE
|
1
When Traffic Flow Prediction Meets Wireless Big
Data Analytics
arXiv:1709.08024v1 [cs.AI] 23 Sep 2017
Yuanfang Chen∗† , Mohsen Guizani‡ , Yan Zhang§ , Lei Wang¶ , Noel Crespi∗ , Gyu Myoung Leek
∗ Institut Mines-Télécom, Télécom SudParis, France
† Department of Computer, Université Pierre et Marie CURIE, France
‡ Electrical and Computer Engineering, University of Idaho, USA
§ University of Oslo, Norway
¶ School of Software, Dalian University of Technology, Dalian, China
k Liverpool John Moores University, Liverpool, UK
Abstract—Traffic flow prediction is an important research
issue for solving the traffic congestion problem in an Intelligent
Transportation System (ITS). Traffic congestion is one of the
most serious problems in a city, which can be predicted in
advance by analyzing traffic flow patterns. Such prediction is
possible by analyzing the real-time transportation data from
correlative roads and vehicles. This article first gives a brief
introduction to the transportation data, and surveys the state-ofthe-art prediction methods. Then, we verify whether or not the
prediction performance is able to be improved by fitting actual
data to optimize the parameters of the prediction model which is
used to predict the traffic flow. Such verification is conducted by
comparing the optimized time series prediction model with the
normal time series prediction model. This means that in the era
of big data, accurate use of the data becomes the focus of studying
the traffic flow prediction to solve the congestion problem. Finally,
experimental results of a case study are provided to verify the
existence of such performance improvement, while the research
challenges of this data-analytics-based prediction are presented
and discussed.
I. I NTRODUCTION
Traffic flow prediction is an important research issue in
an Intelligent Transportation System (ITS), and it can be
used as an important metric to solve the traffic congestion
problem. Traffic congestion is considered a serious problem
in big cities around the world. In a study of 471 U.S. urban
areas in 2014 [1], the extra energy cost due to the traffic
congestion was estimated at $160 billion (3.1 billion gallons
of fuel). In addition, long periods of traffic congestion force
the release of more carbon dioxide (CO2 ) greenhouse gases
into the atmosphere and increase the number of accidents.
This can in turn produce severe public health risks which will
dramatically increase medical treatments’ costs. Predicting
traffic flow patterns can help reduce traffic congestion and
therefore reduce the amount of CO2 emissions as well as save
lives. A scenario of using traffic flow prediction to avoid traffic
congestion is illustrated in Figure 1.
Figure 1 illustrates a scenario: how to avoid congestion
in an Intelligent Transportation System (ITS), and what can
be achieved by avoiding congestion? In this scenario, the
transportation and driving data are collected from various
devices of the ITS. This data is then analyzed to help the
navigation of vehicles. Meanwhile, other vehicles that are
preparing to hit the road want to know the current traffic
conditions to make a decision of: (i) whether to start the
journey and when? and (ii) which routes can be used and
which is the best? On this basis, the congestion can be reduced
to achieve eco-friendly routing (low-carbon transportation).
The traffic flow is predictable by analyzing the relevance of
traffic conditions between different traffic roads, as shown in
Figure 2.
Figure 2 illustrates the possibility of knowing whether or
not there is a potential congestion point in advance. Such
knowledge is acquired by predicting the traffic flow for the
specific route. This is due to the continuity of traffic flow
among different traffic roads. It means that the flow of a traffic
road comes from other roads that connect to it.
It is possible to analyze the relevance of the traffic conditions between different traffic roads by tracing the change of
the traffic flow, and such tracing can be achieved by analyzing
the real-time transportation data submitted by the wireless
devices embedded into the vehicles and road side units (RSUs)
installed in the correlative roads [2].
The trajectory data of vehicles (a kind of transportation
data) is the most commonly used by the traffic flow prediction.
Trajectories provide important information on the mobility of
vehicles, where the moving pattern of vehicles characterizes
the traffic flow of a transportation system. The trajectory data
of vehicles is becoming easily available in current transportation systems due to the prevalence of the global positioning
system (GPS) and other localization technologies. A trajectory
generated by a moving vehicle is usually described by a
temporal sequence of spatial locations with their timestamps.
These trajectories convey underlying information about how
the traffic flow of the transportation system changes.
From the traffic congestion problem to traffic flow prediction, and then further to this kind of transportation data:
trajectory, the research issue becomes how to analyze and mine
the trajectory data to solve the congestion problem. Analyzing
and mining trajectory data can extract and reveal inherent
information or knowledge about potential congestion. It will
benefit broad applications, for example, (i) it is possible to
reduce exhaust emission by congestion knowledge based path
planning to enhance the level of public health; (ii) improving
the public security in transportation systems is possible by
2
To Achieve Eco-friendly Routing
(eRouting): Reducing Congestion
eRouting is based on analyzing
transportation data
Navigation
System
Transportation and
driving data
Smart
Parking
[Big Data Analytics]
based
This system helps
to implement the
eco-friendly routing
Navigation by analyzing
real-time data
Depart? but
when and
which routes?
Fig. 1: Traffic congestion avoidance. This scenario shows how to avoid congestion by analyzing transportation data, where the
data comes from? where does the data go? how the data is used? and what can be achieved by analyzing the data?
Transportation Planning
City Planning
Transportation Controlling
City Management
Applications
Prediction-Supportive
Knowledge
Time-Series Model
Deep Learning based Predictor
Markov Chain Model
Combination of NNs (Neural
Networks) and ARIMA
Prediction:
Analysis and Mining:
Deep Learning, Classification, Ranking, Regression
Feature:
Trajectories:
Spatio-Temporal
Data
Availability:
Time Sequence
Location
Time Stamp
Location Semantics
Geographical
Information
Transportation Data:
GPS, GSM, WiFi, Bluetooth, RFID
Fig. 2: In this scenario, the traffic flow in an intersection is
predictable. The traffic flow of the intersection comes from
the traffic roads connecting at the intersection.
avoiding the traffic accidents caused by congestion. Figure 3
illustrates the general framework of such data analytics based
traffic flow prediction.
Fig. 3: General framework of data analytics based traffic flow
prediction.
The general framework consists of three modules: (i) data
collection module: transportation data is able to be collected
from vehicle-mounted GPS, GSM, WiFi, Bluetooth and RFID,
for example, sharing bikes can be embedded GPS, GSM,
WiFi, Bluetooth, and even RFID to track and record: when
and where the bikes are shared, and the traveling trajectories
of these bikes. The trajectory is a kind of spatio-temporal
data, and this kind of data is with the time sequence feature.
In such data, this information is available: location, location
semantics, geographical information and time stamp; (ii) data
3
analytics based prediction module: the collected transportation
data is analyzed by deep learning, classification, ranking and
regression to acquire useful knowledge as the basis of traffic
flow prediction. These four technologies are able to be used to
conduct the data analytics based prediction: time-series model,
deep learning based predictor, markov chain model, and
combination of Neural Networks (NNs) and Autoregressive
Integrated Moving Average (ARIMA); and (iii) application
module: the predicted results are able to be used to support
many special applications to improve the life experience
in cities, for example, transportation planning, transportation
controlling, city planning, and city management.
II. T RANSPORTATION DATA
What is transportation data and what kind of transportation
data is used in this article?
What is transportation data? Transportation data is a kind
of data that describes the information related to transportation
systems, and it contains various vehicle and road information,
e.g., the trajectory and speed information of vehicles, and the
length information of traffic roads. It is helpful to improve
the performance of transportation related applications by analyzing such data, for example, (i) traffic flow prediction; it
is the attempt to estimate the number of vehicles that will
travel on the traffic roads of a transportation system in the
future. This prediction enables us to understand and develop
an optimal transportation system with efficient movement of
traffic. It will also allow us to minimize the traffic congestion
by designing optimal overall route planning for vehicles based
on the predicted traffic flow. The flow of a traffic road
has a connection with the flows of prior roads which are
connected to this traffic road. It means that it is possible
to predict the flow of the current traffic road by calculating
the potential fractional flows from the prior roads and the
previous time period. The potential fractional flow is able to be
calculated by analyzing the transportation data (the trajectory
and speed data of vehicles); (ii) transportation planning; it is
the process of defining the future design for transportation
systems to prepare for the future needs on moving people
and goods to destinations. In this application, it is possible to
acquire the underlying knowledge of transportation systems
by mining the transportation data from the systems. With the
acquired knowledge, the transportation planning will be more
reasonable.
What kind of transportation data is used in this article?
In this study, trajectory data of vehicles is used, which is a
kind of series transportation data. This data can be used to
calculate and predict the traffic flow of a transportation system.
It is collected from the transportation system in England.
This transportation system has 2501 traffic roads1 covering
300 miles of England highways and arterial roads, which is
illustrated in Figure 4.
The important part of trajectory data is the location information. With the location information, trajectory data can provide
actual semantics to make such data be used by traffic flow
prediction. Flow information of a traffic road is acquired by
1A
two-way highway/arterial road is counted as two traffic roads.
Fig. 4: Flow information of traffic roads is acquired by
analyzing the trajectory data of the vehicles driving on these
traffic roads. The update period of this information is 15
minutes. Highways and arterial roads are marked by different
color depths, and the deeper color denotes the heavier traffic
load on the corresponding road.
counting the number of vehicles being driven on the current
road. So in trajectory data, the location information is required
to filter out the vehicle driving records on the current road
rather than other roads. Moreover, for traffic flow prediction,
the traffic flow of the current road converges from the traffic
flow of prior roads. So the location information becomes important to track the traffic flow changes. Trajectory data relies
on localization technology to provide location information.
The commonly used measurement methods of localization
to generate location-containing trajectory data include GPS,
WiFi, GSM, Bluetooth, and RFID.
III. S TATE - OF - THE - ART P REDICTION M ETHODS
Many kinds of techniques have been proposed to address the
traffic flow prediction problem. The details are summarized in
the following literature list.
Many specific methods have been proposed to make prediction in different situations, and it emerges that traditionally
there is no single best method for every situation. It is better to
combine several suitable techniques to improve the accuracy
of prediction under considering different situations. It means
that the traditional traffic flow prediction methods are not able
to satisfy most real-world application requirements. During the
last five years, some studies have tried to use data analytics
to solve the traffic flow prediction problem, and the results
demonstrate that such schemes are feasible and are able to
4
Techniques
Typical Works and Advantages
Bayesian Networks
[3] this work presented a spatio-temporal Bayesian network
predictor. This work employed the Gaussian Mixture Model
and the Competitive Expectation Maximization algorithm in
order to approximate the joint probability distribution of
the nodes in the Bayesian network.
[4] this work proposed a model: (i) this model has the
ability to create a Bayesian network to the traffic flow on
a given road at a given time; (ii) this model has an
ability to cope with the data-incomplete problem.
[5] the technique of the traffic prediction, which was
presented in this work, used the capability of traffic
flows to reproduce the real behavior of vehicles and to
construct the Gaussian Bayesian network employing the
special characteristics of the variables of the traffic
flow.
[6] this work utilized a linear conditional Gaussian
Bayesian Network (BN) model to consider the impacts of both
spatial and temporal dimensions of traffic as well as speed
information for short-term traffic flow prediction.
Neural Networks
[7] this work proposed a method to make short-term
prediction based on Neural Networks (NNs).
[8] this work provided a traffic flow predictor based on
the architecture of fuzzy neural networks.
Time-Series Models
[9] the statistical time-series model, ARIMA is fitted to
time series data either to better understand the data or to
predict future points in the series. The ARIMA model is not
suited for the prediction which is based on this kind of
data: (i) there is the information missing problem in the
data, and (ii) the filling of data is problematic as the
situation is complex.
[10] this work presented the theoretical basis for
modelling univariate traffic conditions as Seasonal
Autoregressive Integrated Moving Average (SARIMA)
processes. Fitted SARIMA models provide equations which can
be used to produce single and multiple interval prediction.
[11] the focus in this work was on producing time-series
state space models which are flexible and explicitly
multivariate. The models enable to jointly consider the
data from different detectors, and are able to model a wide
variety of univariate models, for instance, ARIMA. The
results were compared to those achieved from the ARIMA
model and were found to be superior.
Kalman Filtering Theory
[12] this work investigated the performance of the adaptive
Kalman Filter for short-term traffic flow prediction.
Moreover, this work presented the empirical results from
the application of the adaptive Kalman Filter to real-world
data. Kalman Filtering is applicable to short-term
stationary or non-stationary stochastic phenomena and it
yields good traffic prediction accuracy.
Markov Chain Model
[13] this work modelled the traffic flow as a high-ordered
Markov Chain. The method proposed in this work employs
current and recent values of the traffic flow, and
describes the future value. This future value is assumed
that the predicted state has a probability distribution,
and both current state and most recent states determine the
next state.
Combination of NNs and ARIMA
[14] this work proposed a hybrid model combining NNs and
ARIMA, which is capable of exploiting the strengths of
traditional time series approaches and artificial neural
networks.
Deep Learning based Predictor
[15] this work proposed a deep learning based traffic flow
prediction method, which considers the spatial and temporal
inherent correlations of traffic records.
improve the accuracy of prediction, as the method proposed
in [15]. Such kind of data analytics based prediction method
is able to satisfy the requirements of different applications by
analyzing the data from each corresponding specific applica-
tion.
IV. P ERFORMANCE C OMPARISON
This section compares the performance of two kinds of time
series prediction models: (i) optimized time series prediction
5
model, and (ii) normal time series prediction model. To verify
this fact: the prediction performance of a prediction model
can be improved by fitting data to optimize model parameters.
ARIMA model is used in this study to make this verification.
Predicting a stationary time series by ARIMA depends on the
parameters (p, d, q) of the ARIMA:
• The parameter p is the number of Auto-Regressive (AR)
terms, for example, if p is 5, the predictors for x(t) will
be x(t − 1), ..., x(t − 5);
• The parameter q is the number of Moving Average
(MA) terms. MA terms are lagged prediction errors
in the prediction equation, for example, if q is 5, the
predictors for x(t) will have such lagged prediction errors
e(t − 1), ..., e(t − 5), where e(i) is the difference between
the moving average and the actual value at the ith instant.
• The parameter d is the number of differences when the
time series becomes stable.
An important concern here is how to determine the values
of ‘p’ and ‘q’. Determining the values of ‘p’ and ‘q’ will
affect the performance of the prediction model. This study
uses Bayesian Information Criterion (BIC) to determine the
optimum values of parameters ‘p’ and ‘q’ to avoid over-fitting,
when fitting a model by training data. BIC is a criterion for
model selection among a finite set of models, and the model
with the lowest BIC value is preferred. It is based, in part, on
the likelihood function.
Comparative results are illustrated in Figure 5. Each line is
the average of 2501 traffic roads.
The RMSE of a model with respect to the “Traffic Flow” is
defined as: the square root of the mean square error between
the values actually observed and the values predicted by a
model. The RMSE estimated from the optimized ARIMA is
less than the estimated result from the normal ARIMA. The
prediction accuracy of the time series model ARIMA is able
to be obviously improved by optimizing relevant parameters.
Figure 5a and Figure 5b illustrate the comparative results
with the data of two selected dates, March 15th and 31st, 2015.
For the RMSE, the optimized ARIMA is much better than the
normal ARIMA. Figure 5c provides the average results for
31 days of March, 2015. On these basis, it is observed that
improving the performance of a prediction model is possible
by optimizing model parameters with analyzing sequential
transportation data.
V. R ESEARCH C HALLENGES
The pattern prediction of traffic flow, which is supported
by data analytics, strongly relies on the transportation data
which records the vehicle mobility of transportation systems.
Moreover, transportation data implies the underlying patterns
and laws of the vehicle mobility. It means that the data is
embedded with the rich information of the traffic dynamics
of a city. Meanwhile, transportation information may cause
privacy issues. This article classifies the research challenges
of data analytics supportive traffic flow prediction, as shown
in Figure 6. There are two important aspects for the research
challenges of this kind of prediction: data and data based
prediction.
Data:
i) Privacy Protection: it is necessary and important to protect the personal privacy information which is contained
in transportation data. Such information includes the
locations, driving trajectories, properties and vehicles’
plate numbers. It is a worthy research challenge to
protect this information when the transportation data is
analyzed to obtain useful information and knowledge.
ii) Mobility: in transportation systems, understanding and
predicting the patterns of mobility is the basis of studying and solving a variety of practical problems, e.g.,
traffic congestion problems. The mobility produces large
amounts of spatio-temporal data, and this kind of data
is accompanied with time and location information. It
is a worthy research challenge to analyze the spatiotemporal data for learning and predicting the patterns of
mobility. Moreover, traffic flow is an important reflection
for the mobility in transportation systems. Therefore,
traffic flow prediction is also an important and worthy
research challenge. Analyzing the spatio-temporal data
is another research challenge as the basis of understanding, learning and predicting the patterns of mobility and
traffic flow. Spatio-temporal data is a kind of time series
data. To analyze this kind of data, a time series analysis
model is necessary, and for different applications, the
objective functions are different because of the different
requirements, and how to combine a time series analysis
model with an objective function for special requirements is an important challenge that requires further
investigation.
iii) Vehicle: the vehicles in a transportation system produce
large amounts of trajectory data. This kind of data can be
used to mine vehicle behaviors and relations. If we know
the behaviors of vehicles and the relations between these
vehicles, it is possible to predict the change of traffic
flow and then to avoid congestion by traffic control and
management. For example, if there is congestion at a
certain intersection and the relevant vehicles have the
same destination, these vehicles have great possibility
to use the same route, and the congestion will happen
again at the next intersection, if there are no control
measures in advance.
Big Data: In these applications, “big data” can be achieved
whether from “mobility” and/or from “vehicles”: (i) the mobility produces large amounts of spatio-temporal data, and (ii) the
vehicles produce large amounts of trajectory data. Such “large
amounts of” is a kind of “big”, and for analyzing this kind of
data, big data analytics is needed. Such analytics is different
from traditional data analysis. For example, deep learning and
NNs can be used to analyze such big data, but how to learn
such data to extract useful information and knowledge, and
learning for what, are worthy challenges which need to be
studied.
Wireless Big Data: Most of the big data of an intelligent transportation system (the trajectory data of vehicles)
is collected using wireless devices. These devices can be
wireless sensor nodes which are embedded on the traffic
6
(a) Actual traffic flow data of March 15th, 2015, and the
predicted results by normal and optimized ARIMA models.
(b) Actual traffic flow data of March 31st, 2015, and the
predicted results by normal and optimized ARIMA models.
(c) Average results for 31 days of March: actual traffic flow and
the predicted results by normal and optimized ARIMA models.
Fig. 5: Comparing with the normal ARIMA, optimized ARIMA achieves better prediction results in the aspect of the changing
trend and values of traffic flow. Root-Mean-Square Errors (RMSE) are marked in each subfigure to indicate the prediction
performance of the normal and optimized ARIMA models.
Mobility
Mobility of Transportation Systems
Privacy
Transportation Data
Vehicle
Protection
Vehicle Behaviors and Relations
Data based
Prediction
City
City Planning and Management
Fig. 6: Research challenges of data supportive prediction for
traffic flow.
road surfaces to detect passing vehicles. There will have data
missing problems in this kind of wireless data because of
some unpredictable problems, e.g., sensor node failure. In
some scenarios, if the critical data is missing or has errors,
it will make the analysis results gravely deviate from making
correct conclusions. For example, if the data from a device is
not correct and the data is used to perform analysis, the result
will be incorrect and cannot be used as useful information or
knowledge to aid decision making. Ensuring the integrity of
the wireless big data information is still a great challenge that
needs to be addressed by experts in the field.
City Data Based Prediction: The prediction results in
analyzing transportation data can help in city planning and
management. The large trajectory data produced by moving
vehicles can help predict the traffic flow that in turn can
greatly improve lives. In addition, this can motivate the design
of intelligent traffic lights that can achieve automatic traffic
flow global scheduling. Such scheduling is based on the
understanding and prediction of global traffic conditions. So,
it is important to know whether the current traffic flow will
impact roads during the next time period or not. These issues
are still a challenge that need researchers’ attention in the near
future.
VI. C ONCLUSION
Transportation data is the sampling of dynamics of moving
objects in the temporal and spatial dimensions. Analysis and
mining of such time series data is becoming a promising way
to discover the underlying knowledge on vehicular activities,
vehicles’ relations and even city dynamics. It helps to well
understand and predict the pattern of traffic flow which is an
important aspect of traffic dynamics. Traffic flow prediction
can be exploited in a wide range of potential applications to
make a city smarter, safer, more livable, and can help reduce
congestion and pollution. However, it remains challenging
7
to analyze time series data to acquire useful knowledge for
problem solving, because of data heterogeneity and data incompleteness in a complex and dynamic transportation system.
So, in this article we have introduced the transportation data.
Then, we discussed state-of-the-art prediction methods. We
run some experiments in order to compare the optimized time
series prediction model with the normal time series prediction
model. Our results show that the prediction performance is
able to be improved by fitting actual data to optimize the
parameters of the prediction model. In addition, we have discussed typical techniques to address the traffic flow prediction
problem. Finally, we listed some challenges that will provide
leads to future researchers in this area.
R EFERENCES
[1] D. Schrank, B. Eisele, T. Lomax, and J. Bak, “2015 urban mobility
scorecard,” The Texas A&M Transportation Institute and INRIX, Tech.
Rep., 2015.
[2] A. Al-Fuqaha, M. Guizani, M. Mohammadi, M. Aledhari, and
M. Ayyash, “Internet of things: A survey on enabling technologies,
protocols, and applications,” IEEE Communications Surveys & Tutorials,
vol. 17, no. 4, pp. 2347–2376, 2015.
[3] S. Sun, C. Zhang, and Y. Zhang, “Traffic flow forecasting using a spatiotemporal bayesian network predictor,” in International Conference on
Artificial Neural Networks. Springer, 2005, pp. 273–278.
[4] S. Sun, C. Zhang, and G. Yu, “A bayesian network approach to
traffic flow forecasting,” IEEE Transactions on intelligent transportation
systems, vol. 7, no. 1, pp. 124–132, 2006.
[5] E. Castillo, J. M. Menéndez, and S. Sánchez-Cambronero, “Predicting
traffic flow using bayesian networks,” Transportation Research Part B:
Methodological, vol. 42, no. 5, pp. 482–509, 2008.
[6] Z. Zhu, B. Peng, C. Xiong, and L. Zhang, “Short-term traffic flow
prediction with linear conditional gaussian bayesian network,” Journal
of Advanced Transportation, vol. 50, no. 6, pp. 1111–1123, 2016.
[7] K. Y. Chan, T. S. Dillon, J. Singh, and E. Chang, “Neural-network-based
models for short-term traffic flow forecasting using a hybrid exponential
smoothing and levenberg–marquardt algorithm,” IEEE Transactions on
Intelligent Transportation Systems, vol. 13, no. 2, pp. 644–654, 2012.
[8] K. Y. Chan and T. S. Dillon, “On-road sensor configuration design for
traffic flow prediction using fuzzy neural networks and taguchi method,”
IEEE Transactions on Instrumentation and Measurement, vol. 62, no. 1,
pp. 50–59, 2013.
[9] R. N. Calheiros, E. Masoumi, R. Ranjan, and R. Buyya, “Workload
prediction using arima model and its impact on cloud applications’ qos,”
IEEE Transactions on Cloud Computing, vol. 3, no. 4, pp. 449–458,
2015.
[10] B. M. Williams and L. A. Hoel, “Modeling and forecasting vehicular
traffic flow as a seasonal arima process: Theoretical basis and empirical
results,” Journal of transportation engineering, vol. 129, no. 6, pp. 664–
672, 2003.
[11] A. Stathopoulos and M. G. Karlaftis, “A multivariate state space approach for urban traffic flow modeling and prediction,” Transportation
Research Part C: Emerging Technologies, vol. 11, no. 2, pp. 121–135,
2003.
[12] J. Guo, W. Huang, and B. M. Williams, “Adaptive kalman filter approach
for stochastic short-term traffic flow rate prediction and uncertainty
quantification,” Transportation Research Part C: Emerging Technologies, vol. 43, pp. 50–64, 2014.
[13] G. Yu, J. Hu, C. Zhang, L. Zhuang, and J. Song, “Short-term traffic
flow forecasting based on markov chain model,” in Intelligent Vehicles
Symposium, 2003. Proceedings. IEEE. IEEE, 2003, pp. 208–212.
[14] A. K. Fard and M.-R. Akbari-Zadeh, “A hybrid method based on
wavelet, ann and arima model for short-term load forecasting,” Journal
of Experimental & Theoretical Artificial Intelligence, vol. 26, no. 2, pp.
167–182, 2014.
[15] Y. Lv, Y. Duan, W. Kang, Z. Li, and F.-Y. Wang, “Traffic flow
prediction with big data: a deep learning approach,” IEEE Transactions
on Intelligent Transportation Systems, vol. 16, no. 2, pp. 865–873, 2015.
| 2cs.AI
|
ELT LINEAR ALGEBRA II
arXiv:1705.00266v1 [math.RA] 30 Apr 2017
GUY BLACHAR AND EREZ SHEINER
Abstract. This paper is a continuation of [4]. Exploded layered tropical (ELT) algebra is an
extension of tropical algebra with a structure of layers. These layers allow us to use classical
algebraic results in order to easily prove analogous tropical results. Specifically we prove and use
an ELT version of the transfer principal presented in [2].
In this paper we use the transfer principal to prove an ELT version of Cayley-Hamilton Theorem,
and study the multiplicity of the ELT determinant, ELT adjoint matrices and quasi-invertible
matrices.
We also define a new notion of trace – the essential trace – and study its properties.
Contents
0. Introduction
0.1. ELT Algebras
0.2. The Element −∞
0.3. Non-Archimedean Valuations and Puiseux Series
0.4. ELT Algebras and Puiseux Series
0.5. Semirings with a Negation Map and ELT Rings
1. ELT Transfer Principle
1.1. The Transfer Principle
1.2. Formulation and Proof of the ELT Transfer Principle
1.3. Multiplicity of the ELT Determinant
1.4. The ELT Adjoint Matrix and Quasi-Invertible Matrices
1.5. The ELT Characteristic Polynomial and ELT Eigenvalues
2. ELT Traces
2.1. Trace of ELT Matrices
2.2. ELT Nilpotent Matrices
2.3. The Essential Trace
References
Page
1
2
3
3
4
5
5
5
7
8
9
11
13
13
13
13
17
0. Introduction
Tropical linear algebra, also known as Max-Plus linear algebra, has been studied for more than
50 years (ref. [5]). While tropical geometry mainly deals with geometric combinatorial problems,
tropical linear algebra deals with algebraic non-linear combinatorial problems (for instance, the assignment problem [18]). Tropical linear algebra may also be used as a mean to study the tropical
algebraic geometry (for instance, the tropical resultant). Notable work in this field can be found at
Date: 19th February 2018.
2010 Mathematics Subject Classification. Primary: 15A03, 15A09, 15A15, 15A63; Secondary: 16Y60, 14T05.
Key words and phrases. Tropical algebra, ELT algebra, matrix theory, characteristic polynomial, trace, transfer
principle.
This article contains work from Erez Sheiner’s Ph.D. Thesis, which was accepted on 1.1.16, and from Guy Blachar’s
M.Sc. Thesis, both submitted to the Math Department at Bar-Ilan University. Both works were carried under the
supervision of Prof. Louis Rowen from Bar-Ilan University, to whom we thank deeply for his help and guidance.
1
2
GUY BLACHAR AND EREZ SHEINER
[5], [6], [13], [15] and [21].
In our previous paper ([4]) we introduced a new structure, which we call exploded layered tropical algebra (or ELT algebra for short). This structure is a generalization of the work of Izhakian
and Rowen ([14]), and is similar to Parker’s exploded structure ([19]). The layers enable us to use
“classical language” even when dealing with tropical questions.
Our work in this paper can be divided into two main parts. The first one uses the theory of
semirings with a negation map to study the ELT structure. We formulate and prove an ELT version
of the transfer principles written in [2], and use them to study ELT matrix theory, such as the ELT
adjoint matrix (Theorem 1.9 and Theorem 1.10).
The second part of our work deals with a new notion of trace. Whereas the trace can be defined
as in the classical theory, it lacks some important properties in the ELT theory. For example, the
trace of an ELT nilpotent matrix need not be of layer zero. We define the essential trace of an ELT
matrix (subsection 2.3) to deal with such cases.
0.1. ELT Algebras.
Definition 0.1. Let L be a semiring, and F a totally ordered semigroup. An ELT algebra is
the pair R = R (L , F ), whose elements are denoted [ℓ]a for a ∈ F and ℓ ∈ L , together with the
semiring (without zero) structure:
[ℓ ]
a1 > a2
1 a1
[ℓ
]
[ℓ
]
[ℓ
]
1
2
2
(1)
a1 + a2 :=
a2
a1 < a2 .
[ℓ1 +L ℓ2 ]
a1 a1 = a2
(2) [ℓ1 ]a1 · [ℓ2 ]a2 := [ℓ1 ·L ℓ2 ] (a1 +F a2 ).
We write . For
[ℓ]a,
ℓ is called the layer, whereas a is called the tangible value.
ELT algebras originate from [19], and are also discussed in [22].
Let R be an ELT algebra. We write s : R → L for the projection on the first component (the
sorting map):
s [ℓ]a = ℓ
We also write τ : R → F for the projection on the second component:
τ [ℓ]a = a
We denote the zero-layer subset
[0]R
= {α ∈ R|s (α) = 0}
and
R× = {α ∈ R|s (α) 6= 0} = R \ [0]R
We note some special cases of ELT algebras.
Example 0.2. Let (G, ·) be a totally ordered group. We denote by Gmax the max-plus algebra
defined over G, i.e. the set G endowed with the operation
a ⊕ b = max {a, b} , a ⊙ b = a · b.
Then Gmax is equivalent to the trivial ELT algebra with F = G and L = {1}.
Example 0.3. Zur Izhakian’s supertropical algebra ([11]) is equivalent to an ELT algebra with a
layering set L = {1, ∞}, where
1 + 1 = ∞, 1 + ∞ = ∞ + 1 = ∞, ∞ + ∞ = ∞
ELT LINEAR ALGEBRA II
3
and
1 · 1 = 1, 1 · ∞ = ∞ · 1 = ∞, ∞ · ∞ = ∞.
The supertropical ”ghost” elements aν correspond to [∞]a in the ELT notation, whereas the
tangible elements a correspond to [1]a.
We define a partial order relation on R in the following way:
x y ⇐⇒ ∃z ∈ [0]R : x = y + z
Lemma 0.4 ([4, Lemma 0.4]). is a partial order relation on R.
Let us point out some important elements in any ELT algebra R:
(1) [1]0, which is the multiplicative identity of R.
(2) [0]0, which is idempotent for both operations of R.
(3) [−1]0, which has the role of “−1” in our theory.
Note that
[0]0
· [ℓ]a = [0]a. Therefore,
[0]R
= [0]0 R. In particular,
[0]R
is an ideal of R.
Throughout this paper, unless otherwise noted, we work under more general assumptions than
in [4]. Out underlying ELT algebras R = R (L , F ) will be commutative ELT rings, meaning
that F is an abelian group, and L is a commutative ring.
0.2. The Element −∞. As in the tropical algebra, ELT algebras lack an additive identity. Therefore, we adjoin a formal element to the ELT algebra R, denoted by −∞, which satisfies ∀α ∈ R:
−∞ + α = α + −∞ = α
−∞ · α = α · −∞ = −∞
We also define s (−∞) = 0. We denote R = R ∪ {−∞}.
We note that R is now a semiring, with the following property:
α + β = −∞ =⇒ α = β = −∞
Such a semiring is called an antiring. Antirings are dealt with in [23] and [7].
0.3. Non-Archimedean Valuations and Puiseux Series. We recall the definition of a (nonArchimedean) valuation (see [8] and [24]).
Definition 0.5. Let K be a field, and let (Γ, +, ≥) be an abelian totally ordered group. Extend Γ
to Γ ∪ {∞} with γ < ∞ and γ + ∞ = ∞ + γ = ∞ for all γ ∈ Γ. A function v : K → Γ ∪ {∞} is
called a valuation, if the following properties hold:
(1) v (x) = ∞ ⇐⇒ x = 0.
(2) ∀x, y ∈ K : v (xy) = v (x) + v (y).
(3) ∀x, y ∈ K : v (x + y) ≥ min {v (x) , v (y)}.
Given a valuation v over a field K, we recall some basic properties:
(1) v (1) = 0.
(2) ∀x ∈ K : v (−x) = v (x).
(3) ∀x ∈ K × : v x−1 = −v (x).
(4) If v (x + y) > min {v (x) , v (y)}, then v (x) = v (y). (For this reason, the equality between
the valuation of two elements is central in out theory.)
One may associate with v the valuation ring
Ov = {x ∈ K|v (x) ≥ 0}
This is a local ring with the unique maximal ideal
mv = {x ∈ K|v (x) > 0}
The quotient kv = Ov mv is called the residue field of the valuation.
4
GUY BLACHAR AND EREZ SHEINER
Let us present another key construction related to valuations. For γ ∈ Γ, let D≥γ = {x ∈ K|v (x) ≥ γ}
and D>γ = {x ∈ K|v (x) > γ}. It is easily seen that D≥γ is an abelian additive group, and
. that
D
D>γ is a subgroup of D≥γ . Note that for γ = 0, D≥0 = Ov and D>0 = mv . Set Dγ = ≥γ D .
>γ
The associated graded ring of K with respect to v is
M
grv (K) =
Dγ
γ∈Γ
Given γ1 , γ2 ∈ Γ, the multiplication in K induces a well-defined multiplication Dγ1 × Dγ2 → Dγ1 +γ2
given by
(x1 + D>γ1 ) · (x2 + D>γ2 ) = x1 x2 + D>(γ1 +γ2 )
This multiplication can be extended to a multiplication map in grv (K), endowing it with a structure
of a graded ring.
We will now focus on Puiseux series, which is the central example for our theory. The field of
Puiseux series with coefficients in a field K and exponents in an abelian ordered group Γ is
(
)
X
i
K{{t}} =
αi t αi ∈ K, I ⊆ Γ is well-ordered
i∈I
The resulting set, equipped with the natural operations, is a field; in addition, if K is algebraically
closed and Γ is divisible, then K{{t}} is also algebraically closed.
Assuming Γ is also totally ordered, one may define a valuation on the field of Puiseux series
v : K{{t}} → Γ ∪ {∞} as follows: v (0) = ∞, and
!
X
i
v
αi t = min {i ∈ I|αi 6= 0}
i∈I
Let us examine the associated graded ring with respect to this valuation. For each γ ∈ Γ, we first
claim that Dγ ∼
= K. Indeed, the kernel of the homomorphism f : D≥γ → K defined by
X
f
αi ti = αγ
γ≤i∈I
is precisely D>γ (since D>γ is the subgroup of D≥γ of Puiseux series whose minimal degree is bigger
than γ).
0.4. ELT Algebras and Puiseux Series. Let R = R (L , F ) be an ELT algebra. In [4] we introduced the EL tropicalization function ELTrop : L {{t}} → R, which is defined in the following
way: if x ∈ L {{t}}\ {0} has a leading monomial ℓta , then
ELTrop (x) = [ℓ] (−a) .
In addition, ELTrop (0) = −∞.
Lemma 0.6 ([4, Lemma 0.9]). The following properties hold:
(1) ∀x, y ∈ L {{t}} : ELTrop (x) + ELTrop (y) ELTrop (x + y).
(2) ∀α ∈ L ∀x ∈ L {{t}} : ELTrop (αx) = [α]0 ELTrop (x).
(3) ∀x, y ∈ L {{t}} : ELTrop (x) ELTrop (y) ELTrop (xy).
We remark that in the case in which R is an ELT integral domain, meaning L is an integral
domain, we have ELTrop (x) ELTrop (y) = ELTrop (xy) for all x, y ∈ L {{t}}.
Let us examine the relation x ELTrop (y) a bit more deeply. If x = ELTrop (y), it means that
x can be lifted to a Puiseux series which has x as its leading monomial. Otherwise, we have that x
is of layer zero, and its tangible value is bigger than the tangible value of ELTrop (y); so one may
say that x can also be lifted to a Puisuex series with leading coefficient x, where we allow it to have
a zero coefficient in its leading monomial.
ELT LINEAR ALGEBRA II
5
0.5. Semirings with a Negation Map and ELT Rings. Semirings need not have additive
inverses to all of the elements. While some of the theory of rings can be copied “as-is” to semirings,
there are many facts about rings which use the additive inverses of the elements. The idea of negation
maps on semirings (sometimes called symmetrized semirings) is to imitate the additive inverse map.
Semirings with negation maps are discussed in [1], [9], [10], [2], [3], [20].
Definition 0.7. Let R be a semiring. A map (−) : R → R is a negation map (or a symmetry)
if the following properties hold:
(1) ∀a, b ∈ R : (−) (a + b) = (−)a + (−)b.
(2) (−)0R = 0R .
(3) ∀a, b ∈ R : (−) (a · b) = a · ((−)b) = ((−)a) · b.
(4) ∀a ∈ R : (−) ((−)a) = a.
We say that (R, (−)) is a semiring with a negation map (or a symmetrized semiring). If (−)
is clear from the context, we will not mention it.
We give several examples of semirings with negation maps:
• A trivial example of a negation map (over any semiring) is (−)a = a.
• If R is a ring, it has a negation map (−)a = −a.
• If R is an ELT algebra, we have a negation map given by (−)a = [−1]0a.
The last example is the central example for our theory, since it shows that any ELT algebra is
equipped with a natural negation map. Thus, the theory of semirings with negation maps can be
used when dealing with ELT algebras.
We now present several notations from this theory:
• a + (−)a is denoted a◦ .
• R◦ = {a◦ |a ∈ R}.
• We define two partial orders on R:
– The relation ◦ defined by
a ◦ b ⇔ ∃c ∈ R◦ : a = b + c
– The relation ∇ defined by
a∇b ⇔ a + (−)b ∈ R◦
If R is an ELT algebra, then some of these notations have already been defined. For example,
a◦ = [0]0a, R◦ = [0]R and the relation ◦ is the relation .
1. ELT Transfer Principle
The transfer principles are two theorems, presented in [2], which allows to conveniently transfer
equalities between polynomial expressions in the classical theory to theorems about semirings with
a negation map. We recall that any ELT algebra R has a negation map
(−) [ℓ]a = [−ℓ]a
and thus we may view each ELT algebra as a semiring with a negation map.
In this section, we use the transfer principles to prove two transfer principles for the ELT theory,
and use these transfer principles to study the ELT adjoint matrix.
1.1. The Transfer Principle. In this subsection, we briefly introduce the two classical transfer
principles given in [2].
Definition 1.1. A positive polynomial expression in the variables λ1 , . . . , λm is a formal expression produced by the context-free grammar E 7→ E +E, (E)×(E), where the symbols 0, 1, λ1 , . . . , λm
are thought of as terminal symbols of the grammar. A monomial in a positive polynomial expression is a sum of expressions of the form cI λi11 . . . λimm , where I = (i1 , . . . , in ) is fixed.
6
GUY BLACHAR AND EREZ SHEINER
That means that 0, 1, λ1 , . . . , λm are positive polynomial expressions. Also, λ1 + (λ2 ) × (λ1 + λ3 )
is a positive polynomial expression. Any positive polynomial expression E can be interpreted as a
polynomial in N [λ1 , . . . , λm ]. We say that a monomial λi11 . . . λimm appears in the expression E if
there exists a positive integer c such that cλi11 . . . λimm appears in the expansion of the polynomial
obtained by interpreting E in N [λ1 , . . . , λm ].
im
If R is a semiring with a negation map, and if cλi11 . . . λm
is a monomial, we define
i1
i1
im
(−) cλ1 . . . λm = ((−) c) λ1 . . . λimm
Definition 1.2. Let R be a semiring with a negation map. A polynomial expression is a
formal expression of the form P + − P − , where P + and P − are positive polynomial expressions. A
monomial in P is a sum of the monomials cI λi11 . . . λimm from P + and c′I λi11 . . . λimm from P − , where
I = (i1 , . . . , in ) is fixed. We say that a monomial appears in the polynomial expression P , if it
appears either in P + or in P − .
Definition 1.3. Let P and Q be polynomial expressions. We say that the identity P = Q is valid
in a semiring with a negation map R, if it holds for any semiring with a negation map R and for
any substitution λ1 = r1 , . . . , λm = rm of r1 , . . . , rm ∈ R.
Recall the relations ◦ and ∇ from subsection 0.5.
Definition 1.4. Let P and Q be polynomial expressions.
(1) We say that the identity P ∇Q holds in all commutative semirings with a negation map, if
for any semiring with a negation map R and for any substitution λ1 = r1 , . . . , λm = rm of
r1 , . . . , rm ∈ R,
P (r1 , . . . , rn ) ∇Q (r1 , . . . , rn )
(2) We say that the identity P ◦ Q holds in all commutative semirings with a negation map,
if for any semiring with a negation map R and for any substitution λ1 = r1 , . . . , λm = rm
of r1 , . . . , rm ∈ R,
P (r1 , . . . , rn ) ◦ Q (r1 , . . . , rn )
We recall the transfer principle ([2, Theorems 4.20 and 4.21]):
Theorem 1.5 (Transfer principle, weak form). Let P and Q be polynomial expressions. If the
identity P = Q holds in all commutative rings, then the identity P ∇Q holds in all commutative
semirings with negation map.
Theorem 1.6 (Transfer principle, strong form). Let P and Q be polynomial expressions. If the
identity P = Q holds in all commutative rings, and if Q = Q+ − Q− for some positive polynomial
expressions such that there is no monomial appearing simultaneously in Q+ and Q− , then the identity
P ◦ Q holds in all commutative semirings with negation map.
The transfer principle allows us to prove several important theorems in a rather convenient way.
Corollary 1.7 (Multiplicativity of determinant). Let R be a semiring with a negation map, and let
A ∈ Rn×n . We define:
X
a1,σ(1) . . . an,σ(n)
det (A)+ =
σ∈An
−
det (A)
=
X
a1,σ(1) . . . an,σ(n)
σ∈Sn \An
det (A) =
+
det (A) − det (A)
−
In [2, Corollary 4.8], it is proven that if R is a semiring with a negation map, then
∀A, B ∈ Rn×n : det (AB) ◦ det (A) det (B)
Corollary 1.8 (Cayley-Hamilton theorem). Let R be a semiring with a negation map, and let
A ∈ Rn×n . We know that over a commutative ring, fA (A) = 0. We can use the strong form of the
transfer principle componentwise, and thus
fA (A) ◦ 0
ELT LINEAR ALGEBRA II
7
In other words,
fA (A) ∈ (R◦ )
n×n
1.2. Formulation and Proof of the ELT Transfer Principle. We first recall that all of our
ELT algebras are commutative ELT rings, meaning R = R (L , F ) where F is an abelian group
and L is a commutative ring.
We would like to have a tool of proving polynomial identities over commutative ELT rings. Recall
that any ELT ring is a semiring with a negation map a 7→ [−1]0a (subsection 0.5), and thus we may
apply the transfer principle. In order to strengthen the general transfer principle, we will use results
from tropical linear algebra. This is formulated in the following theorems:
Theorem 1.9 (ELT Transfer Principle for equality). Let P and Q be polynomial expressions. Assume that the identity P = Q holds in all commutative rings. If the identity P = Q holds in all
commutative tropical algebras, then the identity P = Q holds in all commutative ELT rings.
Theorem 1.10 (ELT Transfer Principle for surpassing). Let P and Q be polynomial expressions.
Assume that the identity P = Q holds in all commutative rings. If the identity P ≥ Q holds in all
commutative tropical algebras, then the identity P Q holds in all commutative ELT rings.
We prove two lemmas which will help us prove the theorems:
Lemma 1.11. Let R be an ELT algebra, and let α, β ∈ R. If s α + [−1]0β = 0, and if τ (α) ≥ τ (β),
then α β.
Proof. Denote α = [ℓ]a, β = [k]b. τ (α) ≥ τ (β) means a ≥ b. We have two options:
(1) If a = b, then α + [−1]0β = [ℓ−k]a ∈ [0]R. Thus, ℓ = k, which implies α = β.
(2) If a > b, then α + [−1]0β = α ∈ [0]R. Thus, ℓ = 0, and α = α + β β.
In any case α β, and thus we are finished.
Lemma 1.12. Let R = R (L , F ) be an ELT algebra. Endow F with a max-plus algebra, Fmax .
Then the function τ : R → Fmax is a “homomorphism”, in the sense that:
(1) ∀x, y ∈ R : τ (x + y) = τ (x) ⊕ τ (y).
(2) ∀x, y ∈ R : τ (xy) = τ (x) ⊙ τ (y).
Proof. Take [ℓ1 ]a1 , [ℓ2 ]a2 ∈ R.
(1) If a1 = a2 , then
τ
[ℓ1 ]a
1
+ [ℓ2 ]a2 = τ
[ℓ1 +L ℓ2 ]a
1
= a1 = a1 ⊕ a2 = τ
[ℓ1 ]a
1
⊕τ
Otherwise, without loss of generality, a1 > a2 , and thus
τ [ℓ1 ]a1 + [ℓ2 ]a2 = τ [ℓ1 ]a1 = a1 = a1 ⊕ a2 = τ [ℓ1 ]a1 ⊕ τ
(2) τ [ℓ1 ]a1 · [ℓ2 ]a2 = τ [ℓ1 ·L ℓ2 ] (a1 +F a2 ) = a1 +F a2 = a1 ⊙ a2 = τ
[ℓ2 ]a
[ℓ2 ]a
[ℓ1 ]a
2
1
2
⊙τ
[ℓ2 ]a
2
We now prove these theorems. We note that since Theorem 1.9 follows from Theorem 1.10, we
will only prove the latter.
Proof of Theorem 1.10. By the weak form of the general transfer principle, s P + [−1]0Q = 0.
Let R = R (L , F ) be a commutative ELT ring. We will now prove that for any substitution
λ1 = r1 , . . . , λm = rm of r1 , . . . , rm ∈ R,
τ (P (r1 , . . . , rm )) ≥ τ (Q (r1 , . . . , rm ))
We endow F with the max-plus operations. By Lemma (Lemma 1.12) τ : R → Fmax is a homomorphism. Therefore, for any ELT polynomial p ∈ R [λ] and for any substitution λ1 = r1 , . . . , λm =
rm of r1 , . . . , rm ∈ R,
τ (p (r1 , . . . , rm )) = p (τ (r1 ) , . . . , τ (rm ))
8
GUY BLACHAR AND EREZ SHEINER
Thus, for any substitution λ1 = r1 , . . . , λm = rm of r1 , . . . , rm ∈ R,
τ (P (r1 , . . . , rm )) = P (τ (r1 ) , . . . , τ (rm )) ≥ Q (τ (r1 ) , . . . , τ (rm )) = τ (Q (r1 , . . . , rm ))
We have proven that s P + [−1]0Q = 0 and that τ (P ) ≥ τ (Q); by Lemma 1.11, we are finished.
Remark 1.13. Throughout the uses of the ELT transfer principle for equality, we need to check
the corresponding identity in commutative tropical algebras. However, since major work has been
done in the supertropical theory (see [15], [16] and [17]), we usually check that one of the following
conditions holds:
(1) The identity P = Q holds in all commutative supertropical algebras.
(2) The identity ν (P ) = ν (Q) holds in all commutative supertropical algebras.
Similarly, to prove a surpassing relation, we usually check that one of the following conditions holds:
(1) The identity P Q holds in all commutative supertropical algebras.
gs
(2) The identity P ≥ν Q holds in all commutative supertropical algebras.
This tool allows us to prove many polynomial surpassing and equalities without effort. An example
is given in the next subsection.
1.3. Multiplicity of the ELT Determinant. We return to Corollary 1.7, which holds in particular
over commutative ELT rings. We first formulate this corollary in the “ELT language”:
Corollary 1.14 (Multiplicativity of the ELT determinant). Let R be a commutative ELT ring.
n×n
If A, B ∈ R
, then
det (AB) det (A) · det (B)
We now present several corollaries from the multiplicativity of the ELT determinant, which are
two cases in which the ELT determinant is strictly multiplicative. First, we prove a lemma that will
be helpful for the first case:
Lemma 1.15. If x, y ∈ R satisfy x y, and if s (x) 6= 0, then x = y.
Proof. Write x = [ℓ]a, y = [k]b. Since x y, there is c ∈ F such that x = y + [0]c. In other words,
[ℓ]a
= [k]b + [0]c
By the definition of addition, a ≥ b, and a = max {b, c}. If a > b, then a = c > b, and thus
[ℓ]a
= [k]b + [0]c = [0]c
in contradiction to the fact that ℓ = s (x) 6= 0. Thus, x = y.
Corollary 1.16. If s (det (AB)) 6= 0, then
det (AB) = det (A) · det (B)
Proof. From Corollary 1.14,
det (AB) det (A) · det (B)
By Lemma 1.15, we get equality.
Another case in which the determinant is multiplicative is when either A or B are invertible:
n×n
Theorem 1.17. If A, B ∈ R
, such that A or B are invertible. Then
det (AB) = det (A) · det (B)
Proof. Assume that B is invertible (the second direction is proved similarly). We note that by
Corollary 1.14,
det (AB) det (A) · det (B)
but also
−1
det (A) = det (AB) B −1 det (AB) · det B −1 = det (AB) · (det (B)) .
ELT LINEAR ALGEBRA II
9
The latter surpassing implies that
det (A) · det (B) det (AB) .
Since is antisymmetric on R, the conclusion follows.
Although the determinant is not multiplicative, a natural question is: if AB is non-singular, is BA
also non-singular? The answer to this question is negative, as the following example demonstrates:
Example 1.18. In R = R (C, R), consider
A=
[1]1
[1]1
[1]2
[1]3
Then det (AAt ) = [1]8, yet det (At A) = [0]10.
1.4. The ELT Adjoint Matrix and Quasi-Invertible Matrices. As we have seen, the invertible
matrices in the ELT sense are limited. So, we shall try to generalize this. Our goal is to find an
equivalent condition to the fact that det (A) is invertible.
Definition 1.19. Let R be a commutative ELT ring. A quasi-identity matrix is a matrix
n×n
, which is idemopotent, nonsingular and defined by
I˜ ∈ R
(
[1]0
i=j
I˜
=
i,j
αi,j i 6= j
where αi,j ∈ R, s (αi,j ) = 0.
n×n
n×n
Definition 1.20. Let R be a commutative ELT ring, and let A ∈ R
. A matrix B ∈ R
is a quasi-inverse for A, if AB and BA are quasi-identity matrices. In this case, A is called
quasi-invertible. Note that AB and BA may differ.
n×n
. The (i, j)-minor A′i,j of
Definition 1.21. Let R be a commutative ELT ring, and let A ∈ R
a matrix A = (ai,j ) is obtained by deleting the i-th row and the j-th column. Its ELT determinant
is a′i,j = det A′i,j .
n×n
Definition 1.22. Let R be a commutative ELT ring, and let A ∈ R
. The adjoint matrix of
A is
(adj (A))i,j = [ε(i,j)]0 a′j,i
where ε (i, j) = (−1)i+j .
−1
We would like to prove that when det (A) is invertible in R, (det (A)) adj (A) is a quasi-inverse
of A. We present here some corollaries from the ELT transfer principle, which together will prove
the assertion (Theorem 1.26). We use the ELT transfer principles componentwise.
Corollary 1.23. A · adj (A) = det (A) IA , where IA I.
Proof. We use the ELT transfer principle for surpassing. This result is known in ring theory, and is
proved in the supertropical theory (see [15, Remark 4.5]).
Corollary 1.24.
det (A · adj (A)) = det (A)
n
Proof. We use the ELT transfer principle for equalities. This result is known in ring theory, and is
proved in the supertropical theory (see [15, Theorem 4.9]).
Corollary 1.25.
2
(A · adj (A)) = det (A) · A · adj (A)
Proof. We use the ELT transfer principle for equalities. This result is known in ring theory, and is
proved in the supertropical theory (see [15, Theorem 4.12]).
Theorem 1.26. If det (A) is invertible in R, then A is quasi-invertible.
10
GUY BLACHAR AND EREZ SHEINER
Proof. By Corollary 1.23, A · adj (A) = det (A) · IA , where IA I. It is left to prove that IA is
nonsingular and idempotent.
By Corollary 1.24, det (A · adj (A)) = det (A)n . But
n
n
det (A) = det (A · adj (A)) = det (det (A) IA ) = det (A) det (IA )
Since det (A) is invertible, det (IA ) = [1]0.
To prove that IA is idempotent, we use Corollary 1.25:
2
−1
−1
2
2
2
=
(det (A)) A · adj (A) = det (A)
IA
(A · adj (A)) =
−1
=
det (A)2
det (A) · A · adj (A) = (det (A))−1 A · adj (A) = IA
as required.
Remark 1.27. Using the same arguments, one may show that adj (A) · A is det (A) times a quasiidentity matrix.
Corollary 1.28. Let R be a commutative ELT ring. Then A is quasi-invertible if and only if det (A)
is invertible.
We will now use the theory of the ELT adjoint matrix to study the connection between matrix
singularity and linear dependency. We recall the following theorem:
Theorem ([4, Theorem 1.6]). Let R = R (F, R) be an ELT algebra, where F is an algebraically
n×n
closed field. Consider A ∈ R
. Then the rows of A are linearly dependent, iff the columns of
A are linearly dependent, iff s (det A) = 0F .
This theorem was proved using the Fundamental Theorem, thus only in the case of R = R (F, R)
where F is an algebraically closed field. We will prove the following:
n×n
an ELT matrix. If det A
Theorem 1.29. Let R be a commutative ELT ring, and let A ∈ R
is invertible in R, then the columns (respectively, rows) of A are linearly independent.
Before proving this theorem, we recall the Hungarian algorithm ([18]):
n×n
Definition 1.30. An entry ai,j of a tropical matrix A ∈ Gmax
is called column-critical if it
is maximal within its columns, i.e., if ∀k : ai,j ≥ ak,j . A matrix A is called critical if there exists a
permutation σ ∈ Sn such that a1,σ(1) , . . . , an,σ(n) are column-critical.
n×n
be a tropical matrix. Then there
Theorem 1.31 (The Hungarian Algorithm). Let A ∈ Gmax
are scalars α1 , . . . , αn ∈ R× such that
α1 ⊙ R1 (A)
..
.
αn ⊙ Rn (A)
n×n
is critical. In other words, there exists a diagonal matrix D ∈ Gmax
, whose diagonal entries
are not −∞, such that A ⊙ D is critical.
n×n
n×n
is critical if τ (A) = (τ (ai,j )) ∈ Gmax
In the ELT case, we say that a matrix A ∈ R
is critical.
n×n
n×n
Corollary 1.32. Let A ∈ R
. Then there exists an invertible diagonal matrix D ∈ R
such that DA is critical.
n×n
Proof. Let D′ = d′i,j ∈ Gmax
be a diagonal matrix such that D′ ⊙ τ (A) is a critical tropical
n×n
matrix. We define D = (di,j ) ∈ R
as
′
[1]d , i = j
i,i
di,j =
−∞, i 6= j
Obviously, DA is critical (since τ (DA) = τ (D) ⊙ τ (A) = D′ ⊙ τ (A) is critical), as required.
ELT LINEAR ALGEBRA II
11
We are now ready to prove Theorem 1.29.
Proof of Theorem 1.29. We prove the assertion on the columns of A. The assertion that the rows
of A are linearly independent can be proven by replacing A with At .
n
t
Suppose that s (Av) = (0, . . . , 0) for some v ∈ R× . If
A∇ = (det A)
−1
adj (A) ,
′
there exists a quasi-identity matrix IA
such that
′
A∇ A = IA
Thus,
t
′
s (IA
v) = s A∇ A · v = (0, . . . , 0) .
n×n
′ t
We apply Corollary 1.32 for (IA
) to find a diagonal invertible matrix D ∈ R
such that
′ t
′
D (IA
) is critical. By transposing this matrix, we get a matrix IA
D such that there is a permutation
′
′
σ ∈ Sn for which (IA
D) i, j is row-critical in IA
D. By Theorem 1.17,
′
′
det (IA
D) = det IA
· det D = det D
′
′
is not of layer zero, i.e. IA
D is nonsingular. We note that the only non-zero layered track in IA
is
′
the diagonal track, since any entry of IA which is not on the diagonal is of layer zero; hence, we may
assume σ = Id, that is the diagonal entries are row-critical.
′
′
′
Returning to the original equation s (IA
v) = (0, . . . , 0)t , we replace IA
by B = IA
D and v
′
−1
by v = D v to get
s (Bv ′ ) = (0, . . . , 0)t
Let 1 ≤ k ≤ n such that
τ (vk′ ) = max τ (vi′ ) .
1≤i≤n
Then
(Bv ′ )k =
n
X
i=1
bk,i vi′ =
n
X
bk,i vi′ + bk,k vk′ .
i=1
i6=k
′
′
We recall that B = IA
D, where IA
is a quasi-identity matrix and D is an invertible diagonal matrix.
Thus, every entry of B which is not on its diagonal is of layer zero. Furthermore, for any i 6= k we
have τ (vi′ ) ≤ τ (vk′ ) and τ (bk,i ) ≤ τ (bk,k ). Therefore each summand bk,i vi′ cannot dominate bk,k vk′ ,
implying
(Bv ′ )k = bk,k vk′
′
which implies s (bk,k vk′ ) = 0. Since bk,k = (IA
)k,k dk,k = dk,k is not of layer zero (because D is
′
′
invertible), we must have s (vk ) = 0. Now, vk = d−1
k,k vk implies s (vk ) = 0.
n
But v ∈ R× , therefore s (vk ) = 0 implies vk = −∞. By the choice of k, we must have
v = (−∞, . . . , −∞), which proves that the columns of A are linearly independent.
1.5. The ELT Characteristic Polynomial and ELT Eigenvalues.
n×n
Definition 1.33. Let A ∈ R
be a matrix. The ELT characteristic polynomial of A is
defined to be
pA (λ) = det λIn + [−1]0A
n×n
be a matrix. A vector v ∈ (R× )n is called an eigenvector of A
Definition 1.34. Let A ∈ R
with an eigenvalue x ∈ R if v 6= (−∞, ..., −∞) and
Av = xv.
Proposition 1.35. Let A ∈ R
n×n
be a matrix with eigenvalue x. Then s (pA (x)) = 0L .
12
GUY BLACHAR AND EREZ SHEINER
Proof. Choose v to be an eigenvector of the eigenvalue v, then Av = xv. Therefore
Av + [−1]0 · xv = [0]0 · xv.
In other words,
s xv + [−1]0Av = (0L , ..., 0L ).
Thus
s (xIn + [−1]0A)v = (0L , ..., 0L ).
By Theorem 1.29, we conclude that s det xIn + [−1]0A = 0L , i.e., s (pA (x)) = 0L .
Note that the other direction is not necessarily true, i.e., there could be an ELT root of the
characteristic polynomial which is not an eigenvalue. Indeed, if one tried to prove that direction, he
would encounter the following difficulty:
Av 6= Av + [−1]0 · xv + xv.
2×2
Example 1.36. Consider the matrix A ∈ R
,
[1]1 [1]2
A = [1]
,
2 [1]3
Its ELT characteristic polynomial is
pA (λ)
[−1]0A
[−1]2
λ + [−1]1
= det
[−1]2
λ + [−1]3
=
det λI2 +
=
(λ + [−1]1)(λ + [−1]3) + [−1]4 = λ2 + [−1]3 · λ + [0]4.
If λ = [0]1, λ = [ℓ]α with α < 1 or λ = [1]3, then s (pA (λ)) = 0.
The only eigenvalue of A is λ = [1]3, with
[1]1 [1]2
[1]0
[1]3
[1]0
[1]3
=
=
.
[1]2 [1]3
[1]1
[1]4
[1]1
One may also define an ELT eigenvalue and an eigenvector in the following way:
n×n
be a matrix. A vector v ∈ (R× )n is called an ELT eigenvector
Definition 1.37. Let A ∈ R
of A with an ELT eigenvalue x ∈ R if v 6= (−∞, ..., −∞) and
s Av + [−1]0xv = (0L , ..., 0L ).
This definition is similar to the concept of ’ghost surpass’ given by Izhakian, Knebusch and Rowen
(ref. [12]).
Proposition 1.38. Let R = R (F, R) be an ELT algebra, where F is an algebraically closed field,
n×n
and let A ∈ R
be a matrix. Then x is an ELT eigenvalue of A if and only if s (pA (x)) = 0L .
Proof. By Definition 1.37, x is an eigenvalue of A if and only if there exists a vector v ∈ (R× )n such
that v 6= (−∞, ..., −∞) and
s xv + [−1]0Av = (0L , ..., 0L ),
if and only if the columns of pA (x) = x · In + [−1]0A are linearly dependent, if and only if pA (x) is
singular ([4, Theorem 1.7]).
We finish by reformulating Cayley-Hamilton theorem (Corollary 1.8) in the “ELT language”.
Corollary 1.39 (ELT Cayley-Hamilton theorem). Let R be a commutative ELT ring, and let
n×n
be an ELT matrix. Then
A∈ R
s (pA (A)) = 0
ELT LINEAR ALGEBRA II
13
2. ELT Traces
2.1. Trace of ELT Matrices.
n×n
Definition 2.1. Let R be a commutative ELT ring, and take A ∈ R
, A = (ai,j ). The trace
of A is
n
X
ai,i
tr (A) =
i=1
Lemma 2.2. The ELT trace satisfies the following relations:
n×n
: tr (A + B) = tr (A) + tr (B).
(1) ∀A, B ∈ R
n×n
(2) ∀α ∈ R ∀A ∈ R
: tr (αA) = αtr (A).
n×n
(3) ∀A, B ∈ R
: tr (AB) = tr (BA).
Proof. These properties can be proved just as the classical theory.
2.2. ELT Nilpotent Matrices.
Definition 2.3. Let R be a commutative ELT ring. A matrix A ∈ R
n×n
if there exists m ∈ N such that Am ∈ [0] R
.
n×n
is called ELT nilpotent,
Similarly to the classical theory, one would expect that the trace of a nilpotent matrix would
be of layer zero; however, this is wrong. For this reason, we define the essential trace in the next
subsubsection.
[1]0 [0]1
Example 2.4. Let R = R (C, R), and consider the following matrix: A = [0]
. Then
0 [1]0
tr (A) = [2]0, while
[1]0 [0]1
[1]0 [0]1
[0]1 [0]1
A2 = [0] [1]
=
.
[0]0 [1]0
[0]0 [0]1
0
0
Therefore, ELT nilpotent matrices don’t necessarily have zero-layered trace.
Another interesting example is an ELT nilpotent matrix, whose determinant is not of layer zero.
Example 2.5. Let R be a commutative ELT ring, and take a, b, c, d ∈ R such that
τ a2 < τ (bc) < τ d2
a b
and s (d) = 0. Consider the matrix A =
. We have
c d
2
a + bc ab + bd
bc bd
2
A =
=
ac + cd bc + d2
cd d2
and
abc + bcd abd + bd2
bcd bd2
=
bc2 + cd2 bcd + d3
cd2 d3
which is of layer zero, since s (d) = 0.
But we may choose a, b, c, d such that τ (bc) > τ (ad) and s (bc) 6= 0; in that case, A is quasiinvertible and ELT nilpotent.
A3 =
2.3. The Essential Trace. Before defining the new notion of trace, we give an important definition,
which is significant in our construction of the new trace.
n
X
hi ∈ R [λ] be an ELT
X
polynomial, where each hi is a monomial. For a monomial h, define ph (λ) =
hi .
Definition 2.6. Let R be a commutative ELT ring, and let p (λ) =
i=1
hi 6=h
(1) The monomial h is called inessential at a point a ∈ R, if p (a) = ph (a) and τ (h (a)) <
τ (p (a)). If h is inessential at every point of R, it is called inessential.
14
GUY BLACHAR AND EREZ SHEINER
(2) The monomial h is called essential at a point a ∈ R, if p (a) = h (a) and τ (ph (a)) <
τ (p (a)). If h is essential at some point of R, it is called essential.
(3) The monomial h is called quasi-essential at a point a ∈ R, if it is neither inessential at
a nor essential at a. If h is quasi-essential at some point of R, it is called quasi-essential.
Throughout the rest of the section, we need a stronger assumption on our ELT algebras. We
require them to be divisible ELT fields, that is ELT algebras of the form R = R (L , F ), where
F is a divisible group and L is a field.
n×n
. We define
Definition 2.7. Let R be a divisible ELT field, and let A ∈ R
τ (cℓ )
τ (ck )
L (A) =
ℓ ≥ 1 ∀k ≤ n :
≥
ℓ
k
µ (A) = min (L (A))
We write µ for µ (A), if A is given.
n
Lemma 2.8. Let R be a divisible ELT field, and let p (λ) = λ +
n
X
αi λn−i ∈ R [λ] be an ELT
i=1
polynomial. Then the first monomial after λn that is not inessential is αµ λn−µ .
Proof. We need to find the monomial for which the intersection between λn and αℓ λn−ℓ is maximal
(in the sense that its tangible value is maximal).
First, we compute the tangible value of the intersection:
λn = αℓ λn−ℓ ⇒ τ (λ) =
τ (αℓ )
ℓ
ℓ)
The tangible value of the value of the polynomial at that point is nτ (α
.
ℓ
So, if
τ (αℓ )
τ (αk )
∀k ≤ n :
≥
ℓ
k
ℓ satisfies our conditions. Take such ℓ minimal, which is µ, and we are done.
Remark 2.9. If |L (A)| ≥ 2, then αµ λn−µ is only quasi-essential in pA (λ).
n×n
Definition 2.10. Let R be a divisible ELT field, and let A ∈ R
be a matrix. Assume
n
X
αi λn−i . αµ is called the dominant characteristic coefficient. The essential
pA (λ) = λn +
i=1
trace of A, denoted etr (A), is given by the formula
(
[−1]0tr (A) λn−1 is essential in p (λ)
tr (A)
A
etr (A) = [0] τ (αµ )
Otherwise
µ
Lemma 2.11. If s (tr (A)) = 0, then s (etr (A)) = 0.
Proof. If [−1]0tr (A) λn−1 is essential in pA (λ), then etr (A) = tr (A). which is of layer zero. Otherwise, [−1]0tr (A) λn−1 is not essential in pA (λ), and thus, by the definition of essential trace,
s (etr (A)) = 0.
n×n
be a matrix over a commutative ELT ring R. A path from i
Definition 2.12. Let A ∈ R
to j is an expression of the form
s = ai,i2 ai2 ,i3 . . . aik ,j
where 1 ≤ i1 , . . . , ik+1 ≤ n. If i = j, we call it a multicycle. The length of s is |s| = k. The
τ (s)
tangible average value of s is
. A simple cycle is a multicycle from i to i, such that ij 6= ij ′
|s|
for j 6= j ′ (with i1 = i).
Fact 2.13. Every multicycle can be written as a product of simple cycles.
ELT LINEAR ALGEBRA II
15
Lemma 2.14. The (i, j) element in Ak is the sum of all paths from i to j. That is,
X
ai,i2 ai2 ,i3 . . . aik ,j
Ak i,j =
1≤i2 ,...,ik ≤n
Proof. By induction on k, where the case k = 1 is clear. If the assertion is true for some k, then
Ak+1
i,j
=
n
X
ℓ=1
ai,ℓ Akℓ,j =
n
X
X
ai,ℓ aℓ,i2 ai2 ,i3 . . . aik ,j =
j=1 1≤i2 ,...,ik ≤n
X
ai,i2 ai2 ,i3 . . . aik+1 ,j
1≤i2 ,...,ik+1 ≤n
Lemma 2.15. The coefficient of λn−k in pA (λ) is
X
X
k
[−1]0
[sgnσ]0 a
i1 ,iσ(1) ai2 ,iσ(2) . . . aik ,iσ(k)
1≤i1 ,...,ik ≤n σ∈Sk
Proof. We must choose n − k indices from which we “take” λ in the expansion of det λI + [−1]0A ;
we are left with a k × k submatrix, with rows i1 , . . . , ik from A. Its determinant is the inner sum.
Lemma 2.16. Any multicycle contributing to the dominant characteristic coefficient must be a
simple cycle.
Proof. Otherwise, assume it is not a simple cycle. Since it can be written as a product of simple
(s)
≥ etr (A), and a shorter length. Thus, s would give
cycles, at least one of which, s, would have τ|s|
a dominant characteristic coefficient of lower degree, in contradiction to our assumption.
Lemma 2.17. If s (etr (A + B)) 6= 0, then etr (A + B) = etr (A) + etr (B).
Proof. The assumption can only happen if [−1]0tr (A + B) λn−1 is essential in pA+B (λ), and etr (A + B) =
tr (A + B).
The multicycles of A + B are products of sums of elements from A and from B. In particular,
every multicycle of A and of B is a part of a multicycle of A + B.
We know that tr (A + B) = tr (A) + tr (B); if τ (tr (A + B)) = τ (tr (A)), then τ (tr (A)) >every
average value of a multicycle of A; so etr (A) = tr (A). We are almost finished:
Case 1.
If τ (tr (A + B)) = τ (tr (A)) = τ (tr (B)), then etr (A) = tr (A), etr (B) = tr (B), and
etr (A + B) = tr (A + B) = tr (A) + tr (B) = etr (A) + etr (B)
Case 2.
If τ (tr (A + B)) = τ (tr (A)) > τ (tr (B)), then etr (A) = tr (A), and
etr (A + B) = tr (A + B) = tr (A) + tr (B) = tr (A) = etr (A)
In particular, τ (etr (A)) > τ (etr (B)); otherwise, there would have been a multicycle from
B with τ (average value) ≥ τ (etr (A)) = τ (etr (A + B)), and thus s (etr (A + B)) = 0,
which is a contradiction. So
etr (A + B) = etr (A) = etr (A) + etr (B)
Example 2.18. In general, it is not true that
etr (A + B)
etr (A) + etr (B). For example, take
[1]0
[1]0
in (R (C, R))2×2 the following matrices: A =
, B = At . Then etr (A) = etr (B) = [2]0.
−∞ [1]0
[2]0 [1]0
However, A + B = [1]
, and pA+B (λ) = λ2 + [−4]0λ + [3]0. The monomial [−4]0λ is quasi0 [2]0
essential, and thus etr (A + B) = [0]0.
Lemma 2.19. etr (AB) = etr (BA).
16
GUY BLACHAR AND EREZ SHEINER
Proof. By Lemma 2.16, it is enough to check only simple cycles. Assume
(AB)i1 ,i2 (AB)i2 ,i3 . . . (AB)ik ,i1
contributes to the dominant characteristic coefficient, where ij 6= ij ′ for j 6= j ′ . So there are ℓ1 , . . . , ℓk
such that
s = ai1 ,ℓ1 bℓ1 ,i2 ai2 ,ℓ2 bℓ2 ,i3 . . . aik ,ℓk bℓk ,i1
contributes to the dominant characteristic coefficient, i.e. τ (s) = etr (AB).
If ℓj = ℓj ′ for j < j ′ , write
s =
ai1 ,ℓ1 bℓ1 ,i2 . . . aij ,ℓj bℓj ,ij′ +1 . . . aik ,ℓk bℓk ,i1 bℓj ,ij+1 aij+1 ,ℓj+1 . . . aij′ ,ℓj′ =
=
ai1 ,ℓ1 bℓ1 ,i2 . . . aij ,ℓj bℓj ,ij′ +1 . . . aik ,ℓk bℓk ,i1 aij+1 ,ℓj+1 . . . aij′ ,ℓj′ bℓj ,ij+1
{z
}|
{z
}
|
s1
s2
Since τ (s) = τ (etr (AB)), τ (s1 ) ≥ τ (etr (AB)) or τ (s2 ) ≥ τ (etr (AB)). Since s1 and s2 are shorter,
we get a contradiction.
So ℓj 6= ℓj ′ for j 6= j ′ . So
s = bℓ1 ,i2 ai2 ,ℓ2 bℓ2 ,i3 . . . aik ,ℓk bℓk ,i1 ai1 ,ℓ1
is a part of the simple cycle
(BA)ℓ1 ,ℓ2 (BA)ℓ2 ,ℓ3 . . . (BA)ℓk ,ℓ1
in BA. So every simple cycle contributing to etr (AB) also contributes to etr (BA).
By symmetry, the opposite is true as well; so etr (AB) = etr (BA).
Lemma 2.20. If there is a multicycle s = ai1 ,i2 ai2 ,i3 . . . aik ,i1 of length k ≥ 2 such that
τ tr (A)k < τ (s)
then [−1]0tr (A) λn−1 is not essential in pA (λ) (meaning, it is either inessential or quasi-essential).
In other words, µ (A) ≥ 2.
Proof. Firstly, we may assume that k ≤ n; otherwise, write s as a product of multicycle, s = s1 . . . sj .
Since
j
X
τ tr (A)k < τ (s) =
τ (sℓ )
|sℓ |
ℓ=1
< τ (sℓ ), and we may replace k by ℓ.
n
X
αi λn−i . By Lemma 2.15,
Therefore, we assume k ≤ n. Write pA (λ) = λn +
At least one sℓ should satisfy τ tr (A)
i=1
ck = ai1 ,i2 ai2 ,i3 . . . aik ,i1 + · · · = s + · · ·
In particular, τ (ck ) ≥ τ (s). Also recall that c1 = [−1]0tr (A). Therefore,
k
τ tr (A)
τ (s)
τ (ck )
τ (c1 ) = τ (tr (A)) =
<
≤
k
k
k
An thus µ (A) ≥ k ≥ 2, as required.
[−1]0tr (A) λn−1
Lemma 2.21. If A is ELT nilpotent, and if s (tr (A)) 6= 0, then
is not essential in
pA (λ). In other words, [−1]0tr (A) λn−1 is either quasi-essential or inessential in pA (λ).
Proof. For this proof, write a ≤τ b if τ (a) ≤ τ (b), a <τ b if τ (a) < τ (b), and a ≡τ b if τ (a) = τ (b).
We take some ai,i such that ai,i ≡τ tr (A) and s (ai,i ) 6= 0; without loss of generality, let i = 1.
n
X
αi λn−i .
Write pA (λ) = λn +
i=1
By Lemma 2.14, Ak 1,1 = ak1,1 + · · · . Take k minimal such that s Ak = 0. In particular,
s Ak 1,1 = 0, so we have two cases:
ELT LINEAR ALGEBRA II
Case 1.
17
There is a multicycle s = a1,i2 ai2 ,i3 . . . aik ,1 with ak1,1 <τ s. Firstly, we may assume that
k ≤ n; otherwise, write s as a product of multicycle, s = s1 . . . sj . Since
τ tr (A)
k
< τ (s) =
j
X
τ (sℓ )
ℓ=1
At least one sℓ should satisfy τ tr (A)|sℓ | < τ (sℓ ), and we may replace k by ℓ.
n
X
αi λn−i . By Lemma 2.15,
Therefore, we assume k ≤ n. Write pA (λ) = λn +
i=1
ck = ai1 ,i2 ai2 ,i3 . . . aik ,i1 + · · · = s + · · ·
Case 2.
In particular, τ (ck ) ≥ τ (s). Also recall that c1 = [−1]0tr (A). Therefore,
k
τ tr (A)
τ (s)
τ (ck )
τ (c1 ) = τ (tr (A)) =
<
≤
k
k
k
An thus µ (A) ≥ k ≥ 2, as required.
There is no multicycle with ak1,1 <τ a1,i2 ai2 ,i3 . . . aik ,1 . Then there has to be a multicycle
for which ak1,1 ≡τ a1,i2 ai2 ,i3 . . . aik ,1 = s (since s Ak 1,1 = 0, and ak1,1 is a summand in
the sum defining Ak 1,1 ).
By writing s as a product of simple cycles, we may assume that each simple cycle
aj1 ,j2 aj2 ,j3 . . . ajℓ ,j1 ≡τ aℓ1,1 (Otherwise, we are finished by the first case).
Since ℓ < n, we get that |L (A)| ≥ 2, meaning [−1]0tr (A) λn−1 is quasi-essential in
pA (λ).
Corollary 2.22. If A is ELT nilpotent, then s (etr (A)) = 0.
Proof. There are two cases:
(1) If s (tr (A)) = 0, then s (etr (A)) = 0 by Lemma 2.11.
(2) Otherwise, s (tr (A)) 6= 0; but then s (etr (A)) = 0 by Lemma 2.21.
References
[1] Marianne Akian, Guy Cohen, Stephane Gaubert, R Nikoukhah, and Jean Pierre Quadrat. Linear systems in
(max,+) algebra. In Decision and Control, 1990., Proceedings of the 29th IEEE Conference on, pages 151–156.
IEEE, 1990.
[2] Marianne Akian, Stéphane Gaubert, and Alexander Guterman. Linear independence over tropical semirings and
beyond. Contemporary Mathematics, 495:1, 2009.
[3] Marianne Akian, Stéphane Gaubert, and Alexander Guterman. Tropical cramer determinants revisited. Tropical
and Idempotent Mathematics and Applications, 616:45, 2014.
[4] Guy Blachar and Erez Sheiner. Elt linear algebra. arXiv preprint arXiv:1603.02204, 2016.
[5] Peter Butkovič. Max-algebra: the linear algebra of combinatorics? Linear Algebra and its applications, 367:313–
335, 2003.
[6] Mike Develin, Francisco Santos, and Bernd Sturmfels. On the rank of a tropical matrix. Combinatorial and
computational geometry, 52:213–242, 2005.
[7] David Dolžan and Polona Oblak. Invertible and nilpotent matrices over antirings. arXiv preprint arXiv:0806.2996,
2008.
[8] Ido Efrat. Valuations, orderings, and Milnor K-theory. Number 124. American Mathematical Soc., 2006.
[9] Stéphane Gaubert. Théorie des systèmes linéaires dans les dioı̈des. PhD thesis, 1992.
[10] Stephane Gaubert. Methods and applications of (max,+) linear algebra. In STACS 97, pages 261–282. Springer,
1997.
[11] Zur Izhakian. Tropical arithmetic and matrix algebra. Communications in Algebra R , 37(4):1445–1468, 2009.
[12] Zur Izhakian, Manfred Knebusch, and Louis Rowen. Dual spaces and bilinear forms in supertropical linear
algebra. Linear and Multilinear Algebra, 60(7):865–883, 2012.
[13] Zur Izhakian and Louis Rowen. The tropical rank of a tropical matrix. Communications in Algebra R ,
37(11):3912–3927, 2009.
[14] Zur Izhakian and Louis Rowen. Supertropical algebra. Advances in Mathematics, 225(4):2222–2286, 2010.
18
GUY BLACHAR AND EREZ SHEINER
[15] Zur Izhakian and Louis Rowen. Supertropical matrix algebra. Israel Journal of Mathematics, 182(1):383–424,
2011.
[16] Zur Izhakian and Louis Rowen. Supertropical matrix algebra II: Solving tropical equations. Israel Journal of
Mathematics, 186(1):69–96, 2011.
[17] Zur Izhakian and Louis Rowen. Supertropical matrix algebra III: Powers of matrices and their supertropical
eigenvalues. Journal of Algebra, 341(1):125–149, 2011.
[18] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(12):83–97, 1955.
[19] Brett Parker. Exploded manifolds. Advances in Mathematics, 229(6):3256–3319, 2012.
[20] Louis Halle Rowen. Algebras with a negation map. arXiv preprint arXiv:1602.00353, 2016.
[21] Sergei Sergeev. Max-plus definite matrix closures and their eigenspaces. Linear algebra and its applications,
421(2-3):182–201, 2007.
[22] Erez Sheiner. Exploded Layered Tropical Algebra. PhD thesis, Bar-Ilan University, 2015.
[23] Yijia Tan. On invertible matrices over antirings. Linear algebra and its applications, 423(2):428–444, 2007.
[24] Jean-Pierre Tignol and R Adrian. Value functions on simple algebras, and associated graded rings. Springer,
2015.
Department of Mathematics, Bar-Ilan University, Ramat-Gan 52900, Israel.
E-mail address: [email protected]
Department of Mathematics, Bar-Ilan University, Ramat-Gan 52900, Israel.
E-mail address: [email protected]
| 0math.AC
|
arXiv:1707.05258v4 [math.AG] 28 Jul 2017
ON RATIONAL CUSPIDAL PLANE CURVES, AND THE LOCAL
COHOMOLOGY OF JACOBIAN RINGS
ALEXANDRU DIMCA1
Abstract. This note gives the complete projective classification of rational, cuspidal plane curves of degree at least 6, and having only weighted homogeneous
singularities. It also sheds new light on some previous characterizations of free and
nearly free curves in terms of Tjurina numbers. Finally, we suggest a stronger form
of Terao’s conjecture on the freeness of a line arrangement being determined by its
combinatorics.
1. Introduction
The main result we prove in this note is the following.
Theorem 1.1. Let C be a rational, cuspidal plane curve of degree d ≥ 6 such that
C has only weighted homogeneous singularities. Then C is projectively equivalent to
exactly one of the following φ(d)/2 models
Cd,k : y d + xk z d−k = 0,
where the integer k satisfies 1 ≤ k < d/2 and k is relatively prime to d. Here
φ denotes the Euler function, with φ(d) counting the number of integers m, where
1 ≤ m ≤ d − 1 and m relatively prime to d.
A slightly stronger result is stated and proved below, see Propositions 4.1 and
4.3. For detailed information on the rational, cuspidal plane curves of degree ≤ 5,
we refer to [14] and the references given there. The curves Cd,k above are called
binomial cuspidal curves in [14, Section 7.1]. The necessary restriction d ≥ 6 in
Theorem 1.1 is discussed in Example 4.4 below.
Theorem 1.1 is quite surprising, given that the classification of rational, cuspidal
plane curves, even of those having at most two singularities, is rather complicated,
see Sakai-Tono paper [16], or Propositions 3.1 and 3.2 in [9] where their results are
quoted. It is known that there are no rational, cuspidal free curves of degree d ≥ 6
having only weighted homogeneous singularities, see [9, Theorem 2.8]. Theorem 1.1
says that if the word free is deleted from the above statement, only the binomial
cuspidal curves may occur, and they are all nearly free. Note also that a weighted
2010 Mathematics Subject Classification. Primary 14H50; Secondary 14B05, 13D02, 32S22.
Key words and phrases. Jacobian ideal, Tjurina number, free curve, nearly free curve.
1
This work has been supported by the French government, through the UCAJEDI Investments
in the Future project managed by the National Research Agency (ANR) with the reference number
ANR-15-IDEX-01.
1
2
ALEXANDRU DIMCA
homogeneous cusp has only one Puiseux pair, but the class of plane curve singularities with exactly one Puiseux pair is much larger. This explains the complicated
classification of rational unicuspidal curves with a unique Puiseux pair in [13].
To prove Theorem 1.1, we use a number of key results, proved by A. A. du Plessis
and C.T.C. Wall in [11, 12], the main one being stated below in Theorem 3.1. To
make our note more self-contained, we include a short proof of a similar result to
Theorem 3.1, which we explain now.
Let S = C[x, y, z] be the graded polynomial ring in three variables x, y, z and let
C : f = 0 be a reduced curve of degree d in the complex projective plane P2 . The
minimal degree of a Jacobian relation, or Jacobian syzygy, for the polynomial f is
the integer mdr(f ) defined to be the smallest integer m ≥ 0 such that there is a
nontrivial relation
(1.1)
afx + bfy + cfz = 0
among the partial derivatives fx , fy and fz of f with coefficients a, b, c in Sm , the
vector space of homogeneous polynomials of degree m. When mdr(f ) = 0, then C is
a union of lines passing through one point, a situation easy to analyse. We assume
from now on that
mdr(f ) ≥ 1.
Denote by τ (C) the global Tjurina number of the curve C, which is the sum of the
Tjurina numbers of the singular points of C. We denote by Jf the Jacobian ideal
of f , i.e. the homogeneous ideal in S spanned by fx , fy , fz , and by M(f ) = S/Jf
the corresponding graded ring, called the Jacobian (or Milnor) algebra of f . Let If
denote the saturation of the ideal Jf with respect to the maximal ideal m = (x, y, z)
in S and consider the local cohomology group
0
N(f ) = If /Jf = Hm
(M(f ).
It was shown in [6] that the graded S-module N(f ) satisfies a Lefschetz type
property with respect to multiplication by generic linear forms. This implies in
particular the inequalities
0 ≤ n(f )0 ≤ n(f )1 ≤ ... ≤ n(f )[T /2] ≥ n(f )[T /2]+1 ≥ ... ≥ n(f )T ≥ 0,
where T = 3d − 6 and n(f )k = dim N(f )k for any integer k. We set
ν(C) = max{n(f )j }.
j
When d = 2m is even, then the above implies that n(f )3m−3 = ν(C). When d =
2m + 1 is odd, then the above and the self duality of the graded S-module N(f ), see
[17, 19], implies that
n(f )3m−2 = n(f )3m−1 = ν(C).
The second main result of this note is the following.
Theorem 1.2. Let C : f = 0 be a reduced plane curve of degree d and let r = mdr(f ).
Then the following hold.
(1) If r < d/2, then
ν(C) = (d − 1)2 − r(d − 1 − r) − τ (C).
RATIONAL CUSPIDAL CURVES AND LOCAL COHOMOLOGY OF JACOBIAN RINGS
3
(2) If r ≥ (d − 2)/2, then
ν(C) = ⌈
3
(d − 1)2 ⌉ − τ (C).
4
Here, for any real number u, ⌈u⌉ denotes the round up of u, namely the smallest
integer U such that U ≥ u. Written down explicitly, this means that for d = 2m is
even and r ≥ m − 1, one has ν(C) = 3m2 − 3m + 1 − τ (C), while for d = 2m + 1 is
odd and r ≥ m, one has ν(C) = 3m2 − τ (C). For (d − 2)/2 ≤ r < d/2, both formulas
in (1) and (2) apply, and they give the same result for ν(C). The relation between
Theorem 1.2 and du Plessis-Wall result in Theorem 3.1 is discussed in Remark 3.2
below.
When C : f = 0 is a line arrangement, examples due to G. Ziegler show that the
invariant mdr(f ) is not combinatorially determined, see [1, Example 4.3]. The above
result suggests that the following stronger version of H. Terao’s conjecture, saying
that the freeness of a line arrangement is combinatorially determined, might be true.
For more on Terao’s conjecture and free hyperplane arrangements we refer to [3].
Conjecture 1.3. Let C : f = 0 be a line arrangement in P2 . Then the invariant
ν(C) is combinatorially determined.
For a reduced plane curve one may state the following.
Conjecture 1.4. Let C : f = 0 be a reduced plane curve P2 . Then the invariant
ν(C) is determined by the degree of C and the list of the analytic types of the isolated
singularities of C.
Some cases where these conjecture hold are described in Example 4.4 (i) (where the
corresponding invariant mdr(f ) is not determined by the data stated in Conjecture
1.4), and in Proposition 3.5.
2. The proof of Theorem 1.2
Consider the graded S−submodule AR(f ) ⊂ S 3 of all relations involving the
partial derivatives of f , namely
ρ = (a, b, c) ∈ AR(f )m
if and only if afx + bfy + cfz = 0 and a, b, c are in Sm . We set ar(f )k = dim AR(f )k
for any integer k.
Consider the rank two vector bundle T hCi = Der(−logC) of logarithmic vector fields along C, which is the coherent sheaf associated to the graded S-module
AR(f )(1). Using the results in the third section of [7], for any integer k one has
d+k+2
k+3
+ τ (C).
−
(2.1)
χ(T hCi(k)) = 3
2
2
Moreover, one has the following for E = T hCi and any integer k, see [7], [17].
(2.2)
h0 ((E(k)) = ar(f )k+1, h1 ((E(k)) = n(f )d+k and h2 ((E(k)) = ar(f )d−5−k .
4
ALEXANDRU DIMCA
Assume that we are in the case d = 2m and apply the formulas (2.1) and (2.2) for
k = m − 3. We get
2ar(f )m−2 − ν(C) = τ (C) − 3m2 + 3m − 1.
Let r = mdr(f ) and note that ar(f )m−2 = 0 if r ≥ m − 1 = [(d − 1)/2] and
m−r
ar(f )m−2 = dim Sm−2−r =
2
if r ≤ m − 2. Indeed, it follows from [18, Lemma 1.4] that the S-module AR(f )
cannot have two independent elemens of degree d1 and d2 satisfying d1 + d2 < d − 1.
Assume next that we are in the case d = 2m + 1 and apply the formulas (2.1) and
(2.2) again for k = m − 3. We get
ar(f )m−2 − ν(C) + ar(f )m−1 = τ (C) − 3m2 .
As above, if r = mdr(f ), we have ar(f )m−2 = ar(f )m−1 = 0 if r ≥ m = [(d − 1)/2]
and
m−r+1
m−r
and ar(f )m−1 = dim Sm−1−r =
ar(f )m−2 = dim Sm−2−r =
2
2
if r ≤ m − 1, with the convention 12 = 0. These formulas, plus some simple
computation, prove Theorem 1.2 stated in the Introduction.
Remark 2.1. An alternative proof of Theorem 1.2 can be obtained using the results
in [1]. Indeed, Proposition 3.2 in [1] implies that dL1 0 = r for r < (d − 2)/2 and
dL1 0 = ⌊(d − 1)/2⌋ for r ≥ (d − 2)/2. Here (dL1 0 , dL2 0 ), with dL1 0 ≤ dL2 0 , denotes the
splitting type of the vector bundle T hCi(−1) along a generic line L0 in P2 . Then [1,
Theorem 1.1] says that
(d − 1)2 − dL1 0 dL2 0 = τ (C) + ν(C).
Since dL1 0 + dL2 0 = d − 1 by [1, Proposition 3.1], these facts give a new proof of
Theorem 1.2.
In view of this remark, Conjecture 1.3 may be restated as follows.
Conjecture 2.2. Let C : f = 0 be a line arrangement in P2 . Then the generic splitting type (dL1 0 , dL2 0 ) of the vector bundle T hCi(−1) is combinatorially determined.
3. Some related results and direct applications
We start this section by recalling the following result due to du Plessis and Wall,
see [11, Theorem 3.2].
Theorem 3.1. For positive integers d and r, define two new integers by
τ (d, r)min = (d − 1)(d − r − 1) and τ (d, r)max = (d − 1)2 − r(d − r − 1).
Then, if C : f = 0 is a reduced curve of degree d in P2 and r = mdr(f ), one has
τ (d, r)min ≤ τ (C) ≤ τ (d, r)max .
RATIONAL CUSPIDAL CURVES AND LOCAL COHOMOLOGY OF JACOBIAN RINGS
5
Moreover, for r = mdr(f ) ≥ d/2, the stronger inequality
2r + 2 − d
τ (C) ≤ τ (d, r)max −
2
holds.
Remark 3.2. Let C : f = 0 be a reduced curve of degree d in P2 and r = mdr(f ).
Note that the function τ (d, r)max , regarded as a function of r, occurs also in Theorem
1.2 (1), which can be restated as
(3.1)
τ (C) + ν(C) = τ (d, r)max ,
for r = mdr(f ) < d/2. The inequality τ (C) ≤ τ (d, r)max in Theorem 3.1 is made
more precise, when r < d/2, by the result in Theorem 1.2 (1).
On the other hand, note that τ (d, r)max is a decreasing (resp. increasing) function
of r on the interval [0, (d − 1)/2] (resp. on the interval [(d − 1)/2, +∞)), and the
number ⌈ 43 (d − 1)2 ⌉ that occurs in Theorem 1.2 (2) is precisely the minimum of the
values of the function τ (d, r)max for r ∈ [0, (d − 1)/2] ∩ N. Hence Theorem 1.2 (2)
implies only
3
τ (C) ≤ ⌈ (d − 1)2 ⌉,
4
for r ≥ d/2, which is weaker than the inequality in Theorem 3.1, though enough for
many applications.
At the end of the proof of Theorem 3.1, in [11], the authors state the following
very interesting consequence (of the proof, not of the statement) of Theorem 3.1.
Corollary 3.3. Let C : f = 0 be a reduced curve of degree d in P2 and r = mdr(f ).
One has
τ (C) = τ (d, r)max
if and only if C : f = 0 is a free curve, and then r < d/2.
Since a plane curve C is free if and only if ν(C) = 0, this characterization of free
curves follows also from Theorem 1.2, as explained in Remark 3.2.
In the paper [4], we have given an alternative proof of Corollary 3.3 and have
shown that a plane curve C is nearly free, which can be defined by the property
ν(C) = 1, if and only if a similar property holds. Namely, one has the following
result, an obvious consequence of Theorem 1.2.
Proposition 3.4. Let C : f = 0 be a reduced curve of degree d in P2 and r = mdr(f ).
One has
τ (C) = τ (d, r)max − 1
if and only if C : f = 0 is a nearly free curve, and then r ≤ d/2.
Concerning Conjecture 1.3 in Introduction we have the following.
6
ALEXANDRU DIMCA
Proposition 3.5. Let C be a plane curve of degree d such that, for d = 2m (resp.
d = 2m + 1) one has τ (C) < τ (d, m − 2)min = (m + 1)(2m − 1) (resp. τ (C) <
τ (d, m − 1)min = 2m(m + 1)). Then
3
(d − 1)2 ⌉ − τ (C).
4
In particular, if C be a line arrangement (resp. a reduced plane curve) satisfying
these inequalities, then Conjecture 1.3 (resp. Conjecture 1.4) holds for C.
ν(C) = ⌈
Proof. We give the proof only in the case d = 2m even. The function τ (d, r)min is
a strictly decreasing function of r on R, so Theorem 3.1 implies that r = mdr(f ) ≥
m − 1. But then Theorem 1.2 (2) implies that ν(C) is determined by d and τ (C).
When C is a line arrangement, d and τ (C) are both combinatorially determined.
For a plane curve in general, τ (C) is determined by the analytic type of the local
singularities, but not by their topological type.
Example 3.6. Let C = A : f = 0 be a line arrangement consisting of d lines. For
d = 6, there are 10 possibilities for the intersection lattice L(A), see for details [5,
Section 5.6]. Out of them, 8 satisfy the inequality τ (A) < 20 from Proposition 3.5.
For d = 6, r = mdr(f ) is determined by the intersection lattice L(A), and there are
exactly 8 lattices with r ≥ 2.
For d = 7, there are 23 possibilities for the intersection lattice L(A). Out of them,
only 4 satisfy the inequality τ (A) < 24 from Proposition 3.5. For d = 7, r = mdr(f )
is again determined by the intersection lattice L(A), and there are exactly 19 lattices
with r ≥ 3, for which ν(C) = ν(A) is given by the formula in Proposition 3.5.
It would be interesting to find a lower bound for τ (A) in terms of d and r, in the
case of line arrangements, which is better than the bound τ (r)min given by Theorem
3.1.
4. The proof of Theorem 1.1
The following result is related to the main conjecture in [10], namely that a rational, cuspidal plane curve is either free or nearly free. It is also the first step in
proving Theorem 1.1.
Recall that a plane curve C has only weighted homogeneous singularities if and
only if µ(C) = τ (C), where µ(C) denotes the sum of all the Milnor numbers of the
singularities of C, see [15]. In general one has the obvious inequality τ (C) ≤ µ(C).
Proposition 4.1. Let C : f = 0 be an irreducible curve in P2 , of degree d ≥ 6. Then
the following properties are equivalent.
(1) τ (C) ≥ d2 − 4d + 8.
(2) r = mdr(f ) = 1
(3) C is a rational, cuspidal plane curve such that µ(C) = τ (C), i.e. C has only
weighted homogeneous singularities.
If any of these properties hold, then C is a nearly free curve.
RATIONAL CUSPIDAL CURVES AND LOCAL COHOMOLOGY OF JACOBIAN RINGS
7
Proof. To prove that (1) implies (2), recall that the function τ (d, r)max is a decreasing
function of r on the interval [1, (d − 1)/2] and note that τ (d, r)max = τ (d, r ′ )max for
r + r ′ = d − 1. For r = 2 we get τ (d, 2)max = d2 − 4d + 7. Then (1) implies that
r = mdr(f ) has to be 1. Indeed, the value r = d − 2 is excluded using the stronger
final inequality in Theorem 3.1. Note that Theorem 1.2 (2) is also enough for this
purpose, as discussed in Remark 3.2. To prove that (2) implies (3), note that
τ (d, 1)min = d2 − 3d + 2 ≤ τ (C) ≤ τ (d, 1)max = d2 − 3d + 3.
The genus g(C) is given by
2g(C) = (d − 1)(d − 2) −
X
(µ(C, p) + r(C, p) − 1),
p
where p runs through the singular points of C, µ(C, p) is the Milnor number of the
singularity (C, p) and r(C, p) is the number of its branches. It follows that
X
µ(C, p) ≤ (d − 1)(d − 2),
µ(C) =
p
with equality if and only if C is a rational cuspidal curve. Then the inequality
τ (C) ≤ µ(C) ≤ (d − 1)(d − 2) forces the equality
τ (d, 1)min = d2 − 3d + 2 = τ (C)
and we see that C is nearly free using Proposition 3.4. The fact that (3) implies (1)
is obvious as soon as d ≥ 6.
Remark 4.2. Irreducible, cuspidal nearly free curves satisfying µ(C) = τ (C), but
not rational have been constructed in [2]. For instance, for any odd integer k ≥ 1, it
is shown that the irreducible curve
C2k : f = x2k + y 2k + z 2k − 2(xk y k + xk z k + y k z k ) = 0,
has 3k singular points of type Ak−1 as singularities, it is a nearly free curve with
mdr(f ) = k and has genus
g(C2k ) =
(k − 1)(k − 2)
.
2
For k = 3, we have in this case τ (C) = 18 < τ (d, 2)max + 1 = 20.
The following result completes the proof of Theorem 1.1.
Proposition 4.3. Let C : f = 0 be an irreducible plane curve of degree d ≥ 6 such
that mdr(f ) = 1. Then C is projectively equivalent to exactly one of the following
φ(d)/2 models
Cd,k : y d + xk z d−k = 0,
where the integer k satisfies 1 ≤ k < d/2 and k is relatively prime to d.
8
ALEXANDRU DIMCA
Proof. Using Proposition 4.1 and [12, Proposition 1.1], we see that C has a 1dimensional symmetry, i.e. it admits a 1-dimensional algebraic subgroup H ′ of
P GL2 (C) as automorphism group. This group lifts uniquely to an algebraic 1parameter subgroup H of GL3 (C), not contained in the center and preserving not
only the curve C, but also its defining equation f . Such a subgroup H may be either
semi-simple or nilpotent, and [12, Proposition 3.1] tells us that only the semi-simple
case can occur for our situation. Indeed, as in the proof above, we know that
τ (C) = (d − 1)(d − 2) = d2 − 3d + 2.
Note that the statement of [12, Proposition 3.1] should be slightly corrected, namely
the part ” if y d has non-zero coefficient in f ” is to be replaced by ” if y d has zero
coefficient in f ”. Once we know that H is semi-simple, we can assume that H = C∗
acting on P2 via
t · [x : y : z] = [tw1 x : tw2 y : tw3 z],
for some integers wj . Then using the discussion and the notation on page 120 in
[12], we see that the only possibilities to get irreducible curves correspond to line
segments Aαk for 1 ≤ k < d/2 and k relatively prime to d.
Example 4.4. (i) Consider the following two rational cuspidal quartics
C : f = y 4 − xz 3 = 0 and C ′ : f ′ = y 4 − xz 3 − y 3z.
Then both curves have an E6 -singularity located at [1 : 0 : 0], hence they have only
weighted homogeneous singularities. However, as noticed in [14, Section 7.1], in the
final part on semi-binomial curves, C and C ′ are not projectively equivalent. This
follows there from the fact that C (resp. C ′ ) has one (resp. two) inflection points.
From our point of view, the difference between C and C ′ is that mdr(f ) = 1, while
mdr(f ′ ) = 2, as a direct computation shows. Note that both curves C and C ′ are
nearly free, since one has
τ (C) = τ (C ′ ) = 6 = τ (4, 1)max − 1 = τ (4, 2)max − 1.
Similar semi-binomial rational cuspidal curves C : f = 0 of arbitrary degree d,
with a unique Puiseux pair and τ (C) < µ(C), and with all the possible values for
r = mdr(f ) ∈ [2, d/2], are discussed in [8].
(ii) For d = 5, any rational cuspidal curve with 4 cusps is projectively equivalent to
C : f = 16x4 y + 128x2 y 2z − 4x3 z 2 + 256y 3z 2 − 144xyz 3 + 27z 5 = 0,
see for instance the discussion of the curve C8 = [(23 ), (2), (2), (2)] in [14, Section
6.3]. This curve has three A2 cusps and one A6 cusp, hence only weighted homogeneous singularities. A direct computation shows that mdr(f ) = 2 and that
τ (C) = τ (5, 2)max , and therefore C is a free curve.
It is surprizing that the condition d ≥ 6 in our results above avoids complicated
situations as (i) and (ii) in this example.
RATIONAL CUSPIDAL CURVES AND LOCAL COHOMOLOGY OF JACOBIAN RINGS
9
References
[1] T. Abe, A. Dimca, On the splitting types of bundles of logarithmic vector fields along plane
curves, arXiv:1706.05146 1, 2.1
[2] E. Artal Bartolo, L. Gorrochategui, I. Luengo, A. Melle-Hernández, On some conjectures
about free and nearly free divisors, in: Singularities and Computer Algebra, Festschrift for
Gert-Martin Greuel on the Occasion of his 70th Birthday, pp. 1–19, Springer (2017) 4.2
[3] A. Dimca, Hyperplane Arrangements: An Introduction, Universitext, Springer, 2017 1
[4] A. Dimca, Freeness versus maximal global Tjurina number for plane curves, Math. Proc.
Cambridge Phil. Soc. 163 (2017), 161–172. 3
[5] A. Dimca, D. Ibadula, A. Măcinic, Numerical invariants and moduli spaces for line arrangements, arXiv:1609.06551. 3.6
[6] A. Dimca, D. Popescu, Hilbert series and Lefschetz properties of dimension one almost complete
intersections, Comm. Algebra 44 (2016), 4467–4482. 1
[7] A. Dimca, E. Sernesi, Syzygies and logarithmic vector fields along plane curves, Journal de
l’École polytechnique-Mathématiques 1(2014), 247-267. 2, 2
[8] A. Dimca, G. Sticlaru, On the exponents of free and nearly free projective plane curves, Rev.
Mat. Complut. 30(2017), 259–268. 4.4
[9] A. Dimca, G. Sticlaru, Free divisors and rational cuspidal plane curves, arXiv:1504.01242, to
appear in MRL. 1
[10] A. Dimca, G. Sticlaru, Nearly free divisors and rational cuspidal curves, arXiv:1505.00666. 4
[11] A.A. du Plessis, C.T.C. Wall, Application of the theory of the discriminant to highly singular
plane curves, Math. Proc. Cambridge Phil. Soc., 126(1999), 259-266. 1, 3, 3
[12] A.A. du Plessis and C.T.C. Wall, Curves in P 2 (C) with 1-dimensional symmetry, Revista Mat
Complutense 12 (1999), 117–132. 1, 4
[13] J. Fernandez de Bobadilla, I. Luengo, A. Melle-Hernandez, A. Nemethi, Classification of rational unicuspidal projective curves whose singularities have one Puiseux pair, in: Real and
Complex Singularities Sao Carlos 2004, Trends in Mathematics, Birkhauser 2006, pp. 31-45. 1
[14] T. K. Moe, Rational Cuspidal Curves, arXiv:1511.02691 (139 pages, Master thesis 2008) 1, 4.4
[15] K. Saito, Quasihomogene isolierte Singularitäten von Hyperflächen, Invent. Math., 14 (1971),
123–142. 4
[16] F. Sakai, K. Tono, Rational cuspidal curves of type (d,d-2) with one or two cusps, Osaka J.
Math. 37(2000), 405-415. 1
[17] E. Sernesi, The local cohomology of the jacobian ring, Documenta Mathematica, 19 (2014),
541-565. 1, 2
[18] A. Simis, S.O. Tohăneanu, Homology of homogeneous divisors, Israel J. Math. 200 (2014),
449-487. 2
[19] D. van Straten, T. Warmt, Gorenstein duality for one-dimensional almost complete
intersections–with an application to non-isolated real singularities, Math. Proc.Cambridge Phil.
Soc.158(2015), 249–268. 1
Université Côte d’Azur, CNRS, LJAD, France
E-mail address: [email protected]
| 0math.AC
|
1
Compact Hash Code Learning with
Binary Deep Neural Network
arXiv:1712.02956v2 [cs.CV] 6 Feb 2018
Thanh-Toan Do, Dang-Khoa Le Tan, Tuan Hoang, Ngai-Man Cheung
Abstract—In this work, we firstly propose deep network models
and learning algorithms for learning binary hash codes given
image representations under both unsupervised and supervised
manners. Then, by leveraging the powerful capacity of convolutional neural networks, we propose an end-to-end architecture
which jointly learns to extract visual features and produce
binary hash codes. Our novel network designs constrain one
hidden layer to directly output the binary codes. This addresses
a challenging issue in some previous works: optimizing nonsmooth objective functions due to binarization. Additionally, we
incorporate independence and balance properties in the direct
and strict forms into the learning schemes. Furthermore, we also
include similarity preserving property in our objective functions.
Our resulting optimizations involving these binary, independence,
and balance constraints are difficult to solve. We propose to
attack them with alternating optimization and careful relaxation.
Experimental results on the benchmark datasets show that our
proposed methods compare favorably with the state of the art.
I. I NTRODUCTION
We are interested in learning binary hash codes for visual
search problems. Given an input image treated as a query, the
visual search systems search for visually similar images in a
database. In the state-of-the-art image retrieval systems [1],
[2], [3], [4], images are represented as high-dimensional feature vectors which later can be searched via classical distance
such as the Euclidean or Cosin distance. However, when
the database is scaled up, there are two main requirements
for retrieval systems, i.e., efficient storage and fast search.
Among solutions, binary hashing is an attractive approach for
achieving those requirements [5], [6], [7], [8] due to its fast
computation and efficient storage.
Briefly, in binary hashing problem, each original high
dimensional vector x ∈ RD is mapped into a very compact
binary vector b ∈ {−1, 1}L , where L D.
Many hashing methods have been proposed in the literature.
They can be divided into two categories: data-independence
methods and data-dependence methods. The former ones [9],
[10], [11], [12] rely on random projections to construct hash
functions. The representative methods in this category are
Locality Sensitive Hashing (LSH) [9] and its kernelized or
discriminative extensions [10], [11]. The latter ones use the
available training data to learn the hash functions in unsupervised [13], [14], [15], [16], [17] or (semi-)supervised [18],
[19], [20], [21], [22], [23], [24] manners. The prepresentative
Thanh-Toan Do is with The University of Adelaide, Australia. Dang-Khoa
Le Tan, Tuan Hoang, Ngai-Man Cheung are with Singapore University of
Technology and Design, Singapore.
Email: thanhtoan [email protected]; {letandang khoa, nguyenanhtuan hoang, ngaiman cheung}@sutd.edu.sg.
unsupervised hashing methods, e.g. Spectral Hashing [13],
Iterative Quantization (ITQ) [14], K-means Hashing [15],
Spherical Hashing [16], try to learn binary codes which
preserve the distance similarity between samples. The prepresentative supervised hashing methods, e.g., ITQ-CCA [14],
Binary Reconstructive Embedding [23], Kernel Supervised
Hashing [19], Two-step Hashing [22], Supervised Discrete
Hashing [24], try to learn binary codes which preserve the
label similarity between samples. The detailed review of
data-independent and data-dependent hashing methods can be
found in the recent surveys [5], [6], [7], [8].
One difficult problem in hashing is to deal with the binary
constraint on the codes. Specifically, the outputs of the hash
functions have to be binary. In general, this binary constraint
leads to an NP-hard mixed-integer optimization problem. To
handle this difficulty, most aforementioned methods relax the
constraint during the learning of hash functions. With this
relaxation, the continuous codes are learned first. Then, the
codes are binarized (e.g., by thresholding). This relaxation
greatly simplifies the original binary constrained problem.
However, the solution can be suboptimal, i.e., the binary codes
resulting from thresholded continuous codes could be inferior
to those that are obtained by directly including the binary
constraint in the learning.
Furthermore, a good hashing method should produce binary
codes with the following properties [13]: (i) similarity preserving, i.e., (dis)similar inputs should likely have (dis)similar
binary codes; (ii) independence, i.e., different bits in the binary
codes are independent to each other so that no redundant
information is captured; (iii) bit balance, i.e., each bit has a
50% chance of being 1 or −1. Note that direct incorporation
of the independent and balance properties can complicate
the learning. Previous works have used some relaxation or
approximation to overcome the difficulties [14], [20], but there
may be some performance degradation.
Recently, the deep learning has given great attention
to the computer vision community due to its superiority
in many vision tasks such as classification, detection,
segmentation [25], [26], [27]. Inspired by the success
of deep learning in different vision tasks, recently, some
researchers have used deep learning for joint learning image
representations and binary hash codes in an end-to-end deep
learning-based supervised hashing framework [28], [29],
[30], [31]. However, learning binary codes in deep networks
is challenging. This is because one has to deal with the
binary constraint on the hash codes, i.e., the final network
outputs must be binary. A naive solution is to adopt the sign
activation layer to produce binary codes. However, due to the
2
non-smoothness of the sign function, it causes the vanishing
gradient problem when training the network with the standard
back propagation [32].
Contributions: In this work, firstly, we propose a novel
deep network model and a learning algorithm for unsupervised hashing. In order to achieve binary codes, instead of
involving the sgn or step function as in recent works [33],
[34], our proposed network design constrains one layer to
directly output the binary codes (therefore, the network is
named as Binary Deep Neural Network). In addition, we
propose to directly incorporate the independence and balance
properties. Furthermore, we include the similarity preserving
in our objective function. The resulting optimization with these
binary and direct constraints is NP-hard. We propose to attack
this challenging problem with alternating optimization and
careful relaxation. Secondly, in order to enhance the discriminative power of the binary codes, we extend our method
to supervised hashing by leveraging the label information so
that the binary codes preserve the semantic similarity between
samples. Finally, to demonstrate the flexibility of our proposed
method and to leverage the powerful capacity of convolution
deep neural networks, we adapt our optimization strategy and
the proposed supervised hashing model to an end-to-end deep
hashing network framework. The solid experiments on various
benchmark datasets show the improvements of the proposed
methods over state-of-the-art hashing methods.
A preliminary version of this work has been reported
in [35]. In this work, we present substantial extension to our
previous work. In particular, the main extension is that we
propose the end-to-end binary deep neural network framework
which jointly learns the image features and the binary codes.
The experimental results show that the proposed end-to-end
hashing framework significantly boosts the retrieval accuracy.
Other minor extensions are we conduct more experiments
(e.g., new experiments on SUN397 dataset [36], comparison
to state-of-the-art end-to-end hashing methods) to evaluate the
effectiveness of the proposed methods.
The remaining of this paper is organized as follows.
Section II presents related works. Section III presents and
evaluates the proposed unsupervised hashing method. Section IV presents and evaluates the proposed supervised hashing
method. Section V presents and evaluates the proposed endto-end deep hashing network. Section VI concludes the paper.
II. R ELATED WORK
Our work is inspired by the recent successful hashing
methods which define hash functions as a neural network [37],
[33], [38], [34], [28], [29], [30]. We propose an improved
design to address their limitations. In Semantic Hashing [37],
the model is formed by a stack of Restricted Boltzmann
Machine, and a pretraining step is required. Additionally, this
model does not consider the independence and balance of
the codes. In Binary Autoencoder [34], a linear autoencoder
is used as hash functions. As this model only uses one
hidden layer, it may not well capture the information of
inputs. Extending [34] with multiple, nonlinear layers is not
straight-forward because of the binary constraint. They also
do not consider the independence and balance of codes. In
Deep Hashing [33], [38], a deep neural network is used as
a hash function. However, this model does not fully take
into account the similarity preserving. They also apply some
relaxation in arriving the independence and balance of codes
and this may degrade the performance. In the end-to-end deep
hashing [29], [30], the independence and balance of codes are
not considered.
In order to handle the binary constraint, Semantic Hashing [37] first solves the relaxed problem by discarding the
constraint and then thresholds the solved continuous solution.
In Deep Hashing (DH) [33], [38], the output of the last
layer, H(n) , is binarized by the sgn function. They include
a term in the objective function
to reduce this binarization
loss: sgn(H(n) ) − H(n) . Solving the objective function
of DH [33], [38] is difficult because the sgn function is
non-differentiable. The authors in [33], [38] overcome this
difficulty by assuming that the sgn function is differentiable
everywhere. In Binary Autoencoder (BA) [34], the outputs of
the hidden layer are passed into a step function to binarize the
codes. Incorporating the step function in the learning leads
to a non-smooth objective function, hence the optimization is
NP-complete. To handle this difficulty, the authors [34] use
binary SVMs to learn the model parameters in the case when
there is only a single hidden layer.
Joint learning image representations and binary hash codes
in an end-to-end deep learning-based supervised hashing
framework [28], [29], [30] has shown considerable boost in
retrieval accuracy. By joint optimization, the produced hash
codes are more sufficient to preserve the semantic similarity
between images. In those works, the network architectures
often consist of a feature extraction sub-network and a subsequent hashing layer to produce hash codes. Ideally, the
hashing layer should adopt a sign activation function to output
exactly binary codes. However, due to the vanishing gradient
difficulty of the sign function, an approximation procedure
must be employed. For example, sign can be approximated
by a tanh-like function y = tanh(βx), where β is a free
parameter controlling the trade off between the smoothness
and the binary quantization loss [29]. However, it is nontrivial to determine an optimal β. A small β causes large
binary quantization loss while a big β makes the output of
the function close to the binary values, but the gradient of the
function almost vanishes, making back-propagation infeasible.
The problem still remains when the logistic-like functions [28],
[30] are used.
III. U NSUPERVISED H ASHING WITH B INARY D EEP
N EURAL N ETWORK (UH-BDNN)
A. Formulation of UH-BDNN
For easy following, we first summarize the notations in
Table I. In our work, the hash functions are defined by a
deep neural network. In our proposed architecture, we use
different activation functions in different layers. Specifically,
we use the sigmoid function as the activation function for
layers 2, · · · , n − 2, and the identity function as the activation
3
codes. The second term is a regularization that tends to decrease the magnitude of the weights, and this helps to prevent
overfitting. Note that in our proposed design, we constrain the
network to directly output the binary codes at one layer, which
avoids the difficulty of the sgn / step function which is nondifferentiability. On the other hand, our formulation with (1)
under the binary constraint (2) is very difficult to solve. It is a
mixed-integer problem which is NP-hard. In order to attack the
problem, we propose to introduce an auxiliary variable B and
use alternating optimization. Consequently, we reformulate the
objective function (1) under constraint (2) as the following
TABLE I
N OTATIONS AND THEIR CORRESPONDING MEANINGS .
Notation
X
Meaning
D×m : set of m training samples;
X = {xi }m
i=1 ∈ R
each column of X corresponds to one sample
L×m : binary code of X
B = {bi }m
i=1 ∈ {−1, +1}
Number of required bits to encode a sample
Number of layers (including input and output layers)
Number of units in layer l
Activation function of layer l
W(l) ∈ Rsl+1 ×sl : weight matrix connecting layer l + 1
and layer l
c(l) ∈ Rsl+1 :bias vector for units in layer l + 1
H(l) = f (l) W(l−1) H(l−1) + c(l−1) 11×m :
output values of layer l; convention: H(1) = X
Matrix has a rows, b columns and all elements equal to 1
B
L
n
sl
f (l)
W(l)
c(l)
H(l)
1a×b
min J
W,c,B
=
1
X − W(n−1) B − c(n−1) 11×m
2m
+
Layer 1
(input layer)
Layer 2
Layer 𝑛 − 2
𝑾(1)
Layer 𝑛 − 1
Output 𝑯(𝑛−1) ∈
−1,1 and is used as
binary code
𝑾(𝑛−2)
Layer 𝑛
(reconstruction layer)
Output 𝐇 (n) ≈ 𝑿
𝑾(𝑛−1)
W,c
1
X − W(n−1) H(n−1) + c(n−1) 11×m
2m
+
n−1
λ1 X
W(l)
2 l=1
2
2
s.t. H(n−1) ∈ {−1, 1}L×m
(3)
(4)
B ∈ {−1, 1}L×m .
function for layer n − 1 and layer n. Our idea is to learn
the network such that the output values of the penultimate
layer (layer n − 1) can be used as the binary codes. We
introduce constraints in the learning algorithm such that the
output values at the layer n − 1 have the following desirable
properties: (i) belonging to {−1, 1}; (ii) similarity preserving;
(iii) independence and (iv) balancing. Fig. 1 illustrates our
network for the case D = 4, L = 2.
Let us start with first two properties of the codes, i.e.,
belonging to {−1, 1} and similarity preserving. To achieve
the binary codes having these two properties, we propose to
optimize the following constrained objective function
=
2
s.t. B = H(n−1) ,
Fig. 1. The illustration of our UH-BDNN (D = 4, L = 2). In our proposed
network design, the outputs of layer n − 1 are constrained to {−1, 1} and are
used as the binary codes. During training, these codes are used to reconstruct
the input samples at the final layer.
min J
n−1
λ1 X
W(l)
2 l=1
2
(1)
(2)
The constraint (2) is to ensure the first property. As the
activation function for the last layer is the identity function,
the term W(n−1) H(n−1) + c(n−1) 11×m is the output of
the last layer. The first term of (1) makes sure that the
binary code gives a good reconstruction of X. It is worth
noting that the reconstruction criterion has been used as an
indirect approach for preserving the similarity in state-ofthe-art unsupervised hashing methods [14], [34], [37], i.e.,
it encourages (dis)similar inputs map to (dis)similar binary
(5)
The benefit of introducing the auxiliary variable B is that
we can decompose the difficult constrained optimization problem (1) into two sub-optimization problems. Consequently, we
are able to iteratively solve the optimization problem by using
alternating optimization with respect to (W, c) and B while
holding the other fixed. Inspired from the quadratic penalty
method [39], we relax the equality constraint (4) by converting
it into a penalty term. We achieve the following constrained
objective function
min J
=
W,c,B
1
X − W(n−1) B − c(n−1) 11×m
2m
+
n−1
λ1 X
W(l)
2 l=1
2
+
2
λ2
H(n−1) − B
2m
2
(6)
s.t. B ∈ {−1, 1}L×m ,
(7)
in which, the third term in (6) measures the (equality) constraint violation. By setting the penalty parameter λ2 sufficiently large, we penalize the constraint violation severely,
thereby forcing the minimizer of the penalty function (6) closer
to the feasible region of the original constrained function (3).
We now consider the two remaining properties of the codes,
i.e., independence and balance. Unlike previous works which
use some relaxation or approximation on the independence and
balance properties [14], [33], [20], we propose to encode these
properties strictly and directly based on the binary outputs of
our layer n − 1. Specifically, we encode the independence
and balance properties of the codes by introducing the fourth
and the fifth terms respectively in the following constrained
objective function
min J
W,c,B
=
1
X − W(n−1) B − c(n−1) 11×m
2m
+
n−1
λ1 X
W(l)
2 l=1
+
λ3
2
2
+
λ2
H(n−1) − B
2m
1 (n−1) (n−1) T
H
(H
) −I
m
s.t. B ∈ {−1, 1}L×m .
2
+
2
2
λ4
H(n−1) 1m×1
2m
2
(8)
(9)
4
The objective function (8) under constraint (9) is our final
formulation. Before discussing how to solve it, let us present
the differences between our work and the recent deep learning
based hashing models Deep Hashing [33], Binary Autoencoder [34], and end-to-end hashing methods [28], [29], [30].
The main important difference between our model and other
deep learning-based hashing methods Deep Hashing [33],
Binary Autoencoder [34], end-to-end hashing [28], [29], [30]
is the way to achieve the binary codes. Instead of involving
the sgn, step function as in [33], [34] or the relaxation of sgn
as sigmoid [28], [30], tanh [29], we constrain the network to
directly output the binary codes at one layer. Other differences
to the most related works Deep Hashing [33] and Binary
Autoencoder [34] are presented as follows.
a) Comparison to Deep Hashing (DH) [33], [38]: the
deep model of DH is learned by the following objective
function
α1 (n) (n) T
tr H (H )
W,c
2m
n−1
n−1
2
2
2
α3 X
α2 X
W(l) (W(l) )T − I +
W(l) + c(l)
+
2 l=1
2 l=1
min J
=
1
sgn(H(n) ) − H(n)
2
2
−
The DH’s model does not have the reconstruction layer. The
authors apply the sgn function to the outputs at the top layer
of the network to obtain the binary codes. The first term aims
to minimize quantization loss by applying the sgn function to
the outputs at the top layer. The balancing and the independent
properties are presented in the second and the third terms. It
is worth noting that minimizing DH’s objective function is
difficult due to the non-differentiability of the sgn function.
The authors work around this difficulty by assuming that the
sgn function is differentiable everywhere.
Contrary to DH, we propose a different model design. In
particular, our model encourages the similarity preserving by
obtaining the reconstruction layer in the network.For the balancing property, they maximize tr H(n) (H(n) )T . According
to [20], maximizing this term is only an approximation in
arriving the balancing property. In our objective function,
the balancing property is directly enforced on the codes
2
by the term H(n−1) 1m×1 . For the independent property,
DH uses a relaxed orthogonality constraint on the network
2
weights W, i.e., W(l) (W(l) )T − I . On the contrary, we
(once again) directly constrain the code independence using
2
1
(n−1)
(H(n−1) )T − I . Incorporating the direct conmH
straints can lead to better performance.
b) Comparison to Binary Autoencoder (BA) [34]: the
differences between our model and BA are quite clear. Firstly,
BA, as described in [34], is a shallow linear autoencoder
network with one hidden layer. Secondly, the BA’s hash
function is a linear transformation of the input followed by the
step function to obtain the binary codes. In BA, by treating
the encoder layer as binary classifiers, they use binary SVMs
to learn the weights of the linear transformation. On the
contrary, our hash function is defined by multiple, hierarchical
layers of nonlinear and linear transformations. It is not clear
if the binary SVMs approach in BA can be used to learn the
weights in our deep architecture with multiple layers. Instead,
we use alternating optimization to derive a backpropagation
algorithm to learn the weights in all layers. Additionally,
another difference is that our model ensures the independence
and balance of the binary codes while BA does not. Note
that independence and balance properties may not be easily
incorporated in their framework, as these would complicate
their objective function and the optimization problem may
become very difficult to solve.
B. Optimization
In order to solve (8) under constraint (9), we propose to use
alternating optimization over (W, c) and B.
1) (W, c) step: When fixing B, the problem becomes the
unconstrained optimization. We use L-BFGS [40] optimizer
with backpropagation for solving. The gradients of the objective function J (8) w.r.t. different parameters are computed as
follows.
At l = n − 1, we have
−1
∂J
(X − W(n−1) B − c(n−1) 11×m )BT + λ1 W(n−1)
=
m
∂W(n−1)
(10)
∂J
−1
(n−1)
(n−1)
(X
−
W
B)1
−
mc
=
(11)
m×1
m
∂c(n−1)
For other layers, let us define
∆(l)
λ2 (n−1)
H
−B
m
2λ3
1 (n−1) (n−1) T
+
H
(H
) − I H(n−1)
m
m
0
λ4 (n−1)
H
1m×m
f (n−1) (Z(n−1) ) (12)
+
m
0
= (W(l) )T ∆(l+1)
f (l) (Z(l) ), ∀l = n − 2, · · · , 2
(13)
∆(n−1)
=
where Z(l) = W(l−1) H(l−1) + c(l−1) 11×m , l = 2, · · · , n;
denotes the Hadamard product.
Then, ∀l = n − 2, · · · , 1, we have
∂J
= ∆(l+1) (H(l) )T + λ1 W(l)
∂W(l)
(14)
∂J
= ∆(l+1) 1m×1
∂c(l)
(15)
2) B step: When fixing (W, c), we can rewrite problem (8)
as
min J
B
=
X − W(n−1) B − c(n−1) 11×m
+λ2 H(n−1) − B
2
2
s.t. B ∈ {−1, 1}L×m .
(16)
(17)
We adaptively use the recent method discrete cyclic coordinate
descent [24] to iteratively solve B, i.e., row by row. The
advantage of this method is that if we fix L − 1 rows of
B and only solve for the remaining row, we can achieve a
closed-form solution for that row.
Let V = X−c(n−1) 11×m ; Q = (W(n−1) )T V+λ2 H(n−1) .
For k = 1, · · · L, let wk be k th column of W(n−1) ; W1 be
matrix W(n−1) excluding wk ; qk be k th column of QT ; bTk
be k th row of B; B1 be matrix of B excluding bTk . We have
closed-form for bTk as
T
T
bT
k = sgn(q − wk W1 B1 ).
(18)
5
Algorithm 1 Unsupervised Hashing with Binary Deep Neural
Network (UH-BDNN)
Input:
D×m : training data; L: code length; T : maximum
X = {xi }m
i=1 ∈ R
iteration number; n: number of layers; {sl }n
l=2 : number of units of layers
2 → n (note: sn−1 = L, sn = D); λ1 , λ2 , λ3 , λ4 .
Output:
Parameters {W(l) , c(l) }n−1
l=1
1: Initialize B(0) ∈ {−1, 1}L×m using ITQ [14]
(l) n−2
2: Initialize {c(l) }n−1
l=1 = 0sl+1 ×1 . Initialize {W }l=1 by getting the
top sl+1 eigenvectors from the covariance matrix of H(l) . Initialize
W(n−1) = ID×L
3: Fix B(0) , compute (W, c)(0) with (W, c) step using initialized
{W(l) , c(l) }n−1
l=1 (line 2) as starting point for L-BFGS.
4: for t = 1 → T do
5:
Fix (W, c)(t−1) , compute B(t) with B step
6:
Fix B(t) , compute (W, c)(t) with (W, c) step using (W, c)(t−1)
as starting point for L-BFGS.
7: end for
8: Return (W, c)(T )
The proposed UH-BDNN method is summarized in Algorithm 1. In the Algorithm 1, B(t) and (W, c)(t) are values
of B and {W(l) , c(l) }n−1
l=1 at iteration t, respectively.
C. Evaluation of Unsupervised Hashing with Binary Deep
Neural Network (UH-BDNN)
This section evaluates the proposed UH-BDNN and compares it with the following state-of-the-art unsupervised hashing methods: Spectral Hashing (SH) [13], Iterative Quantization (ITQ) [14], Binary Autoencoder (BA) [34], Spherical
Hashing (SPH) [16], K-means Hashing (KMH) [15]. For
all compared methods, we use the implementations and the
suggested parameters provided by the authors.
1) Dataset, evaluation protocol, and implementation notes:
a) Dataset: CIFAR10 [41] dataset consists of 60,000
images of 10 classes. The training set (also used as the
database for the retrieval) contains 50,000 images. The query
set contains 10,000 images. Each image is represented by a
800-dimensional feature vector extracted by PCA from 4096dimensional CNN feature produced by the AlexNet [25].
MNIST [42] dataset consists of 70,000 handwritten digit
images of 10 classes. The training set (also used as database
for retrieval) contains 60,000 images. The query set contains
10,000 images. Each image is represented by a 784 dimensional gray-scale feature vector by using its intensity.
SIFT1M [43] dataset contains 128 dimensional SIFT vectors [44]. There are 1M vectors used as database for retrieval;
100K vectors for training (separated from retrieval database)
and 10K vectors for query.
b) Evaluation protocol: We follow the standard setting in
unsupervised hashing methods [14], [16], [15], [34] which use
Euclidean nearest neighbors as the ground truths for queries.
The number of ground truths are set as in [34], i.e., for
CIFAR10 and MNIST datasets, for each query, we use 50 its
Euclidean nearest neighbors as the ground truths; for large
scale dataset SIFT1M, for each query, we use 10, 000 its
Euclidean nearest neighbors as the ground truths. We use the
following evaluation metrics which have been used in the
state of the art [14], [34], [33] to measure the performance
of methods. 1) mean Average Precision (mAP); 2) precision
of Hamming radius 2 (precision@2) which measures precision
on retrieved images having Hamming distance to query ≤ 2
(if no images satisfy, we report zero precision). Note that as
computing mAP is slow on large dataset SIFT1M, we consider
top 10, 000 returned neighbors when computing mAP.
c) Implementation notes: In our deep model, we use n =
5 layers. The parameters λ1 , λ2 , λ3 and λ4 are empirically
set by cross validation as 10−5 , 5 × 10−2 , 10−2 and 10−6 ,
respectively. The max iteration number T is empirically set to
10. The number of units in hidden layers 2, 3, 4 are empirically
set as [90 → 20 → 8], [90 → 30 → 16], [100 → 40 → 24]
and [120 → 50 → 32] for the code length 8, 16, 24 and 32
bits, respectively.
2) Retrieval results: Fig. 2 and Table II show comparative
mAP and precision of Hamming radius 2 (precision@2) of
methods, respectively.
We find the following observations are consistent for all
three datasets. In term of mAP, the proposed UH-BDNN
is comparable to or outperforms other methods at all code
lengths. The improvement is more clear at high code length,
i.e., L = 24, 32. The best competitor to UH-BDNN is
binary autoencoder (BA) [34] which is the current stateof-the-art unsupervised hashing method. Compare to BA,
at high code length, UH-BDNN consistently outperforms
BA on all datasets. In term of precision@2, UH-BDNN
is comparable to other methods at low code lengths, i.e.,
L = 8, 16. At L = 24, 32, UH-BDNN significantly achieve
better performance than other methods. Specifically, the
improvements of UH-BDNN over the best competitor
BA [34] are clearly observed at L = 32 on the MNIST and
SIFT1M datasets.
Comparison with Deep Hashing (DH) [33], [38] As
the implementation of DH is not available, we set up the
experiments on CIFAR10 and MNIST similar to [33] to make
a fair comparison. For each dataset, we randomly sample 1,000
images (i.e., 100 images per class) as query set; the remaining
images are used as training and database set. Follow [33],
for CIFAR10, each image is represented by 512-D GIST
descriptor [45].
Follow Deep Hashing [33], the ground truths of queries
are based on their class labels1 . Similar to [33], we report
comparative results in term of mAP and the precision of
Hamming radius r = 2. The results are presented in the
Table III. It is clearly showed that the proposed UH-BDNN
outperforms DH [33] at all code lengths, in both mAP and
precision of Hamming radius.
IV. S UPERVISED H ASHING WITH B INARY D EEP N EURAL
N ETWORK (SH-BDNN)
In order to enhance the discriminative power of the binary codes, we extend UH-BDNN to supervised hashing by
1 It is worth noting that in the evaluation of unsupervised hashing, instead
of using class label as ground truths, most state-of-the-art methods [14], [16],
[15], [34] use Euclidean nearest neighbors as ground truths for queries.
6
mAP
8
25
UH−BDNN
BA
ITQ
SH
SPH
KMH
20
mAP
10
6
15
25
UH−BDNN
BA
ITQ
SH
SPH
KMH
20
mAP
12
10
15
UH−BDNN
BA
ITQ
SH
SPH
KMH
10
4
5
2
0
8
16
24
0
8
32
5
16
24
0
8
32
16
24
32
number bits (L)
number bits (L)
number bits (L)
(a) CIFAR10
(b) MNIST
(c) SIFT1M
Fig. 2. mAP comparison between UH-BDNN and state-of-the-art unsupervised hashing methods on CIFAR10, MNIST, and SIFT1M.
TABLE II
P RECISION AT H AMMING DISTANCE r = 2 COMPARISON BETWEEN UH-BDNN AND STATE - OF - THE - ART UNSUPERVISED HASHING METHODS ON
CIFAR10, MNIST, AND SIFT1M.
L
UH-BDNN
BA[34]
ITQ[14]
SH[13]
SPH[16]
KMH[15]
8
0.55
0.55
0.54
0.39
0.43
0.53
CIFAR10
16
24
5.79
22.14
5.65
20.23
5.05
18.82
4.23
14.60
3.45
13.47
5.49
19.55
32
18.35
17.00
17.76
15.22
13.67
15.90
8
0.53
0.51
0.51
0.43
0.44
0.50
MNIST
16
24
6.80
29.38
6.44
27.65
5.87
23.92
6.50
27.08
5.02
22.24
6.36
25.68
32
38.50
35.29
36.35
36.69
30.80
36.24
8
4.80
3.85
3.19
4.67
4.25
3.74
SIFT1M
16
24
25.20
62.20
23.19
61.35
14.07
35.80
24.82
60.25
20.98
47.09
20.74
48.86
32
80.55
77.15
58.69
72.40
66.42
76.04
TABLE III
C OMPARISON WITH D EEP H ASHING (DH) [33].
CIFAR10
precision@2
32
16
32
16.62
23.33
15.77
18.52
24.97
18.85
mAP
L
DH [33]
UH-BDNN
16
16.17
17.83
leveraging the label information. There are several approaches
proposed to leverage the label information, leading to different
criteria on binary codes. In [18], [46], binary codes are
learned such that they minimize Hamming distances between
samples belonging to the same class, while maximizing the
Hamming distance between samples belonging to different
classes. In [24], the binary codes are learned such that they
are optimal for linear classification.
In this work, in order to exploit the label information,
we follow the approach proposed in Kernel-based Supervised
Hashing (KSH) [19]. The benefit of this approach is that it
directly encourages the Hamming distances between binary
codes of within-class samples equal to 0, and the Hamming
distances between binary codes of between-class samples
equal to L. In the other words, it tries to perfectly preserve
the semantic similarity. To achieve this goal, it enforces the
Hamming distances between learned binary codes to be highly
correlated with the pre-computed pairwise label matrix.
In general, the network structure of SH-BDNN is similar
to UH-BDNN, excepting that the last layer preserving the
reconstruction of UH-BDNN is removed. The layer n − 1 in
UH-BDNN becomes the last layer in SH-BDNN. All desirable
properties, i.e., semantic similarity preserving, independence,
and balance, in SH-BDNN are constrained on the outputs of
its last layer.
MNIST
precision@2
32
16
32
44.97
66.10
73.29
47.21
69.13
75.26
mAP
16
43.14
45.38
A. Formulation of SH-BDNN
We define the pairwise label matrix S as
1
−1
Sij =
if xi and xj are same class
if xi and xj are not same class
(19)
In order to achieve the semantic similarity preserving property, we learn the binary codes such that the Hamming
distance between learned binary codes highly correlates
with the matrix S, i.e., we want to minimize the quantity
2
1
(n) T
) H(n) − S . In addition, to achieve the indepenL (H
dence and balance properties of codes, we want to minimize
2
2
1
the quantities m
H(n) (H(n) )T − I and H(n) 1m×1 , respectively.
Follow the same reformulation and relaxation as UH-BDNN
(Sec. III-A), we solve the following constrained optimization
which ensures the binary constraint, the semantic similarity
preserving, the independence, and the balance properties of
codes.
min J
W,c,B
=
1
2m
1
(H(n) )T H(n) − S
L
2
λ2
λ3
H(n) − B +
2m
2
2
λ4
(n)
+
H 1m×1
2m
+
s.t. B ∈ {−1, 1}L×m
2
+
n−1
λ1 X
W(l)
2 l=1
2
1 (n) (n) T
H (H ) − I
m
2
(20)
(21)
7
Algorithm 2 Supervised Hashing with Binary Deep Neural
Network (SH-BDNN)
Input:
D×m : labeled training data; L: code length; T :
X = {xi }m
i=1 ∈ R
maximum iteration number; n: number of layers; {sl }n
l=2 : number of
units of layers 2 → n (note: sn = L); λ1 , λ2 , λ3 , λ4 .
Output:
Parameters {W(l) , c(l) }n−1
l=1
It is easy to see that the optimal solution for (26) under
constraint (27) is B = sgn(H(n) ).
The proposed SH-BDNN method is summarized in Algorithm 2. In Algorithm 2, B(t) and (W, c)(t) are values of B
and {W(l) , c(l) }n−1
l=1 at iteration t, respectively.
1: Compute pairwise label matrix S using (19).
2: Initialize B(0) ∈ {−1, 1}L×m using ITQ [14]
(l) n−1
3: Initialize {c(l) }n−1
l=1 = 0sl+1 ×1 . Initialize {W }l=1 by getting the
top sl+1 eigenvectors from the covariance matrix of H(l) .
4: Fix B(0) , compute (W, c)(0) with (W, c) step using initialized
{W(l) , c(l) }n−1
l=1 (line 3) as starting point for L-BFGS.
5: for t = 1 → T do
6:
Fix (W, c)(t−1) , compute B(t) with B step
7:
Fix B(t) , compute (W, c)(t) with (W, c) step using (W, c)(t−1)
as starting point for L-BFGS.
8: end for
9: Return (W, c)(T )
C. Evaluation of Supervised Hashing with Binary Deep Neural Network
(20) under constraint (21) is our formulation for supervised
hashing. The main difference in formulation between UHBDNN (8) and SH-BDNN (20) is that the reconstruction
term preserving the neighbor similarity in UH-BDNN (8) is
replaced by the term preserving the label similarity in SHBDNN (20).
a) Dataset: We evaluate and compare methods on
CIFAR-10 and MNIST, and SUN397 datasets. The descriptions of the first two datasets are presented in section III-C1.
SUN397 [36] dataset contains about 108K images from 397
scene categories. We use a subset of this dataset including
42 largest categories in which each category contains more
than 500 images. This results about 35K images in total.
The query set contains 4,200 images (100 images per class)
randomly sampled from the dataset. The rest images are used
as the training set and also the database set. Each image is
represented by a 800-dimensional feature vector extracted by
PCA from 4096-dimensional CNN feature produced by the
AlexNet [25].
b) Evaluation protocol: Follow the literature [24], [14],
[19], we report the retrieval results in two metrics: 1) mean
Average Precision (mAP) and 2) precision of Hamming radius
2 (precision@2).
c) Implementation notes: The network configuration is
same as UH-BDNN except the final layer is removed. The
values of parameters λ1 , λ2 , λ3 , and λ4 are empirically set
using cross validation as 10−3 , 5, 1, and 10−4 , respectively.
The max iteration number T is empirically set to 5.
Follow the settings in ITQ-CCA [14], SDH [24], all training
samples are used in the learning for these two methods.
For SH-BDNN, KSH [19] and BRE [23] where the label
information is leveraged by the pairwise label matrix, we
randomly select 3, 000 training samples from each class and
use them for learning. The ground truths of queries are defined
by the class labels from the datasets.
2) Retrieval results: Fig. 3 and Table IV show comparative
results between the proposed SH-BDNN and other supervised
hashing methods on CIFAR10, MNIST, and SUN397 datasets.
On the CIFAR10 dataset, Fig. 3(a) and Table IV clearly
show that the proposed SH-BDNN outperforms all compared
methods by a fair margin at all code lengths in both mAP and
precision@2. The best competitor to SH-BDNN on this dataset
is CCA-ITQ [14]. The more improvements of SH-BDNN over
CCA-ITQ are observed at high code lengths, i.e., SH-BDNN
outperforms CCA-ITQ about 4% at L = 24 and L = 32.
B. Optimization
In order to solve (20) under constraint (21), we use alternating optimization, which comprises two steps over (W, c)
and B.
1) (W, c) step: When fixing B, (20) becomes unconstrained optimization. We used L-BFGS [40] optimizer with
backpropagation for solving. The gradients of objective function J (20) w.r.t. different parameters are computed as follows.
Let us define
∆(n)
where V =
∆(l)
=
λ
1
2
H(n) V + VT +
H(n) − B
mL
m
2λ3
1 (n) (n) T
H (H ) − I H(n)
+
m
m
0
λ4 (n)
+
H 1m×m
f (n) (Z(n) )
m
1
(H(n) )T H(n)
L
(22)
− S.
= (W(l) )T ∆(l+1)
0
f (l) (Z(l) ), ∀l = n − 1, · · · , 2
(23)
where Z(l) = W(l−1) H(l−1) + c(l−1) 11×m , l = 2, · · · , n;
denotes the Hadamard product.
Then ∀l = n − 1, · · · , 1, we have
∂J
= ∆(l+1) (H(l) )T + λ1 W(l)
∂W(l)
(24)
∂J
= ∆(l+1) 1m×1
∂c(l)
(25)
2) B step: When fixing (W, c), we can rewrite problem (20) as
min J = H(n) − B
2
B
s.t. B ∈ {−1, 1}L×m
(26)
(27)
This section evaluates our proposed SH-BDNN and compares it to the following state-of-the-art supervised hashing
methods including Supervised Discrete Hashing (SDH) [24],
ITQ-CCA [14], Kernel-based Supervised Hashing (KSH) [19],
Binary Reconstructive Embedding (BRE) [23]. For all compared methods, we use the implementations and the suggested
parameters provided by the authors.
1) Dataset, evaluation protocol, and implementation notes:
8
SH−BDNN
SDH
ITQ−CCA
KSH
BRE
50
40
mAP
mAP
60
100
70
90
60
80
50
mAP
70
70
20
8
SH−BDNN
SDH
ITQ−CCA
KSH
BRE
50
16
24
32
40
8
number bits (L)
40
30
60
30
SH-BDNN
SDH
ITQ-CCA
KSH
BRE
16
24
20
10
8
32
(a) CIFAR10
16
24
32
number bits (L)
number bits (L)
(b) MNIST
(c) SUN397
Fig. 3. mAP comparison between SH-BDNN and state-of-the-art supervised hashing methods on CIFAR10, MNIST and SUN397 datasets.
TABLE IV
P RECISION AT H AMMING DISTANCE r = 2 COMPARISON BETWEEN SH-BDNN AND STATE - OF - THE - ART SUPERVISED HASHING METHODS ON CIFAR10,
MNIST, AND SUN397 DATASETS .
L
SH-BDNN
SDH[24]
ITQ-CCA[14]
KSH[19]
BRE[23]
8
54.12
31.60
49.14
44.81
23.84
CIFAR10
16
24
67.32
69.36
62.23
67.65
65.68
67.47
64.08
67.01
41.11
47.98
32
69.62
67.63
67.19
65.76
44.89
8
84.26
36.49
54.35
68.07
37.67
DR Layer
Binary Optimizer
AlexNet
Fig. 4. The illustration of the end-to-end hashing framework (E2E-BDNN).
On the MNIST dataset, Fig. 3(b) and Table IV show that the
proposed SH-BDNN significant outperforms the current stateof-the-art SDH [24] at low code length, i.e., L = 8. When
L increases, SH-BDNN and SDH [24] achieve similar performance. SH-BDNN significantly outperforms other methods,
i.e., KSH [19], ITQ-CCA [14], BRE [23], on both mAP and
precision@2.
On the SUN397 dataset, the proposed SH-BDNN outperforms other competitors at all code lengths in terms of both
mAP and precision@2. The best competitor to SH-BDNN
on this dataset is SDH [24]. At high code lengths (e.g.,
L = 24, 32), SH-BDNN achieves more improvements over
SDH.
V. S UPERVISED H ASHING WITH E ND - TO -E ND B INARY
D EEP N EURAL N ETWORK (E2E-BDNN)
Even though the proposed SH-BDNN can significantly
enhance the discriminative power of the binary codes, similar
to other hashing methods, its capability is partially depended
on the discriminative power of image features. The recent endto-end deep learning-based supervised hashing methods [28],
MNIST
16
24
94.67
94.69
93.00
93.98
79.99
84.12
90.79
92.86
69.80
83.24
32
95.51
94.43
84.57
92.41
84.61
8
15.52
13.89
13.22
12.64
9.26
SUN397
16
24
41.98
52.53
40.39
49.54
37.53
50.07
40.67
49.29
26.95
38.36
32
56.82
53.25
53.12
46.45
40.36
[29], [30] have shown that joint learning image representations
and binary hash codes in an end-to-end fashion can boost the
retrieval accuracy. Therefore, in this section, we propose to
extend the proposed SH-BDNN to an end-to-end framework.
Specifically, we integrate the convolutional neural network
(CNN) with our supervised hashing network (SH-BDNN) into
a unified end-to-end deep architecture, namely End-to-End
Binary Deep Neural Network (E2E-BDNN), which can jointly
learn visual features and binary representations of images.
In the following, we first introduce our proposed network
architecture. We then describe the training process. Finally,
we present experiments on various benchmark datasets.
A. Network architecture
Fig. 4 illustrates the overall architecture of the end-toend binary deep neural network – E2E-BDNN. In details,
the network consists of three main components: (i) a feature
extractor, (ii) a dimensional reduction layer, and (iii) a binary
optimizer component. We utilize AlexNet [25] as the feature
extractor component of the E2E-BDNN. In our configuration,
we remove the last layer of AlexNet, namely the softmax
layer, and consider its last fully connected layer (fc7) as the
image representation.
The dimensional reduction component (the DR layer) involves a fully connected layer for reducing the high dimensional image representations outputted by the feature extractor
component into lower dimensional representations. We use the
identity function as the activation function for this DR layer.
The reduced representations are then used as inputs for the
following binary optimizer component.
The binary optimizer component of E2E-BDNN is similar
to SH-BDNN. It means that we also constrain the output
codes of E2E-BDNN to be binary. These codes also have
desired properties such as semantic similarity preserving,
9
Algorithm 3 End-to-End Binary Deep Neural Network (E2EBDNN) Learning
Input:
X = {xi }n
i=1 : labeled training images; m: minibatch size; L: code
length; K, T : maximum iteration. λ1 , λ2 , λ3 , λ4 : hyperparameters.
Output:
Network parameters W
(0)
1: Initialize the network W(0)
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
Initialize B(0) ∈ {−1, 1}L×n via ITQ [14]
for k = 1 → K do
for t = 1 → T do
A minibatch X(t) is sampled from X
Compute the corresponding similarity matrix S(t)
From B(k−1) , sample B(t) corresponding to X(t)
(k)
Fix B(t) , optimize W(t) via SGD
end for
(k)
Update B(k) by W(T )
end for
(K)
Return W(T )
WP CA = [v1T v2T · · · vkT ]
cP CA = −WP CA m
independence, and balance. By using the same design as SHBDNN for the last component of E2E-BDNN, it allows us
to observe the advantages of the end-to-end architecture over
SH-BDNN.
The training data for the E2E-BDNN is labelled images,
which contrasts to SH-BDNN which uses visual features such
as GIST [45], SIFT [44] or deep features from convolutional
deep networks. Given the input labeled images, we aim to
learn binary codes with aforementioned desired properties, i.e.,
semantic similarity preserving, independence, and balance.
In order to achieve these properties on the codes, we use
the similar objective function as SH-BDNN. However, it is
important to mention that in SH-BDNN, by its non end-to-end
architecture, we can feed the whole training set to the network
at a time for training. This does not hold for E2E-BDNN.
Due to the memory consuming of the end-to-end architecture,
we can only feed a minibatch of images to the network at
a time for training. Technically, let H be the output of the
last fully connected layer of E2E-BDNN for a minibatch of
size m; S be similarity matrix defined over the minibatch
(using equation (19)); B is an auxiliary variable. Similar to
SH-BDNN, we train the network to minimize the following
constrained loss function
min J
W,B
=
λ1
2m
1 T
H H−S
L
2
λ2
λ3
kH − Bk2 +
2m
2
λ4
2
kH1m×1 k
+
2m
+
L×m
s.t. B ∈ {−1, 1}
1
HHT − I
m
At beginning (line 1 in the Algorithm 3), we initialize
(0)
the network parameter W(0) as follows. For the feature
extractor component, we initialize it by the pretrained AlexNet
model [25]. The dimensional reduction (DR) layer is initialized
by applying PCA on the AlexNet features, i.e. the outputs of
fc7 layer, of the training set. Specifically, let WP CA and
cP CA be the weights and bias of the DR layer, respectively,
we initialize them as follows,
where v1 , v2 , . . . , vk are the top eigenvectors corresponding
to the largest eigenvalues extracted from the covariance matrix
of training features; m is the mean of the training features.
The binary optimizer component is initialized by training
SH-BDNN using AlexNet features (of training set) as inputs.
We then initialize the binary code matrix of the whole dataset
B(0) ∈ {−1, 1}L×n via ITQ [14] (line 2 in the Algorithm 3).
Here, AlexNet features are used as training inputs for ITQ.
In each iteration t of the Algorithm 3, we only sample a
minibatch X(t) from the training set to feed to the network
(line 5 in the Algorithm 3). So after T iterations, we exhaustively sample all training data. In each iteration t, we
first create the similarity matrix S(t) (using equation (19))
corresponding to X(t) as well as the B(t) matrix (line 6
and line 7 in the Algorithm 3). Since B(t) has been already
computed, we can fix that variable and update the network
(k)
parameter W(t) by standard backpropagation with Stochastic
Gradient Descent (SGD) (line 8 in the Algorithm 3). After
T iterations, since the network was exhaustively learned from
the whole training set, we update B(k) = sgn(F) (line 10
in the Algorithm 3), where F is the outputs of the last fully
connected layer for all training samples. We then repeat the
optimization procedure until it reaches a criterion, i.e., after
K iterations.
Implementation details The proposed E2E-BDNN is implemented in MATLAB with MatConvNet library [47]. All
experiments are conducted on a workstation machine with a
GPU Titan X. Regarding the hyperparameters of the loss function, we empirically set λ1 = 10−1 , λ2 = 10−2 , λ3 = 10−2
and λ4 = 10−3 . The learning rate is set to 10−4 and the weight
decay is set to 5 × 10−4 . The minibatch size is 256.
2
C. Evaluation of End-to-End Binary Deep Neural Network
(E2E-BDNN)
(28)
(29)
B. Training
The training procedure of E2E-BDNN is presented in the
Algorithm 3. In the Algorithm 3, X(t) and B(t) ∈ {−1, 1}L×m
are a minibatch sampled from the training set at iteration t and
its corresponding binary codes, respectively. B(k) is the binary
(k)
codes of the whole training set X at iteration k; W(t) is the
network weight when learning up to iterations t and k.
Since we have already compare SH-DBNN to other supervised hashing methods in Section IV-C, in this experiment
we focus to compare E2E-BDNN with SH-BDNN. We also
compare the proposed E2E-BDNN to other end-to-end hashing
methods [29], [28], [30], [31].
a) Comparison between SH-BDNN and E2E-BDNN:
Table V presents comparative mAP between SH-BDNN and
E2E-BDNN. The results shows that E2E-BDNN consistently
improves over SH-BDNN at all code lengths on all datasets.
The large improvements of E2E-BDNN over SH-BDNN are
observed on the CIFAR10 and MNIST datasets, especially at
10
M AP COMPARISON BETWEEN
L
SH-BDNN
E2E-BDNN
8
57.15
64.83
TABLE V
SH-BDNN AND E2E-BDNN ON CIFAR10, MNIST, AND SUN397 DATASETS .
CIFAR10
16
24
66.04
68.81
71.02
72.37
32
69.66
73.56
8
84.65
88.82
low code lengths, i.e., on CIFAR10, E2E-BDNN outperforms
SH-BDNN 7.7% and 5% at L = 8 and L = 16, respectively;
on MNIST, E2E-BDNN outperforms SH-BDNN 4.2% and
3.8% at L = 8 and L = 16, respectively. On SUN397 dataset,
the improvements of E2E-BDNN over SH-BDNN are more
clear at high code length, i.e., E2E-BDNN outperforms SHBDNN 2.5% and 2.7% at L = 24 and L = 32, respectively.
The improvements of E2E-BDNN over SH-BDNN confirm
the effectiveness of the proposed end-to-end architecture for
learning discriminative binary codes.
b) Comparison between E2E-BDNN and other end-toend supervised hashing methods: We also compare our proposed deep networks SH-BDNN and E2E-BDNN with other
end-to-end supervised hashing architectures, i.e., Hashing with
Deep Neural Network (DNNH) [30], Deep Hashing Network
(DHN) [48], Deep Quantization Network (DQN) [31], Deep
Semantic Ranking Hashing (DSRH) [28], and Deep Regularized Similarity Comparison Hashing (DRSCH) [29]. In
those works, the authors propose the frameworks in which
the image features and hash codes are simultaneously learned
by combining CNN layers and a binary quantization layer
into a large model. However, their binary mapping layer only
applies a simple operation, e.g., an approximation of sgn
function (logistic [28], [30], tanh [29]), l1 norm approximation of binary constraints [48]. Our SH-BDNN and E2EBDNN advances over those works in the way to map the
image features to the binary codes. Furthermore, our learned
codes ensure good properties, i.e. independence and balance,
while [30], [29], [31] do not consider such properties, and [28]
only considers the balance of codes. It is worth noting that,
in [30], [28], [29], [48], different settings are used when
evaluating. For fair comparison, follow those works, we setup
two different experimental settings as follows
•
•
Setting 1: according to [30], [31], [48], we randomly
sample 100 images per class to form 1K testing images.
The remaining 59K images are used as database images.
Furthermore, 500 images per class are sampled from the
database to form 5K training images.
Setting 2: following [28], [29], we randomly sample
1K images per class to form 10K testing images. The
remaining 50K images are served as the training set. In
the test phase, each test image is searched through the
test set by the leave-one-out procedure.
Table VI shows the comparative mAP between our methods
and DNNH [30], DHN [48], DQN [31] on the CIFAR10
dataset with the setting 1. Interestingly, even with non end-toend approach, our SH-BDNN outperforms DNNH and DQN
at all code lengths. This confirms the effectiveness of the
proposed approach for dealing with binary constrains and the
having of desire properties such as independence and balance
MNIST
16
24
94.24
94.80
98.03
98.26
32
95.25
98.21
8
33.06
34.15
SUN397
16
24
47.13
57.02
48.21
59.51
32
61.89
64.58
TABLE VI
M AP COMPARISON BETWEEN E2E-BDNN, SH-BDNN AND
DQN[31] ON CIFAR10 (S ETTING 1).
L
SH-E2E
SH-BDNN
DNNH[30]
DQN[31]
DHN[48]
24
60.02
57.30
56.60
55.80
59.40
32
61.35
58.66
55.80
56.40
60.30
DNNH[30],
48
63.59
60.08
58.10
58.00
62.10
TABLE VII
M AP COMPARISON BETWEEN E2E-BDNN, SH-BDNN AND END - TO - END
HASHING METHODS DSRH[28], DRSCH[29] ON CIFAR10 (S ETTING 2).
L
SH-E2E
SH-BDNN
DSRH [28]
DRSCH [29]
24
67.16
65.21
61.08
62.19
32
68.72
66.22
61.74
62.87
48
69.23
66.53
61.77
63.05
on the produced codes. With the end-to-end approach, it helps
to boost the performance of the SH-BDNN. The proposed
E2E-BDNN outperforms all compared methods, DNNH [30],
DHN [48], DQN [31]. It is worth noting that in Lai et
al. [30], increasing the code length does not necessary boost
the retrieval accuracy, i.e., they report a mAP 55.80 at L = 32,
while a higher mAP, i.e., 56.60 is reported at L = 24. Different
from [30], both SH-BDNN and E2E-BDNN improve mAP
when increasing the code length.
Table VII presents the comparative mAP between the
proposed SH-BDNN, E2E-BDNN and the competitors
DSRH [28], DRSCH [29] on the CIFAR10 dataset with the setting 2. The results clearly show that the proposed E2E-BDNN
significantly outperforms the DSRH [28] and DRSCH [29] at
all code lengths. Compare to the best competitor DRSCH [29],
the improvements of E2E-BDNN over DRSCH are from 5%
to 6% at different code lengths. Furthermore, we can see that
even with non end-to-end approach, the proposed SH-BDNN
also outperforms DSRH [28] and DRSCH [29].
VI. C ONCLUSION
In this paper, we propose three deep hashing neural networks for learning compact binary presentations. Firstly, we
introduce UH-BDNN and SH-BDNN for unsupervised and
supervised hashing respectively. Our network designs constrain to directly produce binary codes at one layer. The
designs also ensure good properties for codes: similarity
preserving, independence and balance. Together with designs,
we also propose alternating optimization schemes which allow
to effectively deal with binary constraints on the codes. We
then propose to extend the SH-BDNN to an end-to-end binary
deep neural network (E2E-BDNN) framework which jointly
11
learns the image representations and the binary codes. The
solid experimental results on various benchmark datasets show
that the proposed methods compare favorably or outperform
state-of-the-art hashing methods.
R EFERENCES
[1] R. Arandjelovic and A. Zisserman, “All about VLAD,” in CVPR, 2013.
[2] H. Jégou, M. Douze, C. Schmid, and P. Pérez, “Aggregating local
descriptors into a compact image representation,” in CVPR, 2010.
[3] T.-T. Do and N.-M. Cheung, “Embedding based on function approximation for large scale image search,” TPAMI, 2017.
[4] H. Jégou and A. Zisserman, “Triangulation embedding and democratic
aggregation for image search,” in CVPR, 2014.
[5] K. Grauman and R. Fergus, “Learning binary hash codes for large-scale
image search,” Machine Learning for Computer Vision, 2013.
[6] J. Wang, H. T. Shen, J. Song, and J. Ji, “Hashing for similarity search:
A survey,” CoRR, 2014.
[7] J. Wang, T. Zhang, J. Song, N. Sebe, and H. T. Shen, “A survey on
learning to hash,” TPAMI, 2017.
[8] J. Wang, W. Liu, S. Kumar, and S. Chang, “Learning to hash for indexing
big data - A survey,” Proceedings of the IEEE, pp. 34–57, 2016.
[9] A. Gionis, P. Indyk, and R. Motwani, “Similarity search in high
dimensions via hashing,” in VLDB, 1999.
[10] B. Kulis and K. Grauman, “Kernelized locality-sensitive hashing for
scalable image search,” in ICCV, 2009.
[11] M. Raginsky and S. Lazebnik, “Locality-sensitive binary codes from
shift-invariant kernels,” in NIPS, 2009.
[12] B. Kulis, P. Jain, and K. Grauman, “Fast similarity search for learned
metrics,” TPAMI, pp. 2143–2157, 2009.
[13] Y. Weiss, A. Torralba, and R. Fergus, “Spectral hashing,” in NIPS, 2008.
[14] Y. Gong, S. Lazebnik, A. Gordo, and F. Perronnin, “Iterative quantization: A procrustean approach to learning binary codes for large-scale
image retrieval,” TPAMI, pp. 2916–2929, 2013.
[15] K. He, F. Wen, and J. Sun, “K-means hashing: An affinity-preserving
quantization method for learning binary compact codes,” in CVPR, 2013.
[16] J.-P. Heo, Y. Lee, J. He, S.-F. Chang, and S.-e. Yoon, “Spherical
hashing,” in CVPR, 2012.
[17] W. Kong and W.-J. Li, “Isotropic hashing,” in NIPS, 2012.
[18] C. Strecha, A. M. Bronstein, M. M. Bronstein, and P. Fua, “LDAhash:
Improved matching with smaller descriptors,” TPAMI, pp. 66–78, 2012.
[19] W. Liu, J. Wang, R. Ji, Y.-G. Jiang, and S.-F. Chang, “Supervised
hashing with kernels,” in CVPR, 2012.
[20] J. Wang, S. Kumar, and S. Chang, “Semi-supervised hashing for largescale search,” TPAMI, pp. 2393–2406, 2012.
[21] M. Norouzi, D. J. Fleet, and R. Salakhutdinov, “Hamming distance
metric learning,” in NIPS, 2012.
[22] G. Lin, C. Shen, D. Suter, and A. van den Hengel, “A general two-step
approach to learning-based hashing,” in ICCV, 2013.
[23] B. Kulis and T. Darrell, “Learning to hash with binary reconstructive
embeddings,” in NIPS, 2009.
[24] F. Shen, C. Shen, W. Liu, and H. Tao Shen, “Supervised discrete
hashing,” in CVPR, 2015.
[25] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification
with deep convolutional neural networks,” in NIPS, 2012.
[26] A. S. Razavian, H. Azizpour, J. Sullivan, and S. Carlsson, “CNN features
off-the-shelf: An astounding baseline for recognition,” in CVPRW, 2014.
[27] J. Long, E. Shelhamer, and T. Darrell, “Fully convolutional networks
for semantic segmentation,” in CVPR, 2015.
[28] F. Zhao, Y. Huang, L. Wang, and T. Tan, “Deep semantic ranking based
hashing for multi-label image retrieval,” in CVPR, 2015.
[29] R. Zhang, L. Lin, R. Zhang, W. Zuo, and L. Zhang, “Bit-scalable
deep hashing with regularized similarity learning for image retrieval
and person re-identification,” TIP, pp. 4766–4779, 2015.
[30] H. Lai, Y. Pan, Y. Liu, and S. Yan, “Simultaneous feature learning and
hash coding with deep neural networks,” in CVPR, 2015.
[31] Y. Cao, M. Long, J. Wang, H. Zhu, and Q. Wen, “Deep quantization
network for efficient image retrieval,” in AAAI, 2016.
[32] G. E. Hinton, S. Osindero, and Y. W. Teh, “A fast learning algorithm
for deep belief nets,” Neural Computation, pp. 1527–1554, 2006.
[33] V. Erin Liong, J. Lu, G. Wang, P. Moulin, and J. Zhou, “Deep hashing
for compact binary codes learning,” in CVPR, 2015.
[34] M. A. Carreira-Perpinan and R. Raziperchikolaei, “Hashing with binary
autoencoders,” in CVPR, 2015.
[35] T.-T. Do, A.-D. Doan, and N.-M. Cheung, “Learning to hash with binary
deep neural network,” in ECCV, 2016.
[36] J. Xiao, J. Hays, K. A. Ehinger, A. Oliva, and A. Torralba, “SUN
database: Large-scale scene recognition from abbey to zoo,” in CVPR,
2010.
[37] R. Salakhutdinov and G. E. Hinton, “Semantic hashing,” Int. J. Approx.
Reasoning, pp. 969–978, 2009.
[38] J. Lu, V. E. Liong, and J. Zhou, “Deep hashing for scalable image
search,” TIP, 2017.
[39] J. Nocedal and S. J. Wright, Numerical Optimization, Chapter 17,
2nd ed. World Scientific, 2006.
[40] D. C. Liu and J. Nocedal, “On the limited memory bfgs method for large
scale optimization,” Mathematical Programming, vol. 45, pp. 503–528,
1989.
[41] A. Krizhevsky, “Learning multiple layers of features from tiny images,”
University of Toronto, Tech. Rep., 2009.
[42] Y. Lecun and C. Cortes, “The MNIST database of handwritten digits,”
http://yann.lecun.com/exdb/mnist/.
[43] H. Jégou, M. Douze, and C. Schmid, “Product quantization for nearest
neighbor search,” TPAMI, pp. 117–128, 2011.
[44] D. G. Lowe, “Distinctive image features from scale-invariant keypoints,”
IJCV, pp. 91–110, 2004.
[45] A. Oliva and A. Torralba, “Modeling the shape of the scene: A holistic
representation of the spatial envelope,” IJCV, pp. 145–175, 2001.
[46] V. A. Nguyen, J. Lu, and M. N. Do, “Supervised discriminative hashing
for compact binary codes,” in ACM Multimedia, 2014.
[47] A. Vedaldi and K. Lenc, “Matconvnet: Convolutional neural networks
for matlab,” in ACM Multimedia, 2015.
[48] H. Zhu, M. Long, J. Wang, and Y. Cao, “Deep hashing network for
efficient similarity retrieval,” in AAAI, 2016.
| 1cs.CV
|
Synthesizing Imperative Programs for Introductory Programming Assignments
Hakjoo Oh
Korea University
hakjoo [email protected]
arXiv:1702.06334v2 [cs.PL] 13 Jun 2017
Sunbeom So
Korea University
sunbeom [email protected]
Abstract
We present a novel algorithm that synthesizes imperative programs for introductory programming
courses. Given a set of input-output examples and
a partial program, our algorithm generates a complete program that is consistent with every example. Our key idea is to combine enumerative program synthesis and static analysis, which aggressively prunes out a large search space while guaranteeing to find, if any, a correct solution. We have
implemented our algorithm in a tool, called S IMPL,
and evaluated it on 30 problems used in introductory programming courses. The results show that
S IMPL is able to solve the benchmark problems in
6.6 seconds on average.
1 Introduction
Our long-term goal is to build an intelligent tutoring system that helps students to improve their programming skills.
Our experience in introductory programming courses is that
students, who learn programming for the first time, often
struggle with solving programming problems for themselves.
Manually providing guidance simply does not scale for the increasingly large number of students. To make matters worse,
we found that even instructors sometimes make mistake and
shy students are reluctant to ask questions. Motivated by this
experience, we aim to build an automatic system that helps
students to improve their skills without human teachers.
In this paper, we present a key component of the system,
which automatically generates complete programs from students’ incomplete programs. The inputs of the algorithm
are a partial program with constraints on variables and constants, and input-output examples that specify the program’s
behavior. The output is a complete program whose behavior
matches all of the given input-output examples.
The key novelty of our algorithm is to combine enumerative program synthesis and program analysis techniques. It
basically enumerates every possible candidate program in increasing size until it finds a solution. This algorithm, however, is too slow to be interactively used with students due
to the huge search space of programs. Our key idea to accelerate the speed is to perform static analysis alongside the
enumerative search, in order to “statically” identify and prune
out interim programs that eventually fail to be a solution. We
formalize our pruning technique and its safety property.
The experimental results show that our algorithm is remarkably effective to synthesize introductory imperative programs. We have implemented the algorithm in a tool, S IMPL,
and evaluated its performance on 30 programming tasks used
in introductory courses. With our pruning technique, S IMPL
is fast enough to solve each problem in 6.6 seconds on average. However, without the pruning, the baseline algorithm,
which already adopts well-known optimization techniques,
takes 165.5 seconds (25x slowdown) on average.
We summarize our contributions below:
• We present a new algorithm for synthesizing imperative
programs from examples. To our knowledge, our work
is the first to combine enumerative program synthesis
and static analysis technologies.
• We prove the effectiveness of our algorithm on 30 real
programming problems used in introductory courses.
The results show that our algorithm quickly solves the
problems, including ones that most beginner-level students have hard times to solve.
• We provide a tool, S IMPL , which is publicly available
and open-sourced.1
2 Showcase
In this section, we showcase S IMPL with four programming
problems that most beginners feel difficult to solve. To use
S IMPL, students need to provide (1) a partial program, (2) a
set of input-output examples, and (3) resources that S IMPL
can use. The resources consist of a set of integers, a set of
integer-type variables, and a set of array-type variables. The
goal of S IMPL is to complete the partial program w.r.t. the
input-output examples, using only the given resources.
Problem 1 (Reversing integer) The first problem is to
write a function that reverses a given integer. For example,
given integer 12, the function should return 21. Suppose a
partial program is given as
reverse (n){ r := 0; while(?){?}; return r;}
1
Hidden for double-blind reviewing.
reverse(n){
r := 0;
while ( n > 0 ){
x := n % 10;
r := r * 10;
r := r + x;
n := n / 10;
};
return r;
}
(a) Problem1
sum(n){
r := 0;
while ( n > 0 ){
t := n;
while (t > 0){
r := r + t;
t := t - 1;
};
n := n - 1;
};
return r;
}
(c) Problem 3
count(n,a){
while ( n > 0 ){
t := n % 10;
a[t] := a[t] + 1;
n := n / 10;
};
return a;
}
(b) Problem 2
abssum(a, len){
r := 0;
i := 0;
while (i < len){
if ( a[i] < 0 )
{ r := r - a[i]; }
else
{ r := r + a[i]; }
i := i + 1;
};
return r;
}
(d) Problem 4
Figure 1: Synthesized results by S IMPL (in the boxes)
where ? denotes holes that need to be completed. Suppose
further S IMPL is provided with input-output examples {1 7→
1, 12 7→ 21, 123 7→ 321}, integers {0, 1, 10}, and integer
variables {n, r, x}.
Given this problem, S IMPL produces the solution in Figure 1(a) in 2.5 seconds. Note that, S IMPL finds out that the
integer ‘1’ is unnecessary and the final program does not contain it. Also, S IMPL does not require sophisticated examples,
so that S IMPL can be easily used by inexperienced students.
Problem 2 (Counting) The next problem is to write a function that counts the number of each digit in an integer. The
program takes an integer and an array as inputs, where each
element of the array is initially 0. As output, the program returns that array but now each array element at index i stores
the number of is that occur in the given integer. For example, when a tuple (220, h0, 0, 0i) is given, the function should
output h1, 0, 2i; 0 occurs once, 1 does not occur, and 2 occurs
twice in ‘220’. Suppose the partial program is given as
count(n,a){ while(?){?}; return a;}
with examples {(11, h0, 0i) 7→ h0, 2i, (220, h0, 0, 0i) 7→
h1, 0, 2i}, integers {0, 1, 10}, integer variables {i, n, t}, and
an array variable {a}.
For this problem, S IMPL produces the program in Figure 1(b) in 0.2 seconds. Note that S IMPL uses a minimal set
of resources; i is not used though it is given as usable.
Problem 3 (Sum of sum) The third problem is to compute
1 + (1 + 2) + ... + (1 + 2 + ... + n) for a given integer n.
Suppose the partial program
sum(n){ r := 0; while(?){?}; return r;}
is given with examples {1 7→ 1, 2 7→ 4, 3 7→ 10, 4 7→ 20},
integers {0, 1}, and integer-type variables {n, t, r}.
Then, S IMPL produces the program in Figure 1(c) in 37.6
seconds. Note that S IMPL newly introduced a nested loop,
which is absent in the partial program.
Problem 4 (Absolute sum) The last problem is to sum the
absolute values of all the elements in a given array. We provide the partial program:
abssum(a, len){ r := 0; i := 0;
while(i < len){ if(?){?} else{?}; i:=i+1;};
return r;}
where the goal is to complete the condition and bodies
of the if-statement. Given a set of input-output examples
{(h−1, −2i, 2) 7→ 3, (h2, 3, −4i, 3) 7→ 9}, an integer {0},
integer variables {r, i}, and an array variable {a}, S IMPL
produces the program in Figure 1(d) in 12.1 seconds.
3 Problem Definition
Language We designed an imperative language that is small
yet expressive enough to deal with various programming
problems in introductory courses. The syntax of the language
is defined by the following grammar:
⊕ → + | − | ∗ | / | %, ≺→ =|>|<
l → x | x[y], a → n | l | l1 ⊕ l2 | l ⊕ n | ♦
b → true | false | l1 ≺ l2 | l ≺ n | b1 ∧ b2 | b1 ∨ b2 | ¬b | △
c → l := a | skip | c1 ; c2 | if b c1 c2 | while b c |
An l-value (l) is a variable (x) or an array reference (x[y]).
An arithmetic expression (a) is an integer constant (n), an
l-value (l), or a binary operation (⊕). A boolean expression
(b) is a boolean constant (true, false), a binary relation (≺),
a negation (¬b), or a logical conjunction (∧) and disjunction
(∨). Commands include assignment (l := a), skip (skip), sequence (c1 ; c2 ), conditional statement (if b c1 c2 ), and whileloop (while b c).
A program P = (x, c, y) is a command with input and output variables, where x is the input variable, c is the command,
and y is the output variable. The input and output variables x
and y can be either of integer or array types. For presentation
brevity, we assume that the program takes a single input, but
our implementation supports multiple input variables as well.
An unusual feature of the language is that it allows to write
incomplete programs. Whenever uncertain, any arithmetic
expressions, boolean expressions, and commands can be left
out with holes (♦, △, ). The goal of our synthesis algorithm
is to automatically complete such partial programs.
The semantics of the language is defined for programs
without holes. Let X be the set of program variables, which
is partitioned into integer and array types, i.e., X = Xi ⊎ Xa .
A memory state
m ∈ M = X → V,
v ∈ V = Z + Z∗
is a partial function from variables to values (V). A value
is either an integer or an array of integers. An array a ∈
Z∗ is a sequence of integers. For instance, we write h1, 2, 3i
A[[n]](m) = n
A[[x]](m) = m(x)
A[[x[y]]](m) = m(x)m(y)
A[[l1 ⊕ l2 ]](m) = A[[l1 ]](m) ⊕ A[[l2 ]](m)
A[[l ⊕ n]](m) = A[[l]](m) ⊕ n
B[[true ]](m) = true
B[[false]](m) = false
B[[l1 ≺ l2 ]](m) = A[[l1 ]](m) ≺ A[[l2 ]](m)
B[[l ≺ n]](m) = A[[l]](m) ≺ n
B[[b1 ∧ b2 ]](m) = B[[b1 ]](m) ∧ B[[b2 ]](m)
B[[b1 ∨ b2 ]](m) = B[[b1 ]](m) ∨ B[[b2 ]](m)
B[[¬b]](m) = ¬B[[b]](m)
C[[x := a]](m) = m[x 7→ A[[a]](m)]
A[[a]](m)
C[[x[y] := a]](m) = m[x 7→ m(x)m(y) ]
C[[skip]](m) = m
C[[c1 ; c2 ]](m) = (C[[c2 ]] ◦ C[[c1 ]])(m)
C[[if b c1 c2 ]](m) = cond(B[[b]], C[[c1 ]], C[[c2 ]])(m)
C[[while b c]](m) = (fix F )(m)
where F (g) = cond(B[[b]],
g ◦ C[[c]], λx.x)
f (m) if p(m) = true
cond(p, f, g)(m) =
g(m) if p(m) = false
Figure 2: Semantics of the language
for the array of integers 1, 2, and 3. We write |a|, ai , and
aki for the length of a, the element at index i, and the array
a0 . . . ai−1 kai+1 . . . a|a|−1 , respectively.
The semantics of the language is defined by the functions:
A[[a]] : M → V, B[[b]] : M → B, C[[c]] : M → M
where A[[a]], B[[b]], and C[[c]] denote the semantics of arithmetic expressions, boolean expressions, and commands, respectively. Figure 2 presents the denotational semantics,
where fix is a fixed point operator. Note that the semantics
for holes is undefined.
Synthesis Problem A synthesis task is defined by the five
components:
((x, c, y), E, Γ, Xi , Xa )
where (x, c, y) is an incomplete program with holes, E ⊆
V × V is a set of input-output examples. Γ ⊆ Z is a set of
integers, Xi is a set of integer-type variables, and Xa is a set
of array-type variables. The goal of our synthesis algorithm
is to produce a complete command c without holes such that
• c uses constants and variables in Γ and Xi ∪ Xa , and
• c is consistent with every input-output example:
∀(vi , vo ) ∈ E. C[[c]]([x 7→ vi ]) (y) = vo .
4 Synthesis Algorithm
In this section, we present our synthesis algorithm that combines enumerative search with static analysis.
4.1
Synthesis as State-Search
We first reduce the synthesis task into a state-search problem.
Consider a synthesis task ((x, c, y), E, Γ, Xi , Xa ). The corresponding search problem is defined by the transition system
(S, →, s0 , F ) where S is a set of states, (→) ⊆ S × S is a
transition relation, s0 ∈ S is an initial state, and F ⊆ S is a
set of solution states.
Algorithm 1 Synthesis Algorithm
Input: A synthesis problem ((x, c0 , r), E, Γ, Xi , Xa )
Output: A complete program consistent with E
1: W ← {c0 }
2: repeat
3:
Pick the smallest state s from W
4:
if s is a terminal state then
5:
if solution(s) then return s
6:
else
7:
if ¬prune(s) then W ← W ∪ next(s)
8: until W 6= ∅
• States : A state s ∈ S is a command possibly with holes,
which is defined by the grammar in Section 3.
• Initial state : An initial state s0 is a partial command c0 .
• Transition relation : Transition relation (→) ⊆ S × S
determines the state that is immediately reachable from a
state. The relation is defined as a set of inference rules in
Figure 3. Intuitively, a hole can be replaced by an arbitrary expression (or command) of the same type. Given
a state s, we write next(s) for the set of all immediate
next states, i.e., next(s) = {s′ | s → s′ }. We write s 6→
for terminal states, i.e., states with no holes.
• Solution states : A state s is a solution iff s is a terminal
state and it is consistent with all input-output examples:
solution(s) ⇐⇒
s 6→ ∧ ∀(vi , vo ) ∈ E. C[[s]]([x 7→ vi ]) (y) = vo .
4.2
Baseline Search Algorithm
Algorithm 1 shows the basic architecture of our enumerative
search algorithm. The algorithm initializes the workset W
with c0 (line 1). Then, it picks a state s with the smallest
size and removes the state from the workset (line 3). If s is
a solution state, the algorithm terminates and s is returned
(line 5). For a non-terminal state, the algorithm attempts to
prune the state by invoking the function prune (line 7). If
pruning fails, the next states of s are added into the workset
and the loop repeats. The details of our pruning technique
is described in Section 4.3. At the moment, assume prune
always fails.
The baseline algorithm implicitly performs two wellknown optimization techniques. First, it maintains previously
explored states and never reconsider them. Second, more importantly, it normalizes states so that semantically-equivalent
programs are also syntactically the same. For instance, suppose (r := 0; r := x ∗ 0; ) is the current state. Before
pushing it to the workset, we first normalize it to (r := 0; ).
To do so, we use four code optimization techniques: constant
propagation, copy propagation, dead code elimination, and
expression simplification [Aho et al., 1986]. These two techniques significantly improve the speed of enumerative search.
In addition, the algorithm considers terminating programs
only. Our language has unrestricted loops, so the basic algorithm may synthesize non-terminating programs. To exclude
them from the search space, we use syntactic heuristics to detect potentially non-terminating loops. The heuristics are: 1)
c1 → c′1
c2 → c′2
c1 → c′1
c2 → c′2
b → b b′
b → b b′
a →a a ′
′
′
′
′
′
′
l := a → l := a c1 ; c2 → c1 ; c2 c1 ; c2 → c1 ; c2 if b c1 c2 → if b c1 c2 if b c1 c2 → if b c1 c2 if b c1 c2 → if b c1 c2 ¬b →b ¬b′
b1 →b b′1
b2 →b b′2
b2 →b b′2
b1 →b b′1
b → b b′
c → c′
while b c → while b′ c while b c → while b c′ b1 ∧ b2 →b b′1 ∧ b2 b1 ∧ b2 →b b1 ∧ b′2 b1 ∨ b2 →b b′1 ∨ b2 b1 ∨ b2 →b b1 ∨ b′2
→ l := ♦ → skip
△ →b l ≺ n
→ ;
△ →b △ ∧ △
→ if △
△ →b △ ∨ △
→ while △
△ →b ¬△
♦ →a n
△ →b true △ →b f alse △ →b l1 ≺ l2
♦ →a l
♦ → a l1 ⊕ l2
♦ →a l ⊕ n
Figure 3: Transition Relation (n ∈ Γ, l ∈ Xi ∪ {x[y] | x ∈ Xa ∧ y ∈ Xi })
example1(n){
r := 0;
while (n > 0){
r := n + 1;
n := ♦;
};
return r;
}
example2(n) {
r := 0;
while (n > 0){
;
r := x * 10;
n := n / 10;
};
return r;
}
(a)
(b)
Figure 4: States that are pruned away
we only allow boolean expressions of the form x < y (or
x > n) in loop conditions, 2) the last statement of the loop
body must increase (or decrease) the induction variable x, and
3) x and y are not defined elsewhere in the loop.
4.3
Pruning with Static Analysis
Now we present the main contribution of this paper, pruning
with static analysis. Static analysis allows to safely identify
states that eventually fail to be a solution. We first define the
notion of failure states.
Definition 1. A state s is a failure state, denoted fail(s), iff
every terminal state s′ reachable from s is not a solution, i.e.,
fail(s) ⇐⇒ ((s →∗ s′ ) ∧ s′ 6→ =⇒ ¬solution(s′ )).
Our goal is to detect as many failure states as possible. We
observed two typical cases of failure states that often show up
during the baseline search algorithm.
Example 1. Consider the program in Figure 4(a) and inputoutput example (1, 1). When the program is executed with
n = 1, no matter how the hole (♦) gets instantiated, the output value r is no less than 2 at the return statement. Therefore, the program cannot but fail to satisfy the example (1, 1).
Example 2. Consider the program in Figure 4(b) and inputoutput example (1, 1). Here, we do not know the exact values
of x and r, but we know that 10 ∗ x = 1 must hold at the end
of the program. However, there exists no such integer x, and
we conclude the partial program is a failure state.
Static Analysis We designed a static analysis that aims to
effectively identify these two types of failure states. To do so,
our analysis combines numeric and symbolic analyses; the
numeric analysis is designed to detect the cases of Example 1
and the symbolic analysis for the cases of Example 2. The
abstract domain of the analysis is defined as follows:
b = X → V,
b
m
b ∈M
b =I×S
vb ∈ V
An abstract memory state m
b maps variables to abstract
b
values (V).
An abstract value is a pair of intervals (I)
and symbolic values (S). The domain of intervals is standard [Cousot and Cousot, 1977]:
I = ({⊥} ∪ {[l, u] | l, u ∈ Z ∪ {−∞, +∞} ∧ l ≤ u}, ⊑I).
For symbolic analysis, we define the following flat domain:
S = (SE⊤
⊥ , ⊑S ) where SE → n | βx (x ∈ Xi ) | SE ⊕ SE
A symbolic expression se ∈ SE is a constant (n), a symbol
(βx ), or a binary operation with symbolic expressions. We
introduce symbols one for each integer-type variable in the
program. The symbolic domain is flat and has the partial order: s1 ⊑S s2 ⇐⇒ (s1 = ⊥) ∨ (s1 = s2 ) ∨ (s2 = ⊤). We
b that transforms
define the abstraction function α : V → V
concrete values to abstract values:
α(n) = ([n, n], n)
α(n1 . . . nk ) = ([min{n1 , . . . , nk }, max{n1 , . . . , nk }], ⊤).
The abstract semantics is defined in Figure 5 by the functions:
b → V,
b
b
A[[a]]
:M
b → B,
b
b :M
B[[b]]
b →M
b
b :M
C[[c]]
b = {true,
[ ⊤ is the abstract boolean lattice.
d false}
where B
⊥
Intuitively, the abstract semantics over-approximates the
concrete semantics of all terminal states that are reachable
from the current state. This is done by defining the sound
b
b
b
semantics for holes: A[[♦]](
m),
b B[[△]](
m),
b and C[[]](
m).
b
An exception is that integer variables get assigned symbols,
rather than ⊤, in order to generate symbolic constraints on
integer variables.
In our analysis, array elements are abstracted into a single
b
b
element. Hence, the definitions of A[[x[y]]]
and C[[x[y]
:= a]]
do not involve y. Because an abstract array cell may represent
multiple concrete cells, arrays are weakly updated by joining
(⊔) old and new values. For example, in memory state m
b =
b
[x 7→ ([5, 5], ⊤), ...], C[[x[y]
:= 1]](m)
b evaluates to [x 7→
([1, 5], ⊤), ...].
For while-loops, the analysis performs a sound fixed point
computation. If the computation does not reach a fixed point
after a fixed number of iterations, we apply widening for
b
A[[n]](
m)
b
b m)
A[[l]](
b
=
=
=
=
=
b
B[[true
]](m)
b
b
B[[l1 ≺ l2 ]](m)
b
b ≺ n]](m)
B[[l
b
b
B[[b1 ∧ b2 ]](m)
b
b 1 ∨ b2 ]](m)
B[[b
b
b
B[[¬b]](
m)
b
b
B[[△]](
m)
b
=
=
=
=
=
=
=
b 1 ⊕ l2 ]](m)
A[[l
b
b ⊕ n]](m)
A[[l
b
b
A[[♦]](
m)
b
b := ♦]](m)
C[[x
b
b
C[[x[y] := ♦]](m)
b
b := a]](m)
C[[x
b
b
C[[x[y]
:= a]](m)
b
b
C[[skip]](
m)
b
b 1 ; c2 ]](m)
C[[c
b
b b c1 c2 ]](m)
C[[if
b
b
C[[while
b c]](m)
b
where Fb (g)
b
C[[]](
m)(x)
b
=
=
=
=
=
=
=
=
=
=
([n, n], n)
m(x)
b
(l = x or x[y])
b 1 ]](m)
b 2 ]](m)
b A[[l
A[[l
b ⊕
b
b
bn
A[[l1 ]](m)
b ⊕
([−∞, +∞], ⊤)
b
[
d (B[[false]](
true
m)
b = false)
b
b
b
A[[l1 ]](m)
b ≺ A[[l2 ]](m)
b
b m)
b n
A[[l]](
b ≺
b 1 ]](m)
b 2 ]](m)
b B[[b
B[[b
b ∧
b
b 1 ]](m)
b 2 ]](m)
b B[[b
B[[b
b ∨
b
b
¬
b B[[b]](
m)
b
⊤
m[x
b 7→ ([−∞, +∞], βx )]
m[x
b 7→ ([−∞, +∞], ⊤)]
b
m[x
b 7→ A[[a]](
m)]
b
b
m[x
b 7→ A[[a]](
m)
b ⊔ m(x)]
b
m
b
b 2 ]] ◦ C[[c
b 1 ]])(m)
(C[[c
b
b
b 1 ]], C[[c
b 2 ]])(m)
[
cond(B[[b]], C[[c
b
b
b
(fix F )(m)
b
b
b
[ B[[b]],
cond(
g ◦ C[[c]],
λx.x)
([−∞, +∞], βx ) x ∈ Xi
([−∞, +∞], ⊤) x ∈ Xa
Figure 5: Abstract semantics
infinite interval domain, in order to guarantee the termination of the analysis. We use the standard widening operator
b and cond
d in
in [Cousot and Cousot, 1977]. The function fix
Figure 5 denote a post-fixed point operator and a sound abstraction of cond, respectively.
Pruning Next we describe how we do pruning with the static
analysis. Suppose we are given examples E ⊆ V × V and
a state s with input (x) and output (y) variables. For each
example (vi , vo ) ∈ E, we first run the static analysis with the
input α(vi ) and obtain the analysis result (itvs , se s ) :
b
(itvs , se s ) = (C[[s]]([x
7→ α(vi )])(y).
We only consider the case when itvs = [ls , us ] (when itvs =
⊥, the program is semantically ill-formed and therefore we
just prune out the state). Then, we obtain the interval abstraction [lo , uo ] of the output vo , i.e., ([lo , uo ], −) = α(vo ), and
s
generate the constraints C(v
:
i ,vo )
s
C(v
= (ls ≤ lo ∧ uo ≤ us ) ∧ (se ∈ SE =⇒ lo ≤ se ≤ uo ).
i ,vo )
The first (resp., second) conjunct means that the interval
(resp., symbolic) analysis result must over-approximate the
s
output example. We prune out a state s iff C(v
is unsatisi ,vo )
fiable for some example (vi , vo ) ∈ E:
Definition 2. The predicate prune is defined as follows:
s
prune(s) ⇐⇒ C(v
is unsatisfiable for some (vi , vo ) ∈ E.
i ,vo )
The unsatisfiability can be easily checked, for instance,
with an off-the-shelf SMT solver. Our pruning is safe:
Theorem 1 (Safety). ∀s ∈ S. prune(s) =⇒ fail(s).
That is, we prune out a state only when it is a failure state,
which formally guarantees that the search algorithm with our
pruning finds a solution if and only if the baseline algorithm
(Section 4.2) does so.
5 Evaluation
Experimental setup To evaluate our synthesis algorithm, we
gathered 30 introductory level problems from several online
forums (Table 1).2 The problems consist of tasks manipulating integers and arrays. Some problems are non-trivial for
novice students to solve; they require students to come up
with various control structures such as nested loops and combinations of loops and conditional statements. The partial
programs we used are similar to those shown in Section 2;
they have one boolean expression hole (△), and one or two
command holes (). For each benchmark, we report the
number of integer variables (IVars), array variables (AVars),
integer constants (Ints), and examples (Exs) provided, respectively. All benchmark problems are publicly available with
our tool. Experiments were conducted on MacBook Pro with
Intel Core i7 and 16GB of memory.
Baseline Algorithm Table 1 shows the performance of our
algorithm. The column “Base” shows the running time of our
baseline algorithm that performs enumerative search without
state normalization. In that case, the average runtime was
longer than 616 seconds, and three of the benchmarks timed
out (> 1 hour). The column “Base+Opt” reports the performance of the baseline with normalization. It shows that
normalizing states succeeds to solve all benchmark problems
and improves the speed by more than 3.7 times on average,
although it degrades the speed for some cases due to runtime
normalization overhead.
Pruning Effectiveness On top of “Base+Opt”, we applied
our static-analysis-guided pruning technique (the column
“Ours”). The results show that our pruning technique is remarkably effective. It reduces the average time to 6.6 seconds, improving the speed of “Base+Opt” by 25 times. Note
that S IMPL is able to synthesize the desired programs from a
few examples (Exs), requiring up to 4 examples.
6 Related Work
Computer-aided education Recently, program synthesis
technology has revolutionized computer-aided education.
For instance, the technology has been used in automatic
problem generation [Singh et al., 2012; Ahmed et al., 2013;
Alvin et al., 2014; Polozov et al., 2015], automatic grading [Alur et al., 2013], and automatic solution generation [Gulwani et al., 2011].
Our work is to use program synthesis for automated
programming education system.
A large amount of
work has been done to automate programming education [Adam and Laurent, 1980; Soloway et al., 1981;
Farrell et al., 1984;
Johnson and Soloway, 1984;
Murray, 1989; Singh et al., 2013; Gulwani et al., 2014;
Kaleeswaran et al., 2016; Kim et al., 2016], which focuses
2
E.g., http://www.codeforwin.in
Domain
Integer
Array
Vars
No
Description
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
Given n, return n!.
Given n, return n!! (i.e., double factorial).
P
Given n, return ni=1 i.
P
Given n, return ni=1 i2 .
Q
Given n, return ni=1 i2 .
Given a and n, return an .
P
Given n and m, return m
i.
Q i=n
Given n and m, return m
i=n i.
Count the number of digit for an integer.
Sum the digits of an integer.
Calculate product of digits of an intger.
Count the number of binary digit of an integer.
Find the nth Fibonacci number.
P
P
Given n, return ni=1 ( im=1 m)).
Qn Qi
Given n, return i=1 ( m=1 m)).
Reverse a given integer.
Find the sum of all elements of an array.
Find the product of all elements of an array.
Sum two arrays of same length into one array.
Multiply two arrays of same length into one array.
Cube each element of an array.
Manipulate each element into 4th power.
Find a maximum element.
Find a minimum element.
Add 1 to each element.
Find the sum of square of each element.
Find the multiplication of square of each element.
Sum the products of matching elements of two arrays.
Sum the absolute values of each element.
Count the number of each element.
Average
IVars
AVars
2
3
3
4
4
4
3
3
3
3
3
2
3
3
3
3
3
3
3
3
3
3
3
3
2
3
3
3
2
3
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
1
2
2
1
1
1
1
1
1
1
2
1
1
Ints
Exs
2
3
2
2
2
2
2
2
3
3
3
3
3
2
2
3
2
2
2
2
1
1
2
2
1
2
1
1
1
3
4
4
4
3
3
4
3
3
3
4
3
3
4
4
4
3
2
2
2
2
2
2
2
2
3
2
2
3
2
2
Base
0.0
0.0
0.1
122.4
102.9
0.7
0.2
0.2
0.0
5.2
0.7
0.0
98.7
⊥
⊥
⊥
8.1
7.6
44.6
47.4
1283.3
1265.8
0.9
0.8
0.3
2700.0
1709.8
20.5
45.0
238.9
> 616.8
Time (sec)
Base+Opt
0.0
0.0
0.0
18.1
13.6
0.1
0.0
0.0
0.0
2.2
2.3
0.0
13.9
324.9
316.6
367.3
3.6
3.9
29.9
26.4
716.1
715.5
0.7
0.3
0.0
186.2
1040.3
38.7
50.5
1094.1
165.5
Ours
0.0
0.0
0.0
0.3
0.2
0.1
0.0
0.1
0.0
1.3
0.3
0.0
2.6
37.6
86.9
2.5
0.9
0.9
0.2
0.3
13.0
13.0
0.4
0.1
0.0
11.5
12.6
1.5
12.1
0.2
6.6
Table 1: Performance of S IMPL . ⊥ denotes timeout (> 1 hour). Assume ⊥ as 3,600 seconds for the average of “Base”.
primarily on providing feedback on students’ programming submissions. Our system, S IMPL , has the following
advantages over prior works:
• Feedback on incomplete programs: Existing systems
produce feedback only for complete programs; they cannot help students who do not know how to proceed further. In this case, S IMPL can help by automatically generating solutions starting from incomplete solutions.
• No burden on instructor: Existing systems require
instructor’s manual effort. For example, the system
in [Singh et al., 2013] needs a correct implementation
and a set of correction rules manually designed by the
instructor. On the other hand, S IMPL does not require
anything from the instructor.
An exception is [Farrell et al., 1984], where an automatic LISP feedback system is presented. However, the
system produces feedback by relying on ad-hoc rules.
Programming by example Our work differs from
prior programming-by-example (PBE) techniques in
two ways. First, to our knowledge, our work is the
first to synthesize imperative programs with loops.
Most of the PBE approaches focus on domain-specific
languages for string transformation [Gulwani, 2011;
Kini and Gulwani, 2015;
Raza et al., 2015;
Manshadi et al., 2013; Wu and Knoblock, 2015],
number transformation [Singh and Gulwani, 2012], XML
transformation [Raza et al., 2014], and extracting relational data [Le and Gulwani, 2014], etc.
Several
others have studied synthesis of functional programs
[Albarghouthi et al., 2013;
Osera and Zdancewic, 2015;
Frankle et al., 2016]. Second, our algorithm differs from
prior work in that we combine semantic-based static analysis
technology with enumerative program synthesis. Existing
enumerative synthesis technology used pruning techniques such as type systems [Osera and Zdancewic, 2015;
Frankle et al., 2016] and deductions [Feser et al., 2015],
which are not applicable to our setting.
7 Conclusion
In this paper, we have shown that combining enumerative
synthesis and static analysis is a promising way of synthesizing introductory imperative programs. The enumerative
search allows us to find the smallest possible, therefore general, program while the semantics-based static analysis dra-
matically accelerates the process in a safe way. We demonstrated the effectiveness on 30 real programming problems
gathered from online forums.
References
[Adam and Laurent, 1980] Anne Adam and Jean-Pierre Laurent. Laura, a system to debug student programs. Artificial
Intelligence, 15(1-2), November 1980.
[Ahmed et al., 2013] Umair Z. Ahmed, Sumit Gulwani, and
Amey Karkare. Automatically generating problems and
solutions for natural deduction. In IJCAI, 2013.
[Aho et al., 1986] Alfred V. Aho, Ravi Sethi, and Jeffrey D.
Ullman. Compilers: Principles, Techniques, and Tools.
Addison-Wesley Longman Publishing Co., Inc., Boston,
MA, USA, 1986.
[Albarghouthi et al., 2013] Aws Albarghouthi, Sumit Gulwani, and Zachary Kincaid. Recursive program synthesis.
In CAV, 2013.
[Alur et al., 2013] Rajeev Alur, Loris D’Antoni, Sumit Gulwani, Dileep Kini, and Mahesh Viswanathan. Automated
grading of dfa constructions. In IJCAI, 2013.
[Alvin et al., 2014] Chris Alvin, Sumit Gulwani, Rupak Majumdar, and Supratik Mukhopadhyay. Synthesis of geometry proof problems. In AAAI, 2014.
[Cousot and Cousot, 1977] Patrick Cousot and Radhia
Cousot.
Abstract interpretation: A unified lattice
model for static analysis of programs by construction or
approximation of fixpoints. In POPL, 1977.
[Farrell et al., 1984] Robert G. Farrell, John R. Anderson,
and Brian J. Reiser. An interactive computer-based tutor
for lisp. In AAAI, 1984.
[Feser et al., 2015] John K. Feser, Swarat Chaudhuri, and
Isil Dillig. Synthesizing data structure transformations
from input-output examples. In PLDI, 2015.
[Frankle et al., 2016] Jonathan Frankle, Peter-Michael Osera, David Walker, and Steve Zdancewic. Exampledirected synthesis: A type-theoretic interpretation. In
POPL, 2016.
[Gulwani et al., 2011] Sumit Gulwani, Vijay Anand Korthikanti, and Ashish Tiwari. Synthesizing geometry constructions. In PLDI, 2011.
[Gulwani et al., 2014] Sumit Gulwani, Ivan Radiček, and
Florian Zuleger. Feedback generation for performance
problems in introductory programming assignments. In
FSE, 2014.
[Gulwani, 2011] Sumit Gulwani. Automating string processing in spreadsheets using input-output examples. In POPL,
2011.
[Johnson and Soloway, 1984] W. Lewis Johnson and Elliot
Soloway. Proust: Knowledge-based program understanding. In ICSE, 1984.
[Kaleeswaran et al., 2016] Shalini Kaleeswaran, Anirudh
Santhiar, Aditya Kanade, and Sumit Gulwani. Semisupervised verified feedback generation. In FSE, 2016.
[Kim et al., 2016] Dohyeong Kim, Yonghwi Kwon, Peng
Liu, I. Luk Kim, David Mitchel Perry, Xiangyu Zhang, and
Gustavo Rodriguez-Rivera. Apex: Automatic programming assignment error explanation. In OOPSLA, 2016.
[Kini and Gulwani, 2015] Dileep Kini and Sumit Gulwani.
Flashnormalize: Programming by examples for text normalization. In IJCAI, 2015.
[Le and Gulwani, 2014] Vu Le and Sumit Gulwani. Flashextract: A framework for data extraction by examples. In
PLDI, 2014.
[Manshadi et al., 2013] Mehdi Manshadi, Daniel Gildea,
and James Allen. Integrating programming by example
and natural language programming. In AAAI, 2013.
[Murray, 1989] William R. Murray. Automatic Program DeBugging for Intelligent Tutoring Systems. Morgan Kaufmann Publishers Inc., 1989.
[Osera and Zdancewic, 2015] Peter-Michael Osera and
Steve Zdancewic. Type-and-example-directed program
synthesis. In PLDI, 2015.
[Polozov et al., 2015] Oleksandr
Polozov,
Eleanor
O’Rourke, Adam M. Smith, Luke Zettlemoyer, Sumit
Gulwani, and Zoran Popovic. Personalized mathematical
word problem generation. In IJCAI, 2015.
[Raza et al., 2014] Mohammad Raza, Sumit Gulwani, and
Natasa Milic-Frayling. Programming by example using
least general generalizations. In AAAI, 2014.
[Raza et al., 2015] Mohammad Raza, Sumit Gulwani, and
Natasa Milic-Frayling. Compositional program synthesis
from natural language and examples. In IJCAI, 2015.
[Singh and Gulwani, 2012] Rishabh Singh and Sumit Gulwani. Synthesizing number transformations from inputoutput examples. In CAV, 2012.
[Singh et al., 2012] Rohit Singh, Sumit Gulwani, and Sriram
Rajamani. Automatically generating algebra problems. In
AAAI, 2012.
[Singh et al., 2013] Rishabh Singh, Sumit Gulwani, and Armando Solar-Lezama. Automated feedback generation for
introductory programming assignments. In PLDI, 2013.
[Soloway et al., 1981] Elliot M. Soloway, Beverly Woolf,
Eric Rubin, and Paul Barth. Meno-ii: An intelligent tutoring system for novice programmers. In IJCAI. Morgan
Kaufmann Publishers Inc., 1981.
[Wu and Knoblock, 2015] Bo Wu and Craig A. Knoblock.
An iterative approach to synthesize data transformation
programs. In IJCAI, 2015.
| 2cs.AI
|
Finite subsemigroups of free left regular bands
Artem N. Shevlyakov
arXiv:1406.3429v6 [math.AG] 29 Feb 2016
December 26, 2017
Abstract
We describe all finite subsemigroups of free left regular bands. Moreover, we
show applications of this result in algebraic geometry and model theory.
1
Introduction
A semigroup is a left regular band if it satisfies the identities x2 = x, xyx = xy.
Left regular bands have important applications in hyperplane arrangements, random
walks and matroid theory, see e.g. [1, 2, 3, 4, 5, 6, 7, 8]. In the current paper we
describe all finite subsemigroups of free left regular bands (Theorems 6.1 and 6.2).
Actually, our proof is not short and easy. For example, we had to introduce
about six partial orders over left regular bands (remark that all our orders are not
compatible with the multiplication over left regular bands).
Let us give the plan of our paper. The first tree sections contain all necessary
definitions and simple statements devoted to left regular bands and semilattices.
In particular, we give here the important definitions of local linear order and right
hereditary property. In Section 5 we prove (Theorem 5.4) that any finite subband
of a free left regular band has a local linear order. Thus, the local linear order is
the necessary condition of the embeddability into free left regular bands.
In Section 6 we prove that any right hereditary left regular band S with a local
linear order is embeddable into a free left regular band of an appropriate rank.
Firstly, we define a map h : S → F, where F is a free left regular band of infinite
rank, and prove (Lemma 6.9) that h is a homomorphism. However, h may be not
injective. Therefore, in Subsection 6.2 we define a procedure which takes h and
obtains an embedding h′ after a finite number of iterations. Thus, the embedding
h′ proves the main result of our paper (Theorem 6.1).
In the last section of our paper we talk about open problems and possible applications of Theorem 6.1. Moreover, we discuss how to simplify the proof of Theorem 6.1
and why we do not do it.
2
Semilattices
Below we shall consider join semilattices, and denote multiplication by ∪. Elements
of semilattices are denoted below by bold letters x, y, z, . . .
One can define a partial order over any join semilattice by
x ⊆ y ⇔ y ∪ x = y.
(1)
Let M be a set, and FM be the family of all finite subsets (including ∅) of M .
Let us define over FM a multiplication as a set union. Obviously, FM is isomorphic
1
to the join free semilattice of rank |M | with the adjoined identity. We denote the
free basis of FM by {am |m ∈ M }. Sometimes elements of FM (free generators am )
will be called words (respectively, letters). By am ∈ c we shall denote that a letter
am belongs to a word c ∈ FM .
For any b, c ∈ FM one can define c\b ∈ FM as the set difference. The length |a|
of a ∈ FM is the length of a word representing a. Let F denote the free semilattice
generated by an infinite set of generators {ai |i ∈ I}.
Take a finite join semilattice L. According to semilattice theory, L is embeddable
into a free semilattice of an appropriate rank. The following embedding will play a
crucial role in our study.
Let us define a map ν : L → FL by
[
ν(x) =
at .
x*t
Lemma 2.1. ν is an embedding of L into FL .
Proof. First, we prove that ν is a homomorphism of L. Let us take x, y and z = x∪y.
We have
[
[
ν(x) =
at , ν(y) =
ar ,
x*t
y*r
ν(z) =
[
ap .
z*p
It is sufficient to prove
{t|x * t} ∪ {r|y * r} = {p|z * p}.
The relation x * t (y * r) implies z * t (respectively, z * r). Hence,
{t|x * t} ∪ {r|y * r} ⊆ {p|z * p}.
Let z * p. If x ⊆ p and y ⊆ p then z = x ∪ y ⊆ p, a contradiction. Therefore,
either x * p or y * p. Thus, we obtained
{t|x * t} ∪ {r|y * r} ⊇ {p|z * p}.
Second, we prove that ν is an embedding.
Suppose that Fx = Fy for some distinct x, y, where
Fx = {t|x * t}, Fy = {r|y * r}.
Since x ∈
/ Fx , y ∈
/ Fy , we have x ∈
/ Fy , y ∈
/ Fx . It follows that x ⊆ y and y ⊆ x,
hence x = y.
By the definition of ν, we have the next statement.
Lemma 2.2. If ν(x) does not contain a letter ay , then x ⊆ y in L.
Lemma 2.3. Let x, x′ , y, y′ ∈ L, x ⊂ x′ , y ⊂ y′ and
d(x′ ) = ν(x′ ) \ ν(x) ⊆ ν(y′ ) \ ν(y) = d(y′ ).
Then y ⊆ x.
Moreover, the equality d(x′ ) = d(y′ ) implies x = y and x′ = y′ .
2
Proof. By the definition of ν, ν(x) does not contain ax . However x ⊂ x′ implies
ax ∈ ν(x′ ). Hence,
ax ∈ d(x′ ) ⇒ ax ∈ d(y′ ) ⇒ ax ∈
/ ν(y).
By Lemma 2.2, the last expression gives y ⊆ x.
Suppose d(x′ ) = d(y′ ). By above, it follows x = y and therefore x′ = y′ .
3
Left regular bands
A semigroup S is a left regular band if the next identities
x2 = x, xyx = xy
hold in S.
Remark that we always consider left regular bands with the identity ε adjoined.
By F we denote the free left regular of infinite rank. The elements of F are all
words w in an infinite alphabet {ai |i ∈ I} such that any letter ai occurs at most
one time in w. The product of two elements w1 , w2 ∈ F is defined as follows:
w1 w2 = w1 ◦ (w2 )∃ ,
where ◦ is the word concatenation and the operator ∃ is the deletion of all letters
which occur earlier. For example, (a1 a2 a3 )(a2 a3 a4 ) = a1 a2 a3 a4 .
Below elements of F (free generators ai ) are called words (respectively, letters),
and we shall use the denotation ai ∈ w if a letter ai belongs to a word w ∈ F.
Remark that in our denotations free generators of the semilattice F and the left
regular band F are indexed by the common set I.
For elements x, y of a left regular band S one can define a relation
x ≤ y ⇔ xy = y.
It is easy to check that ≤ is a partial order over S. Remark that in some papers
(e.g. see [4]) ≤ defines the contrary relation: yx = x.
For elements x, y ∈ F the relation x ≤ y means that the word x is a prefix of y.
The next lemma contains simple statements about the structure of free left
regular bands F.
Lemma 3.1.
1. The Hasse diagram of the order ≤ is a tree for F. The root of this tree is
the identity ε. For example, the Hasse diagram for the free left regular band
generated by three elements a1 , a2 , a2 is
a1 a2 a3 a1 a3 a2
a2 a1 a3 a2 a3 a1
a1 a2
❆
❆
✁
✁
a1 a3
a2 a1
❆
❆
✁
✁
a3 a1 a2 a3 a2 a1
a2 a3
❆
a3 a1
✁
❆
✁
❆ ✁
❆ ✁
❆ ✁
❆✁a1
❅
❆✁a2
❆✁a3
✏✏
✁
✏
❅
✏✏
✁
❅
✏✏
✏
❅ ✁ ✏✏
ε❅✁✏✏
Fig. 1
3
a3 a2
2. Let x1 , x2 , y1 , y2 ∈ F, xi ≤ yi and x1 , x2 are not ≤-comparable. Then y1 , y2
are not ≤-comparable.
Define the another relation on elements of a left regular band S by
x y ⇔ yx = y.
This relation is reflexive and transitive, but not necessarily antisymmetric.
For a free left regular band F the relation x y means that the word y contains
all letters of a word x. For example, a1 a2 a2 a1 , a2 a1 a2 a3 . Obviously, the
relation x ≤ y implies x y in any left regular band.
We shall write x ≈ y iff x y and y x. One can get a join semilattice L by
identifying x and y if x ≈ y. Let σ : S → L denote the quotient map. L is called
the support semilattice of S and for all x, y ∈ S we have
σ(xy) = σ(x) ∪ σ(y).
It is easy to see that for any s ∈ F the map σ(s) gets the set of all letters
occurring in s. Thus, we obtain
σ(F) = F,
where σ(ai ) = ai (i ∈ I).
F
Let
S∞us define a linear order < over the semilattice F as follows. We have
F = n=0 Ln , where Ln = {a ∈ F | |σ(a)| = n} and each set Ln has a linear order
<Ln . Thus, we put
a <F b ⇔ |σ(a)| < |σ(b)| or a <Ln b for a, b ∈ Ln .
Notice that x ≺ y implies x <F y for all x, y ∈ F.
→
−
Let us define a map () : F → F as follows. Let {ai1 , ai2 , . . . , ain } be the set
of all letters of a word a ∈ F. The letters of a correspond to free generators
{ai1 , ai2 , . . . , ain } ⊆ F and the order <F is given by
ai1 <F ai2 <F . . . <F ain .
−→
Then the word (a) ∈ F is the product
ai1 ai2 . . . ain .
−→
Roughly speaking, (a) replaces the letters of a to the corresponding generators
of F and sorts them in the given order <F .
Following [4], a left regular band S is called right hereditary if the Hasse diagram
of ≤ is a tree with the root element ε. By Lemma 3.1, any free left regular band is
right hereditary.
Let x, y, z be elements of a left regular band S, and the equivalence relation ∼z
be as follows:
x ∼z y ⇔ xz = yz.
Clearly, the relation x ≤ y is equivalent to x ∼y y.
Lemma 3.2. Let S, T be left regular bands. Suppose a map h : S → T preserves
the relations , ∼c for each c ∈ S, i.e. for any a, b, c ∈ S we have:
a b ⇒ h(a) h(b),
a ∼c b ⇒ h(a) ∼h(c) h(b).
Then h is a homomorphism.
4
Proof. Let us prove h(xy) = h(x)h(y) for all x, y ∈ S. Denote z = xy, hence
zy = xy and z ∼y x. By the condition, h(z)h(y) = h(x)h(y). Since y z, we have
h(y) h(z) and obtain h(z) = h(x)h(y).
Example 3.3. Let us define a right hereditary band B = {x1 , x2 , x3 , y0 , y1 , y2 }
defined by the following multiplication table (the identity ε is omitted):
x1 x2 x3 y 0 y 1 y 2
x1 x1 y 0 y 1 y 0 y 1 y 1
x2 x2 x2 y 2 x2 y 2 y 2
x3 x3 x3 x3 x3 x3 x3
y0 y0 y0 y1 y0 y1 y1
y1 y1 y1 y1 y1 y1 y1
y2 y2 y2 y2 y2 y2 y2
The order ≤ over B is defined by the following Hasse diagram:
y2
y1
y0
x
2
✑
✁
✑
✁ ✑
❅
❅ ✁✑✑
ε❅✁✑
x1
x3
Fig. 2
The support semilattice of B is a linear ordered semilattice 1 < 2 < 3 < 4, where
σ(ε) = 1, σ(x1 ) = 2, σ(x2 ) = σ(y0 ) = 3, σ(x3 ) = σ(y1 ) = σ(y2 ) = 4.
The embedding ν gives:
ν(σ(x1 )) = a1 , ν(σ(x2 )) = ν(σ(y0 )) = a1 a2 , ν(σ(x3 )) = ν(σ(y1 )) = ν(σ(y2 )) = a1 a2 a3 .
Thus, ν embeds σ(B) into F.
The multiplication over B generates the following equivalence classes:
x
Equivalence classes of ∼x
x1 {ε, x1 }, {x2 }, {y2 }, {y0 }, {y1 }, {x3 }
x2
{ε, x2 }, {x1 , y0 }, {x3 }, {y1 }, {y2 }
x3
{ε, x3 }, {x2 , y2 }, {x1 , y0 , y1 }
y0
{ε, x1 , y0 }, {x2 }, {x3 }, {y2 }, {y1 }
y1
{ε, x1 , y0 , y1 }, {x2 , y2 }, {x3 }
y2
{ε, x2 , y2 }, {x3 }, {x1 , y0 , y1 }
4
Properties of right hereditary bands
Below S is a right hereditary band and L = σ(S). The following lemma describes
the main property of left regular bands.
Lemma 4.1. The equality x1 x2 = x3 x4 (xi ∈ S) implies x1 ≤ x3 or x3 ≤ x1 .
Proof. Let x = x1 x2 = x3 x4 and we have x1 ≤ x, x3 ≤ x. Since the Hasse diagram
of the order ≤ is a tree, the elements x1 , x3 should be ≤-comparable.
5
Lemma 4.2. Let S be a right hereditary band and x, y, z ∈ S. Then x ≤ y ≤ xz
or y ≤ x ≤ yz iff x ∼z y.
Proof. Let us prove the “only if” part of the lemma. Let x ≤ y ≤ xz. By the
definition of the order ≤ we have:
xy = y, yxz = xz.
Therefore,
(xy)xz = xz ⇔ (xy)z = xz ⇔ yz = xz ⇔ x ∼z y.
Let us prove the “if” part of the lemma. Since S is right hereditary, the relation
xz = yz implies either x ≤ y or y ≤ x. So
x ≤ y ≤ yz = xz or y ≤ x ≤ xz = yz.
Since the Hasse diagram of the order ≤ is a tree with the root ε, it allows us to
define the ancestor α(s) ∈ S for any s 6= ε as follows:
α(s) = max{x ∈ S|x < s}.
An element s′ with α(s′ ) = s is called a descendant of s.
For any c 6= ε let us define a set Sc ⊆ S by
Sc = {s ∈ S \ {ε}|s ∼c α(s), s ≁c′ α(s) for any c′ < c}
(2)
Notice that Sc 6= ∅, since c ∈ Sc .
Example 4.3. For the left regular band B we have: Sx1 = {x1 }, Sx2 = {x2 , y0 },
Sx3 = {x3 , y0 , y1 , y2 }, Sy0 = {y0 }, Sy1 = {y1 , y2 }, Sy2 = {y1 , y2 }.
Lemma 4.4. For any x ∈ Sb , a < b we have a α(x).
Proof. By the condition, a ≤ α(b) < b.
By Lemma 4.2 we have
α(x) ≤ x ≤ α(x)b.
Also we have
α(x) ≤ α(x)a ≤ α(x)ab = α(x)b.
Since the Hasse diagram of ≤ is a tree, the first chain above should contain
α(x)a.
If α(x) ≤ x ≤ α(x)a Lemma 4.2 gives α(x) ∼a x that contradicts the definition
of the set Sb . Thus,
α(x) ≤ α(x)a < x.
By the definition of the ancestor, it holds α(x) = α(x)a, hence a α(x).
Lemma 4.5.
comparable.
For all x, y ∈ Sc the elements α(x)α(y)x, α(x)α(y)y are ≤-
6
Proof. By the definition of Sc , we have α(x)c = xc, α(y)c = yc. Let us multiply
both equalities by α(x)α(y) and obtain
α(x)α(y)c = α(x)α(y)xc, α(x)α(y)c = α(x)α(y)yc.
Therefore,
α(x)α(y)xc = α(x)α(y)yc.
Since S is right hereditary, the prefixes α(x)α(y)x, α(x)α(y)y are ≤-comparable.
Let s ∈ S \ {ε} and denote
δν (s) = ν(σ(s)) \ ν(σ(α(s))) ∈ FL .
(3)
A band S has a local linear order if for each b the set Sb admits a linear order
⊏b such that for all x, y ∈ Sb it holds:
1.
δν (x) ⊂ δν (y) ⇒ x ⊏b y;
(4)
σ(x) ⊆ σ(α(y)) ⇒ x ⊏b y.
(5)
2.
3. for arbitrary Sc ⊇ Sb we have:
x ⊏b y ⇒ x ⊏c y.
(6)
Finite subbands of F
5
Below in this section we assume S ⊆ F, |S| < ∞, L = σ(S) ⊆ F and ν embeds L
into FL . Let {w} denote the set of all letters of a word w ∈ F. By the properties of
F, for each x ∈ S \ {ε} ⊆ F there exists a unique ∆(x) ∈ F such that x = α(x)∆(x)
and {α(x)} ∩ {∆(x)} = ∅.
Lemma 5.1. The following correspondences between δν and ∆ hold:
1.
δν (x) ⊆ δν (y) ⇒ ∆(x) ∆(y)
(7)
δν (x) ⊂ δν (y) and ∆(x) ≈ ∆(y) ⇒ y ≺ x.
(8)
2.
Proof.
1. By Lemma 2.3, we have α(y) α(x). Let
s1 = α(x)α(y)x = α(x)x = α(x)∆(x),
s2 = α(x)α(y)y = α(x)y = α(x)α(y)∆(y) = α(x)∆(y).
The images of the support mapping are:
ν(σ(s1 )) = ν(σ(α(x))) ∪ ν(σ(α(y)) ∪ ν(σ(x)) = ν(σ(α(x))) ∪ ν(σ(α(y)) ∪ δν (x)
ν(σ(s2 )) = ν(σ(α(x))) ∪ ν(σ(α(y)) ∪ ν(σ(y)) = ν(σ(α(x))) ∪ ν(σ(α(y)) ∪ δν (y)
Since δν (x) ⊆ δν (y), we have ν(σ(s1 )) ⊆ ν(σ(s2 )), and therefore s1 s2 . By
Lemma 4.5 s1 , s2 are ≤-comparable, hence s1 ≤ s2 and
α(x)∆(x) ≤ α(x)∆(y).
Let a be a letter form ∆(x). Since a ∈
/ α(x), we have a ∈ ∆(y). Thus,
∆(x) ∆(y).
7
2. According to Lemma 2.3, we have
α(y) α(x).
(9)
By ∆(x) ≈ ∆(y) we obtain y = α(y)∆(y) α(x)∆(x) = x. Thus, y x.
Assume
x ≈ y,
(10)
then
yx = y ⇒ α(y)∆(y)α(x)∆(x) = α(y)∆(y) ⇒
α(y)∆(y)α(x) = α(y)∆(y) ⇒ α(x) α(y)∆(y).
By ∆(x) ≈ ∆(y), the words α(x) and ∆(y) have no common letters, therefore,
α(x) α(y).
By (9), it follows α(x) ≈ α(y). Using (10), we obtain δν (x) = δν (y) that
contradicts the condition.
Let us define relations ⊏= , ⊏6=
b over a set Sb as follows:
x ⊏= y ⇔ ∆(x) ≈ ∆(y), and y <F x,
where the linear order <F over F was defined above.
Define x ⊏6=
b y iff b contains a letter a ∈ {∆(y)} \ {∆(x)} and there is not any
letter a′ ∈ {∆(x)} \ {∆(y)} occurring after a in the word b. The letter a with such
property is called the (x, y)-separator if b does not contain any letter with the same
property after a.
Example 5.2.
If ∆(x) = a1 a2 a3 a5 , ∆(y) = a2 a4 a5 for some x, y ∈ Sb , b =
a1 a2 a3 a4 a5 then x ⊏6=
b y and a4 is the (x, y)-separator.
Obviously, ⊏6=
b is defined for all x, y ∈ Sb with distinct sets {∆(x)}, {∆(y)}.
The relation x ⊏b y holds iff
1. x ⊏6=
b y or
2. x ⊏= y for undefined x ⊏6=
b y.
Clearly, ⊏b is defined for any pair (x, y) ∈ Sb2 .
Lemma 5.3. If x ∈ Sb then ∆(x) ∆(b).
Proof. We have α(x)b = xb and Lemma 4.4 gives α(b) α(x). Therefore,
α(x)b = xb ⇔ α(x)∆(b) = α(x)∆(x)∆(b).
Let a be a letter from the word ∆(x). Hence, a should occur in the left part of the
last equality. Since α(x), ∆(x) have no common letters, we obtain a ∈ ∆(b). Thus,
∆(x) ∆(b).
Theorem 5.4. Any finite subband S ⊆ F has a local linear order.
Proof. By the definition of the local linear order, one should prove the following.
Let us prove that ⊏b satisfies all conditions of a local linear order.
8
1. ⊏b is a linear order. By the definition, ⊏b is irreflexive and asymmetric and
it is defined for any pair (x, y) ∈ Sb2 . Let us prove the transitivity and take
x ⊏b y, y ⊏b z. We have the following cases.
6=
6=
(a) x ⊏6=
b y, y ⊏b z. Let us prove x ⊏b z.
It follows there exist the (x, y)-separator a and the (y, z)-separator a′ .
By the definition of a separator, we have a ∈ {∆(y)} \ {∆(x)}, a′ ∈
{∆(z)} \ {∆(y)}.
Suppose a is before a′ in the word b (or a = a′ ). It follows that a′ ∈
/ ∆(x).
′
′′
Let us prove that a is the (x, z)-separator. Take an arbitrary a ∈ {∆(x)}
6=
′′
′′
after a′ in the word b. By x ⊏6=
b y, a ∈ {∆(y)}. Since y ⊏b z, a should
belong to {∆(z)}. Thus, a′ is the (x, z)-separator and x ⊏6=
b z.
6=
′
Suppose a is after a in the word b. Since y ⊏b z, then a ∈ {∆(z)}. Let
us prove that a is the (x, z)-separator. Take an arbitrary a′′ ∈ {∆(x)}
6=
′′
after a in the word b. By x ⊏6=
b y, we have a ∈ ∆(y). Since y ⊏b z
′′
′′
and a occurs in b after a, a should belong to {∆(z)}. Thus, a is the
(x, z)-separator, and therefore x ⊏6=
b z.
=
=
(b) x ⊏ y, y ⊏ z. It easily follows ∆(x) ≈ ∆(z), z <F y <F x. Thus,
x ⊏= z.
=
(c) x ⊏6=
b y, y ⊏ z. We have ∆(y) ≈ ∆(z). Let a be the (x, y)-separator.
Hence, there is not any a′ ∈ {∆(x)} \ {∆(y)} = {∆(x)} \ {∆(z)} after a
in the word b. Thus, a is the (x, z)-separator, and x ⊏6=
b z.
(d) x ⊏= y, y ⊏6=
b z. This case is similar to the previous one.
2. Prove (4). Take x, y with δν (x) ⊂ δν (y). By (7), we have ∆(x) ∆(y).
Assume ∆(x) ≺ ∆(y). By Lemma 5.3, all letters of ∆(y) belong to the word b.
Thus, the last letter a ∈ {∆(y)} \ {∆(x)} in the word b is the (x, y)-separator,
and x ⊏6=
b y.
If ∆(x) ≈ ∆(y) then (8) gives
y ≺ x ⇒ y <F x ⇒ x ⊏= y.
3. Prove (5). By Lemma 5.3, all letters of ∆(x), ∆(y) belong to the word b. Since
x α(y), the sets of letters {x}, {∆(y)} do not intersect. Let a (a′ ) be the last
letter from {∆(x)} (respectively, {∆(y)}) in the word b. If a is before a′ in b
we immediately obtain x ⊏b y. Let us assume the converse: ∆(b) = d1 a′ d2 ad3
for some di ∈ F. Using Lemma 4.4, we have α(b) α(x) and
α(x)b = xb ⇒ α(x)∆(b) = x∆(b) ⇒ α(x)d1 a′ d2 ad3 = xd1 a′ d2 ad3 .
The word α(x)d1 a′ does not contain a, and x does not contain a′ . Therefore,
the words α(x)d1 a′ , x are not ≤-comparable, and the last equality does not
hold.
4. Prove (6). Suppose x ⊏= y. Since the relation ⊏= does not depend on b, we
immediately obtain x ⊏c y. Let us put x ⊏6=
b y, and a is the (x, y)-separator.
By Lemma 4.4, α(b) α(x), α(b) α(y), and α(c) α(b) (since b ∈ Sb ⊆ Sc ).
Hence, {∆(x)} ∩ {α(b)} = {∆(y)} ∩ {α(b)} = ∅ and a ∈
/ {α(b)}. The equality
α(b)c = bc gives
α(b)∆(c) = α(b)∆(b)∆(c).
By Lemma 5.3, ∆(y) ∆(y) and therefore a ∈ ∆(b). Therefore, the left part
of the last equality should contain a, hence, a ∈ {∆(c)}. The relation x ⊏6=
c y
fails if there exists a letter a′ ∈ {∆(x)} \ {∆(y)} and a′ occurs after a in ∆(c).
9
′
By Lemma 5.3, a′ ∈ {∆(x)} ⊆ {∆(b)}. Since x ⊏6=
b y, a occurs before a in
the word b.
Thus, ∆(c) = c1 ac2 a′ c3 and ∆(b) = b1 a′ b2 ab3 for some ci , bi ∈ F and we have
α(b) ∼c b ⇔ α(b)c1 ac2 a′ c3 = α(b)b1 a′ b2 ab3 c1 c2 c3 .
The prefixes s1 = α(b)c1 a, s2 = α(b)b1 a′ are not ≤-comparable (s1 does not
contain a′ , s2 does not contain a), a contradiction.
Thus, we proved x ⊏c y.
6
Main result
Recall that F = {ai |i ∈ I} is the free left regular band of infinite rank with the
identity adjoined. The aim of this paper is to prove the next theorem.
Theorem 6.1.
equivalent:
For any finite left regular band S the following conditions are
1. S is embeddable into F;
2. S is right hereditary and has a local linear order.
Obviously, Theorem 6.1 immediately gives the following result for free left regular
bands of finite ranks.
Theorem 6.2.
equivalent:
For any finite left regular band S the following conditions are
1. S is embeddable into a free left regular band Fn of an appropriate rank n;
2. S is right hereditary and has a local linear order.
Let us come back to the proof of Theorem 6.1. The implication (1) ⇒ (2) follows
from Theorem 5.4. The implication (2) ⇒ (1) is quite complicated, so its proof is
divided into two parts. Below S is a finite right hereditary band with a local linear
order.
6.1
Orderings and homomorphisms
Let L = σ(S). One can initially assume that L is a subsemilattice of F. Thus,
below we consider elements of L as words of F. Let s ∈ S \ {ε} and
δ(s) = σ(s) \ σ(α(s)) ∈ F
(obviously, δ(s) coincides with δν (s) defined in (3) if we embed L into F via the
embedding ν).
For any c ∈ S we define a vector
χ(c) = (χ(α(c)), y1 , y2 , . . . , ync )
such that
1. χ(c) is a recursive structure, since it contains a subvector χ(α(c));
2. Sc = {y1 , y2 , . . . , ync };
3. yi ⊏c yi+1 ;
10
(11)
4. χ(ε) is the empty vector.
Example 6.3.
Let us show that the left regular band B from the previous
examples has a local linear order and compute the vectors χ(c) for each c ∈ Sc .
For elements of B we have δ(x1 ) = a1 , δ(x2 ) = a1 a2 , δ(x3 ) = a1 a2 a3 , δ(y0 ) = a2 ,
δ(y1 ) = a3 , δ(y2 ) = a3 (since ν embeds σ(B) into F we have δ(s) = δν (s) for each
s ∈ B).
One of the possible definitions of ⊏b is in the following table (if a set Sb consists
of a unique element the definition of ⊏b is omitted). Using the definition of ⊏b , one
can define vectors χ(b).
x
⊏x over Sx
χ(x)
x1
—
(x1 )
x2
y 0 ⊏ x2
(y0 , x2 )
x3 y0 ⊏ y1 ⊏ y2 ⊏ x3 (y0 , y1 , y2 , x3 )
y0
—
(χ(x1 ), y0 )
y1
y1 ⊏ y2
(χ(y0 ), y1 , y2 )
y2
y1 ⊏ y2
(χ(x2 ), y1 , y2 )
One can recursively apply (11) to α(c) and obtain a new vector
χ(c) = (χ(α(α(c))), x1 , x2 , . . . , xnα(c) , y1 , y2 , . . . , ync ).
It allows us to descend to any element a ≤ c obtaining the vector:
χ(c) = (χ(a), . . . , y1 , y2 , . . . , ync ).
(12)
We shall say y ∈ χ(c) if y ∈ Sc or y ∈ χ(α(c)).
Lemma 6.4. Let χ(c) be the vector (11). Then any element y with y ∼c α(y)
belongs to χ(c).
Proof. Let c′ ≤ c be the maximal element such that α(y) ≁α(c′ ) y and α(y) ∼c′ y
(such element always exists in the set {s ∈ S|ε < s ≤ c}). Therefore, y ∈ Sc′ and
y ∈ χ(c′ ) ⊆ χ(c).
Now we are able to define a map h : S → F of a right hereditary band S.
(
ε, if c = ε,
h(c) =
,
(13)
h(α(c))c1 c2 . . . cn , otherwise
−−→
−
→
where χ(c) = (χ(α(c)), y1 , y2 , . . . , yn ) and ci = δ(yi ) (the operator () was defined in
Section 3).
Example 6.5. Let us compute the images of h for elements of the left regular
band B defined in examples above. Let us fix an order over the free generators
a1 <F a2 <F a3 and obtain
−−→
δ(x)
x
h(x)
x1
a1
a1
x2
a1 a2
(a2 ) · (a1 a2 ) = a2 a1
x3 a1 a2 a3 (a2 ) · (a3 ) · (a3 )(a1 a2 a3 ) = a2 a3 a1
y0
a2
h(x1 ) · (a2 ) = a1 a2
y1
a3
h(y0 ) · (a3 ) · (a3 ) = a1 a2 a3
y2
a3
h(x2 ) · (a3 ) · (a3 ) = a2 a1 a3
11
One can directly check that h embeds B into F.
Since h(c) is recursively defined by the image h(α(c)) and h(α(c)) ≤ h(c), we
obtain the next simple statement.
Lemma 6.6. The map h preserves the order ≤: if a ≤ b then h(a) ≤ h(b).
Lemma 6.7. For any c ∈ S it holds σ(h(c)) = σ(c) (recall that both elements
σ(s), σ(h(s)) belong to F).
Proof. If c = ε the equality obviously holds. Assume the equality holds for the
ancestor α(c):
σ(h(α(c))) = σ(α(c))
Let χ(c) be the vector (11), and the image h(c) is defined by (13). Since c ∈ Sc ,
there exists an index i such that yi = c, hence
σ(h(c)) ⊇ σ(h(α(c))) ∪ δ(c) = σ(α(c)) ∪ δ(c) = σ(c).
Let us prove the contrary inclusion. By the choice of the elements yi , we have
yi c = α(yi )c, hence σ(yi ) ∪ σ(c) = σ(α(yi )) ∪ σ(c) and it follows
δ(yi ) ⊆ σ(c).
Thus,
σ(h(c)) = σ(h(α(c)))
[
δ(yi ) ⊆ σ(α(c)) ∪ σ(c) = σ(c).
i
We need the next auxiliary lemma.
Lemma 6.8.
chain
Let y, z, c ∈ S \ {ε}, z ∼c α(z), s = α(z)yz, s 6= α(z)y and in the
s = s0 > s1 > . . . > sk−1 > sk = α(z)y
each si is the ancestor of si−1 (1 ≤ i ≤ k). Then
1. si ∼c si−1 ;
2. σ(s) \ σ(α(z)y) ⊆ δ(z);
3. δ(si ) ⊆ σ(s) \ σ(α(z)y);
4. δ(si ) ⊆ δ(z);
5.
k
[
δ(si ) = σ(s) \ σ(α(z)α(b));
(14)
i=0
6. the equality δ(si ) = δ(z) implies k = 1;
7. suppose z ∈ Sc′ (it follows c′ ≤ c). Then si ∈ Sci for some ci ≤ c′ ;
8. if δν (si ) ⊂ δν (z) then si occurs before z in the vector χ(c).
Proof.
1. We have the chain
α(z)y = sk < sk−1 < . . . < s0 = α(z)yz ≤ α(z)yzc = α(z)yα(z)c = α(z)yc.
By Lemma 4.2, si ∼c α(z)y and therefore si ∼c si−1 .
12
2. Suppose a letter a belongs to σ(α(z)yz) \ σ(α(z)y) ∈ F. Then a ∈ σ(z) but
a∈
/ σ(α(z)). Thus, a ∈ δ(z).
3. Consider a letter a ∈ δ(si ). Hence, a ∈ σ(si ) ⊆ σ(s), a ∈
/ σ(si+1 ). Since
si+1 ≥ α(z)y, we have a ∈
/ σ(α(z)y). Thus, a ∈ σ(s) \ σ(α(z)y).
4. The statement immediately follows from the previous ones.
5. We have σ(si−1 ) = δ(si−1 ) ⊔ σ(si ) (x ⊔ y is the join of words x, y ∈ F which
do not have common letters). Thus,
σ(s0 ) = δ(s0 ) ⊔ σ(s1 ) = δ(s0 ) ⊔ δ(s1 ) ⊔ σ(s2 ) = . . . =
k
[
δ(si ) ⊔ σ(α(z)y)
i=0
and we obtain (14).
6. Immediately follows from (14) and previous statements of the lemma.
7. We have z ∼c′ α(z). One can repeat the proof of the first statement of the
lemma for c′ instead of c and obtain si ∼c′ si−1 . Therefore, si ∈ Sci for some
ci ≤ c′ ≤ c
8. By the previous statement of the lemma, each si belongs to Sci for some ci ≤ c′ .
Therefore, the element si (z) belongs to the vector χ(ci ) (respectively, χ(c′ )).
If ci < c′ , then the contents of χ(ci ) occur before the contents of χ(c′ ) in the
vector χ(c), and therefore si occur before z in the vector χ(c). If ci = c′ then,
according to (4), the condition δν (si ) ⊂ δν (z) implies si ⊏c′ z. Therefore, si
occurs before z in the vector χ(c′ ). Since the vector χ(c′ ) is included in χ(c),
si occurs before z in χ(c).
Recall that Lemma 6.8 holds for any embedding of σ(S) into F. If we replace all
occurrences of δ to δν we obtain true statements about the embedding ν.
Lemma 6.9.
S.
The map h : S → F is a homomorphism of a right hereditary band
Proof. Using Lemmas 3.2, 6.7, it is sufficient to prove the implication
a ∼c b ⇒ h(a) ∼h(c) h(b).
(15)
α(b) ∼c b ⇒ h(α(b)) ∼h(c) h(b).
(16)
First, we prove
By the definition of h, one can write
h(b) = h(α(b))b1 b2 . . . bn ,
h(c) = c1 c2 . . . cm
(in the last expression we iterate the definition of h(c) to the identity ε), and there
exist vectors
χ(b) = (χ(α(b)), y1 , y2 , . . . , yn ),
χ(c) = (z1 , z2 , . . . , zm ),
−−→
−−−→
such that bi = δ(yi ), cj = δ(zj ).
Lemma 6.4 states zj ∼c α(zj ) (1 ≤ j ≤ m), yi ∼b α(yi ) (1 ≤ i ≤ n). Using
Lemmas 4.2, 4.4, we have the following chain
α(yi ) ≤ yi ≤ α(yi )b ≤ α(yi )bc = (α(yi )α(b))c = α(yi )c.
13
By Lemma 4.2, it follows α(yi ) ∼c yi . Thus, the vector χ(c) contains all yi . In
other words, there exists a function f : {1, 2, . . . , n} → {1, 2, . . . , m} which embeds
{y1 , y2 , . . . , yn } into χ(c) (yi = zf (i) ). By (6), the function f is monotone: i1 < i2 →
f (i1 ) < f (i2 ).
Let z ∈ χ(c) and below we prove that one of the following holds:
1.
δ(z) ⊆
[
δ(zj ) ∪ σ(α(b)),
(17)
j
where elements zj occur in χ(c) before z;
2. z is after b in χ(c);
3. z belongs to χ(b), i.e. z = yi or z ∈ χ(α(b)).
Assume for some z ∈ χ(c) all conditions above fail. Let s = α(z)α(b)z. If
s = α(z)α(b) then δ(z) ⊆ σ(α(b)), and (17) holds. Therefore, we assume below
s > α(z)α(b), and we have a nontrivial chain
s = s0 > s1 > . . . > sk > α(z)α(b),
where each si is the ancestor of the previous one. By Lemma 6.8 we have si ∼c α(si ),
and all si belong to the vector χ(c) (Lemma 6.4).
We have the following cases.
1. Let s = z and denote s′ = α(z)b (we use z ∼c α(z) below):
α(z)α(b)z = z ⇒ α(z)α(b)zc = zc ⇒ α(z)α(b)α(z)c = zc ⇒ α(z)(α(b)c) = zc
⇒ α(z)bc = zc ⇒ s′ c = zc.
Since S is right hereditary, the elements z, s′ are ≤-comparable. Thus, we have
the following cases.
(a) If z ≤ s′ then
α(z) ≤ z ≤ α(z)b
and Lemma 4.2 gives α(z) ∼b z. Thus, z should occur in the χ(b), a
contradiction.
(b) Let s′ = α(z)b < z. By the definition of the ancestor (there are not
elements between α(z) and z), we obtain s′ = α(z). Thus, b α(z).
By (5), we have b ⊏c z, therefore z occurs after b in χ(c), a contradiction
with the choice of z.
2. Let s 6= z and Lemma 6.8 gives δν (s) ⊆ δν (z). If we assume δν (s) = δν (z) then
α(z) ≈ α(s) α(z)α(b) (Lemma 2.3) and we have α(b) α(z). Therefore
s = (α(z)α(b))z = α(z)z = z,
a contradiction.
Thus, δν (s) ⊂ δν (z), and Lemma 6.8 provides δν (si ) ⊂ δν (z). By the last
statement of Lemma 6.8, all si occur before z in χ(c). Let
c = σ(α(b)) ∪
k
[
i=0
By (14), c = σ(α(b)) ∪ (σ(s) \ σ(α(z)α(b))).
14
δ(si ).
Let us prove c ⊇ δ(z). Suppose a letter a belongs to the word δ(z). Therefore,
a ∈ σ(z), a ∈
/ σ(α(z)). If a ∈ σ(α(b)) we immediately obtain a ∈ c. Otherwise
(a ∈
/ σ(α(b))), we have
a ∈ σ(z) \ σ(α(z)α(b)) ⇒ a ∈ c.
Thus, we have proved (17) that contradicts the choice of z.
Let us come back to the equality (16). One should prove
h(α(b))c1 c2 . . . cm = h(α(b))b1 b2 . . . bn c1 c2 . . . cm .
(18)
−−→ −−−−→
Since b ∈ χ(b) and b ∈ χ(c), there exists an index i such that δ(yi ) = δ(zf (i) ) =
−−→
δ(b). For each zk , k < f (i) we have one of the following
1. zk ∈ χ(b) (probably, zk ∈ χ(α(b)));
2. zk satisfies (17).
In the second case all letters of ck are eliminated by letters of previous elements
and one can remove such ck from (18). If zk ∈ χ(α(b)) then all letters of ck occur
in h(α(b)) (Lemma 6.7) and ck can be removed from (18).
Thus one can reduce (18) to
−−→
h(α(b))b1 b2 . . . bi−1 δ(b)cf (i)+1 cf (i)+2 . . . cm =
−−→
−−→
h(α(b))b1 b2 . . . bi−1 δ(b)bi+1 . . . bn b1 b2 . . . bi−1 δ(b)cf (i)+1 cf (i)+2 . . . cm ⇔
−−→
−−→
h(α(b))b1 b2 . . . bi−1 δ(b)cf (i)+1 cf (i)+2 . . . cm = h(α(b))b1 b2 . . . bi−1 δ(b)bi+1 . . . bn cf (i)+1 cf (i)+2 . . . cm .
(19)
Relations α(yi ) ∼b yi implies
σ(yi ) ∪ σ(b) = σ(α(yi )) ∪ σ(b) ⇒ δ(yj ) ⊆ σ(b) = σ(α(b)) ∪ δ(b) = σ(h(α(b))) ∪ δ(b).
Therefore, the elements bi+1 , bi+2 , . . . , bn in (19) are eliminated by letters of
−−→
h(α(b)), δ(b), and (19) becomes a true equality:
−−→
−−→
h(α(b))b1 b2 . . . bi−1 δ(b)cf (i)+1 cf (i)+2 . . . cm = h(α(b))b1 b2 . . . bi−1 δ(b)cf (i)+1 cf (i)+2 . . . cm .
Thus, we have proved (16).
Let us prove (15) using (16). Since S is right hereditary, a < b (similarly, one
can assume b < a). Therefore, there exists a chain
b = b0 > b1 > . . . > br = a,
where bi = α(bi−1 ). Since a ≤ bi ≤ b ≤ ac, Lemma 4.2 gives a ∼c bi . The
transitivity of ∼c provides bi+1 ∼c bi . Above we proved h(bi+1 ) ∼h(c) h(bi ), and the
transitivity of ∼h(c) gives h(a) ∼h(c) h(b).
15
6.2
Improving the homomorphism
The homomorphism h : S → F defined above is not necessarily an embedding of S
into F. In other words, it is possible
ker(h) = {(s1 , s2 )|s1 6= s2 , h(s1 ) = h(s2 )} =
6 ∅.
Example 6.10. Let us define a right hereditary band B ′ = {x1 , x2 , x3 , y1 , y2 , y3 }
by the following multiplication table (the identity ε is omitted):
x1 x2 x3 y 1 y 2 y 3
x1 x1 x1 y 1 y 1 y 2 y 3
x2 x2 x2 y 3 y 3 y 3 y 3
x3 x3 x3 x3 x3 x3 x3
y1 y1 y1 y1 y1 y1 y1
y2 y2 y2 y2 y2 y2 y2
y3 y3 y3 y3 y3 y3 y3
The order ≤ over B ′ is defined by:
y3
y1
y2
x
2
✑
✁
✑
✁ ✑
❅
❅ ✁✑✑
ε❅✁✑
x1
x3
Fig. 3
The support semilattice of B ′ is a linear ordered semilattice 1 < 2 < 3, where
σ(ε) = 1, σ(x1 ) = σ(x2 ) = 2, σ(x3 ) = σ(y1 ) = σ(y2 ) = σ(y3 ) = 3.
The embedding ν gives:
σ(x1 ) = σ(x2 ) = a1 , σ(x3 ) = σ(y1 ) = σ(y2 ) = σ(y3 ) = a1 a2 .
−−→
Thus, the embedding ν embeds σ(B ′ ) into F. We have (computing δ(x), we use
a1 <F a2 ):
−−→
x
Sx
⊏x
χ(x)
h(x)
δ(x)
x1
{x1 }
—
a1
(x1 )
a1
x2
{x2 }
—
a1
(x2 )
a1
x3 {x3 , y1 , y3 } y1 ⊏ y2 ⊏ x3 a1 a2
(y1 , y2 , x3 )
a2 a1
y1
{y1 , y3 }
y1 ⊏ y3
a2
(χ(x1 ), y1 , y3 ) a1 a2
y2
{y2 , y3 }
y2 ⊏ y3
a2
(χ(x1 ), y2 , y3 ) a1 a2
y3
{y2 , y3 }
y2 ⊏ y3
a2
(χ(x2 ), y2 , y3 ) a1 a2
Thus, we have a nonempty kernel
ker(h) = {(x1 , x2 ), (y1 , y2 ), (y1 , y3 ), (y2 , y3 )}.
Below we define a series of homomorphisms h1 , h2 , . . . , hm such that
ker(h) ⊃ ker(h1 ) ⊃ ker(h2 ) ⊃ . . . ⊃ ker(hm ) = ∅.
using the next algorithm.
16
Embedding algorithm
INPUT: a homomorphism h : S → F.
OUTPUT: an embedding h′ : S → F.
1. Put h′ = h, and F ′ is the free left regular band generated by letters {ai | ∃s ∈
S, ai ∈ h(s)}.
2. Take c ∈ S such that b1 , b2 , . . . , bk are all descendants of c (α(bi ) = c), and
h′ (bi ) ≤ h′ (bj ) for some distinct i, j;
Apply Modification(h′ , F ′ , c, b1 , b2 , . . . , bk ) (see below) and obtain a new homomorphism h′ and a new free left regular band F ′ .
3. If there is not any c ∈ S from the previous step, we terminate the algorithm.
Modification(h, F, c, b1 , b2 , . . . , bk )
INPUT: a homomorphism h : S → F , free left regular band F , c ∈ S, and all descendants b1 , b2 , . . . , bk of c.
OUTPUT: a homomorphism h′ : S → F ′ , where F ′ ⊇ F is a free left regular band, and h′
is defined as follows.
Let C ⊆ F be the set of all letters a ∈
/ h(c). Let F ′ be the free left regular band
generated by F and new free generators
{t∅c , tbc1 , tbc2 , . . . , tbck }.
Let
t′c = t∅c tbc1 tbc2 . . . tbck ,
Hc = {s ∈ S | h(s) contains a letter from C},
Scbi = {s | ∃y ≥ bi cs = ys},
!
[
Scbi .
Sc∅ = Hc \
i
First, one should prove the following statement.
Lemma 6.11. For all i 6= j Scbi ∩ Scbj = ∅ and Scbi ⊆ Hc .
Proof. Assume s ∈ Scbi ∩ Scbj . It follows that there exist yi , yj such that bi ≤ yi ,
bj ≤ yj , cs = yi s, cs = yi s. Therefore, yi s = yj s and yi , yj should be ≤-comparable.
However it is impossible, since S is right hereditary.
Let us prove Scbi ⊆ Hc . Any s ∈ Scbi satisfies h(c)h(s) = h(y)h(s) for some
y ≥ bi > c. If assume s ∈
/ Hc then h(s) h(c) ≤ h(y) and the last equality becomes
h(c) = h(y), a contradiction.
Now we are ready to define a new mapping h′ by the following substitutions in
images of h:
(
h(s), if s ∈
/ Hc ,
h′ (s) =
h(s)|a:=atxc t′c , for all a ∈ C if s ∈ Scx (x ∈ {b1 , b2 , . . . , bk , ∅})
17
After the definition of h′ Modification algorithm stops.
Example 6.12. Let us explain the work of Embedding and Modification algorithms at the semigroup B ′ .
First, we take c = ε. Then C = {a1 , a2 }, b1 = x1 , b2 = x2 , b3 = x3 , Sεx1 =
{x1 , y1 , y2 }, Sεx2 = {x2 , y3 }, Sεx3 = {x3 }, Sε∅ = ∅.
By the Modification algorithm, we obtain t′ε = t∅ε txε 1 txε 2 txε 3 . Thus, h′ (x1 ) =
x
a1 tε 1 t∅ε txε 2 txε 3 , h′ (x2 ) = a1 txε 2 t∅ε txε 1 txε 3 ,
h′ (x3 ) = a2 txε 3 t∅ε txε 1 txε 2 a1 txε 3 t∅ε txε 1 txε 2 = a2 txε 3 t∅ε txε 1 txε 2 a1 ,
h′ (y1 ) = h′ (y2 ) = a1 txε 1 t∅ε txε 2 txε 3 a2 txε 1 t∅ε txε 2 txε 3 = a1 txε 1 t∅ε txε 2 txε 3 a2 ,
h′ (y3 ) = a1 txε 2 t∅ε txε 1 txε 3 a2 txε 2 t∅ε txε 1 txε 3 = a1 txε 2 t∅ε txε 1 txε 3 a2 .
We have ker(h′ ) = {(y1 , y2 )}, so one should repeat an iteration of Embedding
algorithm. We take c = x1 and obtain C = a2 b1 = y1 , b2 = y2 , Sx1 y1 = {y1 , x3 }
(since x1 x3 = y1 x3 ), Sx1 y2 = {y2 }, Sx1 ∅ = {y3 }.
Modification algorithm now gives: t′x1 = t∅x1 tyx11 tyx21 , h′′ (x1 ) = h′ (x1 ), h′′ (x2 ) =
h′ (x2 ), h′′ (x3 ) = a2 tyx11 t∅x1 tyx21 txε 3 t∅ε txε 1 txε 2 a1 , h′′ (y1 ) = a1 txε 1 t∅ε txε 2 txε 3 a2 tyx11 t∅x1 tyx21 ,
h′′ (y2 ) = a1 txε 1 t∅ε txε 2 txε 3 a2 tyx21 t∅x1 tyx11 , h′′ (y3 ) = a1 txε 2 t∅ε txε 1 tεx3 a2 t∅x1 tyx11 tyx21 .
Thus, we obtained an embedding h′′ : S → F after two iterations of Embedding
algorithm.
Lemma 6.13. Let B = {b1 , b2 , . . . , bk , ∅}.
1. Let us consider x, y ∈ S such that x ∈ Scd for some d ∈ B. Then
h′ (x)h′ (y) = h(x)|a:=atdc t′c h(y),
(20)
where a is the first letter of h(x) from the set C.
2. Suppose x, y ∈ S satisfy h(x) < h(y). The inequality h′ (x) < h′ (y) fails only
if x ∈ Scd , y ∈ Sce for distinct e, d ∈ B.
3. Suppose x, y, z ∈ S satisfy h(x) ∼h(z) h(y), h(x) ≤ h(y) and h′ (x) ≤ h′ (y).
The relation h′ (x) ∼h′ (z) h′ (y) fails only if y ∈ Scd , z ∈ Sce for distinct e, d ∈ B
and x ∈
/ Hc .
Proof.
1. Let a′ ∈ C be a letter of the word h(x)h(y) which occurs after a. By
the definition of the map h′ , the letter a (a′ ) will be replaced to the expression
a′ tdc t′c for some d ∈ B (respectively, a′ tec t′c for some e ∈ B). However the
subword tec t′c is eliminated by tdc t′c which occurs earlier. Thus, t-letters will be
occur only after a in h′ (x)h′ (y) and we immediately obtain (20).
2. If y ∈
/ Hc then x ∈
/ Hc (since h(x) < h(y)) and h′ (x) = h(x), h′ (y) = h(y). In
this case the inequality h′ (x) < h′ (y) holds.
Since the case x ∈
/ Hc , y ∈ Hc is impossible, it is sufficient to consider the case
x ∈ Hc , i.e. x ∈ Scd for some d ∈ B. Since h(x) < h(y), we have y ∈ Hc and
x ∈ Sce for some e ∈ B. If d = e then the inequality h′ (x) < h′ (y) remains
true. Thus, d 6= e.
3. If x ∈ Hc then x ∈ Scf for some f ∈ B. Since h′ (x) ≤ h′ (y), y belongs to the
same set Scf . Let a be the first letter from h(x) with a ∈ C (since h(x) ≤ h(y),
a is the first letter from C in the word h(y) = h(x)w). By (20) we have
h′ (x)h′ (z) = h(x)|a:=atf t′ h(z),
c c
18
h′ (y)h′ (z) = h(y)|a:=atf t′ h(z) = h(x)|a:=atf t′ wh(z),
c c
c c
and the equality
h(x)|a:=atf t′ h(z) = h(x)|a:=atf t′ wh(z)
c c
c c
obviously holds, since we have the condition h(x)h(z) = h(x)wh(z).
Thus, below we assume x ∈
/ Hc , i.e. h(x) does not contain letters from C and
′
hence h (x) = h(x). Suppose y ∈ Hc , i.e. y ∈ Scd for some d ∈ B. We have
that the right part of the equality h(x)h(z) = h(y)h(z) contains letters from
C. Therefore, h(z) should contain a letter from C, i.e. z ∈ Sce for some e ∈ B.
Let a ∈ C (a′ ∈ C) be the first letter in the word h(y) (respectively, h(z)).
By (20), we have
h′ (y)h′ (z) = h(y)|a:=atdc t′c h(z),
h′ (x)h′ (z) = h(x)h(z)|a′ :=a′ tec t′c .
It is easy to see that for d = e the equality h′ (x)h′ (z) = h′ (y)h′ (z) holds, and
we came to the contradiction. Thus, d 6= e.
It is sufficient to consider the case x, y ∈
/ Hc , i.e. h′ (x) = h(x), h′ (y) = h(y).
′
If assume z ∈
/ Hc then h (z) = h(z), and the equality h′ (x)h′ (z) = h′ (y)h′ (z)
holds. Thus, we put z ∈ Hc , i.e. z ∈ Sce for some e ∈ B. Therefore,
h′ (x)h′ (z) = h(x)h(z)|a:=atec t′c ,
h′ (y)h′ (z) = h(y)h(z)|a:=atec t′c
for some letter a ∈ C of the word h(z). Obviously, we obtain the equality
h′ (x)h′ (z) = h′ (y)h′ (z) (since h(x)h(z) = h(y)h(z)), a contradiction.
Let us prove that the map h′ defined in Modification procedure is a homomorphism.
Lemma 6.14. If x < y (x, y ∈ S) then
h′ (x) < h′ (y)
(21)
Proof. Since h is a homomorphism and h satisfies Lemma 6.7, we have
h(x) < h(y)
(22)
According to Lemma 6.13, the relation (21) fails only if x ∈ Scd , y ∈ Sce , where
e, d ∈ {b1 , b2 , . . . , bk , ∅} and e 6= d.
1. Let x ∈ Scbi , i.e. there exists x′ ≥ bi such that
cx = x′ x.
Let us multiply the last equality by y and obtain cy = x′ y. Thus, y ∈ Scbi .
2. Suppose y ∈ Scbi . There exists y ′ ≥ bi such that cy = y ′ y. Since x ≤ y, we
have cxy = y ′ xy and the elements cx, y ′ x are ≤-comparable.
For the inequality cx ≥ y ′ x we have
y ′ xcx = cx ⇔ y ′ x = cx (we use y ′ ≥ c) ⇒ x ∈ Scbi .
The relation cx ≤ y ′ x implies (we use α(bi ) = c and y ′ x ≥ bi ) either cx = c or
cx ≥ bi . We have:
19
(a) cx ≥ bi , therefore bi x = cx and x ∈ Scbi ;
(b) cx = c (x c), then the word h(x) does not contain any letter from
C. Hence, x ∈
/ Hc that gives a contradiction with x ∈ Scd (Lemma 6.11
provides x ∈ Hc ).
Lemma 6.15. h′ is a homomorphism.
Proof. Following Lemma 3.2, it is sufficient to check that h′ preserves the relations
, ∼x .
Actually h′ is a substitution of letters in images of h. Hence, it obviously preserves -relation. Let us prove
h′ (x) ∼h′ (z) h′ (y)
(23)
if x ∼z y. Below we assume x ≤ y. Since h is a homomorphism, we have
h(x) ∼h(z) h(y),
(24)
h′ (x) ≤ h′ (y).
(25)
where h(x) ≤ h(y).
By Lemma 6.14, we have
According to Lemma 6.13, the relation (23) fails only if the word h(x) does not
contain letters from C, but y ∈ Scd , z ∈ Sce for distinct d, e ∈ {∅, b1 , b2 , . . . , bk }.
By the definition of C, we have
x c.
Suppose y ∈ Scbi . It follows there exist y ′ ≥ bi such that cy = y ′ y. Multiplying
the last equality by z, we obtain
cyz = y ′ yz.
Applying x ∼y z and x c ≤ y ′ , we have
cyz = y ′ yz ⇒ cxz = y ′ xz ⇒ cz = y ′ z.
The last equality states that z ∈ Scbi and (23) holds.
Suppose now z ∈ Scbi . There exists z ′ ≥ bi such that cz = z ′ z. Since x c ≤ z ′ ,
we have
cz = z ′ z ⇒ cxz = z ′ xz ⇒ cyz = z ′ yz.
Since S is right hereditary, the elements cy, z ′ y are ≤-comparable.
1. If z ′ y ≤ cy, then (below we use c z ′ )
z ′ y ≤ cy ⇒ z ′ yc = cy ⇒ z ′ y = cy.
The last equality gives y ∈ Scbi , and (23) holds.
2. If z ′ y ≥ cy, then (we use z ′ y ≥ bi ) either cy = c or cy ≥ bi . The first alternative
gives y c. It follows that h(y) does not contain new letters t∅c , tbc1 , tbc2 , . . . , tbck ,
and (23) holds.
Let us put cy ≥ bi then
cy ≥ bi ⇔ (bi c)y = cy ⇔ bi y = cy ⇒ y ∈ Scbi
and (23) holds.
20
Lemma 6.16.
Let bi , bj ∈ S (i =
6 j) be elements from Modification procedure.
′
′
Then the elements h (bi ), h (bj ) are not ≤-comparable.
Proof. Clearly, we have bi ∈ Scbi , bj ∈ Scbj , and the words h(bi ), h(bj ) are h(bi ) =
h(c)awi , h(bj ) = h(c)a′ wj for some wi , wj ∈ F and some letters a, a′ (probably,
a = a′ ). The images h′ (bi ), h′ (bj ) are
b
h′ (bi ) = h(c)atbci t′c h′ (wi ), h′ (bj ) = h(c)a′ tcj t′c h′ (wj ),
b
and h′ (bi ), h′ (bj ) have ≤-incomparable prefixes h(c)atbci , h(c)atcj . Thus, the words
h′ (bi ), h′ (bj ) are ≤-incomparable.
By the definition of h′ it immediately follows the next statement.
Lemma 6.17. If h(x), h(y) are not ≤-comparable, so are h′ (x), h′ (y).
By Lemmas 6.16, 6.17, it follows
ker(h) ⊃ ker(h′ ).
(26)
Let us come back to the definition of Embedding Algorithm.
Let h′i be the homomorphism defined by the Modification algorithm at i-th
iteration of Embedding algorithm. By (26), we have
ker(h) ⊃ ker(h′1 ) ⊃ ker(h′2 ) ⊃ . . .
(27)
Since the left regular band S is finite, the chain (27) terminates after finite steps.
It means that Embedding algorithm terminates after finite number of iterations.
Lemma 6.18. Let m be the last iteration of Embedding algorithm, then ker(h′m ) =
∅, i.e. h′m is an embedding of S into F ′ .
Proof. Let us take distinct s1 , s2 with s1 < s2 . Lemma 6.7 gives h(s1 ) < h(s2 ), and
Lemma 6.14 provides h′m (s1 ) < h′m (s2 ) .
Let us take two ≤-incomparable elements s1 , s2 . Therefore, there exist the maximal element c with c < si , and distinct b1 , b2 such that bi ≤ si , α(bi ) = c.
By the termination condition of Embedding algorithm, we have that the words
′
hm (b1 ), h′m (b2 ) are not ≤-comparable. Thus, the words h′m (s1 ), h′m (s2 ) have incomparable prefixes h′m (b1 ), h′m (b2 ), therefore h′m (s1 ) 6= h′m (s2 ).
7
Open problems and further research
7.1 Algebraic geometry: equations with no constants.
How to simplify our proof
Theorem 6.1 has a natural application in the study of equations over F. Let us
remind main definitions of universal algebraic geometry (for more details see [9, 10,
11]).
Let X = {x1 , x2 , . . . , xn } be a finite set of variables, and F(X) be the free left
regular band generated by the set X. The elements of F(X) are called terms.
21
An equality of two terms t(X) = s(X) is called an equation. For example, the
next expressions are equations x1 x2 = x2 x1 , x1 x2 x3 = x3 x4 .
A set Y ⊆ F n is called algebraic if there exists a system of equations in variables
X with the solution set Y . An algebraic set is irreducible if it is not a proper finite
union of algebraic sets.
For any algebraic set Y one can define a left regular band Γ(Y ) which is called
the coordinate band of Y (see [9] for more details). Coordinate bands are analogues
of coordinate rings in commutative algebra, so the description of all algebraic sets
over F may be approximated by the description of their coordinate bands. Notice
that any coordinate band Γ(Y ) is always finitely generated, so all Γ(Y ) over F are
finite (the variety of left regular bands is locally finite).
In [3] submonoids of face monoids were studied. However, applying results of [9]
to [3], one can easily obtain the description of all coordinate bands over F.
Theorem 7.1.
band S:
The following conditions are equivalent for a finite left regular
1. S is a coordinate band of an algebraic set over F;
2. S ∈ Qvar(F);
3. S satisfies the formulas Qn (see [4]);
4. S is embedded into a direct power of the semigroup H = {+, −, 0} defined by:
+ − 0
+ + + +
− − − −
0 + − 0
Actually Theorem 7.1 was proved in [3] with the second condition S ∈ Qvar(H),
but one can prove Qvar(F) = Qvar(H) and obtain the theorem above.
A coordinate band which corresponds to an irreducible algebraic set is called
irreducible. The study of irreducible coordinate bands is important, since any algebraic set over F is a finite union of algebraic ones (it follows from the Noetherian
property of F, see [9] for details).
Theorem 7.1 describes all coordinate bands over F, but the class of irreducible
coordinate bands is described by the following theorem.
Theorem 7.2.
band S:
The following conditions are equivalent for a finite left regular
1. S is a coordinate band of an irreducible algebraic set over F;
2. S is discriminated by F;
3. S ∈ Ucl(F);
4. S is right hereditary and has a local linear order;
5. S is embedded into F.
Proof. The implications (1) ⇔ (2) ⇔ (3) follows from the general results of [9]. The
implication (4) ⇔ (5) is the statement of Theorem 6.1.
Let us prove (2) ⇒ (5). By the definition, a band S is discriminated by F if
for any finite set {s1 , s2 , . . . , sn } ⊆ S there exists a homomorphism h : S → F such
that h(si ) 6= h(sj ) for all i 6= j.
Since any finitely generated left regular band is finite, we have S =
{s1 , s2 , . . . , sn }. By the condition, there exists a homomorphism h : S → F which
is injective on S, i.e. h is an embedding.
22
The implication (5) ⇒ (2) clearly follows from the definition of the discrimination. Thus, the theorem is proved.
Let us explain the class Ucl(F). Let ϕ be a first-order sentence (formula) of the
multiplicative language L = {·}. A formula ϕ is universal if it is equivalent to
∀x1 ∀x2 . . . ∀xn ϕ′ (x1 , x2 , . . . , xn ),
where ϕ′ is quantifier-free. The universal closure of F consists of all left regular
bands S such that S |= ϕ for any universal formula ϕ with F |= ϕ.
According to Theorem 7.2, the embedability of a finite band into F can be
defined by a set of universal formulas. Therefore, the right hereditary and local
linear order properties can be written as universal formulas. A universal formula
defining the right hereditary property is simple:
∀x1 , x2 , x3 (x1 ≤ x3 )&(x2 ≤ x3 ) → (x1 ≤ x2 ∨ x2 ≤ x1 ).
Problem 7.3. Find a set Σ of universal formulas in the language L = {·} such
that the following holds: a finite right hereditary band has a local linear order iff it
satisfies Σ.
We guess the positive solution of Problem 7.3 allows to simplify the proof of
Theorem 6.1. In particular, we will be able to prove it with no the properties of the
embedding ν.
7.2 Algebraic geometry: equations with constants. Advantages of the complicated proof
The current proof of Theorem 6.1 is complicated but it has an advantage. In
Lemma 6.9 we found a homomorphism h : S → F with σ(h(s)) = σ(s), where
σ(s), σ(h(s)) ∈ F. Let us show that this homomorphism property is useful in the
study of equations with constants.
Let X = {x1 , x2 , . . . , xn } be a finite set of variables, and F[X] be the free left
regular band generated by F and X. Elements of F[X] are called F-terms.
An equality of two F-terms t(X) = s(X) is called an F-equation. For example,
the next expressions are F-equations x1 a1 x2 a2 a3 = a3 x2 a4 a1 x1 a2 , x1 x2 a1 x3 a2 a3 =
x3 a4 x4 , x1 x2 = a1 a2 a3 .
One can naturally define algebraic sets and coordinate bands for systems of
equations with constants. Following [9], any coordinate band of such algebraic set
contains a subband isomorphic to F (subband of constants), and any homomorphism
between coordinate bands Γ1 , Γ2 should map F ⊆ Γ1 onto F ⊆ Γ2 (constants are
mapped into constants).
Thus, we need homomorphisms which act trivially on F. Fortunately, the homomorphism h has the required property, and we hope h will play a crucial role in
the solution of the next problem.
Problem 7.4.
Describe coordinate bands of irreducible algebraic sets over F
defined by equations with constants.
7.3
“Almost free” left regular bands
Following Theorem 7.2, the universal closure Ucl(F) consists of right hereditary
bands with a local linear order. According to model theory, Ucl(F) is the “logical
23
neighbourhood” of F, i.e. a structure of each S ∈ Ucl(F) is highly close to F. So
we formulate
Problem 7.5. Prove theorems about the structure of bands from Ucl(F) (right
hereditary bands with a local linear order) and solve uniformly algorithmic problems
(e.g. the word problem) for each band from Ucl(F).
We hope this problem should be solved, since we have a similar result in group
theory. In [12] it was proved the analogue of Theorem 7.2 for a free non-abelian
group F .
Theorem 7.6.
group G:
The following conditions are equivalent for a finitely generated
1. G is a coordinate group of an irreducible algebraic set over F ;
2. G is fully residually free;
3. G ∈ Ucl(F );
4. G is embedded into Lyndon group F Z[t] .
There are results about the structure and algorithms for fully residually groups
(see [13, 14]), and it follows that all groups from Ucl(F ) inherits many properties
of the free group F . Thus, we hope all semigroups from Ucl(F) have many nice
properties.
7.4
Approximation with left regular bands
Many objects (e.g. hyperplane arrangements, matroids) are approximated by left
regular bands (see [1]). In other words, any hyperplane arrangement H (matroid
M ) admits a multiplication such that H (respectively, M ) becomes a left regular
band. Below we shall talk only about matroids, but one can easily reformulate the
following problem for hyperplane arrangements and other objects.
Let S(M ) be the left regular band corresponding to a matroid M . Following [1],
S(M ) of a free matroid of rank n is the free left regular band of the same rank.
Free matroids have a simple structure (the free matroid Mn of rank n is isomorphic
to all subsets of n linearly independent vectors). Thus, it is interesting to describe
matroids which are “almost free”.
Problem 7.7. What can we say about a matroid M with S(M ) ∈ Ucl(F) (i.e.
S(M ) is right hereditary with a local linear order)?
References
[1] K. S. Brown, Semigroups, rings, and Markov chains, J. Theoret. Probab., 13(3),
2000, 871-938.
[2] S. Margolis, F. Saliola, B. Steinberg, Poset topology and homological invariants
of algebras arising in algebraic combinatorics, arXiv:1311.6680.
[3] S. Margolis, F. Saliola, B. Steinberg, Semigroups embeddable in hyperplane
monoids, Semigroup Forum, August 2014, Volume 89, Issue 1, pp 236-248.
[4] S. Margolis, F. Saliola, B. Steinberg, Combinatorial topology and the global
dimension of algebras arising in combinatorics, J. Eur. Math. Soc., to appear.
24
[5] F. Saliola, The quiver of the semigroup algebra of a left regular band, Internat.
J. Algebra Comput., 17(8), 2007, 1593–1610.
[6] F. Saliola, The face semigroup algebra of a hyperplane arrangement, Canad. J.
Math., 61(4), 2009, 904–929.
[7] F. Saliola, The quiver of the semigroup algebra of a left regular band, Int. J.
Algebra Comput., 17(8), 2007, 1593–1610.
[8] F. Saliola, The face semigroup algebra of a hyperplane arrangement, Can. J.
Math. 61(4), 2009, 904–929.
[9] E. Daniyarova, A. Miasnikov, V. Remeslennikov, Unification theorems in algebraic geometry, Algebra and Discrete Mathematics, 1, 2008, 80–111.
[10] E. Yu. Daniyarova, A. G. Myasnikov, V. N. Remeslennikov, Algebraic geometry
over algebraic structures. II. Foundations, Fundam. Prikl. Mat., 17:1, 2012, 65–
106
[11] E. Daniyarova, A. Miasnikov, V. Remeslennikov, Algebraic geometry over algebraic structures III: equationally Noetherian property and compactness, Southeast Asian Bulletin of Mathematics, 2011, 35:1, 35-68
[12] O. Kharlampovich, A. Myasnikov, Irreducible affine varieties over free group
II: Systems in trangular quasi-quadratic form and description of residually free
groups, J. Algebra, 200(2), 1998, 517-570.
[13] O. Kharlampovich, A. Myasnikov, V. Remeslennikov, D. Serbin, Subgroups of
fully residually free groups: algorithmic problems, Contemp. Math. series of
the AMS, Group theory, Statistics and Cryptography, 360, 2004, 63–101.
[14] I. Bumagin,O. Kharlampovich, A. Miasnikov, Isomorphism problem for fully
residually free groups , J. Pure and Applied Algebra, 208 (3), 2007, 961–977.
The information of the author:
Artem N. Shevlyakov
Sobolev Institute of Mathematics
644099 Russia, Omsk, Pevtsova st. 13
Phone: +7-3812-23-25-51.
e-mail: a [email protected]
25
| 4math.GR
|
arXiv:1611.06393v1 [math.GR] 19 Nov 2016
THE RELATIVE EXPONENTIAL GROWTH RATE OF
SUBGROUPS OF ACYLINDRICALLY HYPERBOLIC GROUPS
EDUARD SCHESLER
Abstract. In this paper we prove that the relative exponential growth rate
of any subgroup H of a finitely generated group G exists with respect to every
finite generating set of G if H contains a generalized loxodromic element. We
will prove that the relative exponential growth rate of any subexponentially
distorted embedding of an acylindrically hyperbolic group H in a finitely generated group G exists with respect to every finite generating set of G. We
introduce a new invariant of finitely generated groups, the ambiguity function,
and we show that every finitely generated acylindrically hyperbolic group G
possesses a linearly bounded ambiguity function. We also extend these results
to the case of direct products of finitely generated acylindrically hyperbolic
groups.
Contents
1. Introduction
2. Weak Supermultiplicativity
3. Hyperbolic spaces
4. Acylindrically hyperbolic groups
5. Existence of the growth rate
6. The case of products
References
1
3
7
8
9
14
18
1. Introduction
Given a finitely generated group G and a finite symmetric generating set X
X
of G, the growth function βG
: N Ñ N of G with respect to X is defined by
X
X
X
βG pnq “ |BG pnq|, where BG pnq denotes the set of all elements of G which are
represented by words of length ď n in the generators of X. By definition, G has
X
exponential growth if there is a constant a ą 1 such that βG
pnq ą an for every
n P N. Using the submultiplicativity
X
X
X
βG
pm ` nq ď βG
pmqβG
pnq,
1
X
Milnor proved in [M` 68] that the limit lim βG
pnq n exists for every finitely genernÑ8
ated group G and every finite generating set X of G. Thus, if this limit a is larger
than 1, then for every ε ą 0, there is a number N P N such that for every n ě N
we have
(1.1)
X
pa ´ εqn ď βG
pnq ď pa ` εqn .
Date: NOVEMBER 19, 2016.
1
2
EDUARD SCHESLER
This justifies the notion of exponential growth. In this paper we study the following
relative analogue of the growth function.
Definition 1.1. Let G be a finitely generated group with a finite generating set
X and let H ď G be an arbitrary subgroup of G. Then
X
BH
pnq “ th P H : |h|X ď nu
(1.2)
X
X
denotes the dX -ball of radius n P N inside of H. We write βH
pnq “ |BH
pnq| for
X
the cardinality of this ball. The function βH : N Ñ N is called the relative growth
function of H in G with respect to X.
1
X
Again we can ask for the existence of the limit of the sequence pβH
pnq n qnPN .
1
X
n
If the limit lim βH pnq exists we will say that the relative exponential growth rate
nÑ8
of H in G exists with respect to X. Grigorchuk [Gri80] proved that the relative
exponential growth rate of every normal subgroup of a finitely generated free group
F exists with respect to a free generating set of F . This result was extended by
Olshanski [Ols13, Theorem 1.5] by showing that the relative exponential growth
rate of any subgroup of a finitely generated free group F exists with respect to a
free generating set of F . Further Olshanski [Ols13, Remark 3.1] pointed out that
the relative exponential growth rate does not exist in general. He showed that there
is a finitely generated solvable group G and an infinite cyclic subgroup H ď G such
that
(1.3)
lim inf
nÑ8
X
βH
pnq
“ 0 @ε ą 0 and
1`ε
n
1
X
lim sup βH
pnq n ą 1.
nÑ8
The main result of this paper is the following theorem.
Theorem 5.9. Let G be a finitely generated group and H ď G a subgroup of
G. If H contains a generalized loxodromic element h P H, then the relative exponential growth rate of H in G exists with respect to every finite generating set of G.
In Section 2 we introduce a new invariant of a finitely generated group, the ambiguity function, that seems to be useful in studying of the relative exponential growth
rate. We prove the following theorem.
Theorem 5.10. Let G be a finitely generated acylindrically hyperbolic group. Then
G possesses a linearly bounded ambiguity function with respect to every finite generating set of G.
We deduce from this theorem the following result.
Theorem 5.11. Let H be a finitely generated acylindrically hyperbolic group. Let
ι : H Ñ G be a subexponentially distorted embedding in a finitely generated group
G. Then the relative exponential growth rate of H in G exists with respect to every
finite generating set of G.
In Section 6 we give an extension of Theorem 5.10 to direct products of acylindrically hyperbolic groups.
THE RELATIVE EXPONENTIAL GROWTH RATE OF SUBGROUPS OF ACYLINDRICALLY HYPERBOLIC GROUPS
3
Theorem 6.6. Let G “
m
ś
Gi be a finite direct product of finitely generated
i“1
acylindrically hyperbolic groups Gi and let H be a subgroup of G. Suppose that,
for each i “ 1, . . . , m, the image of the canonical projection of H to Gi contains a
generalized loxodromic in Gi . Then the relative exponential growth rate of H in G
exists with respect to every finite generating set X of G.
2. Weak Supermultiplicativity
In this section we give analytical tools which will be used later in the proof of
existence of the relative exponential growth rate. The distortion function of a subgroup, recalled in Definition 2.2, is a well known invariant of a finitely generated
subgroup. In Definition 2.3, we introduce the ambiguity function of a finitely generated group with respect to a finite generating set. In Corollaries 2.12 and 2.13 we
give a sufficient condition for the existence of the relative exponential growth rate
in terms of the distortion and the ambiguity functions of a subgroup.
Definition 2.1. A function f : N Ñ N is called sublinear, if lim
nÑ8
subexponential, if
lim f pnq
n
nÑ8 a
f pnq
n
“ 0 and
“ 0 for every a ą 1.
Definition 2.2. Let G be a finitely generated group with a finite generating set X.
Let H be a finitely generated subgroup of G with a finite generating set Y . The
distortion function of H in G with respect to X and Y is defined by
(2.1)
∆G
H : N Ñ N, n ÞÑ maxt|h|Y : h P H, |h|X ď nu.
The subgroup H is called linearly, polynomially, or subexponentially distorted in G,
if ∆G
H pnq is bounded, respectively, by a linear, polynomial, or subexponential function. A linearly distorted subgroup H of G is also said to be an undistorted subgroup
of G.
It is easy to see that the property of a subgroup H ď G to be linearly, polynomially, or subexponentially distorted in G does not depend on a choice of finite
generating sets of H and G.
The following maps (indexed by m, n P N) are important in studying of exponential growth rates:
X
X
X
pnq Ñ BG
pm ` nq,
Φm,n : BG
pmq ˆ BG
pu, vq ÞÑ uv.
The main difficulty is that, given natural numbers m, n P N and an element w P
X
BG
pm ` nq, there may be many different representations of the form w “ uv
X
X
with u P BG
pmq and v P BG
pnq. It would be useful to give a subexponential
upper bound on the cardinality of the preimage Φ´1
m,n pwq in terms of n only, i.e.,
independent of m and w. However, such a bound does not exist even in the case
of free groups Fk with basis X. Indeed, consider the preimage of the identity
´1
1 P BFXk pm ` nq. Then Φ´1
q P BFXk pmq ˆ BFXk pnq : |u|X ď m, nu
m,n p1q “ tpu, u
´1
X
and thus |Φ2n,n p1q| “ |BFk pnq| “ 2kp2k ´ 1qn´1 has exponential growth with
respect to n. This also shows that we cannot guarantee that |uv|X is approximately
|u|X ` |v|X .
In [Ols13, Theorem 1.5] Olshanski considers the free group case and suggests to
modify the map Φm,n by insertion a word from a ball BFXk pcq between u and v. In
the following definition we promote this idea in the general case.
4
EDUARD SCHESLER
Definition 2.3. Let G be a finitely generated group with a finite generating set
X. A function f : N Ñ N is an ambiguity function of G with respect to X, if there
is a finite subset M Ă G and a map of the form
Φ : G ˆ G Ñ G, pu, vq ÞÑ ugu,v v
with gu,v P M , which, with notation below, satisfies the following property (˚):
Denote c :“ max |g|X , and, for every m, n P N, consider the restriction:
gPM
X
X
X
Φm,n : BG
pmq ˆ BG
pnq Ñ BG
pm ` n ` cq, pu, vq ÞÑ Φppu, vqq.
X
Property p˚q: For every m, n P N and every z P BG
pm ` n ` cq, we have
|Φ´1
m,n pzq| ď f pnq.
The set M is called the set of connecting pieces and the map Φ is called the
concatenation map corresponding to f .
Remark 2.4. It is easily seen that, up to equivalence, an ambiguity function of a
group G does not depend on the choice of the finite generating set of G. Indeed,
let f be an ambiguity function of G with respect to a finite generating set X of G.
Let Φ be the concatenation map and M the set of connecting pieces corresponding
to f . Let c :“ max |g|X . For an arbitrary finite generating set Y of G we define
gPM
p “ max |y|X . Then the inclusion
yPY
Y
Y
X
X
BG
pmq ˆ BG
pnq Ă BG
ppmq ˆ BG
ppnq
is well defined and by definition the preimage of
X
X
X
Φpm,pn : BG
ppmq ˆ BG
ppnq Ñ BG
ppm ` pn ` cq
Y
is bounded by f ppnq. Thus the preimage of the restriction of Φpm,pn to BG
pmq ˆ
Y
BG pnq is also bounded by f ppnq. Therefore the function g : N Ñ N, n ÞÑ f ppnq is
an ambiguity function of G with respect to Y .
We use the following simple remarks.
Remark 2.5. Let X and Y be two sets and f : X Ñ Y a function such that the
cardinality of every preimage f ´1 pyq with y P Y is uniformly bounded by a number
n P N. Then we have |X| ď n ¨ |Y |.
Remark 2.6. Let f : N Ñ N be a subexponential function and let g : N Ñ N be a
polynomially bounded function. Then the composition g ˝ f is subexponential.
Remark 2.7. Let f : N Ñ N be a linearly bounded function and let g : N Ñ N be a
subexponential function. Then the composition g ˝ f is subexponential.
Proposition 2.8. Let G be a finitely generated group with a finite generating set
X and let H ď G be a finitely generated, subexponentially distorted subgroup with a
finite generating set Y . If H possesses a polynomially bounded ambiguity function
with respect to Y , then there is a subexponential function ε : N Ñ N and a constant
c P N such that
(2.2)
X
X
X
βH
pmqβH
pnq ď εpnq ¨ βH
pm ` n ` cq for all m, n P N.
THE RELATIVE EXPONENTIAL GROWTH RATE OF SUBGROUPS OF ACYLINDRICALLY HYPERBOLIC GROUPS
5
Proof. Let f be a subexponential distortion function of H in G with respect to X
and Y . Then the trivial inclusion
(2.3)
X
X
Y
Y
ι : BH
pmq ˆ BH
pnq Ñ BH
pf pmqq ˆ BH
pf pnqq.
is well defined. Let P be a polynomially bounded ambiguity function of H with
respect to Y , let M Ă H be the finite set of connecting pieces and let Φ : HˆH Ñ H
be the concatenation map corresponding to P . We set p
c “ max |h|Y . By restricting
hPM
Y
Y
Φ to subsets of the form BH
pmq ˆ BH
pnq, we obtain the maps
(2.4)
Y
Y
Y
Φm,n : BH
pmq ˆ BH
pnq Ñ BH
pm ` n ` p
cq pu, vq ÞÑ Φppu, vqq “ ugu,v v
Y
such that gu,v P M and |Φ´1
cq and every
m,n pzq| ď P pnq for every z P BH pm ` n ` p
pm, nq P N ˆ N. In particular we get
(2.5)
|Φ´1
f pmq,f pnq pzq| ď P pf pnqq
Y
for all m, n P N and every z P BH
pf pmq ` f pnq ` p
cq. By combining (2.3) and (2.5)
we note that the cardinality of every preimage Ψ´1
m,n pzq of
(2.6)
Y
X
X
pmq ˆ BH
pnq Ñ BH
pf pmq ` f pnq ` p
cq
Ψm,n :“ Φf pmq,f pnq ˝ ι : BH
is bounded by P pf pnqq. Now let c “ max |h|X be the diameter of M in ΓpG, Xq.
hPM
The definition of Φ immediately implies that
(2.7)
X
impΨm,n q Ă BH
pm ` n ` cq.
Thus the cardinality of the preimage of an arbitrary element with respect to the
map
(2.8)
X
X
X
Ψm,n : BH
pmq ˆ BH
pnq Ñ BH
pm ` n ` cq
is bounded by the function εpnq :“ P ˝ f pnq. Due to Remark 2.6 this function is
subexponential and the desired inequalities follow with the help of Remark 2.5.
If the distortion function f is polynomial the ambiguity function g is subexponential, then the composition? ε “ g ˝ f does not have to be subexponential. Just
consider the case of gpnq “ 2 n and f pnq “ n2 .
In the case that the subgroup H is undistorted, we get the following analogue of
Proposition 2.8.
Proposition 2.9. Let G be a finitely generated group and let H ď G be a finitely
generated, undistorted subgroup of G. If H possesses a subexponential ambiguity
function, then there exists a subexponential function ε : N Ñ N and a constant
c ě 0 such that
(2.9)
X
X
X
pmqβH
pnq ď εpnq ¨ βH
pm ` n ` cq for all m, n P N.
βH
The proof is essentially the same as the proof of Proposition 2.8. One only has
to replace Remark 2.6 by Remark 2.7 in the argumentation.
The following lemma provides us with an alternative definition of a subexponential function which will be easier to apply for our purposes.
Lemma 2.10. A function f : N Ñ N is subexponential if and only if
(2.10)
1
lim f pnq n “ 1.
nÑ8
6
EDUARD SCHESLER
Proof. First we assume that f is subexponential. Then for every a ą 1 and every
1
ď ε for all but finitely many n P N. Due to lim ε n “ 1 we
ε ą 0 we have fapnq
n
nÑ8
1
1
therefore obtain f pnq n ď pan εq n for all but finitely many n P N. Since we can chose
1
a ą 1 and ε ą 0 arbitrary close to 1 and 0, respectively, we get lim f pnq n “ 1.
nÑ8
1
Now we assume lim f pnq n “ 1. Thus, for every ε ą 0 we have f pnq ď p1 ` εqn
nÑ8
for all but finitely many n P N. Let δ ą 0 and a “ 1 ` δ ą 1. By setting ε “
obtain
´ 1 ` δ ¯n
f pnq
p1 ` εqn
2
(2.11)
lim sup n ď lim sup
“
lim
sup
“ 0.
a
an
1`δ
nÑ8
nÑ8
nÑ8
f pnq
n
nÑ8 a
Since f pnq ą 0 we get lim
“ 0.
δ
2
we
The crucial analytical ingredient for the proof of the existence of some relative
exponential growth rates is provided by the following proposition. This is essentially
the analytical part of the proof of Lemma 1.6 in [Ols13]. For convenience we shall
include a proof here.
Proposition 2.11. Let B, C ą 0 be two real numbers, ε : N Ñ N a subexponential
function and l : N Ñ N a sublinear function. If f : N Ñ N is a monotonically
increasing function with the properties
1) f pmqf pnq ď εpnqf pm ` n ` lpnqq for all n ě C and all m P N and
2) 1 ď f pnq ď B n for all n P N,
1
then the limit lim f pnq n exists.
nÑ8
1
Proof. First we see that lim sup f pnq n exists due to the second property of f .
nÑ8
1
Without loss of generality we may assume that a :“ lim sup f pnq n ą 1. Indeed,
nÑ8
since f pnq ě 1 for every n P N we would obtain
1
1
1 ď lim inf f pnq n ď lim sup f pnq n ď 1
(2.12)
nÑ8
nÑ8
1
otherwise. Let δ P p0, a ´ 1q and let an :“ f pnq n . With Lemma 2.10 we get
´1
1
lim εpsq s “ 1. Thus we obtain lim εpsq s “ 1. Since l is sublinear, we further
sÑ8
sÑ8
get
s ` lpsq
s lpsq
“ lim `
“ lim 1 ` 0 “ 1.
sÑ8
sÑ8 s
sÑ8
s
s
So there is a natural number s ě C such that
s`lpsq
´
´1
δ
δ
2δ ¯ s
ăa´ .
(2.14)
|as εpsq s ´ a| ď and a ´
3
3
3
In the following, a natural number n will be written an n “ qps ` lpsqq ` r with
r ă s ` lpsq and q P N. Since f is monotonically increasing and
´
2δ ¯n´s´lpsq
(2.15)
a´
ą pa ´ δqn
3
holds for all but finitely many n P N we have
(2.13)
lim
THE RELATIVE EXPONENTIAL GROWTH RATE OF SUBGROUPS OF ACYLINDRICALLY HYPERBOLIC GROUPS
7
f pnq ě f pqps ` lpsqqq
ě εpsq´1 f psqf ppq ´ 1qps ` lpsqqq
ě εpsq´q`1 f psqq´1 f ps ` lpsqq
ě ...
ě εpsq´q f psqq .
´1
1
The definition an “ f pnq n gives us εpsq´q f psqq “ pεpsq s as qsq . In view of the
inequalities above this gives us
´
´
¯sq
´
´1
δ ¯sq
2δ ¯ps`lpsqqq
f pnq ě as εpsq s
ě a´
ě a´
3
3
´
´
2δ ¯n´r
2δ ¯n´s´lpsq
“ a´
ą a´
ą pa ´ δqn .
3
3
Altogether we obtain an ą a ´ δ for almost every n P N. Since δ was chosen
arbitrarily small, the claim follows.
Due to Proposition 2.11 and the two Propositions 2.8 and 2.9 we get immediately
the two following corollaries, which we use later to show the existence of some
relative exponential growth rates.
Corollary 2.12. Let G be a finitely generated group and H ď G a finitely generated, subexponentially distorted subgroup of G. If H possesses a polynomially
bounded ambiguity function, then the relative exponential growth rate of H in G
exists with respect to every finite generating set of G.
It would be interesting to know, whether there is an undistorted subgroup H of
a finitely generated group G such that the relative exponential growth rate of H
in G does not exist with respect to some finite generating set of G. The following
corollary gives an obstruction to this case.
Corollary 2.13. Let G be a finitely generated group and H ď G a finitely generated
undistorted subgroup of G. If H possesses a subexponential ambiguity function, then
the relative exponential growth rate of H in G exists with respect to every finite
generating set of G.
3. Hyperbolic spaces
In this section we summarize some properties of hyperbolic metric spaces. We
follow the book of Bridson and Haefliger [BH99, III.H]. First we recall the definition
of Gromov product.
Definition 3.1. Let pΩ, dq be a metric space and let o, x, y P Ω be arbitrary points
of Ω. The Gromov product of x and y at o is defined by
(3.1)
px.yqo “
1
pdpx, oq ` dpy, oq ´ dpx, yqq.
2
Definition 3.2 (Gromov). A metric space pΩ, dq is called δ-hyperbolic for some
constant δ ě 0 if all points o, x, y, z P Ω satisfy
(3.2)
px.yqo ě minppx.zqo , pz.yqo q ´ δ.
Further pΩ, dq is said to be hyperbolic if pΩ, dq is δ-hyperbolic for some δ. In the
following, the condition (3.2) will be referred to as the Gromov condition.
8
EDUARD SCHESLER
Definition 3.3. Let pΩ, dq be a metric space and I Ă R an (not necessary bounded)
interval. Let λ ě 1 and ε ě 0 be some constants. A map c : I Ñ Ω is called pλ, εqquasigeodesic, if
1 1
(3.3)
|t ´ t| ´ ε ď dpcpt1 q, cptqq ď λ|t1 ´ t| ` ε
λ
for every t1 , t P I. We can replace I by I X Z to get a discrete version of a pλ, εqquasigeodesic which will be used later.
Definition 3.4. Let pΩ, dq be a metric space. For any subset X Ă Ω and any
number ε ą 0 we denote by Vε pXq the ε-neighbourhood of X in Ω. Given two
subsets A, B Ă Ω, the Hausdorff-distance between them is defined by
(3.4)
HdpA, Bq :“ inftε P Rě0 | A Ă Vε pBq, B Ă Vε pAqu.
Although geodesics in hyperbolic spaces are not uniquely determined by their
endpoints in general, we will denote a geodesic segment connecting to points x, y of a
hyperbolic space pΩ, dq by rx, ys. The following proposition gives us a legitimization
for that.
Proposition 3.5. [BH99, III.H.1.7] For all δ ą 0, λ ě 1 and ε ě 0 there is a
constant K “ Kpδ, λ, εq with the following property: Let pΩ, dq be a δ-hyperbolic
space and let c be a finite pλ, εq-quasigeodesic in Ω. Let rp, qs be a geodesic segment
connecting the endpoints of c. Then the Hausdorff-distance between rp, qs and the
image of c is bounded by K.
Using this result, it is not hard do derive the following lemmas.
Lemma 3.6. Let pΩ, dq be a δ-hyperbolic space and let γ : ra, bs Ñ Ω be a pλ, εqquasigeodesic. Then for every m P ra, bs the Gromov product pγpaq.γpbqqγpmq is
bounded by Kpδ, λ, εq.
Lemma 3.7. Let pΩ, dq be a δ-hyperbolic space and let x, y, z, w P Ω such that
(3.5)
px.zqy ď C1 , py.wqz ď C2 and N :“ dpy, zq ą C1 ` C2 ` δ.
Then every geodesic rx, ws runs through the p2C1 ` 5δ ` 1q-neighbourhood of y.
4. Acylindrically hyperbolic groups
In this chapter we recall some known results on acylindrically hyperbolic groups
which will be used in the next section.
Definition 4.1. Let G be a group which acts by isometries on a metric space pΩ, dq.
The action is called acylindrical, if for every ε ą 0 there are constants R, N ą 0
such that for every two points x, y P Ω with dpx, yq ě R, there are at most N
elements g P G such that dpx, gxq ď ε and dpy, gyq ď ε.
Remark 4.2. Let G be a group which acts acylindrically on a metric space pΩ, dq.
Then the induced action of every subgroup K ď G on Ω is acylindrical.
Definition 4.3. Let G be a group which acts on a hyperbolic space pΩ, dq. An
element g P G is called
‚ elliptic, if the set tg n x | n P Zu is bounded for some (equivalently, any)
x P Ω.
THE RELATIVE EXPONENTIAL GROWTH RATE OF SUBGROUPS OF ACYLINDRICALLY HYPERBOLIC GROUPS
9
‚ loxodromic, if the map Z Ñ Ω, n ÞÑ g n x is a quasigeodesic for some
(equivalently, any) x P Ω.
Recall that two loxodromic elements are independent if their fixed point sets in
the Gromov boundary BΩ of Ω are disjoint. Osin [Osi16, Theorem 1.1] classifies
acylindric actions on hyperbolic spaces as follows:
Theorem 4.4. Let G be a group which acts acylindrically on a hyperbolic space
pΩ, dq. Then exactly one of the following statements is true:
1) G has bounded orbits.
2) G is virtually cyclic and contains a loxodromic element.
3) G contains infinitely many independent loxodromic elements.
Definition 4.5. A group G is called acylindrically hyperbolic if G acts acylindrically
on a hyperbolic space pΩ, dq such that G contains infinitely many independent
loxodromic elements.
The following theorem is a part of a result obtained by Osin [Osi16, Theorem
1.4].
Theorem 4.6. Let G be a group and let g P G be an arbitrary element. The
following conditions are equivalent.
1) There exists a (possibly infinite) generating set X of G such that the corresponding Cayley graph ΓpG, Xq is hyperbolic, the natural action of G on
ΓpG, Xq is acylindrical and g is loxodromic.
2) There exists an acylindrical action of G on a hyperbolic space pΩ, dq such
that g is loxodromic.
In general in may happen that there are two non elementary acylindrical actions
of a group G on two hyperbolic spaces such that an element g P G is loxodromic
for exactly one of these actions. For our purposes we need an element which is
loxodromic for some action. Such elements are called generalized loxodromic, see
[Osi16, Definition 6.4].
Definition 4.7. Let G be an arbitrary group. An element g P G is called generalized loxodromic, if there is an acylindrical action of G on a hyperbolic space pΩ, dq,
such that g P G is a loxodromic element for this action.
5. Existence of the growth rate
We start with the following simple observations which follow right from the
definitions.
Lemma 5.1. Let G be a group which acts by isometries on a metric space pΩ, dq.
Then for every g P G and every x, y, z P Ω we get
(5.1)
px.yqz “ pgx.gyqgz .
Lemma 5.2. Let G be a group with a generating set Y such that ΓpG, Y q is hyperbolic and the canonical action of G on ΓpG, Y q is acylindrical. Let g, h P G be
two independent loxodromic elements. Then there is a constant C ě 0, such that
for every n P Z and every two elements x, y P tg n , g ´n , hn , h´n u with x ‰ y ˘1 the
inequality px.yq1 ď C holds.
10
EDUARD SCHESLER
Lemma 5.3. Let G be a group with a generating set Y such that ΓpG, Y q is hyperbolic and the canonical action of G on ΓpG, Y q is acylindrical. Let g, h P G be
two independent loxodromic elements. Then there is a constant A ě 0 such that
for every n P N and every pair of elements pu, vq P G ˆ G there is at least one
xu,v P tg n , g ´n , hn , h´n u such that
(5.2)
pu´1 .xu,v q1 ď A and pv.x´1
u,v q1 ď A.
Proof. First we observe that there is a constant C0 such that
(5.3)
mintpg n .uq1 , pg ´n .uq1 u ď C0 and mintphn .uq1 , ph´n .uq1 u ď C0
for every n P N and every u P G.
Indeed, since the maps n ÞÑ hn and n ÞÑ g n are quasiisometric embeddings of Z
into G, Lemma 3.6 implies that there is a constant C1 such that pg ´n .g n q1 , ph´n .hn q1 ď
C1 for every n P Z. The Gromov condition implies
(5.4)
C1 ě ph´n .hn q1 ě mintph´n .uq1 , pu.hn q1 u ´ δ.
Thus the claim follows by setting C0 :“ C1 ` δ.
Next, there is a constant C2 ě 0, such that for every n P N and every u P G at
least one of the following statements holds.
1) pg n .uq1 ď C2 and pg ´n .uq1 ď C2 .
2) phn .uq1 ď C2 and ph´n .uq1 ď C2 .
Indeed, due to Lemma 5.2 there is a constant C3 ě 0 such that the Gromov product
px.yq1 of two elements x, y P tg n , g ´n , hn , h´n u with x ‰ y ˘1 is bounded by C3 .
The Gromov condition gives us
C3 ě pg n .hn q1 ě mintpg n .uq1 , pu.hn q1 u ´ δ,
C3 ě pg ´n .hn q1 ě mintpg ´n .uq1 , pu.hn q1 u ´ δ,
C3 ě pg n .h´n q1 ě mintpg n .uq1 , pu.h´n q1 u ´ δ,
C3 ě pg ´n .h´n q1 ě mintpg ´n .uq1 , pu.h´n q1 u ´ δ.
If pg n .uq1 ą C3 ` δ or pg ´n .uq1 ą C3 ` δ we thus obtain
(5.5)
C3 ` δ ě phn .uq1 and C3 ` δ ě ph´n .uq1 .
In the same way if phn .uq1 ą C3 ` δ or ph´n .uq1 ą C3 ` δ we obtain
(5.6)
C3 ` δ ě pg n .uq1 and C3 ` δ ě pk ´n .uq1 .
The claim now follows by setting C2 :“ C3 ` δ.
To prove the lemma, let u´1 , v P G be arbitrary. Then, on the one hand there is
an element x P tg n , g ´n , hn , h´n u such that
(5.7)
px.u´1 q1 ď C2 and px´1 .u´1 q1 ď C2 .
And on the other hand we have mintpx.vq1 , px´1 .vq1 u ď C0 . Without loss of generality we may assume that px´1 .vq1 ď C0 . Then the statement of the lemma follows
by setting A :“ maxtC0 , C2 u and xu,v :“ x.
We are now ready to define the concatenation map and the set of connecting
pieces corresponding to the predicted linear ambiguity function for acylindrically
hyperbolic groups.
THE RELATIVE EXPONENTIAL GROWTH RATE OF SUBGROUPS OF ACYLINDRICALLY HYPERBOLIC GROUPS
11
Definition 5.4. Let G be a finitely generated group with a finite generating set X
and a generating set Y such that ΓpG, Y q is hyperbolic and the canonical action of
G on ΓpG, Y q is acylindrical. Let g and h be two independent loxodromic elements
and let
(5.8)
cn :“ maxt|g n |X , |g ´n |X , |hn |X , |h´n |X u for every n P N.
Due to Lemma 5.3 there exists a constant A such that for every n P N there is a
map
Φn : G ˆ G Ñ G, pu, vq ÞÑ uxu,v v
(5.9)
with xu,v P tg , g , h , h´n u and maxtpu´1 .xu,v q1 , pv.x´1
u,v q1 u ď A for every pair
X
X
of elements pu, vq P G ˆ G. By restriction to subsets of the form BG
psq ˆ BG
ptq
we obtain well defined maps
n
(5.10)
´n
n
X
X
X
BG
psq ˆ BG
ptq Ñ BG
ps ` t ` cn q, pu, vq ÞÑ uxu,v v.
To keep the notation simple, these restrictions are also denoted by Φn .
The construction of Φn depends on some specific loxodromic elements g and h.
If it is clear from the context to which elements Φn corresponds to, we will not
mention this correspondence explicitly.
Proposition 5.5. Let G be a group with a generating set Y such that ΓpG, Y q is
hyperbolic and the canonical action of G on ΓpG, Y q is acylindrical. Let g, h P G be
two independent loxodromic elements. Then there are two constants N P N and ε ě
0 such that for every natural number n ě N every geodesic r1, Φn pu, vqs Ă ΓpG, Y q
runs through the ε-neighbourhoods of u and uxu,v with respect to dY .
Proof. We want to apply Lemma 3.7 on the points 1, u, uxu,v and uxu,v v. Let
A be the constant from definition 5.4. Then Lemma 5.1 implies p1.uxu,v qu “
pu´1 .xu,v q1 ď A and
pu.uxu,v vquxu,v
“ ppuxu,v q´1 u.puxu,v q´1 uxu,v vqpuxu,v q´1 uxu,v
“ ppxu,v q´1 .vq1 ď A.
Since g and h are loxodromic, we can choose N big enough to satisfy
(5.11)
mint|g n |Y , |g ´n |Y , |hn |Y , |h´n |Y u ą 2A ` δ
for every n P Z with |n| ě N . Thus we obtain
(5.12)
dY pu, uxu,v q “ |xu,v |Y ą 2A ` δ.
Setting ε “ 2A ` 5δ ` 1, Lemma 3.7 implies that every geodesic r1, uxu,v vs runs
through the ε-neighbourhood of u. Symmetric argumentation shows that every
geodesic ruxu,v v, 1s runs through the ε-neighbourhood of uxu,v .
The following lemma implies in particular that finite changes of generating sets
X and Y of a group G don’t affect the acylindricity of the action of G on its Cayley
graph with respect to X and Y .
Lemma 5.6. [Osi16, Lemma 5.1] For any group G and any generating sets X and
Y of G such that
(5.13)
sup |x|Y ă 8 and sup |y|X ă 8,
xPX
the following hold.
yPY
12
EDUARD SCHESLER
1) ΓpG, Xq is hyperbolic if and only if ΓpG, Y q is hyperbolic.
2) G acts acylindrically on ΓpG, Xq if and only if G acts acylindrically on
ΓpG, Y q.
Proposition 5.7. Let G be a finitely generated group with a finite generating set X
and a generating set Y such that ΓpG, Y q is hyperbolic and the canonical action of
G on ΓpG, Y q is acylindrical. Let g and h be two independent loxodromic elements.
Then there is a number n and a linearly bounded function ln : N Ñ N such that for
X
every pair of natural numbers ps, tq P N ˆ N and every element w P BG
ps ` t ` cn q,
the cardinality of the preimage of w with respect to the concatenation map
X
X
X
psq ˆ BG
ptq Ñ BG
ps ` t ` cn q
Φn : BG
(5.14)
is bounded by ln ptq.
Proof. Due to Lemma 5.6 we can assume that X is a subset of Y . Further Proposition 5.5 provides us with a number N P N and a constant ε ą 0, such that for
every number n ě N and every pair of elements pu, vq P G ˆ G, every geodesic
r1, Φn pu, vqs runs through the ε-neighbourhoods of the elements u and uxu,v with
respect to dY .
Since the action of G on ΓpG, Y q is acylindrical, we can choose positive constants
R and L such that for every two points x, y P ΓpG, Y q with dY px, yq ě R there are at
most L elements g P G satisfying the inequalities dY px, gxq ď 6ε and dY py, gyq ď 6ε.
The definition of a loxodromic element implies immediately that there is a number
Θ ě N such that for every number n ě Θ, and every x P Mn “ tg n , g ´n , hn , h´n u
we have |x|Y ě R. From now on we fix such a number n ě Θ. Let s, t P N and
X
w P BG
ps ` t ` cn q. We want to bound the cardinality of the set of all pairs
X
X
pu, vq P BG
psq ˆ BG
ptq which satisfy the equation Φn pu, vq “ w. For this we fix a
geodesic r1, ws in ΓpG, Y q.
X
X
Let u P BG
psq be an element such that there exists an element v P BG
ptq with
the property Φn pu, vq “ w. Then there are elements z and ζ on the geodesic
r1, Φn pu, vqs “ r1, ws such that dY pu, zq ď ε and dY puxu,v , ζq ď ε. We obtain
dY pu, wq “ dY pu, Φn pu, vqq “ dY pu, uxu,v vq
“ dY p1, xu,v vq ď dX p1, xu,v q ` dX p1, vq
ď t ` cn .
Thus we get
(5.15)
dY pz, wq ď dY pz, uq ` dY pu, wq ď ε ` t ` cn .
So there are at most ε`t`cn points z on the fixed geodesic r1, ws with the property
dY pu, zq ď ε. Now let z P G be an arbitrary point on the geodesic r1, ws. We want
X
Y
to estimate the cardinality of the set of elements u P BG
psq X BG
pz, εq that satisfy
X
the equation Φn pu, vq “ w for some v P BG ptq.
To get a contradiction we assume that there are more than 4 ¨ L such elements u.
By definition Mn contains 4 elements. Thus Dirichlet’s box principle implies that
X
X
there are L ` 1 pairs pu1 , v1 q, . . . , puL`1 , vL`1 q P BG
psq ˆ BG
ptq satisfying xui ,vi “
x0 P Mn and Φn pui , vi q “ w for every i P t1, . . . , L ` 1u. Let i, j P t1, . . . , L ` 1u.
Due to the triangle inequality we get the inequalities
dY pui , uj q ď dY pui , zq ` dY pz, uj q ď 2ε.
THE RELATIVE EXPONENTIAL GROWTH RATE OF SUBGROUPS OF ACYLINDRICALLY HYPERBOLIC GROUPS
13
We fix elements ζi , ζj P r1, ws such that dY pζi , ui x0 q, dY pζj , uj x0 q ď ε. Then
dY pz, ζi q ď dY pz, ui q ` dY pui , ui x0 q ` dY pui x0 , ζi q ď 2ε ` |x0 |Y
and
dY pz, ζi q ě dY pui , ui x0 q ´ dY pz, ui q ´ dY pui x0 , ζi q ď |x0 |Y ´ 2ε.
Analogously we obtain
|x0 |Y ´ 2ε ď dY pz, ζj q ď |x0 |Y ` 2ε
Since ζi and ζj lie on the same geodesic r1, ws, we obtain dY pζi , ζj q ď 4ε. Finally
this implies
dY pui x0 , uj x0 q ď dY pui x0 , ζi q ` dY pζi , ζj q ` dY pζj , uj x0 q ď 6ε.
By defining ûj :“ u´1
j u1 , the obtained inequalities can be rewritten as
dY pûj , 1q ď 2ε and dY pûj x0 , x0 q ď 6ε
for every j P t1, . . . , L ` 1u. This is a contradiction to the acylindricity of the action
X
since dY p1, x0 q ě R. So there are at most 4L ¨ pε ` t ` cn q elements u P BG
psq
X
which satisfy the equation Φn pu, vq “ w for some appropriate elements v P BG ptq.
Due to construction there are just 4 possible choices for xu,v P Mn . But if we fix u
X
ptq satisfying uxu,v v “ w. Thus the
and xu,v , there is at most one element v P BG
X
cardinality of the preimage of an element of BG
ps ` t ` cn q with respect to Φn is
bounded by 16L ¨ pε ` t ` cn q.
Together with Remark 2.5, Proposition 5.7 immediately implies the following
corollary:
Corollary 5.8. Let G be a finitely generated group with a finite generating set X
and a generating set Y such that ΓpG, Y q is hyperbolic and the canonical action of G
on ΓpG, Y q is acylindric. Let H be a subgroup of G which contains two independent
loxodromic elements g and h. Then there is a constant c and a linearly bounded
function l : N Ñ N such that the inequality
X
X
X
βH
psq ¨ βH
ptq ď lptq ¨ βH
ps ` t ` cq
(5.16)
holds for all s, t P N.
Proof. By considering Proposition 5.7, one only has to notice that Φn restricts to
X
X
X
BH
psq ˆ BH
ptq Ñ BH
ps ` t ` cn q since g, h P H.
Theorem 5.9. Let G be a finitely generated group and H ď G a subgroup of G. If
H contains a generalized loxodromic element h P H, then the relative exponential
growth rate of H in G exists with respect to every finite generating set of G.
Proof. First Theorem 4.6 gives us a generating set Y of G such that ΓpG, Y q is
hyperbolic and the canonical action of G on ΓpG, Y q is acylindrical and h is a loxodromic element for this action. Then Theorem 4.4 implies that there are two cases
to consider for H. Either H is virtually cyclic or H contains infinitely many independent loxodromic elements. In the first case H is quasiisometrically embedded
in ΓpG, Y q. In particular H is quasiisometrically embedded in ΓpG, Xq and there is
X
a constant c ą 0 such that βH
pnq ď c ¨ n for every n P N. Thus we see immediately
1
X
that the limit lim βH pnq n exists and is equal to 1. For the second case we can
nÑ8
14
EDUARD SCHESLER
choose a finite generating set X of G. Then we can apply Corollary 5.8 and we get
the inequality
(5.17)
X
X
X
βH
pmq ¨ βH
pnq ď lpnq ¨ βH
pm ` n ` cq
for all m, n P N. Now the statement follows from Proposition 2.11.
In the same way we get the following invariant of acylindrically hyperbolic
groups.
Theorem 5.10. Let G be a finitely generated acylindrically hyperbolic group. Then
G possesses a linearly bounded ambiguity function with respect to every finite generating set of G.
The following result is an immediate consequence of Theorem 5.10 and Corollary 2.12.
Theorem 5.11. Let H be a finitely generated acylindrically hyperbolic group. Let
ι : H Ñ G be a subexponentially distorted embedding in a finitely generated group
G. Then the relative exponential growth rate of H in G exists with respect to every
finite generating set of G.
6. The case of products
In this section we shall see that the relative exponential growth rate of a subgroup
of a direct product of finitely generated acylindrically hyperbolic groups exists if the
projection of the subgroup on each coordinate contains a generalized loxodromic
element. First we shall introduce some notation, which we will use throughout this
section.
Notation. To each element i of an indexset I “ t1, . . . , mu we associate a finitely
generated group Gi and a finite generating set Xi for Gi . For a subset J Ă I we
define
ś
ś
ś
ś
Gj
Gi Ñ
Gi and ψJ :
Gj Ñ
ιJ :
jPJ
iPI
iPI
jPJ
canonically. We will write ιi :“ ιtiu and ψi :“ ψtiu .
Remark 6.1. Let X “
Ť
ιi pXi q. Then |pg1 , . . . , gm q|X “
m
ř
|gi |Xi .
i“1
iPI
The following definition is a relative analogue of the ordinary ambiguity function.
Definition 6.2. Let G be a finitely generated group with a finite generating set X.
Let H ď G be an arbitrary subgroup. A function f : N Ñ N is a relative ambiguity
function of H in G with respect to X, if there is a finite subset M Ă H and a map
of the form
(6.1)
Φ : H ˆ H Ñ H, pu, vq ÞÑ ugu,v v
with gu,v P M such that by setting c :“ max |g|X the following property is satisfied.
gPM
‚ For all m, n P N and all restrictions
(6.2)
X
X
X
Φm,n : BH
pmq ˆ BH
pnq Ñ BH
pm ` n ` cq, pu, vq ÞÑ Φppu, vqq,
X
the cardinality of every preimage Φ´1
m,n pzq with z P BH pm`n`cq is bounded
by f pnq.
THE RELATIVE EXPONENTIAL GROWTH RATE OF SUBGROUPS OF ACYLINDRICALLY HYPERBOLIC GROUPS
15
In the following we will refer to M as the set of connecting pieces and to Φ as the
concatenation map corresponding to f .
Lemma 6.3. Let G be a finitely generated group with a finite generating set X. Let
H, K ď G be subgroups such that there is an epimorphism ψ : H Ñ K satisfying
X
|ψphq|X ď |h|X for every h P H. If βkerpψq
is a subexponential function and K
possesses a subexponential relative ambiguity function ε with respect to X, then H
possesses a subexponential relative ambiguity function with respect to X.
Proof. Let MK be the set of connecting pieces corresponding to ε. For each x P MK
weŤpic a preimage x1 of x with respect to ψ. We will show that the set MH :“
tx1 u gives rise to a subexponential relative ambiguity function δ for H. To do
xPMK
so, we define the concatenation map for H by
ΦH : H ˆ H Ñ H, pu, vq ÞÑ ux1u,v v,
where x1u,v is the chosen preimage of xψpuq,ψpvq P MK . As in the previous sections,
the element xψpuq,ψpvq is the connecting piece defined by the (relative) ambiguity
function
ΦK : K ˆ K Ñ K, pu, vq ÞÑ uxu,v v
corresponding to ε. Let cH :“ max |g|X . As in the non relative case, we shall
gPMH
denote the restrictions
X
X
X
BH
psq ˆ BH
ptq Ñ BH
ps ` t ` cH q, pu, vq ÞÑ ux1u,v v
X
by ΦK . We fix two numbers s, t P N and an element w P BH
ps ` t ` cH q. We
X
want to bound the number of elements v P BH ptq, such that there is an element
X
u P BH
psq satisfying ux1u,v v “ w. Let N :“ kerpψq. Due to our assumption, the
X
set of distinct classes rvs P H{N satisfying ux1u,v v “ w for some u P BH
psq and
X
a representative element v P BH ptq is bounded by εptq. This is because distinct
classes vN are associated to distinct solutions ψpuqxψpuq,ψpvq ψpvq “ ψpwq with
X
ψpvq P BH
ptq. Indeed, due to our assumption we have |ψpvq|X ď |v|X ď t. Now
X
X
we fix an element v P BH
ptq. We shall bound the number of elements v̂ P BH
ptq
X
representing v modulo N . Every such v̂ has the form v̂ “ vg P BH ptq for some
X
g P N . Thus we have to bound the cardinality of the set N X v ´1 BH
ptq. Since
´1 X
X
´1 X
X
|v|X ď t we have v BH ptq Ă BH p2tq and thus |N Xv BH ptq| ď βN p2tq. It is easy
X
X
to verify that t ÞÑ βN
p2tq is a subexponential function since βN
is a subexponential
X
function. Putting these bounds together, we see that the set of elements v P BH
ptq
1
X
X
satisfying uxu,v v for some u P BH psq is bounded by εptqβN p2tq, which is easily seen
to grow subexponentially with respect to t.
Lemma 6.4. Let I “ t1, . . . , mu be a finite set of indices. To each i P I we associate
a group Gi and a generating set Yi such
ś that ΓpGi , Yi q is hyperbolic and Gi acts
Gi be a subgroup, such that ψi pHq ď Gi
acylindrically on ΓpGi , Yi q. Let H ď
iPI
contains a loxodromic element hi for each i P I. Then there is a decomposition of
I into disjoint subsets I1 and I2 , such that the following properties are satisfied.
1) The kernel of the restricted projection ψI1 |H is finite.
2) For every index i P I1 there is a loxodromic element gi P Gi such that H
contains an element of the form pg1 , . . . , gm q with gj “ 1 for j P I1 ztiu.
16
EDUARD SCHESLER
Proof. Let I1 Ă I be a minimal subset such that ψI1 |H has a finite kernel. Thus, if
we take i P I1 , the kernel of ψI1 ztiu|H has to be infinite. This implies that there is
an infinite subset Ai Ă Gi with the following property. For every element h P Ai
there is an element of the form pg1 , . . . , gm q P H with gj “ 1 for j P I1 ztiu and
gi “ h. Due to our assumption there is an element ph1 , . . . , hm q P H such that hi
is loxodromic. We note that
´n
phn1 , . . . , hnm qpg1 , . . . , gm qph´n
1 , . . . , hm q “: pk1 , . . . , km q P H
satisfies kj “ 1 for j P I1 ztiu and ki “ hni gi h´n
“ hni hh´n
i
i . It remains to show
´n
n
that the set thi hhi | n P N, h P Ai u contains a loxodromic element. Due to
Theorem 4.4 this is equivalent to show that there is no uniform bound ε ą 0 such
that |hni hh´n
i |Y ď ε for every element h P Ai and every number n P N. To get
a contradiction, we assume that there is such a bound ε. Since the action of G
on ΓpG, Y q is acylindrical there are constants L, R ą 0 such that for every pair of
elements x, y P G with dY px, yq ě R we have
|tg P G | dY px, gxq ď ε and dY py, gyq ď εu| ď N.
Since hi is loxodromic, we can choose n P N big enough to satisfy |hni |Y ě R. Then,
by setting x “ 1 and y “ hni , we see that every element h P Ai satisfies
dY p1, h1q ď ε and dY phni , hhni q ď ε.
This is a contradiction since Ai is infinite.
Proposition 6.5. Let I “ t1, . . . , mu be a finite set of indices. To each i P I we
associate a finitely generated group Gi and a generating set Yi such
ś that ΓpGi , Yi q
Gi “: G be a
is hyperbolic and Gi acts acylindrically on ΓpGi , Yi q. Let H ď
iPI
subgroup with the following property. For each i P I there are elements g piq “
piq
piq
piq
piq
piq
piq
pg1 , . . . , gm q P H and hpiq “ ph1 , . . . , hm q P H such that gi and hi are inpiq
piq
dependent loxodromic elements and gj , hj “ 1 for j ‰ i. Then the relative
exponential growth rate of H in G exists with respect to every finite generating set
of G.
Proof. Let X be a finite generating set of G. Using Lemma 5.6 we may assume
that Yi contains ψi pXq “ Xi for each i P I. We shall proceed with each factor Gi
as in Proposition 5.7. To do so, we define
ś
ś
Mn “ t pg piq qεn | ε P t1, ´1uu Y t phpiq qεn | ε P t1, ´1uu
iPI
iPI
to be the set of connecting pieces and cn “ max |z|X . For each i P I we define
zPMn
Φnpiq : Gi ˆ Gi Ñ Gi , pu, vq ÞÑ uxpiq
u,v v
as in Definition 5.4. Thus, using Proposition 5.5 for large enough n P N, we see
that there is a constant ε ě 0 such that by defining
piq
piq
Φn : G ˆ G Ñ G, ppui qiPI , pvi qiPI q ÞÑ pΦn pui , vi qqiPI “ pui xu,v vi qiPI ,
the projections of the geodesics
r1, Φn ppui qiPI , pvi qiPI qs Ă ΓpG,
Ť
iPI
piq
ιi pYi qq onto r1, Φn pui , vi qs Ă ΓpGi , Yi q
THE RELATIVE EXPONENTIAL GROWTH RATE OF SUBGROUPS OF ACYLINDRICALLY HYPERBOLIC GROUPS
17
run through the ε-neighbourhoods of ui and ui xui ,vi with respect to dYi . Then, by
the same arguments as in Proposition 5.7, we see that the set of possible choices for
piq
Xi
Yi
Xi
u i P BG
ptq Ă BG
ptq satisfying Φn pui , vi q “ wi for a fixed wi P BG
ps ` t ` cn q,
i
i
i
X
is bounded by a linear function in t. Now let w P BG ps ` t ` cn q be fixed. Then
Xi
we obviously have |ψi pwq|Xi ď |w|X and hence wi :“ ψi pwq P BG
ps ` t ` cn q.
i
Xi
X
Thus, the possible different images ui :“ ψi puq P BGi psq for u P BG
psq satisfying
piq
Xi
Φn pui , vi q “ wi for some appropriate vi P BG
ptq is bounded by a linear function
i
li in t. Since u is uniquely determined by its images ui and there is at most one
X
element v P BG
ptq satisfying uxu,v v “ w for fixed elements u, xu,v and w, wee see
that the cardinality of the preimage of w with respect to the restriction
X
X
X
Φn : BH
psq ˆ BH
ptq Ñ BH
ps ` t ` cn q
is bounded by
|Mn |
ś
iPI
li ptq “ 2n`1
ś
li ptq.
iPI
This is a polynomial function of degree |I|. Using Remark 2.5, the claim follows by
X
applying Proposition 2.11 on the relative growth function βH
.
Theorem 6.6. Let G “
m
ś
Gi be a finite direct product of finitely generated acylin-
i“1
drically hyperbolic groups Gi and let H be a subgroup of G. Suppose that, for each
i “ 1, . . . , m, the image of the canonical projection of H to Gi contains a generalized loxodromic in Gi . Then the relative exponential growth rate of H in G exists
with respect to every finite generating set X of G.
Proof. Let I “ t1, . . . , mu be the set of indices. It remains to reduce the general
case to the case considered in Proposition 6.5. For this we choose generating sets
Yi for Gi such that ΓpGi , Yi q is hyperbolic, Gi acts acylindrically on ΓpGi , Yi q and
hi is a loxodromic element in ψi pHq. Indeed, this is possible due to Theorem 4.6.
Further Lemma 6.4 provides us with a decomposition of the indexset I into two
disjoint subsets I1 and I2 such that the kernel of the map ψI1 |H is finite and for every
index i P I1 , there is a loxodromic element gi P Gi such that H contains an element
of the form ph1 , . . . , hm q with hj “ 1 for j P I1 ztiu and hi “ gi . Due to Lemma 6.3
we can restrict out attention to the case of I “ I1 . We have to consider two cases.
Since ψi pHq contains a loxodromic element, Theorem 4.4 says that ψi pHq is either
virtually cyclic or it contains infinitely many independent loxodromix elements. Let
I “ I1 “ J1 Y J2 be a disjoint decomposition such that ψi pHq is virtually cyclic
if i P J1 , or contains infinitely many independent loxodromic elements if i P J2 .
In the case of i P J1 , the virtually cyclic subgroup ψi pHq contains a loxodromic
element and hence it is quasiisometrically embedded in ΓpGi , Xi q. From this, it is
easily seen that ψi pHq is also quasiisometrically embedded in ΓpG, Xq via ιi . This
provides us with a constant C ě 0 such that βιXi pψi pHqq ptq ď Ct for every t P N. We
X
note that Ni :“ kerpψIztiu|H q is a subgroup of ιi pψi pHqq and hence βN
ptq ď Ct
i
for every t P N. A further application of Lemma 6.3 reduces the problem to the
case of I “ pJ1 ztiuq Y J2 . Now we see, that an inductive application of Lemma 6.3
reduces the general case to the case of I “ J2 . But this case was already proven in
Proposition 6.5.
18
EDUARD SCHESLER
Acknowledgments. I would like to thank my PhD supervisor Prof. Oleg Bogopolski for guidance in writing this paper.
References
[BH99] Martin R Bridson and André Haefliger. Metric spaces of non-positive curvature, volume
319. Springer Science & Business Media, 1999.
[Gri80] Rostislav I Grigorchuk. Symmetrical random walks on discrete groups. Adv. Probab. Related Topics, 6:285–325, 1980.
[M` 68] John Milnor et al. A note on curvature and fundamental group. Journal of Differential
geometry, 2(1):1–7, 1968.
[Ols13] A. Olshanskii. Subnormal subgroups in free groups, their growth and cogrowth. ArXiv
e-prints, November 2013.
[Osi16] Denis Osin. Acylindrically hyperbolic groups. Transactions of the American Mathematical
Society, 368(2):851–888, 2016.
| 4math.GR
|
1
An improvement of the asymptotic Elias bound for
non-binary codes.
Krishna Kaipa
arXiv:1705.07785v3 [cs.IT] 27 Feb 2018
Abstract
For non-binary codes the Elias bound is a good upper bound for the asymptotic information rate at low relative
minimum distance, where as the Plotkin bound is better at high relative minimum distance. In this work, we obtain
a hybrid of these bounds which improves both. This in turn is based on the anticode bound which is a hybrid of
the Hamming and Singleton bounds and improves both bounds. The question of convexity of the asymptotic rate
function is an important open question. We conjecture a much weaker form of the convexity, and we show that
our bounds follow immediately if we assume the conjecture.
Index Terms
information rate, size of a code, anticode.
I. I NTRODUCTION
Let Aq (n, d; L) denote the maximum size of a code of length n, minimum distance at least d, and
contained in a subset L ⊂ F n , where F is an alphabet of finite size q. A central problem in coding theory
is to obtain good upper and lower bounds for Aq (n, d) = Aq (n, d; F n ). The asymptotic version of this
quantity is the asymptotic information rate function:
α(x) = lim sup n−1 logq Aq (n, xn), x ∈ [0, 1].
(1)
n→∞
The quantities Aq (n, d; L) and Aq (n, d) are related by the inequality
Aq (n, d) ≤ q n Aq (n, d; L)/|L|,
(2)
known as the Bassalygo-Elias lemma. Taking L to be a Hamming ball of diameter w, and choosing w
optimally gives, at the asymptotic level the Hamming and the Elias upper bounds.
αH (x) = 1 − Hq (x/2), x ∈ [0, 1].
p
αE (x) = αH (2θ(1 − 1 − x/θ)), x ∈ [0, θ].
(3)
(4)
The bound αE is better than αH for all x. Here Hq (x) is the entropy function (9), and θ := 1 − q −1 .
An anticode of diameter w in F n is any subset of F n with Hamming diameter w. Let A∗q (n, w) denote
the maximum size of an anticode of diameter at most w in F n . In contrast to the situation with Aq (n, d),
the quantity A∗q (n, d) was explicitly determined by Ahlswede and Khachatrian in [1]. From their result,
it is easy to determine the asymptotic quantity α∗ (x) = limn→∞ n−1 logq A∗q (n, xn). We actually do not
need the results of [1], however it is the main inspiration for this work. Taking L to be an A∗q (n, w)
anticode in (2), and choosing w optimally, we get the following two bounds which improve αH and αE
respectively.
Theorem 1. (hybrid Hamming-Singleton bound)
(
1 − Hq ( x2 )
if x ∈ [0, 2/q]
αHS (x) =
(1 − x)Hq (1) if x ∈ [2/q, 1].
K. Kaipa is with the Department of Mathematics at the Indian Institute of Science and Education Research, Pune 411008, India
(email: [email protected]). The author was supported by the Science and Engineering Research Board, Govt. of India, under Project
EMR/2016/005578.
2
The bound αHS improves the Hamming and the Singleton bounds. It is ∪-convex and continuously
differentiable.
Theorem 2. (hybrid Elias-Plotkin bound) Let q > 2.
(
√
2q−3
]
1 − Hq (θ − θ2 − xθ) if x ∈ [0, q(q−1)
αEP (x) =
(q−1)Hq (1)
2q−3
(θ − x) q−2
if x ∈ [ q(q−1) , θ]
The bound αEP improves the Elias and Plotkin bounds. It is ∪-convex and continuously differentiable.
It is not known if the function α(x) itself is ∪-convex, although it is tempting to believe that it is. We
propose a weaker conjecture:
Conjecture 1. The function
α(x)
θ−x
is decreasing. In other words
α(tx + (1 − t)θ) ≤ tα(x) + (1 − t)α(θ), t ∈ [0, 1].
As evidence for this conjecture, we will show that theorems 1 and 2 follow very easily if we admit the
truth of the conjecture.
The bound αEP in Theorem 2 is an elementary and explicit correction to the classical Elias bound.
It does not however improve the upper-bounds obtained by the linear programming approach, like the
second MRRW bound αM RRW 2 (due to Aaltonen [2]) or the further improvement of αM RRW 2 due to
Ben-Haim and Litsyn [3, Theorem 7]. The reasons for this are as follows: For small δ we have αEP (δ) =
αE (δ) ≥ αM RRW 2 (δ). For large δ, the inequality αEP (δ) > αM RRW 2 (δ) follows from the fact that αEP (δ)
has a non-zero slope at δ = 1 − 1/q where as the actual function α(δ) and the bound αM RRW 2 have zero
slope at δ = 1 − 1/q.
The paper is organized as follows. In section II, we collect some results on size of anticodes, which we
use in section III to prove Theorems 1 and 2. We discuss Conjecture 1 in section IV.
II. S IZE OF ANTICODES
We recall that A∗q (n, d) is the maximum size of an anticode of diameter at most d in F n . If we take L
to be an anticode of size A∗q (n, d − 1) then clearly Aq (n, d; L) = 1. Using this in (2), we get a bound
Aq (n, d) ≤ q n /A∗q (n, d − 1),
(5)
known as Delsarte’s code-anticode bound [4]. Taking d = xn where x ∈ [0, 1] we get
n−1 logq Aq (n, xn) ≤ 1 − n−1 logq A∗q (n, xn − 1).
Taking lim supn→∞ we get:
α(x) ≤ 1 − α∗ (x),
(6)
α∗ (x) = lim inf n−1 logq A∗q (n, xn).
(7)
where
n→∞
This is the the asymptotic form of (5). We use the notation B(r; n) and Vq (n, r) to denote a Hamming
ball of radius r in F n and its volume respectively. The ball B(t; n) where t = b(d − 1)/2c in F n is an
anticode of diameter at most d − 1. Let F n = F d−1 × F n−d+1 and let v ∈ F n−d+1 be a fixed word. Sets
of the form F d−1 × {v} of size q d−1 are also anticodes of diameter d − 1. It follows that:
α∗ (x) ≥ max{Hq (x/2), x}.
Here, we have used the well known formula:
lim n−1 logq Vq (n, tn) = Hq (t), t ∈ [0, θ],
n→∞
(8)
3
where,
1
) + (1 − x) logq ( 1−x
), x ∈ [0, 1].
Hq (x) = x logq ( q−1
x
(9)
While the convexity of α(x) is an open question, it is quite easy to see that:
Lemma 1. The function α∗ (x) is ∩-convex.
F
n1
Proof: If S1 ⊂ F n1 and S2 ⊂ F n2 are anticodes of diameters d1 and d2 respectively, then S1 × S2 ⊂
× F n2 is an anticode of diameter d1 + d2 . Taking Si to be A∗q (ni , di ) anticodes, we immediately get
A∗q (n1 + n2 , d1 + d2 ) ≥ A∗q (n1 , d1 )A∗q (n2 , d2 ).
Let n = n1 + n2 go to infinity with n1 /n = t + o(1), d1 /n1 = x + o(1) and d2 /n2 = y + o(1). Applying
lim inf n→∞ n−1 logq to this inequality we get:
α∗ (tx + (1 − t)y) ≥ tα∗ (x) + (1 − t)α∗ (y), t ∈ [0, 1].
We note that with codes we have d(C1 ×C2 ) = min{d(C1 ), d(C2 )}, which is why the above proof method
does not apply to the question of convexity of α(x). From (8) and Lemma 1 we get:
α∗ (tx + (1 − t)y) ≥ tHq (x/2) + (1 − t)y, t ∈ [0, 1].
Let δ = tx + (1 − t)y. We can rewrite this as
α∗ (δ) ≥ f (x, y),
where f : [0, δ) × (δ, 1] is defined by
f (x, y) =
y−δ
(Hq (x/2)
y−x
− x) + δ.
(10)
We note that
(y−x)2 ∂f
(x, y)
δ−x ∂y
x(y−x)2
∂f
(x, y)
y(y−δ) ∂x
= Hq (x/2) − x,
= Hq ( x2 ) − x + (1 − xy ) logq (1 − x2 ).
There is a unique positive number b > 0 satisfying Hq (b/2) = b (where the Hamming and Singleton
bounds intersect). Therefore, Hq (x/2) − x has the same sign as b − x. Using this in (10), we see that
f (x, y) ≤ δ for x ≥ b. Therefore, in order to maximize f (x, y) it suffices to consider x < b. We note that
∂f
(x, y) has the same sign as H(x/2) − x and hence that of b − x. Since x < b, we see that for fixed
∂y
x < b, the function f (x, y) is maximized for y = 1. We are now reduced to maximizing
q (x/2)
f (x, 1) = 1 − (1 − δ) 1−H1−x
, x ∈ [0, δ].
Lemma 2. Let g(x) =
1−Hq (x/2)
1−x
for x ∈ [0, 1].
sign(g 0 (x)) = sign(x − 2q ).
Proof: We calculate:
g 0 (x) =
1
2(1−x)2
2
x(2−x)
logq ( q 4(q−1)
).
2
x(2−x)
Therefore sign(g 0 (x)) = sign( q 4(q−1)
− 1). Next, we note that
q 2 x(2−x)
4(q−1)
− 1 = q(x − 2q ) (q−2)+q(1−x)
4(q−1)
has the same sign as x − 2/q, as was to be shown. A stronger assertion is that g(x) is in fact ∪-convex:
differentiating once more, we get:
2
/4
1
1
ln(q)(1 − x)3 g 00 (x) = ln( qq−1
) + ( 2x−x
2 − 1 − ln( 2x−x2 ))
4
We note that q 2 ≥ 4(q − 1), and hence the first term is non-negative. The remaining parenthetical term is
non-negative using the inequality
t − 1 − ln(t) ≥ 0 for t ≥ 1,
(11)
and the fact that t = 1/(2x − x2 ) ≥ 1 for x ∈ (0, 1].
It follows from Lemma 2 that
q (x/2)
argminx∈[0,δ] 1−H1−x
= min{δ, 2/q}.
(12)
(
Hq (x/2)
if x ∈ [0, 2/q]
β(x) =
1 − (1 − x)Hq (1) if x ∈ [2/q, 1].
(13)
Therefore we obtain the bound:
Theorem 3. α∗ (x) ≥ β(x) where
Moreover, β(x) is continuously differentiable and ∩-convex.
We have used the relation
1−Hq (1/q)
1−2/q
= Hq (1) = logq (q − 1).
(14)
The function β(x) is continuously differentiable because the component for x ≥ 2/q is just the tangent
line at x = 2/q to the component for x ≤ 2/q, i.e. to Hq (x/2). We note that β 0 (x) equals Hq0 (x/2)/2 for
x ≤ 2/q and Hq0 (1/q)/2 for x ≥ 2/q. Since β 0 (x) is non-increasing, it follows that β(x) is ∩-convex.
In the next lemma, we show that there is a sequence of anticodes Sn ⊂ F n of diameter at most δn
such that limn→∞ n−1 logq |Sn | equals β(δ), i.e. the lower bound on α∗ (δ) given in theorem 3.
Lemma 3. Consider the anticodes S(d, n) of diameter d in F n (taken from [1]) given by
S(d, n) = B(rd,n ; n − d + 2rd,n ) × F d−2rd,n , where
rd,n = max{0, min{d d−1
e, d n−d−q+1
e}}.
2
q−2
Then limn→∞ n−1 logq |S(δn, n)| = β(δ).
Proof: We note that
(
rδn,n
n→∞ n
ρ = lim
=
δ
2
1−δ
q−2
if δ ∈ [0, 2/q]
if δ ∈ [2/q, 1].
Also limn→∞ n−1 logq |S(δn, n)| equals
ρ
(1 − δ + 2ρ)Hq ( 1−δ+2ρ
) + (δ − 2ρ),
which simplifies to Hq (δ/2) if δ ≤ 2/q and (on using (14)) to 1 − (1 − δ)Hq (1) if δ ≥ 2/q. This is the
same as β(δ).
We now have all the results we need for proving theorems 1 and 2. However, we will state a remarkable
theorem due to Ahlswede and Khachatrian [1], which we will not need. We also obtain an asymptotic
version of their result and record it as a corollary, as it does not seem to have appeared in literature. In brief
their theorem states that A∗q (n, d) equals |S(d, n)|. Moreover any A∗q (n, d) anticode is Hamming isometric
to the anticode S(d, n) (with some exceptions). At the asymptotic level, the result is again remarkable:
The lower bound β(δ) for α∗ (δ) given in theorem 3 is actually the exact value of α∗ (δ). Moreover α∗ (δ)
need not have been defined using lim inf n→∞ as limn→∞ n−1 logq A∗q (n, δn) already exists.
Theorem. [1] Given q ≥ 2 and integers 0 ≤ d ≤ n, let rd,n and S(d, n) be as in Lemma 3. Then,
A∗q (n, d) = |S(d, n)|.
5
Moreover, up to a Hamming isometry of F n an anticode S of size A∗q (n, d) must be:
• S(d, n)
• or S(d, n) with rd,n replaced with rd,n − 1. This case is possible only if (n − d − 1)/(q − 2) is a
positive integer not exceeding d/2.
Corollary 1.
(
Hq (x/2)
α∗ (x) =
1 − (1 − x)Hq (1)
if 0 ≤ x ≤ 2/q
if 2/q ≤ x ≤ 1.
Proof: It follows from the theorem of Ahlswede and Khachatrian, together with Lemma 3 that
logq A∗q (n,δn)
n
n→∞
lim
Therefore
α∗ (δ) = lim inf
n→∞
= lim
n→∞
logq A∗q (n,δn)
n
logq |S(δn,n)|
n
= βq (δ).
logq A∗q (n,δn)
n
n→∞
= lim
= β(δ).
III. P ROOFS OF THEOREMS 1 AND 2
A. Proof of Theorem 1
If we use the bound α∗ (x) ≥ β(x) of Theorem 3 in the inequality α(x) ≤ 1 − α∗ (x) (see (6)), we
obtain the bound
α(x) ≤ 1 − β(x) =: αHS (x).
Since β(x) is ∩-convex and continuously differentiable (see Theorem 3), it follows that αHS (x) is ∪convex and continuously-differentiable. To show that αHS (x) ≤ αS (x) = 1 − x, we note that αS (x) being
the secant line to αHS (x) between (0, αHS (0)) and (1, αHS (1)), lies above the graph of αHS (x) as the
latter is ∪-convex. To prove that αHS (x) improves αH (x) we note that αHS (x) coincides with αH (x) for
x ≤ 2/q, and for x ≥ 2/q, Lemma 2 implies that αH (x) ≥ (1 − x)Hq (1) = αHS (x). This finishes the
proof of Theorem 1.
It is worth noting that (12) implies the following formula for αHS (δ):
αH (x)(1−δ)
.
1−x
x∈[0,δ]
αHS (δ) = min
Since
θ−δ
θ−x
≤
1−δ
1−x
(15)
for x ∈ [0, δ] and δ ≤ θ, we get
αHS (δ) ≥ αHP (δ) := min
x∈[0,δ]
αH (x)(θ−δ)
.
θ−x
(16)
It can be shown (see subsection III-C) that αHP (δ) is an upper bound for α(δ) which improves both the
Hamming and Plotkin bounds.
B. Proof of Theorem 2
It will be convenient to identify the alphabet F with the abelian group Z/qZ. Given 0 ≤ δ ≤ ω, let
wn = bωnc and dn = bδnc. We take Ln ⊂ F n to be the anticode (from Lemma 3):
Ln = B(rn ; n − wn + 2rn ) × F wn −2rn , where
n −q+1
e}}.
rn = max{0, min{d wn2−1 e, d n−wq−2
(17)
6
We will take the balls B(r; m) to be centered at (0, . . . , 0) ∈ F m . As in Lemma 3, we have
(
ω
if ω ∈ [0, 2/q]
2
ρ := lim rnn = 1−ω
n→∞
if ω ∈ [2/q, 1].
q−2
(18)
We also note that Lemma 3 gives
lim n−1 logq |Ln | = β(ω) = 1 − αHS (ω).
(19)
n→∞
Let Aq (n, dn ; Ln ) be the maximum possible size of a code contained in Ln and having minimum distance
at least dn .
Theorem 4. limn→∞ n−1 logq Aq (n, dn ; Ln ) = 0 if
ρ
θ(1−ω+2ρ)
≤1−
q
1−δ/θ
.
1−ω+2ρ
(20)
Proof: Our proof is similar to the standard proof of the analogous result for the Elias bound (which
corresponds to taking ρ = ω/2 instead of the prescription (18)). First let C ⊂ L be a code of size
M = Aq (n, d; L), where L ⊂ F n is the anticode L = B(r; n − w + 2r) × F w−2r for some r ≤ w/2. Let
γ1 (C) = (M n)
−1
n X
X
m(i, a)2 ,
i=1 a∈F
P
where m(i, a) = #{c ∈ C : ci = a}. We note that M = a∈F m(i, a), and that
XX
γ1
).
M (M − 1)d ≤
d(c, c0 ) = nM 2 (1 − M
c∈C c0 ∈C
We can rewrite this as:
M≤
γ1
M
d/n
, provided
− (1 − nd )
γ1
M
> 1 − nd .
For n − w + 2r < i ≤ n we use Cauchy-Schwarz inequality to get
n
X
1
M 2 (w−2r)
X
P
a∈F
(21)
m(i, a)2 ≥ M 2 /q. In particular
m(i, a)2 ≥ 1q .
(22)
i=n−w+2r+1 a∈F
Let π1 be the projection of F n = F n−w+2r × F w−2r on to the factor F n−w+2r . We note that for c ∈ C,
we have wt(π1 (c)) < r because π1 (C) ⊂ B(r; n − w + 2r). Here wt(v) is the number of nonzero entries
of v. Therefore
n−w+2r
X X
m(i, a) ≤ M r.
i=1
Since
Pn−w+2r P
i=1
a
a6=0
m(i, a) = M (n − w + 2r) we get:
S=
n−w+2r
X
m(i, 0) ≥ (n − w + r)M.
i=1
In particular
S
M (n−w+2r)
We note that
P
a6=0
−
1
q
≥
n−w+r
n−w+2r
−
1
q
=θ−
r
.
n−w+2r
m(i, a) = M − m(i, 0). By Cauchy-Schwarz inequality:
n−w+2r
X
i=1
m(i, 0)2 ≥ S 2 /(n − w + 2r), and
(23)
7
X
m(i, a)2 ≥ (M − m(i, 0))2 /(q − 1)
a6=0
Since
Pn−w+2r P
i=1
a∈F
m(i, a)2 equals
!
n−w+2r
X
m(i, 0)2 +
i=1
X
m(i, a)2 ,
a6=0
we get:
n−w+2r
X X
i=1
n−w+2r
X
2
m(i, a) ≥
qm(i,0)2 +M 2 −2M m(i,0)
q−1
i=1
a∈F
This can be rewritten as:
1
M 2 (n−w+2r)
n−w+2r
X X
i=1
S
m(i, a)2 ≥ 1θ ( M (n−w+2r)
− 1q )2 +
1
q
a∈F
Combining this with (22) we get:
γ1
M
≥
n−w+2r
S
( M (n−w+2r)
nθ
γ1
M
−
− q −1 )2 + 1q .
Using (23) this can be written as:
1
q
≥
n−w+2r
nθ
r
)2 .
n−w+2r
(θ −
Now let Cn ⊂ Ln be a sequence of codes of size Mn = Aq (n, dn ; Ln ). The preceding inequality gives:
γ1 (Cn )
Mn
−
1
q
≥
1−ω+2ρ
θ
(θ −
ρ
)2
1−ω+2ρ
+ o(1).
Using this in (21), we get:
Mn ≤
1−ω+2ρ
θ
δ + o(1)
,
ρ
(θ − 1−ω+2ρ
)2 − (θ − δ) + o(1)
provided the denominator is a positive number. Therefore, limn→∞ n−1 logq Mn = 0 provided
1−ω+2ρ
θ
(θ −
ρ
)2
1−ω+2ρ
This condition is the same as
ρ
θ(1−ω+2ρ)
≤1−
q
≥ θ − δ.
1−δ/θ
1−ω+2ρ
Since Mn = Aq (n, dn ; Ln ) this finishes the proof.
Using (2) we get:
logq Aq (n,dn )
log |L |
≤ 1 − qn n +
n
logq Aq (n,dn ;Ln )
.
n
Taking lim sup as n → ∞ and using the result of Theorem 4 and (19) we get:
α(δ) ≤ αHS (ωmax (δ)),
(24)
where ωmax (δ) is the largest value of ω for which the inequality (20) holds. In order to determine ωmax (δ),
we introduce functions f1 , f2 on [0, θ] defined by:
p
f1 (δ) = 2θ(1 − 1 − δ/θ)
(25)
2
f2 (δ)= 1 − (1 − δ/θ) (q−1)
q(q−2)
Lemma 4. Let q > 2.
1) f1 (δ) ≥ f2 (δ) with equality only at δ =
2q−3
.
q(q−1)
(26)
8
2q−3
.
2) f2 (δ) is the tangent line to f1 (δ) at δ = q(q−1)
2q−3
3) sign(f1 (δ) − 2/q) = sign(f2 (δ) − 2/q) = sign(δ − q(q−1)
).
p
Proof: Let f3 (δ) := 1 − q−1
1 − δ/θ. We observe that
q−2
sign(f3 (δ)) = sign(δ −
2q−3
).
q(q−1)
The three assertions to be proved follow respectively from the following three relations:
f1 (δ)−2/q
2(1−2/q)
=
f1 (δ) − f2 (δ) = (1 − 2/q) f3 (δ)2 ,
p
f10 (δ) − f20 (δ) = f3 (δ)/ 1 − δ/θ,
f2 (δ)−2/q
√
= f3 (δ).
(1−2/q)+θ
1−δ/θ
Proposition 1.
ωmax (δ) =
(
p
2θ(1 − 1 − δ/θ)
2
1 − (1 − δ/θ) (q−1)
q(q−2)
2q−3
if δ ∈ [0, q(q−1)
]
2q−3
if δ ∈ [ q(q−1)
, 1].
The function ωmax (δ) is increasing, continuously differentiable, and ∪-convex on [0, θ].
Proof: The inequality (20) reduces to
(
f1 (δ)
ω≤
f2 (δ)
if ρ = ω/2
if ρ = (1 − ω)/(2 − q),
where ρ is as given in (18). Therefore, for a given δ ∈ [0, θ], the quantity ωmax (δ) is the maximum element
of the set
{ω : δ ≤ ω ≤ min{f1 (δ), 2/q}} ∪ {ω : max{δ, 2/q} ≤ ω ≤ f2 (δ)}.
2q−3
2q−3
If δ ≥ q(q−1)
, then f2 (δ) ≥ 2/q (by Lemma 4) and hence, the maximum of this set is f2 (δ). If δ ≤ q(q−1)
,
then f2 (δ) ≤ f1 (δ) ≤ 2/q (by Lemma 4) and hence, the maximum of this set is f1 (δ). This proves the
asserted formula for ωmax (δ).
2q−3
We note that the second component of ωmax (δ) is the tangent line to the first component at x = q(q−1)
.
p
2q−3
Therefore ωmax (δ) is continuously differentiable. The derivative of ωmax (x) is 1/ 1 − x/θ for x ≤ q(q−1) ,
2q−3
and constant at q−1
for x ≥ q(q−1)
. Since the derivative is positive, the function is increasing. Since the
q−2
derivative is non-decreasing, we see that the function is ∪-convex.
Proof of αEP being an upper bound: We note from lemma 4 that
sign(ωmax (δ) − 2/q) = sign(δ −
2q−3
).
q(q−1)
Therefore αHS (ωmax (δ)) is just the function αEP (δ) defined in of theorem 2. The bound α(x) ≤ αEP (x)
now follows from (24).
Proof of αEP being continuously differentiable: The function αEP (x) = αHS (ωmax (x)) being a composition of continuously differentiable functions, is itself continuously differentiable.
Proof of αEP being ∪-convex: Both the functions αHS and ωmax are ∪-convex, but αHS is decreasing
and hence it is not obvious that αEP (x) = αHS (ωmax (x)) is ∪-convex. We will show instead that the
2q−3
0
0
00
derivative αEP
is non-decreasing. Since αEP
is constant for x ≥ q(q−1)
, it suffices to show that αE
(x) > 0
2q−3
for x ∈ (0, q(q−1) ]. This follows from the next lemma.
9
Lemma 5. The Elias bound αE (x) is ∪-convex on [0, δE ] and ∩-convex on [δE , θ] where δE satisfies:
2q−3
q(q−1)
Proof: Let Z(x) = θ(1 −
< δE < 43 ( q−4/3
).
q−1
p
1 − x/θ). A calculation shows that
4θ ln(q)(1 −
Z(x) 3 00
) αE (x)
θ
Z
ϕ(z) =
θ
z
1−θ
1−z
= ϕ(Z(x)), where
(1 − 1/t)dt.
To see this, we note: αE (x) = Hq (Z(x)) and hence
00
(x) =
ln(q)αE
Z0
Z(1−Z)
+ Z 00 ln( (q−1)(1−Z)
).
Z
Since Z 0 = 1/(2(1 − Z/θ)) and Z 00 = Z 0 /(2θ(1 − Z/θ)2 ), we get
Z θ
Z(x)
Z(x) 3 00
(1 − 1/t)dt,
4θ ln(q)(1 − θ ) αE (x) =
1−θ
1−Z(x)
00
as desired. It follows that sign(αE
(x)) = sign(ϕ(Z(x)). Next we note that Z(x) is increasing on [0, θ]
and
2q−3
) = 1/q, Z( 43 ( q−4/3
)) = 1/2.
Z( q(q−1)
q−1
It now suffices to show that
sign(ϕ(z)) = sign(zE − z), for some zE ∈ ( 1q , 12 ).
We note that
ϕ0 (z) = (z − 21 ) z2(θ−z)
2 (1−z)2 .
Thus ϕ(z) is decreasing on [0, 1/2] and increasing on [1/2, θ]. In order to show sign(ϕ(z)) = sign(zE − z)
for some zE ∈ (1/q, 1/2), it suffices to show that ϕ(1/q) > 0 and ϕ(1/2) < 0. We calculate
1
ϕ(1/q)
2
2q−3
= ( q−1
− 1 − ln( q−1
)) + ( 2q−2
− ln(2)).
2
2
Since q ≥ 3, we have q−1
≥ 1. The inequality (11) implies that the first parenthetical term above is
2
non-negative. Again q ≥ 3 implies
2q−3
2q−2
− ln(2) ≥
3
4
− ln(2) > 0,
and hence the second parenthetical term is positive. Thus ϕ(1/q) > 0.
Next, we note that ϕ(1/2) = 2 − 4/q − ln(q − 1). The function a(t) = 2 − 4/t − ln(t − 1) satisfies
2
a0 (t) = − t(t−2)
2 (t−1) ,
and a(3) = 2/3 − ln(2) < 0. Therefore a(t) < 0 for t ≥ 3, and hence ϕ(1/2) < 0 for all q ≥ 3.
Proof that αEP improves the Plotkin bound: We have already shown that αEP (x) is ∪-convex, and
hence αEP (x) lies below the secant line between x = 0 and x = θ, which is the Plotkin bound.
Proof that αEP improves the Elias bound: This does not readily follow from our results thus far, and
requires more work. The characterization of αEP (x) given in the next theorem clearly implies αEP (x) ≤
αE (x).
αE (x)(θ−δ)
.
θ−x
x∈[0,δ]
Theorem 5. αEP (δ) = min
10
1
0.8
0.6
αE (δ)
αEP (δ)
0.4
0.2
0
0
0.2
0.4
0.6
0.8
δ
Fig. 1. αE (δ) and αEP (δ) for q = 16.
2q−3
E (x)
is decreasing on [0, q(q−1)
] and
Proof: The theorem immediately follows if we show that αθ−x
2q−3
increasing on [ q(q−1) , θ]. We will use the notation from the proof of Lemma 5. Since αE (x) is ∩-convex
for x ≥ δE , it follows that the slope αE (x)/(θ − x) of the secant between x and θ is increasing. It remains
2q−3
2q−3
E (x)
is decreasing on [0, q(q−1)
] and increasing on [ q(q−1)
, δE ]. Since Z(x) is an increasing
to show that αθ−x
2q−3
2
function, with Z(0) = 0, Z( q(q−1) ) = 1/q, and (1 − x/θ) = (1 − z/θ) , it suffices to show that
h(z) =
1−Hq (z)
,
(θ−z)2
is decreasing on [0, 1/q] and increasing on [1/q, zE ] where zE = Z(δE ). A calculation shows that
Z z
3 0
ln(q)(θ − z) h (z) =
ϕ(t)dt.
1/q
z
To see this we note that either side of this equation evaluates to (θ + z) ln( (1−z)(q−1)
) + 2 ln(q(1 − z)).
Since ϕ(t) > 0 for t ∈ (0, zE ), we see
sign(h0 (z)) = sign(z − 1/q), z ∈ [0, zE ].
Thus we have also shown that h(z) is decreasing for z ∈ [0, 1/q] and increasing on [1/q, zE ] as required.
The bounds αHS , αHP and αEP are related as
αEP (δ) ≤ αHP (δ) ≤ αHS (δ).
We have already shown αHP (δ) ≤ αHS (δ) in (16). Since αE (x) ≤ αH (x) for all x, we note that
min
x∈[0,δ]
αE (x)(θ−δ)
θ−x
≤ min
x∈[0,δ]
αH (x)(θ−δ)
.
θ−x
Thus αEP (δ) ≤ αHP (δ). We end this section with a plot comparing αE (x) and αEP (x) for q = 16.
C. Another proof of Theorem 1
Another proof of αHS (x) being an upper bound for α(x) can be given using the following theorem of
Laihonen and Litsyn:
Theorem. [5] Let δ1 , δ2 , µ ∈ [0, 1].
α((1 − µ)δ1 + µδ2 ) ≤ (1 − µ)αH (δ1 ) + µα(δ2 ).
(27)
11
Proof: We give a quick proof. The result follows from the inequality:
q n1 Aq (n2 , d2 )
,
Aq (n1 + n2 , d1 + d2 ) ≤
Vq (n1 , d1 /2)
by taking n1 + n2 = n → ∞, and n1 /n, n2 /n, d1 /n1 and d2 /n2 going to 1 − µ, µ, δ1 and δ2 respectively.
The above inequality in turn comes from the Bassalygo-Elias lemma (2)
q n1 +n2 Aq (n1 + n2 , d1 + d2 ; L)
Aq (n1 + n2 , d1 + d2 ) ≤
,
|L|
by taking L = B(d1 /2; n1 ) × F n2 , and observing that Aq (n1 + n2 , d1 + d2 ; L) ≤ Aq (n2 , d2 ). (If C is a
Aq (n1 + n2 , d1 + d2 ; L) code, and π2 : F n1 × F n2 → F n2 is the projection on the second factor, then the
restriction of π2 to C is injective, and π2 (C) has minimum distance at least d2 .)
If we set δ2 = 1, δ1 = x and µ = (δ − δ1 )/(1 − δ1 ) in (27), we get:
α(y)
αH (x)
≤
for x ≤ y.
1−y
1−x
Thus,
= αHS (δ),
α(δ) ≤ min αH (x)(1−δ)
1−x
x∈[0,δ]
where we have used (15).
We now prove that the function αHP (δ) defined in (16) is an upper bound for α(δ). Taking δ2 = θ,
δ1 = x, and µ = (y − δ1 )/(θ − δ1 ) in (27), we get:
α(y)
αH (x)
≤
for x ≤ y
θ−y
θ−x
Thus
α(δ) ≤ min αH (x)(θ−δ)
= αHP (δ).
(28)
θ−x
x∈[0,δ]
It is not known if the inequality (27) (the theorem of Laihonen-Litsyn) holds if we replace αH by αE .
If such a result were true, then the derivation of the bound αHP (x) above with αH replaced with αE
would immediately yield Theorem 5. We believe that such an inequality
α((1 − µ)δ1 + µδ2 ) ≤ (1 − µ)αE (δ1 ) + µα(δ2 ),
(29)
must be true (it would surely be true if α(x) is ∪-convex), but we believe it cannot be obtained just by a
simple application of the Bassalygo-Elias lemma (2). If (29) holds, we can obtain an upper bound which
improves the Laihonen-Litsyn bound [5]. We recall that the Laihonen-Litsyn bound, which we denote
αHM RRW is a hybrid of the Hamming and MRRW bounds. It coincides with the Hamming bound for
δ ∈ [0, a] and with the MRRW bound for [b, θ] where a < b are points such that the straight line joining
(a, αH (a)) and (b, αM RRW (b)) is a common tangent to both αH at a and αM RRW at b. Since the Hamming
bound is good for small δ and the MRRW bound good for large δ, the Laihonen-Litsyn bound combines
the best features of both bounds in to a single bound. To obtain this bound, we note that (27) implies the
inequality
α((1 − µ)δ1 + µδ2 ) ≤ (1 − µ)αH (δ1 ) + µαM RRW (δ2 ).
We fix δ = (1 − µ)δ1 + µδ2 and choose δ1 and δ2 optimally in order to minimize the right hand side. This
yields the αHM RRW bound. Since the second MRRW bound αM RRW 2 improves the first MRRW bound
αM RRW , a better version αHM RRW 2 of the Laihonen-Litsyn bound (see [3, Theorem 2]) can be obtained
by using αM RRW 2 in place of αM RRW 2 . Since the Elias bound αE (δ) is better than the Hamming bound
αH (δ) for all δ, in case (29) is true, repeating this procedure with αE replacing αH , would yield the
hybrid Elias-MRRW bounds αEM RRW (δ), αEM RRW 2 (δ) which would improve the respective LaihonenLitsyn bounds αHM RRW , αHM RRW 2 (δ). We leave the question of the truth of (29) open.
12
IV. O N THE CONVEXITY OF α(x)
A fundamental open question about the function α(x) is whether it is ∪-convex. In other words is it
true that
α((1 − t)x + ty) ≤ (1 − t)α(x) + tα(y), t ∈ [0, 1].
(30)
It is worth noting that non-convex upper bounds like the Elias bound and the MRRW bound admit
corrections to the non-convex part: the bound αEP (x) for the Elias bound and the Aaltonen straight-line
bound (see the theorem below and the Appendix) for the MRRW bound. This may be viewed as some
kind of evidence supporting the truth of (30). It is known that (30) holds for x = 0 (for example by taking
δ1 = 0 in (27)). Another way to state this is that
(1 − α(x))/x is decreasing on [0, 1].
As a consequence, if αu (x) is any upper bound for α(x) we obtain a better upper bound
(1 − αu (x))δ
x∈[δ,θ]
x
α(δ) ≤ α̃u (δ) = 1 − max
(31)
To see this we use:
1−α(δ)
δ
≥
1−α(x)
x
≥
1−αu (x)
, for
x
x ∈ [δ, θ].
Thus 1−α(δ)
≥ maxx∈[δ,θ] (1−αxu (x)) as desired. If (1 − αu (x))/x is a decreasing function then the improved
δ
bound α̃u (x) coincides with αu (x), but otherwise α̃u (x) improves αu (x). For example let αu (x) be the
first MRRW bound αM RRW (x) =
p
p
Hq (( θ(1 − x) − x(1 − θ))2 ), x ∈ [0, θ].
It can be shown that that (1 − αM RRW (x))/x fails to be decreasing near x = 0, and similarly αM RRW (x)
fails to be ∪-convex near x = 0. This is immediately rectified by passing to the improved bound
α̃M RRW (x), resulting in the following theorem of Aaltonen.
Theorem. (Aaltonen bound) [6] [7, p.53] Let q > 2. α(x) ≤ α̃M RRW (x) where
(
q (1)
1 − xH
if x ∈ [0, (1 − 2q )2 ]
1−2/q
α̃M RRW (x) =
αM RRW (x) if x ∈ [(1 − 2q )2 , θ]
(32)
This bound is ∪-convex, continuously differentiable, and improves the MRRW bound.
We note that for x ≤ (1 − 2/q)2 the bound α̃M RRW (x) coincides with the tangent line to αM RRW (x) at
(1−2/q)2 . In particular α̃M RRW (x) is continuously differentiable. The assertion that α̃M RRW (x) improves
αM RRW (x) follows from the fact that α̃u (x) ≤ αu (x) for any upper bound αu (x) for α(x). The other
assertions are proved in the appendix.
On the other hand, it is not known if the convexity condition (30) holds for y = θ, in other words if
α(x)/(θ − x) is a decreasing function of x. We conjecture that this is true (see Conjecture 1). As evidence
for this conjecture, we now show that the bounds αEP , αHP and αHS can be obtained without doing any
work, if we assume the truth of Conjecture 1: if αu (x) is any upper bound for α(x) we obtain a better
upper bound
αu (x)(θ − δ)
α(δ) ≤ αu† (δ) := min
(33)
x∈[0,δ]
θ−x
To see this we use:
α(δ)
θ−δ
≤
α(x)
θ−x
≤
αu (x)
θ−x
for x ∈ [0, δ].
u (x)
Thus α(δ) ≤ minx∈[0,δ] αu (x)(θ−δ)
as desired. In case αθ−x
is a decreasing function then the improved
θ−x
†
†
bound αu (x) coincides with αu (x), but otherwise αu (x) improves αu (x). Taking αu (x) to be the Elias
13
bound, we get αu† (x) to be the bound αEP . This is the content of Theorem 5. Taking αu (x) to be the
Hamming bound, we get αu† (x) to be the bound αHP . This is the content of (28). Moreover, if α(x)/(θ−x)
is decreasing then α(x)/(1 − x) being the product of the non-negative decreasing functions α(x)/(θ − x)
. Taking αu (x) to be
and (θ − x)/(1 − x) is itself decreasing. Thus we obtain α(δ) ≤ minx∈[0,δ] αu (x)(1−δ)
1−x
αu (x)(1−δ)
is αHS (δ). This is the content of (15).
the Hamming bound, the bound minx∈[0,δ] 1−x
A PPENDIX A
A ALTONEN ’ S STRAIGHT- LINE BOUND
The bound α̃M RRW presented above was obtained by Aaltonen in [6, p.156]. The bound follows from
(31) and the following result
(x)
argmaxx∈[δ,θ] 1−αM RRW
= max{δ, (1 − 2/q)2 }.
x
(34)
The argmax above is not straightforward to obtain, and to quote from [6], was found by a mere chance.
The derivation is not presented in [6]. The purpose of this appendix is to i) record a proof of (34), and ii)
to prove that α̃M RRW (x) is ∪-convex. The author thanks Tero Laihonen for providing a copy of Aaltonen’s
work [6], which is not easily available.
p
p
Let ξ : [0, θ] → [0, θ] be the function defined by ξ(x) = ( θ(1 − x) − x(1 − θ))2 . We note that
αM RRW (x) = Hq (ξ(x)), and that ξ(x) decreases from θ to 0 as x runs from 0 to θ. It is easy to check
that ξ(ξ(x)) = x for x ∈ [0, θ]. Therefore we can invert the relation y = ξ(x) as x = ξ(y). We also note
that ξ((1 − 2/q)2 ) = 1/q. Therefore (34) is equivalent to the assertion:
q (y)
argmaxy∈[0,t] 1−H
= min{t, 1/q}.
ξ(y)
In terms of hA (y) :=
1−Hq (y)
ξ(y)
(35)
we must show
sign(h0A (y)) = sign(1/q − y),
y ∈ (0, θ).
q
q
A calculation shows that:
h0A (y)ξ(y)3/2
y(1−y)
θ(1−θ)
ln(q) =
q
y
1−θ
ln( yθ )
+
1−y
θ
ln( 1−y
) := G(y)
1−θ
Clearly sign(h0A (y)) = sign(G(y)). Therefore, we must show that sign(G(y)) = sign(1/q − y) for y ∈
(0, θ). Clearly G(1/q) = G(1 − θ) = 0. First we will prove that G(y) > 0 on [0, 1/q). We calculate:
r
p
− θ(1 − θ) G0 (y) =
Z
r
r
θ
y
Z
1−θ
1−y
ln(t)dt =
1
r
Z
1−θ
1−y
ln(t)dt +
θ
y
ln(t)dt.
1
We make the substitution t = 1/τ in the first integral to obtain:
r
p
− θ(1 − θ) G0 (y) =
Z
1
1−y
1−θ
r
ln(t)(1 −
1
)dt
t2
Z
+
r
θ
y
1−y
1−θ
ln(t)dt.
We note that t ≥ 1 in both the integrals, and hence both the integrands p
are non-negative.
Consequently,
p
the first integral is positive, and the second integral is also positive when θ/y > (1 − y)/(1 − θ). For
y ∈ [0, θ], this inequality is equivalent to (θ −y)(1−θ −y) > 0 which in turn is equivalentp
to y < 1−θ i.e.
0
y ∈ [0, 1/q). Thus, for y ∈ (0, 1/q), we have shown that G (y) < 0. Since G(0) = ln(q) q/(q − 1) > 0
and G(1/q) = 0, the fact that G(y) is strictly decreasing on [0, 1/q] implies G(y) > 0 on [0, 1/q).
14
Next we prove G(y) < 0 on (1/q, θ). Differentiating the expression for G0 (y) we get:
√
√
p
1−θ
θ
1−θ
4 θ(1 − θ) G00 (y) = ln( yθ ) y3/2
+ ln( 1−y
) (1−y)
3/2 .
Differentiating once more, we get:
p
4 θ(1 − θ) G000 (y) =
√
1−θ
(1
(1−y)5/2
1−θ
+ 23 ln( 1−y
)) −
√
θ
(1
y 5/2
+ 32 ln( yθ ))
√
θ
The second term − y5/2
(1 + 23 ln( yθ )) is negative on [1/q, θ) because θ/y > 1 on this interval. The first term
√
1−θ
(1 + 32
(1−y)5/2
1−θ
ln( 1−y
)) has the same sign as y − (1 − q −1 e2/3 ) for y ∈ [1/q, θ). It follows that G000 (y) < 0
for y ∈ [1/q, 1 − q −1 e2/3 ]. (We note that the condition for 1/q < 1 − q −1 e2/3 , is q ≥ 3, which is the case
here).
√
1−θ
1−θ
3
For y ∈ (1 − q −1 e2/3 , θ), as above (1−y)
5/2 (1 + 2 ln( 1−y )) is positive. It is also an increasing function of
y, because (1−θ)/(1−y) increases
with y. For y ∈ [1−q −1 e2/3 , θ], we note that θ/y decreases with y and
√
θ
θ/y ≥ 1. Therefore the term − y5/2 (1 + 23 ln( yθ )) increases with y. Thus G000 (y) is an increasing function of
y for y ∈ [1 − q −1 e2/3 , θ]. We note the boundary conditions on G000 (y): we have G000 (1/q) < 0 < G000 (θ).
To see this we note that
−4(θ(1 − θ))3 G000 (1/q) = 23 ln(q − 1)(θ3 + (1 − θ)3 ) + (θ3 − (1 − θ)3 ) > 0
because q > 2 is equivalent to θ > 1 − θ as well as ln(q − 1) > 0. Also
G000 (θ) =
2θ−1
4(θ(1−θ))2.5
> 0.
Since G000 (1 − q −1 e2/3 ) < 0 < G000 (θ) and G000 (y) is increasing on [1 − q −1 e2/3 , θ], we conclude that there
is a unique y0 in the interior of this interval such that G000 (y) has the same sign as y − y0 on this interval.
Together with the fact that G000 (y) < 0 on [1/q, 1 − q −1 e2/3 ], we obtain:
sign(G000 (y)) = sign(y − y0 )
on [1/q, θ].
This is illustrated in Figure 2, which shows the graphs of G(y) (dashed plot) and G000 (y) on [1/q, θ] for
q = 8. The point (y, G000 (y)) for y = 1 − q −1 e2/3 is marked. (In this plot, the values of G000 (y) are indicated
on the right-vertical axis, and the values of G(y) are indicated on the left-vertical axis). Thus G00 (y) is
decreasing on [1/q, y0 ] and increasing on [y0 , θ]. Since G00 (θ) = 0, it follows that G00 (y) < 0 on [y0 , θ).
We note that
G00 (1/q) = ln(q−1)(2θ−1)
> 0.
4(θ(1−θ))2
Thus G00 (1/q) > 0 > G00 (y0 ) together with the fact that G00 (y) is decreasing on [1/q, y0 ] implies that there
is a unique y1 in the interior of this interval such that G00 (y) has the same sign as y1 − y on this interval.
We have already shown that G00 (y) < 0 on [y0 , θ]. Thus we conclude
sign(G00 (y)) = sign(y1 − y)
on [1/q, θ).
This implies G0 (y) is increasing on [1/q, y1 ] and decreasing on [y1 , θ]. Since G0 (θ) = 0, we conclude that
G0 (y) > 0 on [y1 , θ). We note that
Z √q−1
G0 (1/q) = √ −1
ln(t)(1 − t12 )dt < 0.
θ(1−θ)
0
1
0
Since G (y) is increasing on [1/q, y1 ] and G (1/q) < 0 < G0 (y1 ), we conclude that there is a unique y2
in the interior of the interval [1/q, y1 ] such that G0 (y) has the same sign as y − y2 on this interval. Also
G0 (y) > 0 on [y1 , θ]. Thus we conclude:
sign(G0 (y)) = sign(y − y2 )
on [1/q, θ).
15
100
-0.02
0
-0.04
-100
-0.06
-200
-0.08
-300
-0.1
-400
-0.12
0.2
0.4
y
0.6
G000(y)
G(y)
0
-500
0.8
Fig. 2. Graphs of G(y) and G000 (y) on [1/q, θ] for q = 8.
This implies that G(y) is decreasing on [1/q, y2 ] and increasing on [y2 , θ]. Since G(1/q) = G(θ) = 0, we
see that G(y) is negative on (1/q, y2 ] as well as [y2 , θ). This finishes the proof of the assertion G(y) < 0
on (1/q, θ), and hence of (35).
0
Next, we prove the ∪-convexity of α̃M RRW (x). We must show that the derivative α̃M
RRW (x) is non2
decreasing. Since the derivative is constant on [0, (1 − 2/q) ], the problem reduces to showing that
αM RRW (x) is ∪-convex for x ∈ [(1 − 2/q)2 , θ]. This follows from the next lemma:
Lemma 6. The first MRRW bound αM RRW (δ) is ∪-convex if q = 2. For q > 2, it is ∩-convex on
[0, δM RRW ] and ∪-convex on [δM RRW , θ] where δM RRW satisfies:
1
2
√
−
q−1
q
< δM RRW < (1 − 2q )2 .
Proof: Let y = ξ(x). Let
χ(y) = 1 − 2y + (2θ − 1)
We calculate:
y 02
2y 00 y(1−y)
=
q
x(1−x)
θ(1−θ)
q
y(1−y)
θ(1−θ)
= χ(y)
(36)
p
Since x(1 − x)/(θ(1 − θ)) is non-negative, we also make the observation that that χ(y) > 0 for all
y ∈ [0, θ). Since αM RRW (x) = Hq (y), we get:
00
ln(q)αM
RRW (x) =
−y 02
y(1−y)
+ y 00 ln (q−1)(1−y)
.
y
Since ξ(ξ(x)) = x, we get y 0 = ξ 0 (x) = 1/ξ 0 (y). Using this we get:
00
0
2
αM
RRW (x)(ξ (y)) y(1 − y) ln(q) = −1 +
2y 00 y(1−y)
y 02
ln
q
(q−1)(1−y)
.
y
Using (36), we obtain:
r
00
0
2
αM
RRW (x)(ξ (y)) y(1
− y) ln(q) = −1 +
ln
(1−y)(q−1)
y
χ(y)
00
Let y ∈ (0, θ). We recall note χ(y) > 0 for y ∈ (0, θ). Thus for αM
RRW (x) has the same sign as
q
G2 (y) := ln( (1−y)(q−1)
) − χ(y).
y
16
We calculate:
G02 (y)y(1 − y) = χ(y)(y − 12 ).
Therefore, sign(G02 (y)) = sign(y − 1/2). In other words G2 (y) is decreasing on [0, 1/2] and increasing
on [1/2, θ). We note G2 (1/q) = ln(q − 1) − 2(1 − 2q ). The function
t 7→ ln(t − 1) − 2(1 − 2/t),
evaluates to 0 at t = 2, and is an increasing function of t for t ≥ 2 (because its derivative (1−2/t)2 /(t−1)
is positive). Thus G2 (1/q) > 0 for q > 2 and G2 (1/q) = 0 for q = 2. Since G2 (0) = +∞ and G2 (y)
is decreasing on [0, 1/q], we conclude that G2 (y) > 0 on [0, 1/q] if q > 2. If q = 2, then G2 (y) ≥ 0 on
[0, 1/q] = [0, θ]. In particular, for q = 2 the bound αM RRW (x) is ∪-convex on [0, θ].
√q−2 ).
q−1
√
2 t−1−t
<
4(t−1)3/2
For q > 2, we note that G2 (1/2) = 21 (ln(q − 1) −
The function b(t) = 12 (ln(t − 1) −
√t−2 )
t−1
satisfies b(3) = 12 (ln(2) − √22 ) < 0 and b0 (t) =
0 for t ≥ 3. Thus G2 (1/2) < 0 for all
q > 2. Since G2 (1/q) > 0 and G2 (1/2) < 0 and G2 (y) is decreasing on [1/q, 1/2], we conclude
that there is a yM RRW ∈ (1/q, 1/2) such that sign(G2 (y)) = sign(yM RRW − y) for y ∈ [0, 1/2]. Also
G2 (θ) = 0, G2 (1/2) < 0 and G2 (y) is increasing on [1/2, θ], which shows that G2 (y) < 0 on [1/2, θ). Thus
00
sign(G2 (y)) = sign(yM RRW − y) for y ∈ (0, θ). Since αM
RRW (x) has the same sign as G2 (y) (where
00
y = ξ(x)), we finally obtain sign(αM RRW (x)) = sign(x − δM RW W ) for
√ x ∈ (0, θ), where δM RW W =
1
ξ(yM RRW ) satisfies ξ(1/2) < δM RW W < ξ(1/q), or in other words: 2 − q−1
< δM RRW < (1 − 2q )2 . This
q
completes the proof of the lemma.
R EFERENCES
[1] R. Ahlswede and L. H. Khachatrian, “The diametric theorem in Hamming spaces—optimal anticodes,” Adv. in Appl. Math., vol. 20,
no. 4, pp. 429–449, 1998.
[2] M. Aaltonen, “A new upper bound on nonbinary block codes,” Discrete Math., vol. 83, no. 2-3, pp. 139–160, 1990.
[3] Y. Ben-Haim and S. Litsyn, “A new upper bound on the rate of non-binary codes,” Adv. Math. Commun., vol. 1, no. 1, pp. 83–92, 2007.
[4] P. Delsarte, “An algebraic approach to the association schemes of coding theory,” Philips Res. Rep. Suppl., no. 10, pp. vi+97, 1973.
[5] T. Laihonen and S. Litsyn, “On upper bounds for minimum distance and covering radius of non-binary codes,” Des. Codes Cryptogr.,
vol. 14, no. 1, pp. 71–80, 1998.
[6] M. Aaltonen, “Bounds on the information rate of a tree code as a function of the code’s feedback decoding minimum distance,” Ann.
Univ. Turku. Ser. A I, no. 181, 1981.
[7] M. Tsfasman, S. Vlăduţ, and D. Nogin, Algebraic geometric codes: basic notions, ser. Mathematical Surveys and Monographs.
Providence, RI: American Mathematical Society, 2007, vol. 139.
| 7cs.IT
|
arXiv:1803.08961v1 [math.AC] 23 Mar 2018
COHEN-MACAULAY CRITERIA FOR PROJECTIVE MONOMIAL
CURVES VIA GRÖBNER BASES
JÜRGEN HERZOG, DUMITRU I. STAMATE
Abstract. We prove new characterizations based on Gröbner bases for the CohenMacaulay property of a projective monomial curve.
Contents
Introduction
1. A Cohen-Macaulay criterion via Gröbner bases
2. Dual sequences
3. A bound for the number of generators of in(I(a)), when C(a) is
arithmetically Cohen-Macaulay
4. Applications
References
1
3
5
10
12
14
Introduction
Let K be any field. For any sequence of distinct positive integers a : a1 , . . . , an
we denote I(a) the kernel of the K-algebra homomorphism φ : S → K[t] where
S = K[x1 , . . . , xn ] and φ(xi ) = tai for i = 1, . . . , n. The image of this map is the
semigroup ring over K of the semigroup H generated by a1 , . . . , an . We no not insist
that a is a minimal generating set for H.
In the following, we assume that gcd(a1 , . . . , an ) = 1 and an > ai for all i < n.
We note that the homogenization of I(a) with respect to the variable x0 is again a
toric ideal, namely it is the kernel of the K-algebra map ψ : S[x0 ] → K[s, t] where
ψ(xi ) = tai san −ai for 1 ≤ i ≤ n and ψ(x0 ) = san . The image of the map ψ is the
subalgebra K[A] of K[t, s] generated by the monomials whose exponents are read
from the columns of the matrix
0
a1
...
an−1
an
(1)
A=
.
an an − a1 . . . an − an−1 0
In case K is an algebraically closed field, I(a) is the vanishing ideal of the affine
curve C(a) given parametrically by t 7→ (ta1 , . . . , tan ), while I(a)h is the vanishing
2010 Mathematics Subject Classification. Primary 13H10, 13P10, 16S36; Secondary 13F20,
14M25.
Key words and phrases. arithmetically Cohen-Macaulay, projective monomial curve, revlex,
Gröbner basis, numerical semigroup, Apéry set.
1
ideal of the projective closure C(a) of C(a) in Pn , given parametrically by [s : t] 7→
[tan : sa1 tan −a1 : · · · : san−1 tan −an−1 : san ]. Curves of this type are called projective
monomial curves.
The projective monomial curve C(a) is called arithmetically Cohen-Macaulay if
its vanishing ideal I(a)h is a Cohen-Macaulay ideal. It is known that this is the case
if and only if tan , san is a regular sequence on K[A]. This fact is a special case of
[10, Theorem 2.6].
Arithmetically Cohen-Macaulay curves are not rare among the projective monomial curves. It follows from Vu’s [18, Theorem 5.7] that for any fixed a, the curve
C(a1 + k, . . . , an + k) is arithmetically Cohen-Macaulay for all k ≫ 0. In small
embedding dimension, Bresinsky, Schenzel and Vogel [4] characterized the arithmetically Cohen-Macaulay projective monomial curves in P3 by the property that
I(a)h is generated by at most 3 elements.
In the context of numerical semigroups, Gröbner bases have been used in algorithms in [16] and [17] to find the Frobenius number of the numerical semigroup H
generated by a, or to characterize when is the tangent cone of the semigroup algebra
K[H] Cohen-Macaulay, see [1], [2].
One of the main results in this paper is that C(a) is arithmetically Cohen–
Macaulay if and only if in(I(a)h ), respectively in(I(a)), is a Cohen-Macaulay ideal,
see Theorem 1.2 (b) and (c). Here the initial ideals are taken for a reverse lexicographic order for which xn , x0 , respectively xn are the smallest variables. These
conditions are also equivalent to condition (f) which says that
in(xn , I(a)) = (xn , in(I(a))).
Yet other equivalent properties are (d) and (e), namely that xn , respectively x0 , xn ,
do not divide any minimal monomial generator of in(I(a)) and of in(I(a)h ), respectively, where the monomial orders are as before.
A Cohen-Macaulay criterion for a simplicial semigroup ring in terms of the Gröbner
basis of its defining ideal is given by Kamoi in [15, Corollary 2.9] and [14, Theorem
1.2]. In the particular case considered in this paper, equivalences (a), (d) and (e) in
Theorem 1.2 sharpen Kamoi’s mentioned results and his [14, Corollary 3.6].
The dual sequence of a is defined to be the sequence a′ : an − a1 , . . . , an − an−1 , an .
The projective monomial curves associated to the sequences a and a′ are obviously
isomorphic. So it is natural to compare the ideals I(a) and I(a′ ) and their reduced
Gröbner bases. That is the focus of Section 2.
P
For w = (w1 , . . . , wn ) ∈ Zn we denote hw, ai = ni=1 wi ai , and we set L(a) =
{w ∈ Zn : hw, ai = 0}. Obviously, I(a) is just the lattice ideal of the lattice
+
−
L(a). Indeed, I(a) is generated by the binomials fw = xw − xw with wP∈ L(a).
Let σ : Zn → Zn be the map given by σ(w1 , . . . , wn ) = (w1 , . . . , wn−1 , − ni=1 wi ).
Then σ is an automorphism of the group Zn such that σ 2 = idZn which induces an
isomorphism between L(a) and L(a′ ). In particular, L(a′ ) = (fσ(w) : w ∈ L(a)).
In general, a minimal set of binomial generators of L(a) is not mapped to a minimal set of binomial generators of L(a′ ), see Remark 3.2. However, in Theorem 2.2 we
show that C(a) is arithmetically Cohen-Macaulay if and only if in(fσ(w) ) = in(fw )
2
for all fw ∈ G, where G denotes the reduced Gröbner basis of I(a) with respect to a
reverse lexicographic monomial order with xn the smallest variable. Moreover, these
conditions are also equivalent to the fact that fw ∈ G if and only if fσ(w) ∈ G ′ , for
all w ∈ Zn , where G ′ is the reduced Gröbner basis of I(a′ ) with respect to the same
monomial order.
Let H denote the numerical semigroup generated by a. For any nonzero element h
in H its Apéry set is defined as Ap(H, h) = {x ∈ H : x−h ∈
/ H}. For h ∈ Ap(H, an )
we denote ϕa (h) the smallest monomial in S for the reverse lexicographic order such
that its a-degree equals h. The close relationship between the ideals I(a) and I(a′ )
is also outlined by the fact that the curve C(a) is arithmetically Cohen-Macaulay if
and only if
in(xn , I(a)) = in(xn , I(a′ )),
see Theorem 2.6. Here one uses a reverse lexicographic order with xn the smallest
variable. For the proof, a key observation is that the latter equation is equivalent
to the fact that for all h in Ap(H, an ) the a′ -degree of ϕa (h) is in Ap(H ′ , an ), where
H ′ denotes the semigroup generated by the dual sequence a′ . As a consequence,
in Corollary 2.7 we recover a criterion of Cavaliere and Niesi ([5]) for C(a) to be
arithmetically Cohen-Macaulay.
When n = 3, it is known from [11] that µ(I(a)) ≤ 3. However, we give examples
showing that a reduced Gröbner basis may be arbitrarily large, see Proposition 3.1.
In case C(a) is arithmetically
Cohen-Macaulay, in Proposition 3.4 we show that
an
µ(in(I(a))) ≤ n−2 .
In Section 4 we apply Theorem 1.2 to test the arithmetically Cohen-Macaulay
property for two families of projective monomial curves in P3 that have appeared
in the literature. For these families of 4-generated numerical semigroups which
were introduced by Arslan ([1]) and by Bresinsky ([3]), respectively, we show that
the corresponding projective monomial curve is (respectively, is not) arithmetically
Cohen-Macaulay.
1. A Cohen-Macaulay criterion via Gröbner bases
The following lemma appears in [7, Exercise 5, page 392]. Lacking reference to a
proof, we prefer to provide one.
Lemma 1.1. Let I be an ideal in the polynomial ring S = K[x1 , . . . , xn ] and I h ⊂
S[x0 ] its homogenization with respect to the variable x0 . We denote < any reverse
lexicographic monomial order on S and <0 the reverse lexicographic monomial order
on S[x0 ] extended from S and such that x0 is the least variable.
If f1 , . . . , fr is the reduced Gröbner basis for I with respect to <, then f1h , . . . , frh
is the reduced Gröbner basis for I h with respect to <0 . Moreover, in<0 (I h ) =
(in< (I))S[x0 ].
Proof. Let F h = {f1h , . . . , frh }. It is proved in [9, Proposition 3.15] that F h is a
Gröbner basis for I h with respect to the block order <′ on S[x0 ] which is defined as
xα xa0 <′ xβ xb0 if (xα < xβ ) or (xα = xβ and a < b)
3
for all α, β ∈ Nn and all nonnegative integers a, b.
Let f be a nonzero polynomial in I. We write f = m1 + · · · + mq as a sum of
monomials with mi > mi−1 for 2 ≤ i ≤ q. Then deg mi ≥ deg mi−1 for 2 ≤ i ≤ q
and
ε
f h = m1 + m2 xε02 + · · · + mq x0q ,
where εi = deg m1 − deg mi for i = 2, . . . , q. Moreover, in the above decomposition
of f h the monomials are listed decreasingly with respect to <0 . Thus in<0 (f h ) =
m1 = in< (f ) = in<′ (f h ) for all f in I. It follows that
in<′ (I h ) = in<′ (fih : 1 ≤ i ≤ r) = (in<′ (fih ) : 1 ≤ i ≤ r)
= (in<0 (fih ) : 1 ≤ i ≤ r) ⊆ in<0 (I h ).
Since the homogeneous ideals in<′ (I h ) and in<0 (I h ) have the same Hilbert function,
we conclude that they are equal and that F h is a Gröbner basis for I h with respect
to <0 .
Assume there exist i, j such that in<0 (fih ) divides a monomial in tail(fjh ), i.e.
in< (fi ) divides mxε0 with m a monomial in tail(fj ). This implies that in< (fi ) divides
m, which contradicts the fact that F is the reduced Gröbner basis for I with respect
to <. Therefore F h is the reduced Gröbner basis for I h with respect to <0 .
The following theorem is one of the main results of this paper.
Theorem 1.2. Let a : a1 , . . . , an be a sequence of positive integers with an > ai for
all i < n. Denote < any reverse lexicographic order on S = K[x1 , . . . , xn ] such that
xn is the smallest variable and <0 the induced reverse lexicographic order on S[x0 ],
where xn > x0 . The following conditions are equivalent:
(a) the projective monomial curve C(a) is arithmetically Cohen-Macaulay;
(b) in<0 (I(a)h ) is a Cohen-Macaulay ideal;
(c) in< (I(a)) is a Cohen-Macaulay ideal;
(d) xn does not divide any element of G(in< (I(a)));
(e) xn and x0 do not divide any element of G(in<0 (I(a)h ));
(f) in< (xn , I(a)) = (xn , in< (I(a)).
Proof. Lemma 1.1 implies that G(in< (I(a))) = G(in<0 (I(a)h )). Therefore (b) ⇐⇒
(c) and (d) ⇐⇒ (e). The implication (b) ⇒ (a) is a general fact, see for example
[12, Corollary 3.3.5]. Assuming (e), we get that x0 , xn is a regular sequence modulo
in<0 (I(a)h ), which implies (b).
Since xn is regular on S/I(a) which is a domain, using [6, Proposition 1.4] we
have that xn is regular on S/ in< (I(a)) if and only if in< (xn , I(a)) = (xn , in< (I(a)).
This shows (d) ⇐⇒ (f).
It remains to prove that (a) ⇒ (e). It is known that the ring K[A] is CohenMacaulay if and only if san , tan is a regular sequence on it, see [10, Lemma 2.4].
That is equivalent to x0 , xn being a regular sequence on S[x0 ]/I(a)h .
Let F = {f1 , . . . , fr } be the reduced Gröbner basis of I(a) with respect to <. For
all i = 1, . . . , r, fi is a binomial of the form ui − vi , where ui = in< (fi ), and ui and
vi have disjoint supports. After a reordering, we may assume that there exists ℓ ≤ r
such that fi is homogeneous if and only if i ≤ ℓ.
4
Using Lemma 1.1, we see that G(in<0 (I(a)h )) = {u1, . . . , ur }. Clearly, x0 does
not divide ui for any i, since deg ui ≥ deg vi .
We first show that for i ≤ ℓ, ui is not divisible by xn . Indeed, if that were not
the case, by the properties of the chosen monomial order <, we would have that xn
divides vi , as well, hence ui and vi do not have disjoint supports, which is false.
It remains to prove that xn does not divide ui for any i > ℓ. For that, we set
gi = fih |x0 =0 for all i = 1, . . . , r and J = (g1 , . . . , gr )S. Then K[A]/(san ) ∼
= S/J.
For i ≤ ℓ, since fi is homogeneous we obtain that x0 does not divide vi . Therefore,
J = (f1 , . . . , fℓ , uℓ+1, . . . , ur ). Since xn is regular on S/J, we infer that xn is not
in the support of any of the monomials ui with i > ℓ. This finishes the proof of
(a) ⇒ (e).
2. Dual sequences
Given a : a1 , . . . , an a sequence of distinct nonnegative integers such that an > ai
for all i < n, the dual sequence is defined to be a′ : an − a1 , . . . , an − an−1 , an . It is
clear that this procedure is a duality: (a′ )′ = a.
The projective monomial curves associated to the sequences a and a′ are isomorphic. Indeed, the ideals I(a)h and I(a′ )h are the kernel of the maps on S[x0 ] sending
x0 , . . . , xn to the monomials having as exponent vectors the columns of the matrix
A in (1), and respectively the columns of the matrix
0 an − a1 . . . an − an−1 an
′
A =
.
an
a1
...
an−1
0
This implies that the polynomials in the ideal I(a′ )h are obtained from those in
I(a)h by switching the variables x0 and xn .
In this section we compare the Gröbner bases of the ideals I(a) and I(a′ ) with
respect to a reverse lexicographic order, in connection to the Cohen-Macaulay property of the associated projective monomial curve.
P
Let σ : Zn → Zn be the map given by σ(w1 , . . . , wn ) = (w1 , . . . , wn−1 , − ni=1 wi ).
It is easy to see that σ is an automorphism P
of the group Zn and that σ 2 = idZn . For
n
w = (w1 , . . . , wn ) ∈ Z we denote hw, ai = ni=1 wi ai . We set
L(a) = {w ∈ Zn : hw, ai = 0}.
Lemma 2.1. With notation as above, the map σ induces an isomorphism between
the groups L(a) and L(a′ ).
P
Proof. If w = (w1 , . . . , wn ) ∈ Zn with ni=1 wi ai = 0 then
!
!
n
n
X
X
(an − ai )wi −
wi an = 0,
i=1
i=1
′
′
equivalently hσ(w), a i = 0 and σ(w) ∈ L(a ). Similarly, if w ′ ∈ L(a′ ) then σ(w ′ ) ∈
L(a′′ ) = L(a).
If the entries of the vector α = (α1 , . . . , αn ) are nonnegative integers we let xα =
· · · xαnn . For w = (w1 , . . . , wn ) ∈ Zn , let w + and w − be the unique vectors with
xα1 1
5
nonnegative entries and disjoint supports such that w = w + − w − . We denote
+
−
fw = xw − xw . It is clear that f−w = −fw . Therefore, a difference of two
monomials with disjoint supports can be identified with a vector w ∈ Zn .
It is known that I(a) = (fw : w ∈ L(a)), hence I(a′ ) = (fσ(w) : w ∈ L(a)).
However, it is not always true that σ maps a minimal generating set (or a Gröbner
basis) for I(a) into a minimal generating set (or a Gröbner basis) for I(a′ ), see
Proposition 3.1 and Remark 3.2.
Theorem 2.2. Let a : a1 , . . . , an be a sequence of nonnegative integers with an > ai
for all i < n. Let G and G ′ be the reduced Gröbner bases of I(a) and I(a′ ), respectively, with respect to a reverse lexicographic monomial order on S = K[x1 , . . . , xn ]
such that xn is the smallest variable. The following conditions are equivalent:
(a) the projective monomial curve C(a) is arithmetically Cohen-Macaulay;
(b) in(fσ(w) ) = in(fw ), for all fw ∈ G;
(c) fw ∈ G ⇐⇒ fσ(w) ∈ G ′ , for all w ∈ Zn .
Proof. We first prove that conditions (a) and (b) are equivalent.
(a) ⇒ (b): We assume that I(a)h is a Cohen-Macaulay ideal in S[x0 ]. We pick
fw in G, where w = (w1 , . . . , wn ). We denote w ′ = σ(w) = (w1′ , . . . , wn′ ). Since the
P
+
leading coefficient
LC(fw ) = 1 we get that in(fw ) = xw , hence d = ni=1 wi ≥ 0
P
and wn′ = − ni=1 wi = −d ≤ 0.
By Theorem 1.2 we obtain that xn does not divide in(fw ), hence wn ≤ 0. ConseP
′ +
′ −
quently, (w ′ )+ = w + . Also, ni=1 wi′ = −wn ≥ 0, and hence deg x(w ) ≥ deg x(w ) .
We distinguish two cases.
′ +
+
′ +
′ −
Firstly, if wn < 0 then deg x(w ) > deg x(w ) , hence in(fw′ ) = x(w ) = xw =
−
′ −
n
in(fw ). Moreover, x(w ) = xw · xd+w
.
n
′ +
′ −
Secondly, in case wn = 0 we get that deg x(w ) = deg x(w ) . Now, if d = 0 then
′ +
′ −
w ′ = w. If d > 0, for the chosen monomial order we obtain that x(w ) > x(w ) ,
′ −
−
because x(w ) = xw · xdn .
+
Thus in(fw′ ) = xw = in(fw ) in all cases, and property (b) holds.
(b) ⇒ (a): If I(a)h is not a Cohen-Macaulay ideal, by Theorem 1.2 there exists
fw in G such that xn divides in(fw ). Let w = (w1 , . . . , wn ) and w ′ = σ(w) =
+
(w1′ , . . . , wn′ ). Since LC(fw ) = 1 we get that in(fw ) = xw , hence wn > 0 and
P
n
i=1 wi ≥ 0.
There exists i0 6= n such that wi0 > 0, otherwise, since hw, ai = 0 we get that
Pn−1
Pn−1
i=1 ai (−wi ) = wn an ≥ −(w1 + · · · + wn−1 )an > −
i=1 ai wi , which is a contradiction. P
′ −
′ +
Since ni=1 wi′ = −wn < 0, we obtain that deg x(w ) > deg x(w ) and in(fw′ ) =
′
−
′ +
x(w ) . As i0 < n, we have that wi′0 = wi0 > 0, hence xi0 divides x(w ) . On the other
′ −
hand, condition (b) implies that in(fσ(w) ) = in(fw ), and therefore xi0 divides x(w )
as well, which gives a contradiction. We conclude that the projective monomial
curve C(a) is Cohen-Macaulay.
Next we prove that (a),(b) ⇒ (c). From the proof above of the equivalence
(a) ⇐⇒ (b), we see that under the assumption that (a) (hence also (b)) holds, for
6
all fw in G one has
(2)
tail(fσ(w) ) = tail(fw ) · xan for some integer a.
From Theorem 1.2 we have that in(I(a′ )) = in(I(a)), therefore property (b) implies
that G ′′ = {fσ(w) : fw ∈ G} is a minimal Gröbner basis for I(a′ ). We show that it is
reduced.
+
Let fw ∈ G with w = (w1 , . . . , wn ) and σ(w) = (w1′ , . . . , wn′ ). Then in(fw ) = xw
+
−
and deg(xw ) ≥ deg(xw ). Condition (a) and Theorem 1.2 imply that wn ≤ 0. Thus
Pn
′ )−
′
(w ′ )+
) ≥ deg(x(w P
).
i=1 wi = −wn ≥ 0, which implies that deg(x
′
′
If wn = 0 then w is homogeneous and either w = w (if ni=1 wi = 0), or xn
P
′ +
′ −
divides x(w ) (if ni=1 wi > 0), hence x(w ) = in(fσ(w) ) and LC(fσ(w) ) = 1.
′ +
′ −
If wn < 0 then deg(x(w ) ) > deg(x(w ) ), and again LC(fσ(w) ) = 1.
If there are fw and fwe in G such that in(fσ(w) ) divides tail(fσ(w)
e ), then, since xn
is not in the support of in(fw ) = in(fσ(w) ), by using (2) we get that in(fw ) divides
tail(fwe ). This contradicts the fact that G is the reduced Gröbner basis for I(a).
Hence G ′′ = G ′ , which proves (c).
For (c) ⇒ (a): If we assume that (c) holds, but I(a)h is not Cohen-Macaulay, then
+
by Theorem 1.2 there exists fw in G such that xn divides in(fw ) = xw . Let σ(w) =
P
−
+
(w1′ , . . . , wn′ ). Since ni=1 wi′ = −wn < 0, it follows that deg(xσ(w) ) > deg(xσ(w) ),
−
hence in(fσ(w) ) = xσ(w) . On the other hand, property (c) implies that fσ(w) ∈ G ′ ,
+
hence in(fσ(w) ) = xσ(w) , which is a contradiction. Therefore, property (a) holds.
This ends the proof of the theorem.
Remark 2.3. The fact that the involution σ maps some (minimal) Gröbner basis
for I(a) into a (minimal) Gröbner basis for I(a′ ) is not enough to imply that the
curve C(a) is arithmetically Cohen-Macaulay.
Indeed, let a : 4, 13, 19. Then a′ : 15, 6, 19. A Singular [8] computation shows
that
G = {y 5 − x2 z 3 , x3 y 2 − z 2 , x5 z − y 3 , x8 − yz}, and
G ′ = {y 5 − x2 , x3 y 2 − z 3 , y 3z 3 − x5 , x8 − yz 6 }
are the reduced Gröbner bases with respect to the reverse lexicographic order with
x > y > z > t for the ideals I(a) and I(a′ ), respectively.
One has G = {fw1 , fw2 , fw3 , fw4 }, where w1 = (−2, 5, −3), w2 = (3, 2, −2), w3 =
(5, −3, 1) and w4 = (8, −1, −1). Since σ(w1 ) = (−2, 5, 0), σ(w2 ) = (3, 2, −3),
σ(w3 ) = (5, −3, −3) and σ(w4 ) = (8, −1, −6) we note that
G ′ = {fσ(w1 ) , fσ(w2 ) , −fσ(w3 ) , fσ(w4 ) }.
This means that {fσ(w1 ) , fσ(w2 ) , fσ(w3 ) , fσ(w4 ) } is a minimal Gröbner basis for I(a′ ),
although different from the reduced one G ′ .
Let a = a1 , . . . , an be a sequence of nonnegative integers with gcd(a1 , . . . , an ) = 1
and an > ai for all i = 1, . . . , n−1. Our next goal is to describe the Cohen-Macaulay
property of C(a) in terms of the Apéry sets of the semigroup generated by a or the
7
dual sequence a′ . We recall that for a numerical semigroup H and 0 6= h in H, the
Apéry set of H with respect to h is
Ap(H, h) = {x ∈ H : x − h ∈
/ H}.
It is known that | Ap(H, h)| = h and that the elements of Ap(H, h) have distinct
residues modulo h.
The sequence a induces a grading on S = k[x1 , . . . , xn ] by letting deg(xi ) = ai
for all i = 1,P. . . , n. For any monomial xα = xr11 · · · xrnn we denote its a-degree by
dega (xα ) = ni=1 ri ai . We denote H the numerical semigroup generated by a. For
any h in Ap(H, an ) we denote ϕa (h) the smallest monomial in S (with respect to
a reverse lexicographic monomial order where xn is the smallest variable) such that
its a-degree equals h. Since h − an ∈
/ H we see that the monomial ϕa (h) is in
′
S = k[x1 , . . . , xn−1 ].
Proposition 2.4. With notation as above, for any h in Ap(H, an ) the monomial
ϕa (h) ∈
/ (xn , in(I(a)))S.
Proof. Let h ∈ Ap(H, an ) and ϕa (h) = xα . Assume that xα ∈ in(xn , I(a)). Then
xα = in(F ) for some F in (xn , I(a)). Since the ideal (xn , I(a)) is generated by
monomials and binomials which are the difference of two monomials with the same
a-degree, without loss of generality we may assume that F is a-homogeneous. Thus
we may write
xα = in(xn f + f1 fw1 + · · · + fq fwq ),
where w1 , . . . , wq ∈ L(a), and f, f1 , . . . , fq are a-homogeneous with dega (xn f ) =
dega (f1 fw1 ) = · · · = dega (fq fwq ).
We notice that f = 0, otherwise h = dega (xα ) = dega (xn f ) = an + dega (f ),
gives that h − an ∈ H, which is false. Hence xα ∈ in(I(a)). The ideal I(a) has
a Gröbner basis of binomials, hence we can write xα = m · in(fw ) for some binomial fw ∈ I(a) and m a monomial in S ′ . Without loss of generality, we may
+
+
−
−
assume in(fw ) = xw . Thus xw > xw , which gives ϕa (h) = xα > xw m. But
−
dega (w + ) = dega (w − ), hence dega (xα ) = dega (xw m), which contradicts the choice
of xα . Therefore, ϕa (h) ∈
/ (xn , in(I(a))).
If we identify a monomial which is in S and not in in(xn , I(a)) with its residue class
modulo the monomial ideal in(xn , I(a)), by Proposition 2.4 the assignment ϕa (−)
defines a map from Ap(H, an ) into Mon(S/ in(xn , I(a))), the K-basis of monomials
of S/ in(xn , I(a)). We prove that this is a bijection.
Proposition 2.5. The map ϕa : Ap(H, an ) → Mon(S/ in(xn , I(a))) is bijective.
Proof. Let h, h′ in Ap(H, an ) with ϕa (h) = ϕa (h′ ). Then h = dega (ϕa (h)) =
dega (ϕa (h′ )) = h′ , and the map ϕa is injective.
By Macaulay’s theorem ([9, Theorem 2.6]) the monomials in S which do not
belong to in(xn , I(a)) form a K-basis for S/(xn , I(a)). Therefore,
| Mon(S/ in(xn , I(a)))| = dimK S/(xn , I(a))
= dimK K[H]/(tan ) = an .
Since | Ap(H, an )| = an , we conclude that the map ϕa is bijective.
8
Theorem 2.6. Let a : a1 , . . . , an be a sequence of distinct positive integers such that
gcd(a1 , . . . , an ) = 1 and an > ai for all i = 1, . . . , n − 1. We denote a′ the dual
sequence of a. Let H and H ′ be the numerical semigroups generated by a and a′ ,
respectively. The following statements are equivalent:
(a) the projective monomial curve C(a) is arithmetically Cohen-Macaulay;
(b) in(xn , I(a)) = in(xn , I(a′ ));
(c) Mon(S/ in(xn , I(a))) = Mon(S/ in(xn , I(a′ )));
(d) dega′ (ϕa (h)) ∈ Ap(H ′ , an ) for all h in Ap(H, an ),
where the initial ideals are taken with respect to the reverse lexicographic term order
on S.
Proof. Assume (a) holds. It follows from Theorem 1.2 that in(xn , I(a)) = (xn , in(I(a)))
and in(xn , I(a′ )) = (xn , in(I(a′ ))). We get from Lemma 1.1 that G(in(I(a))) =
G(in(I h (a))) = G(in(I h (a′ ))) = G(in(I(a))), hence the statement (b) is true.
Clearly, the properties (b) and (c) are equivalent.
We now prove that (b) ⇐⇒ (d). Assume that (b) holds. Let h ∈ Ap(H, an ). By
Proposition 2.4 we have that the monomial ϕa (h) is not in in(xn , I(a)), hence it is
not in in(xn , I(a′ )). Using Proposition 2.5 we get that ϕa (h) = ϕa′ (h′ ) for some h′
in Ap(H ′ , an ). Hence dega′ (ϕa (h)) ∈ Ap(H ′, an ), which proves (d).
Conversely, we assume that (d) holds and we consider the monomial xα not in
in(xn , I(a)). By Proposition 2.5 there exists h in Ap(H, an ) such that ϕa (h) = xα .
Property (d) implies that there exists h′ in Ap(H ′, an ) such that h′ = dega′ (xα ),
which by Proposition 2.4 gives that xα ∈
/ in(xn , I(a′ )). Hence (d) ⇒ (b).
To finish the proof of the theorem we are left to show that (b) ⇒ (a). Assume
in(xn , I(a)) = in(xn , I(a′ )). By Theorem 1.2, it is enough to prove that xn does
not divide any monomial in G(in(I(a))). Assume there exists a monomial u · xcn in
G(in(I(a))) with u not divisible by xn and c > 0. Then u is not a constant, otherwise,
rn−1
in
since I(a) has a Gröbner basis of binomials, there exists f = xcn − xr11 . . . xn−1
Pn−1
c
I(a) with xn = in(f ). This implies that we have a relation c · an = i=1 ri ai with
Pn−1
ri , which is false since an > ai for all i < n.
c ≥ i=1
Let u · xcn = in(fw ), where w = (w1 , . . . , wn ) ∈ L(a). Without loss of generality we
+
−
may assume in(fw ) = xw , hence wn = c. Set v = xw and d = deg(u · xcn ) − deg v.
Then d > 0 by the above discussion. The sum of the components of σ(w) equals
Pn−1
Pn
−
σ(w)+
) < deg(xσ(w) ) and
i=1 wi + (−
i=1 wi ) = −wn = −c < 0, hence deg(x
fσ(w) = u − xdn · v. This gives that
u = xdn · v + fσ(w) ∈ (xn , I(a′ )),
and also that u ∈ in(xn , I(a′ )), which by our hypothesis (b) implies that u ∈
in(xn , I(a)). Since the ideal I(a) is a-homogeneous we can write
u = in(xn f + f1 fz1 + · · · + fq fzq ),
where z1 , . . . , zq ∈ L(a), and f, f1 , . . . , fq are a-homogeneous with dega (xn f ) =
dega (f1 fz1 ) = · · · = dega (fq fzq ). We see that f 6= 0, otherwise u ∈ in(I(a)), which
contradicts the fact that u · xcn is a minimal monomial generator for in(I(a)).
9
Let h = dega (u). Since f 6= 0 we get that h − an ∈ H. We may write h =
h1 + λn an with λn a maximal positive integer and h1 ∈ H, i.e. h1 ∈ Ap(H, an ). Let
λn−1
u1 = ϕa (h1 ) = xλ1 1 . . . xn−1
. Then the binomial f1 = u1 xλnn −u is in I(a). As u · xcn ∈
G(in(I(a))) with c > 1, we get that u ∈
/ in(I(a)), hence in(f1 ) = u1 · xλnn ∈ in(I(a)).
By Proposition 2.4, u1 ∈
/ in(xn , I(a)), hence u1 ∈
/ in(I(a)), as well. This implies
that u1 · xλnn is divisible by a monomial u2 xen ∈ G(in(I(a)) with xn and u2 coprime,
e > 0. Therefore u2 divides u1 , hence dega (u2 ) + h2 = dega (u1 ) for some positive h2
in H. This gives dega (u2 ) ∈ Ap(H, an ).
We may write u2 · xen = in(fwe ) with w
e ∈ L(a), and arguing as before we get that
d′
′
u2 = xn · v1 + fσ(w)
e ∈ in(xn , I(a)) for some positive d . Thus
u2 = in(xn f ′ + f1′ fz1′ + · · · + fℓ′ fzℓ′ ),
where z1′ , . . . , zℓ′ ∈ L(a), and f ′ , f1′ , . . . , fq′ are a-homogeneous with dega (xn f ′ ) =
dega (f1′ fz1′ ) = · · · = dega (fℓ′ fzℓ′ ). If f ′ = 0, then u2 ∈ in(I(a)), which is false since
u2 · xen ∈ G(in(I(a))). On the other hand, f ′ 6= 0 implies that an + dega (f ′ ) =
dega (u2 ), hence dega (u2 ) ∈
/ Ap(H, an ), which is also false. Therefore xn does not
divide any monomial in G(in(I(a))). This concludes the proof of the implication
(b) ⇒ (a) and of the theorem.
Let Ap(H, an ) = {0, ν1 , . . . , νan −1 }. We may assume that νi ≡ i mod an for
e
all i. For each νi , let µi ∈ H ′ be the smallest element such that (νi , µi ) ∈ H.
Note that µi ≡ −i mod an for all i. Cavalieri and Niesi [5] call Ap(H, an ) good, if
{0, µ1 , . . . , µnd −1 } = Ap(H ′ , an ).
As a consequence of Theorem 2.6 we obtain
Corollary 2.7. (Cavaliere-Niesi, [5, Theorem 4.6]) The projective monomial curve
C(a) is arithmetically Cohen-Macaulay if and only if Ap(H, an ) is good.
Pn−1
Pn−1
Proof. Let νi =
j=1 rj (an − aj )
j=1 rj aj with integer coefficients rj ≥ 0 and
Pn−1
Pn−1
Pn−1
minimal. Then µi = j=1 rj (an − aj ). Thus µi = ( j=1 rj )an − νi with j=1
rj
Pn−1
minimal and j=1 rj aj = νi .
Qn−1 sj
Pn−1
xj , then dega′ (ϕa (νi )) = i=1
This shows that if ϕa (νi ) = j=1
sj (an − aj ) =
Pn−1
Pn−1
Pn−1
( j=1 sj )an − j=1 sj aj = ( j=1 rj )an − νi = µi . Hence Theorem 2.6(a) ⇐⇒ (d)
yields the desired conclusion.
3. A bound for the number of generators of in(I(a)), when C(a) is
arithmetically Cohen-Macaulay
In this section we show by examples that the number of generators of in(I(a))
may be arbitrarily large, already if a has only 3 elements.
Proposition 3.1. For the integer h ≥ 2, let a = 4, 6h+1, 6h+7. Then µ(in(I(a))) =
h + 2, where the initial ideal is computed with respect to the reverse lexicographic
monomial order with x1 > x2 > x3 .
10
Proof. We first find I(a) using the method from [11]. For 1 ≤ i ≤ 3 we let ci be the
smallest prositive integer such that
(3)
ci ai = rij aj + rik ak ,
with rij , rik nonnegative integers and {i, j, k} = {1, 2, 3}. Since a1 , a2 , a3 are pairwise
coprime, it is known from [11] that the rij ’s are unique and positive, and ci = rji +rki
for all {i, j, k} = {1, 2, 3}.
From the equations (3h + 2)a1 = a2 + a3 and 2a3 = 3a1 + 2a2 we find c1 = 3h + 2,
c3 = 2 and the corresponding rij ’s from (3). Hence c2 = 3 and 3a2 = (3h−1)a1 +a3 is
the corresponding equation from (3). According to [11], the ideal I(a) is minimally
generated by f1 = x3h+2
− x2 x3 , f3 = x31 x22 − x23 and g1 = x3h−1
x3 − x32 .
1
1
We introduce recursively the polynomials gi+1 = S(gi , f3 ) for 1 ≤ i ≤ h − 1. It
3(h−i)+2 2i−1
follows easily by induction that gi = x1
x3 − x2i+1
, for 1 ≤ i ≤ h. We claim
2
that
(4)
G = {f1 , g1 , . . . , gh , f3 }
is the reduced Gröbner basis of I(a).
To prove that G is a Gröbner basis we need to check that the S-pairs of elements
in G reduce to zero with respect to G, see [9, Theorem 2.14]. Here are the relevant
computations.
G
S(f1 , f3 ) = x22 − x3h−1
f3 = x3 (x32 − x3h−1
x3 ) = −x3 g1 → 0.
1
1
G
S(gh , f1 ) → 0 since gcd(in(gh )), in(f1 )) = 1.
G
S(gh , f3 ) = x31 gh − x2h−1
f3 = x23 gh−1 → 0.
2
For 1 ≤ i ≤ h − 1 :
G
2i−1
3i 2i
2
3 2
S(gi , f1 ) = x3i
f1 = x2 (x2i
1 gi −x3
3 −x1 x2 ) = x2 ·(x3 −x1 x2 )·(. . . ) = x2 ·f3 ·(. . . ) →
0, and
G
S(gi , f3 ) = gi+1 → 0.
For 1 ≤ i < j < h:
2(i+1) 3(j−i) 2(j−i)
2(j−i) 2i+1
2(j−i)
3(j−i)
3(j−i) 2j+1
x2 = x2
(x1
x2
−
S(gi , gj ) = x3
gi −x1
gj = x1
x2 −x3
G
2(j−i)
2(i+1)
x3
) = x2
· f3 · (· · · ) → 0.
G
For 1 ≤ i < h we have that S(gi , gh ) → 0, since gcd(in(gi ), in(gh )) = 1.
By inspecting the binomials in G it follows that they are in fact the reduced
Gröbner basis for I(a). This shows that µ(in(I(a)) = |G| = h + 2.
Remark 3.2. The sequence dual to the one from Proposition 3.1 is
a′ = 6h + 3, 6, 6h + 7.
It is easy to check (using [11]) that the corresponding toric ideal is a complete
intersection I(a′ ) = (x2h+1
− x21 , x31 x22 − x33 ). In particular, this shows that the image
2
through the involution σ of a minimal set of binomial generators for I(a) may no
longer be a minimal generating system for I(a′ ).
Arguing as in the proof of Proposition 3.1, it is routine to verify that for the
reduced Gröbner basis G in (4), the set {fσ(w) : fw ∈ G} is a minimal Gröbner basis
for I(a′ ). The latter set of binomials is fully interreduced, yet it is not the reduced
11
Gröbner basis for I(a′ ) since the leading coefficients of the binomials coming from
g1 , . . . , gh−1 equal −1.
From Theorem 2.2 we infer that C(a) is not arithmetically Cohen–Macaulay. This
can also be seen from the fact that in(g1 ) = x3h−1
x3 ∈ G(in(I(a))) (as h > 1) and
1
using Theorem 1.2.
If C(a) is arithmetically Cohen-Macaulay, we give an explicit bound for µ(in< (I(a)))
depending on an , the largest element of the sequence a. To prove this we first show
Lemma 3.3. Let n ≥ 2 and I ⊂ S = K[x1 , . . . , xn ] be a graded ideal with mk ⊂ I,
where m = (x1 , .. . , xn ). Let < be a monomial
order on S. Then µ(I) ≤ µ(in< (I)) ≤
n+k−1
µ(mk ) = n+k−1
,
and
µ(I)
=
if
and
only
if I = mk .
n−1
n−1
Proof. It suffices to show that for a monomial ideal J ⊂ S with mk ⊂ J, one has
µ(J) ≤ µ(mk ), and µ(J) = µ(mk ), if and only if J = mk . We prove this by induction
on k − a, where a is the least degree of a monomial generator of J. If a = k,
the assertion is trivial. Suppose now that a < k. We denote by G(J) the unique
minimal set of monomial generators of J, and set G(J)j = {u ∈ G(J) : deg u = j}
for all j, and let J ′ = mJa + J≥a+1 , where J≥a+1 = (u ∈ J : deg u ≥ a + 1). Then
G(J ′ )j = 0 for j < a + 1 and mk ⊂ J ′ . Therefore, by our induction hypothesis,
we have µ(J ′ ) ≤ µ(mk ). On the other hand, G(J ′ )a+1 is the disjoint union of
G(mJ)a+1 and G(J)a+1 . Furthermore, G(J ′ )j = G(J)j for j > a + 1. Hence, since
|G(J)a | < |G(mJ)a+1 |, it follows that µ(J) = |G(J)| < |G(J ′ )| = µ(J ′ ) ≤ µ(mk ), as
desired.
Proposition 3.4. Suppose that the monomial curve C(a) is arithmetically Cohen–
an
.
Macaulay. Then µ(in(I(a)) ≤ n−2
Proof. As before we assume that a = a1 , . . . , an with an > ai for all i, and we let
H be the numerical semigroup generated by a. Then I(a) ⊂ S = K[x1 , . . . , xn ] and
G(in< (I(a)) ⊂ S̄ = K[x1 , . . . , xn−1 ], by Theorem 1.2. Therefore,
length(S̄/ in< (I(a)) = length(S/(xn , in< (I(a)))) = length(S/(xn , I(a))) =
length(K[H]/(tan )) = an .
Let k be the smallest number such that (x1 , . . . , xn−1 )k ⊂ in< (I(a)). Then
an =
k−1
X
dimK (S̄/ in< (I(a))j = 1 + (n − 1) +
j=0
k−1
X
dimK (S̄/ in< (I(a))j
j=2
≥ 1 + (n − 1) + (k − 2) = (n − 2) + k.
Thus, k ≤
(n
− 1) + 1, and hence by Lemma 3.3 we get µ(in< (I(a))) ≤
an −
(n−1)+k−1
an
≤ n−2 .
n−2
4. Applications
In this section we use the criteria in Theorem 1.2 to test the arithmetically CohenMacaulay property for two families of projective monomial curves.
12
4.1. Bresinsky semigroups. In [3] Bresinsky introduced the semigroup
Bh = h(2h − 1)2h, (2h − 1)(2h + 1), 2h(2h + 1), 2h(2h + 1) + 2h − 1i,
where h ≥ 2. He showed that the toric ideal IBh ⊂ S = K[x, y, z, t] is minimally
generated by more than 2h binomials. Based on that, in [13, Section 3.3] it is proved
that
F = {xt−yz}∪{z i−1 t2h−i −xi+1 y 2h−i : 1 ≤ i ≤ 2h}∪{x2h+1−j z j −y 2h−j tj : 0 ≤ j ≤ 2h−2}
is a minimal generating set for IBh . Combining the generators corresponding to
i = 2h and j = 1 we get that
u = −z(z 2h−1 − x2h+1 ) − x(x2h z − y 2h−1t) = xy 2h−1 t − z 2h ∈ IBh ,
hence in(u) = xy 2h−1 t ∈ in(IBh ), where one uses the reverse lexicographic monomial
order with x > y > z > t.
If the projective monomial curve associated to Bh were Cohen-Macaulay, then, as
the generators of Bh above are listed increasingly, by Theorem 1.2 we obtain that
xy 2h−1 ∈ in(IBh ). The ideal IBh is generated by binomials with disjoint support,
hence in the reduced Gröbner basis of IBh there exists v = xy d − z α tβ with in(v) =
xy d , 0 < d ≤ 2h − 1 and α, β nonnegative integers. We denote a1 , . . . , a4 the
generators of Bh in the given order. Since a1 < a2 < a3 < a4 we have (d + 1)a2 >
a1 + da2 = αa3 + βa4 > (α + β)a2 , hence α + β ≤ 2h − 1.
If α + β = 2h − 1, after adding to v the binomial z α tβ − y β xα+2 from the given
minimal generating set of IBh , we obtain that xy d − xα+2 y β ∈ IBh . Thus β ≤ d and
y d−β − xα+1 ∈ IBh , which is false, since d < 2h and one can see from F that 2h · a2 is
the smallest positive multiple of a2 which is in the semigroup generated by a1 , a3 , a4 .
Thus α + β < 2h − 1. If we denote I¯ = IBh mod x ⊂ K[y, z, t], then given F it
follows that I¯ = (yz) + (t, z)2h−1 + y 2(y, t)2h−1. It is easy to see that the monomial
¯ which is a contradiction.
v̄ = z α tβ is not in I,
Therefore, we proved the following proposition, which was first obtained by Cavaliere and Niesi in [5, Remark 5.4], as an application of their criterion from Corollary
2.7.
Proposition 4.1. (Cavaliere and Niesi, [5]) The projective monomial curve associated to Bh is not arithmetically Cohen-Macaulay, for any h ≥ 2.
4.2. Arslan semigroups. For h ≥ 2, let
Ah = hh(h + 1), h(h + 1) + 1, (h + 1)2 , (h + 1)2 + 1i.
This family of numerical semigroups was studied by Arslan who shows in [1, Proposition 3.2] that the defining ideal of K[Ah ] is
(5) IAh = (xh−i z i+1 − y h−i+1ti : 0 ≤ i < h)+
(xi+1 y h−i − z i th−i : 0 ≤ i ≤ h) + (xt − yz).
Proposition 4.2. The projective monomial curve associated to Ah is aritmetically
Cohen-Macaulay, for any h ≥ 2.
13
Proof. Letting gi = xh−i z i+1 − y h−i+1ti for 0 ≤ i ≤ h, fi = xi+1 y h−i − z i th−i for
0 ≤ i ≤ h and f = xt − yz, we claim that G = {g0 , . . . , gh , f0 , . . . , fh , f } is the
reduced Gröbner basis of IAh with respect to the reverse lexicographic term order
with x > y > z > t. As a consequence, by inspecting the leading monomials we
may use Theorem 1.2 to conclude the following statement.
We show that all the S-pairs of binomials in G reduce to 0 with respect to G,
and consequently by Buchberger’s criterion ([9, Theorem 2.14]) it follows that G is
a Gröbner basis for IAh .
G
S(g0 , f ) = zg0 + y h f = xh z 2 − xy h t = xg1 → 0.
For 1 ≤ i ≤ h:
G
S(gi , f ) = ygi − xh−i z i f = xh−i+1 z i t − y h−i+2ti = tgi−1 → 0.
For 0 ≤ i < h:
G
S(fi , f ) = zfi − xi+1 y h−i−1f = xi+2 y h−i−1t − z i+1 th−i = tfi+1 → 0.
G
Also, S(fh , f ) → 0 since gcd(in(fh ), in(f )) = gcd(xh+1 , yz) = 1.
For 1 ≤ i ≤ h:
G
S(gi , g0 ) → 0 since gcd(in(gi ), in(g0 )) = 1.
For 0 ≤ i ≤ h:
S(fi , g0 ) = y i+1fi + xi+1 g0 = xh+i+1 z − y i+1 z i th−i = xi (xh+1 − z h )z + z i (xi z h−i+1 −
G
y i+1 th−i ) = xi zfa + z i gh−i → 0.
For 0 ≤ j < i ≤ h:
S(gj , gi ) = z i−j gj − xi−j gi = y h−j+1z i−j tj − xi−j y h−i+1ti = y h−i+1tj (y i−j z i−j −
G
xi−j ti−j ) = y h−i+1tj · f · (. . . ) → 0.
For 1 ≤ i ≤ a, 0 ≤ j ≤ a with i ≤ h − j − 1, i.e. i + j < h:
S(fj , gi ) = xh−i−j−1z i+1 fj −y h−j gi = y 2h−i−j+1ti −xh−i−j−1 z i+j+1 th−j = y h−i−j ti (y h+1−
G
xh z) + xh−i−j−1 zti (xi+j+1 y h−i−j − z j+i th−j−i ) = −ti y h−i−j g0 + xh−i−j−1zti fi+j → 0.
For 1 ≤ i ≤ a, 0 ≤ j ≤ a with i > h − j − 1, i.e. i + j ≥ h:
S(fj , gi ) = z i+1 fj − xi+j+1−h y h−j gi = xi+j+1−h y 2h−i−j+1ti − z i+j+1 th−j =
G
yti (xi+j+1−h y 2h−i−j −z i+j−h t2h−i−j )+z i+j−h th−j (yth −z h+1 ) = yti fi+j−h +z i+j−h th−j gh →
0.
For 0 ≤ j < i ≤ a:
S(fj , fi ) = xi−j fj − y i−j fi = y i−j z i th−i − xi−j z j th−j = z j th−i (y i−j z i−j − xi−j ti−j ) =
G
z j th−i (yz − xt) · (. . . ) = z j th−i · f · (. . . ) → 0.
Acknowledgement. We gratefully acknowledge the use of the Singular ([8])
software for our computations. Dumitru Stamate was supported by the University
of Bucharest, Faculty of Mathematics and Computer Science, through the 2017
Mobility Fund.
References
[1] F. Arslan, Cohen-Macaulayness of tangent cones, Proc. Amer. Math. Soc. 128 (1999), 2243–
2251.
14
[2] F. Arslan, P. Mete, M. Şahin, Gluing and Hilbert functions of monomial curves, Proc. Amer.
Math. Soc. 137 (2009), 2225–2232.
[3] H. Bresinsky, On prime ideals with generic zero xi = tni , Proc. Amer. Math. Soc. 47 (1975),
329–332.
[4] H. Bresinsky, P. Schenzel, W. Vogel, On Liaison, Arithmetical Buchsbaum Curves and Monomial Curves in P3 , J. Algebra 86 (1984), 283–301.
[5] M.P. Cavaliere, G. Niesi, On monomial curves and Cohen-Macaulay type, Manuscripta math.
42 (1983), 147–159.
[6] M. Cimpoeaş, D.I. Stamate, Gröbner–nice pairs of ideals, in preparation.
[7] D. Cox, J. Little, D. O’Shea, Ideals, Varieties, and Algorithms, Third Edition, Springer, 2007.
[8] W. Decker, G.-M. Greuel, G. Pfister, H. Schönemann, Singular 3-1-6 — A computer algebra
system for polynomial computations. http://www.singular.uni-kl.de (2012).
[9] V. Ene, J. Herzog, Gröbner bases in commutative algebra, Graduate Studies in Mathematics
130, American Mathematical Society, 2012.
[10] S. Goto, N. Suzuki, K.-i Watanabe, On affine semigroup rings, Japan J. Math. 2 (1976), 1–12.
[11] J. Herzog, Generators and relations of Abelian semigroups and semigroup rings, Manuscripta
Math. 3 (1970), 175–193.
[12] J. Herzog, T. Hibi, Monomial Ideals, Graduate Texts in Mathematics 260, Springer–Verlag,
London, 2011.
[13] J. Herzog, D.I. Stamate, On the defining equations of the tangent cone of a numerical semigroup ring, J. Algebra 418 (2014), 8–28.
[14] Y. Kamoi, Defining ideals of Cohen-Macaulay semigroup rings, Comm. Alg. 20 (1992), 3163–
3189.
[15] Y. Kamoi, Defining ideals of Buchsbaum semigroup rings, Nagoya Math. J. 136 (1994), 115–
131.
[16] M. Morales, N. Thi Dung, Gröbner basis, a “pseudo-polynomial” algorithm for computing the
Frobenius number, Preprint 2015, arXiv:1510.01973 [math.AC].
[17] B.H. Roune, Solving thousand-digit Frobenius problems using Gröbner bases, J. Symb. Comput. 43 (2008), 1–7.
[18] T. Vu, Periodicity of Betti numbers of monomial curves, J. Algebra 418 (2014), 66–90.
Jürgen Herzog, Fachbereich Mathematik, Universität Duisburg-Essen, Campus
Essen, 45117 Essen, Germany
E-mail address: [email protected]
Dumitru I. Stamate, Faculty of Mathematics and Computer Science, University
of Bucharest, Str. Academiei 14, Bucharest – 010014, Romania
E-mail address: [email protected]
15
| 0math.AC
|
ON FEEBLY COMPACT SEMITOPOLOGICAL SYMMETRIC INVERSE
SEMIGROUPS OF A BOUNDED FINITE RANK
arXiv:1708.02064v2 [math.GR] 30 Jan 2018
OLEG GUTIK
Abstract. We study feebly compact shift-continuous T1 -topologies on the symmetric inverse semigroup
Iλn of finite transformations of the rank 6 n. For any positive integer n > 2 and any infinite cardinal
λ a Hausdorff countably pracompact non-compact shift-continuous topology on Iλn is constructed. We
show that for an arbitrary positive integer n and an arbitrary infinite cardinal λ for a T1 -topology τ
on Iλn the following conditions are equivalent: (i) τ is countably pracompact; (ii) τ is feebly compact;
(iii) τ is d-feebly compact; (iv) (Iλn , τ ) is H-closed; (v) (Iλn , τ ) is Nd -compact for the discrete countable
space Nd ; (vi) (Iλn , τ ) is R-compact; (vii) (Iλn , τ ) is infra H-closed. Also we prove that for an arbitrary
positive integer n and an arbitrary infinite cardinal λ every shift-continuous semiregular feebly compact
T1 -topology τ on Iλn is compact.
We follow the terminology of [6, 7, 8, 26, 27]. If X is a topological space and A ⊆ X, then by clX (A)
and intX (A) we denote the topological closure and interior of A in X, respectively. By |A| we denote
the cardinality of a set A, by A△B the symmetric difference of sets A and B, by N the set of positive
integers, and by ω the first infinite cardinal.
A semigroup S is called inverse if every a in S possesses an unique inverse a−1 , i.e. if there exists an
unique element a−1 in S such that
aa−1 a = a
a−1 aa−1 = a−1 .
and
A map which associates to any element of an inverse semigroup its inverse is called the inversion.
A topological (inverse) semigroup is a topological space together with a continuous semigroup operation (and an inversion, respectively). Obviously, the inversion defined on a topological inverse semigroup
is a homeomorphism. If S is a semigroup (an inverse semigroup) and τ is a topology on S such that
(S, τ ) is a topological (inverse) semigroup, then we shall call τ a semigroup (inverse) topology on S.
A semitopological semigroup is a topological space together with a separately continuous semigroup
operation. If S is a semigroup (an inverse semigroup) and τ is a topology on S such that (S, τ ) is a
semitopological semigroup (with continuous inversion), then we shall call τ a shift-continuous (inverse)
topology on S.
If S is a semigroup, then by E(S) we denote the subset of all idempotents of S. On the set of
idempotents E(S) there exists a natural partial order: e 6 f if and only if ef = f e = e. A semilattice
is a commutative semigroup of idempotents. A topological (semitopological) semilattice is a topological
space together with a continuous (separately continuous) semilattice operation. If S is a semilattice
and τ is a topology on S such that (S, τ ) is a topological semilattice, then we shall call τ a semilattice
topology on S.
Every inverse semigroup S admits a partial order:
a4b
if and only if there exists
e ∈ E(S) such that a = eb.
We shall say that 4 is the natural partial order on S.
Let λ be an arbitrary non-zero cardinal. A map α from a subset D of λ into λ is called a partial
transformation of λ. In this case the set D is called the domain of α and is denoted by dom α. The
image of an element x ∈ dom α under α is denoted by xα Also, the set {x ∈ λ : yα = x for some y ∈ Y }
is called the range of α and is denoted by ran α. The cardinality of ran α is called the rank of α and is
Date: January 31, 2018.
2010 Mathematics Subject Classification. Primary 22A15, 54D45, 54H10; Secondary 54A10, 54D30, 54D40.
Key words and phrases. Semigroup, inverse semigroup, semitopological semigroup, compact, countably compact, countably pracompact, feebly compact, H-closed, infra H-closed, X-compact, semiregular space.
1
2
OLEG GUTIK
denoted by rank α. For convenience we denote by ∅ the empty transformation, a partial mapping with
dom ∅ = ran ∅ = ∅.
Let Iλ denote the set of all partial one-to-one transformations of λ together with the following
semigroup operation:
x(αβ) = (xα)β
if x ∈ dom(αβ) = {y ∈ dom α : yα ∈ dom β},
for α, β ∈ Iλ .
The semigroup Iλ is called the symmetric inverse semigroup over the cardinal λ (see [7]). The symmetric
inverse semigroup was introduced by V. V. Wagner [29] and it plays a major role in the theory of
semigroups.
Put Iλn = {α ∈ Iλ : rank α 6 n}, for n = 1, 2, 3, . . .. Obviously, Iλn (n = 1, 2, 3, . . .) are inverse
semigroups, Iλn is an ideal of Iλ , for each n = 1, 2, 3, . . .. The semigroup Iλn is called the symmetric
inverse semigroup of finite transformations of the rank 6 n. By
x1 x2 · · · xn
y1 y2 · · · yn
we denote a partial one-to-one transformation which maps x1 onto y1 , x2 onto y2 , . . ., and xn onto yn .
Obviously, in such case we have xi 6= xj and yi 6= yj for i 6= j (i, j = 1, 2, 3, . . . , n). The empty partial
map ∅ : λ ⇀ λ is denoted by 0. It is obvious that 0 is zero of the semigroup Iλn .
Let λ be a non-zero cardinal. On the set Bλ = (λ ×λ) ∪{0}, where 0 ∈
/ λ ×λ, we define the semigroup
operation “ · ” as follows
(a, d), if b = c;
(a, b) · (c, d) =
0,
if b 6= c,
and (a, b) · 0 = 0 · (a, b) = 0 · 0 = 0 for a, b, c, d ∈ λ. The semigroup Bλ is called the semigroup of
λ × λ-matrix units (see [7]). Obviously, for any cardinal λ > 0, the semigroup of λ × λ-matrix units Bλ
is isomorphic to Iλ1 .
A subset A of a topological space X is called regular open if intX (clX (A)) = A.
We recall that a topological space X is said to be
• functionally Hausdorff if for every pair of distinct points x1 , x2 ∈ X there exists a continuous
function f : X → [0, 1] such that f (x1 ) = 0 and f (x2 ) = 1;
• semiregular if X has a base consisting of regular open subsets;
• quasiregular if for any non-empty open set U ⊂ X there exists a non-empty open set V ⊂ U
such that clX (V ) ⊆ U;
• compact if each open cover of X has a finite subcover;
• sequentially compact if each sequence {xi }i∈N of X has a convergent subsequence in X;
• countably compact if each open countable cover of X has a finite subcover;
• H-closed if X is a closed subspace of every Hausdorff topological space in which it is contained;
• infra H-closed provided that any continuous image of X into any first countable Hausdorff space
is closed (see [20]);
• countably compact at a subset A ⊆ X if every infinite subset B ⊆ A has an accumulation point
x in X;
• countably pracompact if there exists a dense subset A in X such that X is countably compact
at A;
• feebly compact if each locally finite open cover of X is finite;
• d-feebly compact (or DFCC ) if every discrete family of open subsets in X is finite (see [24]);
• pseudocompact if X is Tychonoff and each continuous real-valued function on X is bounded;
• Y -compact for some topological space Y , if f (X) is compact, for any continuous map f : X → Y .
According to Theorem 3.10.22 of [8], a Tychonoff topological space X is feebly compact if and only
if X is pseudocompact. Also, a Hausdorff topological space X is feebly compact if and only if every
locally finite family of non-empty open subsets of X is finite. Every compact space and every sequentially
compact space are countably compact, every countably compact space is countably pracompact, every
ON FEEBLY COMPACT SEMITOPOLOGICAL SYMMETRIC INVERSE SEMIGROUPS OF ...
3
countably pracompact space is feebly compact (see [3]), every H-closed space is feebly compact too (see
[15]). Also, every space feebly compact is infra H-closed by Proposition 2 and Theorem 3 of [20].
Topological properties of an infinite (semi)topological semigroup λ × λ-matrix units were studied in
[12, 13, 14]. In [13] it was shown that on the infinite semitopological semigroup λ × λ-matrix units Bλ
there exists a unique Hausdorff topology τc such that (Bλ , τc ) is a compact semitopological semigroup
and it was also shown that every pseudocompact Hausdorff shift-continuous topology τ on Bλ is compact.
Also, in [13] it was proved that every non-zero element of a Hausdorff semitopological semigroup λ × λmatrix units Bλ is an isolated point in the topological space Bλ . In [12] it was shown that the infinite
semigroup λ × λ-matrix units Bλ cannot be embedded into a compact Hausdorff topological semigroup,
every Hausdorff topological inverse semigroup S that contains Bλ as a subsemigroup, contains Bλ as
a closed subsemigroup, i.e., Bλ is algebraically complete in the class of Hausdorff topological inverse
semigroups. This result in [11] was extended onto so called inverse semigroups with tight ideal series and,
as a corollary, onto the semigroup Iλn . Also, in [16] it was proved that for every positive integer n the
semigroup Iλn is algebraically h-complete in the class of Hausdorff topological inverse semigroups, i.e.,
every homomorphic image of Iλn is algebraically complete in the class of Hausdorff topological inverse
semigroups. In the paper [17] this result was extended onto the class of Hausdorff semitopological inverse
semigroups and it was shown therein that for an infinite cardinal λ the semigroup Iλn admits a unique
Hausdorff topology τc such that (Iλn , τc ) is a compact semitopological semigroup. Also, it was proved
in [17] that every countably compact Hausdorff shift-continuous topology τ on Bλ is compact. In [14]
it was shown that a topological semigroup of finite partial bijections Iλn with a compact subsemigroup
of idempotents is absolutely H-closed (i.e., every homomorphic image of Iλn is algebraically complete
in the class of Hausdorff topological semigroups) and any countably compact topological semigroup
does not contain Iλn as a subsemigroup for infinite cardinal λ. In [14] there were given sufficient
conditions onto a topological semigroup Iλ1 to be non-H-closed. Also in [10] it was proved that an
infinite semitopological semigroup of λ × λ-matrix units Bλ is H-closed in the class of semitopological
semigroups if and only if the space Bλ is compact.
For an arbitrary positive integer n and an arbitrary non-zero cardinal λ we put
expn λ = {A ⊆ λ : |A| 6 n} .
It is obvious that for any positive integer n and any non-zero cardinal λ the set expn λ with the binary
operation ∩ is a semilattice. Later in this paper by expn λ we shall denote the semilattice (expn λ, ∩). It
is easy to see that expn λ is isomorphic to the subsemigroup of idempotents (the band) of the semigroup
Iλn for any positive integer n. We observe that for every positive integer n the band of the semigroup
Iλn is isomorphic to the semilattice expn λ by the mapping E(Iλn ) ∋ ε 7→ dom ε.
In the paper [18] feebly compact shift-continuous topologies τ on the semilattice expn λ were studied,
and all compact semilattice topologies on expn λ were described. In [18] it was whown that for an arbitrary positive integer n and an arbitrary infinite cardinal λ for a T1 -topology τ on expn λ the following
conditions are equivalent: (i) (expn λ, τ ) is a compact topological semilattice; (ii) (expn λ, τ ) is a countably compact topological semilattice; (iii) (expn λ, τ ) is a feebly compact topological semilattice; (iv)
(expn λ, τ ) is a compact semitopological semilattice; (v) (expn λ, τ ) is a countably compact semitopological semilattice. Also, in [18] there was constructed a countably pracompact H-closed quasiregular
non-semiregular topology τfc2 such that (exp2 λ, τfc2 ) is a semitopological semilattice with the discontinuous semilattice operation and it was proved that for an arbitrary positive integer n and an arbitrary
infinite cardinal λ a semiregular feebly compact semitopological semilattice expn λ is a compact topological semilattice. In [19] it was shown that for an arbitrary positive integer n and an arbitrary infinite
cardinal λ for a T1 -topology τ on expn λ the following conditions are equivalent: (i) τ is countably
pracompact; (ii) τ is feebly compact; (iii) τ is d-feebly compact; (iv) (expn λ, τ ) is an H-closed space.
This paper is a continuation of [11, 13, 16, 17]. We study feebly compact shift-continuous T1 -topologies
on the semigroup Iλn . For any positive integer n > 2 and any infinite cardinal λ a Hausdorff countably
pracompact non-compact shift-continuous topology on Iλn is constructed. We show that for an arbitrary
positive integer n and an arbitrary infinite cardinal λ for a T1 -topology τ on Iλn the following conditions
4
OLEG GUTIK
are equivalent: (i) τ is countably pracompact; (ii) τ is feebly compact; (iii) τ is d-feebly compact; (iv)
(Iλn , τ ) is H-closed; (v) (Iλn , τ ) is Nd -compact for the discrete countable space Nd ; (vi) (Iλn , τ ) is Rcompact; (vii) (Iλn , τ ) is infra H-closed. Also we prove that for an arbitrary positive integer n and an
arbitrary infinite cardinal λ every shift-continuous semiregular feebly compact T1 -topology τ on Iλn is
compact.
Later we shall assume that n is an arbitrary positive integer.
For every element α of the semigroup Iλn we put
↑l α = β ∈ Iλn : αα−1 β = α
and
↑r α = β ∈ Iλn : βα−1α = α .
Then Proposition 5 of [17] implies that ↑l α = ↑r α and by Lemma 6 of [23, Section 1.4] we have that
α 4 β if and only if β ∈ ↑l α for α, β ∈ Iλn . Hence we put ↑4 α = ↑l α = ↑r α for any α ∈ Iλn .
The definition of the semigroup operation of Iλn implies the following trivial lemma.
Lemma 1. Let n be an arbitrary positive integer and λ be any cardinal. Then for any elements α and
β of the semigroup Iλn the sets αIλn β and
↓4 α = {γ ∈ Iλn : γ 4 α}
are finite.
Proof. For any elements α and β of Iλn we have that
αIλn β = αIλn ∩ Iλn β = {γ ∈ Iλn : dom γ ⊆ dom α and ran γ ⊆ ran β} .
Since the sets dom α and ran β are finite, αIλn β is finite, as well.
For every γ ∈ ↓4 α the definition of the natural partial order 4 on the semigroup Iλn (see [23,
Chapter 1]) implies that the finite partial map γ is a restriction of the finite partial map α onto the set
A = dom α ∩ dom ε, where ε is an idempotent of Iλn such that γ = εα. This implies that the set ↓4 α
is finite.
Lemma 2. Let n be an arbitrary positive integer, λ be any infinite cardinal and τ be a shift-continuous
T1 -topology on semigroup Iλn . Then for every element α of the semigroup Iλn the set ↑4 α is open-andclosed in (Iλn , τ ), the space (Iλn , τ ) is functionally Hausdorff and hence it is quasi-regular.
Proof. Fix an arbitrary α ∈ Iλn . Then α ∈ αIλn α and
αIλn α = αIλn ∩ Iλn α = αα−1 Iλn ∩ Iλn α−1 α = αα−1 Iλn α−1 α,
because Iλn is an inverse semigroup. Since the topology τ is T1 , Lemma 1 implies that the set (αIλn α) \
{α} is closed in (Iλn , τ ). By the separate continuity of the semigroup operation in (Iλn , τ ) we have that
there exists an open neighbourhood U(α) of the point α in (Iλn , τ ) such that
αα−1 · U(α) · α−1 α ⊆ Iλn \ ((αIλn ∪ Iλn α) \ {α}) .
The last inclusion implies that U(α) ⊆ ↑α. Again, since the semigroup operation in (Iλn , τ ) is separately
continuous the set ↑4 α is open in (Iλn , τ ) as a full preimage of U(α) and the set ↑4 α is closed in (Iλn , τ )
as a full preimage of the singleton set {α}.
Fix arbitrary distinct elements α and β of the semigroup Iλn . Then either α and β are comparable or
not with respect to the natural partial order on Iλn . If α 4 β or α and β are incomparable in (Iλn , 4)
then it is obvious that the map g : Iλn → [0, 1] defined by the formula
1, if γ ∈ ↑4 β;
(γ)f =
0, if γ ∈
/ ↑4 β
is continuous. We observe that quasi-regularity of (Iλn , τ ) follows from the fact that every non-empty
open subset U of (Iλn , τ ) contains a maximal element δ with respect to the natural partial order 4 on
Iλn such that ↑4 α is an open-and-closed subset of (Iλn , τ ) and hence, since τ is a T1 -topology, {α} ⊆ U
is an open-and-closed subset of (Iλn , τ ).
A topological space X is called
ON FEEBLY COMPACT SEMITOPOLOGICAL SYMMETRIC INVERSE SEMIGROUPS OF ...
5
• totally disconnected if the connected components in X are singleton sets;
• scattered if X does not contain non-empty dense in itself subset, which is equivalent that every
non-empty subset of X has an isolated point in itself.
Lemma 2 implies the following corollary:
Corollary 3. Let n be an arbitrary positive integer, λ be any infinite cardinal and τ be a shift-continuous
T1 -topology on the semigroup Iλn . Then (Iλn , τ ) is a totally disconnected scattered space.
A partial order ≤ on a topological space X is called closed if the relation ≤ is a closed subset of
X × X in the product topology. In this case (X, ≤) is called a pospace [9].
Lemma 2 and Proposition VI-1.4 from [9] imply the following corollary:
Corollary 4. Let n be an arbitrary positive integer, λ be any infinite cardinal and τ be a shift-continuous
T1 -topology on semigroup Iλn . Then (Iλn , τ, 4) is a pospace
The following example shows that the statement of Lemma 2 does not hold in the case when (Iλn , τ )
is a T0 -space.
Example 5. For an arbitrary positive integer n and an arbitrary infinite cardinal λ we define a topology
τ0 on Iλn in the following way:
(i) all non-zero elements of the semigroup Iλn are isolated points in (Iλn , τ0 ); and
(ii) Iλn is the unique open neighbourhood of zero in (Iλn , τ0 ).
Simple verifications show that the semigroup operation and inversion on (Iλn , τ0 ) are continuous.
We need the following example from [17].
Example 6 ([17]). Fix an arbitrary positive integer n. The following family
Bc = Uα (α1 , . . . , αk ) = ↑4 α \ (↑4 α1 ∪ · · · ∪ ↑4 αk ) :
αi ∈ ↑4 α \ {α}, α, αi ∈ Iλn , i = 1, . . . , k
determines a base of the topology τc on Iλn . By Proposition 10 from [17], (Iλn , τc ) is a Hausdorff
compact semitopological semigroup with continuous inversion.
By Theorem 7 from [17], for an arbitrary infinite cardinal λ and any positive integer n every countably
compact Hausdorff semitopological semigroup Iλn is topologically isomorphic to (Iλn , τc ). By Corollary 3 the topological space (Iλn , τc ) is scattered. Since every countably compact scattered T3 -space is
sequentially compact (see [28, Theorem 5.7]), (Iλn , τc ) is a sequentially compact space.
Next we summarise the above results in the following theorem.
Theorem 7. Let n be an arbitrary positive integer, λ be any infinite cardinal and τ be a T1 -shift
continuous topology on the semigroup Iλn . Then the following conditions are equivalent:
(i) τ is compact;
(ii) τ = τc ;
(iii) τ is countably compact;
(iv) τ is sequentially compact.
Since every feebly compact Hausdorff topology on the semigroup Iλ1 is compact, it is natural to ask:
Does there exist a shift-continuous Hausdorff non-compact feebly compact topology τ on the semigroup
Iλn for n > 2?
The following example shows that for any infinite cardinal λ and any positive integer n > 2 there
exists a Hausdorff feebly compact topology τ on the semigroup Iλn such that (Iλn , τ ) is a non-compact
semitopological semigroup.
Example 8. Let λ be any infinite cardinal and τc2 = τc be the topology on the semigroup Iλ2 which
is defined in Example 6. We construct a stronger topology τfc2 on Iλ2 then τc2 in the following way. By
π : λ → Iλ2 : a 7→ εa we denote the map which assigns to any element a ∈ λ the identity partial map
6
OLEG GUTIK
εa : {a} → {a}. Fix an arbitrary infinite subset A of λ. For every non-zero element x ∈ Iλ2 we assume
that the base Bfc2 (x) of the topology τfc2 at the point x coincides with the base of the topology τc2 at x,
and
Bfc2 (0) = UB (0) = U(0) \ ((B)π ∪ {α1 , . . . , αs }) : U(0) ∈ Bc2 (0), α1, . . . , αs ∈ Iλ2 \ {0}
and B ⊆ λ such that |A△B| < ∞}
form a base of the topology τfc2 at zero 0 of the semigroup Iλ2 . Simple verifications show that the family
{Bfc2 (x) : x ∈ Iλ2 } satisfies conditions (BP1)–(BP4) of [8], and hence τfc2 is a Hausdorff topology on
Iλ2 .
Proposition 9. Let λ be an arbitrary infinite cardinal. Then (Iλ2 , τfc2 ) is a countably pracompact
semitopological semigroup with continuous inversion.
Proof. It is obvious that the inversion in (Iλ2 , τfc2 ) is continuous and later we shall show that all translations in (Iλ2 , τfc2 ) are continuous maps. We consider the following possible cases.
(1) 0 · 0 = 0. For every basic open neighbourhood UB (0) of zero in (Iλ2 , τfc2 ) we have that
UB (0) · 0 = 0 · UB (0) = {0} ⊂ Uπ (0).
(2) α · 0 = 0. For all basic open neighbourhoods UB (0) and Uα (β1 , . . . , βk ) of zero and an element
α 6= 0 in (Iλ2 , τfc2 ), respectively, we have that
Uα (β1 , . . . , βk ) · 0 = {0} ⊂ UB (0).
Iλ2
Let VB (0) =
\ (↑4 α1 ∪ · · · ∪ ↑4 αk ∪ (B)π) be an arbitrary basic neighbourhood of zero in (Iλ2 , τfc2 ).
Without loss of generality we may assume that
rank α1 = . . . = rank αk = 1 6 rank α.
Put
Cl = γ ∈ Iλ2 : rank γ = 1 such that αγ = αi for some i = 1, . . . , k
or αγ ∈ E(Iλ2 ) \ {0} .
The definition of the
semigroup
Iλ2 implies that the set Cl is finite. Then we have that α·WB (0) ⊆ VB (0)
S
↑4 γ : γ ∈ Cl .
for WB (0) = Iλ2 \
(3) 0 ·α = 0. For all basic open neighbourhoods UB (0) and Uα (β1 , . . . , βk ) of zero and an element
α 6= 0 in (Iλ2 , τfc2 ), respectively, we have that
0 ·Uα (β1 , . . . , βk ) = {0} ⊂ UB (0).
Iλ2
Let VB (0) =
\ (↑4 α1 ∪ · · · ∪ ↑4 αk ∪ (B)π) be an arbitrary basic neighbourhood of zero in (Iλ2 , τfc2 ).
Without loss of generality we may assume that
rank α1 = . . . = rank αk = 1 6 rank α.
Put
Cr = γ ∈ Iλ2 : rank γ = 1 such that γα = αi for some i = 1, . . . , k
or γα ∈ E(Iλ2 ) \ {0} .
The definition of the
Iλ2 implies that the set Cr is finite. Then we have that WB (0)·α ⊆ VB (0)
S semigroup
2
↑4 γ : γ ∈ Cr .
for WB (0) = Iλ \
(4) α · β = γ 6= 0 and rank α = rank β = rank γ, i.e., ran α = dom β. Then for any open neighbourhoods Uα (α1 , . . . , αk ), Uβ (β1 , . . . , βn ), Uγ (γ1 , . . . , γm ) of the points α, β and γ in (Iλ2 , τfc2 ), respectively,
we have that
Uα (α1 , . . . , αk ) · β = α · Uβ (β1 , . . . , βn ) = {γ} ⊆ Uγ (γ1 , . . . , γm).
(5) α · β = γ 6= 0 and rank α = rank γ = 1 and rank β = 2, i.e., ran α $ dom β. Then for any open
neighbourhoods Uβ (β1 , . . . , βn ) and Uγ (γ1 , . . . , γm ) of the points β and γ in (Iλ2 , τfc2 ), respectively, we
have that
α · Uβ (β1 , . . . , βn ) = {γ} ⊆ Uγ (γ1 , . . . , γm).
ON FEEBLY COMPACT SEMITOPOLOGICAL SYMMETRIC INVERSE SEMIGROUPS OF ...
7
Let Uγ (γ1 , . . . , γk ) be an arbitrary open neighbourhood of the point γ in (Iλ2 , τfc2 ) for some γ1 , . . . , γk ∈
↑4 γ, i.e., rank γ1 = . . . = rank γk = 2. Put
Q = δ ∈ ↑4 α : δβ ∈ {γ1 , . . . , γk } .
The definition of the semigroup Iλ2 implies that the set Q is finite. Then we have that
Uα (Q) · β ⊆ Uγ (γ1 , . . . , γk )
for Uα (Q) = ↑4 α \ δ ∈ ↑4 α : δ ∈ Q .
(6) α · β = γ 6= 0 and rank β = rank γ = 1 and rank α = 2, i.e., dom β $ ran α. In this case the proof
of separate continuity of the semigroup operation on (Iλ2 , τfc2 ) is dual to case (5).
(7) α · β = γ 6= 0, rank γ = 1 and rank α = rank β = 2. Then α and β are isolated points in (Iλ2 , τfc2 )
and hence
α · β = γ ⊆ Uγ (γ1 , . . . , γk ),
for any basic open neighbourhood Uγ (γ1 , . . . , γk ) of γ in (Iλ2 , τfc2 ).
(8) α · β = 0. Then dom β ∩ ran α = ∅ and hence
Uα (α1 , . . . , αk ) · β = α · Uβ (β1 , . . . , βn ) = {0} ⊂ UB (0),
for any basic open neighbourhoods Uα (α1 , . . . , αk ), Uβ (β1 , . . . , βn ) and UB (0) of α, β and zero 0 in
(Iλ2 , τfc2 ), respectivelly.
Thus we have shown that the translations in (Iλ2 , τfc2 ) are continuous maps.
Also, the definition of the topology τfc2 on Iλ2 implies that the set Iλ2 \ Iλ1 is dense in (Iλ2 , τfc2 ) and
every infinite subset of Iλ2 Iλ1 has an accumulation point in (Iλ2 , τfc2 ), and hence the space (Iλ2 , τfc2 ) is
countably pracompact.
Proposition 10. Let n be an arbitrary positive integer and λ be an arbitrary infinite cardinal. Then for
every d-feebly compact shift-continuous T1 -topology τ on Iλn the subset Iλn \ Iλn−1 is dense in (Iλn , τ ).
Proof. Since every quasi-regular d-feebly compact space is feebly compact (see [19, Theorem 2]), by
Lemma 2 the topology τ is feebly compact.
Suppose to the contrary that there exists a feebly compact shift-continuous T1 -topology τ on Iλn
in (Iλn , τ ). Then there exists a point α ∈ Iλn−1 of the space (Iλn , τ )
such that Iλn \ Iλn−1 is not dense
n−1
. This implies that there exists an open neighbourhood U(α) of α in
such that α ∈
/ clIλn Iλn \ Iλ
n
n
(Iλ , τ ) such that U(α) ∩ Iλ \ Iλn−1 = ∅. Lemma 2 implies that ↑4 α is an open-and-closed subset
of (Iλn , τ ) and hence by Theorem 14 of [4], ↑4 α is feebly compact. This implies that without loos of
generality we may assume that U(α) ⊆ ↑4 x ∩ Iλn−1 . By the definition of the semigroup Iλn we have
that there exists a point β ∈ U(α) such that ↑4 β ∩ U(α) = {β}. Again, by Lemma 2 we have that
↑4 β is an open-and-closed subset of (Iλn , τ ) and hence by Theorem 14 of [4], ↑4 β is feebly compact.
Moreover, our choice implies that β is an isolated point in the subspace ↑4 β of (Iλn , τ ).
Suppose that
x1 · · · xk
β=
,
y1 · · · yk
for some finite subsets {x1 , · · · , xk } and {y1 , · · · , yk } of distinct points from λ. Then the above arguments imply that k < n. Put p = n − k. Next we fix an arbitrary infinite sequence {ai }i∈N of distinct
elements of the set λ \ ({x1 , · · · , xk } ∪ {y1 , · · · , yk }).
For arbitrary positive integer j we put
x1 · · · xk ap(j−1)+1 · · · apj
.
βj =
y1 · · · yk ap(j−1)+1 · · · apj
Then βj ∈ Iλn for any positive integer j. Moreover, we have that βj ∈ Iλn \ Iλn−1 and βj ∈ ↑4 β for
any positive integer j.
8
OLEG GUTIK
We claim that the set ↑4 γ ∩ {βj : j ∈ N} is finite for any γ ∈ ↑4 β \ {β}. Indeed, if the set ↑4 γ ∩
{βj : j ∈ N} is infinite for some γ ∈ ↑4 β \ {β} then dom γ contains infinitely many points of the set
{ai : i ∈ N}, which contradicts that γ ∈ Iλn .
By Lemma 2 for every γ ∈ Iλn the set ↑4 γ is open in (Iλn , τ ). Then since β is an isolated point in
↑4 β, our claim implies that the infinite family of isolated points U = {{bj } : j ∈ N} is locally finite in
↑4 β, which contradicts that the subspace ↑4 β of (Iλn , τ ) is feebly compact. The obtained contradiction
implies the statement of the proposition.
Remark 11. The following three examples of topological semigroups of matrix units (Bλ , τmv ), (Bλ , τmh )
and (Bλ , τmi ) from [12] imply that the converse to Proposition 10 is not true for any infinite cardinal λ.
Later by Nd and R we denote the sets of positive integers with the discrete topology and the real
numbers with the usual topology.
Theorem 12. Let n be an arbitrary positive integer and λ be an arbitrary infinite cardinal. Then for
every shift-continuous T1 -topology τ on the semigroup Iλn the following statements are equivalent:
(i) τ is countably pracompact;
(ii) τ is feebly compact;
(iii) τ is d-feebly compact;
(iv) (Iλn , τ ) is H-closed;
(v) (Iλn , τ ) is Nd -compact;
(vi) (Iλn , τ ) is R-compact;
(vii) (Iλn , τ ) is infra H-closed.
Proof. Implications (i) ⇒ (ii) and (ii) ⇒ (iii) are trivial.
(iii) ⇒ (ii) Suppose that a space (Iλn , τ ) is d-feebly compact. By Lemma 2 it is quasi-regular.
Then by Theorem 1 of [19] every quasiregular d-feebly compact space is feebly compact and hence so
is (Iλn , τ ).
(ii) ⇒ (i) Suppose that a space (Iλn , τ ) is feebly compact. By Lemma 2 the topological space (Iλn , τ )
is Hausdorff. Then by Lemma 1 of [19] every Hausdorff feebly compact space with a dense discrete
subspace is countably pracompact (also see Lemma 4.5 of [5] or Proposition 1 from [2] for Tychonoff
spaces) and hence so is (Iλn , τ ).
Implication (iv) ⇒ (ii) follows from Proposition 4 of [15].
(ii) ⇒ (iv) We shall show by induction that if τ is a shift-continuous feebly compact T1 -topology on
the semigroup Iλn then the subspace ↑4 α of (Iλn , τ ) is H-closed for any α ∈ Iλn .
It is obvious that for any α ∈ Iλn with rank α = n the set ↑4 α = {α} is singleton, and since (Iλn , τ )
is a T1 -space, ↑4 α is H-closed.
Fix an arbitrary α ∈ Iλn with rank α = n − 1. By Lemma 2, ↑4 α is an open-and-closed subset of
(Iλn , τ ) and hence by Theorem 14 from [4] the space ↑4 α is feebly compact. Since by Lemma 2 every
point β of ↑4 α with rank α = n is isolated in (Iλn , τ ), the feeble compactness of ↑4 α implies that α is
a non-isolated point of (Iλn , τ ) and the space ↑4 α is compact. This implies that ↑4 α is H-closed.
Next we shall prove the following statement: if for some positive integer k < n for any α ∈ Iλn with
rank α 6 k the subspace ↑4 α is H-closed then ↑4 β is H-closed for any β ∈ Iλn with rank β = k − 1.
Suppose to the contrary that there exists a shift-continuous feebly compact T1 -topology τ on the
semigroup Iλn such that for some positive integer k < n for any α ∈ Iλn with rank α = k the subspace
↑4 α is H-closed and ↑4 β is not an H-closed space for some β ∈ Iλn with rank β = k − 1. Then there
exists a Hausdorff topological space X which contains the space ↑4 β as a dense proper subspace. We
observe that by Lemma 2 and Theorem 14 of [4] the space ↑4 β is feebly compact.
Fix an arbitrary x ∈ X \ ↑4 β. The Hausdorffness of X implies that there exist open neighbourhoods
UX (x) and UX (β) of the points x and β in X, respectively, such that UX (x) ∩ UX (β) = ∅. Then the
assumption of induction implies that without loss of generality we may assume that there do not exist
finitely many α1 , . . . , αm ∈ ↑4 β with rank α1 = . . . = rank αm = k such that
UX (x) ∩ ↑4 β ⊆ ↑4 α1 ∪ · · · ∪ ↑4 αm .
ON FEEBLY COMPACT SEMITOPOLOGICAL SYMMETRIC INVERSE SEMIGROUPS OF ...
9
Fix an arbitrary α1 ∈ ↑4 β such that rank α1 = k and ↑4 α1 ∩ UX (x) 6= ∅. Proposition 1.3.1 of [8],
Lemma 2 and Proposition 10 imply that there exists γ1 ∈ Iλn \ Iλn−1 such that γ1 ∈ ↑4 α1 ∩ UX (x).
Next, by induction using Proposition 1.3.1 of [8], Lemma 2 and Proposition 10 we construct sequences
{αi }i∈N and {γi }i∈N of distinct points of the set ↑4 β such that the following conditions hold:
(a) rank αi+1 = k and ↑4 αi+1 \ ↑4 α1 ∪ · · · ∪ ↑4 αi ∩ UX (x) 6= ∅; and
(b) γi+1 ∈ Iλn \ Iλn−1 and γi+1 ∈ ↑4 αi+1 \ ↑4 α1 ∪ · · · ∪ ↑4 αi ∩ UX (x),
for all positive integers i > 1.
Then Lemma 1 implies that the infinite family of non-empty open subsets U = {{γi } : i ∈ N} is
locally finite, which contradicts the feeble compactness of ↑4 β. The obtained contradiction implies the
statement of induction which completes the proof of the statement that the space (Iλn , τ ) is H-closed.
(iv) ⇒ (v) By Katětov’s Theorem every continuous image of an H-closed topological space into a
Hausdorff space is H-closed (see [8, 3.15.5 (b)] or [22]). Hence the image f (Iλn ) is H-closed for every
continuous map f : (Iλn , τ ) → Nd , which implies that f (Iλn ) is compact (see [8, 3.15.5 (a)]).
(v) ⇒ (ii) Suppose to the contrary that there exists a Hausdorff shift-continuous Nd -compact topology
τ on Iλn which is not feebly compact. Then there exists an infinite locally finite family U = {Ui } of
open non-empty subsets of (Iλn , τ ). Without loss of generality we may assume that the family U = {Ui }
is countable., i.e., U = {Ui : i ∈ N}. Then the definition of the semigroup Iλn and Lemma 2 imply
that for every Ui ∈ U there exists αi ∈ Ui such that ↑4 αi ∩ Ui = {αi } and hence U ∗ = {{αi } : i ∈ N}
is a family of isolated points of (Iλn , τ ). Since the family U is locally finite, without loss of generality
we may assume that αi 6= αj for distinct i, j ∈ N. We claim that the family U ∗ is locally finite.
Indeed, if we assume the contrary then there exists α ∈ Iλn such that every open neighbourhood of α
contains infinitely many elements of the family U ∗ . This implies that the family U is not locally
S finite,
∗
n
a contradiction. Since (Iλ , τ ) is a T1 -space and the family U is locally finite, we have that U ∗ is
a closed subset in (Iλn , τ ) and hence the map f : (Iλn , τ ) → Nd defined by the formula
S
1,
if β ∈ Iλn \ U ∗ ;
f (β) =
i + 1, if β = αi for some i ∈ N,
is continuous. This contradicts that the space (Iλn , τ ) is Nd -compact.
The proofs of implications (iv) ⇒ (vi) and (vi) ⇒ (ii) are same as the proofs of (iv) ⇒ (v) and
(v) ⇒ (ii), respectively.
Implication (ii) ⇒ (vii) follows from Proposition 2 and Theorem 3 of [20].
(vii) ⇒ (ii) Suppose to the contrary that there exists a Hausdorff shift-continuous infra H-closed
topology τ on Iλn which is not feebly compact. Then similarly as in the proof of implication (v) ⇒ (ii)
we choice a locally finite family U ∗ = {{αi } : i ∈ N} of isolated points of (Iλn , τ ). Then the map
f : (Iλn , τ ) → R defined by the formula
S
(
1,
if β ∈ Iλn \ U ∗ ;
1
f (β) =
, if β = αi for some i ∈ N,
i+1
is continuous. This contradicts that the space (Iλn , τ ) is infra H-closed.
Remark 13. By Theorem 5 from [20] conditions (ii) and (vii) of Theorem 12 are equivalent for any
Tychonoff space X.
It is not, however, the case that feebly compact and infra H-closed are equivalent in general. In
[21] Herrlich, beginning with a T1 -space Y , constructs a regular space X such that the only continuous
functions from X into Y are constant. His construction involves the cardinality of Y , but only as the
cardinality of collections of open sets whose intersections are singletons. Thus only the most trivial
modifications are needed in his argument to produce a regular Hausdorff infra H-closed space. It is also
easily shown that the space constructed in this manner is not feebly compact.
Any regular lightly compact space must be a Baire space [25, Lemma 3], and thus it is of interest
to note that the space constructed in [21] can be shown to be the countable union of nowhere dense
subsets using essentially the same argument as can be used to show it is not feebly compact.
10
OLEG GUTIK
Later we need the following technical lemma.
Lemma 14. Let n be an arbitrary positive integer and λ be an arbitrary infinite cardinal. Let τ be
a feebly compact shift-continuous T1 -topology on the semigroup Iλn . Then for every α ∈ Iλn and any
open neighbourhood U(α) of α in (Iλn , τ ) there exist finitely many α1 , . . . , αk ∈ ↑4 α \ {α} such that
Iλn \ Iλn−1 ∩ ↑4 α ⊆ U(α) ∪ ↑4 α1 ∪ · · · ∪ ↑4 αk .
Proof. Suppose to the contrary that there exists a feebly compact shift-continuous T1 -topology on the
semigroup Iλn which satisfies the following property: some element α of the semigroup Iλn has an open
neighbourhood U(α) of α in (Iλn , τ ) such that
Iλn \ Iλn−1 ∩ ↑4 α * U(α) ∪ ↑4 α1 ∪ · · · ∪ ↑4 αk ,
for any finitely many α1 , . . . , αk ∈ ↑4 α \ {α}. We observe that Lemma 2 implies that without loss of
generality we may assume that U(α) ⊆ ↑4 α.
Fix such an element α of Iλn and its open neighbourhood U(α) with the above determined property.
Then our assumption implies that there exists α1 ∈ ↑4 α \ U(α) such that the set
Iλn \ Iλn−1 ∩ ↑4 α1 \ U(α)
is infinite and fix an arbitrary γ1 ∈ Iλn \ Iλn−1 ∩ ↑4 α1 \ U(α). Next, by induction using our assumption
we construct sequences {αi }i∈N and {γi }i∈N of the distinct points of the set ↑4 α such that the following
conditions hold:
(a) αi+1 ∈ ↑4 α \ U(α) ∪ ↑4 α1 ∪ · · · ∪ ↑4 αi and the set
Iλn \ Iλn−1 ∩ ↑4 α ∩ ↑4 α1+1 \ U(α) ∪ ↑4 α1 ∪ · · · ∪ ↑4 αi
is infinite;
(b) γi+1 ∈ Iλn \ Iλn−1 ∩ ↑4 α ∩ ↑4 α1 \ U(α) ∪ ↑4 α1 ∪ · · · ∪ ↑4 αi ,
for all positive integers i.
By Lemma 2 and Theorem 14 of [4] the space ↑4 α is feebly compact. Then Lemma 1 implies that
the infinite family of non-empty open subsets U = {{γi } : i ∈ N} is locally finite, which contradicts
the feeble compactness of ↑4 α. The obtained contradiction implies the statement of the lemma.
Theorem 15. Let n be an arbitrary positive integer and λ be an arbitrary infinite cardinal. Then every
shift-continuous semiregular feebly compact T1 -topology τ on Iλn is compact.
Proof. We shall prove the statement of the theorem by induction. First we observe that for every element
α of a semiregular feebly compact T1 -semitopological semigroup (Iλn , τ ) with rank α = n − 1, n the set
↑4 α is compact. Indeed, by Lemma 2 for every β ∈ Iλn the set ↑4 β is open-and-closed in (Iλn , τ ), and
hence we have that Iλn \ Iλn−1 is an open discrete subspace of (Iλn , τ ) and using Theorem 14 of [4] we
obtain that ↑4 α is feebly compact, which implies that the space ↑4 α is compact.
Next we shall prove a more stronger step of induction: if for every element α of a semiregular feebly
compact T1 -semitopological semigroup (Iλn , τ ) with rank α > l 6 n the set ↑4 α is compact, then ↑4 β is
compact for every β ∈ Iλn with rank α = l.
Suppose to the contrary that there exists a semiregular feebly compact T1 -semitopological semigroup
(Iλn , τ ) such that for some positive integer l 6 n the set ↑4 α is compact for every α ∈ Iλn with
rank α > l, but there exists β ∈ Iλn with rank β = l such that the set ↑4 β is not compact.
First we observe that our assumption that the set ↑4 α is compact and Corollary 3.1.14 of [8] imply
that the following family
Bc (α) = Uα (α1 , . . . , αk ) = ↑4 α \ (↑4 α1 ∪ · · · ∪ ↑4 αk ) : αi ∈ ↑4 α \ {α}, i = 1, . . . , k
is a base of topology at the point α of (Iλn , τ ) for every α ∈ Iλn with rank α > l.
Then the Alexander Subbase Theorem (see [1, Theorem 1] or [8, p. 221, 3.12.2(a)]) and Lemma 2
imply that there exists a base B of the topology τ on Iλn with the following properties:
S
(i) B = {B(γ) : γ ∈ Iλn } and for every γ ∈ Iλn the family B(γ) is a base at the point γ;
ON FEEBLY COMPACT SEMITOPOLOGICAL SYMMETRIC INVERSE SEMIGROUPS OF ...
11
(ii) U(γ) ⊆ ↑4 γ for any U(γ) ∈ B(γ);
(iii) B(γ) = Bc (γ) for every γ ∈ Iλn with rank γ > l;
(iv) there exists a cover U of the set ↑4 β by members of the base B which has not a finite subcover.
We claim that the subspace ↑4 β of (Iλn , τ ) contains an infinite closed discrete subspace X. Indeed,
let U0 be a subfamily of U such that
[
U0 .
{β} ∪ ↑4 β ∩ Iλn \ Iλn−1 ⊆
Since the set ↑4 β is not compact and ↑4 γ is compact for any γ ∈ ↑4 β \ {β}, without loss of generality
we may assume that there exists k > rank β such that the following conditions hold:
S
(a) there
exist
infinitely
many
elements
ζ
∈
↑
β
with
rank
ζ
=
k
such
that
ζ
∈
/
U0 ;
4
S
(b) ς ∈ U0 for all ς ∈ ↑4 β with rank ς < k.
It is obvious that the set
[
[
X = ↑4 β \
U0 \
↑4 ς : rank ς > k
is requested.
Fix an arbitrary regular open neighbourhood U(β) of the point β in (Iλn , τ ) such that U(β) ∩ X = ∅.
By Lemma 14 there exist finitely many β1 , . . . , βs ∈ ↑4 β such that
Iλn \ Iλn−1 ∩ ↑4 β ⊆ U(β) ∪ ↑4 β1 ∪ · · · ∪ ↑4 βs .
It is obvious that the set X \ (↑4 β1 ∪ · · · ∪ ↑4 βs ) is infinite. For every
δ ∈ X the set ↑4 δ is compact and
open, and moreover by Lemma 1 the set ↑4 δ \ ↑4 β1 ∪ · · · ∪ ↑4 βs contains infinitely many points of the
neighbourhood U(β). This implies that intIλn clIλn (U(β)) ∩ X 6= ∅, which contradicts the assumption
that U(0) ∩ X = ∅. The obtained contradiction implies that the subspace ↑4 β of (Iλn , τ ) is compact,
which completes the proof of the theorem.
Acknowledgements. The author acknowledges Taras Banakh and the referee for useful important comments and suggestions.
References
[1] J. W. Alexander, Ordered sets, complexes and the problem of compactification, Proc. Natl. Acad. Sci. USA 25 (1939),
296–298.
[2] A. V. Arhangel’skii, On spaces with point-countable base, Topological spaces and their mappings, Riga, 1985, P. 3–7
(in Russian).
[3] A. V. Arkhangel’skii, Topological function spaces, Kluwer Publ., Dordrecht, 1992.
[4] R. W. Bagley, E. H. Connell, and J. D. McKnight, Jr., On properties characterizing pseudo-compact spaces, Proc.
Amer. Math. Soc. 9 (1958), no. 3, 500–506.
[5] S. Bardyla and O. Gutik, On a semitopological polycyclic monoid, Algebra Discrete Math. 21 (2016), no. 2, 163–183.
[6] J. H. Carruth, J. A. Hildebrant, and R. J. Koch, The theory of topological semigroups, Vols. I and II, Marcell Dekker,
Inc., New York and Basel, 1983 and 1986.
[7] A. H. Clifford and G. B. Preston, The algebraic theory of semigroups, Vols. I and II, Amer. Math. Soc. Surveys 7,
Providence, R.I., 1961 and 1967.
[8] R. Engelking, General topology, 2nd ed., Heldermann, Berlin, 1989.
[9] G. Gierz, K. H. Hofmann, K. Keimel, J. D. Lawson, M. W. Mislove, and D. S. Scott, Continuous lattices and domains,
Cambridge Univ. Press, Cambridge, 2003.
12
OLEG GUTIK
[10] O. Gutik, On closures in semitopological inverse semigroups with continuous inversion, Algebra Discrete Math. 18
(2014), no. 1, 59-85.
[11] O. Gutik, J. Lawson, and D. Repovš, Semigroup closures of finite rank symmetric inverse semigroups, Semigroup
Forum 78 (2009), no. 2, 326–336.
[12] O. V. Gutik and K. P. Pavlyk, On topological semigroups of matrix units, Semigroup Forum 71 (2005), no. 3, 389–400.
[13] O. V. Gutik and K. P. Pavlyk, Topological semigroups of matrix units, Algebra Discrete Math. (2005), no. 3, 1–17.
[14] O. Gutik, K. Pavlyk, and A. Reiter, Topological semigroups of matrix units and countably compact Brandt λ0 extensions, Mat. Stud. 32 (2009), no. 2, 115–131.
[15] O. V. Gutik and O. V. Ravsky, Pseudocompactness, products and topological Brandt λ0 –extensions of semitopological
monoids, Math. Methods and Phys.-Mech. Fields 58 (2015), no. 2, 20–37; reprinted version: J. Math. Sc. 223 (2017),
no. 1, 18–38.
[16] O. V. Gutik and A. R. Reiter, Symmetric inverse topological semigroups of finite rank 6 n, Math. Methods and
Phys.-Mech. Fields 52 (2009), no. 3, 7–14; reprinted version: J. Math. Sc. 171 (2010), no. 4, 425–432.
[17] O. Gutik and A. Reiter, On semitopological symmetric inverse semigroups of a bounded finite rank, Visnyk Lviv
Univ. Ser. Mech. Math. 72 (2010), 94–106 (in Ukrainian).
[18] O. Gutik and O. Sobol, On feebly compact topologies on the semilattice expn λ, Mat. Stud. 46 (2016), no. 1, 29–43.
[19] O. Gutik and O. Sobol, On feebly compact shift-continuous topologies on the semilattice expn λ, Visnyk Lviv Univ.
Ser. Mech. Math. 82 (2016), 128–136.
[20] D. W. Hajek and A. R. Todd, Compact spaces and infra H-closed spaces, Proc. Amer. Math. Soc. 48 (1975), no. 2,
479–482.
[21] H. Herrlich, Wann sind alle stetigen Abbildungen in Y konstant?, Math. Z. 90 (1965), 152–154.
[22] M. Katětov, Über H-abgeschlossene und bikompakte Räume, Čas. Mat. Fys. 69 (1940), no. 2, 36–49.
[23] M. V. Lawson, Inverse semigroups. The theory of partial symmetries, World Scientific, Singapore, 1998.
[24] M. Matveev, A survey of star covering properties, Topology Atlas preprint, April 15, 1998.
[25] R. A. McCoy, A filter characterization of regular Baire spaces, Proc. Amer. Math. Soc. 40 (1973), 268–270.
[26] M. Petrich, Inverse semigroups, John Wiley & Sons, New York, 1984.
[27] W. Ruppert, Compact semitopological semigroups: an intrinsic theory, Lect. Notes Math., 1079, Springer, Berlin,
1984.
[28] J. E. Vaughan, Countably compact and sequentially compact spaces, Handbook of set-theoretic topology, K. Kunen
and J. E. Vaughan (eds.), North-Holland, Amsterdam, 1984, pp. 569–602.
[29] V. V. Wagner, Generalized groups, Dokl. Akad. Nauk SSSR 84 (1952), 1119–1122 (in Russian).
Faculty of Mathematics, National University of Lviv, Universytetska 1, Lviv, 79000, Ukraine
E-mail address: o [email protected], [email protected]
| 4math.GR
|
Soft-Robust Actor-Critic Policy-Gradient
Esther Derman 1 , Daniel J. Mankowitz1 , Timothy A. Mann2 and Shie Mannor1
arXiv:1803.04848v1 [cs.LG] 11 Mar 2018
1
Technion Israel Institute of Technology, Haifa, Israel
2
Google Deepmind, London, UK
[email protected], [email protected], [email protected], [email protected]
Abstract
Robust Reinforcement Learning aims to derive
an optimal behavior that accounts for model uncertainty in dynamical systems. However, previous studies have shown that by considering
the worst case scenario, robust policies can be
overly conservative. Our soft-robust framework
is an attempt to overcome this issue. In this
paper, we present a novel Soft-Robust ActorCritic algorithm (SR-AC). It learns an optimal
policy with respect to a distribution over an
uncertainty set and stays robust to model uncertainty but avoids the conservativeness of robust
strategies. We show convergence of the SR-AC
and test the efficiency of our approach on different domains by comparing it against regular
learning methods and their robust formulations.
1
Introduction
Markov Decision Processes (MDPs) are commonly used
to model sequential decision making in stochastic environments. A strategy that maximizes the accumulated
expected reward is then considered as optimal and can be
learned from sampling. However, besides the uncertainty
that results from stochasticity of the environment, model
parameters are often estimated from noisy data or can
change during testing [Mannor et al., 2007; Roy et al.,
2017]. This second type of uncertainty can significantly
degrade the performance of the optimal strategy from the
model’s prediction.
Robust MDPs were proposed to address this problem
[Iyengar, 2005; Nilim and El Ghaoui, 2005; Tamar et al.,
2014]. In this framework, a transition model is assumed to
belong to a known uncertainty set and an optimal strategy
is learned under the worst parameter realizations. Although the robust approach is computationally efficient
when the uncertainty set is state-wise independent, compact and convex, it can lead to overly conservative results
[Mannor et al., 2012, 2016; Xu and Mannor, 2012; Yu
and Xu, 2016].
For example, consider a business scenario where an
agent’s goal is to make as much money as possible. It can
either create a startup which may make a fortune but may
also result in bankruptcy. Alternatively, it can choose to
live off school teaching and have almost no risk but low
reward. By choosing the teaching strategy, the agent may
be overly conservative and not account for opportunities
to invest in his own promising projects. Our claim is
that one could relax this conservativeness and construct
a softer behavior that interpolates between being aggressive and robust. Ideally, the soft-robust agent should stay
agnostic to outside financing uncertainty but still be able
to take advantage of the startup experience.
This type of dilemma can be found in various domains.
In the financial market, investors seek a good trade-off
between low risk and high returns regarding portfolio
management [Mitchell and Smetters, 2013]. In strategic
management, product firms must choose the amount of resources they put into innovation. A conservative strategy
would then consist of innovating only under necessary
conditions [Miller and Friesen, 1982].
In this paper, we focus on learning a soft-robust policy
(defined below) by incorporating soft-robustness into an
online actor-critic algorithm and show its convergence
properties. Existing works mitigate conservativeness of
robust MDP either by introducing coupled uncertainties
[Mannor et al., 2012, 2016] or by assuming prior information on the uncertainty set [Xu and Mannor, 2012; Yu and
Xu, 2016]. They use dynamic programming techniques to
estimate a robust policy. However, these methods present
some limiting restrictions such as non-scalability and offline estimation. Besides being computationally more efficient than batch learning [Wiering and van Otterlo, 2012],
the use of an online algorithm is of significant interest in
robust MDPs because it can detect non-adversarial stateactions pairs along a trajectory and result in less conservative results, something which cannot be performed when
solving the planning problem [Lim et al., 2016]. Other
works have attempted to incorporate robustness into an online algorithm for policy optimization [Mankowitz et al.,
2018; Tamar et al., 2015]. Although these approaches can
deal with large domains, a sampling procedure is required
for each critic estimate in Tamar et al. [2015], which differs from the strictly-speaking actor-critic. In Mankowitz
et al. [2018], the authors introduce a robust version of
actor-critic policy-gradient but its convergence results are
only shown for the actor updates. Moreover, these works
target the robust solution which may be too conservative.
We review all existing methods in Section 7 and compare
them to our approach.
To the best of our knowledge, our proposed work is the
first attempt to incorporate a soft form of robustness into
an online algorithm that has convergence guarantees besides being computationally scalable. We deal with the
curse of dimensionality by using function approximation
that parameterizes the expected value within a space of
much smaller dimension than the state space. By fixing
a distribution over the uncertainty set, the induced softrobust actor-critic learns a locally optimal policy in an
online manner. Under mild assumptions on the set of
distributions and uncertainty set, we show that our novel
Soft-Robust Actor-Critic (SR-AC) algorithm converges.
Our specific contributions are: (1) A soft-robust derivation of the objective function for policy-gradient; (2) An
SR-AC algorithm that uses stochastic approximation to
learn a variant of distributionally robust policy in an online manner; (3) Convergence proofs of SR-AC; (4) An
experiment of our framework to different domains that
shows the efficiency of soft-robustness. All proofs can be
found in the Appendix.
2
Background
In this section, we introduce the background material
related to our soft-robust approach.
Robust MDP A robust MDP is a tuple hX , A, r, Pi
where X is a finite state-space, A is a finite set of actions, r : X × A → R is the immediate reward function
which is deterministic and bounded and P is a set of
transition matrices. N
We assume that P is structured as
a cartesian product x∈X Px , which is known as the
rectangularity assumption [Nilim and El Ghaoui, 2005].
Given a state x ∈ X , the uncertainty set Px is a family
of transition models px ∈ Px we represent as vectors
in which the transition probabilities of each action are
arranged in the same block. For x, y ∈ X and a ∈ A,
denote by p(x, a, y) the probability of getting from state
x to state y given action a.
At timestep t, the agent is in state xt and chooses an
action at according to a stochastic policy π : X → M(A)
that maps each state to a probability distribution over the
action space, M(A) denoting the set of distributions over
A. It then gets a reward rt+1 and is brought to state xt+1
with probability p(xt , at , xt+1 ).
Policy-Gradient Policy-gradient methods are commonly
used to learn an agent policy. A policy π is parametrized
by θ and estimated by optimizing an objective function
using stochastic gradient descent. A typical objective to
be considered is the average reward function
T −1
1 X
rt+1 | π]
T →+∞
T t=0
X
X
=
dπp (x)
π(x, a)r(x, a)
Jp (π) = lim Ep [
x∈X
a∈A
where rt is the reward at time t, p an aperiodic and irreducible transition model under which the agent operates
and dπp is the stationary distribution of the Markov process
induced by p under policy π. The gradient objective has
previously been shown to be
X
X
∇θ Jp (π) =
dπp (x)
∇θ π(x, a)Qπp (x, a)
x∈X
a∈A
where Qπp (x, a) is the expected differential reward associated with state-action pair (x, a). This gradient is
then used to update the policy parameters according to:
θt+1 = θt + βt ∇θ Jp (π), with βt a positive step-size
[Sutton et al., 2000].
Actor-Critic Algorithm Theoretical analysis and empirical experiments have shown that regular policy-gradient
methods present a major issue namely high variance in the
gradient estimates that results in slow convergence and
inefficient sampling [Grondman et al., 2012]. First proposed by Barto et al. [1983], actor-critic methods attempt
to reduce the variance by using a critic that estimates the
value function. They borrow elements from both value
function and policy-based methods. The value function
estimate plays the role of a critic that helps evaluating the
performance of the policy. As in policy-based methods,
the actor then uses this signal to update policy parameters
in the direction of a gradient estimate of a performance
measure. Under appropriate conditions, the resulting algorithm is tractable and converges to a locally optimal policy
[Bhatnagar et al., 2009; Konda and Tsitsiklis, 2000].
Deep Q-networks Deep Q-Networks (DQNs) have
proven their capability of solving complex learning tasks
such as Atari video games [Mnih et al., 2013]. The Qlearning of Watkins and Dayan [1992] typically learns
a greedy or -greedy policy by updating the Q-function
based on a TD-error. In Deep Q-learning [Mnih et al.,
2013, 2015], a non-linear function such as a neural network is used as an approximator of the Q-function. It
is referred to as a Q-network. The agent is then trained
by optimizing the induced TD loss function thanks to
stochastic gradient descent. Like actor-critic, DQN is an
online algorithm that aims at finding an optimal policy.
The main difference with actor-critic is that it is off-policy:
it learns a greedy strategy while following an arbitrary
behavior [Mnih et al., 2013].
3
Soft-Robustness
Soft-Robust Framework Unlike robust MDPs that maximize the worst-case performance, we fix a prior on how
transition models are distributed over the uncertainty set.
A distribution over N
P is denoted by ω and is structured as
a cartesian product x∈X ωx . We find the same structure
in Xu and Mannor [2012]; Yu and Xu [2016]. Intuitively,
ω can be thought as the way the adversary distributes over
different transition models. The product structure then
means that this adversarial distribution only depends on
the current state of the agent without taking into account
its whole trajectory. This defines a probability distribution
ωx over Px independently for each state.
We further assume that ω is non-diffuse. This implies that
the uncertainty set is non-trivial with respect to ω in a
sense that the distribution does not affect zero mass to all
of the models.
Soft-Robust Objective Throughout this paper, we make
the following assumption:
Assumption 3.1. Under any policy π, the Markov chains
resulting from any of the MDPs with transition laws p ∈
P are irreducible and aperiodic.
Define dπp as the stationary distribution of the Markov
chain that results from following policy π under transition
model p ∈ P.
Definition 3.1. We call soft-robust objective or
¯
soft-robust average reward the function J(π)
:=
Ep∼ω [Jp (π)].
The distribution ω introduces a softer form of robustness
in the objective function because it averages over the
uncertainty set instead of considering the worst-case scenario. It also gives flexibility over the level of robustness
one would like to keep. A robust strategy would then
consist of putting more mass on pessimistic transition
models. Likewise, a distribution that puts all of its mass
on one target model would lead to an aggressive behavior
and result in model misspecification.
The soft-robust differential reward is given by
Q̄π (x, a) := Ep∼ω Qπp (x, a) where
Qπp (x, a)
p
:= E
X
+∞
rt+1 − Jp (π)|x0 = x, a0 = a, π .
t=0
Similarly, we introduce the quantity
V̄ π (x) :=
X
π(x, a)Q̄π (x, a) = Ep∼ω Vpπ (x)
a∈A
P
π
with Vpπ (x) :=
a∈A π(x, a)Qp (x, a). We will interchangeably term it as soft-robust expected differential
reward or soft-robust value function.
Soft-Robust Stationary Distribution The above perfor¯ cannot as yet be written as an expecmance objective J(π)
tation of the reward over a stationary distribution because
of the added measure ω on transition models. Define the
average transition model as p̄ := Ep∼ω [p]. It corresponds
to the transition probability that results from distributing
all transition models according to ω. In analogy to the
transition probability that minimizes the reward for each
given state and action in the robust transition function
[Mankowitz et al., 2018], our average model rather selects the expected distribution over all the uncertainty set
for each state and action. Under Assumption 3.1, we
can show that the transition p̄ as defined is irreducible
and aperiodic, which ensures the existence of a unique
stationary law we will denote by d¯π .
Proposition 3.1 (Stationary distribution in the average
transition model). Under Assumption 3.1, the average
transition matrix p̄ := Ep∼ω [p] is irreducible and aperiodic. In particular, it admits a unique stationary distribution.
As in regular MDPs, the soft-robust average reward satisfies a Poisson equation, as it was first stated in the discounted reward case in Lemma 3.1 of Xu and Mannor
[2012]. The following proposition reformulates this result
for the average reward.
Proposition 3.2 (Soft-Robust Poisson equation).
¯
J(π)
+ V̄ π (x)
!
=
X
a∈A
π(x, a) r(x, a) +
X
0
π
0
p̄(x, a, x )V̄ (x )
x0 ∈X
This Poisson equation enables us to establish an equivalence between the expectation of the stationary distributions over the uncertainty set and the stationary distribution of the average transition model, naming d¯π (x) =
Ep∼ω [dπp (x)] with x ∈ X . Indeed, we have the following:
Corollary 3.1. Recall d¯π the stationary distribution for
the average transition model p̄. Then
X
X
¯
J(π)
=
d¯π (x)
π(x, a)r(x, a).
x∈X
a∈A
The goal is to learn a policy that maximizes the soft-robust
¯ We use a policy-gradient method for
average reward J.
that purpose.
4
approximator of the soft-robust differential reward Q̄π . If
fw minimizes the mean squared error
2
X
X
π
π
π
¯
E (w) :=
d (x)
π(x, a) Q̄ (x, a)−fw (x, a)
x∈X
and is compatible in a sense that ∇w fw (x, a) =
∇θ log π(x, a), then
¯
∇θ J(π)
=
Soft-Robust Policy-Gradient
In policy-gradient methods, we consider a class of
parametrized stochastic policies πθ : X → M(A) with
θ ∈ Rd1 and estimate the gradient of the objective function J¯ with respect to policy parameters in order to update
¯
the policy in the direction of the estimated gradient of J.
The optimal set of parameters thus obtained is denoted by
¯ θ ).
θ∗ := arg max J(π
θ
When clear in the context, we will omit the subscript θ in
πθ for notation ease. We further make the following assumption, which is standard in policy-gradient litterature:
Assumption 4.1. For any (x, a) ∈ X × A, the mapping
θ 7→ πθ (x, a) is continuously differentiable with respect
to θ.
Using the same method as in Sutton et al. [2000], we
can derive the gradient of the soft-robust average reward
thanks to the previous results.
Theorem 4.1 (Soft-Robust Policy Gradient). For any
MDP satisfying previous assumptions, we have
X
X
¯
∇θ J(π)
=
d¯π (x)
∇θ π(x, a)Q̄π (x, a).
x∈X
a∈A
In order to manage with large state spaces, we also
introduce a linear approximation of Q̄π we define as
fw (x, a) := wT ψxa . Sutton et al. [2000] showed that
if the features ψxa satisfy a compatibility condition and
the approximation is locally optimal, then we can use it
in place of Q̄π and still point roughly in the direction of
the true gradient.
a∈A
X
d¯π (x)
x∈X
X
∇θ π(x, a)fw (x, a)
a∈A
We can further improve our gradient estimate by reducing
its variance. One direct method to do so is to subtract a
baseline b(x) from the previous gradient update. It is easy
to show that this will not affect the gradient derivation. In
particular, Bhatnagar et al. [2009] proved that the value
function minimizes the variance. It is therefore a proper
baseline to choose. We can thus write the following:
X
X
¯
∇θ J(π)
=
d¯π (x)
∇θ π(x, a) Q̄π (x, a) − V̄ π (x)
x∈X
=
X
x∈X
a∈A
¯π
d (x)
X
π(x, a)ψxa Āπ (x, a),
a∈A
(1)
where Āπ (x, a) is the soft-robust advantage function defined by Āπ (x, a) := Q̄π (x, a) − V̄ π (x).
5
Soft-Robust Actor-Critic Algorithm
In this section, we present our SR-AC algorithm which
is defined as Algorithm 1. This novel approach incorporates a variation of distributional robustness into an
online algorithm that effectively learns an optimal policy
in a scalable manner. Under mild assumptions, the resulting two-timescale stochastic approximation algorithm
converges to a locally optimal policy.
5.1
SR-AC algorithm
In the case of soft-robust average reward, this defines
a soft-robust gradient update that possesses the ability
to incorporate function approximation, as stated in the
following result. The main difference with that of Sutton
et al. [2000] is that we combine the dynamics of the
system with distributed transitions over the uncertainty
set.
An uncertainty set and a nominal model without uncertainty are provided as inputs. In practice, the nominal
model and the uncertainty set can respectively be an estimate of the transition model resulting from data sampling and its corresponding confidence interval. A distribution ω over the uncertainty set is also provided. It
corresponds to our prior information on the uncertainty
set. The step-size sequences (αt , βt , ξt ; t ≥ 0) consist of
small non-negative numbers properly chosen by the user
(see Appendix for more details).
Theorem 4.2 (Soft-Robust Policy Gradient with Function
Approximation). Let fw : X × A → R be a linear
At each iteration, samples are generated using the nominal model and the current policy. These are utilized to
update the soft-robust average reward (Line 5) and the
critic (Line 7) based on an estimate of a soft-robust TDerror we detail further. In our setting, the soft-robust value
function plays the role of the critic according to which
the actor parameters are updated. We then exploit the
critic to improve our policy by updating the policy parameters in the direction of a gradient estimate for the
soft-robust objective (Line 8). This process is repeated
until convergence.
5.2
Algorithm 1 SR-AC
1: Input: P - An uncertainty set; p̂ ∈ P - A nominal
2:
3:
4:
Convergence Analysis
5:
We establish convergence of SR-AC to a local maxima of
the soft-robust objective function by following an ODE
approach [Kushner and Yin, 1997].
Consider V̂ and Jˆ as unbiased estimates of V̄ and J¯ respectively. Calculating δt (Line 6 in Algorithm 1) requires
an estimate of the soft-robust average-reward that can be
obtained by averaging over samples given immediate reward r and distribution ω (Line 5). In order to get an
estimate of the soft-robust differential value V̂ , we use
linear function approximation. Considering ϕ as a d2 dimensional feature extractor over the state space X , we
may then approximate V̄ π (x) as v T ϕx , where v is a d2 dimensional parameter vector that we tune using linear
TD. This results in the following soft-robust TD-error:
X
δt := rt+1 − Jˆt+1 +
p̄(xt , at , x0 )vtT ϕx0 − vtT ϕxt ,
6:
7:
8:
9:
10:
model; ω - A distribution over P; fx - A feature
extractor for the SR value function;
Initialize: θ = θ0 - An arbitrary policy parameter;
v = v0 - An arbitrary set of value function parameters; α0 , β0 , ξ0 - Initial learning-rates; x0 - Initial
state
repeat
Act under at ∼ πθt (xt , at )
Observe next state xt+1 and reward rt+1
SR Average Reward Update:
Jˆt+1 = (1 − ξt )Jˆt + ξt rt+1
SR TD-Error:
P
δt = rt+1 − Jˆt+1 + x0 ∈X p̄(xt , at , x0 )V̂x0 − V̂xt
Critic Update: vt+1 = vt + αt δt ϕxt
Actor Update: θt+1 = θt + βt δt ψxt at
until convergence
Return: SR policy parameters θ and SR valuefunction parameters v
We used the existing structure of OpenAI Gym environments to run our experiments [Brockman et al., 2016].
6.1
6.1.1
Domains
A single-step MDP
x0 ∈X
where vt corresponds to the current estimate of the softrobust value function parameter.
As in regular MDPs, when doing linear TD learning, the
function approximation of the value function introduces
a bias in the gradient estimate [Bhatnagar et al., 2009].
d
¯
Denoting it as eπ , we have E[∇
θ J(π) | θ] = ∇θ J(π) +
π
e (see Appendix). This bias term then needs to be small
enough in order to ensure convergence.
Convergence of Algorithm 1 can be established by applying Theorem 2 from Bhatnagar et al. [2009] which exploits Borkar’s work on two-timescale algorithms [1997].
The convergence result is presented as Theorem 5.1.
Theorem 5.1. Under all the previous assumptions, given
> 0, there exists δ > 0 such that for a parameter vector
θt , t ≥ 0 obtained using the algorithm, if supπt keπt k < δ,
then the SR-AC algorithm converges almost surely to an
¯
-neighborhood of a local maximum of J.
6
Numerical Experiments
We demonstrate the performance of soft-robustness on
two domains of finite as well as continuous state spaces.
We consider a simplified formulation of the startup vs
teaching dilemma described in Section 1. The problem
is modeled as a 7-state MDP in which one action corresponds to one strategy. An illustration of this construction
is given graphically in Figure 1. At the starting state s0 ,
the agent chooses one of three actions. Action a1 [corresponds to the startup adventure] may lead it to a very
high reward in case of success but can be catastrophic
in case of failure. Action a2 [corresponds to the school
teaching carrier] leads it to low positive reward in case
of success with no possibility of negative reward. Action
a3 [corresponds to an intermediate strategy between entrepreneurship and teaching] can lead to an intermediate
positive reward with a slight risk of negative reward. Depending on the action it chose and if it succeeded or not,
the agent is brought to one of the six right-hand states and
receives the corresponding reward. It is brought back to
s0 at the end of each episode. We assume the probability
of success to be the same for all three actions.
6.1.2
Cart-Pole
In the Cart-Pole system, the agent’s goal consists of balancing a pole atop a cart in a vertical position. It is modeled as a continuous MDP in which each state consists
F1
a1 , R = -105
S1
a1 , R = 105
a2 ,R = 0
F2
a2 ,R = 2000
S2
s0
a3 ,R = −100
a3 ,R = 5000
F3
S3
Figure 1: Illustration of the MDP with initial state s0 .
States F1 , F2 , F3 correspond to failing scenarios for each
action. The succeeding states are represented by states
S1 , S2 , S3 .
6.3
Learning Algorithms
We trained the agent on the nominal model in each experiment. The soft-robust agent was learned using SR-AC in
the single-step MDP. In Cart-Pole, we run a soft-robust
version of a DQN algorithm.
Soft-Robust AC We analyze the performance of SR-AC
by training a soft-robust agent on the single-step MDP.
We run a regular AC algorithm to derive an aggressive
policy and learn a robust behavior by using a robust formulation of AC which consists in replacing the TD-error
with a robust TD-error, as implemented in Mankowitz
et al. [2018]. The derived soft-robust agent is then compared with the resulting aggressive and robust strategies
respectively.
Soft-Robust DQN Robustness has already been incorporated in DQN [Di-Castro Shashua and Mannor, 2017].
The Q-network addressed there performs an online estimation of the Q-function by minimizing at each timestep
t the following robust TD-error:
X
rob
δdqn,t
: = r(xt , at ) + γ inf
p(xt , at , x0 ) max
Q(x0 , a0 )
0
p∈P
x0 ∈X
a ∈A
− Q(xt , at ),
where γ is a discount factor.
of a 4-tuple hx, ẋ, θ, θ̇i which represents the cart position,
the cart speed, the pole angle with respect to the vertical
and its angular speed respectively. The agent can make
two possible actions: apply a constant force either to the
right or to the left of the pole. It gets a positive reward
of 1 if the pole has not fallen down and if it stayed in the
boundary sides of the screen. If it terminates, the agent
receives a reward of 0. Since each episode lasts for 200
timesteps, the maximal reward an agent can get is 200
over one episode.
In our experiments, we incorporate a soft-robust TD-error
inside a DQN that trains a soft-robust agent according to
the induced loss function. The soft-robust TD-error for
DQN is given by:
X
srob
δdqn,t
: = r(xt , at ) + γ
p̄(xt , at , x0 ) max
Q(x0 , a0 )
0
x0 ∈X
− Q(xt , at )
We use the Cart-Pole domain to compare the resulting
policy with the aggressive and robust strategies that were
obtained from a regular and a robust DQN respectively.
6.4
6.2
Uncertainty Sets
For each experiment, we generate an uncertainty set P
before training. In the single-step MDP, we sample from
5 different probabilities of success using a uniform distribution over [0, 1]. In Cart-Pole, we sample 5 different
lengths from a normal distribution centered at the nominal length of the pole which we fix at 0.3. Each length
thus generates a different transition function. We then
sample the average model by fixing ω as a realization of
a Dirichlet distribution. A soft-robust update is applied
by taking the optimal action according to this average
transition function.
a ∈A
Implementation
For each experiment, we train the agent on the nominal
model but incorporate soft-robustness during learning.
A soft-robust policy is learned thanks to SR-AC in the
single-step MDP. We use a linear function approximation
with 5 features to estimate the value function. For CartPole, we run a DQN using a neural network of 3 fullyconnected hidden layers with 128 weights per layer and
ReLu activations. We chose the ADAM optimizer to
minimize all the induced loss functions. We used constant
learning rates which worked well in practice. Each agent
was trained over Mtrain = 3000 episodes. For testing,
we averaged the performance over Mtest = 600 episodes
per parameter setting. Other hyper-parameter values can
be found in the Appendix.
6.5
Results
Shashua and Mannor [2017]; Mankowitz et al. [2018].
However, it is outperformed by the soft-robust agent
around the nominal model. Furthermore, the soft-robust
strategy shows an equilibrium between aggressiveness
and robustness thus leading to better performance than
the non-robust agent on larger pole lengths. We trained
a soft-robust agent on other weighting distributions and
remarked that depending on its structure, soft-robustness
interpolates between aggressive and robust behaviors (see
Appendix).
Single-step MDP
5
1e3
Average Reward
4
3
2
2.0 1e4
1
Aggressive
0
5
Soft-Robust
10
Training Epoch
Robust
1.5
15
Figure 2: Comparison of robust, soft-robust and aggressive agents during training. One training epoch corresponds to 300 episodes. The soft-robust strategy is not as
conservative as the robust one.
1.0
Average reward
0
0.5
0.0
0.5
1.0
Aggressive Policy
Soft-Robust Policy
Robust Policy
1.5
The evaluation of each strategy is represented in Figure 3.
As the probability of success gets low, the performance
of the aggressive agent drops down below the robust and
the soft-robust agents, although it performs best when
the probability of success gets close to 1. The robust
agent stays stable independently of the parameters but
underperforms soft-robust agent which presents the best
balance between high reward and low risk. We remarked
that depending on the weighting distribution initially set,
soft-robustness tends to being more or less aggressive. In
fact, high mass put on low probabilities of success leads to
less aggressive strategies (see Appendix). Incorporating a
distribution over the uncertainty set thus gives significant
flexibility on the level of aggressiveness to be assigned to
the soft-robust agent.
Cart-Pole In Figure 4, we show the performance of all
three strategies over different values of pole length. Similarly to our previous example, the non-robust agent performs well around the nominal model but its reward degrades on more extreme values of pole length. The robust agent keeps a stable reward under model uncertainty
which is consistent with the results obtained in Di-Castro
2.0
0.0
0.2
0.4
0.6
Probability of Success
0.8
1.0
Figure 3: Average reward for different probabilities of
success. Soft-robust policy is less sensitive to uncertainty
than aggressive policy and performs better than robustness
under optimistic scenarios.
250
200
Average reward
Figure 2 shows the evolution of the performance for all
three agents during training. It becomes more stable along
training time, which confirms convergence of SR-AC in
particular. We can see that the aggressive agent performs
best due to the highest reward it can reach on the nominal
model. The soft-robust agent gets rewards in between the
aggressive and the robust agent which performs the worst
due to its pessimistic learning method.
150
100
50
0
DQN
Robust DQN
Soft-Robust DQN
1
2
3
Pole Length
4
5
Figure 4: Average reward performance for DQN, robust
DQN and soft-robust DQN
7
Related work
This paper is related to several domains in RL such as
robust and distributionally robust MDPs, actor-critic methods and online learning via stochastic approximation algorithms. Our work solves the problem of conservativeness
encountered in robust MDPs by incorporating a variational form of distributional robustness. The SR-AC algorithm combines scalability to large scale state-spaces and
online estimation of the optimal policy in an actor-critic
algorithm. Table 1 compares our proposed algorithm with
previous approaches.
Table 1: Comparison of previous approaches with SR-AC
Reference
SR-AC (this paper)
Mankowitz et al.
[2018]
Lim et al. [2016]
Yu and Xu [2016]
Mannor et al. [2012,
2016]
Tamar et al. [2015]
Xu and Mannor
[2012]
Bhatnagar et al.
[2009]
Scalable
3
3
ActorCritic
3
7
SoftlyRobust
3
7
7
7
7
7
7
7
7
3
7
3
7
7
7
7
3
3
3
7
Many solutions have been addressed to mitigate conservativeness of robust MDP. Mannor et al. [2012, 2016] relax
the state-wise independence property of the uncertainty
set and assume it to be coupled in a way such that the planning problem stays tracktable. Another approach tends to
assume a priori information on the parameter set. These
methods include distributionally robust MDPs [Xu and
Mannor, 2012; Yu and Xu, 2016] in which the optimal
strategy maximizes the expected reward under the most
adversarial distribution over the uncertainty set. For finite
and known MDPs, under some structural assumptions on
the considered set of distributions, this max-min problem reduces to classical robust MDPs and can be solved
efficiently by dynamic programming [Puterman, 2009].
However, besides becoming untracktable under largesized MDPs, these methods use an offline learning approach which cannot adapt its level of protection against
model uncertainty and may lead to overly conservative
results [Lim et al., 2016]. The work of Lim et al. [2016]
solutions this issue and addresses an online algorithm that
learns the transitions that are purely stochastic and those
that are adversarial. Although it ensures less conservative results as well as low regret, this method sticks to
the robust objective while strongly relying on the finite
structure of the state-space. To alleviate the curse of dimensionality, we incorporate function approximation of
the objective value and define it as a linear functional of
features.
First introduced in Barto et al. [1983] and later addressed
by Bhatnagar et al. [2009], actor-critic algorithms are online learning methods that aim at finding an optimal policy.
We used the formulation of Bhatnagar et al. [2009] as a
baseline for the algorithm we proposed. The key difference between their work and ours is that we incorporate
soft-robustness. This relates in a sense to the Bayesian
Actor-Critic setup in which the critic returns a complete
posterior distribution of value functions using Bayes’ rule
[Ghavamzadeh and Engel, 2007; Ghavamzadeh et al.,
2015, 2016]. Our study keeps a frequentist approach,
meaning that our algorithm updates return point estimates
of the average value-function which prevents from tracktability issues besides enabling the distribution to be more
flexible. Another major distinction is that the Bayesian
approach incorporates a prior distribution on one model
parameters whereas our soft-robust method considers a
prior on different transition models over an uncertainty
set.
In Mankowitz et al. [2018]; Tamar et al. [2015], the authors incorporate robustness into policy-gradient methods.
A sampling procedure is required for each critic estimate
in Tamar et al. [2015], which differs from the strictlyspeaking actor-critic. A robust version of actor-critic
policy-gradient is introduced in Mankowitz et al. [2018]
but its convergence guarantees are only shown for robust
policy-gradient ascent. Both of these methods target the
robust strategy whereas we seek a soft-robust policy that
is less conservative while protecting itself against model
uncertainty.
8
Discussion
We have presented the SR-AC framework that is able to
learn policies which keep a balance between aggressive
and robust behaviors. SR-AC requires a stationary distribution under the average transition model and compatibility conditions for deriving a soft-robust policy-gradient.
We have shown that this ensures convergence of SR-AC.
This is the first work that has attempted to incorporate a
soft form of robustness into an online actor-critic method.
Our approach has been shown to be computationally scalable to large domains because of its low computational
price. In our experiments, we have also shown that the
soft-robust agent interpolates between aggressive and robust strategies without being overly conservative which
leads it to outperform robust policies under model uncertainty. Subsequent experiments should test the efficiency
of soft-robustness on more complex domains with continuous action spaces.
Garud N. Iyengar. Robust dynamic programming. Mathematics of Operations Research, 30(2):257–280, 2005.
The chosen weighting over the uncertainty set can be
thought as the way the adversary distributes over different
transition laws. In our current setting, this adversarial
distribution stays constant without accounting for the rewards obtained by the agent. Future work should address
the problem of learning the sequential game induced by
an evolving adversarial distribution to derive an optimal
soft-robust policy. Other extensions of our work may also
consider non-linear objective functions such as higher order moments with respect to the adversarial distribution.
Vijay R. Konda and John N. Tsitsiklis. Actor-critic algorithms. In Advances in Neural Information Processing
Systems, volume 12, 2000.
References
Andrew G. Barto, Richard S. Sutton, and Charles W. Anderson. Neuronlike adaptive elements that can solve
difficult learning control problems. IEEE Transactions
on Systems, Man, and Cybernetics, SMC-13(5):834 –
846, 1983.
Shalabh Bhatnagar, Richard Sutton, Mohammad
Ghavamzadeh, and Mark Lee. Natural Actor-Critic
Algorithms. Automatica, elsevier edition, 2009.
Vivek S. Borkar. Stochastic approximation with two
timescales. Systems and Control Letters, 29:291–294,
1997.
Greg Brockman, Vicki Cheung, Ludwig Pettersson, Jonas
Schneider, John Schulman, Jie Tang, and Wojciech
Zaremba. Open{AI} {G}ym. arXiv:1606.01540v1,
2016.
Shirli Di-Castro Shashua and Shie Mannor. Deep robust
{K}alman filter. arXiv preprint arXiv:1703.02310v1,
2017.
Mohammad Ghavamzadeh and Yaakov Engel. Bayesian
actor-critic algorithms. Proceedings of the 24th international conference on Machine learning, pages 297–304,
2007.
Mohammad Ghavamzadeh, Shie Mannor, Joelle Pineau,
and Aviv Tamar. Bayesian reinforcement learning: A
survey. Foundations and Trends in Machine Learning,
8(5-6):359–492, 2015.
Mohammad Ghavamzadeh, Yaakov Engel, and Michal
Valko. Bayesian policy gradient and actor-critic algorithms. Journal of Machine Learning Research, 17(66):
1–53, 2016.
Ivo Grondman, Lucian Busoniu, Gabriel A.D. Lopes, and
Robert Babuska. A survey of actor-critic reinforcement
learning: Standard and natural policy gradients. IEEE
Transactions on Systems, Man, and Cybernetics—Part
C: Applications and Reviews, 42(1291-1307), 2012.
Harold J. Kushner and Dean S. Clark. Stochastic Approximation Methods for Constrained and Unconstrained
Systems. Springer Verlag, 1978.
Harold J. Kushner and G. George Yin. Stochastic Approximation Algorithms and Applications. Springer Verlag,
New York, 1997.
Shiau Hong Lim, Huan Xu, and Shie Mannor. Reinforcement learning in robust markov decision processes. Mathematics of Operations Research, 41(4):
1325–1353, 2016.
Daniel J Mankowitz, Timothy A Mann, Shie Mannor,
Doina Precup, and Pierre-Luc Bacon. Learning robust
options. In AAAI, 2018.
Shie Mannor, Duncan Simester, Peng Sun, and John N.
Tsitsiklis. Bias and variance approximation in value
function estimates. Management Science, 53(2):308–
322, 2007.
Shie Mannor, Ofir Mebel, and Huan Xu. Lightning does
not strike twice: Robust {MDP}s with coupled uncertainty. In ICML, 2012.
Shie Mannor, Ofir Mebel, and Huan Xu. Robust {MDP}s
with k-rectangular uncertainty. Mathematics of Operations Research, 41(4):1484–1509, 2016.
Danny Miller and Peter H. Friesen. Innovation in conservative and entrepreneurial firms: Two models of
strategic momentum. Strategic Management Journal,
1982.
Olivia S. Mitchell and Kent Smetters. The Market for
Retirement Financial Advice. Oxford University Press,
first edition edition, 2013.
Volodymyr Mnih, Koray Kavukcuoglu, David Silver,
Alex Graves, Ioannis Antonoglou, Daan Wierstra, and
Martin Riedmiller. Playing {A}tari with deep reinforcement learning: Technical report. DeepMind Technologies, arXiv:1312.5602, 2013.
Volodymyr Mnih, Koray Kavukcuoglu, David Silver,
Andrei A. Rusu, Joel Veness, Marc G. Bellemare,
Alex Graves, Martin Riedmiller, Andreas K. Fidjeland,
Georg Ostrovski, Stig Petersen, Charles Beattie, Amir
Sadik, Ioannis Antonoglou, Helen King, Dharshan Kumaran, Daan Wierstra, Shane Legg, and Demis Hassabis. Human-level control through deep reinforcement
learning. Nature, 518:529–533, 2015.
Arnab Nilim and Laurent El Ghaoui. Robust control of
{M}arkov decision processes with uncertain transition
matrices. Operations Research, 53(5):783–798, 2005.
Martin L. Puterman. Markov decision processes: Discrete
stochastic dynamic programming, volume 414. Wiley
Series in Probability and Statistics, 2009.
Aurko Roy, Huan Xu, and Sebastian Pokutta. Reinforcement learning under model mismatch. 31st Conference
on Neural Information Processing Systems, 2017.
Richard S. Sutton, David McAllester, Satinger Singh, and
Yishay Mansour. Policy gradient methods for reinforcement learning with function approximation. In
Advances in Neural Information Processing Systems,
volume 12, pages 1057–1063, 2000.
Aviv Tamar, Shie Mannor, and Huan Xu. Scaling up
robust {MDP}s using function approximation. ICML,
32:1401–1415, 2014.
Aviv Tamar, Yinlam Chow, Mohammad Ghavamzadeh,
and Shie Mannor. Policy gradient for coherent risk measures. In Advances in Neural Information Processing
Systems, pages 1468–1476, 2015.
Christopher J. C. H. Watkins and Peter Dayan. Q-learning.
Machine Learning, 8:279–292, 1992.
Marco Wiering and Martijn van Otterlo. Reinforcement
Learning: State-of-the-Art. 12. Springer-Verlag Berlin
Heidelberg, 2012.
Huan Xu and Shie Mannor. Distributionally robust
{M}arkov decision processes. Mathematics of Operations Research, 37(2):288–300, 2012.
Pengqian Yu and Huan Xu. Distributionally robust counterpart in {M}arkov decision processes. IEEE Transactions on Automatic Control, 61(9):2538 – 2543, 2016.
Appendix
A
A.1
Proofs
Proposition 3.1
Proof. Fix x, y ∈ X . For any policy π, we denote by p(x, y) the probability of getting from state x to state y, which
can be written as Ea∼π(x) [p(x, a, y)]. Since ω is non-diffuse, there exists p0 such that ω(p0 ) > 0. Also, by Assumption
3.1, there exists an integer n such that pn0 (x, y) > 0. We thus have
n
p̄ (x, y) = Ep∼ω [p] (x, y)
n
n
p̄ (x, y) ≥ p0 ω(p0 ) (x, y)
n
p̄n (x, y) ≥ pn0 (x, y)ω(p0 )n > 0
which shows that p̄ is irreducible. Using the same reasoning, we show {n ∈ N : pn0 (x, x) > 0} ⊂ {n ∈ N : p̄n (x, x) >
0} and then use the fact that p0 is aperiodic to conclude that p̄ is aperiodic too.
A.2
Proposition 3.2
This recursive equation comes from the same reasoning as in Lemma 3.1 of Xu and Mannor [2012]. We apply it to the
average reward criterion.
Proof. For every p ∈ P, we can apply the Poisson equation to the corresponding model:
!
Jp (π) +
Vpπ (x)
=
X
X
π(x, a) r(x, a) +
p(x, a, x
0
)Vpπ (x0 )
x0 ∈X
a∈A
By integrating with respect to ω we obtain:
!
¯
J(π)
+ V̄ π (x) =
X
π(x, a) r(x, a) +
X
Ep∼ω [p(x, a, x
0
)Vpπ (x0 )]
x0 ∈X
a∈A
We then use the statewise independence assumption on ω to make the recursion explicit. We thus have
(1)
¯
J(π)
+ V̄ π (x) =
X
a∈A
(2)
=
X Z
0
π 0
0
0
π(x, a) r(x, a)+
p(x, a, x )Vp (x )dωx (px )dωx (px )
x0 ∈X
!
X
π(x, a) r(x, a) +
X
Epx ∼ωx [p(x, a, x
a∈A
x0 ∈X
X
X
0
)]Epx0 ∼ωx0 [Vpπ (x0 )]
!
=
a∈A
π(x, a) r(x, a) +
0
π
0
p̄(x, a, x )V̄ (x ) ,
x0 ∈X
where (1) results from the rectangularity assumption on ω. (2) Since p(x, a, x0 ) is an element of vector px that only
depends on the uncertainty set at state x and Vpπ (x0 ) depends on the uncertainty set at state x0 , we can split the integrals.
We slightly abuse notation here because a state can be visited multiple times. In fact, we implicitly introduce
dummy
N+∞
states and treat multiple visits to a state as visiting different states. More explicitely, we write ω as ω = t=0 ωx,t
where ωx,t = ωx , ωx being the distribution at state x. This representation is termed as the stationary model in Xu and
Mannor [2012].
A.3
Corollary 3.1
Proof. According to Proposition 3.2 and summing up both sides of the equality with respect to the stationary distribution
d¯π , we have
!
X
X
X
X
π
π
π
0
π
0
¯
J(π)
+
d¯ (x)V̄ (x) =
d¯ (x)
π(x, a) r(x, a) +
Ep ∼ω [p(x, a, x )]V̄ (x )
x
x∈X
x∈X
a∈A
x0 ∈X
X
X
x
!
X
=
d¯π (x)
x∈X
π(x, a) r(x, a) +
0
π
0
p̄(x, a, x )V̄ (x )
x0 ∈X
a∈A
Since d¯π is stationary with respect to p̄, we can then write
X
¯
J(π)
+
X
d¯π (x)V̄ π (x) =
x∈X
d¯π (x)
x∈X
X
π(x, a)r(x, a) +
X
d¯π (x0 )V̄ π (x0 ).
x0 ∈X
a∈A
It remains to simplify both sides of the equality in order to get the result.
A.4
Theorem 4.1
We use the same technique as in Mankowitz et al. [2018]; Sutton et al. [2000] in order to prove a soft-robust version of
policy-gradient theorem.
Proof.
X
∇θ V̄ π (x) = ∇θ
π(x, a)Q̄π (x, a)
a∈A
X
π
π
π
∇θ V̄ (x) =
∇θ π(x, a)Q̄ (x, a) + π(x, a)∇θ Q̄ (x, a)
a∈A
X
X
π
0
π 0
¯
∇θ V̄ (x) =
∇θ π(x, a)Q̄ (x, a) + π(x, a)∇θ r(x, a) − J(π) +
p̄(x, a, x )V̄ (x )
π
(1)
x0 ∈X
a∈A
X
X
π
π
0
π 0
¯
∇θ V̄ (x) =
∇θ π(x, a)Q̄ (x, a) + π(x, a) −∇θ J(π) +
p̄(x, a, x )∇θ V̄ (x )
x0 ∈X
a∈A
X
X
π
0
π 0
¯
∇θ J(π) =
∇θ π(x, a)Q̄ (x, a) + π(x, a)
p̄(x, a, x )∇θ V̄ (x ) − ∇θ V̄ π (x)
x0 ∈X
a∈A
X
(2)
¯
d¯π (x)∇θ J(π)
=
X
d¯π (x)
X
x∈X
x∈X
a∈A
X
X
X
¯
d¯π (x)∇θ J(π)
=
x∈X
d¯π (x)
x∈X
−
∇θ π(x, a)Q̄ (x, a) +
X
π(x, a)
X
π
∇θ π(x, a)Q̄ (x, a) +
d¯π (x)
x∈X
X
X
p̄(x, a, x )∇θ V̄ (x ) −
d¯π (x)∇θ V̄ π (x)
0
0
π
x0 ∈X
a∈A
a∈A
X
π
X
a∈A
π(x, a)
x∈X
X
0
p̄(x, a, x )∇θ V̄ (x0 )
π
x0 ∈X
d¯π (x)∇θ V̄ π (x)
x∈X
X
(3)
¯
d (x)∇θ J(π)
=
X
¯
∇θ J(π)
=
X
¯π
x∈X
x∈X
d¯π (x)
X
∇θ π(x, a)Q̄π (x, a) +
X
x0 ∈X
a∈A
d¯π (x0 )∇θ V̄ (x0 ) −
X
d¯π (x)∇θ V̄ π (x)
x∈X
π
∇θ π(x, a)Q̄ (x, a)
a∈A
P
¯π
where (1) occurs thanks to the soft-robust Poisson equation.
(2) Multiply
both sides of the Equation by
x∈X d (x). (3)
P
P
P
π
π
0
π 0
¯
¯
¯
Since d (x) is stationary with respect to p̄, we have that x∈X d (x) a∈A π(x, a)p̄(x, a, x ) = x0 ∈X d (x ).
A.5
Theorem 4.2
Proof. Recall the mean squared error:
X
π
E (w) :=
x∈X
d¯π (x)
X
2
π
π(x, a) Q̄ (x, a) − fw (x, a)
a∈A
with respect to the soft-robust state distribution d¯π (x). If we derive this distribution with respect to the parameters w
and analyze it when the process has converged to a local optimum as in Sutton et al. [2000], then we get:
X
X
d¯π (x)
π(x, a) Q̄π (x, a) − fw (x, a) ∇w fw (x, a) = 0
x∈X
a∈A
Additionally, the compatibility condition ∇w fw (x, a) = ∇θ log π(x, a) yields:
X
X
d¯π (x)
π(x, a) Q̄π (x, a) − fw (x, a) ∇θ π(x, a)
1
π(x, a)
a∈A
X
X
d¯π (x)
∇θ π(x, a) Q̄π (x, a) − fw (x, a)
x∈X
x∈X
=
0
=
0
a∈A
Subtract this quantity from the soft-robust policy gradient (Theorem 4.1). We then have:
X
X
X
X
π
π
π
π
¯
¯
¯
∇θ J(π) =
d (x)
∇θ π(x, a)Q̄ (x, a) −
d (x)
∇θ π(x, a) Q̄ (x, a) − fw (x, a)
x∈X
=
X
x∈X
A.6
a∈A
d¯π (x)
X
x∈X
a∈A
∇θ π(x, a)fw (x, a).
a∈A
Convergence Proof for SR-AC
We define as soft-robust TD-error at time t the following random quantity:
X
δt := rt+1 − Jˆt+1 +
p̄(xt , at , x0 )V̂x0 − V̂xt
x0 ∈X
¯
where V̂xt and Jˆt are unbiased estimates that satisfy E[V̂xt | xt , π] = V̄ π (xt ) and E[Jˆt+1 | xt , π] = J(π)
respectively.
π
We can easily show that this defines an unbiased estimate of the soft-robust advantage function Ā [Bhatnagar et al.,
¯
2009]. Thus, using equation (1), an unbiased estimate of the gradient ∇θ J(π)
can be obtained by taking
d
∇
θ J(π) := δt ψxt at .
Similarly, recall the soft-robust TD-error with linear function approximation at time t:
X
δt := rt+1 − Jˆt+1 +
p̄(xt , at , x0 )vtT ϕx0 − vtT ϕxt ,
x0 ∈X
where vt corresponds to the current estimate of the soft-robust value function parameter.
As in regular MDPs, when doing linear TD learning, the function approximation of the value function introduces a bias
in the gradient estimate Bhatnagar et al. [2009].
Define the quantity
Ve π (x) =
X
a∈A
X
¯
π(x, a) r(x, a) − J(π)
+
p̄(x, a, x0 )vπT ϕx0
x0 ∈X
where vπT ϕx0 is an estimate of the value function upon convergence of a TD recursion, that is vπ = limt→∞ vt . Also,
define as δtπ the associated error upon convergence:
X
δtπ := rt+1 − Jˆt+1 +
p̄(xt , at , x0 )vπT ϕx0 − vπT ϕxt .
x0 ∈X
Similarly to Lemma 4 of Bhatnagar et al. [2009], the bias of the soft-robust gradient estimate is given by
eπ :=
X
d¯π (x) ∇θ Ve π (x) − ∇θ vπT ϕx ,
x∈X
π
d
¯
that is E[∇
θ J(π) | θ] = ∇θ J(π) + e . This error term then needs to be small enough in order to ensure convergence
of the algorithm.
Let denote as V̄ (v) := Φv the linear approximation to the soft-robust differential value function defined earlier, where
Φ ∈ Rn×d2 is a matrix and each feature vector ϕx (k) corresponds to the k th column in Φ. We make the following
assumption:
Assumption A.1. The basis functions ϕx ∈ Rd2 are linearly independent. In particular, Φ has full rank. We also have
Φv 6= e for all value function parameters v ∈ Rd2 where e is a vector of all ones.
The learning rates αt and
1)
Pβt (Lines
P7 and 8 in Algorithm
P
P are established such that αt → 0 slower than βt → 0
as t → ∞. In addition, t αt = t βt = ∞ and t αt2 , t βt2 < ∞. We also set the soft-robust average reward
step-size ξt = cαt for a positive constant c. The soft-robust average reward, TD-error and critic will all operate on
the faster timescale αt and therefore converge faster. Eventually, define a diagonal matrix D where the steady-state
distribution d¯π forms the diagonal of this matrix. We write the soft-robust transition probability matrix as:
X
P̄ π (x, x0 ) =
π(x, a)p̄(x, a, x0 ),
a∈A
π
n
where x, P
x0 ∈ X and p̄ designates the
P average transition model. By denoting R ∈ R the column vector of
rewards ( a∈A π(x1 , a)r(x1 , a), · · · a∈A π(xn , a)r(xn , a))T where (x1 , . . . , xn ) is a numbered representation of
the state-space and using the following operator T : Rn → Rn , we can express the soft-robust Poisson equation as:
¯
T (J) = Rπ − J(π)e
+ P̄ π J
The soft-robust average reward iterates (Line 5) and the critic iterates (Line 7) defined in Algorithm 1 converge almost
surely, as stated in the following Lemma which is a straightforward application of Lemma 5 from Bhatnagar et al.
[2009].
Lemma A.1. For any given policy π and {Jˆt }, {vt } as in the soft-robust average reward and critic updates, we have
¯
Jˆt → J(π)
and vt → v π almost surely, where
X
X
¯
J(π)
=
d¯π (x)
π(x, a)r(x, a)
x∈X
a∈A
π
is the average reward under policy π and v is the unique solution to
ΦT DΦv π = ΦT DT (Φv π )
Thanks to all the previous results, convergence of Algorithm 1 can be established by applying Theorem 2 from Bhatnagar
et al. [2009] which exploits Borkar’s work on two-timescale algorithms [1997]. For simplicity, we assume that the
iterates resulting from the actor update (Line 10 of Algorithm 1) in SR-AC remain bounded, although one could prove
convergence without such an assumption by incorporating an operator that projects any policy parameter to a compact
set, as described in Kushner and Clark [1978]. The resulting actor update would then be the projected value of the
predefined iterate. The convergence result is presented as Theorem A.1.
Theorem A.1. Under all the previous assumptions, given > 0, there exists δ > 0 such that for a parameter vector
θt , t ≥ 0 obtained using the algorithm, if supπt keπt k < δ, then the SR-AC algorithm converges almost surely to an
¯
-neighborhood of a local maximum of J.
B
B.1
Experiments
One-step MDP
Model Parameters
Nominal probability of success
Uncertainty set for probabilities of success
Weighting Distribution 1
Weighting Distribution 2
Aggressive rewards
Soft robust rewards
Robust rewards
Value
0.8
[0.1, 0.7, 0.8, 0.3, 0.5]
[0.47, 0.22, 0.1, 0.09, 0.12]
[0.63, 0.04, 0.05, 0.02, 0.26]
105 ; −105
5000; −100
2000; 0
Hyperparameters
Critic Learning rate α
Actor Learning rate β
Step size ξ
Number of linear features
Number of episodes for training Mtrain
Number of episodes for testing Mtest
Value
5e-3
5e-5
3α
5
3000
600
2.0 1e4
1.5
Average reward
1.0
0.5
0.0
0.5
1.0
Aggressive Policy
Soft-Robust Policy 2
Robust Policy
1.5
2.0
0.0
0.2
0.4
0.6
Probability of Success
0.8
1.0
Figure 5: Average reward for different probabilities of success (distribution 2). Soft-robust policy interpolates between
aggressive and robust strategies.
B.2
Cart-Pole example
Hyperparameters
Discount factor γ
Learning rate
Mini-batch size
Final epsilon
Target update interval
Max number of episodes for training Mtrain
Number of episodes for testing Mtest
Value
0.9
1e-4
256
1e-5
10
3000
600
We trained a soft-robust agent on a different weighting over the uncertainty set. Figure 6 shows the performance of the
resulting strategy that presents a similar performance as the robust agent. This stronger form of robustness demonstrates
the flexibility we have on the way we fix the weights, which leads to more or less aggressive behaviors.
250
Average reward
200
150
100
50
0
DQN
Robust DQN
Soft-Robust DQN 2
1
2
3
Pole Length
4
5
Figure 6: Average reward performance for DQN, robust DQN and soft-robust DQN (distribution 2). Soft-robust policy
interpolates between aggressive and robust strategies.
| 2cs.AI
|
Indexing Variation Graphs∗
Jouni Sirén†
arXiv:1604.06605v4 [cs.DS] 12 Jan 2017
Abstract
Variation graphs, which represent genetic variation within
a population, are replacing sequences as reference genomes.
Path indexes are one of the most important tools for working
with variation graphs. They generalize text indexes to graphs,
allowing one to find the paths matching the query string. We
propose using de Bruijn graphs as path indexes, compressing
them by merging redundant subgraphs, and encoding them
with the Burrows-Wheeler transform. The resulting fast,
space-efficient, and versatile index is used in the variation
graph toolkit vg.
1
Introduction
Sequence analysis pipelines typically start with mapping
the reads from the sequenced genome to a reference
genome of the same species. As reference genomes are
usually assembled from the genomes of a small number
of individuals, they are biased towards those individuals.
This reference bias may affect the results of subsequent
analysis, especially when the sequenced individuals are
from different populations than the ones behind the
reference genome.
Variation graphs (also graph genomes, genome
graphs, graph references, or reference graphs), which
encode the genetic variation within a population as a
graph, have been proposed as a solution to the reference
bias [37, 28, 6, 8, 29]. These graphs are expected to
replace sequences as reference genomes. The shift to
graphs will likely take years, as new methods and tools
are needed to replace those based on linear references.
The variation graph toolkit vg [14] is a community
effort to develop such tools. This paper describes
GCSA2, the path index used in vg. A path index is a
generalization of text indexes for labeled graphs. Given
a query string, the index finds the paths with a label
matching the query. Indexing graphs is inherently hard,
as the number of paths increases exponentially with
path length. The design of a path index is hence a tradeoff between maximum query length, index size, query
performance, and pruning complex regions of the graph.
Mapping reads to a graph was first investigated
by Schneeberger et al. [42]. Sirén et al. developed
∗ Supported
† Wellcome
by the Wellcome Trust grant 098051.
Trust Sanger Institute, UK. [email protected].
GCSA [43], which generalized the FM-index [10] (a text
index based on the Burrows-Wheeler transform [4]) to
directed acyclic graphs. GCSA depends on pruning the
complex regions, as there are no limits on query length.
Kim et al. [19] combined GCSA with the HISAT read
aligner [20]. The result was the first practical graphbased read aligner, though it uses the graph for more
accurate mapping to a linear reference.
De Bruijn graphs can be used as k-mer indexes of
other graphs. There are already read aligners based on
them [23, 25]. The succinct de Bruijn graph of Bowe
et. al. [3] encodes the graph with a generalization of
the FM-index. Rødland [39] proposed another similar
generalization. Succinct de Bruijn graphs can simulate
order-k de Bruijn graphs for multiple values of k [2],
but they still need to store the graph explicitly for
the largest value of k. Other representations include
compacted [5] and compressed [28] de Bruijn graphs,
which represent unary paths in the graph as single
nodes or edges. Probabilistic de Bruijn graphs [38] use
Bloom filters to support faster queries, at the expense
of producing false positives. All these space-efficient
representations require several bits per k-mer for the
graph, and more for mapping back to the indexed graph.
Some path indexes store the graph as a collection of
sequences. BWBBLE [17] uses the powerset alphabet for
encoding substitutions and creates new sequences with
a sufficient amount of context for other variants. vBWT
[26] encodes variant sites explicitly in the sequence as
X(A|B|C)Y , using distinct separator symbols for each
site. Queries in both BWBBLE and vBWT are slower
than in ordinary FM-indexes, as variant sites force the
search to branch. The hypertext index [44] works with
graphs that have string labels on the nodes. The strings
are indexed using an FM-index. Partial matches in the
strings are combined into full matches with range queries
in the edge matrix. While matches crossing one edge
are easy to find, the approach becomes impractical with
matches crossing multiple edges.
There are also structures (e.g. [18, 46, 7, 32, 33])
using graphs as a space-efficient way of indexing similar
sequences. While the problem is different from indexing
the paths in a graph, the techniques used are similar.
The above methods can be classified in three
categories based on the data models they use. Many deal
with graphs arising from aligned sequences, assuming
a shared global sequence with local variation. GCSAbased methods can index directed acyclic graphs, as well
as cyclic graphs that are sufficiently similar to de Bruijn
graphs. GCSA2, the hypertext index, and indexes based
on de Bruijn graphs work with arbitrary graphs.
GCSA2 combines ideas from the original GCSA
and from succinct de Bruijn graphs. Conceptually it
uses a de Bruijn graph as a k-mer index of a variation
graph. The de Bruijn graph is pruned (compressed
structurally) by using strings shorter than k characters
as nodes, if the shorter strings identify the start nodes of
the corresponding paths uniquely. The pruned graph is
encoded with a generalization of the FM-index. GCSA2
often uses less space (e.g. less than 1 bit per k-mer)
than other de Bruijn graph-based indexes, which have
to store some information for each k-mer explicitly. The
index also includes extensions based on suffix trees.
The extensions are used for e.g. finding maximal exact
matches in the vg read aligner.
The main differences to the original GCSA are:
1. The graph encoding in GCSA2 has been optimized
for small alphabets, improving query performance
by up to an order of magnitude.
2. The construction algorithm stores the graphs on
disk, reducing the memory requirements of building
a whole-genome human index from hundreds of
gigabytes to tens of gigabytes.
and refer to substrings of length k as k-mers. Substring
S[i, j] is a proper substring of string S, if S 6= S[i, j].
We say that string S 0 is a substring of string collection
S, if it is a substring of a string S ∈ S.
Sometimes we consider infinite character sequences
S = (si )i∈Z , where set Z is a contiguous infinite subset
of Z. The notion of substring generalizes to infinite
sequences in a natural way. A substring of an infinite
sequence S is left-infinite if it extends infinitely to the left,
and right-infinite if it extends infinitely to the right. A
substring of a finite or infinite sequence S is left-maximal
if it is left-infinite or a prefix, and right-maximal if it is
right-infinite or a suffix.
We are primarily interested in sequences over the
DNA alphabet {$, A, C, G, T, N}. Characters A, C, G, and T
are called bases, while character N represents an arbitrary
or unknown base. The alphabet may contain other
technical characters in addition to the endmarker $.
Each character c of the DNA alphabet has a complement
c defined as A = T, C = G, G = C, T = A, and
c = c for other characters c. Given a DNA sequence
←
−
S, its reverse complement is the sequence S obtained
by reversing the non-technical parts of the sequence
and replacing each character with its complement. For
←−−−−−−
example, GATTACA$ = TGTAATC$.
Given a string S[0, n − 1], we define S.rank(i, c) as
the number of occurrences of character c in the prefix
S[0, i − 1]. We also define
S.select(i, c) = max{j ≤ n | S.rank(j, c) < i}
3. GCSA2 can index denser graphs, including cyclic
as the position of the occurrence of character c with
graphs, by limiting maximum query length.
rank i > 0.1 A bitvector is a binary sequence supporting
4. GCSA2 extends the functionality of the FM-index efficient rank/select queries. Wavelet trees [16] are spacewith suffix tree operations.
efficient data structures that use bitvectors to support
rank/select queries on arbitrary strings.
The first two improvements can also be used with the
Let S be a string and S 0 be a string or an infinite
original GCSA. The third point represents a different character sequence over alphabet Σ. We say that
approach to indexing graphs. It was not possible the sequences S and S 0 prefix-match, if S is a prefix of S 0 or
with the original GCSA, as the construction algorithm S 0 is a prefix of S. Set S of strings is prefix-free, if no
required a prefix-range-sorted graph. The extended two strings S, S 0 ∈ S (with S 6= S 0 ) prefix-match.
functionality depends on the new construction algorithm
and on limiting maximum query length.
2.2 Text Indexes The suffix tree [47] is the most
This paper describes the GCSA2 data structure. Its fundamental full-text index supporting substring queries.
uses in the vg toolkit will be discussed in the vg paper. It is formed by taking the suffixes of the text, storing
them in a trie, and compacting unary paths in the trie
into single edges. Although fast and versatile, suffix
2.1 Strings A string S[0, n−1] = s0 · · · sn−1 of length trees are impractical with large texts, as they require
|S| = n is a sequence of characters over an alphabet much more space than the text itself.
Σ = {0, . . . , σ−1}. Text strings T [0, n−1] are terminated
Suffix arrays (SA) [27] were introduced as a spaceby an endmarker T [n − 1] = $ = 0 not found anywhere efficient alternative to the suffix tree. The suffix array
else in the text. A substring of string S is a sequence of
1 These definitions are used in the SDSL library [15]. We assume
the form S[i, j] = si · · · sj . We call substrings of the type
S[0, j] and S[i, n − 1] prefixes and suffixes, respectively, for convenience that S.select(0, c) = −1.
2
Background
Suffixes
$
A$
ATA$
ATCATA$
CATA$
CATCATA$
GCATCATA$
TA$
TCATA$
LF BWT Suffixes
1
7
4
5
8
6
0
2
3
A
T
C
C
T
G
$
A
A
$
A$
ATA$
ATCATA$
CATA$
CATCATA$
GCATCATA$
TA$
TCATA$
SA LCP
8
7
5
2
4
1
0
6
3
0
0
1
2
0
3
0
0
1
pointer. If we find a sample at SA[LFk (i)], we know that
SA[i] = (SA[LFk (i)] + k) mod n.
If we have sampled one out of d suffix array pointers at
regular intervals, finding each occurrence takes O(d) rank
queries. If we also sample one out of d0 inverse suffix
array pointers2 , we can extract an arbitrary substring
X of the text using O(|X| + d0 ) rank queries.
The longest-common-prefix array (LCP array) [27]
is an integer array LCP[0, n − 1], where each value LCP[i]
tells the length of the longest common prefix of suffixes
Figure 1: LF-mapping, BWT, suffix array, and LCP T [SA[i−1], n−1] and T [SA[i], n−1] (with LCP[0] = 0). If
array for text GCATCATA$.
we have the FM-index, the LCP array, and the topology
of the suffix tree, we get the compressed suffix tree, which
supports the full functionality of the suffix tree in a spaceof text T [0, n − 1] is an array of pointers SA[0, n − 1]
efficient manner [40].
to the suffixes of the text in lexicographic order. We
find the occurrences of pattern string X in the text in
2.3 Graphs A graph G = (V, E) consists of a set of
O(|X| log n) time by using binary search in the suffix
nodes V = {0, . . . , |V |−1} and a set of edges E ⊆ V ×V .
array. The suffix array requires n log n bits of space
We say that (u, v) ∈ E is an edge from node u to node
in addition to the text, while its functionality is more
v, and assume that the edges are directed : (u, v) 6= (v, u)
limited than that of the suffix tree. See Figure 1 for an
for u 6= v. The indegree G.in(v) of node v is the number
example of the suffix array and related structures.
of incoming edges to v, while the outdegree G.out(v) is
The Burrows-Wheeler transform (BWT) [4] of text
the number of outgoing edges from v.
T [0, n − 1] is a permutation BWT[0, n − 1] such that
The graphs we use are labeled with alphabet Σ:
BWT[i] = T [(SA[i] − 1) mod n]. Given the lexicographic
each node v ∈ V has a label G.label(v) ∈ Σ. A path in
rank i of suffix T [SA[i], n−1], we can use LF-mapping on
a graph is a sequence of nodes P = v0 · · · v|P |−1 such
the BWT to find the lexicographic rank of the previous
that (vi , vi+1 ) ∈ E for 0 ≤ i < |P | − 1. We say that v0
suffix T [(SA[i] − 1) mod n, n − 1]. Let
is the start node and v|P |−1 is the end node of the path.
The label of path P is the concatenation of node labels
LF(i) = C[BWT[i]] + BWT.rank(i, BWT[i]),
G.label(P ) = G.label(v0 ) · · · G.label(v|P |−1 ).
If the graph has nodes with indegree or outdegree
where C[c] is the number of occurrences of characters
c0 < c in the BWT. Then SA[LF(i)] = (SA[i] − 1) mod n. 0, we add a source node s and a sink node t to it.
To distinguish these technical nodes from the actual
We generalize the definition to any character c ∈ Σ:
nodes, we label them with characters G.label(s) = #
LF(i, c) = C[c] + BWT.rank(i, c).
and G.label(t) = $, which are not used anywhere else in
the graph. We add an edge (s, v) to all nodes v ∈ V \ {s}
0
Let X be a string. If there are i suffixes S of text T with no incoming edges, and an edge (v, t) from all nodes
0
such that S < X in lexicographic order, then there are v ∈ V \ {t} with no outgoing edges. We also add edge
LF(i, c) suffixes S 0 such that S 0 < cX.
(t, s) to guarantee that G.in(v) ≥ 1 and G.out(v) ≥ 1 for
We can use the BWT as a space-efficient text index. all nodes v ∈ V . However, this edge is not considered a
The FM-index [10] combines a representation of the real edge, and no path can cross it.
BWT supporting rank/select queries, the C array, and
We also consider infinite paths P = (vi )i∈Z , genera set of sampled pointers from the suffix array. It uses alizing the definition in a similar way as we did with
backward searching to find the lexicographic range of infinite character sequences in Section 2.1. We say that
suffixes matching pattern X (having X as a prefix). If path P is left-maximal if it starts at the source node or
the lexicographic range matching suffix X[i + 1, |X| − 1] extends infinitely to the left; right-maximal if it ends
of the pattern is SA[sp, ep], the range matching suffix at the sink node or extends infinitely to the right; and
X[i, |X| − 1] is SA[LF(sp, X[i]), LF(ep + 1, X[i]) − 1]. maximal if it is both left-maximal and right-maximal.
Matching the entire pattern takes O(|X|) rank queries.
We use the sampled suffix array pointers for finding
2 The suffix array is a permutation of {0, . . . , n − 1}, and the
the text positions containing the occurrences. If SA[i]
is not sampled, we iterate LF(i) until we find a sampled inverse suffix array is the inverse permutation.
Given a graph G = (V, E), we may want to reason
about the predecessors of a node with the given label.
Let v ∈ V be a node and c ∈ Σ be a character. We write
G.pred(v, c) to denote the set of nodes u ∈ V such that
G.label(u) = c and there is an edge (u, v) ∈ E.
We work with de Bruijn graphs and their generalizations. For that purpose, we define collections of (finite
or infinite) sequences suitable for constructing order-k
de Bruijn graphs.
that includes directed acyclic graphs and de Bruijn
graphs. Before indexing, we transform the graph into
an equivalent graph, where the nodes can be sorted
unambiguously by the labels of the right-maximal paths
starting from them. The transformation increases the
size of the graph exponentially in the worst case. In
addition to sequences BWT and IN, GCSA also uses an
outdegree bitvector OUT, which is encoded in the same
way as IN. LF-mapping uses select queries on bitvector
IN to map nodes to BWT ranges, ordinary LF-mapping
Definition 2.1. (k-collection) Let S be a collection with BWT to map incoming edges to the corresponding
of character sequences over alphabet Σ, and let k > 0 be outgoing edges, and rank queries on bitvector OUT to
a parameter value. We say that S is a k-collection, if map the outgoing edges to the predecessor nodes.
each sequence S ∈ S (a) is left-infinite or begins with #k ;
(b) is right-infinite or ends with $k ; and (c) contains no 3 Path Indexes
other occurrences of characters # and $.
A path index is a generalization of text indexes for labeled
Definition 2.2. (de Bruijn graph) Let k > 0, and graphs. Given a path index for input graph G = (V, E),
let S be a k-collection. The order-k de Bruijn graph of we use the index to find the start nodes v0 ∈ V of the
paths P = v0 · · · v|X|−1 matching pattern X (paths P
S is a graph G = (V, E) such that
with G.label(P ) = X).
• each node vX ∈ V represents a distinct k-mer X
The proofs of the lemmas can be found in Apoccurring in S, with G.label(vX ) = X[0];
pendix A.
• each node vX ∈ V has a key G.key(vX ) = X; and
3.1 De Bruijn Graphs as Path Indexes The
• each edge (vX , vY ) ∈ E represents a (k + 1)-mer k-mer index is the simplest path index. It consists
X[0]Y = XY [k − 1] occurring in S.
of a set of key-value pairs (X, VX ), where X ∈ Σk is a
k-mer and VX ⊆ V is the set of the start nodes of the
We use #k as the source node s and $k as the sink node
paths matching the k-mer. If we store the pairs in a hash
t, adding the edge (t, s) in the usual way.
table, we can quickly search for patterns of length k. If
De Bruijn graphs have several properties that make we use binary search in a sorted list of pairs, queries
them useful for indexing purposes. Node keys prefix- become slower, but we gain the ability to search for
match the labels of all paths starting from the node. patterns shorter than k characters. The main drawback
This makes it possible to sort the nodes unambiguously of these basic k-mer indexes is their size, as they store
by path labels. Every substring of the k-collection is the key-value pairs explicitly.
We can represent k-mer indexes as de Bruijn graphs.
the label of a path in the de Bruijn graph, and every
path label of length at most k + 1 is a substring of the For that purpose, we define the de Bruijn graph of graph
collection. In Section 3, we develop an index structure G = (V, E) by using the collection S of the labels of the
maximal paths in the graph. If sequence S ∈ S is the
based on a generalization of de Bruijn graphs.
label of path P = (vi )i∈Z , we set S.node(S, i) = vi for
2.4 FM-Index for Graphs Suffix trees, suffix ar- all positions i ∈ Z. We transform S into a k-collection
rays, and the FM-index can be generalized to index by inserting characters # to the beginning of each nonmultiple texts. There are also generalizations to other left-infinite sequence when necessary, and characters $
combinatorial structures. The XBW transform [9] is an to the end of each non-right-infinite sequence. If S[i] is a
FM-index for labeled trees. The nodes of the tree are # we inserted, we set S.node(S, i) = s : j, where s ∈ V
sorted by the path labels from the node to the root. BWT is the source node and j is the distance to the nearest
stores the labels of the children of each node, while leaf non-inserted # in S.
nodes are marked with special characters. If a node has
k children, we encode that as a binary sequence 0k−1 1.
We concatenate these sequences to form the indegree
bitvector IN. The labels of the children of the ith node
are found in BWT[IN.select(i, 1) + 1, IN.select(i + 1, 1)].
The generalized compressed suffix array (GCSA)
[43] extends the XBW transform to a class of graphs
Definition 3.1. (de Bruijn graph of a graph)
Let G be a labeled graph, and let S be the k-collection of
maximal path labels in G. The order-k de Bruijn graph
of S is the order-k de Bruijn graph of graph G.
Let G0 = (V 0 , E 0 ) be a de Bruijn graph of graph
G = (V, E), and let S be the k-collection used to define
it. We attach a set of nodes of graph G to each node
v 0 ∈ V 0 as a value G0 .value(v):
{S.node(S, i) | S ∈ S, S[i, i + k − 1] = G0 .key(v)}.
Apart from some technicalities near the source/sink
nodes, G0 .value(v) is the set of the start nodes of the
paths matching pattern G0 .key(v) in graph G. The index
produces no false negatives (path labels that exist in
the input graph but not in the index). There may be
false positives (path labels that exist in the index but
not in the input graph) with patterns longer than k, but
we can avoid them by verifying the results in the input
graph. See Figure 2 for an example.
We can encode the de Bruijn graph as a GCSA
using (|V 0 | + o(|V 0 |))(log σ + 2) bits. By using a similar
sampling scheme for the values as in the FM-index, we
get a k-mer index that uses a couple of bytes per k-mer
for typical variation graphs (see Section 5). While this
is much less than with the basic k-mer indexes, it is still
too much for large variation graphs.
3.2 Path Graphs When a path of length k < k has
a unique label, its extensions become redundant nodes
in the order-k de Bruijn graph, if the path branches
after the first k 0 characters. By using shorter keys when
possible, we can build a smaller graph that is equivalent
to the de Bruijn graph as a path index.
0
Definition 3.2. (Path graph) Let k > 0, let S be the
k-collection of the labels of the maximal paths in graph
G = (V, E), and let K be a prefix-free set of substrings of
length k or less from S. Assume that each right-maximal
substring S ∈ S prefix-matches a string K ∈ K and that
$k ∈ S. The order-k path graph of graph G with key
set K is a graph G0 = (V 0 , E 0 ), where
• Pattern X ∈ Σ∗ matches node v ∈ V 0 , if there
is a path P 0 in G0 with G0 .label(P 0 ) = X. We
use G0 .find(X) to denote the set of nodes VX0 ⊆ V 0
matching the pattern.
• If VX0 ⊆ V 0 is the set of nodes matching pattern X, the set of Soccurrences for the pattern
is G0 .locate(VX0 ) = v0 ∈V 0 G0 .value(v 0 ). We use
X
G0 .locate(X) to denote G0 .locate(G0 .find(X)).
Lemma 3.1. (No false negatives) Let G0 = (V 0 , E 0 )
be a path graph of G = (V, E), and let X ∈ (Σ \ {#, $})∗
be a pattern. Set G0 .locate(X) contains the start nodes
of all paths P in graph G with G.label(P ) = X.
Lemma 3.2. (Context length) Let G0 = (V 0 , E 0 ) be
a path graph, and let X ∈ (Σ \ {#, $})∗ be a pattern.
Set G0 .f ind(X) consists of all nodes v 0 ∈ V 0 such that
X[0, m − 1] is a prefix of G0 .key(v 0 ), for a context length
m, which depends on the graph and the pattern.
Lemma 3.3. (Short keys) Let G0 = (V 0 , E 0 ) be a path
graph with |G0 .key(u0 )| ≤ |G0 .key(v 0 )| + 1 for all edges
(u0 , v 0 ) ∈ E 0 . Then (a) |G0 .pred(v 0 , c)| ≤ 1 for all nodes
v 0 ∈ V 0 and characters c ∈ Σ; and (b) key G0 .key(v 0 )
prefix-matches pattern X ∈ (Σ \ {#, $})∗ for all nodes
v 0 ∈ G0 .find(X).
A path graph may produce false positives with
patterns longer than k 0 characters, where k 0 is the length
of the shortest key. In the next section, we define a class
of path graphs that can be proven to be equivalent to
de Bruijn graphs.
3.3 Pruned de Bruijn Graphs We can compress
de Bruijn graphs structurally by merging keys sharing a
common prefix, if the corresponding values are identical.
• each node vK ∈ V 0 represents a distinct key K ∈ K, These pruned de Bruijn graphs, which arise naturally
with G0 .label(vK ) = K[0];
from GCSA construction, are similar to manifold de
Bruijn graphs [24]. As path indexes, they are equivalent
0
• each node has a key G .key(vK ) = K and a value
to de Bruijn graphs with patterns of length up to k
0
G .value(vK ) = VK , where VK is the set of nodes
characters.
S.node(S, i) ∈ V for S ∈ S and positions i such
that S[i, i + |K| − 1] = K; and
Definition 3.4. (Equivalent path graphs) Let G0
00
let k > 0 be a parameter
• each edge (vK , vK 0 ) ∈ E 0 represents the occurrence and G be two path graphs, and
0
value.
We
say
that
graphs
G
and
G00 are k-equivalent,
0
of substring K[0]K in S such that strings K and
0
00
if we have G .locate(X) = G .locate(X) for all patterns
K[0]K 0 prefix-match.
X ∈ (Σ \ {#, $})∗ with 1 ≤ |X| ≤ k.
We use #k as the source node s and $k as the sink node
Definition 3.5. (Pruned de Bruijn graph) Let G
t, adding the edge (t, s) in the usual way.
be a graph, and let G0 be an order-k path graph of G.
Path graph G0 is an order-k pruned de Bruijn graph, if
Definition 3.3. (Path graph as an index)
0
0
0
Let G = (V, E) be a graph, and let G = (V , E ) be a it is k-equivalent to the order-k de Bruijn graph of G.
path graph of G.
ATC
3
GCA
1
3:A
0:#
1:G
2:C
6:C
5:T
4:T
CAT
2, 6
TCA
5
ATA
7
ATG
3
7:A
9:T
10:A
11:$
###
0:2
##G
0:1
#GC
0
TA$
9
A$$
10
$$$
11
TTC
4
8:G
GCT
1
CTT
2
TGT
5
GTA
8
TTG
4
Figure 2: Left: Input graph G = (V, E), with each node v ∈ V labeled with v : G.label(v). Right: The order-3
de Bruijn graph G0 = (V 0 , E 0 ) of graph G, with each node v 0 ∈ V 0 labeled with G0 .key(v 0 ) and G0 .value(v 0 ). Both:
Edges (t, s) are not shown. The highlighted path in the de Bruijn graph is a false positive, as it consists of two
disjoint paths in the input graph.
Lemma 3.4. (No short false positives)
Let G = (V, E) be a graph, let G0 = (V 0 , E 0 ) be an
order-k pruned de Bruijn graph of graph G, and let
X ∈ (Σ \ {#, $})∗ be a pattern with 1 ≤ |X| ≤ k. Then
G0 .locate(X) is a set of start nodes v ∈ V of paths
matching the pattern in graph G.
and outdegrees in bitvectors IN and OUT, and store
the predecessor labels in BWT. See Figure 3 for an
example. If lexicographic range [spi+1 , epi+1 ] matches
suffix X[i + 1, |X| − 1] of pattern X, we can find the
range [spi , epi ] matching suffix X[i, |X| − 1] as
Lemma 3.5. (Pruning) Let G = (V, E) be a graph, let
G0 = (V 0 , E 0 ) be the order-k pruned de Bruijn graph
of G with key set K, let K ∈ Σ∗ be a string of length
|K| > 0, and let VK0 be the set of nodes v 0 ∈ V 0 having
string K as a proper prefix of G0 .key(v 0 ). If |VK0 | > 0 and
G0 .value(u0 ) = G0 .value(v 0 ) for all u0 , v 0 ∈ VK0 , the path
graph with key set (K \ {G0 .key(v 0 ) | v 0 ∈ VK0 }) ∪ {K} is
an order-k pruned de Bruijn graph of G.
[spin , epin ] = [IN.select(spi+1 , 1) + 1,
We can compress a de Bruijn graph structurally
by merging sets of nodes sharing a common prefix of
their keys, as long as the conditions of Lemma 3.5
hold. Let G0 = (V 0 , E 0 ) be an order-k pruned de Bruijn
graph, and let G00 = (V 00 , E 00 ) be the same graph after
further pruning. Each node v 00 ∈ V 00 is an equivalence
class of nodes V 0 (v 00 ) ⊆ V 0 corresponding to a shared
prefix G00 .key(v 00 ) of keys. For all v 0 ∈ V 0 (v 00 ), we have
G0 .value(v 0 ) = G00 .value(v 00 ). See Figure 3 for an example
of a pruned de Bruijn graph.
Definition 3.6. (Maximally pruned graph)
Let G0 be a a pruned de Bruijn graph of graph G. We
say that G0 is maximally pruned, if we cannot prune it
any further using Lemma 3.5.
Lemma 3.6. (Maximal pruning) Let G0 = (V 0 , E 0 )
be a maximally pruned de Bruijn graph of G = (V, E).
Then |G0 .key(u0 )| ≤ |G0 .key(v 0 )| + 1 for all (u0 , v 0 ) ∈ E 0 .
4
GCSA2
As in the original GCSA, we sort the nodes of the
path graph in lexicographic order, encode the indegrees
IN.select(epi+1 + 1, 1)];
[spout , epout ] = [LF(spin , X[i]),
LF(epin + 1, X[i]) − 1];
[spi , epi ] = [OUT.rank(spout , 1),
OUT.rank(epout , 1)].
In order to support locate queries, we sample the
values of a node, if the node has multiple incoming edges
or the values cannot be derived from the predecessor.
We may also sample other values to improve query
performance. The sampled nodes are marked in bitvector
BS , the number of values in each sample is encoded in
unary in bitvector BV , and the sampled values are stored
in integer array VS . A detailed description of the data
structure can be found in Appendix B.
GCSA2 construction starts from paths of length k in
the input graph. We build a maximally pruned order-2k,
order-4k, or order-8k de Bruijn graph using a similar
prefix-doubling algorithm as in the original GCSA [43],
and encode the result as a GCSA. To avoid excessive
memory usage, we keep the paths and the graphs on
disk, and read only a single chromosome at a time into
memory. The details of the construction algorithm can
be found in Appendix C.
We can improve the query performance with (maximally pruned) de Bruijn graphs by using a simplified
encoding (Appendix D). We replace BWT and IN with
bitvectors Bc for all c ∈ Σ, where Bc [j] = 1 if and only
if node j in lexicographic order has a predecessor with
ATC
3
CA
2, 6
TC
5
ATA
7
ATG
3
###
0:2
##G
0:1
#G
0
GC
1
TA
9
CT
2
TT
4
TG
5
A$
10
$$$
11
GT
8
key
OUT
BWT
IN
key
BS
$$$
A$
ATA
ATC
ATG
CA
CT
GC
GT
TA
TC
TG
TT
#G
##G
###
1
1
1
1
1
0
0
1
1
0
1
1
1
1
1
0
1
1
1
1
A
T
C
C
C
G
T
G
#
T
A
G
A
T
A
T
C
#
#
$
1
1
1
1
1
0
1
1
1
1
0
1
0
1
0
1
1
1
1
1
$$$
A$
ATA
ATC
ATG
CA
CT
GC
GT
TA
TC
TG
TT
#G
##G
###
0
0
1
1
1
1
0
0
1
1
1
1
1
0
0
1
BV
VS
1
1
1
0
1
1
1
1
1
1
1
7
3
3
2
6
8
9
5
5
4
0:2
Figure 3: Left: An order-3 pruned de Bruijn graph G00 3-equivalent to the de Bruijn graph in Figure 2. Right:
GCSA for graph G00 . Leftward arrows illustrate backward searching, with the red arrows showing it from T to AT.
Rightward arrows mark the samples belonging to each node, with the blue ones showing them for node CAT.
label c. This simplifies backward searching to
spout = C[X[i]] + BX[i] .rank(spi+1 , 1);
epout = C[X[i]] + BX[i] .rank(epi+1 + 1, 1) − 1;
[spi , epi ] = [OUT.rank(spout , 1), OUT.rank(epout , 1)].
The compacted trie of keys resembles a suffix tree.
We can simulate it space-efficiently by using the LCP
array [1, 11], and thus extend GCSA2 to support many
suffix tree operations. For example, we can search for
maximal exact matches by using LF-mapping and parent
queries [35], and use that as a basis for a read aligner
similar to BWA-MEM [21]. We can also use document
counting techniques [41] to quickly count the number
of distinct matches in a lexicographic range. Further
details of these extensions can be found in Appendix E.
5
Implementation and Experiments
GCSA2 is the path indexing library of vg [14]. The
implementation is written in C++, and the source code
is available on GitHub.3 It depends on SDSL [15] and
libstdc++ parallel mode. We use the simplified encoding
(Appendix D) with fast non-compressed bitvectors
in most index components. Bitvectors Bc for rare
characters (N, #, and $) are compressed as sparse
bitvectors [36].
We used a system with two 16-core AMD Opteron
6378 processors and 256 gigabytes of memory for the
experiments, and stored all files on a distributed Lustre
file system. The system was running Ubuntu 12.04
on Linux kernel 3.2.0. We used vg version 1.3.0 for
processing the graphs and GCSA2 version 0.8 using
SDSL version 2.1.1 for the benchmarks. All code was
compiled with gcc/g++ version 4.9.2.
3 https://github.com/jltsiren/gcsa2
5.1 Construction Variation graphs, as defined in vg,
use strings as node labels. A node can be traversed
in both forward and reverse complement orientations,
and edges may cross between the orientations. For
indexing, the graph is implicitly converted into an input
graph with single-character labels. We always sample
the input graph nodes corresponding to the initial offsets
of variation graph node labels.
We built vg graphs from the human reference genome
(GRCh37) and 1000 Genomes Project variation [45]. To
avoid excessive growth, we removed paths where 16-mers
crossed more than 4 nontrivial edges with vg mod -p
-l -e 4, and subgraphs shorter than 100 bases with vg
mod -S -l 100. We extracted all paths of length 16
from the forward strand of the graph. There were a total
of 4.80 billion paths with 1.53 billion distinct labels. We
then built GCSA with 1–3 doubling steps, producing
order-32, order-64, and order-128 indexes.
Tables 1 and 2 show construction requirements
and index sizes, respectively. We can build a wholegenome index overnight using less than 96 gigabytes
of memory, including disk cache. The index contains
1.031k ·2.348 billion k-mers, but the path graph only uses
4.4–5.7 billion nodes to represent them. For k = 128,
GCSA2 requires 0.63 bits per k-mer, out of which
0.28 bits is used for the path graph. Extensions based
on suffix trees increase the size to 1.08 bits per k-mer.
Index construction uses more memory with k = 32
than with larger values of k. The order-32 path graph
has more nodes, where we cannot derive the values from
the predecessor node. As we sample more values, we
need more memory in the final phase of construction.
With larger values of k, the path graph resembles the
input graph better, and we sample less values. For the
same reason, index size decreases with larger values of
k, even though the graph requires more space.
k k-mers Nodes
Graph
Index
With extensions
32 6.20G 4.37G 2.89 GB / 4.00 bits 9.50 GB / 13.2 bits 13.2 GB / 18.2 bits
64 16.7G 5.24G 3.46 GB / 1.78 bits 8.64 GB / 4.46 bits 13.6 GB / 6.99 bits
5.73G 3.78 GB / 0.28 bits 8.58 GB / 0.63 bits 14.6 GB / 1.08 bits
128 116G
Table 2: GCSA2 index sizes. Order of the path graph; number of k-mers and nodes in the path graph in billions;
index size in gigabytes and in bits per k-mer for the graph (Bc and OUT), the index (the graph, BS , BV , and VS ),
and the index with the extensions from Appendix E.
k
32
64
128
Time Memory Disk
Read
Write
7.44 h 59.8 GB 387 GB 1.37 TB 0.88 TB
10.4 h 51.9 GB 415 GB 2.03 TB 1.51 TB
14.1 h 52.3 GB 478 GB 2.78 TB 2.25 TB
Table 1: GCSA2 construction. Order of the path graph;
construction time in hours; peak memory and disk usage
in gigabytes; and disk I/O volume for reading and writing
in terabytes.
5.2 Queries We compared the query performance of
the order-128 GCSA2 to several FM-indexes for the
reference sequence. SSA is the SDSL implementation
(csa_wt<>) of the succinct suffix array [30], using a
Huffman-shaped wavelet tree on top of the BWT. As
the default FM-index in SDSL, it prioritizes query
performance over compression. We used SSA with SA
sample period 17 for good locate performance. BWA
is the FM-index in the Burrows-Wheeler Aligner [22]
(version 0.7.15 with the default SA sample period 32).
Optimized for DNA sequences, BWA indexes both the
reference and its reverse complement.
As building the original GCSA requires around 65n
bytes of memory for a path graph with n nodes, we could
not compare GCSA and GCSA2 directly on a system
with 256 gigabytes of memory. Instead, we used RLCSA
[31] (May 2016 version) as a proxy. The RLCSA is an
FM-index for repetitive sequence collections using the
same basic components as the original GCSA. Under a
mixed query load, RLCSA with SA sample period 32 is
1.5x to 3x faster than GCSA, depending on algorithmic
overhead and the mix of find and locate queries [43].
We extracted k-mers for k ∈ {16, 32, 64, 128} from
the (non-pruned) vg graphs by using vg sim, filtered
out k-mers consisting entirely of Ns, and queried for the
remaining k-mers using a single thread. The results can
be seen in Table 3.
Backward searching in an FM-index stops early if
there are no matches. In order to compare the find
performance of the indexes reliably, we must hence
concentrate on the 16-mers, where the fraction of
matching patterns is similar for all indexes. GCSA2 and
the fast FM-indexes (SSA and BWA) all have similar
performance, while RLCSA is several times slower. As a
result, we can estimate that find queries in GCSA2 are
an order of magnitude faster than in GCSA.
When comparing the locate performance of different
FM-indexes, the distribution of the query positions
should be close to uniform. Otherwise the biases from
e.g. different suffix array sampling strategies or the
variation in the number of distinct occurrences per node
in GCSA2 can make the results unreliable. As the k-mers
have been sampled uniformly from the variation graph,
we get the best results with the 16-mers, where all indexes
can match most of the patterns.
GCSA2 uses denser SA sampling than the other
indexes, with effective sample period 10.6. On the
average, GCSA2 calls locate for 2.84 nodes per distinct
value, making the amount of work comparable to sample
period 30.2. SSA with sample period 17 is 2.4x faster
than GCSA2, mostly because it has to do less work.
BWA with sample period 32 is closer to GCSA2 in locate
performance. RLCSA is slower than the other indexes,
but the difference is smaller than with find queries due
to the optimizations for retrieving suffix array ranges.
Assuming that locate queries are 3x slower in GCSA than
in RLCSA, as GCSA does not use the optimizations, we
can estimate that GCSA2 is 4x faster than GCSA.
The remaining queries, parent and count, take a
fraction of a microsecond. As a parent query takes
comparable time to a single step of backward searching,
it will not be a bottleneck in finding maximal exact
matches. Counting the number of distinct occurrences
with a count query is faster than retrieving even a single
occurrence.
6
Discussion
GCSA2 is a path index for variation graphs. It uses a
de Bruijn graph as a k-mer index of the variation graph,
prunes it by merging redundant subgraphs, and encodes
the result with a generalization of the FM-index. The
index supports queries of length up to k exactly, and
longer queries with false positives. GCSA2 also includes
Patterns Index
351584 GCSA2
SSA
BWA
RLCSA
32
351555 GCSA2
SSA
BWA
RLCSA
64
351567 GCSA2
SSA
BWA
RLCSA
128 351596 GCSA2
SSA
BWA
RLCSA
k
16
Found Nodes Occs find() parent()
347453 2477M 872M 4.75 μs 0.42 μs
301538
–
782M 6.00 μs
–
320764
–
1564M 3.64 μs
–
301538
–
782M 23.7 μs
–
333258 112M 34.3M 10.8 μs 0.28 μs
153957
–
26.6M 10.9 μs
–
156080
–
52.9M 6.57 μs
–
153957
–
26.6M 47.6 μs
–
326101 2.63M 1.35M 22.5 μs 0.26 μs
88184
–
0.84M 17.1 μs
–
88786
–
1.60M 10.3 μs
–
88184
–
0.84M 74.3 μs
–
316500 0.32M 0.37M 45.3 μs 0.26 μs
35678
–
0.08M 23.5 μs
–
35741
–
0.12M 14.0 μs
–
35678
–
0.08M 91.7 μs
–
count()
0.87 μs
–
–
–
0.38 μs
–
–
–
0.29 μs
–
–
–
0.26 μs
–
–
–
locate()
5.85 μs
2.43 μs
4.65 μs
8.12 μs
5.44 μs
2.16 μs
3.19 μs
5.87 μs
2.92 μs
1.89 μs
2.34 μs
5.97 μs
3.13 μs
3.47 μs
3.46 μs
12.9 μs
Table 3: Query benchmarks using an order-128 GCSA2 and various FM-indexes. Pattern length; number of
patterns; index type; number matching patterns, matching nodes, and distinct occurrences; average time for
[sp, ep] = find(X), parent(sp, ep), and count(sp, ep) queries in microseconds; and average time per value for
locate(sp, ep) queries in microseconds.
extensions based on suffix trees; other extensions have
been considered but not implemented (see Appendix F).
The index is used in the variation graph toolkit vg for
e.g. read alignment based on maximal exact matches.
We can build a whole-genome index overnight on a
system with 96 gigabytes of memory and a few hundred
gigabytes of fast disk space. The resulting index takes
less than 15 gigabytes, or 1.08 bits per k-mer for the
order-128 index with extensions. Query performance is
comparable to that of fast FM-indexes for sequences.
The primary design goals for GCSA2 were query
performance and index size. The index works with
arbitrary graphs, supports queries that are long enough
to map short reads in one piece without false positives,
and provides several options for dealing with complex
regions. Other path indexes work with a more restricted
class of graphs [43, 17, 19, 26], are at least an order
of magnitude slower [17, 26], require much more space
[3, 38, 5, 28], or are theoretical proposals that have never
been implemented [44].
We may want to determine whether a pattern
matches known haplotypes or only their recombinations.
As GCSA2 does not support this directly, vg must
determine it afterwards using a separate structure
[34]. The FM-index of alignment [32, 33] embeds the
haplotypes directly in a GCSA-like index and reports the
haplotypes matching the find query. While the solution
depends on specific properties of the graph, it could be
possible to extend it to work with any GCSA.
Acknowledgements. The author thanks Erik Garrison, Richard Durbin, and Adam Novak for the fruitful
discussions while developing the GCSA2 index.
A
Proofs of Lemmas
Proof. [Lemma 3.1: No false negatives] Let S be the
k-collection used for building the path graph, and let P
be a path starting from v0 ∈ V with G.label(P ) = X.
The collection contains a sequence S ∈ S such that
S[i, i + |X| − 1] = X and S.node(S, i) = v0 .
For all positions j with i ≤ j ≤ i + |X| − 1,
there is a node vj0 ∈ V 0 with G0 .label(vj0 ) = S[j] and
G0 .key(vj0 ) = S[j, j +|G0 .key(vj0 )|−1]. By definition, path
0
graph G0 has an edge (vj0 , vj+1
) ∈ E 0 for all such positions
0
0
0
j. Hence P = vi · · · vi+|X|−1 is a path in G0 with
G0 .label(P 0 ) = X. As path P 0 starts from node vi0 ∈ V 0 ,
node vi0 is included in the set G0 .find(X). Furthermore,
v0 = S.node(S, i) ∈ G0 .value(vi0 ) ⊆ G0 .locate(X).
Proof. [Lemma 3.2: Context length] If |X| ≤ 1, the
statement is true by definition for m = |X|. Now let
Mi+1 = G0 .find(X[i + 1, |X| − 1]) be the set of all nodes
v 0 ∈ V 0 such that substring X[i + 1, i + mi+1 ] is a prefix
of key G0 .key(v 0 ), and assume that the set is nonempty.
Consider the set
[
Mi =
G0 .pred(v 0 , X[i]) = G0 .find(X[i, |X| − 1]).
v 0 ∈Mi+1
There is an edge (u0 , v 0 ) ∈ E 0 if and only if key G0 .key(u0 )
prefix-matches string G0 .label(u0 ) · G0 .key(v 0 ). Hence key
G0 .key(u0 ) prefix-matches string X[i, i + mi+1 ] for all
nodes u0 ∈ Mi .
Now let u0 ∈ V 0 be a node with key G0 .key(u0 ) prefixmatching string X[i, i + mi+1 ]. If the key is a prefix of
string X[i, i + mi+1 ], there is an edge (u0 , v 0 ) ∈ E 0 to
all nodes v 0 ∈ Mi+1 , and hence u0 ∈ Mi . Otherwise let
S[j, j + |G0 .key(u0 )| − 1] = G0 .key(u0 ) be a substring of
the k-collection used for building the path graph. As
S[j + 1, j + mi+1 ] = X[i + 1, i + mi+1 ], the substring
starting at S[j + 1] is represented by a node v 0 ∈ Mi+1 ,
and hence u0 ∈ Mi .
Set Mi is the set of all nodes u0 ∈ V 0 such that
substring X[i, i + mi+1 ] prefix-matches key G0 .key(u0 ).
If |Mi | > 1, string X[i, i + mi+1 ] is a proper prefix of key
G0 .key(u0 ) for all nodes u0 ∈ Mi due to the prefix-free
property, and we can set mi = mi+1 + 1. Otherwise we
set mi = min(mi+1 + 1, |G0 .key(u0 )|) for the only node
u0 ∈ Mi .
Proof. [Lemma 3.3: Short keys] (a) If node v has
multiple predecessors with label c, the keys of the
predecessors must be longer than string c · G0 .key(v 0 ), as
the key set is prefix-free.
(b) By the construction in the proof of Lemma 3.2,
the context length for pattern X in graph G0 is
min(|X|, |G0 .key(v 0 )|).
0
00
or (b) u00 = vK
and u0 ∈ VK0 , and the same holds for
00
0
nodes v and v . We can hence transform any path
in graph G0 into an equivalent path in graph G00 by
00
replacing nodes v 0 ∈ VK0 with node vK
.
00
00
00
Let P = v0 · · · v|P 00 |−1 be a in graph G00 . We
0
transform it into an equivalent path P 0 = v00 · · · v|P
00 |−1
0
00
in graph G . There are two cases for |P | = 1. If
00
v000 = vK
, we can replace it with any v00 ∈ VK0 . Otherwise
we use the node v00 ∈ V 0 with G0 .key(v00 ) = G00 .key(v000 ).
In the general case |P 00 | > 1, assume that we have
00
00
transformed the suffix v100 · · · v|P
into
00 |−1 of path P
0
0
0
an equivalent path v1 · · · v|P 00 |−1 in graph G . Because
G0 .value(v10 ) = G00 .value(v100 ), node v10 must have a
predecessor with label c = G00 .label(v000 ). We can choose
any such predecessor as node v00 .
Consider the predecessors v 0 ∈ G0 .pred(v10 , c) and
00
v ∈ G00 .pred(v100 , c). Their keys prefix-match string
Y = c · K1 . There are three cases:
1. If Y is a prefix of G0 .key(v 0 ), the key G0 .key(w0 ) of
every successor w0 of node v 0 prefix-matches K1 and
hence has K1 as a prefix. Therefore G0 .value(v 0 ) is
the union of sets G.pred(v, c) over v ∈ G0 .value(v10 ).
If Y is also a prefix of G00 .key(v 00 ), nodes v 0 and v 00
have identical value sets by the same reasoning.
2. If Y is a prefix of G0 .key(v 0 ) and G00 .key(v 00 ) is
a proper prefix of Y , there is only one possible
00
predecessor v 00 = v000 . Hence v 0 ∈ VK0 and v000 = vK
.
Proof. [Lemma 3.4: No short false positives] We may
assume without loss of generality that graph G0 is a
3. If key G0 .key(v 0 ) is a proper prefix of string Y , there
de Bruijn graph. Let v 0 ∈ G0 .find(X) be a node. By
is only one possible predecessor v 0 = v00 . Because
Lemma 3.3, pattern X prefix-matches key G0 .key(v 0 ).
key G00 .key(v000 ) prefix-matches string Y , it must
For every node v ∈ G0 .value(v 0 ), there is a substring
also be a proper prefix of the string. Hence either
S[i, i + k − 1] = G0 .key(v00 ) in the k-collection S used
00
.
G0 .key(v00 ) = G00 .key(v000 ) or v00 ∈ VK0 and v000 = vK
0
for building graph G , with S.node(S, i) = v. Prefix
0
0
00
00
S[i, i + |X| − 1] = X of the substring corresponds to a In every case, G .value(v0 ) = G .value(v0 ), and we can
path with label X starting from node v in graph G.
use shorter of strings Y and G00 .key(v000 ) as K0 .
We can transform any path P 0 in graph G0 into
Definition A.1. (Equivalent paths)
an equivalent path P 00 in graph G00 , and the other
0
00
Let G and G be path graphs of the same graph, and way around. Because the labels of equivalent paths
0
00
let P 0 = v00 · · · vn−1
and P 00 = v00 · · · vn−1
be paths in and the value sets of their start nodes are identical,
0
00
graphs G and G , respectively. We say that paths P 0 we have G0 .locate(X) = G00 .locate(X) for all patterns
and P 00 are equivalent, if for 0 ≤ i < n, keys G0 .key(vi0 ) X ∈ (Σ \ {#, $})∗ with |X| > 0.
and G00 .key(vi00 ) have a common prefix Ki such that
G0 .value(v 0 ) = G00 .value(v 00 ) for all nodes v 0 ∈ V 0 and Proof. [Lemma 3.6: Maximal pruning] Let Gd = (Vd , Ed )
the de Bruijn graph of graph G with the same order
v 00 ∈ V 00 having Ki as a prefix of their keys.
k as graph G0 , and let S be the k-collection used for
00
00
00
Proof. [Lemma 3.5: Pruning] Let G = (V , E ) be building the path graphs. If v 0 ∈ V 0 is a node, then
the path graph corresponding to the new key set, let Gd .value(v) = G0 .value(v 0 ) for all nodes v ∈ Vd (v 0 ).
00
vK
∈ V 00 be the node with key K, and let S be the
Assume that |G0 .key(u0 )| > |G0 .key(v 0 )| + 1 for an
k-collection used to define the path graphs.
edge (u0 , v 0 ) ∈ E 0 . String G0 .label(u0 ) · G0 .key(v 0 ) must
0 0
0
Consider the edge (u , v ) ∈ E defined by substring then be a prefix of key G0 .key(u0 ). There cannot be edges
0
S of S. The same substring also defines an edge (u0 , w0 ) to other nodes w0 =
6 v 0 , as keys G0 .key(v 0 ) and
00 00
00
00
00
0
0
0
0
(u , v ) ∈ E , where either (a) G .key(u ) = G .key(u ) G .key(w ) would prefix-match.
Let S[i, i + |G0 .key(u0 )| − 1] = G0 .key(u0 ) be a the range [spi , epi ] = G0 .find(X[i, |X| − 1]):
substring of S. Because G0 .key(v 0 ) is a substring of
[spin , epin ] = [IN.select(spi+1 , 1) + 1,
G0 .key(u0 ) and G0 is a pruned de Bruijn graph, the
IN.select(epi+1 + 1, 1)];
set of nodes S.node(S, i + 1) over all occurrences of
0
0
0
0
0
substring G .key(u ) in S is G .value(v ). As node u has
[spout , epout ] = [LF(spin , X[i]),
no other successors, set G0 .value(u0 ) is the union of sets
LF(epin + 1, X[i]) − 1];
G.pred(v, G0 .label(u0 )) over all nodes v ∈ G0 .value(v 0 ).
[spi , epi ] = [OUT.rank(spout , 1),
The above is true for all x0 ∈ G0 .pred(v 0 , G0 .label(u0 )).
0
OUT.rank(epout , 1)].
Hence we can prune graph G further using string
0
0
0
0
G .label(u ) · G .key(v ) as the new key in Lemma 3.5.
We can think this as a generalization of LF-mapping:
[spi , epi ] = G0 .LF([spi+1 , epi+1 ], X[i]).
B GCSA for Path Graphs
Query G0 .locate(X) retrieves the values G0 .value(v 0 )
0
0
0
Let G = (V , E ) be a path graph. We sort the nodes for nodes v 0 ∈ G0 .find(X) and filters out duplicates.
V 0 by their keys in lexicographic order and generate the Instead of storing the values explicitly for all nodes,
sequences BWT, IN, and OUT from the nodes in that GCSA uses a sampling scheme to save space. We
order. For each node v 0 ∈ V 0 , we append BWT with the assume that the nodes of the input graph G = (V, E) are
predecessor labels G.label(u0 ) for all edges (u0 , v 0 ) ∈ E 0 ; conveniently chosen integers. If (u, v) ∈ E is the only
0
0
IN with the indegree encoded as 0G .in(v )−1 1; and OUT outgoing edge from node u and the only incoming edge
0
0
with the outdegree as 0G .out(v )−1 1.
to node v, it should be that v = u + 1.
If node v 0 ∈ V 0 has lexicographic rank i, the range of
We sample the values G0 .value(v 0 ) for a node v 0 ∈ V 0 ,
0 0
0
incoming edges (u , v ) ∈ E to that node is [spin , epin ] = (a) if there are multiple incoming edges to node v 0 ;
[IN.select(i, 1) + 1, IN.select(i + 1, 1)]. The labels of (b) if v 0 is the source node s; or (c) if G0 .value(v 0 ) 6=
the predecessor nodes are encoded in BWT[spin , epin ]. {u + 1 | u ∈ G0 .value(u0 )} for the only incoming edge
Sorting the incoming edges by pairs (BWT[j], i), where (u0 , v 0 ) ∈ E 0 . We may also sample the values for some
BWT[j] corresponds to edge (u0 , v 0 ) ∈ E 0 , is equivalent nodes on long unary paths for performance reasons. If
to sorting them by strings G0 .label(u0 ) · G0 .key(v 0 ). As the set G0 .value(v 0 ) has not been sampled, we can derive
multiple edges may have the same sort key, our sorting it from sampled values by following the incoming edges.
algorithm must be stable. We get the desired sorting
If node v 0 ∈ V 0 with lexicographic rank i has only
order by using LF-mapping: j 7→ LF(j).
one predecessor, the lexicographic rank of the predecessor
The range of outgoing edges (u0 , v 0 ) ∈ E 0 from node is G0 .LF(i) = OUT.rank(LF(IN.select(i, 1) + 1), 1). If
u0 ∈ V 0 with lexicographic rank i0 is [spout , epout ] = the lexicographic rank G0 .LFk (i) corresponding to node
[OUT.select(i0 , 1)+1, OUT.select(i0 +1, 1)]. The edges are w0 ∈ V 0 is the first sampled node we encounter, we know
already sorted by keys G0 .key(u0 ). Because graph G0 is a that G0 .value(v 0 ) = {w + k | w ∈ G0 .value(w0 )}.
path graph, we know that key G0 .key(u0 ) prefix-matches
Let BS [0, |V 0 |−1] be a bitvector. If we have sampled
0
0
0
0
string G .label(u ) · G .key(v ). The sorting orders are the values for the node v 0 ∈ V 0 with lexicographic
therefore compatible. For every j ∈ [spin , epin ] for a rank i, we mark that as BS [i] = 1. We can then
node v 0 ∈ V 0 , having LF(j) ∈ [spout , epout ] for a node determine the rank of node v 0 among the sampled nodes
u0 ∈ V 0 implies an edge (u0 , v 0 ) ∈ E 0 .
as j = BS .rank(i, 1). For each sampled node v 0 ∈ V , we
0
We use backward searching for query G .find(X). store the size of the value set |G0 .value(v 0 )| in another
Let X ∈ (Σ \ {#, $})∗ be a pattern. If |X| = 0, query bitvector BV , using the same encoding as for bitvectors
G0 .find(X) returns the lexicographic range [0, |V 0 | − 1] IN and OUT. We store the samples in array VS in the
containing all nodes. Now assume that |X| ≥ 1 and same order, using log|V | bits each. The sampled values
that G0 .find(X[i + 1, |X| − 1]) = [spi+1 , epi+1 ]. We want for node v 0 with rank j among the sampled nodes can
to find the lexicographic range G0 .find(X[i, |X| − 1]), be found at VS [BV .select(j, 1) + 1, BV .select(j + 1, 1)].
which is the union of sets G0 .pred(v 0 , X[i]) over nodes
v 0 ∈ G0 .find(X[i + 1, |X| − 1]). We map the node C Index Construction
range [spi+1 , epi+1 ] to the range [spin , epin ] of incoming GCSA construction [43] is based on the prefix-doubling
edges; the incoming edges to the corresponding range of algorithm for suffix array construction [27]. The original
outgoing edges [spout , epout ]; and the outgoing edges to GCSA started from paths of length 1 in the input graph,
and then repeatedly joined paths of length k into paths
of length 2k, until each path had a distinct label. The
resulting path graph was essentially an order-∞ pruned
de Bruijn graph and supported queries of any length.
We use a variant of that algorithm with GCSA2.
Let G = (V, E) be the input graph. We extract all paths
of length k (typically with k = 16) from graph G. For
each path P = v0 · · · v|P |−1 , we store several fields. Key
P.key encodes G.label(P ) as a sequence of lexicographic
ranks of k-mers. If |P | is not an integer multiple of k, the
key consists of the k-mer ranks for the lexicographically
smallest (d|P |/ke · k)-mer having G.label(P ) as a prefix,
followed by the rank of the last k-mer in the largest
such (d|P |/ke · k)-mer. Value P.value is the start node
v0 of the path. We store the set of predecessor labels
{c ∈ Σ | |G.pred(v0 , c)| > 0} as P.pred. For each possible
extension node v ∈ {v ∈ V | (v|P |−1 , v) ∈ E}, we
create a separate copy of the path and store the node as
P.ext = v.
The construction uses several supporting structures.
We build an order-k de Bruijn graph Gd = (Vd , Ed ) of
the path labels and encode it as a GCSA, using the
predecessor labels P.pred for determining the edges. Let
v0 , . . . , v|Vd |−1 be the nodes of the de Bruijn graph in
lexicographic order by their keys. We use two additional
arrays: the LCP array LCP[0, |Vd | − 1], where LCP[i]
is the length of the longest common prefix of keys
Gd .key(vi−1 ) and Gd .key(vi ) (with LCP[0] = 0), and
the last character array L[0, |Vd | − 1], where L[i] =
Gd .key(vi )[k − 1]. The LCP array is stored as a wavelet
tree for fast range minimum queries [13].
Because we store path labels explicitly, we only do a
limited number of doubling steps, typically 2 or 3. After
d doubling steps, the length of the paths is 2d k, and we
can use them to build a maximally pruned order-(2d k)
de Bruijn graph. Each doubling step consists of a pruning
step, followed by an extension step. The pruning step
applies a limited form of Lemma 3.5 to lexicographic
ranges of paths. Given two paths P and P 0 , we can
determine the length of the longest common prefix of
the path labels by using the keys P.key and P 0 .key and
the LCP array. If all paths sharing a prefix start from
the same node, we merge them into a single path Q with
Q.key based on the shared prefix and Q.ext = −1.
The extension step transforms the current set of
paths of length (up to) k 0 into a set of paths of length
(up to) 2k 0 . If P is a path with P.ext = −1, we use
it as such. If P.ext = P 0 .value for paths P and P 0 , we
create a new path P P 0 . We set (P P 0 ).key according to
the concatenation of the path labels, take value and pred
from path P , and take ext from path P 0 . If we have
another path Q with Q.key = P.key such that QP 0 is a
path, and if P 0 .ext = −1, all possible 2k 0 -mer extensions
of label G.label(P P 0 ) are also labels of paths starting
from node Q.value, and the other way around. Hence
paths P P 0 and QP 0 can be represented by a single node
in a pruned de Bruijn graph.
The doubling steps are followed by the merging
step, which transforms the paths into the nodes of a
maximally pruned de Bruijn graph G00 = (V 00 , E 00 ).
We merge the paths with identical keys into the nodes
of a pruned de Bruijn graph G0 = (V 0 , E 0 ). If paths
P0 , · · · , Pm−1 all have the same key, we create a node
v 0 ∈ V 0 with the shared key as G0 .key(v 0 ) and with
Sm−1
G0 .value(v 0 ) = i=0 Pi .value. We also store the union
Sm−1
of predecessor labels as v 0 .pred = i=0 Pi .pred. We
then apply Lemma 3.5 maximally, transforming graph
G0 into graph G00 .
Storing the paths and the graphs may require
hundreds of gigabytes of memory when indexing wholegenome variation graphs. To avoid that, we keep them
on disk when possible. The subgraph corresponding to
each chromosome is stored in a separate file, with paths
sorted by their labels in lexicographic order. Extension
steps are done separately for each of the chromosomes.
The pruning step merges the sorted files. It keeps reading
paths into a buffer, until it has found a maximal range of
paths that can be merged. The merged path is written
into the new file for that chromosome, and the original
range of paths is removed from the buffer. The merging
step works in a similar way.
After creating the nodes V 00 of the maximally pruned
order-k 0 de Bruijn graph G00 = (V 00 , E 00 ), we build
the index. Sequences BWT and IN can be generated
from the predecessor sets v 00 .pred, while the outdegree
sequence OUT requires further processing. There is an
edge (u00 , v 00 ) ∈ E 00 if and only if string c · G00 .key(v 00 )
prefix-matches key G00 .key(u00 ) and c ∈ v 00 .pred. We
determine the edges and produce the outdegree sequence
by scanning the node file sequentially with σ + 1 pointers.
The pointer corresponding to node v 00 scans the entire
file, while each of the remaining σ pointers scans only the
range of nodes u00 with G00 .label(u00 ) = c for a character
c ∈ Σ. We also sample the nodes for locate queries
during the same scans.
Checking whether key G00 .key(u00 ) prefix-matches
string c · G00 .key(v 00 ) can be done by using the GCSA for
de Bruijn graph Gd = (Vd , Ed ) and the last character
array L. If the lexicographic rank of k-mer X is i, the
lexicographic rank of k-mer (cX)[0, k − 1] is Gd .LF(i, c).
If 2k-mer X is encoded with k-mer ranks (i, j), we can encode string cX as a lexicographic range of k-mer rank sequences, with (Gd .LF(i, c), Gd .LF(j, L[i]), Gd .LF(0, L[j]))
as the lower bound and Gd .LF(|Vd | − 1, L[j]) as the last
rank in the upper bound. String c · G00 .key(v 00 ) prefixmatches key G00 .key(u00 ) if and only if the lexicographic
ranges of the k-mer rank sequences overlap.
Complex regions of the variation graph must be
pruned before indexing. While this happens before index
construction begins, the construction algorithm has fea-
tures that can make the pruning less destructive. Pruning heuristics often create regions that are completely
missing from the index. The same mechanism that saves
memory by having each chromosome in a separate file
can be used to index overlapping subgraphs without indexing any paths between them. By having the pruned
graph in one file and the reference path in another file,
we can guarantee that no region is completely missing
from the index. We can also index selected paths in complex regions by duplicating nodes for prefix-doubling and
mapping the duplicates back to the original nodes during
the merging step. If we index the paths corresponding to
known haplotypes in complex regions, we can guarantee
that the index contains all observed variation.
D
Simplified GCSA Encoding
Let G0 = (V 0 , E 0 ) be a path graph with G0 .pred(v 0 , c) ≤ 1
for all nodes v 0 ∈ V 0 and characters c ∈ Σ. This is true
for de Bruijn graphs, and also for maximally pruned
de Bruijn graphs (Lemmas 3.6 and 3.3). We can use the
simplified encoding of the original GCSA [43] with such
path graphs.
We replace the sequences BWT and IN with indicator
bitvectors Bc [0, |V 0 | − 1] for all c ∈ Σ. If node v 0 ∈ V
with lexicographic rank i has a predecessor with label
c ∈ Σ, we set Bc [i] = 1. The backward step becomes:
spout = C[X[i]] + BX[i] .rank(spi+1 , 1);
epout = C[X[i]] + BX[i] .rank(epi+1 + 1, 1) − 1;
[spi , epi ] = [OUT.rank(spout , 1), OUT.rank(epout , 1)].
Two expensive queries (IN.select() and BWT.rank()) are
replaced with a cheap Bc .rank().
Computing G0 .LF(i) can expensive, as we have to
look at Bc [i] for all c ∈ Σ to determine the character
used in the backward step. If the alphabet is small,
this is still faster than the select queries in the general
encoding. We can further improve the time/space tradeoff by compressing the bitvectors Bc for rare characters
(e.g. N, #, and $) and checking Bc [i] first for the frequent
characters (e.g. bases) when computing G0 .LF(i).
E
Suffix Tree of a Path Graph
resented by the lexicographic range [sp, ep] matching the
path label `(v) from the root to the node. These ranges
can be determined from the LCP array. If we use the
FM-index with an LCP array supporting next/previous
smaller value queries and range minimum queries, we
can support the full functionality of the suffix tree [11].
If we build an x-ary tree over the LCP array, with
each internal node storing the minimum LCP value in
the corresponding range, we can support the required
queries in O(x logx |V 0 |) time with O(logx |V 0 |) random
x
memory accesses, while using x−1
|V 0 | log k bits of space.
Because the nodes of a path graph may match patterns
that do not prefix-match their keys, we have to be careful
with the suffix tree operations we use. By Lemma 3.2,
the ranges returned by find queries always correspond to
prefixes of the pattern. Hence we can safely use parent
queries with such ranges.4
BWA-MEM [21] aligns reads to a reference genome
by finding maximal exact matches between the read
and the reference. It indexes both the reference and its
reverse complement, allowing it to extend the pattern in
both directions. GCSA cannot use similar techniques, as
we cannot guarantee that the length of the lexicographic
range matching pattern X is the same as the length of
←
−
the range matching its reverse complement X (that the
←
−
key set contains key K if and only if it contains key K ).
However, we can search for maximal exact matches by
using LF-mapping and the parent operation [35].
In an ordinary FM-index, the length of a lexicographic range tells the number of distinct pointers in the
range. In GCSA, each node v 0 ∈ V 0 may have multiple
values (pointers), and a value may occur in multiple
nodes. As locate queries can be slow, we need another
way to support efficient counting queries. The problem
is similar to determining the document frequency of a
pattern. Given a collection of documents, the document
frequency of pattern X is the number of distinct documents that contain occurrences of the pattern. We can
use a bitvector of length 2n − d − 1, where n is the total
size of the document collection and d is the number of
documents, to compute the frequencies efficiently [41].
Let v be a suffix tree node corresponding to lexicographic range [sp, ep], and let count(v) be the document
frequency of the label `(v). If nodes v0 , . . . , vm−1 are the
children of node v,Pthe number of redundant documents
m−1
in them is R(v) = i=0 count(vi )−count(v). We create
an array R[0, n − 2] based on the inorder traversal of the
suffix tree. If the ith internal node we encounter is our
first visit to node v, we set R[i] = R(v). We set R[j] = 0
for any subsequent visits to the same node. Range
Let G0 = (V 0 , E 0 ) be an order-k path graph, and let
0
v00 , . . . , v|V
0 |−1 be its nodes in lexicographic order. The
LCP array of graph G0 is an array LCP[0, |V 0 |−1], where
LCP[i] is the length of the longest common prefix of keys
0
G0 .key(vi−1
) and G0 .key(vi0 ), with LCP[0] = 0. If we
build a trie of keys G0 .key(v 0 ) for v 0 ∈ V 0 and compact
the unary paths into single edges, the resulting tree can
be considered the suffix tree of graph G0 .
4 The shorter query in the variable-order de Bruijn graph [2] is
The LCP interval tree [1] is an alternate representation of the suffix tree. Each node v of the suffix tree is rep- essentially a parent query.
R[sp, ep − 1] covers the internal nodes in the subtree
with node v as the root. We can determine
P document
frequencies as count(v) = (ep + 1 − sp) − R[sp, ep − 1].
If we encode array R in unary, with value x becoming
x
0P
1, we get a bitvector BR , where we can compute sums
b
i=a R[i] with select queries as
(BR .select(b + 1, 1) − b) − (BR .select(a, 1) + 1 − a).
For value counting, we use array R to store the
number of redundant values as above. We use another
array A[0, |V 0 | − 1] to store the number of additional
values in each node vi0 ∈ V 0 as A[i] = |G0 .value(vi0 )| − 1,
and encode it as a bitvector BA in the same way as
array R above. The number of distinct values in range
[sp, ep] = G0 .find(X) for a pattern X is
G0 .count(X) = G0 .count(sp, ep)
=
ep
X
(A[i] + 1) −
i=sp
ep−1
X
R[i].
While graphs are a natural formalism for representing genetic variation, they cannot adequately represent
certain types of rearrangements. For example, if sequence S can occur in different positions of the genome
(e.g. ASBC and ABSC), we can either have a single
copy or multiple copies of S in the graph. Neither option
is good in a reference genome. With a single copy, we
create paths in the reference that do not correspond to
any valid genome. With the second option, we lose the
information that both copies of S are the same sequence.
One solution is to use context-free grammars. As long
as the grammar is non-nested, we can handle it with
the hypertext index. We build a hypertext index for
a high level graph, where each node is labeled with a
nonterminal symbol, while each nonterminal expands
into a subgraph indexed in GCSA.
References
i=sp
The bitvectors are often highly compressible [12], but
GCSA already uses one of the compression schemes
implicitly when it prunes the de Bruijn graph.
F Using the Hypertext Index
The hypertext index [44] is based on graphs G = (V, E),
where the label G.label(v) of a node v ∈ V is a string
over alphabet Σ. The labels are indexed in FM-index
F , while the reverse labels are indexed in FM-index
R. When we search for a pattern X ∈ Σ∗ of length
|X| ≥ 2, some of the matches may cross edges. In
order to find matches crossing one edge, we search for
suffixes X[i, |X| − 1] in the forward index F and the
reverses of prefixes X[0, i − 1] in the reverse index R,
for all 1 ≤ i ≤ |X| − 1. For each value of i, we combine
the partial matches into complete matches with a twodimensional range query in the edge matrix E, using the
lexicographic ranges for the reverse of X[0, i − 1] and for
X[i, |X| − 1] as the query ranges.
We often have to prune complex regions of the input
graph before indexing it. This causes false negatives:
paths that exist in the input graph but not in the index.
We can avoid the false negatives with a generalization of
the hypertext index. Instead of pruning the input graph
heuristically, we create a primary graph based on known
haplotypes and build a GCSA index for both strands
of the graph. We then create a matrix of additional
edges corresponding to potential recombinations in the
path graph, always crossing from the reverse complement
strand to the forward strand. We search for pattern X
←
−
and its reverse complement X in the index, and combine
←
−
the results find(X[i, |X|−1]) and find( X [|X|−i, |X|−1])
with a range query.
[1] M. I. Abouelhoda, S. Kurtz, and E. Ohlebusch. Replacing suffix trees with enhanced suffix arrays. Journal of
Discrete Algorithms, 2(1):53–86, 2004.
[2] C. Boucher et al. Variable-order de Bruijn graphs. In
Proc. 2015 IEEE Data Compression Conference (DCC
2015), pages 383–392. IEEE, 2015.
[3] A. Bowe, T. Onodera, K. Sadakane, and T. Shibuya.
Succinct de Bruijn graphs. In Proc. 12th Workshop
on Algorithms in Bioinformatics (WABI 2012), volume
7534 of LNCS, pages 225–235. Springer, 2012.
[4] M. Burrows and D. J. Wheeler. A block sorting lossless
data compression algorithm. Technical Report 124,
Digital Equipment Corporation, 1994.
[5] B. Cazaux, T. Lecroq, and E. Rivals. From indexing
data structures to de bruijn graphs. In Proc. 25th
Annual Symposium on Combinatorial Pattern Matching
(CPM 2014), volume 8486 of LNCS, pages 89–99.
Springer, 2014.
[6] D. M. Church et al. Extending reference assembly
models. Genome Biology, 16:13, 2015.
[7] A. Danek, S. Deorowicz, and S. Grabowski. Indexing
large genome collections on a PC. PLOS ONE,
9(10):e109384, 2014.
[8] A. Dilthey et al. Improved genome inference in the
MHC using a population reference graph. Nature
Genetics, 47(6):682–688, 2015.
[9] P. Ferragina, F. Luccio, G. Manzini, and S. Muthukrishnan. Compressing and indexing labeled trees, with
applications. Journal of the ACM, 57(1):4, 2009.
[10] P. Ferragina and G. Manzini. Indexing compressed text.
Journal of the ACM, 52(4):552–581, 2005.
[11] J. Fischer, V. Mäkinen, and G. Navarro. Faster entropybounded compressed suffix trees. Theoretical Computer
Science, 410(51):5354–5364, 2009.
[12] T. Gagie et al. Document counting in compressed space.
In Proc. 2015 IEEE Data Compression Conference
(DCC 2015), pages 103–112. IEEE, 2015.
[13] T. Gagie, G. Navarro, and S. J. Puglisi. New algorithms on wavelet trees and applications to information
retrieval. Theoretical Computer Science, 426-427:25–41,
2012.
[14] E. Garrison et al. vg: the variation graph toolkit, 2014–
2016. https://github.com/vgteam/vg.
[15] S. Gog, T. Beller, A. Moffat, and M. Petri. From theory
to practice: Plug and play with succinct data structures.
In Proc. 13th International Symposium on Experimental
Algorithms (SEA 2014), volume 8504 of LNCS, pages
326–337. Springer, 2014.
[16] R. Grossi, A. Gupta, and J. S. Vitter. High-order
entropy-compressed text indexes. In Proc. fourteenth
annual ACM-SIAM symposium on Discrete algorithms
(SODA 2003), pages 841–850. SIAM, 2003.
[17] L. Huang, V. Popic, and S. Batzoglou. Short read
alignment with populations of genomes. Bioinformatics,
29(13):i361–i370, 2013.
[18] S. Huang et al. Indexing similar DNA sequences. In
Proc. The Sixth International Conference on Algorithmic Aspects in Information and Management (AAIM
2010), volume 6124 of LNCS, pages 180–190. Springer,
2010.
[19] D. Kim et al. HISAT2, 2015–2016. https://ccb.jhu.
edu/software/hisat2/index.shtml.
[20] D. Kim, B. Langmead, and S. L. Salzberg. HISAT:
a fast spliced aligner with low memory requirements.
Nature Methods, 12(4):357–360, 2015.
[21] H. Li. Aligning sequence reads, clone sequences and
assembly contigs with BWA-MEM. arXiv:1303.3997,
2013.
[22] H. Li and R. Durbin. Fast and accurate short read alignment with Burrows-Wheeler transform. Bioinformatics,
25(14):1754–1760, 2009.
[23] A. Limasset, B. Cazaux, E. Rivals, and P. Peterlongo.
Read mapping on de bruijn graphs. BMC Bioinformatics, 17:237, 2016.
[24] Y. Lin and P. A. Pevzner. Manifold de Bruijn graphs. In
Proc. 14th Workshop on Algorithms in Bioinformatics
(WABI 2014), volume 8701 of LNCS, pages 296–310.
Springer, 2014.
[25] B. Liu, H. Guo, M. Brudno, and Y. Wang. deBGA:
read alignment with de Bruijn graph-based seed and
extension. Accepted to Bioinformatics, 2016.
[26] S. Maciuca, C. del Ojo Elias, G. McVean, and Z. Iqbal.
A natural encoding of genetic variation in a BurrowsWheeler transform to enable mapping and genome
inference. bioRxiv, 2016.
[27] U. Manber and G. Myers. Suffix arrays: A new method
for on-line string searches. SIAM Journal on Computing,
22(5):935–948, 1993.
[28] S. Marcus, H. Lee, and M. C. Schatz. SplitMEM: a
graphical algorithm for pan-genome analysis with suffix
skips. Bioinformatics, 30(24):3476–3483, 2014.
[29] T. Marschall et al. Computational pan-genomics:
Status, promises and challenges. BioRxiv, 2016.
[30] V. Mäkinen and G. Navarro. Succinct suffix arrays
based on run-length encoding. Nordic Journal of
Computing, 12(1):40–66, 2005.
[31] V. Mäkinen, G. Navarro, J. Sirén, and N. Välimäki.
Storage and retrieval of highly repetitive sequence collections. Journal of Computational Biology, 17(3):281–308,
2010.
[32] J. C. Na et al. FM-index of alignment: A compressed
index for similar strings. Theoretical Computer Science,
638:159–170, 2016.
[33] J. C. Na et al. FM-index of alignment with gaps.
arXiv:1606.03897, 2016.
[34] A. M. Novak, E. Garrison, and B. Paten. A graph
extension of the positional Burrows-Wheeler transform
and its applications. bioRxiv, 2016.
[35] E. Ohlebusch, S. Gog, and A. Kügel. Computing
matching statistics and maximal exact matches on
compressed full-text indexes. In Proc. 17th Symposium
on String Processing and Information Retrieval (SPIRE
2010), volume 6393 of LNCS, pages 347–358. Springer,
2010.
[36] D. Okanohara and K. Sadakane. Practical entropycompressed rank/select dictionary. In Proc. Ninth
Workshop on Algorithm Engineering and Experiments
(ALENEX 2007), pages 60–70. SIAM, 2007.
[37] B. Paten, A. Novak, and D. Haussler. Mapping to a
reference genome structure. arXiv:1404.5010, 2014.
[38] J. Pell et al. Scaling metagenome sequence assembly
with probabilistic de Bruijn graphs. Proc. National
Academy of Sciences of the United States of America,
109(33):13272–13277, 2012.
[39] E. A. Rødland. Compact representation of k-mer
de Bruijn graphs for genome read assembly. BMC
Bioinformatics, 14:313, 2013.
[40] K. Sadakane. Compressed suffix trees with full functionality. Theory of Computing Systems, 41(4):589–607,
2007.
[41] K. Sadakane. Succinct data structures for flexible
text retrieval systems. Journal of Discrete Algorithms,
5(1):12–22, 2007.
[42] K. Schneeberger et al. Simultaneous alignment of
short reads against multiple genomes. Genome Biology,
10(9):R98, 2009.
[43] J. Sirén, N. Välimäki, and V. Mäkinen. Indexing graphs
for path queries with applications in genome research.
IEEE/ACM Transactions on Computational Biology
and Bioinformatics, 11(2):375–388, 2014.
[44] C. Thachuk. Indexing hypertext. Journal of Discrete
Algorithms, 18:113–122, 2013.
[45] The 1000 Genomes Project Consortium. A global
reference for human genetic variation. Nature, 526:68–
64, 2015.
[46] S. Wandelt, J. Starlinger, M. Bux, and U. Leser. RCSI:
Scalable similarity search in thousand(s) of genomes.
Proc. VLDB Endowment, 6(13):1534–1545, 2013.
[47] P. Weiner. Linear pattern matching algorithms. In
Proc. 14th Annual IEEE Symposium on Switching and
Automata Theory (FOCS 1973), pages 1–11. IEEE,
1973.
| 8cs.DS
|
arXiv:1710.02034v2 [cs.IT] 16 Feb 2018
Simpler proof for nonlinearity of majority function
Thomas W. Cusick
∗
Department of Mathematics, University at Buffalo, Buffalo, NY
14260, USA
Abstract
Given a Boolean function f , the (Hamming) weight wt(f ) and the
nonlinearity N (f ) are well known to be important in designing functions that are useful in cryptography. The nonlinearity is expensive to
compute, in general, so any formulas giving the nonlinearity for particular functions f are significant. The well known majority function
has been extensively studied in a cryptographic context for the last
dozen years or so, and there is a formula for its nonlinearity. The
known proofs for this formula rely on many detailed results for the
Krawtchouk polynomials. This paper gives a much simpler proof.
Key words: Hamming weight, nonlinearity, Boolean functions, majority
function, Walsh transform.
Mathematics Subject Classifications (2010) 94C10 94A60 06E30
1
Introduction
Boolean functions have many applications, particularly in coding theory
and cryptography. A detailed account of the latter applications can be
found in the book [4]. If we define Vn to be the vector space of dimension
n over the finite field GF (2) = {0, 1}, then an n variable Boolean function
f (x1 , x2 , ..., xn ) = f (x) is a map from Vn to GF (2). Every Boolean function
f (x) has a unique polynomial representation (usually called the algebraic
normal form [4, p. 8]), and the degree of f is the degree of this polynomial.
A function of degree at most 1 is called affine, and if the constant term is
0 such a function is called linear. We let Bn denote the set of all Boolean
∗
Email: [email protected]
1
functions in n variables, with addition and multiplication done mod 2. If
we list the 2n elements of Vn as v0 = (0, ..., 0), v1 = (0, ..., 0, 1), ... in lexicographic order (we abbreviate this as lexico order below), then the 2n -vector
(f (v0 ), f (v1 ), . . . , f (v2n −1 )) is called the truth table of f. The weight (also
called Hamming weight) wt(f ) of f is defined to be the number of 1′ s in
the truth table for f . In many cryptographic uses of Boolean functions, it
is important that the truth table of each function f has an equal number of
0′ s and 1′ s; in that case, we say that the function f is balanced.
The distance d(f, g) between two Boolean functions f and g in the same
number of variables is defined by
d(f, g) = wt(f + g),
where the polynomial addition is done mod 2. An important concept in
cryptography is the nonlinearity N (f ) defined by
N (f ) = min d(f, a) .
a af f ine
In order for a Boolean function to be useful in a cryptographic application, it
is usually necessary that the function has high nonlinearity (see, for example,
[4, p. 122]). So-called Fourier analysis of Boolean functions (see [4, Chapter
2]) is very important in cryptography and other contexts. The efficient
computation of values of the nonlinearity is important here, and for this a
very important tool is the Walsh transform of a Boolean function fn in n
variables. The Walsh transform of fn is the map Wf : Vn → R defined by
X
(−1)f (x)+w·x ,
Wf (w) =
x∈Vn
where the values of f = fn are taken to be the real numbers 0 and 1. This
Walsh transform has also been important in physics and other sciences for
at least 40 years. In this paper we want the Walsh transform because of the
well known formula (see [4, Th. 2.21, p. 17], where fˆ is used instead of f )
N (fn ) = 2n−1 −
1
max |Wf (u)|.
2 u∈Vn
We shall use the obvious fact
Wf (0) = 2n − 2wt(f )
without comment in the rest of the paper.
2
(1)
2
Weight = nonlinearity if weight is small enough
It follows from the definitions that the nonlinearity is always ≤ the weight.
The following lemma gives a useful sufficient condition for the weight and
nonlinearity to be equal. This lemma is well known to experts, but I have
not found a statement of it in the literature. The proof below was shown
to me by Claude Carlet; it is much simpler than the proof using the Walsh
transform given in [3], which is a preliminary version of this paper. Before
[3] was written, this lemma was stated in a version of [7].
Lemma 1. Suppose f is a Boolean function in n variables with wt(f ) ≤
2n−2 . Then wt(f ) = N (f ).
Proof. We give a proof by contradiction. Suppose a is an affine function
such that
d(a, f ) < wt(f ) ≤ 2n−2 .
Then a is not constant and by the triangle inequality (0 is the zero function)
wt(a) = d(a, 0) ≤ d(a, f ) + d(f, 0) < 2n−2 + 2n−2 = 2n−1 .
This is a contradiction since the nonconstant affine function a is balanced,
and so has weight 2n−1 .
Lemma 1 is sharp in the sense that it is no longer true if we replace the
upper bound on wt(f ) by 2n−2 + 1. For example, the function with n = 3
having truth table (0, 0, 0, 1, 0, 0, 1, 1) (its algebraic normal form is x1 x2 x3 +
x1 x2 + x2 x3 ) has weight 3 and nonlinearity 1. However, sometimes we can
use Lemma 1 to evaluate the nonlinearity of functions gn in n variables even
if wt(gn ) 6= N (gn ), and even if wt(gn ) > 2n−2 . An example of this (Theorem
6) is given in the next section.
3
Finding nonlinearity of the majority function
The Boolean majority function Mk (x) can be defined in several slightly
different ways, but here we use the most common one, namely
Mk (x) = 1 if and only if wt(x) ≥ k/2.
The majority function seems to have been first mentioned in a cryptographic
context in the 1991 book [8, pp. 70-80], where only the case of k odd was
considered. The function has been of special interest in cryptography for
3
the last dozen years or so because this function and variants of it can be
proven to have optimal algebraic immunity. We do not need to explain this
technical concept here; interested readers can find a discussion of it in [4, pp.
174-176]. A useful recent survey of the applications of majority functions
to algebraic immunity is given in [2]. The papers [1, 6, 9, 11, 12, 13] and
others referenced in those papers all contain recent work on cryptographic
applications of these functions. It is necessary to know the nonlinearity of
the majority function to begin such studies.
A frequently quoted determination of the nonlinearity of the majority
function was given in [6, Th. 3, p. 52], using many detailed results on the
Krawtchouk polynomials (these date back to 1929; see [10, pp. 150-154] for
a study of these polynomials and proofs of some of their properties). It is
convenient to deal with the cases of odd and even k separately, so we define:
M2n+1 (x) = 1 if and only if wt(x) ≥ n + 1
(2)
M2n (x) = 1 if and only if wt(x) ≥ n.
(3)
and
For any k, we let M (k) denote the truth table of Mk (x). We let A(k) and
B(k) denote, respectively, the left and right halves of the truth table, so
M (k) = A(k)B(k)
(juxtaposition of the truth tables, each one being thought of simply as a
bitstring). We give a much simpler proof for the value of the nonlinearity of
the majority function in Theorem 5 below, but first we need three preliminary lemmas. We use the notation C(f ) for the complement of f, that is,
the function obtained by switching 0 to 1 and 1 to 0 for every entry in the
truth table of f. Similarly, if S is a bitstring, then C(S) is its complement.
Also, S ∗ denotes the bitstring S in reverse order.
Lemma 2. We have M (2n + 1) = C(M (2n))∗ M (2n) for n ≥ 2.
Proof. By (2), we can describe the two halves of M (2n + 1) by
A(2n + 1) = {0y : wt(y) ≥ n + 1, y in lexico order}
(4)
B(2n + 1) = {1y : wt(y) ≥ n, y in lexico order}.
(5)
and
Now (4) implies
A(2n + 1)∗ = {0y : wt(y) ≥ n + 1, y in reverse lexico order},
4
so (5) gives
B(2n + 1) = C(A(2n + 1))∗ .
(6)
It follows from (3) and (4) that
A(2n + 1) = {C(1y) : wt(y) ≤ n, y in lexico order}
= {C(y1) : wt(y) ≤ n, y in reverse lexico order}
= C(M (2n))∗
and then (6) gives B(2n + 1) = M (2n).
Below is an example, in which we use the notation ai for a string of i
symbols a = 0 or 1.
Example 1. M (5) = 07 103 1013 03 1013 017
Note that it follows from (4) and (6) that wt(M (2n + 1)) = 22n , so
M2n+1 (x) is a balanced function.
Lemma 3. For any truth table A, the nonlinearity of the function with truth
table A C(A)∗ is given by
N (A C(A)∗ ) = 2N (A).
Proof. For any affine function a we have
d(A, a) = d(C(A), C(a)) = d(C(A)∗ , C(a)∗ ).
Hence the lemma follows from the fact that aC(a)∗ is an affine function
whenever a is (see the well known Folklore Lemma [4, Lemma 2.2, p. 8]).
Lemma 4. We have wt(A(2n + 1)) = 22n−1 − 21 2n
n for n ≥ 2.
Proof. If we let b2n denote a bitstring of length 2n, then (4) gives
n
X
2n
2n
2n
wt(A(2n + 1)) = 2 − |{b2n : wt(b2n ) ≤ n}| = 2 −
j
j=0
and now the Binomial Theorem implies the formula in the lemma.
Our next example illustrates Lemmas 2 and 4.
Example 2.
M (6) = 07 103 1013 03 1013 017 03 1013 017 0115
A(7) = 015 1 07 103 1013 07 103 1013 03 1013 017
5
Now the nonlinearity formulas for Mk (x) are given in the following theorem.
Theorem 5. The nonlinearity of the majority function for n ≥ 2 is given
by
2n
2n
N (M (2n + 1)) = 2wt(A(2n + 1)) = 2 −
(7)
n
and
1
1 2n
N (M (2n)) = N (M (2n + 1)) = 22n−1 −
.
(8)
2
2 n
Proof. The value 22n − 2n
for M (2n + 1) (in a different notation) was
n
obtained via (1) in [8, Th. 4.20, p. 74]. The proof uses only elementary
properties of binomial coefficients but is long enough so that we will not
repeat it here. Now Lemma 4 completes the proof of (7). Finally, Lemmas
2 and 3 imply (8).
Note that Lemma 2 gives A(2n + 1) = C(M (2n))∗ (see Example 2) and
with (8) plus Lemma 4 we obtain
N (A(2n + 1)) = N (M (2n)) = wt(A(2n + 1),
which is a weight = nonlinearity result not implied by Lemma 1.
We can use Lemma 1 to obtain even more information about the majority
function, for example we can determine N (B(2n)), which is given in the next
theorem. The next example, which illustrates how the truth tables of M (i)
change as i increases, may be helpful in reading the proof.
Example 3.
M (2n − 1) = A(2n − 1) B(2n − 1)
M (2n)
= A(2n − 1) B(2n − 1) B(2n)
M (2n + 1) = Q1(2n + 1) M (2n − 1) M (2n − 1) B(2n)
Theorem 6. The nonlinearity of B(2n) for n ≥ 3 is equal to
∗
wt(C(B(2n)) ) = 2
2n−2
X
j=n+1
=
2n−2
X
j=n+1
2n − 2
2n − 2
+
j
n
2n − 2
1 2n − 2
2n−3
.
+2
−
2 n−1
j
6
(9)
(10)
Proof. Let Q1(2n+1) denote the first quarter of the truth table of M (2n+1).
It follows from Lemma 2 that
Q1(2n + 1) = C(B(2n))∗ .
(11)
If we let Q1A(2n + 1) and Q1B(2n + 1) denote, respectively, the left and
right halves of the truth table for Q1(2n + 1) and we let b2n−2 denote a
bitstring of length 2n − 2, then
Q1A(2n + 1) = {000b2n−2 : wt(b2n−2 ) ≥ n + 1, b2n−2 in lexico order}
and
Q1B(2n + 1) = {001b2n−2 : wt(b2n−2 ) ≥ n, b2n−2 in lexico order}.
Therefore
wt(Q1A(2n + 1)) =
2n−2
X
j=n+1
and
wt(Q1B(2n + 1)) =
2n−2
X
j=n
2n − 2
j
(12)
2n − 2
.
j
(13)
Now (9) follows from (11), (12) and (13), and (10) follows from (9) by
elementary properties of binomial coefficients.
Simple estimates using (10) show that wt(C(B(2n))∗ ) < 22n−2 . Hence
Lemma 1 gives wt(C(B(2n))∗ ) = N (C(B(2n))∗ ) = N (B(2n)).
The methods in Section 3 can be used to clarify various earlier results
in the literature. For example, it is easy to show that the functions φ2k
discussed in [5, p. 106] satisfy φ2k = M (2k).
References
[1] A. Braeken, B. Preneel, On the algebraic immunity of symmetric
Boolean functions, Indocrypt 2005 (Berlin: Springer LNCS 3797,
2005), 35–48. Extended version http://eprint.iacr.org/2005/245.
[2] C. Carlet: A survey on nonlinear Boolean functions with optimal
algebraic immunity suitable for stream ciphers, Vietnam J. Math.
41 (4) (2013), 527–541.
7
[3] T. W. Cusick, Weight=nonlinearity for all small weight Boolean
functions, https://arxiv.org/abs/1710.02034, 10 pp., 2017.
[4] Thomas W. Cusick, Pantelimon Stănică: Cryptographic Boolean
Functions and Applications, Second Edition, Elsevier-Academic
Press, San Diego, 2017.
[5] D. K. Dalai, K. C. Gupta, S. Maitra: Cryptographically significant
Boolean functions: construction and analysis in terms of algebraic
immunity, Fast Software Encryption (FSE 2005) (Berlin: Springer
LNCS 3557, 2005), 98–111.
[6] D. K. Dalai, S. Maitra, S. Sarkar: Basic theory in the construction
of Boolean functions with maximum possible algebraic immunity,
Designs, Codes Crypto. 40 (2006), 41–58.
[7] E. M. Sanger, Rotation Symmetric Boolean Functions, Ph. D. thesis,
University at Buffalo, 2017.
[8] C. Ding, G. Xiao, W. Shan: The Stability Theory of Stream Ciphers
(Berlin: Springer LNCS 561, 1991).
[9] S. Fu, C. Li, K. Matsuura, L. Qu: Construction of rotation symmetric Boolean functions with maximum algebraic immunity, Cryptology
and Network Security (CANS 2009) (Berlin: Springer LNCS 5888,
2009), 402–412.
[10] F. J. McWilliams and N. J. A. Sloane, The Theory of ErrorCorrecting codes, North Holland, Amsterdam, 1978.
[11] S. Sarkar, S. Maitra: Construction of rotation symmetric Boolean
functions with optimal algebraic immunity, Proceedings of Applied Algebra, Algebraic Algorithms and Error-Correcting Codes
(AAECC–17) (Berlin: Springer LNCS 4851, 2007), 271–280.
[12] S. Su: Construction of balanced even-variable Boolean functions on
odd number of variables with maximum algebraic immunity, Internat. J. Computer Math. 92 (2015), 2219–2232.
[13] S. Su, X. Tang: Construction of rotation symmetric Boolean functions with optimal algebraic immunity and high nonlinearity, Designs, Codes Crypto. 71 (2014), 183–199.
8
| 7cs.IT
|
Sentence Ordering and Coherence Modeling using Recurrent Neural Networks
Lajanugen Logeswaran1 , Honglak Lee1 , Dragomir Radev2
arXiv:1611.02654v2 [cs.CL] 22 Dec 2017
1
Department of Computer Science & Engineering, University of Michigan
2
Department of Computer Science, Yale University
[email protected], [email protected], [email protected]
Abstract
Modeling the structure of coherent texts is a key NLP problem.
The task of coherently organizing a given set of sentences
has been commonly used to build and evaluate models that
understand such structure. We propose an end-to-end unsupervised deep learning approach based on the set-to-sequence
framework to address this problem. Our model strongly outperforms prior methods in the order discrimination task and a
novel task of ordering abstracts from scientific articles. Furthermore, our work shows that useful text representations can
be obtained by learning to order sentences. Visualizing the
learned sentence representations shows that the model captures
high-level logical structure in paragraphs. Our representations
perform comparably to state-of-the-art pre-training methods
on sentence similarity and paraphrase detection tasks.
1
Introduction
Modeling the structure of coherent texts is an important problem in NLP. A well-written text has a particular high-level
logical and topical structure. The actual word and sentence
choices and their transitions come together to convey the
purpose of the text. Our primary goal is to build models that
can learn such structure by arranging a given set of sentences
to make coherent text.
Multi-document Summarization (MDS) and retrievalbased Question Answering (QA) involve extracting information from multiple documents and organizing it into a
coherent summary. Since the relative ordering of sentences
from different sources can be unclear, being able to automatically evaluate a particular order is essential. Barzilay and
Elhadad (2002) discuss the importance of an ordering component in MDS and show that finding acceptable orderings
can enhance user comprehension.
More importantly, by learning to order sentences we can
model text coherence. It is difficult to explicitly characterize
the properties of text that make it coherent. Ordering models
attempt to understand these properties by learning high-level
structure that causes sentences to appear in a specific order
in human-authored texts. Automatic methods for evaluating
human/machine generated text have great importance, with
applications in essay scoring (Miltsakaki and Kukich 2004;
Burstein, Tetreault, and Andreyev 2010) and text generation
Copyright c 2018, Association for the Advancement of Artificial
Intelligence (www.aaai.org). All rights reserved.
(Park and Kim 2015; Kiddon, Zettlemoyer, and Choi 2016).
Coherence models aid the better design of these systems.
Exploiting unlabelled corpora to learn semantic representations of data has become an active area of investigation.
Self-supervised learning is a typical approach that uses information naturally available as part of the data as supervisory
signals (Wang and Gupta 2015; Doersch, Gupta, and Efros
2015). Noroozi and Favaro (2016) attempt to learn visual
representations by solving image jigsaw puzzles. Sentence
ordering can be considered as a jigsaw puzzle in the language
domain and an interesting question is whether we can learn
useful textual representations by performing this task.
Our approach to coherence modeling is driven by recent
success in capturing semantics using distributed representations and modeling sequences using Recurrent Neural Nets
(RNN). RNNs are now the dominant approach to sequence
learning and mapping problems. The Sequence-to-sequence
(Seq2seq) framework (Sutskever, Vinyals, and Le 2014) and
its variants have fueled RNN based approaches to a range of
problems such as language modeling, text generation, MT,
QA, and many others.
In this work we propose an RNN-based approach to the
sentence ordering problem which exploits the set-to-sequence
framework of Vinyals, Bengio, and Kudlur (2015). A wordlevel RNN encoder produces sentence embeddings, and a
sentence-level set encoder RNN iteratively attends to these
embeddings and constructs a context representation. Initialized with this representation, a sentence-level pointer network
selects the sentences sequentially.
The most widely studied task relevant to sentence ordering and coherence modeling is the order discrimination task.
Given a document and a permuted version of it, the task involves identifying the more coherent ordering. Our proposed
model achieves state-of-the-art performance on two benchmark datasets for this task, outperforming several classical
approaches and recent data-driven approaches.
Addressing the more challenging task of ordering a given
collection of sentences, we consider the novel and interesting
task of ordering sentences from abstracts of scientific articles.
Our model strongly outperforms previous work on this task.
We visualize the learned sentence representations and show
that our model captures high-level discourse structure. We
provide visualizations that help understand what information
in the sentences the model uses to identify the next sentence.
Finally, we demonstrate that our ordering model learns
coherence properties and text representations that are useful
for several downstream tasks including summarization, sentence similarity and paraphrase detection. In summary, our
key contributions are as follows:
• We propose an end-to-end trainable model based on the
set-to-sequence framework to address the problem of coherently ordering a collection of sentences.
• We consider the novel task of understanding structure in
abstract paragraphs and demonstrate state-of-the-art results
in order discrimination and sentence ordering tasks.
• We show that our model learns sentence representations
that perform comparably to recent unsupervised pretraining methods on downstream tasks.
2
Related Work
Coherence modeling & sentence ordering. Coherence
modeling and sentence ordering have been approached by
closely related techniques. Many approaches propose a measure of coherence and formulate the ordering problem as
finding an order with maximal coherence. Recurring themes
from prior work include linguistic features, centering theory,
local and global coherence.
Local coherence has been modeled by considering properties of local windows of sentences such as sentence similarity
and transition structure. Lapata (2003) represent sentences by
vectors of linguistic features and learn the transition probabilities between features of adjacent sentences. The Entity-Grid
model (Barzilay and Lapata 2008) captures local coherence
by modeling patterns of entity distributions. Sentences are
represented by the syntactic roles of entities appearing in the
document. Features extracted from the entity grid are used to
train a ranking SVM. These two methods are motivated from
centering theory (Grosz, Weinstein, and Joshi 1995), which
states that nouns and entities in coherent discourses exhibit
certain patterns.
Global models of coherence typically use HMMs to model
document structure. The content model (Barzilay and Lee
2004) represents topics in a particular domain as states in an
HMM. State transitions capture possible presentation orderings within the domain. Words of a sentence are modeled
using a topic-specific language model. The content model
inspired several subsequent work to combine the strengths
of local and global models. Elsner, Austerweil, and Charniak (2007) combine the entity grid and the content model
using a non-parametric HMM. Soricut and Marcu (2006)
use several models as feature functions and define a loglinear model to assign probability to a text. Louis and
Nenkova (2012) model the intentional structure in documents
using syntax features.
Unlike previous approaches, we do not use any handcrafted
features and adopt an embedding-based approach. Local coherence is taken into account by a next-sentence prediction
component in our model, and global dependencies are naturally captured by an RNN. We demonstrate that our model
can capture both logical and topical structure by several evaluation benchmarks.
Data-driven approaches. Neural approaches have gained
attention recently. Li and Hovy (2014) model sentences as
embeddings derived from recurrent neural nets and train a
feed-forward neural network that takes an input window of
sentence embeddings and outputs a probability which represents the coherence of the sentence window. Coherence
evaluation is performed by sliding the window over the text
and aggregating the score. Li and Jurafsky (2016) study the
same model in a larger scale task and also use a sequence-tosequence approach in which the model is trained to generate
the next sentence given the current sentence and vice versa.
Nguyen and Joty (2017) learn to model coherence using a
convolutional network that operates on the Entity-Grid representation of an input document. These models are limited
by their local nature; our experiments show that considering
larger contexts is beneficial.
Hierarchical RNNs for document modeling. Word-level
and sentence-level RNNs have been used in a hierarchical
fashion for modeling documents in prior work. Li, Luong,
and Jurafsky (2015) proposed a hierarchical autoencoder for
generation and summarization applications. More relevant to
our work is a similar model considered by Lin et al. (2015).
A sentence-level RNN predicts the bag of words in the next
sentence given the previous sentences and a word-level RNN
predicts the word sequence conditioned on the sentence RNN
hidden state. Our model has a hierarchical structure similar
to these models, but takes a discriminative approach.
Combinatorial optimization with RNNs. Vinyals, Bengio, and Kudlur (2015) equip sequence-to-sequence models
with the ability to handle input and output sets, and discuss
experiments on sorting, language modeling and parsing. This
is called the read, process and write (or set-to-sequence)
model. The read block maps input tokens to a fixed length
vector representation. The process block is an RNN encoder
which, at each time-step, attends to the input token embeddings and computes an attention readout, appending it to the
current hidden state. The write block is an RNN which produces the target sequence conditioned on the representation
produced by the process block. Their goal is to show that
input and output orderings can matter in these tasks, which
is demonstrated using small scale experiments. Our work
exploits this framework to address the challenging problem
of modeling logical and hierarchical structure in text. Vinyals,
Fortunato, and Jaitly (2015) proposed pointer-networks for
combinatorial optimization problems where the output dictionary size depends on the number of input elements. We
use a pointer-network as the decoder to sequentially pick the
next sentence.
3
Approach
Our proposed model is inspired by the way a human would
solve this task. First, the model reads the sentences to capture
their meaning and the general context of the paragraph. Given
this knowledge, the model tries to pick the sentences one by
one sequentially till exhaustion.
Our model is based on the read, process and write framework of Vinyals, Bengio, and Kudlur (2015) briefly discussed
in the previous section. We use the encoder-decoder terminology that is more common in the following discussion.
1) This is preliminary information, ...
2) Any errors in this report will be ...
3) On March 7, 2000, at 1020 hrs …
…
𝑎1
𝑎2
LSTM
LSTM
𝑦ො1 ~ 𝑎1′
𝑦ො2 ~ 𝑎2′
𝑦ො𝑛 ~ 𝑎𝑛′
𝑎𝑚
𝑎1′
𝑎2′
𝑎𝑛′
LSTM
LSTM
LSTM
n) The aircraft sustained substantial ..
<start>
…
𝑎1𝑖 𝑠𝑖
𝑖
𝑎𝑚−1
𝑠𝑖 <start>
…
LSTM
𝑠𝑦ො1
𝑠𝑦ො𝑛−1
Sentence Encoder
Pointer Network Decoder
Set Encoder
Figure 1: Model Overview: The input set of sentences are represented as vectors using a sentence encoder. At each time step of
the model, attention weights are computed for the sentence embeddings based on the current hidden state. The encoder uses the
attention probabilities to compute the input for the next time-step and the decoder uses them for prediction.
The model is comprised of a sentence encoder RNN, an
encoder RNN and a decoder RNN (Fig. 1). The sentence
encoder takes as input the words of a sentence s sequentially
and computes an embedding representation of the sentence.
Henceforth, we use s to refer to a sentence or its embedding
interchangeably. The embeddings {s1 , s2 , ..., sn } of a given
set of n sentences constitute the sentence memory, available
to be accessed by subsequent components.
The encoder LSTM is identical to the originally proposed
process block, defined by Eqs 1-4. At each time step the input
to the LSTM is computed by taking a weighted sum over the
memory elements, the weights being attention probabilities
obtained using the previous hidden state as query (Eqs. 1,
2). This is iterated for a fixed number of times called the
read cycles. Intuitively, the model identifies a soft input order
to read the sentences. As described in Vinyals, Bengio, and
Kudlur (2015) the encoder has the desirable property of being invariant to the order in which the sentence embeddings
reside in the memory.
t
et,i
enc = f (si , henc ); i ∈ {1, ..., n}
(1)
atenc
(2)
=
statt =
t+1
ht+1
enc , cenc
=
Softmax(etenc )
n
X
at,i
enc si
i=1
LSTM(htenc , ctenc , statt )
t−1
t−1
htdec , ctdec = LSTM(ht−1
)
dec , cdec , x
et,i
dec
atdec
=
=
f (si , htdec ); i ∈
Softmax(etdec )
{1, ..., n}
(4)
The decoder is a pointer network that takes a similar form
with a few differences (Eqs. 5-7). The LSTM takes the embedding of the previous sentence as input instead of the attention readout. At training time the correct order of sentences
(so1 , so2 , ..., son ) = (x1 , x2 , ..., xn ) is known (o represents
the correct order) and xt−1 is used as the input. At test time
the predicted assignment x̂t−1 is used instead. The attention
computation is identical to that of the encoder, but now at,i
dec is
interpreted as the probability for si being the correct sentence
choice at position t, conditioned on the previous sentence
assignments p(St = si |S1 , ..., St−1 ). The initial state of the
decoder LSTM is initialized with the final hidden state of the
encoder. x0 is a vector of zeros.
(6)
(7)
Scoring Function. We consider two choices for the scoring
function f in Eqs. 1, 6. The first (Eq. 8) is a single hidden
layer feed-forward net that takes s, h as inputs (W, b, W 0 , b0
are learnable parameters). The structure of f is similar to the
window network of Li and Hovy (2014). While they used a
local window of sentences to capture context, this scoring
function exploits the entire history of sentences encoded
in the RNN hidden state to score candidates for the next
sentence.
f (s, h) = W 0 tanh(W [s ; h] + b) + b0
(8)
We also consider a bilinear scoring function (Eq. 9). Compared to the previous scoring function, this takes a generative
approach to regress the next sentence given the current hidden state (W h + b), enforcing that it be most similar to the
correct next sentence. We observed that this scoring function
led to better sentence representations (Sec. 4.4).
f (s, h) = sT (W h + b)
(3)
(5)
(9)
Contrastive Sentences. In its vanilla form, we found that
the set-to-sequence model tends to rely on certain word clues
to perform the ordering task. To encourage holistic sentence
understanding, we add a random set of sentences to the sentence memory when the decoder makes classification decisions. This makes the problem more challenging for the
decoder since now it has to distinguish between sentences
that are relevant and irrelevant to the current context in identifying the correct sentence.
Coherence modeling. We define the coherence score of
an arbitrary partial/complete assignment (sp1 , ..., spk ) to the
first k sentence positions as
Pk
(10)
i=1 log p(Si = spi |S1,...,i−1 = sp1 ,...,pi−1 )
where S1 , .., Sk are random variables representing the sentence assignment to positions 1 through k. The conditional
Table 1: Mean Accuracy comparison on the Accidents and Earthquakes data for the order discrimination task. The reference
models are Entity-Grid (Barzilay and Lapata 2008), HMM (Louis and Nenkova 2012), Graph (Guinaudeau and Strube 2013),
Window network (Li and Hovy 2014) and sequence-to-sequence (Li and Jurafsky 2016), respectively.
Entity-Grid
HMM
Graph
0.904
0.872
0.842
0.957
0.846
0.635
Accidents
Earthquakes
probabilities are derived from the network. This is our measure of comparing the coherence of different renderings of a
document. It is also used as a heuristic during decoding.
Training Objective. The model is trained using the maximum likelihood objective
P
P|x|
max x∈D t=1 log p(xt |x1 , ..., xt−1 )
(11)
where D denotes the training set and each training instance is given by an ordered document of sentences x =
(x1 , ..., x|x| ).
4
Experimental Results
We first consider the order discrmination task that has been
widely used in the literature for evaluating coherence models.
We then consider the more challenging ordering problem
where a coherent order of a given collection of sentences
needs to be determined. We then demonstrate that our ordering model learns coherence properties useful for summarization. Finally, we show that our model learns sentence
representations that are useful for downstream applications.
For all tasks discussed in this section we train the model
with the maximum likelihood objective on the training data
relevant to the task. We used the single hidden layer MLP
scoring function for the order discrimination and sentence
ordering tasks. Models are trained end-to-end. We use pretrained 300 dimensional GloVe word embeddings (Pennington, Socher, and Manning 2014) to initialize word vectors.
All LSTMs use a hidden layer size of 1000 and the MLP in
Eq. 8 has a hidden layer size of 500. The number of read cycles in the encoder is set to 10. The same model architecture
is used across all experiments. We used the Adam optimizer
(Kingma and Ba 2014) with batch size 10 and learning rate
5e-4 for learning. The model is regularized using early stopping. Hyperparameters were chosen using the validation set.
4.1
Order Discrimination
The ordering problem is traditionally formulated as a binary
classification task: Given a reference paragraph and its permuted version, identify the more coherent one (Barzilay and
Lapata 2008).
The datasets widely used for this task in previous work
are the Accidents and Earthquakes news reports. In each of
these datasets the training and test sets include 100 articles
and approximately 20 permutations of each article.
In Table 1 we compare our results with traditional approaches and recent data-driven approaches. The entity grid
model provides a strong baseline on the Accidents dataset,
Window
(Recurrent) (Recursive)
0.840
0.864
0.951
0.976
Seq2seq
Ours
0.930
0.992
0.944
0.997
only outperformed by our model and Li and Jurafsky (2016).
On the Earthquakes data the window approach of Li and Jurafsky (2016) performs strongly. Our approach outperforms
prior models on both datasets, achieving near perfect performance on the Earthquakes dataset.
While these datasets have been widely used, they are quite
formulaic in nature and are no longer challenging. We hence
turn to the more challenging task of ordering a given collection of sentences to make a coherent document.
4.2
Sentence Ordering
In this task we directly address the ordering problem. We do
not assume the availability of a set of candidate orderings
to choose from and instead find a good ordering from all
possible permutations of the sentences.
The difficulty of the ordering problem depends on the nature of the text, as well as the length of paragraphs considered.
Evaluation on text from arbitrary text sources makes it difficult to interpret the results, since it may not be clear whether
to attribute the observed performance to a deficient model
or ambiguity in next sentence choices due to many plausible
orderings.
Text summaries are a suitable source of data for this task.
They often exhibit a clear flow of ideas and have little redundancy. We specifically look at abstracts of conference papers
and research proposals. This data has several favorable properties. Abstracts usually have a particular high-level format
- They begin with a brief introduction, a description of the
problem and proposed approach and conclude with performance remarks. This would allow us to identify if the model
can capture high-level logical structure. Second, abstracts
have an average length of about 10, making the ordering task
more accessible. This also gives us a significant amount of
data to train and test our models.
We use the following sources of abstracts for this task.
• NIPS Abstracts. We consider abstracts from NIPS papers
in the past 10 years. We parsed 3280 abstracts from paper
pdfs and obtained 3259 abstracts after omitting erroneous
extracts. The dataset was split into years 2005-2013 for
training and 2014, 2015 respectively for validation, testing.
• ACL Abstracts. A second source of abstracts are papers
from the ACL Anthology Network (AAN) corpus (Radev
et al. 2009). We extracted 12,157 abstracts from the text
parses using simple keyword matching for the strings ‘Abstract’ and ‘Introduction’. We use all extracts of papers
published up to year 2010 for training, year 2011 for validation and years 2012-2013 for testing.
Table 2: Comparison against prior methods on the abstracts data.
NIPS Abstracts
NSF Abstracts
Accuracy
τ
Accuracy
τ
Accuracy
τ
Random
15.59
0
19.36
0
9.46
0
Entity Grid (Barzilay and Lapata 2008)
Seq2seq (Uni) (Li and Jurafsky 2016)
Window network (Li and Hovy 2014)
RNN Decoder
Proposed model
20.10
27.18
41.76
48.22
51.55
0.09
0.27
0.59
0.67
0.72
21.82
36.62
50.87
52.06
58.06
0.10
0.40
0.65
0.66
0.73
13.68
18.67
25.79
28.33
0.10
0.28
0.48
0.51
• NSF Abstracts. We also used the NSF Research Award
Abstracts dataset (Lichman 2013). It comprises abstracts
from a diverse set of scientific areas in contrast to the previous two sources of data and the abstracts are also lengthier,
making this dataset more challenging. Years 1990-1999
were used for training, 2000 for validation and 2001-2003
for testing. We capped the parses of the abstracts to a maximum length of 40 sentences. Unsuccessful parses and
parses of excessive length were discarded.
Further details about the data are provided in the supplement.
The following metrics are used to evaluate performance.
Accuracy measures how often the absolute position of a sentence was correctly
predicted. Kendall’s tau (τ ) is computed
as 1 − 2 · N/ n2 , where N is the number of pairs in the
predicted sequence with incorrect relative order and n is the
sequence length. Lapata (2006) discusses that this metric
reliably correlates with human judgements.
The following baselines are used for comparison:
• Entity Grid. Our first baseline is the Entity Grid model
of Barzilay and Lapata (2008). We use the Stanford
parser (Klein and Manning 2003) and Brown Coherence
Toolkit1 to derive Entity grid representations. A ranking
SVM is trained to score correct orderings higher than incorrect orderings as in the original work. We used 20 permutations per document as training data. Since the entity
grid only provides a means of feature extraction we evaluate the model in the ordering setting as follows. We choose
1000 random permutations for each document, one of them
being the correct order, and pick the order with maximum
coherence. We experimented with transitions of length at
most 3 in the entity-grid.
• Seq2seq. The second baseline we consider is a sequenceto-sequence model which is trained to predict the next
sentence given the current sentence. Li and Jurafsky (2016)
consider similar methods and our model is the same as their
uni-directional model. These methods were shown to yield
sentence embeddings that have competitive performance
in several semantic tasks in Kiros et al. (2015).
• Window Network. We consider the window approach of
Li and Hovy (2014) and Li and Jurafsky (2016) which
demonstrated strong performance in the order discrimination task as our third baseline. We adopt the same coherence score interpretation considered by the authors. In
1
AAN Abstracts
bitbucket.org/melsner/browncoherence
both the above models we consider a special embedding
vector which is padded at the beginning of a paragraph and
learned during training. This vector allows us to identify
the initial few sentences during greedy decoding.
• RNN Decoder. Another baseline is our proposed model
without the encoder. The decoder hidden state is initialized
with zeros. We observed that using a special start symbol
as for the other baselines helped obtain better performance
with this model. However, a start symbol did not help when
the model is equipped with an encoder as the hidden state
initialization alone was good enough.
We do not place emphasis on the particular search algorithm in this work and thus use beam search with the coherence score heuristic for all models. A beam size of 100 was
used. During decoding, sentence candidates that have been
already chosen are pruned from the beam. All RNNs use a
hidden layer size of 1000. For the window network we used
a window size of 3 and a hidden layer size of 2000. We initialize all models with pre-trained GloVe word embeddings.
We assess the performance of our model against baseline
methods in Table 2. The window network performs strongly
compared to the other baselines. Our model does better by a
significant margin by exploiting global context, demonstrating that global context is important in this task.
While the Entity-Grid model has been fairly successful for
the order discrimination task in the past we observe that it
fails to discriminate between a large number of candidates.
One reason could be that the feature representation is less
sensitive to local changes in sentence order (such as swapping
adjacent sentences). The computational expense of obtaining parse trees and constructing grids on a large amount of
data prohibited experimenting with this model on the NSF
abstracts data.
The Seq2seq model performs worse than the window network. Interestingly, Li and Jurafsky (2016) observe that the
Seq2seq model outperforms the window network in an order discrimination task on Wikipedia data. However, the
Wikipedia data considered in their work is an order of magnitude larger than the datasets considered here, and that could
have potentially helped the generative model. These models are also expensive during inference since they involve
computing and sampling from word distributions.
Fig. 2 shows t-SNE embeddings of sentence representations learned by our sentence encoder. These are sentences
from test sets, color coded by their positions in the source
First
Sentence
(a) NIPS Abstracts
(b) AAN Abstracts
Last
Sentence
(c) NSF Abstracts
Figure 2: t-SNE embeddings of representations learned by the model for sentences from the test set. Embeddings are color coded
by the position of the sentence in the document it appears.
Table 3: Comparison on extractive summarization Table 4: Performance comparison for semantic similarity and paraphrase
between models trained from scratch and models detection. The first row shows the best performing purely supervised methpre-trained with the ordering task.
ods. The last section shows our models.
Model
From scratch
Pre-train ord.
From scratch
Pre-train ord.
ROUGE-1
ROUGE-2
ROUGE-L
Summary length = 75b
18.29
47.56
12.79
18.77
50.32
13.25
Summary length = 275b
35.82
10.67
33.69
36.47
10.99
34.27
Model
Supervised
Uni-ST (Kiros et al. 2015)
Ordering model
+ BoW
+ Uni-ST
abstract. This shows that our model learns high-level structure in the documents, generalizing well to unseen text. The
structure is less apparent in the NSF dataset due to its data
diversity and longer documents. While approaches based on
the Barzilay and Lee (2004) model explicitly capture topics
by discovering clusters in sentences, our neural approach
implicitly discovers such structure.
4.3
Sentence Ordering and Summarization
In this section we show that sentence ordering models learn
coherence properties useful for summarization. We consider
a variation of our model where the model takes a set of sentences from several documents as input and sequentially picks
summary sentences until it predicts a special ‘stop’ symbol.
A key distinction between this model and recent work (Cheng
and Lapata 2016; Nallapati, Zhou, and Ma 2016) is that the
input order of sentences is assumed to be unknown, making
it applicable to multi-document summarization.
We train a model from scratch to perform extractive summarization in the above fashion. We then consider a model
that is pre-trained on the ordering task and is fine-tuned on
the above task. The DailyMail and CNN datasets (Cheng
and Lapata 2016) were used for experimentation. We use
DailyMail for pre-training purposes and CNN for fine-tuning
and evaluation. The labels in DailyMail are not used. We
compare ROUGE scores of the two models in Table 3 under
standard evaluation settings.
We observe that the model pre-trained with the ordering
task scores consistently better than the model trained from
scratch. The results can be improved further by using larger
news corpora. This shows that sentence ordering is an attractive unsupervised objective for exploiting large unlabelled
r
0.868
0.848
0.807
0.842
0.860
SICK
ρ
0.808
0.778
0.742
0.775
0.795
MSE
0.253
0.287
0.356
0.299
0.270
MSRP
(Acc) (F1)
80.4 86.0
73.0 81.9
72.3 81.1
74.0 81.9
74.9 82.5
corpora to improve summarization systems. It further shows
that the coherence scores obtained from the ordering model
correlates well with summary quality.
4.4
Learned Sentence Representations
One of the original motivations for this work is the question
of whether we can learn high-quality sentence representations
by learning to model text coherence. To address this question
we trained our model on a large number of paragraphs using
the BookCorpus dataset (Kiros et al. 2015).
To evaluate the quality of sentence embeddings derived
from the model, we use the evaluation pipeline of Kiros
et al. (2015) for tasks that involve understanding sentence
semantics. These evaluations are performed by training a
classifier on top of the embeddings derived from the model
(holding the embeddings fixed) so that the performance is indicative of the quality of sentence representations. We present
a comparison for the semantic relatedness and paraphrase
detection tasks in Table 4. Results for only uni-directional
versions of models are discussed here for a fair comparison.
Similar to the skip-thought (ST) paper, we train two models
- one predicting the correct order in the forward direction
and another in the backward direction. The numbers shown
for the ordering model were obtained by concatenating the
representations from the two models.
Concatenating the above representation with the bag-ofwords representation (using the fine-tuned word embeddings)
of the sentence further improves performance. This is because the ordering model can choose to pay less attention to
specific lexical information and focus on high-level document
structure. Hence, the two representations capture complementary semantics. Adding ST features improves performance
Table 5: Visualizing salient words (Abstracts are from the AAN corpus).
In this paper , we propose a new method for semantic class induction .
First , we introduce a generative model of sentences , based on dependency trees and which takes into account homonymy .
Our model can thus be seen as a generalization of Brown clustering .
Second , we describe an efficient algorithm to perform inference and learning in this model .
Third , we apply our proposed method on two large datasets ( 108 tokens , 105 words types ) , and demonstrate that classes
induced by our algorithm improve performance over Brown clustering on the task of semisupervised supersense tagging and
named entity recognition .
Representation learning is a promising technique for discovering features that allow supervised classifiers to generalize from
a source domain dataset to arbitrary new domains .
We present a novel , formal statement of the representation learning task .
We argue that because the task is computationally intractable in general , it is important for a representation learner to be able
to incorporate expert knowledge during its search for helpful features .
Leveraging the Posterior Regularization framework , we develop an architecture for incorporating biases into representation
learning .
We investigate three types of biases , and experiments on two domain adaptation tasks show that our biased learners identify
significantly better sets of features than unbiased learners , resulting in a relative reduction in error of more than 16% for
both tasks , with respect to state-of-the-art representation learning techniques.
further. We observed that the bilinear scoring function and
introducing contrastive sentences to the decoder improved
the quality of learned representations significantly.
Our model has several key advantages over ST. ST has a
word-level reconstruction objective and is trained with large
softmax output layers. This limits the vocabulary size and
slows down training (they use a vocabulary size of 20k and report two weeks of training). Our model achieves comparable
performance and does not have such a word reconstruction
component. We train with a vocabulary of 400k words; the
above results are based on a training time of two days on a
GTX Titan X GPU.
4.5
Word Influence
We attempt to understand what text-level clues the model uses
to perform the ordering task. Inspired by Li et al. (2015), we
use gradients of prediction decisions with respect to words of
the correct sentence as a proxy for the salience of each word.
We feed sentences to the decoder in the correct order and at
each time step compute the derivative of the score e (Eq. 6)
of the correct next sentence s = (w1 , .., wn ) with respect to
its word embeddings. The importance of word wi in correctly
∂e
predicting s as the next sentence is defined as k ∂w
k. We
i
assume the hidden states of the decoder to be fixed and only
back-propagate gradients through the sentence encoder.
Table 5 shows visualizations of two abstracts. Darker
shades correspond to higher gradient norms. In the first example the model appears to be using the word clues ‘first’,
‘second’ and ‘third’. A similar observation was made by Chen,
Qiu, and Huang (2016). In the second example we observe
that the model pays attention to phrases such as ‘We present’,
‘We argue’, which are typical of abstract texts. It also focuses
on the word ‘representation’ appearing in the first two sentences. These observations link to centering theory which
states that entity distributions in coherent discourses exhibit
certain patterns. The model implicitly learns these patterns
with no syntax annotations or handcrafted features.
5
Conclusion
This work investigated the challenging problem of coherently organizing a set of sentences. Our RNN-based model
performs strongly compared to baselines and prior work on
sentence ordering and order discrimination tasks. We further
demonstrated that it captures high-level document structure
and learns useful sentence representations when trained on
large amounts of data. Our approach to the ordering problem
deviates from most prior work that use handcrafted features.
However, exploiting linguistic features for next sentence classification can potentially further improve performance. Entity
distribution patterns can provide useful features about named
entities that are treated as out-of-vocabulary words. The ordering problem can be further studied at higher-level discourse
units such as paragraphs, sections and chapters.
6
Acknowledgments
This material is based in part upon work supported by IBM
under contract 4915012629. Any opinions, findings, conclusions or recommendations expressed above are those of
the authors and do not necessarily reflect the views of IBM.
We thank the UMich/IBM Sapphire team and Junhyuk Oh,
Ruben Villegas, Xinchen Yan, Rui Zhang, Kibok Lee and
Yuting Zhang for helpful comments and discussions.
References
Barzilay, R., and Elhadad, N. 2002. Inferring strategies for
sentence ordering in multidocument news summarization.
Journal of Artificial Intelligence Research 35–55.
Barzilay, R., and Lapata, M. 2008. Modeling local coherence: An entity-based approach. Computational Linguistics
34(1):1–34.
Barzilay, R., and Lee, L. 2004. Catching the drift: Probabilistic content models, with applications to generation and
summarization. arXiv preprint cs/0405039.
Burstein, J.; Tetreault, J.; and Andreyev, S. 2010. Using
entity-based features to model coherence in student essays.
In Human language technologies: The 2010 annual conference of the North American chapter of the Association for
Computational Linguistics, 681–684. Association for Computational Linguistics.
Chen, X.; Qiu, X.; and Huang, X. 2016. Neural sentence
ordering. arXiv preprint arXiv:1607.06952.
Cheng, J., and Lapata, M. 2016. Neural summarization by extracting sentences and words. arXiv preprint
arXiv:1603.07252.
Doersch, C.; Gupta, A.; and Efros, A. A. 2015. Unsupervised
visual representation learning by context prediction. In Proceedings of the IEEE International Conference on Computer
Vision, 1422–1430.
Elsner, M.; Austerweil, J. L.; and Charniak, E. 2007. A
unified local and global model for discourse coherence. In
HLT-NAACL, 436–443.
Grosz, B. J.; Weinstein, S.; and Joshi, A. K. 1995. Centering:
A framework for modeling the local coherence of discourse.
Computational linguistics 21(2):203–225.
Guinaudeau, C., and Strube, M. 2013. Graph-based local
coherence modeling. In ACL (1), 93–103.
Kiddon, C.; Zettlemoyer, L.; and Choi, Y. 2016. Globally
coherent text generation with neural checklist models. In
Proceedings of the 2016 Conference on Empirical Methods
in Natural Language Processing (EMNLP).
Kingma, D., and Ba, J. 2014. Adam: A method for stochastic
optimization. arXiv preprint arXiv:1412.6980.
Kiros, R.; Zhu, Y.; Salakhutdinov, R. R.; Zemel, R.; Urtasun,
R.; Torralba, A.; and Fidler, S. 2015. Skip-thought vectors. In
Advances in Neural Information Processing Systems, 3276–
3284.
Klein, D., and Manning, C. D. 2003. Accurate unlexicalized
parsing. In Proceedings of the 41st Annual Meeting on Association for Computational Linguistics-Volume 1, 423–430.
Association for Computational Linguistics.
Lapata, M. 2003. Probabilistic text structuring: Experiments with sentence ordering. In Proceedings of the 41st Annual Meeting on Association for Computational LinguisticsVolume 1, 545–552. Association for Computational Linguistics.
Lapata, M. 2006. Automatic evaluation of information ordering: Kendall’s tau. Computational Linguistics 32(4):471–
484.
Li, J., and Hovy, E. H. 2014. A model of coherence based on
distributed sentence representation. In EMNLP, 2039–2048.
Li, J., and Jurafsky, D. 2016. Neural net models
for open-domain discourse coherence. arXiv preprint
arXiv:1606.01545.
Li, J.; Chen, X.; Hovy, E.; and Jurafsky, D. 2015. Visualizing and understanding neural models in nlp. arXiv preprint
arXiv:1506.01066.
Li, J.; Luong, M.-T.; and Jurafsky, D. 2015. A hierarchical
neural autoencoder for paragraphs and documents. arXiv
preprint arXiv:1506.01057.
Lichman, M. 2013. UCI machine learning repository.
Lin, R.; Liu, S.; Yang, M.; Li, M.; Zhou, M.; and Li, S. 2015.
Hierarchical recurrent neural network for document modeling. In Proceedings of the 2015 Conference on Empirical
Methods in Natural Language Processing, 899–907.
Louis, A., and Nenkova, A. 2012. A coherence model
based on syntactic patterns. In Proceedings of the 2012
Joint Conference on Empirical Methods in Natural Language
Processing and Computational Natural Language Learning,
1157–1168. Association for Computational Linguistics.
Miltsakaki, E., and Kukich, K. 2004. Evaluation of text
coherence for electronic essay scoring systems. Natural
Language Engineering 10(01):25–55.
Nallapati, R.; Zhou, B.; and Ma, M. 2016. Classify or select:
Neural architectures for extractive document summarization.
arXiv preprint arXiv:1611.04244.
Nguyen, D. T., and Joty, S. 2017. A neural local coherence
model. In Proceedings of the 55th Annual Meeting of the
Association for Computational Linguistics (Volume 1: Long
Papers), volume 1, 1320–1330.
Noroozi, M., and Favaro, P. 2016. Unsupervised learning of
visual representations by solving jigsaw puzzles. In European
Conference on Computer Vision, 69–84. Springer.
Park, C. C., and Kim, G. 2015. Expressing an image stream
with a sequence of natural sentences. In Advances in Neural
Information Processing Systems, 73–81.
Pennington, J.; Socher, R.; and Manning, C. D. 2014. Glove:
Global vectors for word representation. In EMNLP, volume 14, 1532–43.
Radev, D. R.; Joseph, M. T.; Gibson, B.; and Muthukrishnan,
P. 2009. A Bibliometric and Network Analysis of the field of
Computational Linguistics. Journal of the American Society
for Information Science and Technology.
Soricut, R., and Marcu, D. 2006. Discourse generation
using utility-trained coherence models. In Proceedings of the
COLING/ACL on Main conference poster sessions, 803–810.
Association for Computational Linguistics.
Sutskever, I.; Vinyals, O.; and Le, Q. V. 2014. Sequence
to sequence learning with neural networks. In Advances in
neural information processing systems, 3104–3112.
Vinyals, O.; Bengio, S.; and Kudlur, M. 2015. Order
matters: Sequence to sequence for sets. arXiv preprint
arXiv:1511.06391.
Vinyals, O.; Fortunato, M.; and Jaitly, N. 2015. Pointer
networks. In Advances in Neural Information Processing
Systems, 2674–2682.
Wang, X., and Gupta, A. 2015. Unsupervised learning of
visual representations using videos. In Proceedings of the
IEEE International Conference on Computer Vision, 2794–
2802.
| 2cs.AI
|
Fine-grained Event Learning of Human-Object
Interaction with LSTM-CRF
arXiv:1710.00262v1 [cs.CV] 30 Sep 2017
Tuan Do and James Pustejovsky
∗
Brandeis University - Department of Computer Science
Waltham, Massachusetts - United States of America
Abstract.
Event learning is one of the most important problems in AI. However, notwithstanding significant research efforts, it is still a very complex task, especially
when the events involve the interaction of humans or agents with other objects,
as it requires modeling human kinematics and object movements. This study
proposes a methodology for learning complex human-object interaction (HOI)
events, involving the recording, annotation and classification of event interactions. For annotation, we allow multiple interpretations of a motion capture
by slicing over its temporal span; for classification, we use Long-Short Term
Memory (LSTM) sequential models with Conditional Randon Field (CRF) for
constraints of outputs. Using a setup involving captures of human-object interaction as three dimensional inputs, we argue that this approach could be
used for event types involving complex spatio-temporal dynamics.
1
Introduction
The study of events has long involved many disciplines, including philosophy, cognitive psychology, linguistics, computer science, and AI. The Gestalt school of philosophy characterized events as whole processes that emerge from the relations between
their components. Cognitive psychologists, such as Tulving [1], recognized the importance of events by postulating a separate cognitive process called episodic memory. The representation of events in natural language has been studied from many
different approaches, from formal logic and AI [2], and frames [3], to computational
linguistics [4]. Combining perspectives from computer science, logic, and linguistics,
some recent work suggests that events can be effectively modeled as programs within
a dynamic logic (DITL) [5], enabling computer simulations of linguistic expressions
[6].
In computer science, there is little consensus about how events should be modeled
for learning. They can be represented atomically, i.e., entire events are predicted
in a classification manner [7], or as combinations of more primitive actions [8], i.e.,
complex event types are learned based on recognition of combined primitive actions.
For the former type of event representation, there are quantitative approaches based
on low-level pixel features such as in [9] and qualitative approaches such as induction
from relational states among event participants [10]. For the latter approach, systems such as [11], use state transition graphical models such as Dynamic Bayesian
Networks (DBN).
∗ This work is supported by a contract with the US Defense Advanced Research Projects Agency
(DARPA), Contract W911NF-15-C-0238. Approved for Public Release, Distribution Unlimited.
The views expressed are those of the authors and do not reflect the official policy or position of
the Department of Defense or the U.S. Government. We would like to thank Nikhil Krishnaswamy
and Keigh Rim for their discussion and input on this topic. All errors and mistakes are, of course,
the responsibilities of the authors..
While learning events as a whole works best for human motion signatures such as
running, sitting etc., it poses a problem for event types that require distinctions in
spatio-temporal relationships between objects. As pointed in [10], it is also difficult
to model events as strict orderings of subevents, especially when there are overlapping
or during relations between them. Moreover, if the purpose of event learning is to
facilitate communication and interaction between human and computational agents,
such as robots, to achieve some common goals, these agents need to keep track
of multiple events at the same time, involving themselves, other humans, as well
as the surrounding environment. From a practical point of view, this calls for a
finer-grained treatment of event modeling.
It is also the case that a fine-grained analysis of events is strongly supported
from a theoretical point of view. For example, it has long been known that event
classification needs to take into account what is called extra-verbal factors. Event
types should not be semantically defined only by a base verbal expression, such as
running or walking, but need to incorporate other components of the expression
compositionally, such as objects and adjuncts, which can change the event type of
the overall verb phrase or sentence [12].
Motivated by those arguments, we suggest a different approach to event learning.
Instead of treating events as whole, or as programs of subevents, we allow multiple
interpretations of a motion capture by slicing over its temporal span and give a
separate annotation for each slice.
In particular, we use an event capture and annotation tool called ECAT [13],
which employs Microsoft Kinectr to capture sessions of performers interacting with
two types of objects, a cube (which can be slid on a flat surface) and a cylinder
(which can be rolled). Objects are tracked using markers fixed to their sides facing
the camera. They are then projected into three dimensional space using depth of
field. Performers are tracked using the Kinect API, which provides three dimensional inputs of a performer’s joint points (e.g., wrist, palm, shoulder). Sessions are
first sliced, and each slice is annotated with a textual description using our event
language. Our sequence learning algorithms (LSTM-CRF) will input sequences of
feature vectors and output a representation of an event.
The main contributions of our study are twofolds. Firstly we created a framework
for event recording and annotation that takes into account their temporal dynamics,
i.e., different interpretations of events on different temporal spans. Applying a flavor
of the popular sequential learning method LSTM that accommodates to output
constraints, we achieved good performance in our human-object interaction setup.
2
Learning Framework
We used three dimensional coordinates of bodies tracked by KinectrSDK to model
human kinematics. Only joint points on the upper body of performers (13 joint
points) are kept and concatenated, because their lower parts are occluded when they
interact with objects on top of a table. In addition, for each marker detected (using
Glyph detection algorithm[14]), we generate 12 features (4 3-D corners). Features of
objects are concatenated into a vector of a fixed size (there is always one performer
and two objects tracked). A sample in the dataset consists of a sequence of fixed
length of feature vectors. Its label is mapped from the textual annotation into an
output structure (Subject, Object, Locative, Verb, Preposition). We will call this
output structure (Y1 , . . . , Y5 ).
2.1
LSTM
LSTM is a flavor of deep Recursive neural network (RNN) that has generally solved
the problem of “vanishing gradients” in traditional RNN learning [15, 16] and has
found their application in a wide range of problems involving sequential learning,
such as hand written recognition, speech recognition, gesture recognition, etc.
p1
Y1
Y2
Y3
p2
p3
p4
Y4
Y5
Fig. 1: LSTM model with possible constraints of outputs with CRF. CRF layer is
represented as dashed links among predicted labels.
We will not describe in detail here our LSTM implementation, as we provide
online access to the code and approach 1 . Briefly, however, the model passes each
feature vector through a linear layer before feeding each sequence into an LSTM.
Each label Yi requires a separate LSTM cell, Xi . Depending on whether we predict
each label independently and combine them for final prediction, or we predict on
the basis of the sum of outputs from the last layer, two variants are considered,
correspondingly LSTM-I and LSTM-W.
2.2
CRF
SU BJ
V ERB
OBJ
OBJ
ST ART
LOC
LOC
SU BJ
P REP
V ERB
P REP
Fig. 2: Reformation from general CRF (left) to Tree-CRF (right)
CRF has been used extensively to learn structured output as it allows specification of constraints of output labels[17]. In this model we wish to constrain the
1 https://github.com/tuandnvn/ecat
learning
outputs so that: one object (Performer or the other objects) is not allowed to fill two
different syntactic slots; when there is no verb, all the other slots should be None;
locative and preposition are dependent, because if locative is None, preposition must
also be None and vice versa. The edges between nodes on the left side of Figure 2
show the dependencies on output labels that we wish to model.
However, training and classifying using a full CRF model would be more difficult,
especially when implemented with a neural network architecture. We modified the
model into a tree-CRF structure (right side of Figure 2) to make the model learnable
using dynamic programming. The complexity of the algorithm reduced from O(n5 )
to O(n2 ) ∗ 5) where n is the size of our vocabulary. The learning problem is thereby
changed to learning the weights along the edges on the tree-CRF, for example,
P locative preposition (together with parameters of LSTM). The directionality of
edges is the forward direction of the message passing algorithm used for learning
(and in reverse, for testing using the backward direction).
2.3
LSTM-CRF
LSTM-CRF is a natural extension of LSTM applied for constrained outputs. For
instance it is used for named entities recognition task to model constraints on BIO
labels[18]. To put CRF learning on top of LSTM-W, we modify the term t (the term
before softmax) produced by outputs of LSTM as followings.
t(l, s, o, p, v) = tl + ts + to + tp + tv Original LSTM-W
t(l, s, o, p, v) = tl + ts + to + tp + tv Modified
+ Pstart l + Pls + Plo + Plp + Psv
where l, s, o, p, v stand for Locative, Subject, Object, Preposition and Verb respectively.
In training, sof tmax is calculated for a predicted label combination, namely
(l′ , s′ , o′ , p′ , v ′ ) as below. We can calculate the log of sum using message passing over
the tree nodes of the CRF tree. We use cross entropy between predicted distribution
and correct output as the cost in training.
sof tmax = exp[t(l′ , s′ , o′ , p′ , v ′ ) − log[
XXXXX
l
= exp[t(l′ , s′ , o′ , p′ , v ′ ) − log[
X
v
exp(tv + Psv )][
X
o
X
s
o
p
exp(t(l, s, o, p, v))]]
v
exp(tl + Pstart l )[
l
exp(to + Plo )][
X
X
exp(ts + Pls )
s
exp(tp + Plp )]]
p
In evaluation, a similar message passing algorithm is used, but instead of log sum,
we use max to calculate the probabilities and argmax to keep track of the best
combination.
3
3.1
Experiments
Event Capture and Annotation
To demonstrate our model’s capability to learn the spatio-temporal dynamics of
object interactions in events, we use a collection of four action types: push, pull, slide,
and roll, along with three different spatial prepositions used for space configurations
between objects, namely toward (when the trajectory of a moving object is straightly
lined up with a destination static object and makes it closer to that target), away
from (makes it further from that object) and past (moving object getting closer to
static object then further again).
Afterwards, for each session, we sliced the events into short segments of 20 frames.
Two annotators were assigned to watch and annotate them (segments can be played
back). To speed up annotation, only event types related to original captured types
are shown for selection. For instance, if the event type of the captured session is
“The performer pushes A toward B”, other available event types are “The performer
pushes A”, “A slides toward B” or “None”.
3.2
Classification Results
Our LSTM models have one hidden layer of 200 features. Two methods are used
to combat over-fitting: (i) dropout in LSTM cell with probability of 0.8, and (ii)
gradient clipping by a global norm. The network is trained with mini-batch gradient
descent optimization for 200 epochs on the Tensorflow library.
Most frequent label tagging is used as the baseline for this study: that is, we
simply predict any sample with the most frequent tuple seen in the training corpus.
Captured sessions are split into training and testing sets on the proportion of
60/40, i.e., 18 sessions of training and 12 sessions for testing for each event type.
That gives a total of 2680 training samples and 1840 testing samples. Precisions
reported are averaged over 5 runs (each run is obtained with a random initialization).
Breakdowns of precision for each label show that verb precision is the lowest, with
high confusion between the following pairs: push vs roll, pull vs roll, and slide vs
roll. It is likely because of poor tracking result when objects are rolled. In fact,
the capture tool could not recognize objects in many frames when objects roll fast,
and it compensated by using interpolation. Improvement on tracking of objects,
however, is not the target of our study.
Model
Baseline
LSTM-I
LSTM-W
LSTM-CRF
Precision
6%
38%
39%
43%
Table 1: Evaluation
Label
Subject
Object
Locative
Verb
Preposition
Precision
86%
87%
73%
68%
72%
Table 2: Label precision breakdown
We observed significant improvement of learning using LSTM over baseline, especially when it is coupled with CRF. Moreover, we also observed a reduction of invalid
outputs from 20% to 3% when CRF is used. We consider these results to be quite
good, particularly since the sequential learning model we used is both simple and
fast, and we did not employ any feature engineering method.
4
Conclusion and Future Directions
In this paper, we have demonstrated a methodology that provides reasonable learning results for a set of complex event types. We hope that our study will provide a
starting point for further investigations into fine-grained event learning. Currently
our learning method requires a fix number of objects in inputs, which could be overcome by incrementally adding object features into a fix size feature vector, possibly
by using a recursive neural network. Regarding our annotation framework, a natural extension is that spans of different lengths could be annotated with appropriate
re-sampling methods. We leave these as some of our future research topics.
We are currently applying this learning pipeline (simple interval annotation +
sequential learning with constraint outputs) on a large movie dataset that have event
annotations. We are looking to publish our results in the near future.
References
[1] Endel Tulving. Elements of Episodic Memory. Oxford University Press, 1983.
[2] James Allen. Towards a general theory of action and time. Arificial Intelligence, 23:123–154,
1984.
[3] C. Fillmore. Santa Cruz Lectures on Deixis. Indiana University Linguistics Club. Bloomington,
IN, 1975.
[4] James Pustejovsky, Kiyong Lee, Harry Bunt, and Laurent Romary. Iso-timeml: A standard
for annotating temporal information in language. In LREC, 2011.
[5] James Pustejovsky and Jessica L Moszkowicz. The qualitative spatial dynamics of motion in
language. Spatial Cognition & Computation, 11(1):15–44, 2011.
[6] James Pustejovsky and Nikhil Krishnaswamy. Generating simulations of motion events from
verbal descriptions. Lexical and Computational Semantics (* SEM 2014), page 99, 2014.
[7] Amir Shahroudy, Jun Liu, Tian-Tsong Ng, and Gang Wang. Ntu rgb+ d: A large scale dataset
for 3d human activity analysis. arXiv preprint arXiv:1604.02808, 2016.
[8] Harini Veeraraghavan, Nikolaos Papanikolopoulos, and Paul Schrater. Learning dynamic event
descriptions in image sequences. In 2007 IEEE Conference on Computer Vision and Pattern
Recognition, pages 1–6. IEEE, 2007.
[9] Quoc V Le, Will Y Zou, Serena Y Yeung, and Andrew Y Ng. Learning hierarchical invariant
spatio-temporal features for action recognition with independent subspace analysis. In Computer Vision and Pattern Recognition (CVPR), 2011 IEEE Conference on, pages 3361–3368.
IEEE, 2011.
[10] Krishna Sandeep Reddy Dubba, Anthony G Cohn, David C Hogg, Mehul Bhatt, and Frank
Dylla. Learning relational event models from video. Journal of Artificial Intelligence Research,
53:41–90, 2015.
[11] Anthony Hoogs and AG Amitha Perera. Video activity recognition in the real world. In AAAI,
pages 1551–1554, 2008.
[12] James Pustejovsky. The Generative Lexicon. MIT Press, Cambridge, MA, 1995.
[13] Tuan Do, Nikhil Krishnaswamy, and James Pustejovsky. Ecat: Event capture annotation tool.
Proceedings of ISA-12: International Workshop on Semantic Annotation, 2016.
[14] Andrew
Kirillov.
From
glyph
recognition
to
augmented
reality.
https://www.codeproject.com/articles/258856/from-glyph-recognition-to-augmented-reality.
Accessed: 2016-03-15.
[15] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9
(8):1735–1780, 1997.
[16] Jürgen Schmidhuber. Deep learning in neural networks: An overview. Neural Networks, 61:
85–117, 2015.
[17] Charles Sutton and Andrew McCallum. An introduction to conditional random fields for
relational learning, volume 2. Introduction to statistical relational learning. MIT Press, 2006.
[18] Zhiheng Huang, Wei Xu, and Kai Yu. Bidirectional lstm-crf models for sequence tagging.
arXiv preprint arXiv:1508.01991, 2015.
| 1cs.CV
|
Dual Recurrent Attention Units for Visual Question Answering
Ahmed Osman 1,2 , Wojciech Samek 1
1
Fraunhofer Heinrich Hertz Institute, Berlin, Germany
2
University of Freiburg, Freiburg, Germany
arXiv:1802.00209v1 [cs.AI] 1 Feb 2018
{ahmed.osman, wojciech.samek}@hhi.fraunhofer.de
Abstract
We propose an architecture for VQA which utilizes recurrent layers to generate visual and textual attention. The
memory characteristic of the proposed recurrent attention
units offers a rich joint embedding of visual and textual features and enables the model to reason relations between
several parts of the image and question. Our single model
outperforms the first place winner on the VQA 1.0 dataset,
performs within margin to the current state-of-the-art ensemble model. We also experiment with replacing attention
mechanisms in other state-of-the-art models with our implementation and show increased accuracy. In both cases,
our recurrent attention mechanism improves performance
in tasks requiring sequential or relational reasoning on the
VQA dataset.
Figure 1. Diagram of the DRAU network.
tial manner which recurrent layers are better suited due to
their ability to capture relevant information over an input
sequence.
In this paper, we propose a RNN-based joint representation to generate visual and textual attention. We argue
that embedding a RNN in the joint representation helps the
model process information in a sequential manner and determine what is relevant to solve the task. We refer to the
combination of RNN embedding and attention as Recurrent
Textual Attention Unit (RTAU) and Recurrent Visual Attention Unit (RVAU) respective of their purpose. Furthermore,
we employ these units in a fairly simple network, referred
to as Dual Recurrent Attention Units (DRAU) network, and
show improved results over several baselines. Finally, we
enhance state-of-the-art models by replacing the model’s
default attention mechanism with RVAU.
Our main contributions are the following:
1. Introduction
Although convolutional neural networks (CNNs) and recurrent neural networks (RNNs) have been successfully
applied to various image and natural language processing
tasks (cf. [1, 2, 3, 4]), these breakthroughs only slowly
translate to multimodal tasks such as visual question answering (VQA) where the model needs to create a joint understanding of the image and question. Such multimodal
tasks require joint visual and textual representations.
Since global features can hardly answer questions about
certain local parts of the input, attention mechanisms have
been extensively used in VQA recently [5, 6, 7, 8, 9, 10,
11, 12]. It attempts to make the model predict based on
spatial or lingual context. However, most attention mechanisms used in VQA models are rather simple, consisting
of two convolutional layers followed by a softmax to generate the attention weights which are summed over the image features. These shallow attention mechanisms may fail
to select the relevant information from the joint representation of the question and image. Creating attention for complex questions, particularly sequential or relational reasoning questions, requires processing information in a sequen-
• We introduce a novel approach to generate soft attention. To the best of our knowledge, this is the first attempt to generate attention maps using recurrent neural
networks. We provide quantitative and qualitative results showing performance improvements over the de1
fault attention used in most VQA models.
• Our attention modules are modular, thus, they can substitute existing attention mechanisms in most models
fairly easily. We show that state-of-the-art models with
RVAU “plugged-in” perform consistently better than
their vanilla counterparts.
2. Related Work
This section discusses common methods that have been
explored in the past for VQA.
Bilinear representations Fukui et al. [7] use compact bilinear pooling to attend over the image features and combine it with the language representation. The basic concept
behind compact bilinear pooling is approximating the outer
product by randomly projecting the embeddings to a higher
dimensional space using Count Sketch projection [13] and
then exploiting Fast Fourier Transforms to compute an efficient convolution. An ensemble model using MCB won first
place in VQA (1.0) 2016 challenge. Kim et al. [5] argues
that compact bilinear pooling is still expensive to compute
and shows that it can be replaced by element-wise product (Hadamard product) and a linear mapping (i.e. fullyconnected layer) which gives a lower dimensional representation and also improves the model accuracy. Recently,
Ben-younes et al. [14] proposed using Tucker decomposition [15] with a low-rank matrix constraint as a bilinear
representation. They propose this fusion scheme in an architecture they refer to as MUTAN which as of this writing
is the current state-of-the-art on the VQA 1.0 dataset.
Attention-based Closely related to our work, Lu et al. [9]
were the first to feature a co-attention mechanism that applies attention to both the question and image. Nam et al.
[6] use a Dual Attention Network (DAN) that employs attention on both text and visual features iteratively to predict
the result. The goal behind this is to allow the image and
question attentions to iteratively guide each other in a synergistic manner.
RNNs for VQA Using recurrent neural networks (RNNs)
for VQA has been explored in the past. Xiong et al. [16]
build upon the dynamic memory network from Kumar and
Varaiya [17] and proposes DMN+. DMN+ uses episodic
modules which contain attention-based Gated Recurrent
Units (GRUs). Note that this is not the same as what we
propose; Xiong et al. generate soft attention using convolutional layers and then uses it to substitute the update gate of
the GRU. In contrast, our approach uses the recurrent layers
to generate the attention. Noh and Han [8] propose recurrent answering units in which each unit is a complete module that can answer a question about an image. They use
joint loss minimization to train the units. However during
testing, they use the first answering unit which was trained
from other units through backpropagation.
Notable mentions Kazemi and Elqursh [18] show that a
simple model can get state-of-the-art results with proper
training parameters. Wu et al. [19] construct a textual representation of the semantic content of an image and merges
it with textual information sourced from a knowledge base.
Ray et al. [20] introduce a task of identifying relevant questions for VQA. Kim et al. [21] apply residual learning techniques to VQA and propose a novel attention image attention visualization method using backpropagation.
3. Dual Recurrent Attention in VQA
We propose our method in this section. Figure 1 illustrates the flow of information in the DRAU model. Given
an image and question, we create the input representations
v and q. Next, these features are combined by 1 × 1 convolutions into two separate branches. Then, the branches
are passed to an RTAU and RVAU. Finally, the branches are
combined using a fusion operation and fed to the final classifier. The full architecture of the network is depicted in
Figure 2.
3.1. Input Representation
Image representation We use the 152-layer “ResNet”
pretrained CNN from He et al. [1] to extract image features.
Similar to [7, 6], we resize the images to 448 × 448 and
extract the last layer before the final pooling layer (res5c)
with size 2048 × 14 × 14. Finally, we use l2 normalization on all dimensions. Recently, Anderson et al. [22] have
shown that object-level features can provide a significant
performance uplift compared to global-level features from
pretrained CNNs. Therefore, we experiment with replacing
the ResNet features with FRCNN [23] features with a fixed
number of proposals per image (K = 36).
Question representation We use a fairly similar representation as [7]. In short, the question is tokenized and encoded using an embedding layer followed by a tanh activation. We also exploit pretrained GloVe vectors [24] and
concatenate them with the output of the embedding layer.
The concatenated vector is fed to a two-layer unidirectional
LSTM that contains 1024 hidden states each. In contrast
to Fukui et al., we use all the hidden states of both LSTMs
rather than concatenating the final states to represent the final question representation.
3.2. 1 × 1 Convolution and PReLU
We apply multiple 1 × 1 convolution layers in the network for mainly two reasons. First, they learn weights from
Figure 2. The proposed network. ⊕ denotes concatenation.
the image and question representations in the early layers.
This is important especially for the image representation,
since it was originally trained for a different task. Second,
they are used to generate a common representation size.
To obtain a joint representation, we apply 1 × 1 convolutions followed by PReLU activations [1] on both the image
and question representations. Through empirical evidence,
PReLU activations were found to reduce training time significantly and improve performance compared to ReLU and
tanh activations. We provide these results in Section 4.
3.3. Recurrent Attention Units
The result from the above-mentioned layers is concatenated and fed to two separate recurrent attention units
(RAU). Each RAU starts with another 1 × 1 convolution
and PReLU activation:
ca = PReLU Wa x
(1)
where Wa is the 1 × 1 convolution weights, x is the input to
the RAU, and ca is the output of the first PReLU.
Furthermore, we feed the previous output into an unidirectional LSTM:
ha,n = LSTM ca,n
(2)
Next, we use the attention weights to compute a
weighted average of the image and question features.
atta,n =
N
X
Watt,n fn
(4)
n=1
where fn is the input representation and atta,n is the attention applied on the input. Finally, the attention maps are fed
into a fully-connected layer followed by a PReLU activation. Figure 3 illustrates the structure of a RAU.
yatt,n = PReLU Wout atta,n
(5)
where Wout is a weight vector of the fully connected layer
and yatt,n is the output of each RAU.
3.4. Reasoning layer
A fusion operation is used to merge the textual and visual
branches. For DRAU, we experiment with using elementwise multiplication (Hadamard product) and MCB [7, 25].
The result of the fusion is given to a many-class classifier
using the top 3000 frequent answers. We use a single-layer
softmax with cross-entropy loss. This can be written as:
Pa = softmax fusion op ytext , yvis Wans
(6)
where ha,n is the hidden state at time n.
To generate the attention weights, we feed all the hidden
states of the previous LSTM to a 1 × 1 convolution layer
followed by a softmax function. The 1×1 convolution layer
could be interpreted as the number of glimpses the model
sees.
Watt,n = softmax PReLU Wg ha,n
(3)
where ytext and yvis are the outputs of the RAUs, Wans
represents the weights of the multi-way classifier, and Pa is
the probability of the top 3000 frequent answers.
The final answer â is chosen according to the following:
where Wg is the glimpses’ weights and Watt,n is the attention weight vector.
Experiments are performed on the VQA 1.0 and 2.0
datasets [26, 27]. These datasets use images from the
â = argmax Pa
(7)
4. Experiments and Results
VQA 1.0 Validation Split
Open Ended Task
Baselines
Y/N Num. Other
Language only
Simple MCB
Joint LSTM
Figure 3. Recurrent Attention Unit.
MS-COCO dataset [28] and generate questions and labels
(10 labels per question) using Amazon’s Mechanical Turk
(AMT). Compared to VQA 1.0, VQA 2.0 adds more imagequestion pairs to balance the language prior present in the
VQA 1.0 dataset. The ground truth answers in the VQA
dataset are evaluated using human consensus.
P
a is in human annotation
,1
(8)
Acc(a) = min
3
We evaluate our results on the validation, test-dev, teststd splits of each dataset. Models evaluated on the validation set use train and Visual Genome for training. For
the other splits, we include the validation set in the training
data. However, the models using FRCNN features do not
use data augmentation with Visual Genome.
To train our model, we use Adam [29] for optimization
with β1 = 0.9, β2 = 0.999, and an initial learning rate
of = 7 × 10−4 . The final model is trained with a small
batch size of 32 for 400K iterations. We did not fully explore tuning the batch size which explains the relatively
high number of training iterations. Dropout (p = 0.3) is
applied after each LSTM and after the fusion operation. All
weights are initialized as described in [30] except LSTM
layers which use an uniform weight distribution. The pretrained ResNet was fixed during training due to the massive
computational overhead of fine-tuning the network for the
VQA task. While VQA datasets provide 10 answers per
image-question pair, we sample one answer randomly for
each training iteration.
4.1. VQA 1.0 Experiments
During early experiments, the VQA 2.0 dataset was not
yet released. Thus, the baselines and early models were
evaluated on the VQA 1.0 dataset. While building the final model, several parameters were changed, mainly, the
learning rate, activation functions, dropout value, and other
modifications which we discuss in this section.
Baselines We started by designing three baseline architectures. The first baseline produced predictions solely
from the question while totally ignoring the image. The
78.56
78.64
79.90
27.98
32.98
36.96
30.76
39.79
49.58
All
48.3
54.82
59.34
Table 1. Evaluation of the baseline models on the VQA 1.0 validation split.
model used the same question representation described in
[7] and passed the output to a softmax 3000-way classification layer. The goal of this architecture was to assess the
extent of the language bias present in VQA.
The second baseline is a simple joint representation of
the image features and the language representation. The
representations were combined using the compact bilinear
pooling from [25]. We chose this method specifically because it was shown to be effective by Fukui et al. [7]. The
main objective of this model is to measure how a robust
pooling method of multimodal features would perform on
its own without a deep architecture or attention. We refer to
this model as Simple MCB.
For the last baseline, we substituted the compact bilinear pooling from Simple MCB with an LSTM consisting
of hidden states equal to the image size. A 1 × 1 convolutional layer followed by a tanh activation were used on the
image features prior to the LSTM, while the question representation was replicated to have a common embedding size
for both representations This model is referred to as Joint
LSTM.
We begin by testing our baseline models on the VQA
1.0 validation set. As shown in Table 1, the language-only
baseline model managed to get 48.3% overall. More impressively, it scored 78.56% on Yes/No questions. The Simple MCB model further improves the overall performance,
although little improvement is gained in the binary Yes/No
tasks. Replacing MCB with our basic Joint LSTM embedding improves performance across the board.
Modifications to the Joint LSTM Model We test several variations of the Joint LSTM baseline which are highlighted in Table 2. Using PReLU activations has helped
in two ways. First, it reduced time for convergence from
240K iterations to 120K. Second, the overall accuracy has
improved, especially in the Other category. The next modifications were inspired by the results from [18]. We experimented with appending positional features which can be described as the coordinates of each pixel to the depth/feature
dimension of the image representation. When unnormalized
with respect to the other features, it worsened results significantly, dropping the overall accuracy by over 2 points.
Model
VQA 1.0 Validation Split
Open Ended Task
Y/N Num. Other
Joint LSTM baseline
PReLU
Pos. features
Pos. features (norm.)
High dropout
Extra FC
79.90
79.61
79.68
79.69
79.03
78.86
36.96
36.21
36.52
36.36
34.84
33.51
49.58
50.77
46.59
50.69
47.25
45.57
All
59.34
59.74
57.71
59.75
57.59
56.51
Table 2. Evaluation of the Joint LSTM model and its modifications
on the VQA 1.0 validation split.
Normalizing positional features did not have enough of a
noticeable improvement (0.01 points overall) to warrant its
effectiveness. Next, all dropout values are increased from
0.3 to 0.5 deteriorated the network’s accuracy, particularly
in the Number and Other categories. The final modification was inserting a fully connected layer with 1024 hidden
units before the classifier, which surprisingly dropped the
accuracy massively.
4.2. VQA 2.0 Experiments
After the release of VQA 2.0, we shifted our empirical
evaluation towards the newer dataset. First, we retrain and
retest our best performing VQA 1.0 model Joint LSTM as
well as several improvements and modifications.
Since VQA 2.0 was built to reduce the language prior
and bias inherent in VQA, the accuracy of Joint LSTM drops
significantly as shown in Table 3. Note that all the models that were trained so far do not have explicit visual or
textual attention implemented. Our first network with explicit visual attention, RVAU, shows an accuracy jump by
almost 3 points compared to the Joint LSTM model. This
result highlights the importance of attention for good performance in VQA. Training the RVAU network as a multilabel task (RVAUmultilabel ), i.e. using all available annotations
at each training iteration, drops the accuracy horribly. This
is the biggest drop in performance so far. This might be
caused by the variety of annotations in VQA for each question which makes the task for optimizing all answers at once
much harder.
DRAU Evaluation The addition of RTAU marks the creation of our DRAU network. The DRAU model shows favorable improvements over the RVAU model. Adding textual attention improves overall accuracy by 0.56 points.
Substituting the PReLU activations with ReLU (DRAUReLU )
massively drops performance. While further training might
have helped the model improve, PReLU offers much faster
1 Concurrent
Work
Model
VQA 2.0 Validation Split
Open Ended Task
Y/N Num.
Other
All
Joint LSTM w/PReLU
72.04
37.95
48.58
56.00
RVAU
RVAUmultilabel
74.59
77.53
37.75
36.05
52.81
40.18
59.02
53.67
DRAUHadamard fusion
DRAUanswer vocab = 5k
DRAUReLU
DRAUno final dropout
DRAUhigh final dropout
76.62
76.33
72.69
77.02
76.47
38.92
38.21
34.92
38.26
38.71
52.09
51.85
45.05
50.17
52.52
59.58
59.27
54.11
58.69
59.71
-
-
-
59.14
59.67
MCB [26]
Kazemi and Elqursh [18]1
Table 3. Evaluation of RVAU and DRAU-based models on the
VQA 2.0 validation split.
convergence. Increasing the value of the dropout layer after
the fusion operation (DRAUhigh final dropout ) improves performance by 0.13 points, in contrast to the results of the Joint
LSTM model on VQA 1.0. Note that on the VQA 1.0 tests,
we changed the values of all layers that we apply dropout
on, but here we only change the last one after the fusion
operation. Totally removing this dropout layer worsens accuracy. This suggests that the optimal dropout value should
be tuned per-layer.
We test a few variations of DRAU on the test-dev set. We
can observe that VQA benefits from more training data; the
same DRAU network performs better (62.24% vs. 59.58%)
thanks to the additional data. Most of the literature resize
the original ResNet features from 224 × 224 to 448 × 448.
To test the effect of this scaling, we train a DRAU variant
with the original ResNet size (DRAUsmall ). Reducing the
image feature size from 2048 × 14 × 14 to 2048 × 7 × 7
adversely affects accuracy as shown in Table 4. Adding
more glimpses significantly reduces the model’s accuracy
(DRAUglimpses = 4 ). A cause of this performance drop could
be related to the fact that LSTMs process the input in a onedimensional fashion and thus decide that each input is either
relevant or non-relevant. This might explain why the attention maps of DRAU separate the objects from the background in two glimpses as we will mention in Section 5.
2D Grid LSTMs [31] might help remove this limitation.
Removing the extra data from Visual Genome hurts the
model’s accuracy. That supports the fact that VQA is very
diverse and that extra data helps the model perform better.
Finally, substituting Hadamard product of MCB in the final
fusion operation boosts the network’s accuracy significantly
by 1.17 points (DRAUMCB fusion ).
As mentioned in Section 3.1, we experiment replacing
the global ResNet features with object-level features as sug-
VQA 2.0 Test-Dev Split
Open Ended Task
Y/N
Num. Other
Model
DRAUHadamard fusion
DRAUsmall
DRAUglimpses = 4
DRAUno genome
DRAUMCB fusion
DRAUFRCNN features
78.27
77.53
76.82
79.63
78.97
82.85
40.31
38.78
39.15
39.55
40.06
44.78
53.57
49.93
51.07
51.81
55.47
57.4
4.4. DRAU versus the state-of-the-art
All
62.24
60.03
60.32
61.88
63.41
66.45
Table 4. Evaluation of later DRAU-based models on the VQA 2.0
test-dev split.
VQA 2.0 Test-dev Split
Open Ended Task
Y/N Num. Other
Model
All
3
MCB [7]
MCB w/RVAU
78.41
77.31
38.81
40.12
53.23
54.64
61.96
62.33
MUTAN [14]
MUTAN w/RVAU
79.06
79.33
38.95
39.48
53.46
53.28
62.36
62.45
Table 5. Results of state-of-the-art models with RVAU.
gested by [22]. This change provides a significant performance increase of 3.04 points (DRAUFRCNN features ).
4.3. Transplanting RVAU in other models
To verify the effectiveness of the recurrent attention
units, we replace the attention layers in MCB and MUTAN
[14] with RVAU.
For MCB [7] we remove all the layers after the first MCB
operation until the first 2048-d output and replace them with
RVAU. Due to GPU memory constraints, we reduced the
size of each hidden unit in RVAU’s LSTM from 2048 to
1024. In the same setting, RVAU significantly helps improve the original MCB model’s accuracy as shown in Table 5. The most noticeable performance boost can be seen
in the number category, which supports our hypothesis that
recurrent layers are more suited for sequential reasoning.
Furthermore, we test RVAU in the MUTAN model [14].
The authors use a multimodal vector with dimension size of
510 for the joint representations. For coherence, we change
the usual dimension size in RVAU to 510. At the time of this
writing, the authors have not released results on VQA 2.0
using a single model rather than a model ensemble. Therefore, we train a single-model MUTAN using the authors’
implementation.2 The story does not change here, RVAU
improves the model’s overall accuracy.
2 https://github.com/Cadene/vqa.pytorch
3 http://www.visualqa.org/roe_2017.html
VQA 1.0 Table 6 shows a comparison between DRAU
and other state-of-the-art models. Excluding model ensembles, DRAU performs favorably against other models. To
the best of our knowledge, [5] has the best single model
performance of 65.07% on the test-std split which is very
close our best model (65.03%). Small modifications or hyperparameter tuning could push our model further. Finally,
the FRCNN image features boosts the model’s performance
close to the state-of-the-art ensemble model.
VQA 2.0 Our model DRAUMCB fusion landed the 8th place
in the VQA 2.0 Test-standard task.4 . Currently, all reported
submissions that outperform our single model use model
ensembles. Using FRCNN features boosted the model’s
performance to outperform some of the ensemble models
(66.85%). The first place submission [22] reports using
an ensemble of 30 models. In their report, the best single
model that uses FRCNN features achieves 65.67% on the
test-standard split which is outperformed by our best single
model DRAUFRCNN features .
5. DRAU versus MCB
In this section, we provide qualitative results that highlight the effect of the recurrent layers compared to the MCB
model.
The strength of RAUs is notable in tasks that require sequentially processing the image or relational/multi-step reasoning. In the same setting, DRAU outperforms MCB in
counting questions. This is validated in a subset of the validation split questions in the VQA 2.0 dataset as shown in
Figure 4. Figure 5 shows some qualitative results between
DRAU and MCB. For fair comparison we compare the first
attention map of MCB with the second attention map of our
model. We do so because the authors of MCB [7] visualize
the first map in their work5 . Furthermore, the first glimpse
of our model seems to be the complement of the second
attention, i.e. the model separates the background and the
target object(s) into separate attention maps. We have not
tested the visual effect of more than two glimpses on our
model.
In Figure 5, it is clear that the recurrence helps the model
attend to multiple targets as apparent in the difference of the
attention maps between the two models. DRAU seems to
also know how to count the right object(s). The top right example in Figure 5 illustrates that DRAU is not easily fooled
by counting whatever object is present in the image but
rather the object that is needed to answer the question. This
4 https://evalai.cloudcv.org/web/challenges/
challenge-page/1/leaderboard
5 https://github.com/akirafukui/vqa-mcb/blob/
master/server/server.py#L185
Model
VQA 1.0 Open Ended Task
Test-dev
Y/N Num. Other
All
Y/N
Test-standard
Num. Other
SAN [10]
DMN+ [16]
MRN [21]
HieCoAtt [9]
RAU [8]
DAN [6]
MCB [7] (e = 7)
MLB [5] (1 model)
MLB [5] (e = 7)
MUTAN [14] (e = 5)
79.3
80.5
82.28
79.7
81.9
83.0
83.4
84.57
85.14
36.6
36.8
38.82
38.7
39.0
39.1
39.8
39.21
39.81
46.1
48.3
49.25
51.7
53.0
53.9
58.5
57.81
58.52
58.7
60.3
61.68
61.8
63.3
64.3
66.7
66.77
67.42
82.39
81.7
82.8
83.24
84.02
84.61
84.91
38.23
38.2
38.1
39.47
37.90
39.07
39.79
49.41
52.8
54.0
58.00
54.77
57.79
58.35
58.9
60.4
61.84
62.1
63.2
64.2
66.47
65.07
66.89
67.36
DRAUHadamard fusion
DRAUMCB fusion
DRAUFRCNN features
82.73
82.44
84.92
38.18
38.22
39.16
54.43
56.30
57.70
64.3
65.1
66.86
82.41
84.87
38.33
40.02
55.97
57.91
65.03
67.16
All
Table 6. DRAU compared to the state-of-the-art on the VQA 1.0 dataset.e = n corresponds to a model ensemble of size n.
Model
VQA 2.0 Open Ended Task
Test-dev
Y/N Num. Other
All
Y/N
Test-standard
Num. Other
UPC
MIC TJ
neural-vqa-attention[10]
CRCV REU
VQATeam MCB [26]
DCD ZJU[32]
VQAMachine [33]
POSTECH
UPMC-LIP6[14]
LV NUS[34]
DLAIT
HDU-USYD-UNCC[35]
Adelaide-Teney ACRV MSR[36]
67.1
69.02
70.1
73.91
78.41
79.84
79.4
78.98
81.96
81.95
82.94
84.39
85.24
31.54
34.52
35.39
36.82
38.81
38.72
40.95
40.9
41.62
48.31
47.08
45.76
48.19
25.46
35.76
47.32
54.85
53.23
53.08
53.24
55.35
57.07
59.99
59.94
59.14
59.88
43.3
49.32
55.35
60.65
61.96
62.47
62.62
63.45
65.57
67.71
67.95
68.02
69.00
66.97
69.22
69.77
74.08
78.82
79.85
79.82
79.32
82.07
81.92
83.17
84.5
85.54
31.38
34.16
35.65
36.43
38.28
38.64
40.91
40.67
41.06
48.38
46.66
45.39
47.45
25.81
35.97
47.18
54.84
53.36
52.95
53.35
55.3
57.12
59.63
60.15
59.01
59.82
43.48
49.56
55.28
60.81
62.27
62.54
62.97
63.66
65.71
67.64
68.22
68.09
69.13
DRAUHadamard fusion
DRAUMCB fusion
DRAUFRCNN features
78.27
78.97
82.85
40.31
40.06
44.78
53.58
55.48
57.4
62.24
63.41
66.45
78.86
79.27
83.35
39.91
40.15
44.37
53.76
55.55
57.63
62.66
63.71
66.85
All
Table 7. DRAU compared to the current submissions on the VQA 2.0 dataset.
property also translates to questions that require relational
reasoning. The second column in Figure 5 demonstrates
how DRAU can attend the location required to answer the
question based on the textual and visual attention maps.
6. Conclusion
We proposed an architecture for VQA with a novel attention unit, termed the Recurrent Attention Unit (RAU).
The recurrent layers help guide the textual and visual atten-
tion since the network can reason relations between several
parts of the image and question. We provided quantitative
and qualitative results indicating the usefulness of a recurrent attention mechanism. Our DRAU model showed improved performance in tasks requiring sequential/complex
reasoning such as counting or relational reasoning over [7],
the winners of the VQA 2016 challenge. In VQA 1.0, we
achieved near state-of-the-art results for single model performance with our DRAU network (65.03 vs. 65.07 [5]).
Adding the FRCNN features gets the model within margin
Question Type (Occurrence)
Accuracy(%)
47.28
46.31
43.98
42.06
DRAU
MCB
44.30
42.57
Visual Question Answering and Visual Grounding,” in Empirical Methods in Natural Language Processing (EMNLP),
2016, pp. 457–468.
[8] H. Noh and B. Han, “Training Recurrent Answering Units
with Joint Loss Minimization for VQA,” arXiv:1606.03647,
2016.
[9] J. Lu, J. Yang, D. Batra, and D. Parikh, “Hierarchical
Question-Image Co-Attention for Visual Question Answering,” in Advances in Neural Information Processing Systems
(NIPS), 2016, pp. 289–297.
how many
people are in
(905)
how many
people are
(2,005)
how many
(20,462)
Figure 4. Results on questions that require counting in the VQA
2.0 validation set.
of the state-of-the-art 5-model ensemble MUTAN [14]. Finally, we demonstrated that substituting the visual attention
mechanism in other networks, MCB [7] and MUTAN [14],
consistently improves their performance. In future work we
will investigate implicit recurrent attention mechanism using recently proposed explanation methods [37, 38].
References
[1] K. He, X. Zhang, S. Ren, and J. Sun, “Delving Deep into
Rectifiers: Surpassing Human-Level Performance on ImageNet Classification,” in IEEE International Conference on
Computer Vision (ICCV), 2015, pp. 1026–1034.
[2] S. Bosse, D. Maniry, K.-R. Müller, T. Wiegand, and
W. Samek, “Deep neural networks for no-reference and fullreference image quality assessment,” IEEE Transactions on
Image Processing, vol. 27, no. 1, pp. 206–219, 2018.
[10] Z. Yang, X. He, J. Gao, L. Deng, and A. Smola, “Stacked Attention Networks for Image Question Answering,” in IEEE
Conference on Computer Vision and Pattern Recognition
(CVPR), 2016, pp. 21–29.
[11] K. Chen, J. Wang, L.-C. Chen, H. Gao, W. Xu, and
R. Nevatia, “ABC-CNN: An Attention Based Convolutional Neural Network for Visual Question Answering,”
arXiv:1511.05960, 2015.
[12] H. Xu and K. Saenko, “Ask, Attend and Answer: Exploring Question-Guided Spatial Attention for Visual Question
Answering,” in European Conference on Computer Vision
(ECCV), 2016, pp. 451–466.
[13] M. Charikar, K. Chen, and M. Farach-Colton, “Finding frequent items in data streams,” Theoretical Computer Science,
vol. 312, no. 1, pp. 3–15, 2004.
[14] H. Ben-younes, R. Cadene, M. Cord, and N. Thome, “MUTAN: Multimodal Tucker Fusion for Visual Question Answering,” arXiv:1705.06676, 2017.
[15] L. R. Tucker, “Some mathematical notes on three-mode factor analysis,” Psychometrika, vol. 31, no. 3, pp. 279–311,
1966.
[3] D. Bahdanau, K. Cho, and Y. Bengio, “Neural Machine
Translation by Jointly Learning to Align and Translate,”
in International Conference on Representation Learning
(ICLR), 2015.
[16] C. Xiong, S. Merity, and R. Socher, “Dynamic Memory Networks for Visual and Textual Question Answering,” in International Conference on Machine Learning (ICML), 2016,
pp. 2397–2406.
[4] R. Nallapati, B. Zhou, C. Gulcehre, B. Xiang et al., “Abstractive text summarization using sequence-to-sequence rnns
and beyond,” arXiv:1602.06023, 2016.
[17] P. R. Kumar and P. Varaiya, Stochastic systems: Estimation,
identification, and adaptive control. SIAM, 2015.
[5] J.-H. Kim, K.-W. On, J. Kim, J.-W. Ha, and B.-T. Zhang,
“Hadamard Product for Low-rank Bilinear Pooling,” in International Conference on Representation Learning (ICLR),
2017.
[6] H. Nam, J.-W. Ha, and J. Kim, “Dual Attention Networks
for Multimodal Reasoning and Matching,” in IEEE International Conference on Computer Vision (CVPR), 2017, pp.
299–307.
[7] A. Fukui, D. H. Park, D. Yang, A. Rohrbach, T. Darrell, and
M. Rohrbach, “Multimodal Compact Bilinear Pooling for
[18] V. Kazemi and A. Elqursh, “Show, Ask, Attend, and Answer: A Strong Baseline For Visual Question Answering,”
arXiv:1704.03162, 2017.
[19] Q. Wu, P. Wang, C. Shen, A. Dick, and A. van den Hengel, “Ask Me Anything: Free-form Visual Question Answering Based on Knowledge from External Sources,” in IEEE
Conference on Computer Vision and Pattern Recognition
(CVPR), 2016, pp. 4622–4630.
[20] A. Ray, G. Christie, M. Bansal, D. Batra, and D. Parikh,
“Question Relevance in VQA: Identifying Non-Visual And
False-Premise Questions,” in Empirical Methods in Natural
Language Processing (EMNLP), 2016, pp. 919–924.
How many
How many
camels are in
lanterns hang
the photo?
DRAU: 0
off the clock
tower ?
DRAU: 4
How many
lanterns hang
off the clock
tower?
GT: 4
MCB: 1
How many
camels are in
the photo?
GT: 0
What is on the
floor leaning on
the bench in
How many
people are
between the
shown ?
DRAU: 3
How many
people are
shown?
GT: 3
MCB: 2
people ?
DRAU: racket
What is on the
floor leaning on
the bench in
between the
people?
GT: racket
does the vase
animals are
depict ?
DRAU: owl
there ?
DRAU: 2
MCB: one
MCB: backpack
What animal
How many
How many
animals are
there?
GT: 2
MCB: 1
What animal
does the vase
depict?
GT: owl
MCB: elephant
Figure 5. DRAU vs. MCB Qualitative examples. Attention maps for both models shown, only DRAU has textual attention.
[21] J.-H. Kim, S.-W. Lee, D. Kwak, M.-O. Heo, J. Kim, J.-W.
Ha, and B.-T. Zhang, “Multimodal Residual Learning for
Visual QA,” in Advances in Neural Information Processing
Systems 29, 2016, pp. 361–369.
[22] P. Anderson, X. He, C. Buehler, D. Teney, M. Johnson,
S. Gould, and L. Zhang, “Bottom-Up and Top-Down At-
tention for Image Captioning and VQA,” arXiv:1707.07998,
2017.
[23] S. Ren, K. He, R. Girshick, and J. Sun, “Faster R-CNN:
Towards Real-Time Object Detection with Region Proposal
Networks,” arXiv:1506.01497, Jun. 2015.
[24] J. Pennington, R. Socher, and C. D. Manning, “GloVe:
Global Vectors for Word Representation,” in Empirical
Methods in Natural Language Processing (EMNLP), 2014,
pp. 1532–1543.
[25] Y. Gao, O. Beijbom, N. Zhang, and T. Darrell, “Compact
bilinear pooling,” in IEEE Conference on Computer Vision
and Pattern Recognition (CVPR), 2016, pp. 317–326.
[26] Y. Goyal, T. Khot, D. Summers-Stay, D. Batra, and
D. Parikh, “Making the V in VQA Matter: Elevating the
Role of Image Understanding in Visual Question Answering,” in IEEE Conference on Computer Vision and Pattern
Recognition (CVPR), 2017, pp. 6904–6913.
[27] S. Antol, A. Agrawal, J. Lu, M. Mitchell, C. L. Zitnick,
D. Batra, and D. Parikh, “VQA: Visual Question Answering,” in IEEE International Conference on Computer Vision
(CVPR), 2015, pp. 2425–2433.
[28] T.-Y. Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Dollár, and C. L. Zitnick, “Microsoft coco: Common objects in context,” in European Conference on Computer Vision – ECCV 2014, 2014, pp. 740–755.
[29] D. P. Kingma and J. Ba, “Adam: A Method for Stochastic
Optimization,” arXiv:1412.6980, 2014.
[30] X. Glorot and Y. Bengio, “Understanding the difficulty of
training deep feedforward neural networks,” in AISTATS,
2010, pp. 249–256.
[31] N. Kalchbrenner, I. Danihelka, and A. Graves, “Grid long
short-term memory,” arXiv:1507.01526, 2015.
[32] Y. Lin, Z. Pang, D. Wang, and Y. Zhuang, “Task-driven Visual Saliency and Attention-based Visual Question Answering,” arXiv:1702.06700, 2017.
[33] P. Wang, Q. Wu, C. Shen, and A. van den Hengel, “The
VQA-Machine: Learning How to Use Existing Vision Algorithms to Answer New Questions,” arXiv:1612.05386, 2016.
[34] I. Ilievski and J. Feng, “A Simple Loss Function for Improving the Convergence and Accuracy of Visual Question Answering Models,” arXiv:1708.00584, 2017.
[35] Z. Yu, J. Yu, C. Xiang, J. Fan, and D. Tao, “Beyond Bilinear:
Generalized Multi-modal Factorized High-order Pooling for
Visual Question Answering,” arXiv:1708.03619, 2017.
[36] D. Teney, P. Anderson, X. He, and A. van den Hengel, “Tips
and Tricks for Visual Question Answering: Learnings from
the 2017 Challenge,” arXiv:1708.02711, 2017.
[37] G. Montavon, W. Samek, and K.-R. Müller, “Methods for interpreting and understanding deep neural networks,” Digital
Signal Processing, vol. 73, pp. 1–15, 2018.
[38] L. Arras, G. Montavon, K.-R. Müller, and W. Samek, “Explaining recurrent neural network predictions in sentiment
analysis,” in EMNLP’17 Workshop on Computational Approaches to Subjectivity, Sentiment & Social Media Analysis
(WASSA), 2017, pp. 159–168.
| 1cs.CV
|
Evorus: A Crowd-powered Conversational Assistant
Built to Automate Itself Over Time
Ting-Hao (Kenneth) Huang, Joseph Chee Chang, and Jeffrey P. Bigham
Language Technologies Institute and Human-Computer Interaction Institute
Carnegie Mellon University
{tinghaoh, josephcc, jbigham}@cs.cmu.edu
arXiv:1801.02668v2 [cs.HC] 10 Jan 2018
ABSTRACT
Crowd-powered conversational assistants have been shown to
be more robust than automated systems, but do so at the cost
of higher response latency and monetary costs. A promising
direction is to combine the two approaches for high quality,
low latency, and low cost solutions. In this paper, we introduce
Evorus, a crowd-powered conversational assistant built to automate itself over time by (i) allowing new chatbots to be easily
integrated to automate more scenarios, (ii) reusing prior crowd
answers, and (iii) learning to automatically approve response
candidates. Our 5-month-long deployment with 80 participants and 281 conversations shows that Evorus can automate
itself without compromising conversation quality. Crowd-AI
architectures have long been proposed as a way to reduce cost
and latency for crowd-powered systems; Evorus demonstrates
how automation can be introduced successfully in a deployed
system. Its architecture allows future researchers to make further innovation on the underlying automated components in
the context of a deployed open domain dialog system.
ACM Classification Keywords
H.5.m. Information Interfaces and Presentation (e.g. HCI):
Miscellaneous
Author Keywords
crowd-powered system; crowdsourcing; real-time
crowdsourcing; conversational assistant; chatbot
INTRODUCTION
Conversational assistants, such as Apple’s Siri, Amazon’s
Echo, and Microsoft’s Cortana, are becoming increasingly
popular, but are currently limited to specific speech commands
that have been coded for pre-determined domains. As a result, substantial effort has been placed on teaching people
how to talk to these assistants, e.g., via books to teach Siri’s
language [36], and frequent emails from Amazon advertising
Alexa’s new skills [1]. To address the problem of users not
knowing what scenarios are supported, AI2 recently built an
Alexa skill designed to help people find skills they could use,
only to have it rejected by Amazon [8].
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for components of this work owned by others than the
author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or
republish, to post on servers or to redistribute to lists, requires prior specific permission
and/or a fee. Request permissions from [email protected].
CHI 2018, April 21–26, 2018, Montreal, QC, Canada
© 2018 Copyright is held by the owner/author(s). Publication rights licensed to ACM.
ACM 978-1-4503-5620-6/18/04...$15.00
https://doi.org/10.1145/3173574.3173869
Figure 1. Evorus is a crowd-powered conversational assistant that automates itself over time by (i) learning to include responses from chatterbots and task-oriented dialog systems over time, (ii) reusing past responses, and (iii) gradually reducing the crowd’s role in choosing highquality responses by partially automating voting.
Crowd-powered assistants are more robust to diverse domains,
and are able to engage users in rich, multi-turn conversation. Some systems use professional employees, such as Facebook M [15], while others use crowd workers, such as Chorus [25]. Despite their advantages, crowd-powered agents remain largely impractical for deployment at large scale because
of their monetary cost and response latency [3, 19]. On the
other hand, crowd-powered systems are often touted as a path
to fully automated systems, but transitioning from the crowd
to automation has been limited in practice. The most straightforward approach is to use data from prior conversations to
train an automated replacement. This can work in specific
domains [46], or on so-called “chit-chat” systems [2]. Fully
automating a general conversational assistant this way can be
difficult because of the wide range of domains to cover and the
large amount of data needed within each to train automated
replacements. Such automated systems only become useful
once they can completely take over from the crowd-powered
system. Such abrupt transition points mean substantial upfront
costs must be paid for collecting training examples before any
automation can be tested in an online system.
In this paper, we explore an alternative approach of a crowdpowered system architecture that supports gradual automation
over time. In our approach, the crowd works with automated
components as they continue to improve, and the architecture
provides narrowly scoped points where automation can be
introduced successfully. For instance, instead of waiting until an automated dialog system is able to respond completely
on its own, one component that we developed recommends
responders from a large set of possible responders that might
be relevant based on the on-going conversation. Those responses are then among the options available to the crowd to
choose. Another component learns to help select high-quality
responses. Each problem is tightly scoped, and thus potentially
easier for machine learning algorithms to automate.
This paper introduces Evorus, a crowd-powered conversational
agent that provides a well-scoped path from crowd-powered
robustness to automated speed and frugality. Users can converse with Evorus in open domains, and the responses are
chosen from suggestions offered by crowd workers and any
number of automated systems that have been added to Evorus.
Evorus supports increased automation over time in three ways
(Figure 1): (i) allowing third-party developers to easily integrate automated chatterbots or task-oriented dialog systems
to propose response candidates, (ii) reusing crowd-generated
responses from previous conversations as response candidates,
and (iii) learning to automatically select high-quality response
candidates to reduce crowd oversight over time.
In Evorus, existing dialog systems can be incorporated via
simple REST (REpresentational State Transfer) interfaces that
take in the current conversation context, and respond with a
response candidate. Over time, Evorus learns to select a subset
of the automated components that are most likely to generate
high-quality responses for different context. The responses are
then forwarded to crowd workers as candidates. Workers then
choose which of the responses to present to the users. Evorus
sees workers selecting responses from candidates as signals
that enable it to learn to select both automated components
and response candidates in the future. It is important to note
that while Evorus is a functioning and deployed system, we
do not see the current version and its constituent components
to be final. Rather, its architecture is designed to allow future
researchers to improve on its performance and the extent to
which it is automated, by working on constituent problems,
which are each challenging in their own right. The structure of
Evorus provides distinct learning points that can be bettered by
other researchers. Others may include additional dialog systems or chatterbots, and improve upon its learning components,
driven by the collected data and its modular architecture.
We deployed the current version of Evorus over time to better
understand how well it works. During our deployment, automated response were chose 12.44% of time, Evorus reduced
the crowd voting by 13.81%, and the cost of each non-user
message reduced by 32.76%. In this paper, we explore when
the system was best able to automate itself, and present clear
opportunities for future research to improve on these areas.
This paper makes four primary contributions:
• Evorus Architecture: a crowd-powered conversational assistant that is designed to gradually automate itself over
time by including more responses from existent chatbots
and reduce the oversight needed from the crowd;
• Learning to Choose Chatbots Over Time: we introduced
a learning framework that uses crowd votes and prior accepted message to estimate the likelihood of each chatbots
when receiving a user message;
• Automatic Voting: we implemented a machine learning
model for automatically reducing the amount of crowd oversight needed, evaluated its performance on a dataset of real
conversations, and developed a mathematical framework to
estimate the expected reward of using the model; and
• Deployment: we deployed Evorus for over 5 months with
80 participants and 281 conversations to understand how
the automatic components we developed could gradually
take over from the crowd in a real setting.
RELATED WORK
Our work draws from research on conversational agents, general purpose dialog systems and crowd-machine systems.
General Purpose Dialog System: A number of general purpose dialog systems such as IRIS [2] have been proposed.
Wen et al. [43] designed a neural network language generation model for multi-domain dialog systems. A deeplearning-based domain adaptation model was also proposed
recently [11]. Project DialPort [50] introduced a multi-agent
framework that has the capability to include multiple taskoriented dialog systems to hold a multiple domain conversation. On the other hand, in the field of natural language
processing, general response generation technologies were
also developed. Ritter et al. [35] generated responses based on
phrase-based statistical machine translation based on Twitter
data. Li et al. [26] introduced a response generator based
on speaker model that encodes personas with background information and speaking style. Recently, researchers started
exploring end-to-end joint learning of language understanding
in dialogue systems [6, 48, 27]. However, after decades of
developments, sophisticated artificial “conversational intelligence” are largely absent in modern digital products.
Crowd-powered Conversational Agents: Building fullyautomated, open-domain conversational assistants is a widely
researched topic in the field of artificial intelligence (AI), but
has thus far remained an open challenge. In response to this,
the Chorus [25] conversational agent is powered by a crowd
of human actors, which enables it to work robustly across
domains [24]. To help users manage information and services
through crowd-powered conversational agents, Guardian takes
as input a Web API and a desired task from the user and the
crowd determines the parameters necessary to complete the
task [20, 18]; IntructableCrowd helps users automate the management of sensors and tasks in their mobile phones through
a conversational agent [16]; and WearMail enables users to
access their emails by talking to the crowd-powered assistant
via smartwatch [39]. Conversational assistants powered by
trained human operators, such as Facebook M [15] and Magic
Assistant [29], have emerged in the recent years.
While most crowd-powered conversational systems function
well in laboratory settings, Chorus was deployed in the real
world [19] and revealed a range of problems such as determining when to terminate a conversation or protecting workers
from abusive content introduced by end users. Microsoft
Tey [45] introduced an AI-powered agent which encountered
problems when deployed publicly, because some users realized that they could influence what Tey would say because
it mimicked them. Unlike Tey, Evorus does not learn only
by mimicking users, and paid crowd workers are kept in the
loop to verify responses in order to maintain quality. Our
deployment did not reveal such problems and we believe the
structure of Evorus makes such problems less likely.
Crowd-Machine Hybrid Systems: Crowd and machine hybrid systems have enabled us to solve a wide range of tasks
that were difficult for machines or humans to solve alone,
making impacts in areas including crowdsourcing, machine
learning, databases and computer vision [10, 21, 37, 34, 22].
For instance, Flock and Alloy [7, 5] use crowds to suggest
predictive features, label data, and weigh these features with
machine learning techniques to produce coherent categories
and accurate models. The Knowledge Accelerator [12] uses
crowds to synthesize such crowd-machine structures into coherent articles. Zensors creates custom computer vision sensors bootstrapped by the crowd [23]. Similarly, CrowdDB [10]
uses human input for providing information that is missing
from the database, for performing computationally difficult
functions, and for matching, ranking, or aggregating results
based on fuzzy criteria. JellyBean [37] introduces a suite of
crowd-vision hybrid counting algorithms that can perform in
independent or hybrid modes returning more accurate counts
that either workers or computer vision could do alone.
EVORUS’ CONVERSATIONAL ASSISTANT FRAMEWORK
Evorus obtains multiple responses from multiple sources, including crowd workers and chatbots, and uses a voting mechanism to decide which responses to send to the end-user.
Worker Interface
Evorus’ worker interface contains two major parts (Figure 2):
the chat box in the middle and the fact board on the side. Chat
box’s layout is similar to an online chat room. Crowd workers
can see the messages sent by the user and the responses candidates proposed by workers and bots. The role label on each
message indicates it was sent by the user (blue label,) a worker
(red label,) or a bot (green label.) Workers can click on the
check mark (4) to upvote on the good responses, click on the
cross mark (6) to downvote on the bad responses, or type text
to propose their own responses. Beside the chat box, workers
can use the fact board to keep track of important information
of the current conversation. To provide context, chat logs and
the recorded facts from previous conversations with the same
user were also shown to workers.
The score board on the upper right corner displays the current
reward points the worker have earned in this conversation. If
the conversation is over, the worker can click the long button
on the top of the interface to leave and submit this task.
Selecting Responses using Upvotes and Downvotes
Crowd workers and bots can upvote or downvote on a response
candidate. As shown in Figure 2, on the interface, the upvoted
responses turned to light green, and the downvoted responses
turned to gray. Crowd workers automatically upvote their own
candidates whenever they propose new responses. Upon calculating the voting results, we assigned a negative weight to a
downvote while an upvote have a positive weights. We empirically set the upvote weight at 1 and downvote’s weight at 0.5,
which encourages the system to send more responses to the
user. We inherited the already-working voting threshold from
deployed Chorus [19], which accepts a response candidate
when it accumulates a vote weight that is larger or equal to 0.4
times number of active workers in this conversation. Namely,
Evorus accepts a response candidate and sends it to the user
when Equation (1) holds:
#upvote ×Wupvote − #donwvote ×Wdownvote
> #active_workers × threshold
Wupvote = 1.0, Wdownvote = 0.5, threshold = 0.4
(1)
We formally defined the #active_workers in the later subsection of real-time recruiting. Evorus does not reject a message,
so it does not have a threshold for negative vote weight.
Expiring Unselected Messages to Refresh Context
When Evorus accepts a response, the system turns the accepted
message to a white background, and also expires all other
response candidates that have not been accepted by removing
them from the chat box in the worker interface. This feature
ensures all response candidates displayed on the interface
were proposed based on the latest context. We also created a
“proposed chat history” box on the left side of worker interface,
which automatically records the worker’s latest five responses.
Workers can copy his/her previously-proposed response and
send it again if the message expired too fast.
A Proposed, Accepted, or Expired Message
In Evorus, non-user messages are in one of three states: [Proposed], [Accepted], or [Expired]. [Proposed] messages are
open to be up/downvoted. These messages were proposed by
either a worker or a bot, has not yet received sufficient votes
to be accepted, and has not yet expired; [Accepted] messages
received sufficient votes before they expired and were sent
to the user; and [Expired] messages did not receive sufficient
votes before they expired. These messages were not sent to
the users, and were removed from the worker interface. A
[Rejected] state does not exist since Evorus does not reject a
message proactively.
Worker’s Reward Point System
To incentivize workers, Evorus grants reward points to workers for their individual actions such as upvoting on a message
or proposing a response candidate, and also for their collective
decisions such as agreeing on accepting a message or proposing a message that were accepted. The score box on the right
top corner of the interface shows the current reward points to
the worker in real-time. Reward points are later converted to
bonus pay for workers. Without compromising output quality,
if some of these crowd actions can be successfully replaced
by automated algorithms, the cost of each conversation can
be reduced. Evorus’ reward point schema was extended from
the Chorus reward schema, which was previously used during
Figure 2. The Evorus worker interface allows workers to propose responses and up/down vote candidate responses. The up/down votes give Evorus
labels to use to train its machine-learning system to automatically gauge the quality of responses. Evorus automatically expires response candidates
upon acceptance of another in order to prevent workers from voting through candidate responses that are good but no longer relevant. Workers can
tell each message is sent by the end-user (blue label), a worker (red label), or a chatbot (green label) by the colored labels.
its year-long deployment [19]. This schema encodes the importance of each action, and thus provides a good guide for
algorithms to estimate the benefit and risk when automating a
crowd action. Moreover, this reward schema will be used to
estimate the expected reward points (and corresponding costs)
an automatic voting bot can save, which we describe later.
votes. To coordinate with human workers’ speed, Evorus often
needs to set constraints on the frequency or capability (e.g.,
in which condition can a chatbot propose responses) of bots.
More importantly, Evorus can learn from the crowd feedback
to automate itself over time via three primary mechanisms: (i)
the chatbot selector, (ii) the retrieval-based chatbot that can
reuse old responses, and (iii) the automatic voting bot.
Real-time Recruiting & Connecting to Google Hangouts
When a conversation starts, Evorus uses the Ignition model
[17] to recruit workers quickly and economically from Amazon Mechanical Turk, and uses the Hangoutsbot [14] library
to connect with the Google Hangout servers so that users can
use its clients to talk with Evorus on computers or mobile
devices. Each conversation starts with 1 worker and incorporates 5 workers at most. Workers may reach a conversation at
different times, but typically stay to the end of the conversation (average duration ' 10 minutes). The #active_workers
in Equation 1 is defined as “the number of crowd workers
who were working on this conversation when the message was
proposed,” which varies as workers arrive (or drop out) at different times. In our deployment, the average #active_workers
of all crowd messages is 3.56 (SD=1.29,) and 77.56% of the
crowd messages had #active_workers >= 3.
EVORUS’ AUTOMATION AND LEARNING FRAMEWORK
Evorus is a conversational assistant that is collaboratively run
by real-time crowdsourcing and artificial intelligence. The
core concept of Evorus is to have crowd workers work with automated virtual agents (referred to as “bots”) on the fly to hold
sophisticated conversations with users. To test this, we developed two types of bots: the automatic response generators, i.e.,
chatbots, and the automatic voting algorithms, i.e., vote bots.
Evorus monitored all ongoing conversations, and periodically
called chatbots and vote bots to participate in active conversations. Both chatbots and vote bots take the entire chat log
as input, and based on the chat log to generate responses or
Part I: Learning to Choose Chatbots Over Time
In Evorus, existing dialog systems or chatbots can be incorporated by defining a simple REST interface on top of them
that accepts information about the current conversation state,
and responses with a suggested response. When a sufficient
amount of chatbots are included in the bot pool, selecting
the most appropriate chatbots to answer different questions
becomes critical. For instance, a simple “ping-pong” chatbot
that always responds with what it was told can be selected
to reply echo questions such as “Hi” or “How are you?”; a
restaurant recommendation bot can be selected when the user
is looking for food; and a chatbot that was built on a friend’s
chat log can be selected when the user feels lonely [31]. In
this paper, we introduce a learning framework that uses crowd
votes and prior accepted messages to estimate the likelihood
of each chatbot when receiving a user message. The learning
framework naturally assigns a slightly higher likelihood to
newly-added chatbots to collect more data. The beauty of this
design is that any chatbot can contribute, as long as it can
effectively respond to – even a small – set of user messages.
Part II: Reusing Prior Answers
Upon receiving a message from the user, Evorus uses a
retrieval-based approach to find the most similar message
in prior conversations and populates its prior response for the
crowd to choose from. By doing so, Evorus is capable to reuse
the answer of prior similar questions to respond to users. The
advantage of using a retrieval-based method is that it naturally
increases its capability of answering questions with the growth
of the collected conversations, without the need of recreation
or retraining of machine-learning models. With the oversight
of the crowd, the retrieval-based approaches also do not need
to be perfect to contribute. As long as it find good responses
to a portion of user conversations, the learning framework
described in Part I can gradually learn when to use it.
Part III: Automatic Voting
Closing the loop of automating the entire system, the last piece
is to automate the oversight of the crowd that are necessary
for quality control, i.e., the voting process in Evorus. We
formulated response voting as a classification task and tackled
it with a supervised machine-learning approach. A set of
features based on literature, including the word, the speaker,
and the time of the proposed messages are used to develop a
machine-learning model, and the prior collected crowd votes
are used as gold-standard labels. While the overall classifier
performance is efficient in the dataset, a misfired vote (a falsepositive) that mistakenly accepts a low-quality response will
not only disturb the conversation, but also waste extra bonus
money to crowd workers who proposed and voted for it. In
this paper we propose a mathematical framework to estimate
expected benefits of using an automatic voting classifier.
In Evorus, both workers and the vote bot can upvote suggested
responses. When a new suggestion is offered, the vote bot is
called. It first calculates its confidence score, and, if the confidence is greater than a threshold, which is estimated by our
proposed mathematical framework, the vote bot automatically
upvotes the message. Evorus monitors the latest down/upvotes
and calculates voting results in real-time. When a crowd
message collects sufficient vote weight, Evorus (i) accepts it
and sends it to the user, and (ii) removes all other candidate
messages from the worker interface to refresh context.
In the following three sections, we describe in detail the three
main parts of the Evorus framework.
PART I: LEARNING TO CHOOSE CHATBOTS OVER TIME
Evorus’ chatbot selector learns over time from the crowd’s
feedback to choose the right chatbots to respond to user messages. Evorus also regularly populates lower-ranking chatbots to allow the model to learn about new chatbots and tp
keep the model up-to-date.
Ranking and Sampling Chatbots
Upon receiving a message from a user, Evorus uses both the
text and prior collected data to estimate how likely each chatbot is capable of responding the user (i.e., P(bot|message)).
We used a conditional probability, as shown in Equation 2, to
characterize the likelihood of selecting a chatbot (bot) after
receiving a user message.
P(bot|message) = P(bot) × P(message|bot)
≈ P(bot) × similarity(message, historybot )
(2)
P(bot) is the prior probability of the chatbot, and
P(message|bot) is the likelihood of the user message given the
chatbot’s history (i.e., previous user messages that the bot has
successfully responded). While training an n-gram language
model using previous messages to estimate P(message|bot)
is intuitive [13], sufficient data for building such model is
often unavailable for newly-added bots. To generalize, we
used a similarity measure based on distance between word
vectors (similarity(message, historybot )) to approximate this
likelihood. We will explain how we calculate these two components in the following subsection.
Equipped with the estimates, Evorus ranks all the chatbots
based on the likelihood values, and always calls the firstranking chatbot to provide its response. More interestingly,
in addition to the top chatbot, Evorus also randomly selects
a lower-ranking chatbot to provide responses. By doing so,
Evorus is capable to gradually update its estimates of each
bot based on the crowd feedback and learn over time the best
scenario to call each chatbot. Similar strategies, such as the
epsilon-greedy strategy that yields a small portion of probability for random outcomes and collects feedback, have been
used in models that learn to select crowd workers [40] and
dialogue actions [38]. For the new chatbot, Evorus initially
assigns a starting probability to it to allow the system to collect
data about it, which we describe in the following subsection.
Estimating Likelihood of a Chatbot
We aimed at designing a learning framework that is (i) inexpensive to update, since we want the model to be updated
every single time when the system receives a new label, and
(ii) allows new bots to be added easily.
Prior Probability of Chatbots: To generate more reliable
prior estimation for newly-added bots with limited histories,
we used a beta distribution with two shape parameters α and
β (Equation 3) to model the prior probability of each chatbot.
P(bot) ≈
(#accepted messages from bot) + α
(#user messages since bot online) + α + β
(3)
P(bot) can be interpreted as the overall acceptance rate of
the chatbot without conversation contexts. The two shaping
parameters α and β can be viewed as the number of accepted
(positive) and not-accepted (negative) messages that will be assigned to each new chatbot to begin with, respectively. Namely,
any new chatbot’s prior probability P(bot) will be initially assigned as α/(α + β ), and then later be updated over time. The
beta distribution’s α and β are both functions of the mean (µ)
and variance (σ 2 ) of the distribution. In our pilot study, in
which each automatic response requires only one vote to be
accepted, four chatterbots had an average message acceptance
rate of 0.407 (SD=0.028.) Since we increased the required
vote count from 1 to 2 in the final deployment, a lower acceptance rate is expected. We used µ = 0.3 and σ = 0.05 to
estimate the shape parameters, where α = 24.9 and β = 58.1.
Similarity between Messages and Chatbots: To estimate
similarity(message, bot), we first used the pre-trained 200dimension GloVe word vector representation trained on
Wikipedia and Gigaword [32] to calculate the average word
vector of each message. We then used previous user messages
that were successfully responded by the chatbot as the bot vector ~wbot that represents the chatbot in the word-vector space.
We also calculated the centroid vector of all user messages,
~woverall , to represent general user messages. Finally, as shown
in Equation 4, the similarity between a message and a chatbot
is defined as the distance ratio between the vectors.
similarity(message, historybot )
(4)
dist(~wmessage ,~woverall )
B
dist(~wmessage ,~wbot ) + dist(~wmessage ,~woverall )
While ~wbot can be calculated as the centroid vector of prior
user messages that were successfully responded to by the chatbot, in cold-start scenarios, a chatbot will not have sufficient
accepted messages to calculate the vector. We provide two solutions for chatbot developers: First, the developer can provide
a small set of example messages where their chatbots should
be called. For instance, the developer of an Yelp chatbot can
list “Find me a sushi restaurant in Seattle!” as an example.
Evorus will treat these example messages as the user messages that the chatbot successfully responded to, and use their
centroid vector as the initial ~wbot . When more messages are
accepted, they will be added into this set and update the vector.
Second, for some chatbots, especially non-task chatterbots, it
could be difficult to provide a set of examples. Therefore, if
the developer decided not to provide any example messages,
we set the initial dist(~wmessage ,~wbot ) = 0 for new chatbots.
PART II: REUSING PRIOR RESPONSES
Evorus uses an information-retrieval-based (IR-based) method
to find answers to similar queries in prior conversations to
suggest as responses to new queries. To do so, Evorus first
extracts query-response pairs from all the old conversations,
and then performs a similarity-based sorting over these pairs.
Extracting Query-Response Pairs: One advantage Evorus
has is that each accepted response has crowd votes, which can
be used as a direct indicator of the response’s quality. For each
turn of a conversation between one user and Evorus, we extracted the accepted crowd response which did not receive any
downvotes, along with the user message (query) it responded
to, as a (query, response) pair. Since the deployed Evorus has
not had any prior conversation with users yet, we obtained
conversation data that were collected by the deployed Chorus, which also used crowd voting to select responses, during
May, 2016 to March, 2017 to start with. We further removed
the messages from known malicious workers and users, also
removed all conversations where the users are co-authors or
collaborators of Chorus [19]. At the beginning of the Evorus
deployment, 3,814 user messages were included, and each of
these user message is paired with 1 to 5 crowd responses.
Searching for the Most Similar Query: For the query message in each query-response pair, we calculated its average
word vector by using the pre-trained 200-dimension GloVe
word vector representation based on Wikipedia and Gigaword [32] and stores the vector in the database. When Evorus
receives a user message, the system first calculates its average
word vector ~wmessage using the same GloVe representation, and
then searches in the database to look for the top k responses
that their corresponding queries’ word vectors had the shortest Euclidean distances with ~wmessage . Finally, for increasing
answer’s diversity, the system randomly selects one from top
k responses to send back to Evorus for the crowd to choose
from. We empirically set k = 2 in our deployment.
PART III: AUTOMATIC VOTING
Evorus uses supervised learning to vote on responses.
Data Preparation: The voting mechanism has been proven
to be useful in selecting good responses and holding conversations in the lab prototype [25] and deployed system [19].
The final status (i.e., accepted or not) of a messages is a strong
signal to indicate its quality. However, when we used this data
to develop an AI-powered automated voting algorithm, it is
noteworthy that expired messages were not all of lower quality.
In some cases the proposed response was good and fitted in the
old context well, but the context changed shortly after the message was sent; Some messages were automatically accepted
and bypassed the voting process because Evorus does not have
enough active workers, i.e., when 0.4 × #active_workers < 1
in Equation 1); Furthermore, since downvotes can cancel out
upvotes, the voting results in Evorus could be influenced by
race conditions among workers. For instance, when two upvotes of a message has been sent to the server, Evorus might
decide that this message’ vote weight is sufficient and sent it
to the user, in which a belated downvote would deduct its vote
weight to lower than the threshold. Therefore, the training
data needs to be carefully developed.
Similar to Part II, we used voting data collected during the
Chorus deployment [19] to train the initial machine learning
model for voting. We first extracted the expired messages
with one or more downvote(s) as examples of “downvote”, and
extracted the accepted crowd messages with both one or more
upvote(s) and zero downvote as examples of “upvote.” We
excluded the automatically-accepted messages that were sent
when the task did not have sufficient number of active workers
(0.4 × #active_workers < 1) From all the messages collected
by the deployed Chorus during September 2016 to March
2017, 1,682 “upvote” messages 674 “downvote” messages
were extracted to from the dataset.
Model & Performance: We then used this dataset to train a
LibLinear [9] classifier. For each message, Evorus extracted
features in message, turn, and conversation levels to capture
the dialogic characteristics [33], and also used GloVe word
vector, which is identical as that of our retrieval-based response
generation, to represent the content. Our approach reached to
a precision of 0.740 and a recall of 0.982 (F1-score = 0.844)
on the “upvote” class in a 10-fold cross-validation experiment.
The feature analysis using the Weka toolkit [47] showed that
the top 3 features were about the historical performance of
the worker who proposed the message, and also 13 out of
20 top features were of a particular dimension of one of the
word vectors. On the other hand, the performance of the
“downvote” class is less effective. Its precision is 0.714 but
recall is only 0.134. This could be caused by the insufficient
amount of training data, since Chorus promoted upvote more
than downvote by design. According to this result, in the
deployed Evorus, the system only automatically upvoted when
the classifier ouput “upvote,” but did not downvote otherwise.
Optimizing Automatic Voting
Automatic voting directly participates in the process of deciding which messages to send. While our machine-learning
model resulted in good performance on our dataset, we would
upvoted message is sent ([Misfire]). Therefore, E[Bad] is decided by (i) how often one mistaken upvote triggers a misfire,
and (ii) how expensive is one misfire, as follows:
E[Bad] = P(Misfire|Bad) × E[RMisfire ]
P(Misfire|Bad) is the conditional probability of a message
being sent to the user given the classifier has mistakenly upvoted on it. E[RMisfire ] is the expected reward points that were
granted to workers in a single [Misfire] event.
Figure 3. (A) The precision-recall curve of the LibLinear classifier for
automatic upvoting. (B) Using our model (Equation (5)) to estimate the
precision and recall at different thresholds and their corresponding expected reward amount saved.
like to use Evorus’ worker reward point schema to find the
right confidence threshold for the automatic voting classifier.
If the threshold is set too low, the classifier would vote frequently even when it is not confident about the prediction,
and thus many low-quality responses would be accepted and
disturb the conversation; if the threshold is set too high, the
classifier would rarely vote, and the system will not gain much
from using it. Liblinear can output the probability estimates
of each class when performing prediction, which we used as
the notion of confidence of the classifier. Thresholding out
the predictions with lower confidences increased the precision
but reduced the recall of the classifier. Figure 3(A) shows the
Precision-Recall curve of Evorus’ upvoting classifier.
Possible Outcomes When the Classifier Upvotes: To find
confidence thresholds for Evorus, we introduced the following
heuristics to estimate reward points saved per message by
using the upvoting classifier. Consider the following cases:
1. [Good Vote] The classifier upvoted on a message that
would originally be selected by the crowd. It saves 1 upvote
reward (Rupvote × 1) and 1 agreement reward (Ragreement × 1)
that would originally be granted to one human worker.
2. [Bad Vote] The classifier upvoted on a message that would
originally not be selected by the crowd. In this case, one
of the two following consequences will occur: (i) [Misfire]
The message is sent to the user. The system grants agreement rewards to all human workers who upvoted on this
message (Ragreement × #upvoted_workers) and 1 successful
proposal reward to the worker who proposed the message
(R proposal × 1); and (ii): [No Difference] The message remains not sent. Even with one extra upvote, this message’s
vote count was still insufficient to get accepted.
Our training dataset only used the not-accepted messages with
at least one downvote to form the “Downvote” class, which
were less likely to be misfired after adding one extra automatic
vote. For better estimating P(Misfire|Bad), we first ran the
classifier on all messages that were not included in our training
set, and within all the messages that the classifier decided to
upvote, we then calculated the proportion of messages that
would be sent to the user if one extra upvote were added. The
rate was 0.692, which we used to approximate P(Misfire|Bad).
Furthermore, based on Evorus’ mechanism, E[RMisfire ] can be
calculated as follows :
E[RMisfire ] = Ragreement × E[#upvoted_workers] + R proposal
E[#upvoted_workers] is the expected number of human workers who upvoted on the message in an [Misfire] event. Similarly, we ran the classifier on the unlabelled data, and calculated the average number of workers who upvoted on the
messages that the classifier decided to upvote on. The number is 0.569 (SD = 0.731), which we used to approximate
E[#upvoted_workers].
Finally, E[Hit] = 100 + 500 = 600, and E[RBad ] = 0.692 ×
(500 × 0.569 + 1000) = 888.874. Using Equation (5), we can
estimate the precision and recall at different thresholds and
their corresponding reward amount (Figure 3(B)). According
to the estimation, the best confidence threshold is at 0.65. In
the deployed Evorus we selected a slightly higher precision
and set the threshold at 0.7 (P = 0.823 and R = 0.745.)
DEPLOYMENT STUDY AND RESULTS
T PR is the true positive rate and FPR is the false positive
rate of the classifier. E[Good] is the expected reward points
saved per [Good Vote] event, and E[Bad] is the expected reward points wasted per [Bad Vote] event.
Evorus was launched to the public as a Google Hangouts chatbot in March 2017. While the end-users were not aware of the
changes of the system from the client side, behind the scenes,
our deployment had 3 phases: (i) Phase 1, (ii) Control Phase,
and (iii) Phase 2. Phase-1 deployment started in March, 2017.
We launched the system with only four chatterbots and one
vote bot, without the learning component described in Part I,
to understand the basics of having virtual bots working with
human workers on the fly. For comparison, in May 2017 we
then temporarily turned off all automation components and
had the system solely run by the crowd till late August 2017,
which we referred to as the Control Phase. Finally, for testing
the capability of learning to select chatbots, we started Phase
2 deployment in early September. The Phase-2 deployment
included several significant changes: (i) increasing the frequency of calling chatbots for responses, (ii) increasing the
vote count needed to accept a response from 1 to 2, and (iii)
incorporating the Part I learning.
In Evorus, E[Good] is a constant (Rupvote + Ragreement ). Meanwhile, [Bad Vote] event costs reward points only when the
To recruit users, we periodically sent emails to mailing lists
at several universities and posted on social media sites, such
Estimating System’s Expected Gain: Given these setups,
the expected reward points E[Rsave ] saved per message by
using the classifier can be estimated as follows:
E[Rsave ] = T PR × E[Good] − FPR × E[Bad]
(5)
as Facebook and Twitter. Participants who volunteered to use
our system were asked to sign a consent form first, and no
compensation was offered. After the participants submitted
the consent form, a confirmation email was automatically sent
to them to instruct them how to send messages to Evorus via
Google Hangouts. The users can use Evorus as many times as
they want to, for anything, via any devices that are available
to them. Eighty users total talked with Evorus during 281
conversations. The Phase-1 deployment had 34 users talked
to Evorus during 113 conversations, and Phase-2 deployment
(till 17th September, 2017) had 26 users with 39 conversations.
The Control Phase had 42 users with 129 conversations.
Phase 1: Chatterbots & Vote bot
Our Phase-1 deployment explored how chatbots and our vote
bot could work together synchronously with crowd workers.
We implemented four chatterbots (including the IR-based chatterbot using Chorus conversation data described in Part II) and
a vote bot. During the Phase-1 deployment, the system only
randomly selects one of four chatterbots to respond every half
a minute, where the learning component described in Part I
will be later included in Phase-2 deployment.
Implementing Four Chatterbots: In our Phase-1 deployment, we implemented the following four chatterbots.
1. Chorus Bot (shown as Part II): A chatterbot that is powered by a retrieval-based method to reuse prior conversations to respond to users, which was described in Part II.
2. Filler Bot: A chatterbot that randomly selects one response
from a set of candidates, regardless of context. We manually
selected 13 common “conversation filler” in the Chorus
dataset (e.g., “Is there anything else I can help you with?”,
or “Thanks”) to form the candidate pool.
3. Interview Bot: A chatterbot that uses a retrieval-based
method, which is identical to Chorus Bot, to find the best
response from 27,894 query-response pairs extracted from
767 transcripts of TV celebrity interview [30].
4. Cleverbot: Cleverbot is a third-party AI-powered chatbot
which reuses more than 200 million conversations it had
with users to generate responses [44, 4].
Vote Bot Setup: Currrently, The vote bot only votes on
human-proposed messages, but not messages proposed by
chatbots. In Phase 1, Evorus requires each automatic vote to
have at least one extra human upvote to be accepted. Vote bots
also skipped messages if the same worker proposed identical
content earlier in the conversation but did not get accepted.
We believe that worker’s re-sending is a strong signal of the
poor quality of the message. Vote bots can decide not to vote
if the confidence is too low (Part III.)
Automating Human Labors: During Phase-1 deployment, a
conversation on average contained 9.90 user messages (SD
= 11.69), 13.6 accepted messages proposed by the crowd
workers (SD = 10.44), and 13.58 accepted messages proposed
by automatic chatterbots (SD = 2.81). Thus, automated responses were chosen 12.44% of the time. As a comparison
(Figure 4(A)), in the Control Phase (42 users and 129 conversations), a conversation on average contained 8.73 user
messages (SD = 10.05) and 12.98 accepted crowd messages
Figure 4. Phase-1 Deployment: (A) Average number of accepted messages per conversation. Automated responses were chosen 12.44% of
the time. (B) Average number of upvotes per accepted non-user message.
Human upvotes were reduced by 13.81% by using automatic voting.
(SD = 11.39). In terms of upvotes, each accepted non-user
message received 1.06 human upvotes (SD = 0.73) and 0.11
automatic upvotes (SD = 0.18) In comparison, in the Control
Phase, each accepted non-user message received 1.23 human
upvotes (SD = 0.70.) Crowd voting was thus reduced by
13.81%. The comparison is shown in Figure 4(B). Moreover,
an accepted non-user message sent by Evorus costed $0.142 in
Phase-1 deployment on average, while it costed $0.211 during
the Control Phase. Namely, with automated chatbots and the
vote bot, the cost of each message is reduced by 32.76%.
We also calculated the acceptance rate of messages proposed
by each chatbot. The Filler Bot, which ignores context and
proposes responses randomly, had the highest acceptance rate,
41.67%. The Chorus Bot’s acceptance rate was 30.99%, that
of the Interview Bot was 33.33%, and that of the Cleverbot
was 30.99%. This might be because Filler Bot’s commonplace responses (e.g., “I don’t know”) were often considered
acceptable by human raters. In Phase 2, where an automatic
response needed 2 human votes to go through, the Chorus
Bot had the highest acceptance rate among all four chatterbots
(Figure 6). While human workers, whose acceptance rate was
72.04% during Phase 1, still outperformed all chatbots by a
large margin, chatbots with low accuracy can still contribute to
the conversation. For instance, the Filler Bot, while being very
simple and ignoring any context, nevertheless, often produces
reasonable responses:
[The user asked information about the wildfire and smoke in
Emory university campus.]
user Do you know where they are happening exactly? (The wildfires I
mean)
bot Can you provide some more details?
Compared with Filler Bot, Chorus Bot better targets its responses because it chooses messages based on similarity with
previous human responses:
user Hey how many people like bubble tea here?
bot Ask for their feedback when you talk with them
Conversation Quality: We sampled conversations with accepted automatic responses and a matching set without automated contributions. For each, 8 MTurk workers rated [Satisfaction, Clarity, Responsiveness, Comfort], which was based
on the PARADISE’s objectives for evaluating dialogue performance [42] and the Quality of Communication Experience
metric [28], on a 5-point Likert scale (5 is best.) The original
conversations (N=46) had an average rating of [3.47, 4.04,
3.88, 3.56], while those with automatic responses (N=54) had
[3.57, 3.74, 3.52, 3.66]. The similar results suggest that the
automatic components did not make conversations worse.
Figure 5. (A) An actual conversation of Evorus. Conversations in Evorus tend to combine multiple chatbots and workers together. (B) User questionnaire used in Phase-2 deployment. The average user satisfaction rating of automated and non-automated conversations had no significant difference.
Phase 2: Learning to Select Chatbots
Our Phase-2 deployment explored how the learning component, described in Part I, can select the right chatbots in context,
and how integrating additional chatbots affects performance.
We implemented two additional utility bots, a Yelp Bot and
a Weather Bot, that can perform information inquiry tasks
for different contexts, in addition to the four chatterbots and
one vote bot from Phase 1. We first launched the learning
system with four chatterbots for two days, and then added
the two utility bots for observing the changes of the model.
Furthermore, in order to directly compare user satisfaction
levels, all the automated components (including the chatbots,
the vote bot, and the learning component) were only applied
to 50% of the conversations randomly, and the other half of
the conversations were solely run by the crowd as a baseline.
To efficiently collect crowd feedbacks to update our model, we
increased the frequency Evorus called chatbots from randomly
calling one chatbot and one vote bot every 30 seconds (Phase
1) to calling two chatbots (top-1 ranked plus random) and one
vote bot every 10 seconds (Phase 2.) To compensate for the
possible drop in quality caused by higher calling frequency
and randomly selecting one of the two bots, we increased the
required upvote count for accepting an automatic response
from 1 vote to 2 votes. Namely, while Evorus obtained more
automatic responses with a much higher frequency in Phase 2,
it also required more human upvotes to approve each automatic
response at the same time. As a result, among the conversations that had automation in Phase 2, automated responses
were chosen 13.25% of the time, in which a conversation on
average contained 10.68 user messages (SD = 8.90), 15.74 accepted messages proposed by the crowd workers (SD = 11.92),
and 2.40 accepted messages proposed by automatic chatterbots (SD = 2.40). Each accepted non-user message received
1.90 human upvotes (SD = 1.13) and 0.30 automatic upvotes
(SD = 0.22). We did not compare Phase 2’s results in detail to
that of the Control Phase because the high-frequency setup of
Phase 2 is primarily for experimental exploration.
Implementing Two Utility Bots: In addition to the four chatterbots in the Phase-1 deployment, we implemented the following two task-oriented utility bots:
1. Yelp Bot: A chatbot that suggests restaurants near the location mentioned by the user powered by the Yelp API [49]. If
the user did not mention any location, it replies with “You’re
looking for a restaurant. What city are you in?”.
2. Weather Bot: A chatbot that reports the current weather
of a mentioned city powered by the WeatherUnderground
API [41]. If the user did not mention any city names, it
replies, “Which city’s weather would you like to know?”
The chatbots’ developer (the first author) also provided three
example chat messages for each that should trigger the corresponding chatbot, to initiate the learning process. For example,
“Any restaurant recommendations in NYC?” for the Yelp Bot.
Similar User Satisfaction Level: In Phase 2 we implemented
an exit survey to measure end user satisfaction (Figure 5(B)).
At the end of each conversation, the user had 10 minutes to report their satisfaction using a Likert-scale ranging from 1 (Very
Dissatisfied) to 5 (Very Satisfied). 13 of 30 users provided
feedback (response rate = 43%). The automated conversations’
average user satisfaction rating was 4.50 (SD=0.5, N=4); the
crowd conversations’ average user satisfaction rating was 4.00
(SD=0.47, N=9), a difference that was not significant.
Updating Estimates of Chatbot’s Prior Over Time: While
our deployment is of a medium scale, the dynamics of our likelihood model can still be observed. For instance, the estimated
prior probability described in Equation 3 was continuously
updated with the growth of conversation that Evorus had. Our
model assigned a starting probability of 0.3 to each chatbots
(Figure 6). When users started talking with Evorus, crowd
workers provided their feedback by upvoting and downvoting,
and thus changed the estimation over time. When new chatbots were added, Evorus intentionally assigned them a higher
prior probabilities to allow quicker crowd feedback.
user hi! Can you summarize the features of the new iPhone for me?
[Multiple messages list the features of the iPhone X.]
bot There is no iPhone 7.
[The crowd lists more features, and the user says thank you.]
crowd No problem!
crowd Some Auto replies don’t even make sense
DISCUSSION AND CONCLUSION
Figure 6. The estimated prior probability (Equation 3) of each chatbot
was continuously updated with the growth of user messages.
Utility Bots in Cold-Start Scenarios: We would like to understand if Evorus can select appropriate chatbots to obtain
responses in corresponding context. Since non-task chatterbots such as the Cleverbot could be difficult for humans to
judge if it should be called given a message, we focused only
on task-oriented utility bots in the evaluation. For each user
message in the automated conversations in Phase 2, the researchers manually annotated if it is relevant to the topic of
“weather” and “restaurant,” respectively. With the assumption
that each utility bot should be called when its topic comes
up, we compared the human-labelled topic against the top
chatbot that were suggested by Evorus and calculated the precision, recall, and F1-score. It is noteworthy that we only
evaluated when the appropriate chatbot was called, regardless
of the quality of the responses it generated. As a result, two
newly-added utility bots both had a high precision and a lower
recall. The Weather Bot’s precision was 1.00 and the recall
was 0.47 (F1=0.64); and the Yelp Bot’s precision was 0.67 and
the recall was 0.20 (F1=0.31.) This result shows the nature of
new bots in our learning framework: Evorus uses the expertgenerated small set of examples to calculate the initial vector
of each chatbot, which could result in precise predictions but
with lower coverage. Over time, when Evorus collects more
examples that each bot has successfully responded to and the
recall increases. Moreover, conversations in Evorus tend to
combine multiple chatbots and workers together. For instance,
in the following conversation, the crowd had the user to narrow
down the query, and then the Weather Bot was able to answer
it. An additional detailed example is shown in Figure 5(A).
[The user asked about the weather in Afghanistan.]
crowd What city in Afghanistan?)
[The crowd sends a website about Afghanistan weather.]
user Kabul
bot Friday’s weather forecast for [Kabul, Afghanistan]: Cloudy
with a few showers. High 79F. Winds NW at 5 to 10 mph.
Chance of rain 30%.
Error Cases: Our bot-selection algorithm starts with high
precision and low recall, and increases recall as it gradually
gathers examples. Therefore, most errors we observed were
false-negatives, where a chatbot should have been triggered
but was not. Other errors came from the chatbot, where the bot
was correctly triggered but its response was invalid. Workers
usually downvoted or ignored these suggestions. In the rare
cases where invalid automatic responses were mistakenly sent
to the user, the crowd often tried to explain how the response
was automatically generated to the user afterward.
We introduced Evorus, a crowd-powered system conversational assistant built to automate itself over time. Informed by
two phases of public field deployment and testing with real
users, we iteratively designed and refined its flexible framework for open-domain dialog. We imagine a future where
thousands of online service providers can develop their own
chatbots, not only to serve their own users in a task-specific
context, but also to dynamically integrate their services into
Evorus with the help of the crowd, allowing users to interact freely with thousands of online services via a universal
portal. Supporting this scale offers opportunities for future
research. For example, one direction is to improve the learning
framework to support third-party chatbots that also improves
overtime, or to better balance between the exploitation and
exploration phases (like in a multi-armed bandit problem).
Evorus could also be used to collect valuable fail cases to
enable third-party developers to improve their bots (i.e., when
a bot was triggered, but its proposed response was rejected).
Evorus has three main advantages as compared to previous
approaches. First, it is a working system that can serve as a
scaffold for automation over time. A core advantage of starting
with a working system is that users can talk to Evorus naturally
from day one, ensuring conversation quality while collecting
training data for automation. Second, given the oversight of
the crowd, Evorus has a high tolerance for errors from its
automated components. Even an imperfect automation component (e.g., chatbots) can contribute to a conversation without
hurting quality, which yields more space for algorithms to “explore” different actions (e.g., selecting a chatbot with medium
confidence.) Finally, Evorus allows a mixed group of humans
and bots to collaboratively hold open conversations.
Most automated systems created from crowd work simply use
the crowd for data; Evorus tightly integrates crowds and machine learning, and provides specific points where automated
components can be introduced. This architecture allows each
component to be improved, providing a common research
harness on which researchers specializing in different areas
may innovate and compete. For instance, “response generation” has long been developed in the NLP community; Evorus
provides a natural evaluate it within a larger conversational
system. The flexibility of the Evorus framework potentially
allows for low cost integration between many online service
providers and fluid collaboration between chatbots and human workers to form a single user-facing identity. Given the
complexity of conversational assistance, Evorus is likely to
be crowd-powered in part for some time, but we expect it to
continue to increasingly rely on automation.
ACKNOWLEDGEMENTS
We thank Walter S. Lasecki and also the workers who participated in our studies. This project has been funded as part of the
Yahoo!/Oath InMind Project at Carnegie Mellon University.
REFERENCES
1. Amazon. 2017. Meet Alexa. (2017).
https://www.amazon.com/meet-alexa/b?ie=
UTF8&node=16067214011
2. Rafael E Banchs and Haizhou Li. 2012. IRIS: a
chat-oriented dialogue system based on the vector space
model. In Proceedings of the ACL 2012 System
Demonstrations. Association for Computational
Linguistics, 37–42.
3. Jeffrey P. Bigham, Chandrika Jayant, Hanjie Ji, Greg
Little, Andrew Miller, Robert C. Miller, Robin Miller,
Aubrey Tatarowicz, Brandyn White, Samual White, and
Tom Yeh. 2010. VizWiz: Nearly Real-time Answers to
Visual Questions. In Proceedings of the 23Nd Annual
ACM Symposium on User Interface Software and
Technology (UIST ’10). ACM, New York, NY, USA,
333–342. DOI:
http://dx.doi.org/10.1145/1866029.1866080
4. Rollo Carpenter. 2006. Cleverbot. (2006).
https://www.cleverbot.com/ [Online; accessed
08-March-2017].
5. Joseph Chee Chang, Aniket Kittur, and Nathan Hahn.
2016. Alloy: Clustering with crowds and computation. In
Proceedings of the 2016 CHI Conference on Human
Factors in Computing Systems. ACM, 3180–3191.
6. Yun-Nung Chen, Dilek Hakkani-Tür, Gökhan Tür,
Jianfeng Gao, and Li Deng. 2016. End-to-End Memory
Networks with Knowledge Carryover for Multi-Turn
Spoken Language Understanding.. In INTERSPEECH.
3245–3249.
7. Justin Cheng and Michael S Bernstein. 2015. Flock:
Hybrid crowd-machine learning classifiers. In
Proceedings of the 18th ACM Conference on Computer
Supported Cooperative Work & Social Computing. ACM,
600–611.
8. Josh Constine. 2017. Amazon rejects AI2’s Alexa skill
voice-search engine. Will it build one? (May 2017).
https://techcrunch.com/2017/05/31/amazon-skillsearch-engine/
9. Rong-En Fan, Kai-Wei Chang, Cho-Jui Hsieh, Xiang-Rui
Wang, and Chih-Jen Lin. 2008. LIBLINEAR - A Library
for Large Linear Classification. (2008).
http://www.csie.ntu.edu.tw/~cjlin/liblinear/ The
Weka classifier works with version 1.33 of LIBLINEAR.
10. Michael J. Franklin, Donald Kossmann, Tim Kraska,
Sukriti Ramesh, and Reynold Xin. 2011. CrowdDB:
Answering Queries with Crowdsourcing. In Proceedings
of the 2011 ACM SIGMOD International Conference on
Management of Data (SIGMOD ’11). ACM, New York,
NY, USA, 61–72. DOI:
http://dx.doi.org/10.1145/1989323.1989331
11. Milica Gasic, Nikola Mrksic, Lina M Rojas-Barahona,
Pei-Hao Su, Stefan Ultes, David Vandyke, Tsung-Hsien
Wen, and Steve Young. 2016. Dialogue manager domain
adaptation using Gaussian process reinforcement
learning. arXiv preprint arXiv:1609.02846 (2016).
12. Nathan Hahn, Joseph Chang, Ji Eun Kim, and Aniket
Kittur. 2016. The Knowledge Accelerator: Big picture
thinking in small pieces. In Proceedings of the 2016 CHI
Conference on Human Factors in Computing Systems.
ACM, 2258–2270.
13. Bo Han and Timothy Baldwin. 2011. Lexical
Normalisation of Short Text Messages: Makn Sens a
#Twitter. In Proceedings of the 49th Annual Meeting of
the Association for Computational Linguistics: Human
Language Technologies - Volume 1 (HLT ’11).
Association for Computational Linguistics, Stroudsburg,
PA, USA, 368–378.
http://dl.acm.org/citation.cfm?id=2002472.2002520
14. Hangoutsbot. 2017. hangoutsbot/hangoutsbot. (Apr
2017). https://github.com/hangoutsbot/hangoutsbot
15. Jessi Hempel. 2015. Facebook Launches M, Its Bold
Answer to Siri and Cortana. (Aug 2015).
https://www.wired.com/2015/08/facebook-launchesm-new-kind-virtual-assistant/
16. Ting-Hao Kenneth Huang, Amos Azaria, and Jeffrey P.
Bigham. 2016. InstructableCrowd: Creating IF-THEN
Rules via Conversations with the Crowd. In Proceedings
of the 2016 CHI Conference Extended Abstracts on
Human Factors in Computing Systems (CHI EA ’16).
ACM, New York, NY, USA, 1555–1562. DOI:
http://dx.doi.org/10.1145/2851581.2892502
17. Ting-Hao K. Huang and Jeffrey P. Bigham. 2017. A
10-Month-Long Deployment Study of On-Demand
Recruiting for Low-Latency Crowdsourcing. In In
Proceedings of The fifth AAAI Conference on Human
Computation and Crowdsourcing (HCOMP 2017). AAAI,
AAAI.
18. Ting-Hao K. Huang, Yun-Nung Chen, and Jeffrey P.
Bigham. 2017. Real-time On-Demand Crowd-powered
Entity Extraction. In In Proceedings of the 5th Edition Of
The Collective Intelligence Conference (CI 2017, oral
presentation).
19. Ting-Hao Kenneth Huang, Walter S. Lasecki, Amos
Azaria, and Jeffrey P. Bigham. 2016. “Is there anything
else I can help you with?”: Challenges in Deploying an
On-Demand Crowd-Powered Conversational Agent. In
Proceedings of AAAI Conference on Human Computation
and Crowdsourcing 2016 (HCOMP 2016). AAAI.
20. Ting-Hao Kenneth Huang, Walter S Lasecki, and
Jeffrey P Bigham. 2015. Guardian: A Crowd-Powered
Spoken Dialog System for Web APIs. In Third AAAI
Conference on Human Computation and Crowdsourcing.
21. Ece Kamar, Severin Hacker, and Eric Horvitz. 2012.
Combining human and machine intelligence in
large-scale crowdsourcing. In Proceedings of the 11th
International Conference on Autonomous Agents and
Multiagent Systems-Volume 1. International Foundation
for Autonomous Agents and Multiagent Systems,
467–474.
22. Ece Kamar and Lydia Manikonda. 2017. Complementing
the Execution of AI Systems with Human Computation.
In AAAI Workshop on Crowdsourcing, Deep Learning
and Artificial Intelligence Agents 2017. AAAI.
23. G. Laput, W. S. Lasecki, J. Wiese, R. Xiao, J. P. Bigham,
and C. Harrison. 2015. Zensors: Adaptive, Rapidly
Deployable, Human-Intelligent Sensor Feeds. In
Proceedings of the SIGCHI Conference on Human
Factors in Computing Systems (CHI ’15). ACM, New
York, NY, USA, 10. http://www.cs.cmu.edu/~jbigham/
pubs/pdfs/2015/zensors.pdf
24. Walter S. Lasecki, Phyo Thiha, Yu Zhong, Erin Brady,
and Jeffrey P. Bigham. 2013a. Answering Visual
Questions with Conversational Crowd Assistants. In
Proceedings of the 15th International ACM SIGACCESS
Conference on Computers and Accessibility (ASSETS
’13). ACM, New York, NY, USA, Article 18, 8 pages.
DOI:http://dx.doi.org/10.1145/2513383.2517033
25. Walter S. Lasecki, Rachel Wesley, Jeffrey Nichols,
Anand Kulkarni, James F. Allen, and Jeffrey P. Bigham.
2013b. Chorus: A Crowd-powered Conversational
Assistant. In Proceedings of the 26th Annual ACM
Symposium on User Interface Software and Technology
(UIST ’13). ACM, New York, NY, USA, 151–162. DOI:
http://dx.doi.org/10.1145/2501988.2502057
26. Jiwei Li, Michel Galley, Chris Brockett, Jianfeng Gao,
and Bill Dolan. 2016. A Persona-Based Neural
Conversation Model. arXiv preprint arXiv:1603.06155
(2016).
27. Xuijun Li, Yun-Nung Chen, Lihong Li, Jianfeng Gao,
and Asli Celikyilmaz. 2017. End-to-End
Task-Completion Neural Dialogue Systems. In
Proceedings of The 8th International Joint Conference on
Natural Language Processing (IJCNLP 2017). AFNLP.
28. Leigh Anne Liu, Chei Hwee Chua, and Günter K Stahl.
2010. Quality of communication experience: definition,
measurement, and implications for intercultural
negotiations. Journal of Applied Psychology 95, 3 (2010),
469.
29. Matthew Lynley. 2016. Make Magic’s Assistants Do
Almost Anything With $100/Hour And A Text Message.
(Jan 2016). https://techcrunch.com/2016/01/05/
make-magics-assistants-do-almost-anything-with100hour-and-a-text-message/
30. Cable News Network. 2017. (2017).
http://transcripts.cnn.com/TRANSCRIPTS/
31. Casey Newton. 2016 (accessed October 24th, 2016).
SPEAK, MEMORY: When her best friend died, she rebuilt
him using artificial intelligence.
https://www.theverge.com/a/luka-artificialintelligence-memorial-roman-mazurenko-bot
32. Jeffrey Pennington, Richard Socher, and Christopher D.
Manning. 2014. GloVe: Global Vectors for Word
Representation. In Empirical Methods in Natural
Language Processing (EMNLP). 1532–1543.
http://www.aclweb.org/anthology/D14-1162
33. Antoine Raux and Maxine Eskenazi. 2008. Optimizing
endpointing thresholds using dialogue features in a
spoken dialogue system. In Proceedings of the 9th
SIGdial Workshop on Discourse and Dialogue.
Association for Computational Linguistics, 1–10.
34. Daniela Retelny, Sébastien Robaszkiewicz, Alexandra To,
Walter S Lasecki, Jay Patel, Negar Rahmati, Tulsee
Doshi, Melissa Valentine, and Michael S Bernstein. 2014.
Expert crowdsourcing with flash teams. In Proceedings of
the 27th annual ACM symposium on User interface
software and technology. ACM, 75–85.
35. Alan Ritter, Colin Cherry, and William B Dolan. 2011.
Data-driven response generation in social media. In
Proceedings of the conference on empirical methods in
natural language processing. Association for
Computational Linguistics, 583–593.
36. Erica Sadun and Steve Sande. 2012. Talking to Siri:
Learning the Language of Apple’s Intelligent Assistant.
Que Publishing.
37. Akash Das Sarma, Ayush Jain, Arnab Nandi, Aditya
Parameswaran, and Jennifer Widom. 2015. Surpassing
humans and computers with JELLYBEAN:
Crowd-vision-hybrid counting algorithms. In Third AAAI
Conference on Human Computation and Crowdsourcing.
38. Konrad Scheffler and Steve Young. 2002. Automatic
Learning of Dialogue Strategy Using Dialogue
Simulation and Reinforcement Learning. In Proceedings
of the Second International Conference on Human
Language Technology Research (HLT ’02). Morgan
Kaufmann Publishers Inc., San Francisco, CA, USA,
12–19.
http://dl.acm.org/citation.cfm?id=1289189.1289246
39. Saiganesh Swaminathan, Raymond Fok, Fanglin Chen,
Ting-Hao K. Huang, Irene Lin, Rohan Jadvani, Walter
Lasecki, and Jeffrey Bigham. 2017. WearMail:
On-the-Go Access to Information in Your Email with a
Privacy-Preserving Human Computation Workflow. In
30th ACM Symposium on User Interface Software and
Technology (UIST 2017).
40. Long Tran-Thanh, Sebastian Stein, Alex Rogers, and
Nicholas R Jennings. 2014. Efficient crowdsourcing of
unknown experts using bounded multi-armed bandits.
Artificial Intelligence 214 (2014), 89–111.
41. Weather Underground. 2017. A Weather API Designed
for Developers. (2017).
https://www.wunderground.com/weather/api/
42. Marilyn A Walker, Diane J Litman, Candace A Kamm,
and Alicia Abella. 1997. PARADISE: A framework for
evaluating spoken dialogue agents. In Proceedings of the
eighth conference on European chapter of the
Association for Computational Linguistics. Association
for Computational Linguistics, 271–280.
43. Tsung-Hsien Wen, Milica Gasic, Nikola Mrksic, Lina M
Rojas-Barahona, Pei-Hao Su, David Vandyke, and Steve
Young. 2016. Multi-domain Neural Network Language
Generation for Spoken Dialogue Systems. arXiv preprint
arXiv:1603.01232 (2016).
44. Wikipedia. 2017a. Cleverbot — Wikipedia, The Free
Encyclopedia.
http://en.wikipedia.org/w/index.php?title=
Cleverbot&oldid=771836990. (2017). [Online; accessed
02-April-2017].
45. Wikipedia. 2017b. Tay (bot) — Wikipedia, The Free
Encyclopedia.
http://en.wikipedia.org/w/index.php?title=
Tay%20(bot)&oldid=769762463. (2017). [Online;
accessed 04-April-2017].
46. Jason D Williams and Steve Young. 2007. Partially
observable Markov decision processes for spoken dialog
systems. Computer Speech & Language 21, 2 (2007),
393–422.
47. Ian H Witten, Eibe Frank, Mark A Hall, and
Christopher J Pal. 2016. Data Mining: Practical machine
learning tools and techniques. Morgan Kaufmann.
48. Xuesong Yang, Yun-Nung Chen, Dilek Hakkani-Tür,
Paul Crook, Xiujun Li, Jianfeng Gao, and Li Deng. 2017.
End-to-end joint learning of natural language
understanding and dialogue manager. In Acoustics,
Speech and Signal Processing (ICASSP), 2017 IEEE
International Conference on. IEEE, 5690–5694.
49. Yelp. 2017. Yelp API Documentation. (2017).
https://www.yelp.com/developers/documentation/
v2/overview
50. Tiancheng Zhao, Kyusong Lee, and Maxine Eskenazi.
2016. DialPort: Connecting the Spoken Dialog Research
Community to Real User Data. arXiv preprint
arXiv:1606.02562 (2016).
| 2cs.AI
|
Property-driven State-Space Coarsening for
Continuous Time Markov Chains ?
Michalis Michaelides1 , Dimitrios Milios1 ,
Jane Hillston1 , and Guido Sanguinetti12
arXiv:1606.01111v2 [cs.SY] 29 Oct 2016
2
1
School of Informatics, University of Edinburgh
SynthSys, Centre for Synthetic and Systems Biology, University of Edinburgh
Abstract. Dynamical systems with large state-spaces are often expensive to thoroughly explore experimentally. Coarse-graining methods aim
to define simpler systems which are more amenable to analysis and exploration; most current methods, however, focus on a priori state aggregation based on similarities in transition rates, which is not necessarily
reflected in similar behaviours at the level of trajectories. We propose
a way to coarsen the state-space of a system which optimally preserves
the satisfaction of a set of logical specifications about the system’s trajectories. Our approach is based on Gaussian Process emulation and
Multi-Dimensional Scaling, a dimensionality reduction technique which
optimally preserves distances in non-Euclidean spaces. We show how
to obtain low-dimensional visualisations of the system’s state-space from
the perspective of properties’ satisfaction, and how to define macro-states
which behave coherently with respect to the specifications. Our approach
is illustrated on a non-trivial running example, showing promising performance and high computational efficiency.
1
Introduction
Reasoning about behavioural properties of dynamical systems is a central goal
of formal modelling. Recent years have witnessed considerable progress in this
direction, with the definition of formal languages [10,9] and logics [12] which
enable compact representations of dynamical systems, and mature reasoning
tools to model-check properties in an exact [15] or statistical way [20,14].
While such advances are indubitably improving our understanding of dynamical
systems, the applicability of these techniques in practical scenarios is still largely
hindered by computational issues. In particular, systems with large state-spaces
quickly become infeasible to analyse via exact methods due to the phenomenon
of state-space explosion; even statistical methods may require computationally
expensive and extensive simulations. State-space reduction methodologies aim
?
M.M., D.M. and G.S. are supported by the European Research Council under grant
MLCS 306999. J.H. is supported by the EU project, QUANTICOL 600708.
to construct more compact representations for complex systems. Such reducedstate systems are generally amenable to more effective analysis and may yield
deeper insights into the structure and dynamics of the system.
Broadly speaking, state-space reduction can be achieved by either model simplification, usually by abstracting some system behaviours into a simpler system,
or state aggregation, often by exploiting symmetries or approximate invariances.
A prime example of model simplification is the technique of time-scale separation, which replaces a large system with multiple weakly dependent sub-systems
[5]. Most aggregation methods, instead, are based on grouping different states
with similar behaviour with respect to their transition probabilities. This idea is
at the core of the concept of approximate lumpability, which extends the exact
lumpability relationship by aggregating states based on a pre-defined metric on
the outgoing exit rates [11,7,1,19,17].
In this paper we propose a novel state-space reduction paradigm by shifting the
focus from the infinitesimal properties of states (i.e. their transition rates) to
the global properties of trajectories. Namely, we seek to aggregate states that
yield behaviourally similar trajectories according to a set of pre-defined logical
specifications. Intuitively, two states will be aggregated if trajectories starting
from either state exhibit similar probabilities of satisfying the logical specifications. We define a statistical algorithm based on statistical model checking
and Gaussian Process emulation to define this behavioural similarity across the
whole state-space of the system. We then propose a dimensionality reduction and
clustering pipeline to aggregate states and define reduced (non-Markovian) dynamics. To illustrate our approach, we give a running example of model reduction
for the Susceptible-Infected-Recovered-Susceptible (SIRS) model, a non-trivial,
non-linear stochastic system widely used in epidemiology. Our results show that
property-driven aggregation can yield an effective tool to reduce the complexity
of stochastic dynamical systems, leading to non-trivial insights in the structure
of their state-space.
2
2.1
Background
Population Continuous Time Markov Chains
A Continuous Time Markov Chain (CTMC) is a continuous-time Markovian
stochastic process over a discrete state-space S. We will consider only population models, where the state-space is organised along populations: in this case,
the state-space is indexed by the counts of each population ni ∈ N ∪ {0}. Population CTMCs (pCTMCs) are frequently used in many scientific and engineering
domains; we will use here the notation of chemical reactions as it is widespread
and intuitively appealing. Transitions in a pCTMC are denoted as
τ (X)
r1 X1 + . . . rn Xn −−−→ s1 X1 + . . . sn Xn
2
meaning that ri particles of type Xi are consumed and sj particles of type Xj
are produced when the specific transition takes place. τ (X) is a transition rate
which depends on the current state of the system.
It is easy to show that waiting times between transitions are exponentially distributed random variables; this observation is the basis of exact simulation algorithms for pCTMCs, such as the celebrated Gillespie algorithm [13]. The Gillespie algorithm generates trajectories of a pCTMC by randomly choosing the
next reaction to occur and the time to elapse until the reaction occurs.
Example 1.1 We introduce here our running example, the Susceptible-InfectedRecovered-Susceptible (SIRS) model of epidemic spreading. The SIRS model is a
discrete stochastic model of disease spread in a population, where individuals in
the population can be in one of three states, Susceptible, Infected and Recovered.
There are different variations of the model, some open (individuals can enter and
exit the system), others with individuals relapsing to a susceptible state after
having recovered. Here, we consider a relapsing, closed system, which evolves
in a discrete, 2-dimensional state-space, where dimensions are the number of
Susceptible and Infected individuals in the population (Recovered numbers are
uniquely determined since the total population is constant). We also introduce a
spontaneous infection of a susceptible individual with constant rate, independent
of the number of infected individuals, to eliminate absorbing states.
With a population size of N , states in the 2D space can be represented by
x = (S, I), S ∈ {0, · · · , N }, I ∈ {0, · · · , N − S} for a total of (N + 1)(N + 2)/2
states. The chemical reactions for this system are:
α
infection S + I −
→ 2I;
β/5
spontaneous infection S −−→ I;
β
recovery I −
→ R;
β
relapsing R −
→ S.
We set the infection rate α = 0.005, recovery rate β = 0.01, and population size
N = S + I + R = 100, for a total of 5151 states in this SIRS system. Sample
trajectories of the system were simulated using the Gillespie algorithm.
2.2
Temporal Logic and Model Checking
We formally specify trajectory behaviours by using temporal logic properties. We
are particularly interested in properties that can be verified on single trajectories,
and assume metric bounds on the trajectories, so that they are observed only
for a finite amount of time. Metric Interval Temporal logic (MITL) offers a
convenient way to formalise such specifications.
3
Formally, MITL has the following grammar:
φ ::= tt | µ | ¬φ | φ1 ∧ φ2 | φ1 U[T1 ,T2 ] φ2 ,
where tt is the true formula, conjunction and negation are the standard boolean
connectives, and the time-bounded until U[T1 ,T2 ] is the only temporal modality. Atomic propositions µ are (non-linear) inequalities on population variables.
A MITL formula is interpreted over a function of time x, and its satisfaction relation is given as in [16]. More temporal modalities, such as the timebounded eventually and always, can be defined in terms of the until operator:
F[T1 ,T2 ] φ ≡ ttU[T1 ,T2 ] φ and G[T1 ,T2 ] φ ≡ ¬F[T1 ,T2 ] ¬φ.
MITL formulae evaluate as true or false on individual trajectories; when trajectories are sampled from a stochastic process, the truth value of a MITL formula
is a Bernoulli random variable. Computing the probability of such a random variable is a model checking problem. Model checking for MITL properties evaluated
on trajectories from a CTMC requires the computation of transient probabilities; despite major computational efforts [15], this is seldom possible exactly due
to state-space explosion. Statistical model checking (SMC) methods circumvent
such problems by adopting a Monte Carlo perspective: by drawing repeatedly
and independently sample trajectories, one may obtain an unbiased estimate of
the truth probability, and statistical error bounds can be obtained by employing
either frequentist or Bayesian statistical approaches [20,14]. It should be pointed
out that such bounds do not carry the same guarantees as numerical results obtained say by transient analysis; however, simply by drawing more samples one
may reduce the uncertainty in the bounds arbitrarily.
Example 1.2 MITL formulae can be used effectively to obtain behavioural characterisations of the system’s trajectory. We turn again to the SIRS model to
illustrate this concept.
Assume one may want to express a global bound on the virulence of the infection,
so that the fraction of infected population never exceeds λ. This can be done by
considering the formula φ1 , defined as
φ1 ::= G[0,100] (I < λN )
(1)
which translates to:
(
φ1 (x) =
tt
if It < λN ∀t ∈ [0, 100],
¬tt otherwise.
Statistical model checking of this formula is trivial: one simply draws a trajectory
using Gillespie’s algorithm, and monitors that the maximal number of infected
does not exceed the specified threshold in the [0, 100] interval.
4
NI
3
1
Initial state space
NS
2
φ-calculation
GP imputation
NR
P (φ2 )
1
Property space (φ-space)
2
1
3
P (φ1 )
1
MDS projection
JSD as metric
1
P (φn )
MDS extension
θm<n
2
3
JSD space
1
θ1
Clustering
Cluster labelling
θm<n
2
3
JSD space
1
θ1
Fig. 1. The sequence of transformations from space to space are shown in the figure. States from the original state-space (blue circles 1-3) are projected to φ-space
according to satisfaction rate of set properties (found via simulation of the system).
MDS is used to project from φ-space to a space where JSD of φ satisfaction probability distributions between states is preserved as Euclidean distance (in the figure,
JSD[Pφ (2) k Pφ (3)] < JSD[Pφ (1) k Pφ (2)], JSD[Pφ (1) k Pφ (3)] so states 2, 3 are placed
closer together than 1). The states are then clustered to produce macro-states. Out-ofsample states (red cross) can be projected to φ-space, using GP imputation to estimate
satisfaction probabilities. MDS extension allows projecting from φ-space to JSD space
without moving the sampled states. The most likely cluster for the state to belong to
(nearest centroid) is the macro-state it belongs to.
5
3
3.1
Methodology
High level method description
We first present a high-level description of the proposed methodology; the technical ingredients will be introduced in the following subsections. Figure 1 provides an intuitive roadmap of the approach. The overarching idea is to provide
a state-space aggregation algorithm which uses behavioural similarities as an
aggregation criterion.
The input to the approach is a CTMC model and a set of MITL formulae
φ1 , . . . , φn which define the behavioural traits we are interested in. We formalise
some of the key concepts through the following definitions.
Definition 1. A coarsening map C for a CTMC M is a surjective map
M:S→
− R,
(2)
from the state-space S of M to a finite set R, such that card(S) ≥ card(R).
Definition 2. The macro-states of the coarsened system are the elements of the
image of the coarsening map C.
Therefore, the set of all macro-states is a partition of the set of initial states S,
where each element in the partition is a macro-state. In general, there is no way to
retrieve the initial state configuration of the system only from information of the
macro-state configuration, i.e., the coarsening entails an information loss.
We illustrate the various steps of the proposed procedure in Figure 1. The first
step is to take a sample of possible initial states; we then evaluate the joint
satisfaction of the n formulae, given a particular state as initial condition. This
implicitly defines a map
n
Φ : S → [0, 1]2
(3)
which associates each initial state with the probability of each possible satisfaction pattern of the n formulae. Notice that all of the 2n possible truth values
are needed to ensure correlations between properties are captured. Constructing
such a property map by exhaustive exploration of the state-space is clearly computationally infeasible; we therefore evaluate it (by SMC) on a subset of possible
initial states, and then extend it using a statistical surrogate, a Gaussian Process
(Figure 1 top).
The property representation contains the full information over the dependence
of the properties of interest on the initial state. It can be endowed with an
information-theoretic metric by using the JSD between the resulting probability distributions. However, the high dimensionality and likely very non-trivial
structure of the property representation may make this unwieldy. We therefore
propose a dimensionality reduction strategy which maintains approximately the
6
metric structure of the property representation using Multi-Dimensional Scaling
(MDS; Figure 1 middle). MDS will also have the advantage of automatically
identifying potentially redundant characterisations, as implied for example by
logically dependent formulae.
The low-dimensional output of the MDS projection can then be visually inspected for groups of initial states (macro-states) with similar behaviours with
respect to the properties. This operation is a coarsening map, which can also be
automated by using a variety of clustering algorithms.
The model dynamics induce, in principle, a dynamics on this reduced space R.
In practice, such dynamics will be non-Markovian and not easily expressible in
a compact form; we propose a simple, simulation-based alternative definition
which re-uses some of the computation performed in the previous steps to define
an empirical, coarse-grained dynamics on the macro-states.
3.2
Satisfaction probability as a function of initial conditions
The starting point for our approach consists of embedding the initial state-space
into the property space, φ-space. This is achieved by computing satisfaction
probabilities for the 2n possible truth patterns of the n properties we consider.
As in general these satisfaction probabilities can only be computed via SMC, this
is potentially a tremendous computational bottleneck. To obviate this problem,
we turn the computation of the property map into a machine learning problem:
we evaluate the 2n functions on a (sparse) subset of initial states, and predict
their values on the remaining initial states using a Gaussian Process (GP).
GPs have extensively been used in machine learning for regression purposes and
it is in this context they are used here. A GP is a generalisation of the multivariate normal distribution to function spaces with infinitely many dimensions;
within a regression context, GPs are used to provide a flexible prior distribution
over the set of candidate functions underpinning the hypothesised input-output
relationship. Given a number of input-output observations (training set), one
can use Bayes’s rule to condition the GP on the training set, obtaining a posterior distribution over the regression function at other input points. For a review
of GPs and their uses in machine learning, we refer the reader to [18].
In our setting, the input-output relationship is the property map from initial
states to satisfaction probabilities of the properties. This function is defined over
a discrete space, but we can use the population structure of the pCTMC to embed
the state-space S in a (subset) of RD for some D. We can then treat the problem
n
as a standard regression problem, learning a function fφ : RD → R2 .
Remark. GPs have already been used to explore the dependence of the satisfaction probability of a formula on model parameters in the so-called Smoothed
Model Checking approach [6]. There, the authors proved a smoothness result
7
which justified the use of smoothness-inducing GPs for the problem. It is easy
to see that such smoothness does not hold in general for the function fφ ; for
example, the probability of satisfying the formula x(0) > N has a discontinuity
at x = N . However, since we only ever evaluate fφ on a discrete set of points,
the lack of smoothness is not an issue, as a continuous function can approximate
arbitrarily well a discontinuous function when restricted to a discrete set.
Example 1.3 We exemplify this procedure on the SIRS example. We consider
here three properties of interest: the global bound encoded in formula φ1 defined
in equation (1), and two further properties encoded as
φ2 ::= F[0,60] G[0,40] (0.05N ≤ I ≤ 0.2N ),
(4)
φ3 ::= F[30,50] (I > 0.3N ).
(5)
Satisfaction of φ2 requires that the infection has remained within 5 to 20% of
the total population for 40 consecutive time units, starting anytime in the first
60 time units; satisfaction of φ3 requires that the infection peaks at above 30%
between time 30 and time 50.
The property map in this case would have an 8-dimensional co-domain, representing the probability of satisfaction for each of the 23 possible truth values of
the three formulae. Figure 2 plots the probability of satisfaction for the three
formulae individually, as we vary the initial state. In this case, 10% of all possible
initial states were randomly selected and numerically mapped to the property
space via SMC, while the satisfaction probabilities for the remaining 90% were
imputed using GPs. We see that throughout most of the state-space the second property has low probability. Also it is of interest to observe the strong
anti-correlation between the first and third properties: intuitively, if there is
very high probability that the infection will be globally bounded below 40% of
individuals, it becomes more difficult to reach a peak at above 30%.
3.3
Dimensionality reduction of behaviours
Once states are mapped onto φ-space, reducing dimensionality of this space is
useful to remove correlations and redundancies in the properties tracked. Properties may often capture similar behaviour, leading to strong correlations in
their satisfaction probability. Reducing the dimensionality of the property space
mostly retains the information of how behaviour differs from state to state, eliminating redundancies. Moreover, reduced dimensional mappings can aid practitioners to visually identify structures within the state-space of the system.
In order to quantify the similarity of different initial states with respect to property satisfaction, the Jensen-Shannon Divergence (JSD) between the probability
distributions of property satisfaction is used as a metric. JSD is an information
theoretic symmetric distance between probability distributions — the higher the
8
difference between the distributions, the higher JSD is. Between two distributions, P, Q, JSD is defined as
JSD[P k Q] =
1
(KL[P k M ] + KL[Q k M ]),
2
where M = 0.5(P + Q) the average of the distributions, and KL[P k Q] =
P
P (i)
i P (i) log Q(i) , the Kullback-Leibler divergence.
The JSD enables us to derive a matrix of pairwise distances in property space
between different initial states. Such a distance is not Euclidean, and is defined
in the high-dimensional property space. To map the initial states in a more convenient, low-dimensional space, we employ a dimensionality reduction technique
known as Multi-Dimensional Scaling (MDS) [4].
MDS has its roots in the social science literature; it is a valuable and widely
used tool in psychology and similar fields where data is collected by assessing
similarity between pairs.
Given some points X in an m-dimensional space, MDS finds the position of corresponding points Z in an n-dimensional space, where usually n < m, such that
a given metric between points is optimally preserved. In the most common case,
(also known as Torgerson–Gower scaling or Principal Component Analysis), the
metric to be preserved is the Euclidean distance, and is preserved by minimisation of a loss function. This function is generally known as stress for metric
MDS, but specifically for classical MDS as strain.
For the classical MDS case, the projection is achieved by eigenvalue decomposition of a distance matrix of the (normalised) points XX > , and subsequently
reconstructing the points from the n largest (eigenvector, eigenvalue) pairs. This
results in Z, a projection of the points to an n-dimensional space, where Euclidean distance is optimally preserved.
In the classical MDS definition, the MDS projection is defined statically for the
available data points, and needs ab initio re-computation if new points become
available. In [2], the method is extended to new points by constructing a new
dissimilarity matrix of new points to old ones, by which the projection of new
points will be consistent to that of the old points. The kernel for this new matrix
achieves this by replacing the means required for centring with expectations over
the old points; such that for points x, y ∈ X
1X 2 0
1X 2
1 X 2 0 0
1
d (x , y) −
d (x, y 0 ) + 2
d (x , y )] ,
K̃(x, y) = − d2 (x, y) −
2
n 0
n 0
n 0 0
x
y
x ,y
where K̃(x, y) is the kernel used for the dissimilarity matrix, is replaced by
1 2
2
2
0
2
0
K̃(a, b) = − d (a, b) − Ex [d (x, a)] − Ex0 [d (b, x )] + Ex,x0 [d (x, x )] ,
2
where a can be an out-of-sample point (a ∈
/ X, b ∈ X).
9
This reconstructs the dissimilarity matrix for the original points exactly, and
allows us to generalise to out-of-sample points and find their positions in the
embedding learned, as described in [2]. Extending MDS allows us to create
macro-states based on samples of points, and then project new points on the
space created by MDS to find in which clusters they belong.
Example 1.4 We have introduced three properties in Equations (1), (4) and (5),
and the associated property map. This has an eight-dimensional co-domain, but
already some of its properties can be gleaned by the three-dimensional plot of the
single-formula probabilities shown in Figure 2. Particularly, these reveal strong
negative correlations, indicating that MDS may prove fruitful.
Fig. 2. Left: Projection of states in φ-space via SMC (trajectory simulations for each
initial state). Notice the non-trivial state distribution structure. Right: Projection of
states in φ-space using SMC for 10% of the states, and GP regression to estimate P (φ)
for the rest 90% of states (red crosses).
Fig. 3. Left: P (φ1 , φ2 , φ3 ) estimated via SMC for each state. MDS was then used to
project them from an 8D to a 2D space. Right: GP estimates of P (φ1 , φ2 , φ3 ) for 90%
of states (red crosses) produce an almost identical MDS projection.
10
Figure 3 shows the states projected to a 2D space were proximity implies similar probability distribution over property satisfaction. This was achieved using
MDS to project the states, with JSD used as the metric to be preserved as Euclidean distance in the new 2D space. Elements of the square-shaped structure
visible in φ-space (figure 2) are preserved, with the subset of states giving rise
to higher probabilities for property φ2 (top of Figure 2) appearing further from
the connected outline (bottom left group in Figure 3).
3.4
Clustering and structure discovery
The MDS projection enables us to visually appreciate the existence of non-trivial
structures within the state-space, such as clusters of initial states that produce
similar behaviours with respect to the property specification. Our intuition is
that such structures should form the basis to define macro-states of the system,
groups of states that will exhibit similar satisfaction probabilities for the properties defined. To automate this process, we propose to use a clustering algorithm
to define macro-states. Since our goal is to group states with similar behaviours,
we adopt k-means clustering [3], which is based on the Euclidean distance of
the states in the MDS space (representative of the JSD between the probability
satisfaction distributions). k-means requires specification of the desired number
of clusters (the k parameter); this allows the user to select the level of coarsening
required. Figure 4 shows the clusters produced in the reduced MDS space for the
running SIRS model example, where we set the number of clusters k = 10.
k-Means
0.2
0.1
0.0
0.1
0.2
0.0
0.2
0.4
0.6
0.8
Fig. 4. The states were clustered in the space created by the MDS projection and
coloured accordingly, using k-means (10 clusters). Since the Euclidean distance in this
space is representative of distance in probability distributions over properties, states
with different behaviour should be in different clusters.
11
3.5
Constructing coarse dynamics
Once states have been grouped into macro-states, a major question is how to
construct dynamics for the now coarsened system. The coarsened system naturally inherits dynamics from the original (fine-grained) system; however, such
dynamics are non-Markovian, and in general fully history dependent so that
transition probabilities would have the form
p(k 0 , t|k, h) = p(k 0 |k, t, h)p(t|k, h),
(6)
where h denotes the history of the process. Simulating such a non-Markovian
system is very difficult and likely to be much more computationally expensive
than simulating the original system.
We therefore seek to define approximate dynamics which are amenable to efficient simulation, but still capture aspects of the non-Markovian dynamics. The
most natural approximation is to replace the system with a semi-Markov system:
transitions are still history-independent, but the distribution of sojourn times
is non-exponential. To evaluate the sojourn-time distribution, we resort to an
empirical strategy, and construct an empirical distribution of sojourn times by
re-using the simulated trajectories of the fine system that were drawn to define
the coarsening. In other words, once a clustering is defined, we retrospectively
inspect the trajectories to construct a histogram distribution of sojourn times,
approximating p(t|k).
A possible drawback of this semi-Markov approximation is that it may introduce transitions which are actually impossible in the original state-space. This
is because states were clustered based on behaviour rather than transition rates,
and therefore states that are actually quite far in the original state-space may
end up being clustered together. Since the identity of the original states is lost
after the coarse graining, impossible transitions may be introduced.
Retrospectively inspecting whole system trajectories, rather than agnostically
examining cluster transitions of the original system with a uniform initial state
distribution within the cluster, ameliorates this problem. Similarly, estimates
of p(k 0 |t, k) are produced from the same trajectories; these are the macro-state
transition frequencies in each bin of the sojourn time probability histogram.
This method avoids a lot of impossible trajectories one might generate, if the
above probabilities were estimated by sampling randomly from initial states in a
macro-state and looking at when the macro-state is exited and to which macrostate the system transitions. Assuming the original system has a steady state,
the empirical dynamics constructed here capture this steady state macro-state
distribution; however, accuracy of transient dynamics suffers, and the coarsened
system enters the steady state faster than the original system.
Example 1.5 We illustrate and evaluate the quality of the coarsened trajectories
with respect to the original ones on the SIRS example. In particular, we examine
12
the probability distribution over the macro-states at different times in the evolution of the system. The macro-state distribution has been estimated empirically
by sampling trajectories using the Gillespie algorithm for the fine system, and
our coarse simulation scheme for the coarsened system. We have then constructed
histograms to capture the distribution of the categorical random variables that
represent the macro-state. Finally, we measure the histogram distance between
histograms obtained from the fine and the coarse systems. Figure 5 depicts the
evolution of the macro-state histograms over time.
cluster
Fine system cluster distribution
10
9
8
7
6
5
4
3
2
1
0.6
0.4
0.2
20
40
60
80
100
120
140
0
time
cluster
Coarse system cluster distribution
10
9
8
7
6
5
4
3
2
1
0.6
0.4
0.2
20
40
60
80
100
120
140
0
time
Fig. 5. Evolution of the macro-state histograms over time
Quality of Approximation In order to put any distance between empirical distributions into context, this has to be compared with the corresponding average
self-distance, which is the expected distance value when we compare two samples
from the same distribution. In this work, we estimate the self-distance using the
result of [8]: given N samples and K bins in the p
histogram, an upper bound
for the average histogram self-distance is given by (4K)/(πN ). In our example, we have K = 10 histogram bins, which are as many as the macro-states.
In practice, a distance value smaller than the self-distance implies that the distributions compared are virtually identical for a given number of samples. In
Figure 6, we see the estimated distances for N = 10000 simulation runs for
times t ∈ [0, 150]. It can be seen that the steady-state behaviour of the system is
captured accurately, as the majority of the distances recorded after time t = 60
lie below the self-distance threshold. However, the transient behaviour of the system is not captured as accurately. Upon a more careful inspection of the shape
of the histograms in Figure 5, we see that the coarsened system simply converges
more quickly to steady-state. To conclude, we think that the the approximation
13
quality of the steady-state dynamics is a promising result, but a more accurate
approximation of the transient behaviour is subject of future work.
0.25
Estimated distance
Upper bound for average self−distance
Distance
0.2
0.15
0.1
0.05
0
0
50
100
150
Time
Fig. 6. Evolution of the macro-state histogram distances over time
Computational Savings State-space coarsening results in a more efficient simulation process, since the coarse system is characterised by lower complexity
as opposed to the fine system. We demonstrate these computational savings
empirically in terms of the average number of state transitions invoked during
simulation. More specifically, we consider a sample of 5000 trajectories of the
fine and the coarse system. We have recorded 320 ± 25 initial state transitions
on average in each trajectory of the fine system, compared to 56±31 macro-state
transitions in trajectories of the coarsened system. The number of transitions in
the coarse system is an order of magnitude lower than in the fine one, owing
to the reduction of states in the system from a total of 5151 to 10 (the number
of macro-states). Clearly, our procedure, particularly the GP imputation, incurs
some computational overheads. Table 1 presents the computational savings of
using GPs to estimate satisfaction probability distributions for most states, instead of exhaustively exploring the state-space. All simulations were performed
using a Gillespie algorithm implementation, taking 1000 trajectories starting at
each examined state, running on 10 cores.
Sample size
GP & MDS
time (s)
Simulation
time (s)
Total
time (s)
100%
50%
40%
30%
20%
10%
1616*
1133
884
595
354
170
6900
3450
2760
2070
1380
690
8516
4583
3644
2665
1734
860
Percentage of
exhaustive total time
(Total time/8516s)
100%
54%
43%
31%
20%
10%
* No GP was performed here, just the MDS.
Table 1. Real running times for simulations of varying sample size (percentage of
state-space) and GP estimation of remaining states.
14
4
Discussion
We presented a novel approach to the coarsening of a CTMC, in order to gain a
stochastic process with a much smaller state-space. Unlike previous approaches
to CTMC aggregation, which are based on structural properties of the statespace, our approach is based on property satisfaction, allowing the coarse-grained
system to focus on abstracting the dynamics in terms of aspects of behaviour
that are important in the modelling study. The further steps are to identify
key clusters of states in property space, or a lower-dimensional representation of
it, and approximate the transition dynamics between them. For example, this
approach might be used within multi-scale modelling to reduce the state-space
of a lower level model before embedding in a higher-level representation.
Common aggregation techniques, such as exact or approximate lumpability, often
impose stringent conditions on the symmetries and transition rates within the
original state-space. Moreover, the macro-states produced can be difficult to interpret when the reduction is applied directly at the state-space level (i.e. without
a corresponding bisimulation over transition labels). In contrast, the propertybased approach allows macro-states to be defined by high-level behaviour, rather
than them emerging from an algorithm applied to low-level structure.
The GP regression we employed for estimating satisfaction probability of properties for out-of-sample states proved quite accurate; simulation estimates for 10%
of the states were sufficient to reconstruct the state distribution in the space
defined by the probability of property satisfaction, φ-space, without substantial
loss of structure. Therefore, the proposed approach may be helpful in effectively
understanding the behavioural structure of large and complex Markovian systems, with implications for design and verification.
Initial experiments on a simple system show that our approach can be practically
deployed, with considerable computational savings. The approach induces coarsened dynamics which empirically match the original system’s dynamics in terms
of steady-state behaviour. However, the recovery of transient coarse-grained dynamics poses more of a challenge and this will provide a focus for future work. In
particular, we will seek to explore the possibility of quantifying the information
lost through the coarsening approach, at least asymptotically, for systems which
admit a steady state. Exploring the scalability of the approach on more complex,
higher dimensional examples will also be an important priority. In general, we
expect our approach to be beneficial when simulation costs dominate the overheads incurred by the GP regression approach. This condition will be mostly
met for systems with moderately large state spaces but complex (e.g. stiff) dynamics. For extremely large state spaces, the cubic complexity (in the number
of retained states) of GP regression may force users to adopt excessively sparse
sub-sampling schemes, and it may be preferable to replace the GP regression
step with alternative schemes with better scalability. Exploration of these computational trade-offs would likely prove insightful for the methodology.
15
References
1. A. Abate, L. Brim, M. Ceska, and M. Z. Kwiatkowska. Adaptive aggregation of
Markov chains: Quantitative analysis of chemical reaction networks. In Proc. of
CAV, pages 195–213, 2015.
2. Y. Bengio, J.-F. Paiement, P. Vincent, O. Delalleau, N. Le Roux, and M. Ouimet.
Out-of-sample extensions for lle, isomap, mds, eigenmaps, and spectral clustering.
In Proc. of NIPS, pages 177–184, 2004.
3. C. M. Bishop. Pattern Recognition and Machine Learning. Springer-Verlag New
York, Inc., Secaucus, NJ, USA, 2006.
4. I. Borg and P. Groenen. Modern Multidimensional Scaling: Theory and Applications. Springer, 2005.
5. L. Bortolussi, D. Milios, and G. Sanguinetti. Efficient stochastic simulation of
systems with multiple time scales via statistical abstraction. In Proc. of CMSB,
pages 40–51, 2015.
6. L. Bortolussi, D. Milios, and G. Sanguinetti. Smoothed model checking for uncertain continuous-time Markov chains. Inf. Comput., pages 235–253, 2016.
7. P. Buchholz and J. Kriege. Approximate aggregation of Markovian models using
alternating least squares. Perform. Eval., 73:73–90, 2014.
8. Y. Cao and L. Petzold. Accuracy limitations and the measurement of errors in the
stochastic simulation of chemically reacting systems. J. Comput. Phys., 212(1):6–
24, 2006.
9. F. Ciocchetta and J. Hillston. Bio-PEPA: A framework for the modelling and
analysis of biological systems. Theor. Comput. Sci., 410(33):3065–3084, 2009.
10. V. Danos, J. Feret, W. Fontana, R. Harmer, and J. Krivine. Rule-based modelling
of cellular signalling. In In Proc. of CONCUR, pages 17–41, 2007.
11. K. Deng, P. G. Mehta, and S. P. Meyn. Optimal Kullback-Leibler aggregation
via spectral theory of Markov chains. Automatic Control, IEEE Transactions on,
56(12):2793–2808, 2011.
12. A. Donzé and O. Maler. Robust satisfaction of temporal logic over real-valued
signals. Springer, 2010.
13. D. Gillespie.
Exact stochastic simulation of coupled chemical reactions.
J. Phys. Chem., 81(25), 1977.
14. S. K. Jha, E. M. Clarke, C. J. Langmead, A. Legay, A. Platzer, and P. Zuliani. A
Bayesian approach to model checking biological systems. In Proc. of CMSB, pages
218–234, 2009.
15. M. Kwiatkowska, G. Norman, and D. Parker. PRISM 4.0: Verification of probabilistic real-time systems. In Proc. of CAV, pages 585–591, 2011.
16. O. Maler and D. Nickovic. Monitoring temporal properties of continuous signals.
In Proc. of FORMATS, pages 152–166, 2004.
17. D. Milios and S. Gilmore. Component aggregation for pepa models: An approach
based on approximate strong equivalence. Perform. Eval., 94:43–71, 2015.
18. C. E. Rasmussen and C. K. I. Williams. Gaussian Processes for Machine Learning.
MIT Press, 2006.
19. M. Tschaikowski and M. Tribastone. A unified framework for differential aggregations in Markovian process algebra. J. Log. Algebr. Meth. Program., 84(2):238–258,
2015.
20. H. L. S. Younes and R. G. Simmons. Statistical probabilistic model checking with
a focus on time-bounded properties. Inf. Comput., 204(9):1368–1409, 2006.
16
| 3cs.SY
|
Inverse Modeling of Climate Responses of Monumental Buildings
R.P. Kramer, A.W.M. van Schijndel & H.L. Schellen
Eindhoven University of Technology
ABSTRACT
The indoor climate conditions of monumental buildings are very important for the
conservation of these objects. Simplified models with physical meaning are desired
that are capable of simulating temperature and relative humidity. In this paper we
research state-space models as methodology for the inverse modeling of climate
responses of unheated monumental buildings. It is concluded that this approach is
very promising for obtaining physical models and parameters of indoor climate
responses. Furthermore state space models can be simulated very efficiently: the
simulation duration time of a 100 year hourly based period take less than a second
on an ordinary computer.
INTRODUCTION
Effects of climate change on ecosystems and on the global economy have been
researched intensively during the past decades but almost nothing is known about
the influence to our cultural heritage. Although these historical monuments are
exposed to extensive loads caused by stampedes of visitors, there are many other
factors deteriorating World Heritage Sites. The impacts of climate change are a longterm and substantial menace to the sites. Lots of monumental buildings are used as
museum or storage for paintworks, books and artefacts. The indoor climate
conditions of monumental buildings are very important for the conservation of these
objects (ASHRAE, 2007). The influence of the changing climate on the indoor climate
of monumental buildings is unknown. It is impossible to prepare adequately for the
future, by anticipating on this change and adapting the installations, because of this
lack of knowledge (CIBSE, 2005). The result is that the conservation of the buildings
and the collections are at risk. Furthermore, the worldwide energy problem also
affects these monumental buildings. Due to the ancient building techniques used,
which result in huge transmission and infiltration losses, the energy consumption of
these buildings is high. Three problems can be identified with respect to the current
research methods: Firstly, due to the long simulation period (hundred years with
time step 1h), combined with detailed physical models, the simulation run time is
long. Secondly, the detailed modeling of the buildings itself requires much effort: the
monumental buildings are old and protected. Therefore, blueprints are hard to find
and; destructive methods to obtain building material properties are not allowed.
Thirdly, the used modeling approach does neither facilitate in an easy
characterization of the building nor in an easy characterization of the energy
performance. A simplified model with physical meaning is desired which is capable of
simulating both temperature and relative humidity. The objective is the successful
application of inverse modeling on a simplified thermal and hygric building model, in
order to determine the parameters with physical meaning. The simplified model is
needed for the prediction of the indoor temperature and the indoor relative
humidity and for characterization of the building parameters and the energy
performance. The paper is organised as follows. We start with a literature study on
simplified models. The methodology of the inverse modeling development for
climate responses is presented and applied for a group of four unheated
monumental buildings in The Netherlands and Belgium. The approach is evaluated
and conclusions are drawn.
Inverse Modeling of Climate Responses of Monumental Buildings
Page 1 of 12
LITERATURE
Due to the increase of computational power, the attention for simplified models has
decreased. However, through the years it became clear that simplified models have
benefits over complex models (Wang & Chen, 2001), (Mathews, et al., 1994): user
friendliness, straight forward, fast calculation. The response factor method and
lumped capacitance method are suitable for simplified modeling. More recently,
linear parametric models and neural network models are used for simplified models.
Neural network models, e.g. (Mustafaraj, et al., 2011), can be classified as black box
models. The parameters have no direct physical meaning, but the output is
generated by the hidden layers (black box) from the input. Some models are referred
to as gray box models. An example in the field of simplified building models is the
use of linear parametric models (Mustafaraj, et al., 2010).The linear model itself is a
black box model, but the parameters can be determined using physical data
(Jimenez, et al., 2008). Some researchers stress out the importance of simplified
models with physical meaning (Kopecký, 2011), so called white box models. The
lumped capacitance model can be classified as a white box model. Another
advantage of this approach is the representation of building elements using R
(resistance) and C (capacitance), according to the electrical analogy, which makes a
graphical representation of the model possible. Most of the simplified building
models are based on this approach. There are three approaches to create a
simplified model: (1) Create a detailed comprehensive model from known building
properties and perform afterwards a model order reduction technique, e.g. (Gouda,
et al., 2002); (2) Create directly a simplified model from building properties, e.g.
(Nielsen, 2005); (3) Create a simplified model and identify the parameter values with
an inverse modeling technique (Balan, et al., 2011). Technique 1 is obviously the
most labor intensive: building a detailed model and simplifying it afterwards.
Detailed construction properties need to be available together with a methodology
for simplifying an existing model. The lumped capacitance model can be used for this
model order reduction (Mathews, et al., 1994) and neural network models can be
used to filter out unimportant parameters (Mustafaraj, et al., 2011), called pruning.
Technique 2 is faster, but a validated methodology should be known how to identify
the parameters. Therefore, it is difficult to achieve good results with this technique.
(Fraisse, et al., 2002) has demonstrated a methodology how to incorporate multiple
walls into one single order model. Technique 3 is not labor intensive and
identification of the model parameters is done by an optimization algorithm. This
technique can be used with the lumped capacitance model (Wang & Xu, 2006),
neural network model, e.g. (Mustafaraj, et al., 2011), and linear parametric model,
e.g. (Moreno, et al., 2007). Technique 3 is used for the work in this paper.
INVERSE MODEL DEVELOPMENT FOR CLIMATE RESPONSES
A linear model is often sufficient to accurately describe the system dynamics and, in
most cases, one should first try to fit linear models (Ljung, 1999). State Space models
are Linear Time Invariant (LTI) models. State Space models are models that use state
variables to describe the system dynamics by a set of first order differential
equations. To understand the concept of State Space, think of the system as
spanning a space where the axes (i.e. dimensions, i.e. orders) represent the state
variables. Even if some of the systems differential equations are higher order, they
should be converted to multiple first order equations. The state of the system can be
represented as a vector within that space. The system of first order differential
equations can be represented according to:
Inverse Modeling of Climate Responses of Monumental Buildings
Page 2 of 12
𝒙̇ (𝑡) = 𝑨𝒙(𝑡) + 𝑩𝒖(𝑡)
𝒚(𝑡) = 𝑪𝒙(𝑡) + 𝑫𝒖(𝑡)
The first part of equation (1) is known as state equation where x(t) is the state vector
and u(t) is the input vector. The second equation is referred to as the output
equation. A is the state matrix, B is the input matrix, C the output matrix and D the
direct transition matrix. The main advantage is that the calculation speed is very
high, especially compared to solving the differential equations. To show this
advantage, a first order RC-network is simulated as a State Space model and with the
ode23 routine for different simulation periods. The results are shown in Table 1.
Table 1: calculation time of State-Space and ode23 for different periods.
1month 1 year
10 year
100 year
time [s] time [s]
time [s]
time [s]
ODE23
5
89
State-Space
0.016
0.016
0.050
0.45
The results show the huge advantage of the State Space model regarding simulation
time. While the simulation time increases almost linearly when using the ode23
routine to solve the differential equation, the simulation time of the State Space
model is less predictable. However, especially if the task requires a repeated
simulation for a long period, e.g. hundred years, the State Space model has a very
limited calculation time of half a second on an ordinary computer.
We proceed with the thermal and hygric modeling. Due to space limitation
we summarize the work of Kramer (2012). He investigated several thermal models
including solar irradiation. Also several hygric models have been developed and
tested. In this paper we present the optimal models according to Kramer (2012).
The thermal model is shown in Figure 1. There are thermal capacitances for
the interior (Cint), indoor air (Ci) and envelope (Cw). The sun irradiation is connected
to Cint (interior) and Ci (indoor air). Thermal resistances represent ventilation
(1/Gfast), wall to external air (1/Gw), wall to indoor air (1/Gi) and indoor air to
interior constructions (1/Gint).
Figure 1. The thermal network
The ODEs are,
Inverse Modeling of Climate Responses of Monumental Buildings
Page 3 of 12
𝑑𝑇𝑤
= 𝐺𝑤 (𝑇𝑒 − 𝑇𝑤 ) − 𝐺𝑖 (𝑇𝑤 − 𝑇𝑖 )
𝑑𝑡
𝑑𝑇𝑖
���⃗ • ������������⃗
𝐶𝑖
= 𝐺𝑖 (𝑇𝑤 − 𝑇𝑖 ) − 𝐺𝑓 �𝑇𝑖 − 𝑇𝑓 � − 𝐺𝑖𝑛𝑡 (𝑇𝑖 − 𝑇𝑖𝑛𝑡 ) + 𝐺𝑓𝑎𝑠𝑡 (𝑇𝑒 − 𝑇𝑖 ) + 𝑓𝐼
𝐼𝑟𝑟𝑎𝑑 (𝑏)
𝑑𝑡
𝑑𝑇𝑖𝑛𝑡
���⃗ • �����������⃗
𝐶𝑖𝑛𝑡
= 𝐺𝑖𝑛𝑡 (𝑇𝑖 − 𝑇𝑖𝑛𝑡 ) + 𝑓𝐼
𝐼𝑟𝑟𝑎𝑑(𝑎)
𝑑𝑡
𝐶𝑤
The State Space matrices are,
−𝐺𝑤 −𝐺𝑖
𝐺𝑖
⎡
0 ⎤
𝐶𝑤
⎢ 𝐶𝑤
⎥
−𝐺𝑖 −𝐺𝑓 −𝐺𝑖𝑛𝑡 − 𝐺𝑓𝑎𝑠𝑡
𝐺𝑖𝑛𝑡 ⎥
⎢ 𝐺𝑖
𝐴=⎢
𝐶𝑖
𝐶𝑖
𝐶𝑖 ⎥
⎢
⎥
𝐺𝑖𝑛𝑡
−𝐺𝑖𝑛𝑡 ⎥
⎢
0
⎣
𝐶𝑖𝑛𝑡
𝐶𝑖𝑛𝑡 ⎦
𝐺𝑤
⎡
0
0
0
0
0⎤
⎢ 𝐶𝑤
⎥
𝐺𝑓 ⎥
𝑓𝐼1
𝑓𝐼2
𝑓𝐼3
𝑓𝐼4
𝐵 = ⎢𝐺𝑓𝑎𝑠𝑡
⎢
⎥
𝐶𝑖
𝐶𝑖
𝐶𝑖
𝐶𝑖
𝐶𝑖 ⎥
⎢ 𝐶𝑖
⎣ 0
0
0
0
0
0⎦
0
𝐶=� 0
0
0
1
0
0
0
𝐷=� 0
0 �
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0 �
0
The hygric model is shown in Figure 2. The hygric capacitances and resistances
represent are analog to the thermal network, but now for vapour pressures instead
of temperatures.
Figure 2. The hygric model
The ODEs are
𝑑𝑃𝑤
= 𝐺𝑤 (𝑃𝑒 − 𝑃𝑤 ) − 𝐺𝑖 (𝑃𝑤 − 𝑃𝑖 )
𝑑𝑡
𝑑𝑃𝑖
𝐶𝑖
= 𝐺𝑖 (𝑃𝑤 − 𝑃𝑖 ) + 𝐺𝑓𝑎𝑠𝑡 (𝑃𝑒 − 𝑃𝑖 )
𝑑𝑡
𝐶𝑤
Inverse Modeling of Climate Responses of Monumental Buildings
Page 4 of 12
The State Space matrices are,
−𝐺𝑤 −𝐺𝑖
𝐶𝑤
𝐺𝑖
𝐶𝑖
⎡
𝐴=⎢
⎢
⎣
𝐺𝑖
𝐺𝑤
⎤
⎡
⎤
𝐶𝑤
⎥ 𝐵 = ⎢ 𝐶𝑤 ⎥ 𝐶 = � 0
−𝐺𝑖 −𝐺𝑓𝑎𝑠𝑡 ⎥
0
⎢ 𝐺𝑓𝑎𝑠𝑡 ⎥
𝐶𝑖
⎦
⎣ 𝐶𝑖 ⎦
0
0
�𝐷 = � �
1
0
This completes the modeling part. We proceed with the inverse modeling
methodology.
Inverse modeling is the inverse of traditional modeling. In traditional
modeling, the system is known and the output is unknown. By modeling the system,
the output can be simulated. In inverse modeling, the output is known (e.g.
measured), but little is known about the system. The objective is to identify the
parameter values of the model by repeatedly trying different parameter values and
comparing the simulated output with the measured output. The goal is to minimize
the simulation error, formulated as an objective function (e.g. summed squared
error). The process of finding the parameter set which minimizes the objective
function is called optimization. If the solution space includes multiple minima, the
goal is to find the global minimum, called global optimization. These solvers are all
very different, each having its advantages and disadvantages. One important aspect
is whether a solver is gradient based or gradient free. Gradient based solvers are the
most efficient in finding quickly a minimum. However, the solution space should be
smooth and continuous. If not, the solver fails. The second aspect is whether the
solver handles constraints or it is only intended for unconstrained problems. The
inverse problem in this research is a typical example of a constrained problem, since
all variables are not allowed to be negative. Moreover, constraining the problem
helps in finding the global minimum since it scales down the solution space. The
third aspect is whether a solver is deterministic or stochastic: all solvers are
deterministic except for the Genetic Algorithm. To maximize the speed of the
optimization process, all calculations which do not need to be repeated are executed
in the initialization step: preparing climate data, include measured data, set
constraints, set initial values and assign a function handle for the optimization
algorithm. Then the optimization algorithm determines the parameter set, the first
time it’s the initial value vector, and passes the parameter set to the function file:
the function file includes the model and simulates the model with the given
parameter set and calculates the objective function by comparing the simulated
output with the measured data. The objective function is passed to the Global
Optimization Algorithm which calculates the new parameter set which is likely to
minimize the objective function.
RESULTS FOR A GROUP OF UNHEATED MONUMENTAL BUILDINGS
An important validation method is the performance assessment of the developed
models on multiple other buildings: By fitting both the thermal and hygric model to
these buildings, their general performance is tested. General applicability is an
important aspect in this study. Moreover, it is important to chose a strategic set of
buildings and rooms to gain maximum added value of this validation method:
included in the set of buildings are (i) Room which is surrounded by a water canal; (ii)
Room with much thermal mass and hygric mass (iii) Room with significant sun
irradiation (iv) Room with sun irradiation on roof, but no windows (v) Rooms with
Inverse Modeling of Climate Responses of Monumental Buildings
Page 5 of 12
ground contact and no sun irradiation. The physical aspects of the buildings and
rooms and the argumentation why they are included, is explained in the next
Section. The performance per room is visualized in graphs and explained in word.
The performance of the rooms is compared by three performance criteria (MSE,
MAE & Goodness of Fit) at the end of the section. Figure 3 presents the monumental
buildings involved in the study.
Figure 3. The Monumental buildings involved in this study. Top Left: Castle of
Amerongen, Top Right: Castle of Gaasbeek, Bottom Left: Castle Keukenhof, Bottom
Right: St. Baafs Cathedral
Castle of Amerongen: Washing room
The washing room is situated in the basement of the castle of Amerongen. The castle
is surrounded by a canal: the external walls of the washing room are in direct contact
with water. Moreover, the basement has been flooded multiple times. The
performance of the thermal model is shown in Figure 4 (top). The model performs
good over the seasons resulting in a good fit. The detail at the right side of the figure
shows that the delicate fluctuations of the measured signal are not reproduced by
the model, which is a good thing: the sensors have an accuracy of ± 0.5°C meaning
that the observed fluctuations are cover by the uncertainty of the measurement. The
reason of these delicate fluctuations is unclear and considered to be unimportant for
the performance assessment of the model. The performance of hygric model 3 is
shown in Figure 4 (bottom). There seems to be a problem analogous to the thermal
model which has been observed earlier: the signal lies locally above, but mostly
below the measured signal. The result is that the model can’t be fitted accurately. A
new parameter is introduced for the hygric model, which is analogous to the thermal
model: a node with a fixed vapour pressure. The performance of this hygric model
with fixed vapour pressure is shown in Figure 4 (middle). The fixed vapour pressure
Inverse Modeling of Climate Responses of Monumental Buildings
Page 6 of 12
improves the fit significantly. The necessity of such a fixed vapour pressure suggest
the existence of a vapour source. Physically, the external wall which is connected to
a canal is analogous to the thermal situation with ground contact.
Figure 4. The thermal model (top), hygric model with additional fixed vapour
pressure node in model (middle) and hygric model without fixed pressure (bottom)
fitted to washing room (This figure relies on color see digital version of the paper).
Castle of Gaasbeek - Scockaert chamber
The Scockaert chamber is situated at the second floor in the castle of Gaasbeek
(Belgium). It has windows which allow a significant amount of sun irradiation into the
room. The room is richly decorated with antique furniture which adds to the
thermal, and mostly, the hygric capacity. The difficulty here: the irregular
disturbances by visitors. Because this effect can’t be captured by a linear time
invariant model, the Scockaert chamber is a good candidate for the models
performance assessment. The result of the thermal model is shown in Figure 5 (top).
The overall performance is good, but incidentally, some measured peaks are
observed which are not reproduced by the model. These peaks might be a result of
internal heat sources like visitors. The result of the hygric model (with fixed vapour
pressure node) is shown in Figure 6 (bottom). The same observation holds as for the
thermal result. The overall performance is good, but identical series of peaks are not
reproduced. These peaks are due to inputs which are not included, e.g. moisture
production by visitors.
Inverse Modeling of Climate Responses of Monumental Buildings
Page 7 of 12
Figure 5. The thermal model (top) and hygric model with additional fixed vapour
pressure node (bottom) fitted to Scockaert chamber (This figure relies on color see
digital version of the paper).
Castle of Gaasbeek - Gothic chamber
The Gothic chamber is also situated at the second floor of the castle of Gaasbeek.
The room is decorated richly with much furniture and decorations which contribute
to the thermal, but mostly hygric capacity. Also this room is frequently visited by
tourists. The performance of the thermal model is shown in Figure 6 (top). The
incidental peaks are less intense compared to those in the Scockaert chamber, but
are still visible. The detail at the right side of the figure shows that the model
reproduces the measured signal accurately. The delicate quick fluctuations which are
present in the measurements of the washing room are totally absent in the
measurements of the Gothic chamber.
Figure 6. The thermal model (top) and hygric model with additional fixed vapour
pressure node (bottom) fitted to Gothic chamber (This figure relies on color see
digital version of the paper).
Inverse Modeling of Climate Responses of Monumental Buildings
Page 8 of 12
The hygric model performance is shown in Figure 6 (bottom). The overall
performance is good, but small disturbances which are not reproduced are probably
due to time-variant phenomena. At least, the phenomena which could change over
the seasons turn out to be no source of error: think of possible use of sun blinds in
summer.
Castle Keukenhof - the loft
The castle Keukenhof is situated a few kilometers from the Dutch coast. Therefore
the building is exposed to a sea climate. The observed room is the loft which is
situated directly underneath the roof. This room is specifically interesting because
the overheating risk during summer. This situation will be a real test for the model
since the influence of sun irradiation is significantly present, but windows are absent.
The thermal model performance is visualized in Figure 7(top). The risk of overheating
is clearly present and is reproduced good by the model. Although the sun irradiation
is modeled by 4 input signals, each representing the sun irradiation on a vertical wall
oriented respectively to the north, east, south and west, the sun irradiation on the
roof and the resulting heat flux to the loft is simulated correctly. This significant heat
flux resulting from sun irradiation results in an indoor temperature of reaching 35°C
in summer. The detail at the right side of the figure shows how accurate the models
output has been fitted to the measured temperature. The performance assessment
of the hygric model is shown in Figure 7 (bottom). The hygric reproducibility is poor.
Besides the indoor temperature, also the indoor moisture content (or vapour
pressure) fluctuates heavily. These fluctuations good not be reproduced with the
only given input of the outdoor vapour pressure. This casus is interesting for a
follow-up research.
Figure 7. thermal model (top) and hygric model with additional fixed vapour pressure
node (bottom) fitted to Keukenhof-Loft (This figure relies on color see digital version
of the paper).
St. Baafs cathedral – South transept
The St. Baafs cathedral is situated in Gent (Belgium). The building consists of several
parts, e.g. the choir, the entrance, the transepts. Most of the parts are openly
connected. The building is totally free floating and has huge thermal mass and hygric
mass.
Inverse Modeling of Climate Responses of Monumental Buildings
Page 9 of 12
The measurements to which the model has been fitted are performed by a sensor in
the south transept. The south transept includes a huge window with colored glass
which is oriented to the south resulting in a significant amount of incoming sun
irradiation. The performance of the thermal model is shown in Figure 8 (top). The
measured indoor temperature is reproduced very accurately, also visualized clearly
in the detailed graph on the right side. Although there are visitors, the influence on
the indoor climate seems to be small due to the large dimensions of the cathedral.
The performance of hygric model is shown in Figure 8 (bottom). The measured signal
is reproduced perfectly by the model. The same holds for the hygric part: although
there are visitors, they hardly influence the indoor moisture content due to the vast
space in the cathedral. Moreover, the quality of the measurements are high without
signal noise.
Figure 8. The thermal model(top) and hygric model with additional fixed vapour
pressure node (bottom) fitted to St. Baafs cathedral-South transept (This figure relies
on color see digital version of the paper).
ASSESMENT OF THE METHODOLOGY
This section deals with the performance assessment of the models and the used
methods. The performance of the models relates to two things:
i) how accurately can the measured signal be reproduced
ii) how accurately can the parameters be identified (in validation section)
This implies a contradiction. Usually, the rule holds that a higher order model with
more parameters yields a higher accuracy. On the other hand, more parameters
result in a higher uncertainty in parameter identification (Ljung, 1999).
To be able to compare and rate how accurately the different models can reproduce a
measured temperature or vapour pressure, three performance criteria are used: the
MSE (Mean Squared Error), MAE (Mean Absolute Error) and FIT (goodness of FIT).
The MSE is calculated according to,
𝑁
1
𝑀𝑆𝐸 = �(𝑦 ′ − 𝑦)2
𝑁
𝑘=1
where 𝑦 ′ is the measured signal and 𝑦 is the simulated signal.
Inverse Modeling of Climate Responses of Monumental Buildings
Page 10 of 12
The MAE is calculated according to,
𝑁
1
𝑀𝐴𝐸 = �|𝑦 ′ − 𝑦|
𝑁
𝑘=1
The Goodness of Fit is calculated according to,
𝐹𝐼𝑇 = 100 ∙ �1 −
𝑛𝑜𝑟𝑚(𝑦 ′ − 𝑦)
�
𝑛𝑜𝑟𝑚(𝑦 ′ − 𝑦�′)
𝑛𝑜𝑟𝑚(𝑦) is the Euclidean length of the vector 𝑦, also known as the magnitude. The
above equation therefore calculates in the numerator the magnitude of the error
between measured and simulated signal. This is divided by the denominator,
calculating how much the measured signal fluctuates around its mean.
Consequently, the Goodness of Fit criterion is robust with respect to the fluctuation
level of the signal. All three are used independently in several studies, e.g.
(Boaventura Cunha, et al., 1997) (Crabb, et al., 1987) (Frausto, et al., 2003).
However, (Mustafaraj, et al., 2010) stresses the strength of using the three together.
For example, the MSE gives more weight to bigger errors. Consequently, it is a good
criterion to express the amount of big errors. The MAE expresses the overall mean
error.
Table 2. results of fit summarized in three performance criteria (MSE, MAE & FIT).
thermal model 4a
hygric model 3
MSE
MAE
FIT MSE
MAE FIT
2
2
building
room
[°C ]
[°C]
[%] [Pa ]
[Pa] [%]
washing room*
4938
58
72
Amerongen
0.09
0.24
91
washing room
2468
39
81
Scockaert
1.03
0.73
84 5369
57
77
Gaasbeek
chamber
Gothic chamber 0.66
0.59
87 5316
56
78
Keukenhof loft
1.33
0.88
84 23785
113
58
south transept
0.17
0.32
91
St.
Baafs
south
1870
32
86
cathedral
0.74
0.69
81
transept**
dungeon of pain 0.30
0.41
87 3686
47
84
museum
iron chamber
0.10
0.26
82 462
18
85
Gevangenknight chamber 0.34
0.45
88 3710
47
84
poort
stock loft
1.41
0.96
81 6954
65
78
*model without node for fixed vapour pressure
**no solar input
CONCLUSIONS
It is concluded that this approach is very promising for obtaining physical models and
parameters of indoor climate responses. Furthermore by using state space models,
the simulation duration time of a 100 year hourly based period take less than a
second on an ordinary computer.
Inverse Modeling of Climate Responses of Monumental Buildings
Page 11 of 12
REFERENCES
ASHRAE, 2007. Museums, Archives and Libraries. In: ASHRAE handbook Heating, Ventilation and Air
Conditioning applications. SI edition ed. s.l.:American Society of Heating Refrigaration and AirConditioning Engineers, pp. 21.1-21.23.
Balan, R. et al., 2011. Parameter identification and model based predictive control of temperature
inside a house. Energy and Buildings, Volume 43, pp. 748-758.
Boaventura Cunha, B., Couto, C. & Ruano, A., 1997. Real time estimation of dynamic temperature
models for greenhouse environment control. Control Engineering Practice, 5(10), p. 1473–1481.
CIBSE, T., 2005. Climate change and the indoor environment impacts and adaption. s.l.:s.n.
Crabb, J., Murdoch, N. & Pennman, J., 1987. A simplified thermal response model. Building Services
Engineering Research and Technology, Volume 8, pp. 13-19.
de Wit, M., 2006. Heat Air and Moisture model for building and systems evaluation. Bouwstenen,
Volume 100, p. 100 pages.
de Wit, M. & Driessen, H., 1998. ELAN: a computer model for building energy design. Building and
Environment, Volume 23, pp. 285-289.
Fraisse, G., Viardot, C., Lafabrie, O. & Achard, G., 2002. Development of a simplified and accurate
building model based on electrical analogy. Energy and Buildings, Volume 34, pp. 1017-1031.
Frausto, H., Pieters, J. & Deltour, J., 2003. Modelling greenhouse temperature by means of auto
regressive models. Biosystems Engineering, 84(2), p. 147–157.
Gouda, M., Danaher, S. & Underwood, C., 2002. Building thermal model reduction using nonlinear
constrained optimization. Building and Environment, Volume 37, pp. 1255-1265.
Jimenez, M., Madsen, H. & Andersen, K., 2008. Identification of the main thermal characteristics of
building components using Matlab. Building and Environment, Volume 43, pp. 170-180.
Kramer, R.P., 2012. The application of inverse modeling for the prediction and characterization of
indoor climates and energy performances, MSc Report, Eindhoven University of Technology
Kopecký, P., 2011. Experimental validation of two simplified thermal zone models. Tampere (Finland),
s.n.
Ljung, L., 1999. System Identification: Theory for the User. 2nd edition ed. Upper Saddle River(New
York): Prentice-Hall.
Mathews, E., Richards, P. & Lombard, C., 1994. A first-order thermal model for building design. Energy
and Buildings, Volume 21, pp. 133-145.
Moreno, G. et al., 2007. Modelling temperature in intelligent buildings by means of autoregressive
models. Automation in Construction, 16(5), p. 713–722.
Mustafaraj, G., Chen, J. & Lowry, G., 2010. Development of room temperature and relative humidity
linear parametric models for an open office using BMS data. Energy and Buildings, Volume 42, pp.
348-356.
Mustafaraj, G., Lowry, G. & Chen, J., 2011. Prediction of room temperature and relative humidity by
autoregressive linear and nonlinear neural network models for an open office. Energy and
Buildings, Volume 43, pp. 1452-1460.
Nielsen, T., 2005. Simple tool to evaluate energy demand and indoor environment in the early stages
of building design. Solar Energy, Volume 78, pp. 73-83.
van Schijndel, A., 2007. Integrated heat air and moisture modeling and simulation, Eindhoven: s.n.
van Schijndel, A., 2009. The exploration of an inversed modeling technique to obtain material
properties of a building construction. Istanbul, s.n., pp. 91-98.
van Schijndel, A. & Schellen, H., 2011. Inverse modeling of the indoor climate using a 2-state 5parameters model in Matlab. Visby, s.n., pp. 1-12.
van Schijndel, A., Schellen, H., Martens, M. & van Aarle, M., 2010. Modeling the effect of climate
change in historic buildings at several scale levels. Eindhoven, s.n., pp. 161-180.
Wang, S. & Chen, Y., 2001. A novel and simple building load calculation model for building and system
dynamic simulation. Applied Thermal Engineering, Volume 21, pp. 683-702.
Wang, S. & Xu, X., 2006. Parameter estimation of internal thermal mass of building dynamic models
using genetic algorithm. Energy Conversion and Management, Volume 47, pp. 1927-1941.
Wang, S. & Xu, X., 2006. Simplified building model for transient thermal performance estimation using
GA-based parameter identification. International Journal of Thermal Sciences, Volume 45, pp. 419432.
Inverse Modeling of Climate Responses of Monumental Buildings
Page 12 of 12
| 5cs.CE
|
A family of OWA operators based on Faulhaber’s
formulas
arXiv:1801.10545v1 [cs.AI] 31 Jan 2018
O.G. Duartea,∗, S.M. Télleza,∗∗
a Universidad
Nacional de Colombia, Facultad de Ingeniería, Bogotá, Colombia
Abstract
In this paper we develop a new family of Ordered Weighted Averaging (OWA)
operators. Weight vector is obtained from a desired orness of the operator.
Using Faulhaber’s formulas we obtain direct and simple expressions for the
weight vector without any iteration loop. With the exception of one weight,
the remaining follow a straight line relation. As a result, a fast and robust
algorithm is developed. The resulting weight vector is suboptimal according with
the Maximum Entropy criterion, but it is very close to the optimal. Comparisons
are done with other procedures.
Keywords: Data fussion, Aggregation operators, Ordered Weighted Averaging
1.
Introduction
Ordered Weighted Averaging (OWA) operators are well known aggregation
operators [1] that have received great attention in recent years. Although OWA
operators were proposed by Yager since 1988 [3], a great amount of research
arises since 2008. [9] provides a collection of recent developments around OWA
operators. A systematic review of literature about OWA is done in [2]; it reveals
some great research lines, one of them is to solve the weight vector determination
problem. Another wide area of research deals with generalizations to fuzzy set
theory.
The weight vector determines the way in which the OWA operator aggregates
information. For some applications we need to adjust the behaivor of the OWA
operator to resemble the Minimum or Maximum operators. Those operators are
numerical implementations of logical operators AN D and OR. To handle that,
the orness of the OWA operator was defined by Yager as a measure of how close
to the OR operator the OWA is.
A common problem is to determine the vector weight of an OWA operator
that meets a desired orness. A lot of methods have been proposed to solve this
problem [7]. In this paper we propose another one. Our method is fast, robust,
∗ Corresponding
∗∗ This
author
work was partially supported by Colciencias, call number 617
Preprint submitted to Elsevier
1 de febrero de 2018
has good performance and is easy to interpret by non-experts. Some of these
features appear in some methods, and other features in others, but our method
accomplishes all of them. As an example, consider the exponential method [3]:
it is fast and robust, but its performance (measured with the entropy index)
is not the best; our method is faster, as robust and performs better than the
exponential. As another example, consider the Maximum Entropy method [5],
whose performance is optimal, but it is slow and have numerical problems; our
method is faster and more robust than it, and its performance is very close to
the optimal.
The structure of the paper is the following: section 2 summarizes the background about OWA operators. In section 3 we propose a new method to obtain
the weight vector of an OWA operator that meets a desired orness. We explore
the behaivor and results of the proposed method in section 4. Some conclusions
are done in section 5.
2.
OWA operators
OWA operators are aggregation operators defined by [3, 1]:
OW A : Rn → R
y = f (x1 , x2 , · · · , xn )
y=
Pn
i=1
wi x̄i
Pn
i=1
wi = 1
(1)
0 ≤ wi ≤ 1 ∀i
Where x̄i is the i−th element of {x1 , x2 , · · · , xn } previously ordered from
highest to lowest. The selection of the weight vector w = {w1 , · · · , wn } stablishes
different kind of aggregations; remarkable OWA operators are:
Maximum: with w = {1, 0, 0, · · · , 0}
Minimum: with w = {0, 0, · · · , 0, 1}
Simple average: with w = {1/n, 1/n, · · · , 1/n}
2.1. Orness and entropy
Yager define orness(w) and disp(w) (dispersion or entropy) as:
n
orness(w) =
1 X
(n − i)wi
n − 1 i=1
disp(w) = −
n
X
wi ln wi
(2)
(3)
i=1
Orness refers to the degree to which the aggregation is like an or operation
where as dispersion measures the degree to which w takes into account all
information in the aggregation.
2
2.2.
Families of OWA operators
There are a lot of strategies to compute the weight vector of an OWA operator. An extensive review is found in [7] where they are classified in 5 groups:
1. Optimization methods. The problem is stated as the optimization of a performance index subject to restrictions. Shanon entropy, dispersion, Rényi
entropy are some of the indexes used. One of the most remarkable method is
the Maximum Entropy method proposed by O’Hagan [8] that was reformulated in [5] in an analytical way.
2. Empirical data methods. This group of methods are usually intended to model the risk attitude of a decision maker: from a set of recorderd decisions we
must find the OWA that better reflects them.
3. Regular Increasing Monotone (RIM) quantifier methods. RIMs are lingüistic
quantifiers that can be used to guide the selection of the weight vector.
4. The argument dependent methods. These are methods that use the input data vector x as useful information in the weight vetor determination problem.
Unlike the classical OWA operators, inputs do not have to be ordered.
5. Preference relation methods. These methods are an extension of the empirical
data methods. Empirical data is replaced by a preference matrix provided by
the experts.
A sixth group of methods can be added. It can be called Shaped based
methods, because they use a predefined shape of the weight distribution to
solve the problem. Notable methods are the exponential proposed by Yager in
[3] and the linear proposed by Lamata in [6].
3.
A family of OWA operators
We now address the issue to find the weights of an OWA operator with
a desired orness. The operator must aggregate the information of n individual
values. We define α ∈ [0, 1] as a variable that equals the desired orness. However,
in a more general perspective, α = g(orness) ∈ [0, 1] with g(·) a monotonically
increasing function with g(0) = 0 and g(1) = 1.
We impose the following restrictions, intended to construct a family whose
behaivor is easy to understand:
OWA.1 if α = 0.0 the operator must be equivalent to the Mimimum operator.
OWA.2 if α = 0.5 the operator must be equivalent to the Simple Average
operator.
OWA.3 if α = 1.0 the operator must be equivalent to the Maximum operator.
3
OWA.4 operator’s behaivor must be symetrical respect to α = 0.5. In other
words, the way in which the operator evolves from Minimum to Simple
average when α varies from 0.0 to 0.5 must be the same way in which
it evolves from Maximum to Simple Average when α varies from 1.0
to 0.5.
Last restriction is easy to acomplish if we define a family for 0 ≤ α ≤ 0.5
and then we use simetry to define the behaivor for 0.5 < α ≤ 1.0.
3.1.
And-like operators (0 ≤ α ≤ 0.5)
Be α the desired orness, and consider first the case 0 ≤ α ≤ 0.5. The rationale
of our proposal is the following: as α varies from 0 to 0.5 the weights must
vary from those shown in figure 1(a) to those in figure 1(b). We first vary
wn monotonically from 1.0 to 1/n. The amount in which we diminishes wn
is distributed into the other weights to keep unchanged the sum of weights.
We distribute this amount using an straight line whose parameters are easy to
compute, as we show in the following paragraphs.
We define f (x) : [0, 0.5] → [0, 1] a monotonically increasing function with
f (0) = 0 y f (0.5) = 1. Other conditions for f (x) are developed later. The value
of wn must vary from 1 to 1/n when α varies from 0 to 0.5. We use f (·):
wn = 1 − ∆
∆ = f (α)(1 − 1/n) = f (α)(n − 1)/n
(4)
Under this conditions, the sum of the other weights must be
n−1
X
i=1
wi =
m
X
wi = ∆
m=n−1
(5)
i=1
According with equation 2 the orness is
n
orness(w) =
m
1 X
1 X
(n − i)wi =
(n − i)wi = α
n − 1 i=1
m i=1
(6)
We propose a distribution of ∆ following the linear relation:
wi = Ki + b
i = 1, 2, · · · , m
(7)
Notice that this proposal is different from the linear proposal of Lamata [6],
because we use the linear relation for n − 1 weights and Lamata uses it for all
the n weights.
Equations 5 and 6 turns into
Pm
i=1 (Ki + b) = ∆
(8)
1 Pm
(n
−
i)(Ki
+
b)
=
α
i=1
m
4
That can be arranged as
K
m
n
Pm
i=1
i−
K
Pm
i=1
b
m
i2 +
(
Pm
i=1
Pm
i=1
i+
n−
Pm
i=1
Pm
i=1
b
=∆
(9)
i) = α
We compute the sums using Faulhaber’s formulas
m
X
i=
i=1
K
m
m(m+1)2
2
m
X
m(m + 1)
2
−
i2 =
i=1
m(m+1)(2m+1)
6
b
m
+
m(m + 1)(2m + 1)
6
+ mb
K m(m+1)
2
=∆
=α
m(m + 1) −
m(m+1)
2
(10)
And rearrange the terms in order to show explicitly a system of 2 linear
equations and 2 unknows
+
bm = ∆
K m(m+1)
2
(11)
m+1
+
b
=
α
K (m+1)(m+2)
6
2
From 11 we get direct expressions for K and b
∆n/m−2α
K = 6 m2 −1
b=
∆
m
(12)
− K (m+1)
2
Using equation 4 we can compute K and b from α and m or n
f (α)−2α
(α)−2α
= 6 fn(n−2)
K = 6 m2 −1
b=
f (α)
m+1
−
K (m+1)
2
5
=
f (α)
n
−
K n2
(13)
3.2.
Selection of f (·)
As all weights must be non-negative, lets consider the first weight w1 =
K ×1+b
K +b ≥0
K+
∆
m
− K m+1
2
K 1−m
2 +
∆
m
≥0
≥0
K 1−m
2
∆
≥ −m
3(1 − m) ∆n/m−2α
m2 −1
∆
≥ −m
n
− 2α)
− n3 (∆ m
∆
≥ −m
2mα − n∆
(14)
≥ −n∆/3
2mα ≥ 2n∆/3
3α ≥ ∆n/m
3α ≥ f (α)
To ensure that the other weights are also non-negative, we impose the condition K ≥ 0. Directly from equation 13 we get f (α) ≥ 2α.
We have found the following restrictions for f (α):
2α ≤ f (α) ≤ 3α
(15)
If we choose f (α) = 1 − (1 − 2α)β with some 0 ≤ β ≤ 1 (figure 2), those
restrictions are satisfied if the slope at 0 is in [2, 3]
2≤
df (α)
dα |α=0
≤3
2 ≤ 2β(1 − 2α)β−1 |α=0
≤3
2≤
2β
≤3
1.0 ≤
β
≤ 1.5
(16)
Notice that the limit case with β = 1.0 makes f (α) = 2α. It makes K = 0 and
b = 1/n meaning that ∆ is homogeneously distributed between w1 , · · · , wn−1 .
6
3.3. Or-like operators (0.5 ≤ α ≤ 1)
To ensure a symetrical behaivor of the family, we need to adapt the previous
result when 0.5 ≤ α ≤ 1.0 in the following way:
Compute the weights for ᾱ = 1 − α
Reverse the order of the weights.
Algorithm 1 implements the above results in pseudocode.
4.
Examples and analysis
In order to analyze the algorithm performance, we compare their results for
3 values of β (1.0, 1.25 and 1.5) against two well known algorithms: exponential
method [3] and maximum entropy method [5]. Exponential method is similar
to our proposal in the sense that it also propose a shape for the distribution
of ∆ (an exponential distribution); however, to fit the desired orness, α must
be preset according to pre-computed figures. Maximum entropy method has an
optimization approach, and it involves the numerical solution of an equation.
Figure 3 shows the weights obtained for n = 5 when the desired orness is 0.6.
Notice the linear distribution from w2 to w5 in figure 3(a) and the Homogeneous
distribution when β = 1.0. Notice also that the linear method approaches better
than the exponential the weight distribution of the optimal entropy method.
In fact, the exponential method provides an increasing distribution where as
entropy and linear methods provide decreasing distributions.
Figure 4 shows how fast varies w1 when we change the desired orness for
n = 5. We have added the case of the exponential method without preset, in
order to stress the presetting effect. Curves in figure 4(b) has a change when
orness is 0.5, caused by the strategy adopted to ensure simetry. These figures also
show that linear method behaves closer to entropy method than the exponential
one.
Figure 5 shows the change in all the weights when we change the desired
orness for n = 5. It is important to remark the simetry of the entropy and linear
methods. Similarity between figures 5(a) and 5(b) reinforces the idea that linear
method with β = 1.5 is very similar to maximum entropy method.
Figure 6 shows how varies dispersion (entropy) when we change the desired
orness for n = 5. Variations of entropy with the linear method for different β
are minimal. Notice also that the entropy obtained by the linear method with
β = 1.5 is very close to the optimal obtained by the entropy method.
Table 1 shows a comparison of computing times. Algorithms have been implemented in OpenModelica [4]. We have run every algorithm 20 times for n = 10
and n = 100. The average simulation times are in table 1, as relative times to
the minimal average time. Exponential method, as described in [3], involves a
graphical procedure for the preset of α; we have implemented it using an exhaustive search procedure, which explains the high times for this method. It
is an unfair comparison, and we prefer to compare exponentials method times
using an implementation without preseting α.
7
Table 1 shows that the linear method is the fastest one. It is faster than the
entropy method, because it doesn’t need any iteration loop and it is faster than
the exponential method because arithmetic operations are simpler.
Moreover, entropy method has important lacks (see remark 1 in [7]): objective function definition implies wi > 0 and as Maximum and Minimum operators
are OWA’s with most of the weights equal to zero, those cases can not be included. Even for α close to 0 or 1 numerical problems may arise: Figure 7 shows the
dispersion (entropy) and orness for the Maximum entropy method for desired
orness ∈ [0.9, 1.0] obtained with our OpenModelica implementation. Notice the
numerical problems thah appear near α = 0.98. Those numerical problems are
not present in the linear method, because there are not iteration loops.
5.
Conclusions
The linear method is a fast and reliable method for computing weights of
OWA operators with a desired orness. It is as fast or faster than the exponential
method. It is also more suitable for software implementations than the exponential method because does not use precomputed curves for presettings. It is
more robust than optimization methods because there are no iterations loops.
And it provides a suboptimal result, very close to the optimal.
References
Referencias
[1] Detyniecki, M., 2001. Fundamentals on Aggregation Operators. Computer
Science division, University of California, Berkeley.
[2] Emrouznejad, A., Marra, M., 2014. Ordered weighted averaging operators
1988–2014: A citation-based literature survey. International Journal of Intelligent Systems 29 (11), 994–1014.
URL http://dx.doi.org/10.1002/int.21673
[3] Filev, D., Yager, R. R., 1998. On the issue of obtaining owa operator weights.
Fuzzy Sets and System 94, 157–169.
[4] Fritzson, P., 2014. Principles of Object Oriented Modeling and Simulation
with Modelica 3.3. Wiley-IEEE Press.
[5] Fullér, R., Majlender, P., 2000. An analytic approach for obtaining maximal entropy owa operator weights. Tech. rep., Turku Centre for Computer
Science.
[6] Lamata, M. T., Perez, E. C., 2009. Owa weights determination by means of
linear functions. Mathware and soft computing 16 (2).
8
[7] Liu, X., 2011. Recent Developments in the Ordered Weighted Averaging
Operators: Theory and Practice. Vol. 265 of Studies in Fuzziness and Soft
Computing. Springer, Ch. A Review of the OWA Determination Methods:
Classification and Some Extensions, pp. 60–90.
[8] O’Hagan, M., 1988. Aggregating template or rule antecedents in real-time
expert systems with fuzzy set logic. In: Twenty-Second Asilomar Conference
on Signals, Systems and Computers. Vol. 2. pp. 681–689.
[9] Yager, R. R., Kacprzyk, J., Beliakov, G. (Eds.), 2011. Recent Developments
in the Ordered Weighted Averaging Operators: Theory and Practice. Vol.
265 of Studies in Fuzziness and Soft Computing. Springer.
9
input : n the size of the OWA operator
input : orness the desired orness of the OWA operator
output: w a vector of size n with the weights of the OWA operator
Define f (x) = 1 − (1 − 2x)β with β ∈ [1, 1.5]. By default β = 1.5
2 if orness ≤ 0.5 then
3
α ← orness
4 else
5
α ← 1 − orness
6 end
7 m←n−1
8 ∆ ← f (α)m/n
2
9 K ← 6(f (α) − 2α)/(m − 1)
10 b ← f (α)/n − Kn/2
11 if orness ≤ 0.5 then
12
wn ← 1 − ∆
13
for i ← 1 to m do
14
wi ← Ki + b
15
end
16 else
17
w1 ← 1 − ∆
18
for i ← 1 to m do
19
wi+1 ← K(n − i) + b
20
end
21 end
Algorithm 1: Algorithm to get the weights of an OWA operator using
the linear method
1
10
Weight
Weight
Weight
1.0
1.0
1.0
1
n
1
n
1
n
w1
w2
w3
w4
w5
w1
(a) Mínimum
w2
w3
w4
w5
(b) Simple average
f (α)
3α
1
2α
α
0
0.5
Figura 2: Function f (α) = 1 − (2α − 1)β and its restrictions
11
w2
w3
(c) Maximum
Figura 1: Weigths for 3 OWAs with n = 5
0
w1
w4
w5
Weight
1.0
Maximum entropy
Exponential with preset
0.8
0.6
0.4
0.2
0
w1
w2
w3
w4
w5
(a) Exponential and Entropy methods
Weight
1.0
Linear with β = 1.00
Linear with β = 1.25
Linear with β = 1.50
0.8
0.6
0.4
0.2
0
w1
w2
w3
w4
w5
(b) Linear method
Figura 3: Weights for n = 5 and desired orness 0.6.
12
w1
Maximum entropy
Exponential with preset
Exponential without preset
1.00
0.75
0.50
0.25
0
Orness
0
0.2
0.4
0.6
0.8
1.0
(a) Exponential and Entropy methods
w1
Linear with β = 1.00
Linear with β = 1.25
Linear with β = 1.50
1.00
0.75
0.50
0.25
0
Orness
0
0.2
0.4
0.6
0.8
1.0
(b) Linear method
Figura 4: Variation of w1 for n = 5 as a function of the desired orness
13
W eight
1.00
w1
w2
w3
w4
w5
0.75
0.50
0.25
W eight
1.00
w1
w2
w3
w4
w5
0.75
0.50
0.25
0
Orness
0
0.2
0.4
0.6
0.8
0
1.0
Orness
0
(a) Maximum entropy
0.2
0.4
0.6
0.8
1.0
(b) Linear with β = 1.50
W eight
1.00
w1
w2
w3
w4
w5
0.75
0.50
0.25
W eight
1.00
w1
w2
w3
w4
w5
0.75
0.50
0.25
0
Orness
0
0.2
0.4
0.6
0.8
0
1.0
Orness
0
(c) Exponential with preset
0.2
0.4
0.6
0.8
1.0
(d) Linear with β = 1.25
W eight
1.00
w1
w2
w3
w4
w5
0.75
0.50
0.25
W eight
1.00
w1
w2
w3
w4
w5
0.75
0.50
0.25
0
Orness
0
0.2
0.4
0.6
0.8
1.0
0
Orness
0
(e) Exponential without preset
0.2
0.4
0.6
0.8
(f) Linear with β = 1.00
Figura 5: Comparison of Weights for n = 5 as a function of desired Orness
14
1.0
Dispersion
2.00
1.75
1.50
1.25
1.00
0.75
0.50
0.25
0
0
0.2
Maximum entropy
Exponential with preset
Exponential without preset
Orness
0.4
0.6
0.8
1.0
(a) Exponential and Entropy methods
Dispersion
2.00
1.75
1.50
1.25
1.00
0.75
0.50
0.25
0
0
0.2
Linear with β = 1.00
Linear with β = 1.25
Linear with β = 1.50
Orness
0.4
0.6
0.8
1.0
(b) Linear method
Dispersion
2.00
1.75
1.50
1.25
1.00
0.75
0.50
0.25
0
0
0.2
Maximum entropy
Exponential with preset
Linear with β = 1.50
Orness
0.4
0.6
0.8
1.0
(c) Exponential, Entropy and Linear method
Figura 6: Comparison of Orness and Dispersion (Entropy) for n = 5 as a function of desired
Orness
15
Dispersion
4.0
Real Orness
1.0
3.5
3.0
2.5
2.0
0.5
1.5
1.0
0.5
0
Orness
0.90
0.92
0.94
0.96
0.98
0
Orness
1.0
0.90
(a) Dispersion
0.92
0.94
0.96
0.98
1.0
(b) Real orness
Figura 7: Dispersion (Entropy) and Real Orness for the Maximum entropy method with
desired orness close to Maximum, for n = 100
16
Cuadro 1: Relative computing times. Averages from 20 samples
Method
Maximum entropy
Exponential with preset
Exponential without preset
Linear with β = 1.00
Linear with β = 1.25
Linear with β = 1.50
17
n = 10
2.91
108.81
1.05
1.05
1.03
1.00
n = 100
2.26
366.22
2.04
1.01
1.02
1.00
| 2cs.AI
|
Fast Shortest Path Routing in Transportation
Networks with Time-Dependent Road Speeds
Costas K. Constantinou∗ , Georgios Ellinas∗† , Christos Panayiotou∗† and Marios Polycarpou∗†
∗ KIOS
Research Center for Intelligent Systems and Networks
of Electrical and Computer Engineering
University of Cyprus
Emails: {constantinou.k.costas, gellinas, christosp, mpolycar}@ucy.ac.cy
† Department
arXiv:1408.4113v6 [cs.DS] 7 Jan 2016
Abstract
The current paper deals with the subject of shortest path routing in transportation networks (in terms of travelling time), where
the speed in several of the network’s roads is a function of the time interval. The main contribution of the paper is a procedure
that is faster compared to the conventional approaches, that derives the road’s traversal time according to the time instant of
departure, for the case where the road’s speed has a constant value inside each time interval (in general, different value for each
time interval). Furthermore, the case where the road’s speed is a linear function of time inside each time interval (in general,
different linear function for each time interval) is investigated. A procedure that derives the road’s traversal time according to
the time instant of departure is proposed for this case as well. The proposed procedures are combined with Dijkstra’s algorithm
and the resulting algorithms, that are practically applicable and of low complexity, provide optimal shortest path routing in the
networks under investigation.
I. I NTRODUCTION
In the current paper, the subject of shortest path routing in transportation networks where the roads’ speed is function of
time, is investigated. This category of networks is met in practice, since the speed in network’s roads can be available for
several time instants, either directly measured, or derived from vehicle density measurements [1].
The objective of this work was the development of low-complexity, practically applicable algorithms for fast derivation of
shortest paths (in terms of travelling time) in the networks under investigation.
Section II gives the notation, definitions and assumptions used throughout the paper. In general, a transportation network is
modelled as a directed graph, where the nodes, arcs of the graph represent the network junctions, roads respectively. A cost
is assigned to each arc, that in the current work represents the travelling time on the arc (called “traversal time” as well). For
the classical case, where the network is time-independent (called static as well), this cost is constant over time for every arc,
derived by dividing the length of the arc with the speed in it. For the case investigated here, the network is time-dependent, i.e.,
this cost is a function of time, depending on the time instant of departure. Therefore, for shortest path routing in time-dependent
networks an additional calculation must be performed compared to static networks: the calculation of the travelling time on
the network’s arcs according to the time instant of departure.
Throughout the paper it is considered that the time horizon is split into non-overlapping time intervals, and the speed in a
network’s arc depends on the time interval. This model was firstly presented in [2]. More details on this, as well as the reasons
why this model was adopted for the work proposed here, can be found in Section III, where the main models for the networks
under investigation, are presented.
The main contribution of the current paper can be found in Sections IV and V.
In Section IV, a procedure is proposed, that derives the arc’s (road’s) traversal time according to the time instant of departure,
for the case where the speed in the arc has a constant value inside each time interval (in general, a different value for each time
interval). The proposed procedure is faster compared to the conventional approaches. More precisely, if we consider that the
time horizon is split into K time slots, the computational complexity of the proposed procedure is of order O(log K), whereas
the complexity of the fastest existing approach is of order O(K). Special cases are also investigated, where the complexity of
the proposed procedure can be further decreased if certain constraints are valid for the network graph.
In Section V, the general case of speed being an arbitrary function of time inside the time interval is investigated, and
a procedure is proposed that derives the road’s traversal time according to the time instant of departure, for the case where
the speed in the arc is linear function of time inside each time interval (in general, a different linear function for each time
interval). To the best of our knowledge, relevant methods for this case do not exist in the literature.
The proposed methods are combined with Dijkstra’s algorithm and the resulting algorithms, that are practically applicable
and of low complexity, provide optimal shortest path routing in the networks under investigation.
The paper finishes with Section VI, where the conclusions are presented, as well as possible future research.
II. N OTATION , D EFINITIONS , A SSUMPTIONS
Throughout the paper the network is modelled as a directed graph G = (N, A), consisting of n = |N | nodes (road junctions)
and m = |A| arcs (roads). The arc originating from node x and end at node y is denoted by < xy >, and its length (i.e., actual
length of the corresponding road) by dxy . The speed in < xy > is denoted by v xy . A node y is considered to be adjacent to
x if < xy > exists in the graph.
The cost cxy (or “traversal time” or “travelling time”) of arc < xy > is defined as the time needed to traverse it, i.e., to
move from node x to y. Consequently, the term “shortest path” from node x to node y refers to the path of the minimum
travelling time from x to y. For the case of static networks, the cost of an arc is constant over time, equal to dxy /v xy for arc
< xy >. For time-dependent networks, it is a function of the time instant of departure (τ ) from node x, and it is denoted by
cxy (τ ). Therefore, for shortest path routing in time-dependent networks an additional calculation must be performed compared
to static networks: the calculation of cxy (τ ) according to the time instant of departure from x.
Throughout the paper, it is considered that the speed in the network’s roads has been measured for certain time instants1 ,
during a large time interval (e.g., an entire year). In this way, a speed-pattern can be derived for each road. These patterns
constitute an estimation of the time-dependent network graph.
III. E XISTING M ODELS OF T IME -D EPENDENT T RANSPORTATION N ETWORKS
In this Section, the two main models of time-dependent transportation networks are presented.
A. Flow Speed Model (FSM)
One approach to model a time-dependent network is to use the Flow Speed Model (FSM), proposed in [2]. In the FSM, for
each network arc a temporal area from t = 0 to t = T is partitioned into (in general not equal) non-overlapped time intervals,
and the speed depends on the time interval. The time division is the same for every network arc. The (k + 1)th time interval
is denoted by [τk , τk+1 ), with k ∈ N and 0 ≤ k ≤ K − 1, τ0 = 0 and τK = T . The number of time intervals is equal to K.
The speed in arbitrary arc < xy > is considered to be constant inside each time interval and it is denoted by vkxy for time
interval [τk , τk+1 ). Let the set of speeds vkxy and time intervals [τk , τk+1 ), 0 ≤ k ≤ K − 1, for arc < xy > be denoted by V xy
and T xy respectively, and V = ∪∀<xy>∈A V xy and T = ∪∀<xy>∈A T xy . Then, for the FSM the network graph is denoted by
G = (N, A, T, V ).
xy
For t > T , it can be considered either that vkxy = vK−1
(i.e., the network is static for t > T ), or that vkxy is periodic with
period equal to T .
The FSM always satisfies the First-In-First-Out (FIFO) property, as proven in [2]. In simple words, the FIFO property has
the following meaning: Consider two vehicles A and B that depart from node x and traverse arc < xy >, with B having
greater time instant of departure from x, compared to A. Then, B will have greater time instant of arrival at y, compared to
A.
As stated in Section II, for routing purposes in time-dependent networks, the derivation of arc traversal time must be
performed. In the FSM, this is performed as follows.
1) Procedure for Derivation of Arc Traversal Time (AT T ): The traversal time cxy (τ ) of arc < xy > if τ is the time instant
of departure from x can be derived as follows (AT T xy (τ ) procedure2 ):
1) Locate index k such that τk ≤ τ < τk+1
xy
2) If (vkxy · (τk+1 − τ ) ≥ dxy ) cxy (τ ) = vdxy
k
Else {
3) a) i) a = dxy − vkxy · (τk+1 − τ )
ii) k ∗ = k + 1
b) While (vkxy∗ · (τk∗ +1 − τk∗ ) < a) {
i) a ← a − vkxy∗ · (τk∗ +1 − τk∗ )
ii) k ∗ ← k ∗ + 1 }
4) cxy (τ ) = (τk∗ − τ ) + vaxy∗ }
k
The necessity of the aforementioned procedure is due to the fact that a single time interval may not be enough for the
derivation of the arc traversal time. This occurs when the distance that can be traversed from the time instant of departure till
the end of the corresponding time interval, is less than the length of the arc.
Step 1 of the AT T procedure needs O(K) time, if the time intervals are checked sequentially. The order of the number
of the time intervals that are checked during the while loop (step 3b) is O(K). Therefore, the order of the computational
complexity of the AT T procedure is O(K).
1 Either
2 This
directly measured, or derived from vehicle density measurements [1]
procedure was initially proposed in [2]; the one presented in the current paper is an equivalent form of the one in [2].
m/s 12
v
10
8
6
4
2
0
Fig. 1.
0 5 10 15 20 25 30 35 40 45 50 t
s
Example of derivation of arc traversal time (d = 250m, τ = 2s)
The example of Figure 1 illustrates the operation of the AT T procedure. Here, the length dxy of arc < xy > is equal to
170m and the departure time τ from x is equal to 6s. The first five time intervals are [τ0 , τ1 ) = [0, 10), [τ1 , τ2 ) = [10, 15),
[τ2 , τ3 ) = [15, 30), [τ3 , τ4 ) = [30, 40), [τ4 , τ5 ) = [40, 50). These along with their corresponding speed are given in Figure 1.
For this example, the AT T procedure performs as follows:
1) The departure time τ = 6s lies into time interval [τ0 , τ1 ) = [0, 10) ⇒ k = 0
2) The distance that can be traversed from τ = 6s until the end of this time interval is equal to v0xy ·(τ1 −τ ) = 10·(10−6) =
40m. Since it is smaller than the length of the arc, the procedure is continued.
3) a) i) a = dxy − v0xy · (τ1 − τ ) = 170 − 40 = 130m
ii) k ∗ = 1
b) • v1xy · (τ2 − τ1 ) = 6 · (15 − 10) = 30m < a = 130m ⇒
i) a = 130 − 30 = 100m
ii) k ∗ = 2
xy
• v2 · (τ3 − τ2 ) = 8 · (30 − 15) = 120m > a = 100m ⇒ exit from the while loop
xy
4) c (6) = (τ2 − τ ) + a/v2xy = (15 − 6) + 100/8 = 9 + 12.5 = 21.5s
The traversal time is equal to 21.5s and the time instant of arrival at node y is equal to 6 + 21.5 = 27.5s.
2) Applicability of the FSM for Networks under Investigation: The application of the FSM for the networks investigated
in the current paper, is straightforward. The time instants for partitioning the temporal area are the ones where the speed has
been measured. As stated previously, in [2] the time division is the same for every network arc. In general, it may be different.
In the current paper, for simplicity, we make the same assumption as in [2], i.e., that the time division is the same for every
arc. Nevertheless, the generalisation is straightforward, as can be seen from the description of the AT T procedure.
In [2], as stated previously, the speed is considered to be constant in every time interval. Up to Section IV, this assumption
is adopted. The general case of the speed being a function of time inside a time interval, is investigated in Section V.
In [2], the AT T procedure is combined with Dijkstra’s algorithm [3], and the resulting algorithm (called Time-DependentDijkstra (TD-Dijkstra) hereafter) is as follows.
3) Time-Dependent Dijkstra’s Algorithm: As stated in Section II, for the networks under investigation, the calculation of
cxy (τ ) must be performed during the execution of any shortest path routing algorithm. In [2], where the FSM model and the
AT T procedure were proposed, the latter is combined with Dijkstra’s algorithm [3] and the resulting algorithm (TD-Dijkstra)
is suitable for shortest path routing in networks with time-dependent road speeds. The input of TD-Dijkstra is the network
graph G = (N, A, T, V ) and the source node, and the output is the shortest path from the source to every other network node.
For the execution of TD-Dijkstra, the following are used.
• s: Source.
• W (x): Label of node x.
• p(x): Predecessor of node x.
• Gx : Set of nodes adjacent to node x.
• gx : Number of nodes adjacent to node x (i.e., gx = |Gx |).
The exact steps of TD-Dijkstra are:
1) a) W (s) = 0
b) p(s) = 0
c) N ∗ = N − {s}
d) ∀x ∈ N ∗ :
i) If (x ∈ Gs )
{ Run AT T sx (W (s)); W (x) = csx (W (s)); p(x) = s }
ii) Else { W (x) = ∞; p(x) = 0 }
2) While (N ∗ 6= ∅){
a) Find x ∈ N ∗ such that ∀x0 ∈ N ∗ : W (x) ≤ W (x0 )
b) N ∗ ← N ∗ − {x}
c) ∀x0 ∈ (N ∗ ∩ Gx ):
0
i) Run AT T xx (W (x))
0
ii) If (W (x) + cxx (W (x)) < W (x0 ))
0
{ W (x0 ) = W (x) + cxx (W (x)); p(x0 ) = x } }
The TD-Dijkstra algorithm functions as the classical Dijkstra’s algorithm, with the difference that the AT T procedure is
used in steps 1(d)i and 2(c)i for the derivation of the arc’s cost according to the time instant of departure.
On termination of the algorithm, the label W (x) of a node x gives the cost of the shortest path from the source to this
node, and p(x) gives its predecessor in this path.
Since the computational complexity of the AT T procedure is of order
Pn O(K), step 1 requires O(nK) time and each iteration
of step 2 requires O(n + gx K) time. Therefore, due to the fact that x=1 gx = m, the complexity of TD-Dijkstra is of order
O(nK + n2 + mK) = O(n2 + mK).
The D-Dijkstra is, to the best of our knowledge, the fastest existing algorithm for shortest path routing in the networks
investigated in the current paper.
B. Traversal Time Model (TTM)
Another approach to model a time-dependent network, is to use the Traversal Time Model (TTM). In the TTM, the traversal
time function f xy (τ ) is utilised, where f xy (τ ) is equal to the time needed to traverse arc < xy >, if τ is the time instant of
departure from node x. The TTM may or may not satisfy the FIFO property. This model was initially proposed in [4] and
was exploited in several other papers, such as in [5], [6], [7], [8], [9], [10], [11], [12].
Function f xy (τ ) can be defined so as to have either integer-valued or real-valued domain and range, leading to discrete- or
continuous-time-dependent networks respectively.
If f xy (τ ) is defined so as to have integer-valued domain and range (first variation of the TTM), the network is discrete-time
and f xy (τ ) is known for every τ ∈ N , 0 ≤ τ ≤ T . This variation can be found in [12] as well as other sources.
For real-valued domain and range of f xy (τ ) (second variation of the TTM), the network is continuous-time and f xy (τ ) is
known for several time instants denoted by τk , with 0 ≤ τk ≤ T , k ∈ N and 0 ≤ k ≤ K, τ0 = 0 and τK = T .
These time instants, in general, are different for each arc. Work that utilises this variation of the TTM can be found in
[5]-[11] and possibly in other sources. In the aforementioned papers, f xy (τ ) is considered to be a piecewise linear function
of time, having the time instants τk where it is known, as breakpoints. For a time instant τ for which f xy (τ ) is unknown, it
is derived by linear interpolation between the consecutive breakpoints τk , τk+1 such that τk < τ < τk+1 :
⇒
xy
)−f xy (τk )
f xy (τ )−f xy (τk )
= f (ττk+1
τ −τk
k+1 −τk
xy
)−f xy (τk )
f xy (τ ) = f (ττk+1
· (τ − τk ) + f xy (τk )
k+1 −τk
(1)
(2)
Equation 1 can be derived from the tangent of angle θ of Figure 2.
For both variations, for τ > T either it can be considered that f xy (τ ) = f xy (T ) (i.e., the network is static for τ > T ), or
it can be considered that f xy (τ ) is periodic with period equal to T .
1) Applicability of the TTM for Networks Under Investigation: For optimal routing in the networks under investigation the
continuous-time variation of the TTM must be used, since the time instant of departure from a network node can have any
arbitrary value. The derivation of f xy (τ ) for the time instants the speed has been measured, can be performed using the AT T
procedure of the FSM. For the calculation of cxy (τ 0 ) for τ 0 not equal to one of the aforementioned time instants, the assumption
that f xy (τ ) is piecewise linear, must be valid. However, the following example shows that for the networks investigated in the
current paper, this is not always true.
Consider the example of Figure 1. Here, if the AT T procedure is applied for τ equal to 0s, 10s, the derived f xy (τ ) is
equal to 20s, 22s, respectively. Using equation 2, f xy (6) = 21.2s. However, cxy (6) = 21.5s 6= f xy (6), as derived in Section
III-A1. Therefore, for this example, the assumption that f xy (τ ) is piecewise linear, is not valid.
Consequently, the FSM is utilised for the routing algorithms proposed in the following Section.
f(t)
f(τk+1)
f(τ)
θ
f(τk)
τk
τ
Fig. 2.
τk+1
t
Derivation of equation 1
IV. FAST S HORTEST PATH ROUTING A LGORITHMS
The design of a procedure that functions as the AT T 3 but with lower computational complexity compared to the AT T , can
lead to algorithms for faster shortest path routing in the networks under investigation, compared to TD-Dijkstra.
As stated in Section III-A1, the complexity of the AT T procedure is of order O(K) due to steps 1 and 3b. When the AT T
procedure is executed, the index of the time interval where the time instant of arrival lies in, can be kept [2]. In more detail,
after the execution of the AT T for the derivation of the time instant τ ∗ = τ + cxy (τ ) of arrival at node y of the arbitrary
arc < xy >, the index k ∗ such that tk∗ ≤ τ ∗ < tk∗ +1 can be kept, since it is known from the last iteration of step 3(b)ii of
the AT T (let the index that is kept, be denoted by I). For each one of the upcoming executions of the AT T in step 2(c)i of
TD-Dijkstra, performed for every z such that z ∈ (N ∗ ∩ Gy ), index k at step 1 is equal to I. Therefore,
• For the first execution of the AT T (step 1(d)i of TD-Dijkstra) the index k of step 1 of the AT T is known (it is equal to
0), and
• For any upcoming execution of the AT T (step 2(c)i of TD-Dijkstra) the index k of step 1 of the AT T is known as
described previously
The result is that the complexity of step 1 of the AT T is of order O(1) if for every execution of the AT T , the index of
the time interval where the time instant of arrival lies in, is kept.
Consequently, any attempt to decrease the complexity of the AT T procedure must concentrate on the reduction of the
complexity of step 3b of it. At this step, variable k ∗ is increased by one at each iteration, (sequential search is performed over
k ∗ ) until the last necessary time interval is located, i.e., the time interval where the time instant of arrival τ ∗ at node y (of
< xy >), lies in (i.e., [tk∗ , tk∗ +1 ) such that tk∗ ≤ τ ∗ < tk∗ +1 ).
An initial thought for decreasing the complexity would be to perform binary search [13] over variable k ∗ instead of sequential
search, to locate the last necessary time interval. However, this will not lead to lower complexity, since at each iteration of the
binary search, the check whether tk∗ ≤ τ ∗ < tk∗ +1 will need O(K) time4 , thus not decreasing the complexity of the AT T
procedure.
To achieve reduction of the complexity of the AT T procedure using binary search over k, a pair of new variables, named
effective length and additive effective length, are introduced in the current paper. They are defined as follows.
A. Effective Length, Additive Effective Length
•
The effective length lkxy is defined as the distance that can be traversed on arc < xy > during the time interval [tk , tk+1 ).
Therefore,
lkxy = vkxy · (tk+1 − tk )
Note that in general, lkxy can be greater, equal or less than the actual length dxy of arc < xy >.
3 i.e.,
that derives the cost of an arc according to the time instant of departure
at each iteration of this check, the initial while loop, i.e., with sequential search on k∗ , must be executed.
4 Since
(3)
•
Pi
xy
xy
The additive effective length Lxy
i of the time interval [ti , ti+1 ) for arc < xy > is defined as Li =
k=0 lk , i.e., is the
sum of the effective lengths from the first time interval [t0 , ti+1 ) up to the [ti , ti+1 ), including the latter. In other words,
it is equal to the distance that is traversed from time instant t0 to ti+1 . For i ≤ j,
Lxy
j
−
Lxy
i
=
j
X
k=0
lkxy
−
i
X
k=0
lkxy
=
j
X
lkxy
(4)
k=i+1
xy
Therefore, the difference Lxy
j − Li is equal to the distance that is traversed from time instant ti+1 to tj+1 .
In the following Section, a new procedure is proposed, named Fast AT T procedure (F AT T ), that utilises the additive
effective length for fast calculation of the arc traversal time.
B. Fast AT T Procedure (F AT T )
The exact steps of the F AT T xy (τ ) procedure for arc < xy > and τ as time instant of departure from node x, are as follows.
1) Steps of the F AT T xy (τ ) Procedure:
1) Locate index k such that τk ≤ τ < τk+1
xy
2) If (vkxy · (τk+1 − τ ) ≥ dxy ) cxy (τ ) = vdxy
k
Else {
3) a) i) a = dxy − vkxy · (τk+1 − τ )
c
ii) k ∗ = b (k+1)+K
2
iii) a0 = 0
b) While (a0 = 0) {
(k+1)+k∗
xy
xy
∗
• If (a < Lk∗ −1 − Lk ) k ← b
c
2
xy
xy
k∗ +K
∗
• Else If (a > Lk∗ − Lk ) k ← b 2 c
xy
xy
xy
xy
xy
xy
0
• Else If (Lk∗ −1 − Lk ≤ a ≤ Lk∗ − Lk ) a = Lk∗ −1 − Lk }
0
}
4) cxy (τ ) = (τk∗ − τ ) + a−a
xy
vk
∗
τ
2) Description of F AT Txy
Procedure:
• Steps 1 and 2: They are identical to the ones of the AT T procedure.
• Step 3(a): (i) If the procedure is continued to step 3, a is set so as to be equal to the residual distance of arc < xy >,
if the distance traversed from time instant τ to τk+1 is subtracted from dxy . The traversal of distance dxy with τ as the
time instant of departure, is equivalent to traversing distance a with τk+1 as the time instant of departure.
(ii) For the binary search over k ∗ (i.e., the index of the time interval [τk∗ , τk∗ +1 ) where the time instant of arrival lies
in), k ∗ is initialised with the value b (k+1)+K
c.
2
(iii) Variable a0 is set to zero and it is used in steps 3(b) and 4.
∗
• Step 3(b): Binary search over k is performed, to locate the time interval [τk∗ , τk∗ +1 ) where the time instant of arrival
xy
∗
lies in. According to equation 4, Lxy
k∗ − Lk is equal to the distance traversed on < xy > from τk+1 to τk +1 .
xy
xy
– If a < Lk∗ −1 − Lk , then distance a (with τk+1 as time instant of departure) will be traversed prior to τk∗ . Index k ∗
∗
c.
is too large, consequently, k ∗ ← b (k+1)+k
2
∗
xy
– If a > Lxy
−
L
,
then
distance
a
will
be
traversed
after τk∗ +1 . Index k ∗ is too small, consequently, k ∗ ← b k +K
k∗
k
2 c.
xy
xy
xy
xy
∗
∗
– If Lk∗ −1 − Lk ≤ a ≤ Lk∗ − Lk , index k is the “correct” one (i.e., for the time instant of arrival τ , τk∗ ≤ τ ∗ ≤
xy
τk∗ +1 ). Consequently, binary search is terminated and variable a0 is set to Lxy
k∗ −1 − Lk . The procedure exits the
while loop.
0
0
• Step 4: Distance a is the part of a that is traversed from τk+1 to τk∗ . The residual distance, i.e., a − a , will be traversed
in time interval [τk∗ , τk∗ +1 ), i.e., with speed vkxy∗ . Therefore, the traversal time of < xy > with τ as time instant of
0
0
xy
departure, is equal to the traversal time up to time instant τk∗ , plus a−a
(τ ) = (τk∗ − τ ) + a−a
xy , i.e., c
xy .
vk
vk
∗
∗
To further clarify the proposed procedure, note that:
xy
xy
xy
xy
xy
xy
a−a0
a−a0
a−a0
a−a0
• For the quantity
xy , 0 ≤
xy ≤ l ∗ . If a = L ∗
xy = 0. If a = L ∗ − L
xy = l ∗ .
k
k −1 − Lk , then vk
k
k , then vk
k
vk
vk
∗
∗
∗
∗
∗
• The way values are assigned to k , permits the latter to vary from k + 1 to K − 1, i.e., being in accordance to the fact
that the arrival time (if the procedure has moved to step 3) will lie in a time interval from [τk+1 , τk+1 ) to [τK−1 , τK ).
The utilisation of the additive effective length in the proposed F AT T procedure leads to O(1) time for each iteration of
the while loop of step 3(b). Therefore, the complexity of the F AT T procedure is of order O(log K), since O(log X) is the
complexity of binary search over X ordered objects [13].
If the F AT T procedure is used in TD-Dijkstra instead of the AT T , the resulting algorithm, called Fast Time-DependentDijkstra (FTD-Dijkstra), has complexity of order O(n2 + m log K), compared to O(n2 + mK) of TD-Dijkstra.
The derivation of the additive effective length for arc < xy > can be performed using the following proposed AELxy
procedure, applied to arc < xy >.
TABLE I
C OMPLEXITY OF EXISTING AND PROPOSED ALGORITHMS
ALGORITHM
Exist.
Prop.
Prop.
Prop.
Prop.
TD-Dijkstra
FTD-Dijkstra
B-FTD-Dijkstra (Q ≥ K)
B-FTD-Dijkstra (1 < Q < K)
B-FTD-Dijkstra (Q ≤ 1)
PREPROC.
SPACE
QUERY
O(mK)
O(mK)
O(mK)
O(mK)
O(mK)
O(mK)
O(mK)
O(mK)
O(mK)
O(mK + n2 )
O(m log K + n log n)
O(m log K + n log n)
O(m log Q + n log n)
O(m + n log n)
C. AELxy Procedure
xy
1) Lxy
0 = l0 , i = 1
2) While (i ≤ K − 1)
xy
xy
a) Lxy
i = Li−1 + li
b) i ← i + 1
Procedure AEL has O(K) computational complexity; the same stands for space complexity. Therefore, the preprocessing
computational complexity of FTD-Dijkstra is O(mK).
The space complexity is also O(mK), equal to the case where the AEL procedure is not applied, since the network consists
of m arcs, each one consisting of K time intervals with their corresponding speeds.
A faster implementation can be performed, if Fibonacci heaps are utilised [14], [13]. Under this data structure, the amortised
complexity of the selection of the node with the minimum cost (step 2a of TD-Dijkstra) is O(log n), leading to computational
complexity of O(m log K + n log n) for FTD-Dijkstra. The latter complexity can be further decreased if specific constraints
are valid for the network graph, as detailed below.
D. Special Cases
xy
Consider the case where the effective lengths are bounded below, i.e., lixy ≥ dQ ∀ < xy >∈ A and ∀i : 0 ≤ i ≤ K − 1.
Without loss of generality, consider that Q is integer. Then, in step 3(b) of the F AT T procedure, Q consecutive time intervals
are enough for the traversal of distance a5 , since:
P(k+1)+(Q−1)
k+1
lixy ≥
≥ ((k + 1) + (Q − 1) − (k + 1) + 1) ·
=Q·
dxy
Q
=d
xy
dxy
Q
=
>a
(5)
Therefore, the binary search in step 3(b) of the F AT T procedure can be constrained in the area from k + 1 up to k + 1 + Q,
instead of up to K. This leads to computational complexity of order O(log Q) for the F AT T procedure and, consequently
to O(m log Q + n log n) for FTD-Dijkstra. If Q < K, then this complexity is lower compared to O(m log K + n log n). The
resulting algorithm is named Bounded-FTD-Dijkstra (B-FTD-Dijkstra).
If the network has the property that for every arc, every time interval has effective length at least as large as the distance
i
of the arc, i.e., lxy
≥ dxy ∀ < xy >∈ A and ∀i : 0 ≤ i ≤ K − 1, then Q ≤ 1 and B-FTD-Dijkstra has complexity of
O(m + n log n), i.e., equal to the one of classical Dijkstra’s algorithm [3], [14].
The complexity of the existing TD-Dijkstra algorithm as well as of the proposed FTD-Dijkstra and B-FTD-Dijkstra
algorithms, is given in Table I. It must be stated that all the algorithms of Table I (existing and proposed) are optimal
for the networks under investigation.
V. G ENERAL C ASE OF S PEED AS F UNCTION OF T IME I NSIDE THE T IME I NTERVAL
Up to this point, it was assumed (as in [2]) that the speed is considered to be constant inside every time interval. In this
Section, the manipulation of the general case of speed being a function
of time inside a time interval is proposed. Let this
R
function be gkxy (t) for time interval [τk , τk+1 ) and arc < xy >, and gkxy (t) = Gxy
k (t). This general case can be handled by
generalising the notion of effective length (equation 3) as follows:
Z τk+1
xy
xy
lk =
gkxy (t)dt = Gxy
(6)
k (τk+1 ) − Gk (τk )
τk
5 with
τk+1 as time instant of departure, as explained previously
A. Special Case: Constant Speed Inside the Time Interval
The (already investigated) special case of speed being constant in every time interval [τk , τk+1 ), equal to the one measured
at time instant τk (i.e., gkxy (t) = vkxy ) is derived from equation 6 as follows:
lkxy
Z
τk+1
=
τk
gkxy (t)dt
=
vkxy
τk+1
Z
τk
dt = vkxy · (tk+1 − tk )
(7)
Let the F AAT procedure for this case be Constant F AAT (C − F AAT ). Then, C − F AAT is exactly the one presented
in Section IV-B.
B. Special Case: Linear Speed Inside the Time Interval
For the case of speed measurements (as described in Section II), it is more natural to assume that the speed in the arbitrary time
xy
interval [τk , τk+1 ) of arc < xy >6 is a linear function of time (gkxy (t)), taking the values gkxy (τk ) = vkxy and gkxy (τk+1 ) = vk+1
,
xy
rather than being constant, equal to vk for the entire time interval [τk , τk+1 ). To the best of our knowledge, work on this case
cannot be found in the literature for the networks under investigation.
Under the aforementioned assumption, gkxy (t) is derived as follows:
xy
xy
vk+1
−vk
τk+1 −τk
xy
vk+1
−v xy
xy
xy
⇒ gk (t) − vk = τk+1 −τkk · (t − τk )
xy
xy
xy
v xy −vk
vk+1
−vk
xy
⇒ gkxy (t) = τk+1
t
+
(v
−
k
τk+1 −τk τk )
k+1 −τk
xy
xy
xy
xy
v
v
−v
τ
k+1 −vk+1 τk
k
k
⇒ gkxy (t) = τk+1
t
+
τk+1 −τk
k+1 −τk
xy
xy
gk
(t)−vk
t−τk
=
(8)
(9)
If we set,
xy
xy
vk+1
−vk
τk+1 −τk =
xy
xy
vk
τk+1 −vk+1
τk
τk+1 −τk
Rkxy
(10)
= Skxy
(11)
gkxy (t) = Rkxy t + Skxy
(12)
then,
⇒
Gxy
k (t)
=
2
Rkxy t2
+
Skxy t
(13)
From equations 6 and 13, the effective length is derived as follows:
lkxy = Rkxy
2
τk+1
2
+ Skxy τk+1 − Rkxy
τk2
2
− Skxy τk
(14)
Let the F AAT procedure for the current case be Linear F AAT (L − F AAT ). Then, L − F AAT is derived from F AAT
as follows.
•
•
Step 1 remains the same as in F AAT .
In step 2 of the F AAT , vkxy · (τk+1 − τ ) is equal to the distance traversed from τ to τk+1 . This distance in L − F AAT
is given by
R τk+1
τ
gkxy (t)dt = Rkxy
2
τk+1
2
2
+ Skxy τk+1 − Rkxy τ2 − Skxy τ
(15)
– If this distance is equal to or larger than dxy , then the traversal time cxy (τ ) is given by the solution of equation 16,
as follows (where for simplicity, in the intermediate steps for the derivation of equation 17 from equation 16, cxy (τ ),
Rkxy , Skxy and dxy are written as c, R, S and d respectively).
6 where
xy
the speed has been measured at time instants τk and τk+1 with measured values equal to vkxy and vk+1
respectively
R τ +cxy (τ )
⇒
τ
(τ +c)2
R 2
2
gkxy (t)dt = dxy
k
+ S(τ + c) −
2
R τ2 −
2
(16)
Sτ = d
⇒ R(τ + c) + 2S(τ + c) − Rτ − 2Sτ = 2d
⇒ Rτ 2 + Rc2 + 2Rτ c + 2Sτ + 2Sc − Rτ 2 − 2Sτ = 2d
⇒ Rc2 + 2(Rτ + S)c − 2d = 0
√
−2(Rτ +S)± 4(Rτ +S)2 +8Rd
⇒c=
2R
√
xy
xy xy
xy
τ +Skxy )2 +2Rk
d
−(Rk
τ +Skxy )± (Rk
xy
⇒ c (τ ) =
xy
R
(17)
k
– If this distance is less than dxy , the procedure continues to step 3.
•
•
τ2
2
In step 3(a)i, a is set to dxy − Rkxy k+1
Skxy τk+1 − Rkxy τ2 − Skxy τ . Steps 3(a)ii, 3(a)iii and 3(b) remain the same.
2 +
0
is the time needed to traverse distance a − a0 , where this distance is traversed
In step 4 of the F AAT , the quantity a−a
xy
vk
∗
inside time interval [τk∗ , τk∗ +1 ). This time in L − F AAT is given by
R τk∗ +cxy (τ )
τk ∗
⇒ cxy (τk∗ ) =
gkxy∗ (t)dt = a − a0
√ xy
xy
xy
xy
−(Rk
∗ τk∗ +Sk∗ )±
xy
0
(Rk∗ τk∗ +Sk∗ )2 +2Rk
∗ (a−a )
xy
Rk
∗
(18)
(19)
The intermediate steps for derivation of equation 19 from eq. 18 are omitted since they are analogous to derivation of
equation 17 from eq. 16.
Considering the above analysis, the exact steps of the proposed L − F AAT procedure are as follows.
1) Steps of the L − F AT T xy (τ ) Procedure:
1) Locate index k such that τk ≤ τ < τk+1
τ2
xy
xy
xy τ 2
xy
2) If (Rkxy k+1
2 + Sk τk+1 − Rk 2 − Sk τ ≥ d )
cxy (τ ) =
xy
−(Rk
τ +Skxy )±
√
xy
xy xy
(Rk
τ +Skxy )2 +2Rk
d
xy
Rk
Else {
τ2
xy
xy τ 2
xy
3) a) i) a = dxy − Rkxy k+1
2 + Sk τk+1 − Rk 2 − Sk τ
(k+1)+K
∗
ii) k = b
c
2
iii) a0 = 0
b) While (a0 = 0) {
(k+1)+k∗
xy
xy
∗
• If (a < Lk∗ −1 − Lk ) k ← b
c
2
xy
xy
k∗ +K
∗
• Else If (a > Lk∗ − Lk ) k ← b 2 c
xy
xy
xy
xy
xy
xy
0
• Else If (Lk∗ −1 − Lk ≤ a ≤ Lk∗ − Lk ) a = Lk∗ −1 − Lk }
4)
√ xy
xy
xy
xy
2
0
−(Rk
(Rk∗ τk∗ +Skxy
∗ τk∗ +Sk∗ )±
∗ ) +2Rk∗ (a−a )
cxy (τk∗ ) =
xy
R ∗
k
}
The computational complexity of the proposed L − F AAT procedure is O(log K). The routing algorithms for this case are
derived from the ones proposed for the case of constant speeds inside the time interval (Section IV) just by substituting the
F AAT procedure with the L − F AAT .
Note that the F AAT procedure can be modified analogously for the case of speed being an alternative function of time
(rather than linear) inside the time interval.
C. Validity of the FIFO Property for the Generalised Case
In [2], it was proven that the FSM satisfies the FIFO property for the case of constant speeds inside the time interval. In
this Section, it is proven that the FIFO property is also valid for the general case of speed being an arbitrary function of time
(gkxy (t)) inside the time interval.
Consider that for the arbitrary arc < xy >, two vehicles 1, 2 depart from node x on time instants τ1 , τ2 > τ1 respectively,
and arrive at node y on time instants τ10 , τ20 respectively. Obviously, τ10 > τ1 and τ20 > τ2 .
TABLE II
C OMPLEXITY OF EXISTING AND PROPOSED PROCEDURES
Speed inside the time interval
Existing procedure
Proposed procedure
constant value
linear function of time
O(K)
−
O(log K)
O(log K)
To prove that the FIFO property is valid, it must be proven that τ20 > τ10 . The following cases are possible:
•
•
τ 0 >τ2
2
τ2 ≥ τ10 ==
=⇒ τ20 > τ10
τ1 < τ2 < τ10 . The distance dxy is traversed from time instant τ1 to τ10 , and it can be split into distances a and b
(a + b = dxy ), where a is the distance traversed from τ1 to τ2 and b is the distance traversed from τ2 to τ10 . Then,
τ2 > τ1 ⇒ a > 0 ⇒ b < dxy
(20)
From equation 20, it is concluded that the time from τ2 to τ10 is not enough to traverse the whole arc < xy >. Therefore,
vehicle 2 that departs from node x at time instant τ2 will arrive at node y at τ20 > τ10 .
The computational complexity of the existing and proposed procedures for the derivation of the road’s traversal time according
to the time instant of departure, for both the investigated cases of speed having constant value or being a linear function of
time inside the time interval, are presented in Table II.
VI. C ONCLUSIONS
In the current paper, the problem of shortest path routing in transportation networks (in terms of travelling time) where the
speed in several of the network’s roads is a function of the time interval, was investigated. More precisely, it was considered that
the time horizon was split into time intervals, and the speed depended on the time interval. Fast procedures for the derivation
of the road’s traversal time according to the time instant of departure have been proposed. For the case of the speed having
a constant value inside each time interval (in general, different value for each time interval), the proposed procedure is faster
compared to the conventional approaches. Furthermore, for the case of the speed being a linear function of time inside each
time interval (in general, different linear function for each time interval), a procedure is also proposed and, to the best of our
knowledge, relevant approaches cannot be found in the literature.
The proposed procedures were combined with Dijkstra’s algorithm and the resulting algorithms, that are practically applicable
and of low complexity, provide optimal shortest path routing in the networks under investigation.
Future research will focus on the combination of the proposed procedures with more advanced shortest path routing algorithms
that were designed for static networks, in order to derive their versions for time-dependent networks as modelled in the current
paper. Examples of such algorithms are A* [15], Contraction Hierarchies [16], SHARC routing [17]. Although time-dependent
versions of these algorithms can be found in the literature (e.g., [11] for the time-dependent version of Contraction Hierarchies),
these have been designed under the assumption that the traversal time function f xy (τ ) (Section III-B) is a piecewise linear
function of τ . However, as shown in Section III-B1, for time-dependent networks as modelled in the current paper, this is not
always valid.
ACKNOWLEDGMENTS
This work was co-funded by the European Regional Development Fund and the Republic of Cyprus through the Research
Promotion Foundation (Project New Infrastructure/Strategic/0308/26).
R EFERENCES
[1] A. D. May, Traffic flow fundamentals, Prentice Hall, 1990.
[2] K. Sung, M. G. Bell, M. Seong, and S. Park, Shortest paths in a network with time-dependent flow speeds, European Journal of Operational Research,
vol. 121, no. 1, pp. 3239, 2000.
[3] E. W. Dijkstra, A note on two problems in connexion with graphs, Numerische Mathematik, vol. 1, pp. 269271, 1959
[4] K. Cooke and E. Halsey, The shortest route through a network with time-dependent internodal transit times, Journal of Mathematical Analysis and
Applications, vol. 14, no. 3, pp. 493498, 1966.
[5] D. Delling, Time-dependent SHARC-routing, Algorithmica, vol. 60, no. 1, pp. 6094, 2011.
[6] G. Nannicini, D. Delling, D. Schultes, and L. Liberti, Bidirectional A* search on time-dependent road networks, Networks, vol. 59, no. 2, pp. 240251,
2012.
[7] D. Delling and G. Nannicini, Core routing on dynamic time-dependent road networks, INFORMS J. on Computing, vol. 24, no. 2, pp. 187201, Apr.
2012.
[8] D. Delling, P. Sanders, D. Schultes, and D. Wagner, Algorithmics of large and complex networks, J. Lerner, D. Wagner, and K. A. Zweig, Eds. Berlin,
Heidelberg: Springer-Verlag, 2009, ch. Engineering Route Planning Algorithms, pp. 117139.
[9] D. Delling and D. Wagner, Time-dependent route planning, in Robust and Online Large-Scale Optimization, ser. Lecture Notes in Computer Science,
R. Ahuja, R. Mhring, and C. Zaroliagis, Eds. Springer Berlin Heidelberg, 2009, vol. 5868, pp. 207230.
[10] B. Ding, J. X. Yu, and L. Qin, Finding time-dependent shortest paths over large graphs, in Proceedings of the 11th International Conference on Extending
Database Technology: Advances in Database Technology, ser. EDBT 08. New York, NY, USA: ACM, 2008, pp. 205216.
[11] G. V. Batz, R. Geisberger, P. Sanders, and C. Vetter, Minimum time-dependent travel times with contraction hierarchies, J. Exp. Algorithmics, vol. 18,
pp. 1.4:1.11.4:1.43, Apr. 2013.
[12] I. Chabini and S. Lan, Adaptations of the A* algorithm for the computation of fastest paths in deterministic discrete-time dynamic networks, Trans.
Intell. Transport. Sys., vol. 3, no. 1, pp. 6074, Mar. 2002.
[13] R. K. Ahuja, T. L. Magnanti, and J. B. Orlin, Network Flows: Theory, Algorithms, and Applications, Upper Saddle River, NJ, USA: Prentice-Hall, Inc.,
1993.
[14] M. L. Fredman and R. E. Tarjan, Fibonacci heaps and their uses in improved network optimization algorithms, J. ACM, vol. 34, no. 3, pp. 596615, Jul.
1987.
[15] P. E. Hart, N. J. Nilsson, and B. Raphael, A formal basis for the heuristic determination of minimum cost paths, IEEE Transactions on Systems Science
and Cybernetics, vol. SSC-4(2), pp. 100107, 1968.
[16] R. Geisberger, P. Sanders, D. Schultes, and D. Delling, Contraction hierarchies: Faster and simpler hierarchical routing in road networks, in Proceedings
of the 7th International Conference on Experimental Algorithms, ser. WEA08. Berlin, Heidelberg: Springer-Verlag, 2008, pp. 319333.
[17] R. Bauer and D. Delling, SHARC: Fast and robust unidirectional routing, J. Exp. Algorithmics, vol. 14, pp. 4:2.44:2.29, Jan. 2010.
| 8cs.DS
|
Mon. Not. R. Astron. Soc. 000, 1–18 (2007)
Printed 7 January 2013
(MN LATEX style file v2.2)
Clustered Calibration: An Improvement to Radio Interferometric
Direction Dependent Self-Calibration
S. Kazemi1⋆, S. Yatawatta2 , S. Zaroubi1
1 Kapteyn
Astronomical Institute, University of Groningen, P.O. Box 800, 9700 AV Groningen, the Netherlands
Postbus 2, 7990 AA Dwingeloo, the Netherlands
arXiv:1301.0633v1 [astro-ph.IM] 3 Jan 2013
2 ASTRON,
7 January 2013
ABSTRACT
The new generation of radio synthesis arrays, such as LOFAR and SKA, have been designed to
surpass existing arrays in terms of sensitivity, angular resolution and frequency coverage. This
evolution has led to the development of advanced calibration techniques that ensure the delivery of accurate results at the lowest possible computational cost. However, the performance
of such calibration techniques is still limited by the compact, bright sources in the sky, used
as calibrators. It is important to have a bright enough source that is well distinguished from
the background noise level in order to achieve satisfactory results in calibration. This paper
presents “clustered calibration” as a modification to traditional radio interferometric calibration, in order to accommodate faint sources that are almost below the background noise level
into the calibration process. The main idea is to employ the information of the bright sources’
measured signals as an aid to calibrate fainter sources that are nearby the bright sources. In the
case where we do not have bright enough sources, a source cluster could act as a bright source
that can be distinguished from background noise. For this purpose, we construct a number
of source clusters assuming that the signals of the sources belonging to a single cluster are
corrupted by almost the same errors. Under this assumption, each cluster is calibrated as a
single source, using the combined coherencies of its sources simultaneously. This upgrades
the power of an individual faint source by the effective power of its cluster. The solutions thus
obtained for every cluster are assigned to each individual source in the cluster. We give performance analysis of clustered calibration to show the superiority of this approach compared
to the traditional un-clustered calibration. We also provide analytical criteria to choose the
optimum number of clusters for a given observation in an efficient manner.
Key words: methods: statistical, methods: numerical, techniques: interferometric
1 INTRODUCTION
Low frequency radio astronomy is undergoing a revolution as a
new generation of radio interferometers with increased sensitivity and resolution, such as the LOw Frequency ARray (LOFAR)1 ,
the Murchison Wide-field Array (MWA) 2 and the Square Kilometer Array (SKA)3 are being devised and some are already becoming operational. These arrays form a large effective aperture by the
combination of a large number of antennas using aperture synthesis (Thompson et al. 2001). In order to achieve the full potential
of such an interferometer, it is essential that the observed data is
properly calibrated before any imaging is done. Calibration of radio interferometers refers to the estimation and reduction of errors
introduced by the atmosphere and also by the receiver hardware,
⋆
1
2
3
E-mail: [email protected]
http://www.lofar.org
http://www.mwatelescope.org
http://www.skatelescope.org
c 2007 RAS
before imaging. We also consider the removal of bright sources
from the observed data part of calibration, that enable imaging the
weak background sources. For low frequency observations with a
wide field of view, calibration needs to be done along multiple directions in the sky. Proper calibration across the full field of view is
required to achieve the interferometer’s desired precision and sensitivity giving us high dynamic range images.
Early radio astronomy used external (classical) calibration
which estimates the instrumental unknown parameters using a radio source with known properties. This method has a relatively
low computational cost and a fast execution time, but it generates
results strongly affected by the accuracy with which the source
properties are known. In addition, existence of an isolated bright
source, which is the most important requirement of external calibration, in a very wide field of view is almost impractical, and
even when it does, external calibration would only give information along the direction of the calibrator. The external calibration
was then improved by self-calibration (Pearson & Readhead 1984).
Self-calibration has the advantage of estimating both the source
2
Kazemi et al.
and instrumental unknowns, without the need of having a prior
knowledge of the sky, only utilizing the instrument’s observed data.
Moreover, it can achieve a superior precision by iterating between
the sky and the instrumental parameters.
The accuracy of any calibration scheme, regardless of the
used technique, is determined by the level of Signal to Noise Ratio (SNR). This limits the feasibility of any calibration scheme
to only bright sources that have a high enough SNR to be distinguished from the background noise (Bernardi et al. 2010; Liu et al.
2009; Pindor et al. 2010). Note that interferometric images are produced using the data observed during the total observation (integration) time. However, calibration is done at shorter time intervals of that total duration. This increases the noise level of the data
for which calibration is executed compared to the one in the images. In other words, there are some faint sources that appear well
above the noise in the images while they are well below the noise
level during calibration. It has been a great challenge to calibrate
such faint sources having a very low SNR. In this paper we present
clustered self-calibration, introduced in Kazemi et al. (2011a), and
emphasize its better performance compared to un-clustered calibration below the calibration noise level. Existing un-clustered calibration approaches (Intema et al. 2009; Smirnov 2011) can only handle a handful of directions where strong enough sources are present
and we improve this, in particular for subtraction of thousands of
sources over hundreds of directions in the sky, as in the case for
LOFAR (Bregman 2012).
The implementation of clustered calibration is performed by
clustering the sources in the sky, assuming that all the sources in a
single cluster have the same corruptions, and calibrating each cluster as a single source. At the end, the obtained calibration solutions
for every source cluster is assigned to all the sources in that cluster.
This procedure improves the information used by calibration by incorporating the total of signals observed at each cluster instead of
each individual source’s signal. Therefore, when SNR is very low, it
provides a considerably better result compared to un-clustered calibration. Moreover, calibrating for a set of source clusters instead
of for all the individual sources in the sky reduces the number of
directions along which calibration has to be performed, thus considerably cutting down the computational cost. However, there is
one drawback in clustered calibration: The corruptions of each individual source in one given cluster will almost surely be slightly
different from the corruptions estimated for the whole cluster by
calibration. We call this additional error as “clustering error” and
in order to get the best performance in clustered calibration, we
should find the right balance between the improvement in SNR as
opposed to degradation by clustering error.
Recently, clustering methods have gained a lot of popularity in
dealing with large data sets (Kaufman & Rousseeuw 1990). There
are various clustering approaches that could be applied to calibration based on their specific characteristics. An overview of different
clustering methods is given in Xu & Wunsch (2008). Clustering of
radio sources should take into account (i) their physical distance
from each other and (ii) their individual intensity. The smaller the
angular separations of sources are, the higher the likelihood that
they share the same corruptions in their radiated signals. Moreover, in order to get the best accuracy in the calibration results,
there should be a balance between effective intensities of different
clusters. Thus, in the clustering procedure, every source should be
weighted suitably according to its brightness intensity.
The brightness distribution of radio source in the sky is a
power law and the spatial distribution is Poisson. Therefore, clustering the sources via probabilistic clustering approaches is com-
putationally complex. Hierarchical clustering (Johnson 1967) is a
well-known clustering approach with a straight forward implementation suitable for our case. But, its computational cost grows exponentially with the size of its target data set which can be a disadvantage when the number of sources is huge. Weighted K-means
clustering (Kerdprasop et al. 2005; MacQueen 1967) is also one of
the most used of clustering schemes applicable in clustered calibration. The advantage of this clustering technique is its low computational cost, which is proportional to the number of clusters and the
size of the target data set. Nonetheless, we emphasize that the computational time taken by any of the aforementioned clustering algorithms is negligible compared with the computational time taken
by the actual calibration. Therefore, we pursue all clustering approaches in this paper. However, the use of Fuzzy C-means clustering (Bezdek 1981) in clustered calibration requires major changes
in the calibration data model and will be explored in future work.
This paper is organized as follows: First, in section 2 we
present the general data model used in clustered calibration. In section 3, we present modified weighted K-means and divisive hierarchical clustering for clustering sources in the sky. Next, in section 4, we focus on analyzing the performance of clustered calibration, with an a priori clustered sky model, and compare the results with un-clustered calibration. In clustered calibration, there is
contention between the improvement of SNR by clustering sources
and the errors introduced by the clustering of sources itself. Thus,
we relate the clustered calibration’s performance to the effective
Signal to Interference plus Noise Ratio (SINR) obtained for each
cluster. For this purpose, we use statistical estimation theory and
the Cramer-Rao Lower Bounds (CRLB) (Kay 1993). In section 5,
we derive criteria for finding the optimum number of clusters for a
given sky. We use the SINR analysis and adopt Akaike’s Information Criterion (AIC) (Akaike 1973) and the Likelihood Ratio Test
(LRT) (Graves 1978) to estimate the optimum number of clusters.
We present simulation results in section 6 to show the superiority
of clustered calibration to un-clustered calibration and the performance of the presented criteria in detecting the optimum number
of clusters. Finally, we draw our conclusions in section 7. Through
this paper, calibration is executed by the Space Alternating Generalized Expectation maximization (SAGE) (Fessler & Hero 1994;
Yatawatta et al. 2009; Kazemi et al. 2011b) algorithm. Moreover,
in our simulations, radio sources are considered to be uniformly
distributed in the sky and their flux intensities follow Raleigh distribution, which is the worst case scenario. In real sky models, there
usually exist only a few (two or three) number of bright sources
which dominate the emission. In the presence of these sources and
the background noise, it is impractical to solve for the other faint
sources in the field of view individually. Therefore, obtaining a better calibration performance via the clustered calibration approach,
compared to the un-clustered one, is guaranteed. Therefore, in section 6 we illustrate this using simulations in which the brightness
distribution of sources is a power law with a very steep slope. On
top of that, Kazemi et al. (2012); Yatawatta et al. (submitted 2012)
also present the performance of clustered calibration on real observations using LOFAR.
The following notations are used in this paper: Bold, lowercase letters refer to column vectors, e.g., y. Upper case bold letters
refer to matrices, e.g., C. All parameters are complex numbers, unless stated otherwise. The inverse, transpose, Hermitian transpose,
and conjugation of a matrix are presented by (.)−1 , (.)T , (.)H , and
(.)∗ , respectively. The statistical expectation operator is referred to
as E{.}. The matrix Kronecker product and the proper (strict) subset are denoted by ⊗ and (, respectively. In is the n × n identity
c 2007 RAS, MNRAS 000, 1–18
Clustered Radio Interferometric Calibration
matrix and ∅ is the empty set. The Kronecker delta function is presented by δij . R and C are the sets of Real and Complex numbers,
respectively. The Frobenius norm is shown by ||.||. Estimated pac All logarithmic calculations are
rameters are denoted by a hat, (.).
to the base e. The multivariate Gaussian and Complex Gaussian
distributions are denoted by N and CN , respectively.
2 CLUSTERED SELF-CALIBRATION DATA MODEL
In this section, we present the measurement equation of a polarimetric clustered calibration in detail (Hamaker et al. 1996;
Hamaker 2006). Suppose we have a radio interferometer consisting of N polarimetric antennas where each antenna is composed
of two orthogonal dual-polarization feeds that observe K compact
sources in the sky. Every i-th source signal, i ∈ {1, 2, . . . , K},
e pi = [vXpi vY pi ]T at X and Y
causes an induced voltage of v
dipoles of every p-th antenna, p ∈ {1, 2, . . . , N }. In practice,
T
epi = Jpi ei ,
v
(1)
where ei = [eXi eY i ] is the source’s electric field vector and
Jpi represents the 2 × 2 Jones matrix (Hamaker et al. 1996) corresponding to the direction-dependent gain corruptions in the radiated signal. These corruptions are originated from the instrumental
(the beam shape, system frequency response, etc.) and the propagation (tropospheric and ionospheric distortions, etc.) properties
which later on, in this section, will be explained in more detail.
The signal vp obtained at every antenna p is a linear suepi where i ∈
perposition of the K sources corrupted signals, v
{1, 2, . . . , K}, plus the antenna’s thermal noise. The multitude of
ignored fainter sources also contributes to this additive noise.
The voltages collected at the instrument antennas get corrected
for geometric delays, based on the location of their antennas, and
some instrumental effects, like the antenna clock phases and electronic gains. Then, they are correlated in the array’s correlator to
generate visibilities (Hamaker et al. 1996). The visibility matrix of
the baseline p − q, E{vp ⊗ vqH }, is given by
!
K
X
H
Vpq = Gp
Jpi (θθ )Ci{pq} Jqi (θθ ) GH
(2)
q + Npq .
i=1
In Eq. (2), θ ∈ CP , P = 4KN , is the unknown instrumental
and sky parameter vector, Npq is the additive 2 × 2 noise matrix
of the baseline p − q, and Ci{pq} is the Fourier transform of the
i-th source coherency matrix Ci = E{ei ⊗ eH
i } (Born & Wolf
1999; Hamaker et al. 1996). If the i-th source radiation intensity
is Ii , then Ci = I2i I2 . Considering this source to have equatorial
coordinate, (Right Ascension α, Declination δ), equal to (αi , δi ),
and the geometric components of baseline p − q to be (u, v, w),
then
√
2
2
(3)
Ci{pq} = e−2π(ul+vm+w( 1−l −m −1)) Ci ,
where 2 = −1 and,
l = sin(αi − α0 )cos(δi ),
m = cos(δ0 )sin(δi ) − cos(αi − α0 )cos(δi )sin(δ0 ),
are the source direction components corresponding to the observation phase reference of (α0 , δ0 ) (Thompson et al. 2001). The errors
common to all directions, such as the receiver delay and amplitude
errors, are given by Gp and Gq . Initial calibration at a finer time
c 2007 RAS, MNRAS 000, 1–18
3
and frequency resolution is performed to estimate and correct for
Gp -s and the corrected visibilities are obtained as
−H
e pq = G−1
V
.
p Vpq Gq
(4)
The remaining errors are unique to a given direction, but residual
errors in Gp -s are also absorbed into these errors, which are denoted by Jpi in the usual notation.
Vectorizing Eq. (4), the final visibility vector of the baseline
p − q is given by
vpq =
K
X
i=1
J∗qi (θθ ) ⊗ Jpi (θθ )vec(Ci{pq} ) + npq .
(5)
Stacking up all the cross correlations (measured visibilities) and
noise vectors in y and n, respectively, the un-clustered selfcalibration measurement equation is given by
y=
K
X
si (θθ ) + n.
(6)
i=1
In Eq. (6), y, n ∈ CM , M = 2N (N − 1), the noise vector is considered to have a zero mean Gaussian distribution with covariance
Π , n ∼ N (0, Π M ×M ), and the nonlinear function si (θθ ) is defined
as
J∗2i (θθ ) ⊗ J1i (θθ )vec(Ci{pq} )
J∗3i (θθ ) ⊗ J1i (θθ )vec(Ci{pq} )
(7)
si (θθ ) ≡
.
..
.
J∗Ni (θθ ) ⊗ J(N−1)i (θθ )vec(Ci{pq} )
Calibration is essentially the Maximum Likelihood (ML) estimation of the unknown parameters θ (P complex values or 2P real
values), or of the Jones matrices J(θθ ), from Eq. (6) and removal of
the K sources. Note that calibration methods could also be applied
to the uncorrected visibilities of (4) to estimate Gp and Gq errors
as well.
The Jones matrix Jpi , for every i-th direction and at every p-th
antenna, is given as
Jpi ≡ Epi Zpi Fpi .
(8)
In Eq. (8), Epi , Zpi , and Fpi are the antenna’s voltage pattern,
ionospheric phase fluctuation, and Faraday Rotation Jones matrices, respectively. In practice, the E, Z, and F Jones matrices obtained for nearby directions and for a given antenna are almost the
same. Thus, for every antenna p, if the i-th and j-th sources have a
small angular separation from each other, we may consider
Jpi ∼
= Jpj .
(9)
This is the underlying assumption for clustered calibration.
Clustered calibration first assigns source clusters, Li for i ∈
{1, 2, . . . , Q} where Q ≪ K, on which the sky variation is considepi which
ered to be uniform. Then, it assumes there exists a unique J
is shared by all the sources of the i-th cluster Li , i ∈ {1, 2, . . . , Q},
at receiver p, p ∈ {1, 2, . . . , N }. Based on that, the visibility at every baseline p − q, given by Eq. (4), is reformulated as
e pq =
V
Q
X
i=1
epi (θe){
J
X
l∈Li
e
eH
e
Cl{pq} }J
qi (θ ) + Npq .
(10)
e pq is the clustered calibration’s effective noise at
In Eq. (10), N
baseline p − q which will be explicitly discussed at section 4. Note
that clustered calibration estimates the new unknown parameter
4
Kazemi et al.
e
θe ∈ CP where Pe = 4QN . Denoting the effective signal of every i-th cluster at baseline p − q by
X
e i{pq} ≡
C
Cl{pq} ,
(11)
l∈Li
the clustered calibration visibility vector at this baseline (vectorized
form of Eq. (10)) is
vpq =
Q
X
i=1
epi (θe)vec(C
e i{pq} ) + n
e∗qi (θe) ⊗ J
e pq .
J
(12)
Finally, stacking up the visibilities of all the instrument’s baselines
in vector y, clustered calibration’s general measurement equation
is resulted as
y=
Q
X
i=1
e
e.
si (θe) + n
(13)
In Eq. (13) e
si is defined similar to si in Eq. (7) where J and C are
e and C,
e respectively.
replaced by J
Because of the similarity between the clustered and the unclustered calibration’s measurement equations presented by Eq.
(13) and Eq. (6), respectively, they could utilize the same calibration techniques. Thus, the only difference between these two
types of calibration is that clustered calibration solves for clusters
of sources instead of for the individual ones. That upgrades the signals which should be calibrated for from Eq. (3) to Eq. (11).
3 CLUSTERING ALGORITHMS
Clustering is grouping a set of data so that the members of the same
group (cluster) have some similarities (Kaufman & Rousseeuw
1990). This similarity is defined based on the application of the
clustering method.
We need to define clustering schemes in which two radio
sources merge to a single cluster based on the similarity in their
direction dependent gain errors ( Eq. (8) ). Radiations of sources
that are close enough to each other in the sky are assumed to be
affected by almost the same corruptions ( Eq. (9) ). Based on that,
we aim to design source clusters with small angular diameters. On
the other hand, every cluster’s intensity is the sum of the intensities
of its members ( Eq. (11) ). In order to keep a balance between different clusters’ intensities, we intend to apply weighted clustering
techniques in which the sources are weighted proportional to their
intensities.
Suppose that the K sources, x1 , x2 , . . . , xK have
(α1 , δ1 ), (α2 , δ2 ), . . . , (αK , δK ) equatorial coordinates, respectively. TheP
aim is to provide Q clusters so that the objective
Q
function f =
q=1 D(Lq ) is minimized. D(Lq ) is the angular
diameter of cluster Lq , for q ∈ {1, 2, . . . , Q}, defined as
D(Lq ) ≡ max {d(xi , xj )|xi , xj ∈ Lq },
For defining the centroids, we associate a weight to every source
xi , as
wi = w(xi ) ≡
Ii
,
I∗
for i ∈ {1, 2, . . . , K},
(16)
where Ii is the source’s intensity and I ∗ = min {I1 , I2 , . . . , IK }.
Applying the weights to the clustering procedure, the centroids of
the clusters lean mostly towards the brightest sources. That causes
a tendency in faint sources to gather with brighter sources close
to them into one cluster. Thus, their weak signals are promoted
being added up with some brighter sources’ signals. Moreover, very
strong sources will be isolated such that their signals are calibrated
individually, without being affected by the other faint sources.
We cluster radio sources using weighted K-means
(Kerdprasop et al. 2005) and divisive hierarchical clustering
(Johnson 1967) algorithms. Since the source clustering for calibration is performed offline, its computational complexity is
negligible compared with the calibration procedure itself.
3.1 Weighted K-means clustering
Step 1. Select the Q brightest sources, x1∗ , x2∗ , . . . , xQ∗ , and initialize the centroids of Q clusters by their locations as
for q ∈ {1, 2, . . . , Q}, q ∗ ∈ {1∗ , 2∗ , . . . , Q∗ }.
(17)
Step 2. Assign each source to the cluster with the closest centroid,
defining the membership function
(
1, if d(xi , cq ) = min{d(xi , cj )|j = 1, . . . , Q}
mLq (xi ) =
0, Otherwise
cq ≡ [αq∗ , δq∗ ],
Step 3. Update the centroids by
PK
mLq (xi ) wi xi
cq = Pi=1
,
K
i=1 mLq (xi ) wi
for q ∈ {1, 2, . . . , Q}.
(18)
Repeat steps 2 and 3 until there are no reassignments of sources to
clusters.
3.2 Divisive hierarchical clustering
Step 1. Initialize the cluster counter Q′ to 1, assign all the K
sources to a single cluster L1 and ∅ to a set of null clusters A.
Step 2. Choose cluster Lq∗ , for q ∗ ∈ {1, 2, . . . , Q′ } − A, with the
largest angular diameter
D(Lq∗ ) = max{D(Lq )|q ∈ {1, 2, . . . , Q′ } − A}.
(19)
Step 3. Apply the presented weighted K-means clustering technique to split Lq∗ into two clusters, L′q∗ and L′′q∗ .
Step 4. If D(L′q∗ ) + D(L′′q∗ ) < D(Lq∗ ), then set Q′ = Q′ + 1,
Lq∗ ≡ L′q∗ , LQ′ ≡ L′′q∗ , and A = ∅, otherwise set A = A∪{q ∗ }.
Repeat steps 2, 3, and 4 until Q′ = Q.
(14)
and d(., .) is the angular separation between any two points on the
celestial sphere. Having two radio sources a and b with equatorial
coordinates (αa , δa ) and (αb , δb ), respectively, the angular separation d(a, b), in radians, is obtained by
p
cos2 δb sin2 ∆α + [cosδa sinδb − sinδa cosδb cos∆α]2
−1
tan
,
sinδa sinδb + cosδa cosδb cos∆α
(15)
where ∆α = αb − αa .
3.3 Comparison of clustering methods
Hierarchical clustering method tends to design clusters with almost the same angular diameters, whereas, the K-means clustering
method tends to keep the same level of intensity at all its clusters.
In practice, since hierarchical clustering method makes less errors
in dedicating the same solutions to sources in small clusters, it performs better than Weighted K-means clustering in a clustered calibration procedure. But, when the number of sources (and clusters)
c 2007 RAS, MNRAS 000, 1–18
42
42
41
41
40
40
δ [deg.]
δ [deg.]
Clustered Radio Interferometric Calibration
39
38
39
38
37
37
36
36
35
278
276
274
α [deg.]
272
35
270
Figure 1. A simulated 8 by 8 degrees sky of fifty point sources with intensities below 3 Jy. The source positions and their brightness are following
uniform and Rayleigh distributions, respectively. The marker sizes are proportional to sources intensities.
In this section, we explain the reasons for clustered calibration’s
better performance, compared to un-clustered calibration, at a low
c 2007 RAS, MNRAS 000, 1–18
276
274
α [deg.]
272
270
42
41
3.3.1 Example 1: Weighted K-means and hierarchical clustering
4 PERFORMANCE ANALYSIS
278
Figure 2. Fifty point sources are clustered into ten source clusters by
Weighted K-means clustering technique. There is not a good balance between different clusters angular diameters.
is very large (Q > 100), its prohibitive computational costs makes
the fast K-means clustering method preferable.
40
δ [deg.]
We simulate an 8 by 8 degrees sky with fifty point sources with intensities below 3 Jansky (Jy). The source positions and their brightness follow uniform and Rayleigh distributions, respectively. The
result is shown by Fig. 1 in which the symbol sizes are proportional
to intensities of sources. Weighted K-means and divisive hierarchical clustering methods are applied to cluster the fifty sources into
ten source clusters. The results are presented in Fig. 2 and Fig. 3,
respectively. Fig. 2 shows that the Weighted K-means clustering
could design source clusters with considerably large angular diameters. Assigning the same calibration solutions to the sources of
these large clusters could cause significant errors. However, as Fig.
3 shows, this is not the case for the hierarchical clustering and it
constructs clusters with almost the same angular diameters.
Since the number of sources in this simulation is not that large
(K = 50), the difference between execution time of the two clustering methods is not significant. Hence in such a case, the use of hierarchical clustering method, rather than the Weighted K-means, is
advised. However, this is not the case when we have a large number
of sources, and subsequently a large number of source clusters, in
the sky. To demonstrate this, we use the two clustering techniques
for clustering thousand of sources (K = 1000) into Q source clusters, Q ∈ {3, 4, . . . , 100}. The methods’ computational times versus the number of clusters are plotted in Fig. 4. As Fig. 4 shows,
for large Qs, the computational cost of Weighted K-means is much
cheaper than the one of the hierarchical clustering. That can make
the Weighted K-means clustering method more suitable than the
hierarchical clustering for such a case.
5
39
38
37
36
35
278
276
274
α [deg.]
272
270
Figure 3. Fifty point sources are clustered into ten source clusters via hierarchical clustering method. Different clusters have almost the same angular
diameters.
SNR (Kazemi et al. 2012). This superiority is in the sense of
achieving an unprecedented precision in solutions with a considerably low computational complexity, given the optimum clustering
scheme. In the next section, we present different criteria for finding
the optimum number of clusters at which the clustered calibration
performs the best.
4.1 Cramer-Rao Lower Bounds
The most fundamental assumption in clustered calibration is that
the sources at the same cluster have exactly the same corruptions
in their radiated signals. This assumption is of course incorrect,
nonetheless, it provides us with a stronger signal, the sum of the
signals in the whole cluster. We present an analytic comparison of
clustered and un-clustered calibration where we use the Cramer-
6
Kazemi et al.
Computational time (s)
800
In other words, the variance of any unbiased estimator of the unknown parameter vector θ is bounded from below by the diagonal
elements of [I(θθ )]−1 .
Using (23) and (26), the Fisher information matrix of the visibility vector y is obtained as
Weighted K−means
Divisive Hierarchical
600
I(θθ ) = 2σ −2 Re(JsH Js ),
400
where Js is the Jacobian matrix of s with respect to θ
Js (θθ ) =
200
0
0
(27)
20
40
60
Number of clusters
80
100
Figure 4. Weighted K-means and divisive hierarchical clustering methods
computational costs. For small number of source clusters, there is no difference between execution times of the two clustering methods. But, when
the number of source clusters is large, the computational cost of Weighted
K-means becomes much cheaper than the one of the hierarchical clustering.
Rao Lower Bound (CRLB) (Kay 1993) as a tool to measure the
performance of the calibration.
4.1.1 Estimations of CRLB for two sources at a single cluster
For simplicity, first consider observing two point sources at a single
baseline, lets say baseline p − q. Based on Eq. (4), the visibilities
are given by
H
e pq = Jp1 C1{pq} JH
V
q1 + Jp2 C2{pq} Jq2 + Npq ,
(20)
e pq , the visiin the un-clustered calibration strategy. Vectorizing V
bility vector is
y = J∗q1 ⊗ Jp1 vec(C1{pq} ) + J∗q2 ⊗ Jp2 vec(C2{pq} ) + npq .
Assuming npq ∼ CN (0, σ 2 I4 ), we have
y ∼ CN (s(θθ ), σ 2 I4 ),
(21)
where
s(θθ ) ≡
X
i=1,2
J∗qi (θθ ) ⊗ Jpi (θθ )vec(Ci{pq} ).
(22)
Using Eq. (21), the log-likelihood function of the visibility vector
y is given by
π
L(θθ |y) = −4ln{ 2 } − σ −2 (y − s(θθ ))H (y − s(θθ )).
(23)
σ
CRLB is a tight lower bound on the error variance of any unbiased parameter estimators (Kay 1993). Based on its definition, if
the log-likelihood function of the random vector y, L(θθ |y), satisfies the “regularity” condition
∂
Ey
L(θθ |y) = 0, for all θ ,
(24)
∂θθ
then for any unbiased estimator of θ , θb,
var(θbi ) > [I −1 (θθ )]ii ,
for i ∈ {1, . . . , M },
where I(θθ ) is the Fisher information matrix defined as
2
∂ L(θθ |y)
I(θθ ) = − Ey
.
∂θθ ∂θθ T
(25)
(26)
2
X
∂
{J∗ ⊗ Jpi }[I4 ⊗ vec(Ci{pq} )].
θ qi
∂θ
i=1
(28)
Thus, variations of any unbiased estimator of parameter vector θ ,
lets say θb, is bounded from below by the CRLB as
Var(θb) > [2σ −2 Re(JsH Js )]−1 .
(29)
Lets try to bound the error variations of the clustered calibration parameters assuming that the two sources construct a single
cluster, called cluster number 1. We reform Eq. (20) as
e p1 (C1{pq} +C2{pq} )J
eH
Vpq = J
q1 +Γ1{pq} +Γ2{pq} +Npq , (30)
where Γi{pq} , referred to as the “clustering error” matrices, are
given by
e
eH
Γi{pq} = Jpi Ci{pq} JH
qi − Jp1 Ci{pq} Jq1 ,
(31)
e pq ≡ Γ1{pq} + Γ2{pq} + Npq .
N
(32)
e∗q1 ⊗ J
ep1 vec(C1{pq} + C2{pq} ) + n
e pq ,
y=J
(33)
ep1 (θe) is the clustered calibration solution at receiver p.
and J
e pq in
Eq. (30) implies that what is considered as the noise matrix N
the clustered calibration data model, Eq. (10), is in fact
Vectorizing Eq. (30), the clustered calibration visibility vector is
obtained by
e pq ).
e pq = vec(N
where n
We point out that depending on the observation as well as the
positions of the two sources on the sky, the clustering error Γi{pq}
will have different properties. However, in order to study the performance of the clustered calibration in a statistical sense, and to
simplify our analysis, we make the following assumptions.
(i) Consider statistical expectation over different observations
and over different sky realizations where the sources are randomly
e → E{J}
distributed on the sky. In that case, almost surely E{J}
and consequently
E{Γi{pq} } → 0.
(34)
In other words, we assume the clustering error to have zero mean
over many observations of different parts of the sky.
(ii) We assume that the closer the sources are together in the sky,
the smaller the errors introduced by clustering would be. Therefore,
given a set of sources, the clustering error will reduce as the number of clusters increase. In fact this error introduced by clustering
vanishes when the number of clusters is equal to the number of
sources (each cluster contains only one source). Therefore, given a
set of sources, the variance of Γi{pq} will decrease as the number
of clusters increase.
Using Eq. (34) and bearing in mind that E{Npq } = 0, we
eH
e pq ∼ CN (0, σ
e 2 I4 .
can consider n
e 2 I4 ) where E{e
npq n
pq } = σ
Therefore,
y ∼ CN (e
s, σ
e 2 I4 ),
c 2007 RAS, MNRAS 000, 1–18
Clustered Radio Interferometric Calibration
The CRLBs obtained for the un-clustered and clustered calibrations in Eq. (29) and Eq. (35), respectively, are both almost
equal to the inverse of the Signal to Interference plus Noise Ratio
(SINR), SINR−1 . In un-clustered calibration, the effective signal
for the faintest source is C2{pq} where the noise is Npq . Therefore, SINR for this source is
2000
Clustered calibration
Un−clustered calibration
CRLB
1500
1000
||C2{pq} ||2
.
||Npq ||2
SINR2 =
(36)
But, in clustered calibration, the effective signal and noise are
e {pq} ≡ C1{pq} + C2{pq} and N
e pq , respectively. Thus, SINR
C
for the cluster is
500
0
0
7
1000
2000
3000
4000
Clustered calibration effective noise power
5000
Figure 5. Clustered and un-clustered calibrations CRLB. When the effece 2 , is small enough, then
tive noise power of the clustered calibration, ||N||
its CRLB is lower than of the un-clustered calibration’s and it reveals a
superior performance.
e∗q1 ⊗ J
e p1 vec(C1{pq} + C2{pq} ),
e
s≡J
and similar to Eq. (29), we have
b
σ −2 Re(JesH Jes )]−1 .
Var(θe) > [2e
(35)
We use numerical simulations to compare the un-clustered and
clustered calibrations performances via their CRLBs which are
given by Eq. (29) and Eq. (35), respectively.
4.1.2 Example 2: CRLB for two sources and one cluster
We simulate a twelve hour observation of two point sources with
intensities I1 = 11.25 and I2 = 2.01 Jy at sky coordinates (l, m)
equal to (−0.014, −0.005) and (−0.011, −0.010) radians, respectively. We use the uv-coverage of Westerbork Synthesis Radio Telescope (WSRT) with 14 receivers in this simulation.
We Consider the J Jones matrices in Eq. (20) to be diagonal.
Their amplitude and phase elements follow U(0.75, 0.95) and
U(0.003, 0.004) distributions, respectively. The background noise
is N ∼ CN (0, 10I). Jones matrices of the clustered calibration,
e p1 for p = 1, 2, are obtained as Jp1 + U(0.02, 0.40)ejU (0.5,2) .
J
e matrices, we calculated CRLB of the unFor 20 realizations of J
clustered and clustered calibrations using Eq. (29) and Eq. (35),
respectively. The results are presented in Fig. 5. As shown in this
figure, for small enough errors matrices Γ of Eq. (31), the clustered calibration’s CRLB stands below the un-clustered calibration’s CRLB. On the other hand, with increasing power of error mae the un-clustered calibratrices, or the power of effective noise N,
tion’s CRLB becomes lower than the clustered calibration’s CRLB.
4.1.3 Analysis of CRLB
Generally, if source 1 is considerably brighter than source 2,
||C1{pq} || ≫ ||C2{pq} ||, and if the weak source power is much
lower than the noise level, ||C2{pq} || ≪ ||Npq ||, then clustered
calibration’s performance is better than un-clustered calibration.
Note that the worst performance of both calibrations is at the
faintest source and we are more concerned to compare the CRLBs
for this source.
c 2007 RAS, MNRAS 000, 1–18
SINRc =
e {pq} ||2
||C
.
e pq ||2
||N
(37)
Clustered calibration has an improved performance when
SINRc ≫ SINR2 .
(38)
Consider the two possible extremes in a clustered calibration
procedure:
(i) Clustering many sources in a large field of view to a very
small number of clusters. In this case, the angular diameter of a
cluster is probably too large for the assumption of uniform corruptions to apply. Subsequently, dedicating a single solution to all the
sources of every cluster by clustered calibration introduces clustering error matrices Γ with a large variance (see Eq. (31)). Having
high interference power, the clustered calibration effective noise
e of Eq. (32) becomes very large. Therefore, clustered calibration
N
SINR will be very low and it does not produce high quality results.
(ii) Clustering sources in a small field of view to a very large
number of clusters. In this case, the variance of Γ matrices are almost zero while the signal powers of source clusters are almost as
low as of the individual sources. Therefore, the SINR of clustered
calibration is almost equal to the un-clustered calibration SINR and
the calibration performance is expected to be almost the same as
well.
Thus, the best efficiency of clustered calibration is obtained at the
smallest number of clusters for which Eq. (38) is satisfied. We use
the SINR of Eq. (38) as an efficient criterion for detecting the optimum number of clusters.
4.1.4 Generalization to many sources and many clusters
For the visibility vector y of un-clustered calibration’s general data
model, presented by Eq. (6), we have
y ∼ CN (
K
X
si (θθ ), Π ).
(39)
i=1
Therefore, the CRLB of un-clustered calibration is
X
−1
K
K
X
var(θθ ) > 2 Re (
,
Jsi (θθ ))H Π −1 (
Jsi (θθ ))
i=1
i=1
where Jsi is the Jacobian matrix of si with respect to θ .
Computing the exact CRLB is more complicated when we
have clustered calibration. In the clustered calibration measurement
equation, given by Eq. (13), we have
e≡
n
K
X
i=1
Γ i + n,
(40)
8
Kazemi et al.
−10
where n is the un-clustered calibration’s noise vector,
(41)
and Γ i{pq} is given by Eq. (31). Due to the existence of the nuisance
parameters Γ i in the clustered calibration data model, calculation
of its conventional CRLB is impractical. This leads us to the use of
Cramer-Rao like bounds devised in the presence of the nuisance parameters (Gini & Reggiannini 2000). We apply the Modified CRLB
(MCRLB) (Gini et al. 1998) to the performance of clustered calibration.
b̃
The MCRLB for estimating the errors of θ in the presence of
the nuisance parameters Γ (clustering error) is defined as
−1
∂ ∂
b̃
Γ
θ
var(θ ) > Ey,ΓΓ − Ey|ΓΓ
ln{P
(y|Γ
;
θ̃
)}
,
T
∂θ̃θ ∂θ̃θ
(42)
Γ; θ̃θ ) is the Probability Density Function (PDF) of the
where P (y|Γ
visibility vector y assuming that the Γ matrices of Eq. (41) are a
priori known. Since n ∼ CN (0, Π ), from Eq. (13) we have
Q
K
X
X
Γ ∼ CN ([
Γ i ], Π ),
y|Γ
s̃i +
2
1
0
∂
∂
T
θ ∂θ̃
∂θ̃
θ
Γ
θ
ln{P (y|Γ ; θ̃ )} , which
20
30
Number of clusters
40
50
−6
10
Q
Q
K
K
X
X
X
X
JΓ i (θ̃θ )]}.
JΓ i (θ̃θ )]H Π −1 [
Js̃i (θ̃θ ) +
2Re{[
Js̃i (θ̃θ ) +
i=1
Clustered calibration
Un−clustered calibration
i=1
Note that Ey,ΓΓ in Eq. (42) could be estimated by Monte-Carlo
method.
As a rule of thumb, reducing the heavy computational cost of
MCRLB, one can interpret the SINR test of Eq. (38) as follows: If
in average the effective SINR of clustered calibration, SINRc , gets
higher than the effective SINR of un-clustered calibration obtained
for the weakest observed signal, SINRw ,
E{SINRc } ≫ E{SINRw },
x 10
9
8
SINR
i=1
10
Figure 6. Clustered calibration’s MCRLB. For very small Q, where the
effect of interference is large, MCRLB is high. By increasing the number
of clusters, MCRLB decreases and reaches its minimum where the best
performance of the clustered calibration is expected. After that, due to the
dominant effect of the background noise, MCRLB starts to increase until it
reaches the un-clustered calibration CRLB.
is called the modified Fisher information, is equal to
i=1
3
i=1
i=1
and therefore in Eq. (42), − Ey|ΓΓ
(43)
Clustered calibration
Un−clustered calibration
4
MCRLB
Γi{12} )T . . . vec(Γ
Γi{(N−1)N} )T ]T ,
Γ i = [vec(Γ
5 x 10
7
6
5
4
(44)
then clustered calibration can achieve a better results. In Eq. (44),
the expectation is taken with respect to the noise N, error matrices
Γ , and all the baselines.
4.1.5 Example 3: MCRLB and SINR estimations
We simulate WSRT including N = 14 receivers which observe
fifty sources with intensities below 15 Jy. The source positions
and their brightness follow uniform and Rayleigh distributions, respectively. The background noise is N ∼ CN (0, 15IM ), where
M = 2N (N − 1) = 364. We cluster sources using divisive
hierarchical clustering, into Q number of clusters where Q ∈
e are gener{3, 4, . . . , 50}. Clustered calibration’s Jones matrices, J,
jU (0,0.2)
ated as U(0.9, 1.1)e
. Since for smaller number of clusters,
we expect larger interference (errors)
soP in clustered calibration’s
150
lutions, for every Q, we consider 50
i=1 Γi ∼ CN (0, Q IM ). The
choice of the complex Gaussian distribution for the error matrices
Γ is due to the central limit theorem and the assumptions made in
section 4.1.1.
We proceed to calculate the clustered calibration’s MCRLB,
given by Eq. (42), and E{SINRc }, utilizing the Monte-Carlo
method. Jacobian matrices for MCRLB are calculated numerically
and in computation of E{SINRc }, signal power of every cluster is
obtained only using the cluster’s brightest and faintest sources. The
3
0
10
20
30
Number of clusters
40
50
Figure 7. Clustered calibration’s SINR. SINR is low for small Q, when
the interference is large. By increasing the number of clusters the SINR
increases and gets its highest level for which the best performance of the
calibration is expected. After that, it decreases due to the dominant effect of
the background noise, and converges to the un-clustered calibration SINR.
estimated results of MCRLB and E{SINRc } are presented by Fig.
6 and Fig. 7, respectively. As we can see in Fig. 6, for very small
Q, where the effect of interference is large, MCRLB is high. By increasing the number of clusters, MCRLB decreases and reaches its
minimum where the best performance of the clustered calibration is
expected. After that, due to the dominant effect of the background
noise, MCRLB starts to increase until it reaches the CRLB of unclustered calibration. The same result is derived from E{SINRc }
plot of Fig. 7. As Fig. 7 shows, E{SINRc } is low for very small Q,
when the interference (i.e., the error due to clustering) is large. By
increasing the number of clusters, E{SINRc } increases and gets its
peak for which the clustered calibration performs the best. After
that, it decreases and converges to the E{SINR} of un-clustered
calibration.
c 2007 RAS, MNRAS 000, 1–18
Clustered Radio Interferometric Calibration
4.2 Computational cost
In the measurement equation of un-clustered calibration, presented
in Eq. (6), we have M = 2N (N − 1) constraints given by the
visibility vector y, and need to solve for P = 4KN unknown
parameters θ . If P > M , then Eq. (6) will be an under-determined
non-linear system. This clarifies the need of having a small enough
N (number of antennas) and a large enough K (number of sources)
for estimating θ . However, clustered calibration, Eq. (13), has the
advantage of decreasing the number of directions, K, relative to the
number of source clusters, Q ≪ K. This considerably cuts down
the number of unknown parameters P that needs to be calibrated,
thus reducing the computational cost of calibration.
5 SELECTION OF NUMBER OF CLUSTERS
Consider a clustered calibration procedure with a predefined clustering scheme. There is no guarantee that the calibration results
for Q number of clusters, where Q ∈ {1, 2, . . . , K} is randomly
chosen, is the most accurate. Thus, we seek the optimum number
of clusters at which the clustered calibration performs the best. In
this section, we describe the use of: (i) Akaike’s Information Criterion (AIC) (Akaike 1973), as well as (ii) Likelihood-Ratio Test
(LRT) (Graves 1978), in finding this optimum Q for a given observation. Some other alternative criteria could also be found in
Wax & Kailath (1985). Note that for different clustering schemes
the optimum Q is not necessarily the same.
We utilize Akaike’s Information Criterion (AIC) to find the optimum Q for clustered calibration.
e ∼ CN (0, σ
Consider having n
e 2 IM ) in the general data model of
clustered calibration, Eq. (13). Then, the log-likelihood of the visibility vector y is given by
=
−
−M log π − M log σ
e2
Q
Q
X
X
1
e))H (y −
e
e
(y
−
s
(
θ
si (θe)).
i
σ
e2
i=1
i=1
1
c
(y −
σ
e2 =
M
i=1
e
si (θe))H (y −
Q
X
i=1
=
−M log π − M
−
M log{
e
si (θe)).
(46)
Q
Q
X
X
1
e
e
(y −
si (θe))H (y −
si (θe))}. (47)
M
i=1
i=1
Using Eq. (47), the AIC is given by
b
AIC(Q) = −2L(θe|y) + 2(2Pe).
(48)
The optimum Q is selected as the one that minimizes AIC(Q).
5.2 Likelihood-Ratio Test (LRT)
Errors in clustered calibration originate from the system (sky and
instrumental) noise, “clustering errors” introduced in section 4.1.1,
c 2007 RAS, MNRAS 000, 1–18
θepi = [Re(Je11,p ) Im(Je11,p ) . . . Re(Je22,p ) Im(Je22,p )]Ti . (50)
T
T
zpkl = [θepk θepl ]T .
(51)
In fact, we are concatenating the solutions of the same antenna for
two different directions (clusters) together.
We define the null H0 model as
(45)
Substituting Eq. (46) in Eq. (45), we arrive at the maximum likelihood estimation of θe,
b
L(θe|y)
i
Then, the parameter vector θepi (corresponding to the i-th direction
and p-th antenna) is obtained by
H0 : zpkl ∼ N (m, Σ0 ).
(52)
m = [m̄(θepk )T m̄(θepl )T ]T ,
(53)
where
The maximum likelihood estimation of the noise variance σ
e 2 is
Q
X
and “solver noise” which is referred to as errors produced by the
calibration algorithm itself. We assume that the true Jones matrices
along different directions (clusters) at the same antenna are statistically uncorrelated. Note that this assumption is only made for
the LRT to produce reasonable results and this assumption is not
needed for clustered calibration to work. Therefore, if such correlations exist, they are caused by the aforementioned errors. Consequently, the more accurate the clustered calibration solutions are,
the smaller their statistical similarities would be. Based on this general statement, the best number of clusters in a clustered calibration
procedure is the one which provides us with the minimum correlations in the calibrated solutions. Note that for a fixed measurement,
the correlation due to the system noise is fixed. Therefore, differences in the statistical similarities of solutions obtained by different
clustering schemes are only due to “clustering errors” and “solver
noise”.
To investigate the statistical interaction between the gain solutions we apply the Likelihood-Ratio Test (LRT).
e pi (θe) for directions i,
Consider the clustered calibration solution J
i ∈ {1, 2, . . . , Q}, at antennas p, where p ∈ {1, 2, . . . , N },
"
#
Je11,p Je12,p
e
.
(49)
Jpi =
Je21,p Je22,p
Let us define for each antenna p and each pair of directions k and
l, where k and l are belong to {1, 2, ..., Q}, a vector zpkl as
5.1 Akaike’s Information Criterion (AIC)
L(θe|y)
9
and
Σ0 =
"
s2 (θepk )
0
0
s2 (θepl )
#
.
(54)
In Eq. (53) and Eq. (54), m̄ and s2 are denoting sample mean and
sample variance, respectively. Note that having a large number of
samples in hand, the assumption of having a Gaussian distribution
for solutions is justified according to the Central Limit theorem.
The structure of the variance matrix Σ0 tells us that the statistical
correlation between the components of the random vector zqkl , or
between the solutions θepk and θepl , is zero. This is the ideal case in
which there are no estimation errors.
To investigate the validity of the null model compared with the
case in which there exists some correlation between the solutions,
we define the alternative H1 model as
H1 : zpkl ∼ N (m, Σ1 ),
where the variance matrix Σ1 is given by
s2 (θepk )
Cov(θepk , θepl )
.
Σ1 =
Cov(θepk , θepl )T
s2 (θepl )
(55)
(56)
10
Kazemi et al.
Cov(θepk , θepl ) in Eq. (56) is the 8 × 8 sample covariance matrix.
Using the above models, the Likelihood-Ratio is defined as
Likelihood for null model
Λ = −2ln
,
(57)
Likelihood for alternative model
which has a χ2 distribution with 64 degrees of freedom. As Λ becomes smaller, the null model, in which the statistical correlation
between the solutions is zero, becomes more acceptable rather than
the alternative model. Therefore, the smaller the Λ is, the less the
clustered calibration’s errors are, and vice-versa.
6 SIMULATION STUDIES
We use simulations to compare the performance of un-clustered
and clustered calibration. Working with simulations has the advantage of having the true solutions available, which is not the case in
real observations. That makes the comparison much more objective. Nevertheless, the better performance of the clustered calibration in comparison with the un-clustered ones in calibrating for real
observations of LOFAR is also shown by Kazemi et al. (2011a);
Yatawatta et al. (submitted 2012).
We simulate an East-West radio synthesis array including 14
antennas (similar to WSRT) and an 8 by 8 degrees sky with fifty
sources with low intensities, below 3 Jy. The source positions and
their brightness follow uniform and Rayleigh distributions, respectively. The single channel simulated observation at 355 MHz is
shown in Fig. 8.
We proceed to add gain errors, multiplying source coherencies
by the Jones matrices, as it is shown in Eq. (2), to our simulation.
The amplitude and phase of the Jones matrices’ elements are generated using linear combination of sine and cosine functions. We aim
at simulating a sky with almost uniform variations on small angular
scales. In other words, we provide very similar Jones matrices for
sources with small angular separations. To accomplish this goal, for
every antenna, we first choose a single direction as a reference and
simulate its Jones matrix as it is explained before. Then, for the
remaining forty nine sources, at that antenna, the Jones matrices
(amplitude and phase terms) are that initial Jones matrix multiplied
by the inverse of their corresponding angular distances from that
reference direction. The result of adding such gain errors to our
simulation is shown in Fig. 9.
6.1 Performance comparison of the Clustered and
un-clustered calibrations at SNR=2
We add noise n ∼ CN (0, σ 2 I) with σ 2 = 28, as it is shown in Eq.
(6), to our simulation. The result has a SN R = 2 and is presented
in Fig. 10. We have chosen to present the case of SN R = 2 since
for this particular simulated observation both the divisive hierarchical and the weighted K-means clustered calibrations achieve their
best performances at the same number of clusters, as will be shown
later in this section.
We apply un-clustered and clustered calibration on the simulation to compare their efficiencies. The fifty sources are grouped
into Q ∈ {3, 4, . . . , 49} number of clusters, using the proposed
divisive hierarchical and weighted K-means clustering algorithms.
Self-calibration is implemented via Space Alternating Generalized Expectation Maximization (SAGE) algorithm (Fessler & Hero
1994; Yatawatta et al. 2009; Kazemi et al. 2011b) with nine iterations. Plots of the averaged Frobenius distance between the simulated (true) Jones matrices and the obtained solutions is shown in
Fig. 11. As we can see in Fig. 11, for both clustering schemes, increasing the number of clusters decreases this distance and the minimum is reached at approximately thirty three clusters (Q = 33).
Beyond this number of clusters, it increases until the fifty individual
sources become individual clusters. This shows that the best performance of both the divisive hierarchical and the weighted K-means
clustered calibrations is approximately at thirty three clusters and
is superior to that of the un-clustered calibration.
The Frobenius distance curves in Fig. 11, the MCRLB curve in
Fig. 6, and the SINR curve in Fig. 7 illustrate that clustered calibration with an extremely low number of clusters does not necessarily
perform better than the un-clustered calibration. The reason is that
when there are only a small number of clusters, the interference, or
the so-called “clustering errors” introduced in section 4.1.1, is relatively large. Therefore, the effect of this interference dominates the
clustering of signals. On the other hand, we reach the theoretical
performance limit approximately after twenty five number of clusters and therefore increasing the number beyond this point gives
highly variable results, mainly because we are limited by the number of constraints as opposed to the number of parameters that we
need to solve for. But, this is not the case for the plots in Fig. 6 and
Fig. 7. The reason is that the MCRLB results of Fig. 6 as well as the
SINR results of Fig. 7 are obtained by Monte-Carlo method with
iterations over fifty different sky and noise realizations. However,
Fig. 11 is limited to the presented specific simulation with only one
sky and one noise realization.
The residual images of the un-clustered calibration as well as
the divisive hierarchical and weighted K-means clustered calibrations for Q = 33 are shown by Fig. 12 and Fig. 13, respectively.
As it is shown by Fig. 12, in the result of un-clustered calibration,
the sources are almost not subtracted at all and there is a significant residual error remaining. The residuals have asymmetric Gaussian distribution with variance σ 2 = 82.29 which is much larger
than the simulated (true) noise variance σ 2 = 10.85. On the other
hand, the sources have been perfectly subtracted in the case of clustered calibration, Fig. 13, and the residuals converge to the simulated background noise distribution. The residuals of the divisive
hierarchical and Weighted K-means clustered calibrations follow
symmetric zero mean Gaussian distributions with σ 2 = 20.17 and
σ 2 = 18.76, respectively. These variances are closer to the simulated one σ 2 = 10.85 and this indicates the promising performance
of clustered calibration. As we can see, hierarchical clustered calibration provides a slightly better result compared to the K-means
one. This is due to the fact that hierarchical clustering constructs
clusters of smaller angular diameters and thus it assigns the same
calibration solutions to the sources with smaller angular separations.
We also calculate the Root Mean Squared Error of Prediction
(RMSEP) to assess the performance of clustered and un-clustered
calibrations’ non-linear regressions. The results of log(RMSEP),
presented by Fig. 14, also justify that the best efficiencies of the hierarchical and K-means clustered calibrations are obtained at thirty
three number of clusters. But, note that there is a difference between the behavior of log(RMSEP) plot of Fig. 14 and the plots
of MCRLB, SINR, and Frobenius distance between the simulated
Jones matrices and solutions in Fig. 6, Fig. 7, and Fig. 11, respectively. In Fig. 14, log(RMSEP) of clustered calibration is less than
that of un-clustered calibration, even for extremely low number of
clusters. This means that even with those low number of clusters,
c 2007 RAS, MNRAS 000, 1–18
Clustered Radio Interferometric Calibration
11
Jy
1.4
1.2
1
0.8
0.6
0.4
0.2
0
−0.2
Figure 8. Single channel simulated observation of fifty sources, with intensities below 3 Jy. The source positions and their brightness are following uniform
and Rayleigh distributions, respectively. The image size is 8 by 8 degrees at 355 MHz. There are no gain errors and noise in the simulation.
Jy
0.6
0.5
0.4
0.3
0.2
0.1
0
−0.1
Figure 9. Simulated image with added gain errors. The errors, the complex 2 × 2 Jones matrices, are generated as linear combinations of sin and cos functions.
The variation of the sky is almost uniform on small angular scales.
c 2007 RAS, MNRAS 000, 1–18
12
Kazemi et al.
Jy
0.6
0.5
0.4
0.3
0.2
0.1
0
−0.1
−0.2
Figure 10. Simulated image of sky, corrupted by gain errors and by noise, as in Eq. (6). The simulated noise vector n has zero mean complex Gaussian
distribution and the SNR is equal to 2.
Jy
0.1
0.08
0.06
0.04
0.02
0
−0.02
−0.04
−0.06
−0.08
−0.1
Figure 12. Residual image of the un-clustered SAGE calibration for fifty sources. The sources are almost not subtracted at all and there are significant residual
errors around them. The residuals have asymmetric Gaussian distribution with variance σ2 = 82.29 which is much larger than the true noise variance
σ2 = 10.85.
c 2007 RAS, MNRAS 000, 1–18
Clustered Radio Interferometric Calibration
13
Jy
0.1
0.08
0.06
0.04
0.02
0
−0.02
−0.04
−0.06
−0.08
−0.1
||True Jones−Calibration solutions||
Figure 13. The residual images of the clustered calibration using hierarchical (right) and Weighted K-means (left) clustering methods with thirty three
source clusters. Calibration is implemented by SAGE algorithm. The sources are subtracted perfectly and the residuals converge to the simulated background
noise distribution. The hierarchical clustering and Weighted K-means residuals follow symmetric zero mean Gaussian distributions with σ2 = 20.17 and
σ2 = 18.76, respectively, where the simulated noise distribution is, CN (0, 10.85I).
we have
35
Weighted K−means
Divisive Hierarchical
un−clustered
30
y=
i=1
25
20
e
si (θe) +
K
X
Γ i + n,
(58)
i=1
After executing a calibration for the above data model, there is a
distance between the target parameters θe and the estimated solub
tions θe. This is the so-called “solver noise”, mentioned in section
15
5.2. Thus, the residuals are given as
10
5
0
0
Q
X
10
20
30
Number of Clusters
40
50
Figure 11. The average Frobenius distance between the simulated (true)
Jones matrices and the solutions of clustered and un-clustered calibrations.
The two curves represent clustered calibration via divisive hierarchical and
weighted K-means clustering algorithms. By increasing the number of clusters, for both clustering methods, this distance is decreased and gets its minimum approximately at thirty three clusters. After that, it is increased till the
fifty individual sources. That shows that the best performance of the clustered calibration is at around thirty three clusters.
clustered calibration performs better than the un-clustered calibration. This is somewhat in disagreement with the scenarios shown in
Fig. 6, Fig. 7, and Fig. 11. For a better understanding of the reason
behind this contrast, first lets see how the residual errors are originated.
Based on Eq. (13) and Eq. (40), in the clustered calibration strategy,
c 2007 RAS, MNRAS 000, 1–18
y−
Q
X
i=1
Q
K
i=1
i=1
X
X
b
b
e
Γ i + n.
{e
si (θe) − e
si (θe)} +
si (θe) =
(59)
From Eq. (59), we immediately see that the background noise n is
fixed
Pand the “clustering errors” are calculated for all the sources
as K
i=1 Γ i . However, since we solve only for Q directions and
not for all the K sources individually, the “solver noise” part,
PQ
b
{e
si (θe) − e
si (θe)}, is also calculated only for Q clusters and
i=1
not for all the K sources. It is clear that for a very small Q, this
term could be much less than for Q ≃ K. Therefore, in Fig. 11, the
result of RMSEP at a very low number of clusters is still less than
the ones at Q = K. When Q is not at the two extremes of being
very small or very large (almost equal to the number of individual
sources), then the result of RMSEP is promising. Moreover, applying more accurate calibration methods or increasing the number of
iterations, the “solver noise” will decrease and subsequently, we
expect the RMSEP curve to have the same behavior as the curves
of Fig. 7 and Fig. 11.
14
Kazemi et al.
6
−0.5 x 10
3
Weighted K−means
Divisive Hierarchical
un−clustered
Weighted K−means
Divisive Hierarchical
un−clustered
−1
1
−1.5
AIC
log(RMSEP)
2
0
−2
−1
−2.5
−2
−3
0
10
20
30
Number of clusters
40
−3
0
50
Figure 14. The RMSEP for clustered and un-clustered calibrations. The results are obtained using a base ten logarithmic scale. The two curves are corresponding to clustered calibration via divisive hierarchical and weighted
K-means clustering algorithms. By increasing the number of clusters, the
results are decreased and the minimum result is obtained at around thirty
three clusters. After that, the results are increased till the fifty individual
sources. That shows the superior performance of the clustered calibration
compared to the un-clustered one. The best performance of the clustered
calibration for both of the applied clustering methods is at around thirty
three number of clusters.
40
50
1800
We utilize AIC and LRT to select the optimum number of clusters for which clustered calibration achieves its best performance.
The methods are applied to our simulation for the case of SNR=2.
The AIC and LRT results are shown in Fig. 15 and Fig. 16, respectively. They both agree on Q = 33 as the optimum number of
clusters for the divisive hierarchical and Weighted K-means clustered calibrations. Likelihood-Ratio plot of Fig. 16 has almost the
same behavior as the plot of Frobenius distance between the simulated Jones matrices and the obtained solutions presented by Fig.
11. The reason is that the results of the both plots are obtained using
the solutions themselves as the input data. However, since AIC results are computed using the residual errors as inputs, which is also
the case for obtaining the RMSEP curves of Fig. 14, AIC curves of
Fig. 15 are slightly steeper than the Frobenius distance between the
simulated Jones matrices and solutions and the Likelihood-Ratio
curves of Fig. 11 and Fig. 16, respectively.
1400
Likelihood−Ratio
1600
We start changing the noise in our simulation to see how it effects the clustered calibration’s efficiency. We simulate the cases
for which SNR∈ {1, 2, . . . , 15} and apply clustered calibration on
them. Since the sky model does not change, the clusters obtained by
divisive hierarchical and weighted K-means methods for the case of
SNR= 2 remain the same. Fig. 17 shows the optimum number of
clusters, on which the best performances of clustered calibrations
are obtained for those different SNRs. As we can see in Fig. 17,
for low SNRs, the optimum Q is small. By increasing the SNR, the
optimum Q is increased till it becomes equal to the number of all
the individual sources that we have in the sky, i.e., K. This means
that when the SNR is very low, the benefit of improving signals
by clustering sources is much higher than the payoff of introducing
“clustering errors” in a clustered calibration procedure. Therefore,
clustered calibration for Q ≪ K has a superior performance com-
20
30
Number of clusters
Figure 15. AIC plot for clustered and un-clustered calibrations. Both the
weighted K-means and divisive hierarchical clustered calibrations get their
minimum AIC at about thirty three clusters. This illustrates that their best
performances are obtained at this number of clusters. Also, their AIC results at thirty three clusters is lower than the un-clustered calibration’s AIC,
which shows their better performances compared to the un-clustered calibration.
6.2 Optimum number of clusters for SNR=2
6.3 Clustered calibration’s efficiency at different SNRs
10
Weighted K−means
Divisive Hierarchical
un−clustered
1200
1000
800
600
400
200
0
10
20
30
Number of clusters
40
50
Figure 16. Likelihood-ratio of the gain solutions obtained by clustered and
un-clustered calibrations. In the both cases of weighted K-means and divisive hierarchical clustered calibrations, the minimum Likelihood-ratio values belong to approximately thirty three number of clusters. These minimums are also lower than the un-clustered calibration’s Likelihood-ratio
result. Therefore, clustered calibration via both the clustering methods performs better than the un-clustered calibration and it achieves the best accuracy in its solutions at thirty three clusters.
pared to the un-clustered calibration. While for a high enough SNR,
the situation becomes the opposite. In this case, the un-clustered
calibration achieves better results compared to clustered calibration
having the disadvantage of introducing “clustering errors”.
6.3.1 Empirical estimation of SINR
Having the results of Fig. 17 in hand, we could find an empirical
model for estimating the optimum number of clusters for various
SNRs. Note that by changing the observation and the instrument
characteristics, this model will also be changed.
As it is explained in section 4.1.3, the best performance of
clustered calibration is obtained when the SINR is at its highest
c 2007 RAS, MNRAS 000, 1–18
Divisive Hierarchical optimum Q
55
50
45
40
35
30
25
Weighted K−means
Divisive Hierarchical
un−clustered
20
15
0
5
SNR
10
Figure 17. The optimum number of clusters, on which the best performance
of clustered calibration is obtained, at different SNRs. For low SNRs, the
efficiency of the clustered calibration is superior to the un-clustered calibration. As the SNR gets higher, clustered calibrations achieve their best
solutions utilizing a higher number of clusters. Finally, when the SNR is
high enough, the performance of un-clustered calibration becomes better
than the clustered one.
level. Fig. 17 shows the number of clusters on which the maximum
SINR was obtained, where the signal (sky) and the noise powers are
known a priori. Thus, the only unknown for estimating the SINR
is the interference, or the “clustering errors”, for which we need
to have a prediction model. After estimating the SINR using this
model, finding the optimum Q will be straightforward.
Consider the definition of “clustering errors” given by Eq.
(31). It is logical that for every source, the difference between its
e is
true Jones matrix and the clustered calibration solution, ||J − J||,
a function of the angular distance between the source and the centroid of the cluster that it belongs to. Based on this and using Eq.
(31) and Eq. (34), for the interference of the i-th cluster at baseline
p − q we assume that
X
e i{pq} ||2 {D(Li )}ν I2 ),
Γl{pq} ∼ CN (0, η||C
(60)
l∈Li
where η and ν are unknowns. Eq. (60), in fact, considers an intere i{pq} ||2 {D(Li )}ν for every i-th
ference power (variance) of η||C
cluster, i ∈ {1, 2, . . . , Q}, at baseline p − q. Assuming the interferences of different clusters to be statistically independent from each
other, and bearing in mind that the baseline’s additive noise Npq
has also a complex Gaussian distribution independent from those
interferences’, then the noise plus interference power for the i-th
cluster at baseline p − q is obtained by
e i{pq} ||2 {D(Li )}ν + ||Npq ||2 .
η||C
(61)
Fitting suitable η and ν to Eq. (61), the SINR for the i-th cluster at
e i{pq} ||2 ,
baseline p − q is equal to the cluster’s signal power, ||C
divided by the result of Eq. (61). Subsequently, estimation of
E{SINRc } will be straight forward where the expectation is calculated with respect to all the source clusters and all the baselines. Note that simulation provides us with the true noise power,
||Npq ||2 . In the case of having a real observation, this power could
be estimated by Eq. (46).
Fig. 18 shows the number of clusters on which divisive hierarchical and weighted K-means clustered calibrations achieve their
maximum estimated E{SINRc }. The results are calculated for
SNR∈ {1, 2, . . . , 15}. For the hierarchical clustering η = 1550
c 2007 RAS, MNRAS 000, 1–18
15
50
40
30
20
True value
SINR estimation
10
0
15
Weighted K−means optimum Q
Optimum number of clusters (Q)
Clustered Radio Interferometric Calibration
5
SNR
10
15
50
40
30
20
True value
SINR estimation
10
0
5
SNR
10
15
Figure 18. Optimum number of clusters at which the divisive hierarchical (top) and Weighted K-means (bottom) clustered calibrations perform
the best. For both of the clustering methods, the results obtained by SINR
estimations mostly match the true optimum number of clusters.
and ν = 0.3, and for the K-means clustering η = 2500 and
ν = 0.003. As we can see, for both the clustering methods, these
maximum E{SINRc }s are mostly obtained at the true optimum
number of clusters for which clustered calibration performed the
best. Introducing more refined models compared to Eq. (60) could
even improve the current result.
6.4 Realistic sky models
So far, we have limited our studies to sky models in which the
brightness and position of the radio sources follow Rayleigh and
uniform distributions, respectively. These characteristics provide us
with a smooth and uniform variation of flux intensities in our simulated skies. In such a case, the effects of the background noise on
the faintest and the strongest signals are almost the same. Therefore, if clustered calibration performs better than the un-clustered
calibration that would be only based on upgrading the signals
against the noise. Although, in nature, we mostly deal with the sky
models in which the distribution of the flux intensities is a power
law, with a steep slope, and the spatial distribution is Poisson.
Hence, there exist a few number of very bright sources, whose signals are considerably stronger than the others, and they are sparse in
the field of view. The corruptions of the background noise plus the
interferences of the strong signals of those few bright sources make
Kazemi et al.
the calibration of the other faint point sources impractical. Thus,
there is the need for utilizing the clustered calibration which applies
the solutions of the bright sources to their closed by fainter ones or
solves for upgraded signals obtained by adding up a group of faint
signals together. This has been shown by Kazemi et al. (2011a);
Yatawatta et al. (submitted 2012), when comparing the efficiency
of the clustered and un-clustered calibrations on LOFAR real observations. In this section, using simulations, we also reveal the
superiority of clustered calibration compared to the un-clustered
calibration for such sky models.
We simulate a sky of 52 radio point sources which are obtained
by modified Jelić et al. (2008) foreground model. The brightness
distribution of the point sources follows the source count function
obtained at 151 MHz (Willott et al. 2001), while the angular clustering of the sources are characterized by a typical two-point correlation function,
ρ(d) = Ad−0.8 .
(62)
In (62), ρ is the two point correlation function, d is the angular separation, and A is the normalization amplitude of ρ. The flux cut off
is 0.1 Jy.
We corrupt the signals with gain errors which are linear combinations of sin and cos functions, as in our previous simulations. At
the end, a zero mean Gaussian thermal noise with a standard deviation of 3 mJy is added to the simulated data. The result is shown
in Fig. 19. In Fig. 19, all the bright sources are concentrated on the
right side of the image, rather than being uniformly distributed in
the field of view, and the rest of the sources are so faint that are
almost invisible.
We apply the clustered and un-clustered calibrations on Q ∈
{3, 4, . . . 51} number of clusters and K = 52 number of individual sources, respectively. The clustering method used is the divisive hierarchical and the calibrations are executed via SAGE algorithm with nine number of iterations. The residual noise variances obtained are demonstrated in Fig. 20. As Fig. 20 shows, the
level of the residual noise obtained by the clustered calibration for
Q ∈ {15, 16, . . . 45} number of clusters is always below the result
of the un-clustered calibration. This proves the better performance
of the clustered calibration. The best result of the clustered calibration, with the minimum noise level, is achieved for Q = 27 number
of clusters.
The residual images of the clustered calibration with Q = 27
number of source clusters, and the un-clustered calibration for
K = 52 individual sources are shown by Fig. 21. In the right side
of the residual image of the un-clustered calibration there exist artificial stripes caused by over and under estimating the brightest
sources of the field of view. That shows the problematic performance of the un-clustered calibration. However, clustered calibration has generated much less artificial effects after subtracting these
sources. On top of that, the zoomed in window in the left side of
the images of Fig. 21 show that the faint sources are not removed
by the un-clustered calibration at all, while being almost perfectly
subtracted by the clustered calibration. Moreover, the residual noise
of the clustered calibration follows a symmetric zero mean Gaussian distributions with a standard deviation of 4.2 mJy, while the
one from the un-clustered calibration has an asymmetric Gaussian
distribution with mean and standard deviation equal to -1.2 and 5.3
mJy, respectively. Taking to account that the simulated noise distribution is a zero mean Gaussian distribution with a variance of
3 mJy, the superior performance of the clustered calibration compared to the un-clustered one is evident.
As the final conclusion of this simulation, calibrating below
8
Divisive Hierarchical
Un−clustered
Noise variance
16
7
6
5
4
0
10
20
30
Number of clusters
40
50
Figure 20. The noise variances of the residual images, obtained by clustered and un-clustered calibrations, in mJy. The level of the residual noise
obtained by the clustered calibration for Q ∈ {15, 16, . . . 45} number of
clusters stands below the result of the un-clustered calibration. That reveals
the superior performance of the clustered calibration in comparison with the
un-clustered one. The best result of the clustered calibration which achieves
the minimum noise level is at Q = 27 number of clusters.
the noise level, clustered calibration always performs better than
the un-clustered calibration. This is regardless of the sky model and
is only based on the fact that solving for individual sources with
very poor signals is impractical. Nevertheless, when some sources
are very close to each other, the sky corruption on their signals
would be exactly the same and there is no point in solving for every
of them individually.
7 CONCLUSIONS
In this paper, we demonstrate the superior performance of “clustered calibration” compared to un-clustered calibration especially
in calibrating sources that are below the calibration noise level. The
superiority is in the sense of having more accurate results by the
enhancement of SNR as well as by the improvement of computational efficiency by reducing the number of directions along which
calibration has to be performed.
In a “clustered calibration” procedure, sky sources are
grouped into some clusters and every cluster is calibrated as a single source. That replaces the coherencies of individual sources by
the total coherency of the cluster. Clustered calibration is applied to
these new coherencies that carry a higher level of information compared with the individual ones. Thus, for the calibration of sources
below the noise level it has a considerably better performance compared to un-clustered calibration. Another way of looking at clustering is to consider the distribution of source flux densities, i.e.
the number of sources vs. the source flux density curve. Regardless
of the intrinsic flux density distribution, clustering makes the number of clusters vs. the cluster flux density curve more uniform, thus
yielding superior performance. An analytical proof of this superiority, for an arbitrary sky model, is presented using MCRLB and
SINR analysis.
KLD and LRT are utilized to detect the optimum number of
clusters, for which the clustered calibration accomplishes its best
performance. A model for estimating SINR of clustered calibration
is also presented by which we could find the optimum number of
clusters at low computational cost.
c 2007 RAS, MNRAS 000, 1–18
Clustered Radio Interferometric Calibration
17
Jy
0.8
0.6
0.4
0.2
0
−0.2
Figure 19. Simulated observation of fifty two sources which are obtained by modified Jelić et al. (2008) foreground model. The corrupting gain errors are
generated as linear combinations of sin and cos functions. The image size is 8 by 8 degrees and the additive thermal noise is a zero mean Gaussian noise with
a standard deviation of 3 mJy.
Divisive hierarchical as well as Weighted K-means clustering
methods are used to exploit the spatial proximity of the sources.
Simulation studies reveal clustered calibration’s improved performance at a low SNR, utilizing these clustering algorithms. Both the
clustering methods are hard clustering techniques which divide data
to distinct clusters. However, we expect more accurate results using
fuzzy (soft) clustering, which constructs overlapping clusters with
uncertain boundaries. Application and performance of this type of
clustering for clustered calibration will be explored in future work.
ACKNOWLEDGMENT
The first author would like to gratefully acknowledge NWO grant
436040 and to thank N. Monshizadeh for the useful discussions. We
also thank the referee for a careful review and valuable comments
that helped enhance this paper.
REFERENCES
Akaike H., 1973, in Second International Symposium on Information Theory (Tsahkadsor, 1971), Akadémiai Kiadó, Budapest,
pp. 267–281
Bernardi G., Mitchell D. A., Ord S. M., Greenhill L. J., Pindor B.,
Wayth R. B., Wyithe J. S. B., 2010, ArXiv e-prints
Bezdek J. C., 1981, Pattern Recognition with Fuzzy Objective
Function Algorithms. Kluwer Academic Publishers, Norwell,
MA, USA
Born M., Wolf E., 1999, Principles of Optics. Cambridge University Press
Bregman J. D., 2012, PhD Thesis, University of Groningen
c 2007 RAS, MNRAS 000, 1–18
Fessler J. A., Hero A. O., 1994, IEEE Transactions on Signal Processing, 42, 2664
Gini F., Reggiannini R., 2000, IEEE Transactions on Communications, 48, 2120
Gini F., Reggiannini R., Mengali U., 1998, IEEE, 46, 52
Graves S., 1978, British J. Philos. Sci., 29, 1
Hamaker J. P., 2006, A&A, 456, 395
Hamaker J. P., Bregman J. D., Sault R. J., 1996, A&AS, 117, 137
Intema H. T., van der Tol S., Cotton W. D., Cohen A. S., van
Bemmel I. M., Röttgering H. J. A., 2009, A&A, 501, 1185
Jelić V., Zaroubi S., Labropoulos P., Thomas R. M., Bernardi G.,
Brentjens M. A., de Bruyn A. G., Ciardi B., Harker G., Koopmans L. V. E., Pandey V. N., Schaye J., Yatawatta S., 2008, MNRAS, 389, 1319
Johnson S. C., 1967, Psychometrika, 32, 241
Kaufman L., Rousseeuw P., 1990, Finding Groups in Data An Introduction to Cluster Analysis. Wiley Interscience, New York
Kay S., 1993, Fundamentals of Statistical Signal Processing: Estimation theory, Prentice Hall signal processing series. PrenticeHall PTR
Kazemi S., Yatawatta S., Zaroubi, 2012, IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP),
2533
Kazemi S., Yatawatta S., Zaroubi S., 2011a, IEEE Workshop on
StatisticalSignal Processing, SSP2011, 414, 597
Kazemi S., Yatawatta S., Zaroubi S., de Bruyn A. G., Koopmans
L. V. E., Noordam J., 2011b, MNRAS, 414, 1656
Kerdprasop K., Kerdprasop N., Sattayatham P., 2005, in In
DaWaK, pp. 488–497
Liu A., Tegmark M., Zaldarriaga M., 2009, MNRAS, 394, 1575
MacQueen J. B., 1967, in Proc. of the fifth Berkeley Symposium
on Mathematical Statistics and Probability, Cam L. M. L., Ney-
18
Kazemi et al.
Jy
0.03
0.025
0.02
0.015
0.01
0.005
0
−0.005
−0.01
−0.015
Figure 21. The residual images of the clustered calibration for Q = 27 number of clusters (right) and the un-clustered calibration for K = 52 (left).
Calibration is implemented by SAGE algorithm with nine number of iterations and the clustering method applied is divisive hierarchical. In the right side of
the residual image of the un-clustered calibration we see stripes of over and under estimations due to problematic performance of the un-clustered calibration
in subtracting the brightest sources. The zoomed in window in the left side of the image also shows that the faint sources are not removed at all. However,
clustered calibration could remove all the faint sources almost perfectly and has generated much less artefacts after subtracting for the brightest sources in the
right side of the field of view. Moreover, the residual noise of the clustered calibration follows a symmetric zero mean Gaussian distributions with a standard
deviation of 4.2 mJy, while the one from the un-clustered calibration has an asymmetric Gaussian distribution with mean and standard deviation equal to -1.2
and 5.3 mJy, respectively. Taking to account that the simulated noise distribution is a zero mean Gaussian distribution with a standard deviation of 3 mJy, the
better performance of the clustered calibration compared to the un-clustered one is evident.
man J., eds., Vol. 1, University of California Press, pp. 281–297
Pearson T. J., Readhead A. C. S., 1984, ARA&A, 22, 97
Pindor B., Wyithe J. S. B., Mitchell D. A., Ord S. M., Wayth R. B.,
Greenhill L. J., 2010, PASA, 28, 46
Smirnov O. M., 2011, A&A, 527, A108
Thompson A. R., Moran J. M., Swenson G. W., 2001, Interferometry and Synthesis in Radio Astronomy, 2nd edn. Wiley-VCH
Wax M., Kailath T., 1985, Acoustics, Speech and Signal Processing, IEEE Transactions on, 33, 387
Willott C. J., Rawlings S., Blundell K. M., Lacy M., Eales S. A.,
2001, MNRAS, 322, 536
Xu R., Wunsch D., 2008, Clustering. John Wiley and Sons
Yatawatta S., de Bruyn A. G., LOFAR Collaboration, submitted
2012, Astronomy and Astrophysics
Yatawatta S., Zaroubi S., de Bruyn G., Koopmans L., Noordam J.,
2009, Digital Signal Processing Workshop and 5th IEEE Signal
Processing Education Workshop, 2009. DSP/SPE 2009. IEEE
13th, 150
This paper has been typeset from a TEX/ LATEX file prepared by the
author.
c 2007 RAS, MNRAS 000, 1–18
| 5cs.CE
|
Checks and Balances: A Low-complexity High-gain
Uplink Power Controller for CoMP
arXiv:1610.08491v1 [cs.SY] 26 Oct 2016
Fangzhou Chen†§ , Yin Sun†§ , Yiping Qin‡ , and C. Emre Koksal†
† Dept. of ECE, The Ohio State University, Columbus, OH
‡ Huawei Technologies Co., Shanghai, China
§ Co-primary authors
Abstract—Coordinated Multipoint (CoMP) promised substantial throughput gain for next-generation cellular systems. However,
realizing this gain is costly in terms of pilots and backhaul
bandwidth, and may require substantial modifications in physicallayer hardware. Targeting efficient throughput gain, we develop
a novel coordinated power control scheme for uplink cellular
networks called Checks and Balances (C&B), which checks the
received signal strength of one user and its generated interference
to neighboring base stations, and balances the two. C&B has some
highly attractive advantages: C&B (i) can be implemented easily
in software, (ii) does not require to upgrade non-CoMP physicallayer hardware, (iii) allows for fully distributed implementation
for each user equipment (UE), and (iv) does not need extra pilots
or backhaul communications. We evaluate the throughput performance of C&B on an uplink LTE system-level simulation platform,
which is carefully calibrated with Huawei. Our simulation results
show that C&B achieves much better throughput performance,
compared to several widely-used power control schemes.
Index Terms—Uplink power control, coordinated multipoint
(CoMP), LTE, system-level simulation, throughput improvement
I. I NTRODUCTION
Next generation of cellular communication, e.g., Long Term
Evolution Advanced (LTE-A) and 5G, is expected to significantly improve the average throughput and cell-edge throughput
for serving user equipments (UEs). One important candidate
technique for achieving such throughput improvement is Coordinated Multipoint (CoMP), which refers to the cooperation
among different Base Stations (BSs).
The promised benefits of CoMP are hard to realize because
of many issues in practical systems [1], [2]. In particular,
the uplink CoMP techniques (such as distributed interference
cancellation/alignment, joint detection, coordinated scheduling)
all require nearby BSs to communicate received signals, control
messages, and channel state information through backhaul
links. In addition, these CoMP techniques are costly in terms
of power and pilot resources, which considerably decreases
the resources allocated for data transmissions. Therefore, the
realized throughput performance is greatly degraded.
Aiming to realize the potential benefits of CoMP in practical
systems, we propose a novel coordinated power control design
for uplink cellular networks. The task of uplink power control
is to make the signal received at the base station sufficiently
This work was supported in part by Huawei, Inc. under Agreement YB
2013110091, National Science Foundation under Grants CNS-1054738 and
CNS-1514260.
TABLE I: Performance of FPC [5], Max Power [6], RLPC [7]
and C&B in Macrocell system-level simulations.
Average Throughput (Mbits/s)
5%-Edge Throughput (Mbits/s)
Power Efficiency (Mbits/J)
FPC
8.05
0.23
751
Max Power
12.01
0.09
6.77
RLPC
9.78
0.22
226
C&B
12.23
0.23
387
strong, and in the meanwhile keep the interference generated
to nearby base stations not severe. In practice, overly high
and low transmission powers are both harmful. Specifically,
increasing the transmission power of one UE can increase its
throughput, but it causes some strong interference to nearby
cells, which will degrade the throughput of other UEs. Hence,
finding the correct balance between a UE’s own performance
and its incurred cost to the other UEs is crucial to achieve a
satisfying performance.
In practical uplink cellular networks, each BS receiver experiences the interference from hundreds of UEs from neighboring cells. Even if perfect CSI knowledge of the signal and
interference channels are available, the optimal power control
problem is strongly NP hard [8], [9]. To make things worse,
the base station in current systems typically estimates the signal
channel of its served UEs, but the channel coefficients of interfering UEs are mostly unavailable. These practical limitations
make the power control problem even more challenging.
In our research, we develop a low-complexity Coordinated
power control scheme that provides significant throughput
gains, with minimum cost and modifications. To that end, the
following are the contributions of this paper:
• We develop a novel coordinated power control scheme,
named Checks and Balances (C&B). C&B requires very
little information, including the large-scale path loss from
one UE to several nearby BSs, coarse power level of cochannel interference, and the throughput vs SINR curve
of Adaptive Modulation and Coding (AMC). Based on
this information, C&B checks the SNR of one UE and its
generated INR to nearby BSs, and balances the two.
• C&B has some highly attractive advantages: C&B (i) can
be implemented easily in software, (ii) does not require to
upgrade non-CoMP physical-layer hardware, (iii) allows
for fully distributed implementation for each UE, and (iv)
does not need extra pilots or backhaul communications.
• We evaluate the throughput performance of C&B on a
system-level simulation platform for LTE uplink, which
is carefully calibrated with Huawei. As shown in Table
I, C&B increases the average throughput by 51.9% over
Fractional Power Control (FPC) [5] and 21.5% over
Reverse Link Power Control (RLPC) [7], and achieves
similar cell-edge throughput with FPC and RLPC. Compared to Max Power Control [6], C&B increases the
average throughput and cell-edge throughput by 1.8% and
156%, respectively, together with greatly improved power
efficiency.
We expect C&B to achieve even better throughput performance when working with physical-layer CoMP techniques,
which will be considered in future work.
Related Studies: Non-coordinated power control is standardized in 3GPP protocols [3] and has attracted vast research interests [5], [6], [10]–[15]. There exist three mainstream schemes:
1) Full Compensation Power Control (FCPC) [10] allocates
transmission power to one UE by making full compensation
of its large-scale path loss such that all UEs have the same
received signal strength, which results in poor performance
in per-cell average throughput and inter-cell interference management. 2) Max Power scheme [6] let all UEs transmit at
their maximum allowable power. It provides high average percell throughput, but performs poorly in power efficiency and
throughput of UEs at the cell edge. 3) Fractional Power Control
(FPC) [5] is currently the most widely adopted scheme [12]–
[15], which allocates transmission power by making fractional
compensation of UEs’ large-scale path losses, such that UEs in
the interior of one cell have stronger received signal strength
than UEs at the cell edge. The key drawback remains in the
unsatisfying average throughput per cell.
Two coordinated power control schemes have been proposed
[7], [16], which make additional compensation for large-scale
path losses from one UE to its neighboring cells. These schemes
are essentially variations of FCPC. Hence, they partially inherit
the drawbacks of FCPC which significantly limits the throughput gain.
II. S YSTEM M ODEL AND P ROBLEM D ESCRIPTION
We consider an LTE uplink multicellular network. In such
network, each UE transmits to its serving cell and meanwhile
generates interference to its neighboring cells. Consider UE u
transmits signal x at power P on a single subcarrier, its received
signal y at the serving cell c can be expressed as:
Xp
√
y = P · hx +
Pj · hj xj + n,
(1)
j
where h and hj denotes the instantaneous complex channel
gain from UE u and uj (served in neighboring cell cj ) to cell
c, respectively, n denotes the experienced noise.P
Hence
p the total
inter-cell interference experienced by UE u is j Pj · hj xj .
Assuming the transmitted signals x and xj have unit variance,
the signal-to-interference-plus-noise ratio (SINR) can thereby
be calculated as:
sinr = P
P · |h|2
,
2
2
j Pj · |hj | + σn
(2)
where σn2 denotes the variance of noise.
The power controller decides the transmission power of UEs
across all cells, which heavily affects their SINR. Unlike Noncoordinated power control that only exploits the CSI of UEs
to their serving cells, coordinated power control additionally
utilizes the CSI of UEs to multiple neighboring cells. The
problem we are tackling is to come up with a coordinated power
control design with low complexity and high throughput gain
over all existing solutions.
Besides the strong impact from power controller, the
throughput performance is also influenced by how each cell
picks the modulation and coding scheme (MCS) for active
UEs. In LTE networks, the existing Turbo-coded modulation
techniques are paired with associated bit rate selections to
form 29 different available MCS options [3]. For each SINR
value, one of these MCS options is chosen by an Adaptive
Modulation and Coding (AMC) module. The selected MCS
should provide a sufficient high throughput and meanwhile
guarantee a low decoding error probability. Usually, the block
decoding error rate is required to be less than 10%, which
will be later compensated by hybrid automatic repeat request
(HARQ). The stairs in the throughput curve are due to the AMC
module. In particular, when the SINR is lower than −6.5 dB,
no MCS can decode successfully and hence the throughput is
zero. When the SINR is higher than 18 dB, the maximum MCS
can decode perfectly, achieving a maximum throughput. Hence,
the SINR region for effective AMC selection is [−6.5 dB, 18
dB].
III. C HECKS
AND
BALANCES : A P OWER C ONTROL D ESIGN
In this section, we present a novel power controller design,
called Checks and balances (C&B), for uplink cellular systems.
This power controller requires very little information, including
the throughput versus SINR curve of the receiver design, the
large-scale path loss from one UE to several nearby BSs,
and some coarse distribution information of the co-channel
interference in the cellular system. The key idea in C&B is
to cooperatively balance the SNR of a UE and its generated
INR to nearby BSs. The complexity of C&B is very low, and
the throughput gain is huge. One can consider C&B as the
simplest implementation of CoMP, which provides significant
throughput improvement without incurring huge cost in pilots
or backhaul. There is no upgrade of the physical layer design,
except for the change of uplink transmission power.
A. Approximations of SINR and Throughput
C&B can operate in the open-loop mode, when only large
scale CSI is utilized, whereas small scale CSI is unavailable
due to the lack of instantaneous channel estimation pilot. The
large-scale path loss between UE uj and cell c is defined as:
−1
,
(3)
PLj , E |hj |2
1
SINR-Throughput Mapping
Piece-wise Approximation f(SINR)
4
0.8
3
0.6
SNR
IoT
cdf
Spectrum Efficiency (bits/s/Hz)
5
2
0.4
1
0.2
0
-20
-10
0
SINR (dB)
10
0
-10
20
0
10
Time-average (dB)
20
Fig. 1: Throughput Curve Approximation.
Fig. 2: Time-average SNR and IoT distributions.
where E[·] denotes the expectation. When we only consider
such large-scale path losses, u’s received signal-to-noise ratio
(SNR) and interference over thermal noise (IoT) at c, and generated interference-to-noise ratio (INR) to cj can be respectively
approximated as:
Hence, we propose a method to estimate IoTS . In the simulation
environment described in Section IV, the distribution of the
IoT in the uplink cellular system is illustrated in Fig. 2. We
select IoTS to be the 95th percentile of IoT that experienced
by an arbitrary UE, which is IoTS = 9 dB. Note that this
statistical distribution can be collected at the BS. In practical
applications, we recommend to initialize the system with this
value and update it according to the measured IoT distribution.
The approximated throughput RS (P ) is plotted in Fig. 3a,
which increases with the transmission power P .
SNR(P )
=
IoT =
INRj (P )
=
PL−1 · P
,
N0
P
N0 + j Pj · PL−1
j
PL−1
u→j
N0
N0
·P
,
(4)
,
(5)
(6)
where N0 denotes the average noise power and is assumed to
be the same and known for all UEs. We also define PLu→j as
the large-scale path loss from UE u to cell cj . Derived from
Eq. (4) and (5), we can approximate the received SINR of UE
u as:
SNR(P )
.
(7)
SINR =
IoT
As C&B only acquires large scale CSI and approximated
SINR information, there is no need to use an accurate throughput curve to determine the transmission power. Therefore, we
introduce a piece-wise function to approximate the foregoing
throughput curve:
f (SINR) = min Tmax , a · log2 (1 + b · SINR) ,
(8)
where Tmax = 4.18, a = 0.7035 and b = 0.7041. These
parameters are achieved by curve fitting as shown in Fig. 1.
Note that for different physical layer designs, we can always
find such an approximated throughput function f (SINR). Next,
we use function f (SINR) to evaluate the throughput of one UE
u and the throughput of the UEs interfered by UE u.
B. Checking the Influence of SNR and INR
Consider an arbitrarily chosen UE u, its uplink throughput
can be approximated as
SNR(P )
RS (P ) = f
,
(9)
IoTS
where IoTS denotes a prior estimation of the interference
experienced by UE u. In fully distributed power control, the
instantaneous interference power that UE u will experience
during its transmissions is not available at the power controller.
Consider the UEs that are notably interfered by UE u, their
sum uplink throughput can be approximated as:
X
SNRI
.
(10)
f
RI (P ) =
IoTI + INRj (P )
PL
<PL
u→j
th
Note that we only consider the non-negligible interference
generated by UE u. To this end, we set the threshold PLth
in (10) as the path loss of a interference link such that the
interference power generated by UE u with the maximum
power Pmax is stronger than noise, i.e.,
PL−1
th · Pmax
> 1,
N0
(11)
where Pmax = 200 mW (23 dBm) [18]. Furthermore, in fully
distributed power control, the received SNR and IoT (excluding
the interference generated by u) of these UEs, denoted by SNRI
and IoTI , are unknown to the power controller of UE u. We
also propose methods to estimate SNRI and IoTI by using
the SINR region [−6.5 dB, 18 dB] in Fig. 1 for effective
AMC selection. Their values are given by SNRI = 24 dB
and IoTI = 5 dB. Due to space limitations, the details for
computing these parameters will be explained in the journal
version. The approximated sum throughput RI (P ) is plotted
in Fig. 3b, which decreases with the transmission power P .
C. Balancing the Two
In order to find a proper balance between SNR and INR, we
consider the following weighted sum throughput maximization
problem:
max
0≤P ≤Pmax
RS (P ) + ζ · RI (P ),
(12)
27
4
26
28
27.5
Weighted Sum Utility
5
3
UI(P)
US(P)
25
2
24
0
-5
22
0
5
10
15
P (dBm)
20
25
(a) Approximated throughput of UE u.
21
-5
26
25.5
23
1
27
26.5
25
24.5
0
5
10
15
P (dBm)
20
25
24
-5
0
5
10
15
P (dBm)
20
25
(b) Approximated total throughput of UEs that (c) Weighted sum throughput approximation.
are interfered by UE u.
Fig. 3: Approximated throughput functions for C&B power controller design.
where ζ is the weight parameter that adjusts the relative
importance of SNR and INR, which is chosen around 1. Large ζ
(e.g. ζ > 1) indicates more focus on mitigating INR rather than
enhancing SNR, which leads to more conservative transmission
power. This benefits the UEs with poor channel states due to
the reduced interference, while constraining the transmission
power of UEs with good channel states that prevents them from
achieving better throughput. It works the other way around
when we choose small ζ (e.g. ζ < 1). Tending to suppress
strong interference, we initially select ζ to be 1.3. As shown in
Fig. 3c, the weighted sum throughput is maximized at a unique
transmission power, i.e., the balance we choose between SNR
and INR.
IV. P ERFORMANCE E VALUATION ON S YSTEM - LEVEL
S IMULATION P LATFORM
A. Simulation Platform Configuration
(13)
We consider an LTE uplink cellular network [19], where the
BSs are located on a typical hexagonal lattice including 19
BS sites, 3 sectors per site. Each sector is regarded as a cell.
The minimal distance between two neighboring sites is 500
meters (Macrocell) [19]. The UEs are uniformly distributed in
the entire network area. The distance from a UE to a nearest
BS is no smaller than 35 meters [18]. Each UE has a single
antenna, and the BSs are equipped with 2 antennas per sector.
The wireless channel coefficients and BS antenna pattern are
generated by following the SCM model for Urban Macro environments in 3GPP TR 25.996 [18], where 3D antenna pattern
and Rayleigh fading [23] are adopted. We set the maximum
doppler shift frequency at 7 Hz according to moving speed 3
km/h and carrier frequency 2.5 GHz. The receivers employ nonCoMP maximum mean square error (MMSE) estimation [25]
with interference rejection combination (IRC) techniques [24].
To obtain the channel coefficients, we estimate the pilots, i.e.,
demodulation reference signals (DMRS) [3], with DFT-based
estimation [22]. Adaptive Transmission Bandwidth (ATB) [21]
non-CoMP packet scheduling scheme is implemented. The
frequency bandwidth is 10MHz, and the noise figure of each BS
receiver is 5 dB [3]. We set a uniform penetration loss of 20 dB
[18] for all users. The delay of control signaling is uniformly
set as 6 ms (i.e., 6 time slots) [17]. The wrap around technique
is employed to avoid the border effect. The parameters of our
system level simulation platform are listed in Table II.
We use the proportional-fair (PF) policy [25] in stochastic
network control in which the weight of UE u used in the ATB
scheduler at time-slot t is
(ru [t])α
,
(14)
(r̄u )β
We use bisection method to find the solution of (13), with
the detailed steps presented in Algorithm 1. This algorithm is
easy to implement in software, and the number of iterations
for convergence is no more than 10. We note that Algorithm 1
is a fully distributed algorithm, where each UE can choose its
transmission power independently.
where ru [t] and r̄u denote the UE u’s potentially achieved data
rate in time slot t and long-term average data rate, respectively.
We set the two associated parameters at: α = 1, β = 1.
We compare C&B with three reference policies: One policy
is the widely used fractional power control (FPC) scheme,
which determines the transmission power P F P C of UE u for
Algorithm 1: Bisection method for solving (13).
1
2
3
4
5
6
7
8
9
10
11
12
13
14
Given l = −10 dBm, r = Pmax , tolerance ǫ = 0.1;
if r − l < ǫ then
m := Pmax ;
else
while r − l ≥ ǫ do
m := (l + r)/2;
if RS′ (m) + ζ · RI′ (m) < 0 then
l := m;
else
r := m;
end
end
end
return P := m.
It is easy to show that Problem (12) is a one-dimensional
quasi-convex optimization problem, and thereby can be obtained by solving the following problem:
RS′ (P ) + ζ · RI′ (P ) = 0.
TABLE II: SYSTEM LEVEL SIMULATION PARAMETERS
Parameter
Deployment Scenario
Inter-site Distance
System Bandwidth
Avg. UEs per Cell
UE/BS Antennas
Distance-dependent Path Loss
Shadowing Standard Deviation
Antenna Pattern
Penetration Loss
Scheduling Decision Delay
Target BLER
Traffic Model
Scheduling Algorithm
Power Control (PC) scheme
Stochastic Network Control Scheme
Link Adaptation
BS Receiver Type
Channel Estimation
Maximum Doppler Shift
α (PF)
β (PF)
Pmax
P0F P C (FPC)
κ (FPC)
P0RL (RLPC)
φ (RLPC)
Setting
19 BS sites, 3 sectors (cells) per site,
wrap-around
500m (Macrocell)
10MHz [50 PRBs, 2 used for control]
10
1/2 per cell
According to 3GPP 36.814 [19]
8 dB
3D
20 dB
6 ms
10%
Fully backlogged queues
ATB
FPC, Max Power, RLPC, C&B
Proportional Fair (PF)
AMC, based on 3GPP TS 36.213 [3]
IRC MMSE
DFT-based Estimation
7Hz
1
1
23 dBm
−87 dBm
0.8
−102 dBm
0.8
each resource block by [3]:
P F P C = min(Pmax , P0F P C + κ · PL),
(15)
P0F P C
where Pmax is the maximum power constraint,
is the
default transmission power, and PL is the large-scale path loss
from UE u to its serving cell. The values of the parameters are
Pmax = 23 dBm, P0F P C = −87 dBm and κ = 0.8. The second
policy is Max Power, which sets all UEs at their maximum
allowable transmission power Pmax . The third reference policy
is the coordinated Reverse Link Power Control (RLPC) scheme,
where the transmission power P RL is decided by [7]:
P RL = min Pmax , P0RL + φ · PL + (1 − φ) · PLmin , (16)
where PLmin denotes the measured minimum path loss from
u to its neighboring cells. The rest parameters are selected as
P0RL = −102 dBm and φ = 0.8. The parameters of C&B are
chosen as we discussed in Section III.
B. Simulation Results
We compare the performance of different power control
schemes in terms of three key metrics: cell average throughput
(i.e. sum average throughput per cell), cell edge throughput and
power efficiency. In particular, cell-edge throughput is defined
as the 5th percentile throughput performance among all UEs,
denoted as 5%-Edge, which is widely used in evaluating the
performance of UE fairness [6], [11], [15].
1) Performance Comparison in Macrocell Scenario: First,
we investigate the performance in enhancing received signal
strength, which is evaluated by time-average SNR as illustrated
in Fig. 4a. It can be seen that C&B, with the weight parameter
set as ζ = 1.3, is able to significantly boost UEs’ timeaverage SNR over FPC. Particularly, more than 40% UEs
have gained at least 3.5 dB and approximately 20% UEs even
achieve more than 10 dB SNR increase. C&B also improves
the SNR of the top 20% UEs compared to RLPC, which
greatly benefit the UEs with good channel conditions. It is
clear that Max Power has the best SNR performance due to the
maximum transmission power. Nevertheless, it simultaneously
incurs severe interference which is quite undesirable.
Next, we switch our attention on the interference mitigation
performance, which is represented by the time-average IoT as
shown in Fig. 4b. Compared with FPC, C&B only slightly
increase the UEs’ average IoT by 1.5 dB. On the other hand,
C&B performs a lot better than Max Power and shows clear
advantage over RLPC in suppressing the inter-cell interference.
Now we concentrate on the throughput performance, which
is the comprehensive result of signal enhancement and interference mitigation. In comparison with FPC and RLPC, C&B
noticeably improves the throughput of UEs with good channel
condition (the top 30%), as illustrated in Fig. 4c. In particular,
the top 20% UEs have even achieved throughput gain of at least
1 Mbits/s. As for UEs with poor channel condition (bottom
10%), C&B shows great advantage over Max Power.
Finally, we present C&B’s advantages over all other candidate power control schemes based on the detailed performance
summary in Table I. C&B shows desirable advantage over FPC
by providing 51.9% improvement on cell average throughput,
while keeping the same 5%-Edge throughput performance and
only dropping the power efficiency by 48%. Further, C&B beats
Max Power in significantly improving 5%-Edge throughput
and power efficiency by 156% and 5, 716% respectively, with
slightly increased cell average throughput. Comparison with a
CoMP power control scheme called RLPC, C&B achieves appreciable gains in cell average throughput and power efficiency
respectively by 25.1% and 71.2%, and simultaneous keeps 5%Edge throughput slightly improved.
Note that the performance improvement of C&B is solely
obtained by power control. One can further incorporate coordinated transceiver and scheduling techniques to achieve even
higher gain. For example, it is known that CoMP receiving techniques can enhance the edge throughput significantly [2]. As
shown in Table I, the throughput gain of C&B is more evident
in the high throughput UEs. Hence, additional improvement is
promising by combining C&B with other CoMP techniques.
2) Tradeoff between cell average throughput and cell edge
throughput: Recall the utility maximization problem (12), we
can vary the weight parameter ζ to adjust the balance between
received signal enhancement and interference mitigation. As
we gradually decrease ζ from 1.3 to 0.7, the cell average
throughput keeps improving in pair with the continuously
degrading 5%-Edge throughput, as summarized in Table III.
This tradeoff between average and edge throughput can be
explained as follows: (a) the UEs with good channel conditions
contribute most part of the average throughput. Increasing
such UEs’ transmission power by decreasing ζ greatly helps
them to achieve stronger received signal, which results in their
throughput gains. Such gains lead the cell average throughput
1
0.8
0.8
0.8
0.6
0.6
0.6
0.4
cdf
cdf
1
cdf
1
0.4
FPC
Max Power
RLPC
C&B
0.2
0
-20
0
20
SNR (dB)
0.4
FPC
Max Power
RLPC
C&B
0.2
40
60
(a) Time-average received SNR.
0
0
5
10
UE IoT (dB)
15
FPC
Max Power
RLPC
C&B
0.2
20
0
0
(b) Time-average experienced IoT.
1
2
3
4
Average Throughput (Mbits/s)
5
(c) Time-average throughput.
Fig. 4: Simulation results in Macrocell scenario.
to improve. (b) In contrast, boosting transmission power incurs
stronger inter-cell interference, which especially jeopardizes
those vulnerable UEs with poor channel states. As a result,
the 5%-Edge throughput is decreased.
TABLE III: Throughput performance comparison between different weight factor ζ selections.
Average Throughput (Mbits/s)
5%-Edge Throughput (Mbits/s)
ζ = 1.3
12.23
0.23
ζ = 1.1
12.41
0.21
ζ = 0.9
12.95
0.18
ζ = 0.7
13.17
0.15
Note that when ζ = 0.7, C&B is remarkably better than
the Max Power policy in terms of both average and edgethroughput. In addition, when ζ = 1.3, C&B is significantly
better than the FPC and RLPC in term of cell average throughput. Therefore, one can adjust the weight parameter in order to
adapt different system requirements.
V. C ONCLUSION
AND
F UTURE W ORK
We investigate how to use limited large scale CSI to achieve
the throughput gain of CoMP through the design of a power
controller named C&B. The optimal power control by itself
is an NP hard problem, which has been open up to date.
Further, as very limited coordination is possible in the openloop mode, the power controllers of different BSs are not
allowed to communicate. C&B satisfies all practical constraints
in cellular systems, and can significantly improve the average
and edge throughput over existing power control schemes with
very low complexity and almost no cost. Further throughput
enhancement is promising by combining C&B with other
coordinated transceiver and scheduling techniques.
R EFERENCES
[1] A. Lozano, R. W. Heath, and J. G. Andrews, “ Fundamental limits of
cooperation, ” IEEE Trans. Inform. Theory, vol. 59, no. 9, pp. 5213-5226,
Sep. 2013.
[2] M. H. C. Suh and D. N. C. Tse, “ Downlink interference alignment, ”
IEEE Trans. Commun., vol. 59, no. 9, pp. 2616-2626, Sep. 2011.
[3] 3GPP TS 36.213, “ E-UTRA Physical Layer Procedures.”
[4] A. Barbieri, P. Gaal, S. Geirhofer, T. Ji, D. Malladi, Y. Wei, and F. Xue,
“ Coordinate downlink multi-point communications in heterogeneous 4g
cellular networks, ” in Proc. Inform. Theory and Appl. Workshop, San
Diego, CA, Feb. 2012.
[5] J. F. Whitehead, “ Signal-Level-Based Dynamic Power Control for Cochannel Interference Management, ” 43rd IEEE Vehicular Technology
Conference, pp. 499-502, May 1993.
[6] A. Simonsson and A. Furuskar, “ Uplink Power Control in LTE Overview and Performance, Subtitle: Principles and Benefits of Utilizing
rather than Compensating for SINR Variations, ” Vehicular Technology
Conference, 2008.
[7] A. M. Rao, “ Reverse Link Power Control for Managing Inter-cell Interference in Orthogonal Multiple Access Systems, ” Vehicular Technology
Conference, 2007.
[8] D. Shah, D. N. C. Tse, and J. N. Tsitsiklis, “ Hardness of low delay
network scheduling, ” IEEE Trans. Inf. Theory, vol. 57, no. 12, pp. 78107817, Dec. 2011.
[9] Z.-Q. Luo and S. Zhang, “ Dynamic spectrum management: Complexity
and duality, ” IEEE Journal of Selected Topics in Signal Processing, vol.
2, no. 1, pp. 57-73, Feb 2008.
[10] J.G.Gipson, “ The Communication Handbook, ” CRC Press, IEEE Press
1996.
[11] C. U. Castellanos, D. L. Villa, C. Rosa, K. I. Pedersen, F. D. Calabrese,
P. Michaelsen, and J. Michel, “ Performance of Uplink Fractional Power
Control in UTRAN LTE, ” Vehicular Technology Conference, 2008.
[12] T. D. Novlan, H. S. Dhillon and J. G. Andrews, “ Analytical Modeling
of Uplink Cellular Networks, ” IEEE Transactions on Wireless Communications, vol. 12, no. 6, Jun 2013.
[13] T. D. Novlan and J. G. Andrews, “ Analytical Evaluation of Uplink
Fractional Frequency Reuse, ” IEEE Transactions on Communications,
vol. 61, no. 5, May 2013.
[14] W. Xiao, R. Ratasuk, A. Ghosh, R. Love, Y. Sun, and R. Nory, “ Uplink
Power Control, Interference Coordination and Resource Allocation for
3GPP E-UTRA, ” Vehicular Technology Conference, 2006.
[15] M. Coupechoux and J. -M. Kelif, “ How to Set the Fractional Power
Control Compensation Factor in LTE, ” IEEE Sarnoff Symposium, 2011.
[16] R. Yang, A. Papathanassiou, W. Guan, A. T. Koc, H. Yin, Y. Chio, “
Supporting Material for UL OLPC Proposal, ” IEEE C80216m-09-0844,
2009.
[17] 3GPP TS 36.321, “ E-UTRA Medium Access Control (MAC) Protocol
Specification. ”
[18] 3GPP TR 25.996 v9.0.0., “ Channel Model for Multiple Input Multiple
Output (MIMO) Simulations, ” 2009.
[19] 3GPP TR 36.814, “ E-UTRA Further Advancements for E-UTRA Physical Layer Aspects. ”
[20] 3GPP TS 36.133, “ E-UTRA Requirements for Support of Radio Resource Management. ”
[21] F. Calabrese, C. Rosa, M. Anas, P. Michaelsen, K. Pedersen, and P.
Mogensen, “ Adaptive transmission bandwidth based packet scheduling
for lte uplink, ” in VTC 2008-Fall, Sept 2008, pp. 1-5.
[22] G. Huang, A. Nix, and S. Armour, “ DFT-Based Channel Estimation and
Noise Variance Estimation Techniques for Single-Carrier FDMA, ”in VTC
2010-Fall, 2010.
[23] Y. R. Zheng and C. Xiao, “ Simulation Models with Correct Statistical Properties for Rayleigh Fading Channels, ” IEEE Transactions on
Communications, vol. 51, issue 6, pp. 920-928, 2003.
[24] Y. Leost, M. Abdi, R. Richter and M. Jeschke, “ Interference Rejection
Combining in LTE Networks, ” Bell Labs Technical Journal, pp. 25-50,
2012.
[25] D. Tse and P. Viswanath, “ Fundamentals of Wireless Communication, ”
Cambridge University Press, 2005.
| 3cs.SY
|
1
Superimposed Pilots are Superior for Mitigating
Pilot Contamination in Massive MIMO
arXiv:1603.00648v3 [cs.IT] 2 Aug 2017
Karthik Upadhya, Student Member, IEEE, Sergiy A. Vorobyov, Senior Member, IEEE,
Mikko Vehkapera, Member, IEEE
Abstract—In this paper, superimposed pilots are introduced as
an alternative to time-multiplexed pilot and data symbols for mitigating pilot contamination in massive multiple-input multipleoutput (MIMO) systems. We propose a non-iterative scheme for
uplink channel estimation based on superimposed pilots and
derive an expression for the uplink signal-to-interference-plusnoise ratio (SINR) at the output of a matched filter employing
this channel estimate. Based on this expression, we observe
that power control is essential when superimposed pilots are
employed. Moreover, the quality of the channel estimate can
be improved by reducing the interference that results from
transmitting data alongside the pilots, and an intuitive iterative
data-aided scheme that reduces this component of interference is
also proposed. Approximate expressions for the uplink SINR are
provided for the iterative data-aided method as well. In addition,
we show that a hybrid system with users utilizing both timemultiplexed and superimposed pilots is superior to an optimally
designed system that employs only time-multiplexed pilots, even
when the non-iterative channel estimate is used to build the
detector and precoder. We also describe a simple approach to
implement this hybrid system by minimizing the overall inter
and intra-cell interference. Numerical simulations demonstrating
the performance of the proposed channel estimation schemes and
the superiority of the hybrid system are also provided.
Index Terms—Massive MIMO, pilot contamination, superimposed pilots.
I. I NTRODUCTION
Massive multiple-input multiple-output (MIMO) systems,
proposed in [1], have received significant interest in recent
years as a candidate for fifth-generation mobile communication technologies [2]–[4]. These systems are a variation of
multi-user MIMO (MU-MIMO) and have a large number of
base station (BS) antennas that result in an improved spectral
efficiency through spatial multiplexing [5], [6]. Under favorable propagation conditions [1], significant gains in throughput
can be achieved by employing simple linear processing at the
Manuscript received June 10, 2016; revised Dec. 5, 2016; accepted Feb.
8, 2017. The associate editor coordinating the review of this manuscript and
approving it for publication was Dr. Marios Kountouris.
Copyright (c) 2017 IEEE. Personal use of this material is permitted.
However, permission to use this material for any other purposes must be
obtained from the IEEE by sending a request to [email protected].
Part of this work has been presented at the 2016 IEEE International Conf.
Acoustics, Speech, and Signal Processing, Shanghai, China.
This work was supported in part by the Academy of Finland research grant
No. 299243.
K. Upadhya and S. A. Vorobyov are with the Department of Signal
Processing and Acoustics, Aalto University, FI-00076 Aalto, Finland (E-mails:
[email protected], [email protected]).
M. Vehkapera is with the Department of Electronic and Electrical
Engineering, University of Sheffield, Sheffield, S1 3JD, UK (E-mail:
[email protected]).
BS [7], [8]. In addition, large numbers of antennas result in
an improved uplink (UL) energy-efficiency [9] and render the
system performance resilient to hardware impairments [10].
However, all the above mentioned benefits of a massive
MIMO communication system hinge on the assumption that
the BS has access to accurate estimates of the channel state
information (CSI). For systems that employ either frequency
division duplexing (FDD) or time division duplexing (TDD),
the channel estimates are obtained using orthogonal pilot
sequences. In FDD systems, each antenna at the BS transmits
a pilot sequence that is orthogonal to the pilot sequences
transmitted by the other antennas. Since the number of orthogonal pilot sequences required becomes proportional to the
number of BS antennas, FDD is considered impractical for
channel estimation in massive MIMO [5], [11]. Moreover,
since the CSI corresponding to each antenna is estimated
by the users, it has to be fedback from the users to the
BS, consuming additional bandwidth. Consequently, massive
MIMO systems are typically assumed to employ TDD with
full frequency-reuse and utilize channel reciprocity to obtain
CSI. In these systems, each user in a cell is assigned a different
pilot sequence and these pilots are time-multiplexed with
data in each coherence block. When using time-multiplexed
pilots and data, the requirement for high spectral efficiency
necessitates sharing of pilot sequences between adjacent cells,
resulting in the channel estimates of the users in a cell being
corrupted by the channel vectors of users in the adjacent cells.
This phenomenon called ‘pilot contamination’ [12] introduces
interference in both the UL and downlink (DL), degrading the
overall performance of the system.
Existing methods to mitigate pilot contamination for massive MIMO are designed for the case wherein the pilots
are time-multiplexed with the data, henceforth referred to as
time-multiplexed pilots. In [13], it has been observed that
the eigenvectors of the autocorrelation matrix of the received
data correspond to the channel vectors of the desired and
interfering users, and a method for channel estimation has been
developed based on this observation. Pilot decontamination
has been performed in [14] by projecting the contaminated
channel estimate on an interference-free subspace spanned by
the channel vectors of the desired users, whereas [15] derives
asymptotic conditions for separability between the subspaces
of the desired and interfering users. In [16], a coordinated
method for pilot allocation has been proposed for separating
desired and interfering users in correlated channels. A pilot
decontamination method based on the array processing model
has been proposed in [17] for use in parametric channels with
2
a finite number of discrete paths. In [18], a resource allocation
approach has been proposed for optimizing the number of
users scheduled in each cell in order to minimize the effect
of pilot contamination. A common theme for the approaches
described above, except for [16] and [18], is that they focus on
decontaminating the channel estimate at the receiver, which in
this case is the BS. However, since pilot contamination results
from interfering pilot transmissions, there is a scope for better
separating the desired and interfering users by optimizing the
pilot transmissions at the user terminal as well.
In this paper, we propose using superimposed pilots as an
alternative to, and in combination with, time-multiplexed pilots
in massive MIMO systems. Methods for channel estimation
based on pilots that are embedded in data, such as superimposed pilots, have been extensively studied for MIMO systems
[19]–[24]. However, these papers have focused on embedded
and superimposed pilots in the context of accommodating a
loss in signal-to-noise ratio (SNR) in exchange for a reduced
pilot transmission overhead [20], [21]. Particularly, scenarios
with high user-mobility, wherein it is impractical to allocate
dedicated symbols for training, have been of interest for
employing superimposed pilots. In the context of multi-cell
massive MIMO, provided that the number of users in the
system is smaller than the number of symbols in the UL,
superimposed pilots allow for each user in the system to be
assigned a unique pilot sequence, enabling the receiver to
estimate the channel vectors of both the desired and interfering
users. In addition, these pilots mitigate pilot contamination
by time-averaging over long sequences and offer a higher
efficiency due to a reduced transmission overhead.
We obtain expressions for the signal-to-interference-plusnoise ratio (SINR) at the output of a matched filter (MF)based detector when a non-iterative least-squares (LS)-based
channel estimate is employed for channel estimation. Based
on the SINR expression, we highlight the need for power
control when superimposed pilots are employed in a massive
MIMO system. Moreover, we discuss the shortcomings of the
non-iterative channel estimator and propose an intuitive lowcomplexity iterative channel estimation scheme for superimposed pilots.1 In addition, we introduce the concept of a hybrid
system and show by means of theoretical arguments that the
hybrid system is superior to its counterpart that employs only
time-multiplexed pilots, even when the non-iterative channel
estimator is used to obtain the channel estimate from superimposed pilots. A simple approach to design and implement
this hybrid system is also detailed. Although the use of superimposed pilots requires some coordination between the BSs
in assigning pilot sequences to the users and estimating their
path-loss coefficients, these are minor impediments compared
to the performance improvements provided by the proposed
scheme.
The article in the existing literature that is closest to this
paper is [26], wherein superimposed pilots have been employed in the context of multi-cell multiuser MIMO systems.
1 The work in this paper is a significant extension of our relevant conference
paper [25]. In addition to a detailed exposition, we have included additional
results that demonstrate the superiority of massive MIMO systems that use
superimposed pilots instead of time-multiplexed pilots.
However, unlike [26], the focus of our paper is to demonstrate the superiority of superimposed pilots when used in
conjunction with time-multiplexed pilots in a hybrid system.
The theoretical results and simulations that have been obtained
are in line with this objective.
In Section II, the system model for the massive MIMO
UL is described. In Section III, time-multiplexed pilots are
described and the pilot contamination problem is detailed.
Section IV introduces the superimposed pilot scheme and
describes the non-iterative method for channel estimation and
Section V discusses the iterative data-aided scheme. In Section
VI, the concept of a hybrid system that employs both timemultiplexed and superimposed pilots is introduced and in
Section VII, a simple approach for implementing this hybrid
system is discussed. Section VIII presents simulation results
demonstrating the effectiveness of employing superimposed
pilots for pilot decontamination. Section IX concludes the
paper. Some of the proofs are given in Appendix.
Notation : Lower case and upper case boldface letters denote column vectors and matrices, respectively. The notations
(·)∗ , (·)T , (·)H , and (·)−1 represent the conjugate, transpose,
Hermitian transpose, and inverse, respectively. The notation
µ, Σ) stands for the complex normal distribution with
CN (µ
mean µ and covariance matrix Σ and E {·} is used to denote
the expectation operator. The notation IN denotes an N × N
identity matrix, and k·k and k·kF denote the Euclidean norm of
a vector and Frobenius norm of a matrix, respectively. Upper
case calligraphic letters denote sets, and ∅ denotes the empty
set. The notation 1{S} represents the indicator function over
the set S, whereas Card (S) is used to represent its cardinality.
The notation δn,m denotes the Kronecker delta function, and
η(·) stands for an element-by-element decision function that
replaces each element of the input vector with the constellation
point that is closest in Euclidean distance to that element. The
big O notation f (x) = O(g(x)) implies that |f (x)|/|g(x)| is
bounded as x → ∞.
II. S YSTEM M ODEL
We consider a TDD massive MIMO UL with L cells and
K single-antenna2 users per cell. Each cell has a BS with
M K antennas. The number of symbols C, over which
the channel is coherent, is assumed to be divided into Cu
and Cd , which are the number of symbols in the UL and DL
time slots, respectively. The matrix of received measurements
Yj ∈ CM ×Cu at BS j can be written as
Yj =
L−1
X K−1
X
√
µ`,k hj,`,k sT`,k + Wj
(1)
`=0 k=0
where µ`,k denotes the power with which the vector of
symbols s`,k ∈ CCu ×1 are transmitted by user k in cell `,
Wj ∈ CM ×Cu is the additive white Gaussian noise at BS
j with each column distributed as CN (0, σ 2 IM ). Moreover,
the columns of Wj are mutually independent of each other.
The vector hj,`,k ∈ CM ×1 represents the channel response
2 For training and channel estimation, users with T > 1 antennas can be
treated as T separate single-antenna users.
3
between the antennas at BS j, and user k in cell `, and is
assumed to be distributed as3
hj,`,k ∼ CN (0, βj,`,k IM )
the LS estimate of the channel of user m in cell j can be
obtained as [1]
L−1
(2)
where βj,`,k denotes the large-scale path-loss coefficient which
depends on the user location in the cell. In addition, the
channel vectors hj,`,k are assumed to be mutually independent
of each other ∀j, `, k. The aforementioned statistics of the
channel vector correspond to the non-line-of-sight scenario
with rich scattering [1]. By virtue of their zero mean and
mutual independence, the channel vectors are asymptotically
orthogonal and the following equation holds almost surely [1]
hH
j,`,k hm,n,p
= βj,`,k δj,m δ`,n δk,p , ∀ j, k, `, m, n, p .
M →∞
M
(3)
Moreover, hj,`,k is assumed to be constant for the duration
of C symbols, and βj,`,k is constant for a significantly longer
duration which depends on the user mobility. For the sake of
simplicity, the effects of shadowing are not taken into account
in this paper.4 The transmitted symbols s`,k contain both
pilots and data. The pilots could either be time-multiplexed
or superimposed pilots, and the elements of the data vector
x`,k are assumed to be independent and identically distributed
(i.i.d) random variables with zero-mean and unit variance and
take values from an alphabet χ, which is a realistic assumption.
lim
III. T IME -M ULTIPLEXED P ILOTS A ND T HE P ILOT
C ONTAMINATION P ROBLEM
With time-multiplexed pilots, each user in a cell transmits a
τ ≥ K length orthogonal pilot sequence for channel estimation
followed by Cu − τ symbols of uplink data. In order to
minimize the overhead incurred, it is necessary to reuse these
pilot sequences in the adjacent cells. However, this pilot-reuse
results in the channel estimates of the desired users being
contaminated by the channel vectors of users in adjacent cells,
causing interference and in turn, a loss in spectral efficiency.
It is assumed here that the transmission of the pilot sequences by the users in the L cells are synchronized, which
corresponds to the worst-case scenario for pilot contamination.5 Consider a unitary matrix Φ ∈ Cτ ×τ whose columns
φ1 , . . . , φ τ } are the orthogonal pilot sequences that are
{φ
transmitted by the users, i.e., φ H
n φ p = τ δn,p . If φ b`,k is
the pilot sequence transmitted by user k of cell `, where
b`,k ∈ {1, . . . , τ } is the index of the transmitted pilot, and if
each pilot sequence is reused once every r , τ /K cells [18],
3 While, for the sake of simplicity, an environment with rich scattering is
assumed, the conclusions made in this paper are independent of the channel
distribution and only require the channel vectors of any pair of users to be
asymptotically orthogonal.
4 The algorithms and analysis in this paper remain the same in the presence
of shadowing, provided that the users are allocated to the strongest BSs.
However, the geometric interpretations that are made based on the location
of the user in the cell will no longer be valid.
5 No additional improvement in the UL performance can be gleaned by
separating the pilot and data transmissions across cells [1], [9], [18].
X
1
(p)
b TP , √1 Y(p)φ ∗ = hj,j,m +
h
hj,`,m + √ Wj φ ∗bj,m
bj,m
j,j,m
j
τ pu
τ pu
`=0
`6=j
`∈Lj (r)
(4)
where the superscript TP indicates that the estimates are computed when using time-multiplexed pilots, the superscript p
indicates that the observations are made during pilot transmission, and Lj (r) is the subset of the L cells that use the same set
of pilots as cell j. In addition, it is assumed in (4) without loss
of generality that the transmit powers are same for all users
employing time-multiplexed pilots, i.e., µ`,k = pu , ∀`, k, and
any variation in the transmit power of an individual user is
absorbed into the corresponding path-loss coefficient β. It can
be observed from (4) that the estimates of the channel vectors
of the users in cell j are contaminated by the channel vectors
of the users in the remaining Card (Lj (r) − 1) cells. When
M → ∞, the UL SINR of user m in cell j, at the output of
an MF that uses the channel estimate in (4) for detection, can
be written as [1]
SINRTP−ul
=
j,m
β2
P j,j,m2
.
βj,`,m
(5)
`6=j
`∈Lj (r)
The corresponding throughput of the user using Gaussian
signaling in the UL can then be expressed as [1]
(Cu − τ )
TP−ul
TP−ul
Rj,m
=
log2 1 + SINRj,m
.
(6)
C
From the above equation, it can be observed that the rate per
user is a function of both the overhead τ as well as the loss in
SINR due to pilot contamination. A larger value of r would
reduce the effect of pilot contamination and increase the SINR
at the cost of a reduced transmission efficiency (Cu − τ )/C.
IV. S UPERIMPOSED P ILOTS
With superimposed pilots, the pilot symbols are transmitted
at a reduced power alongside the data symbols, and in its
simplest version, the pilot and data symbols are transmitted
alongside each other for the entire duration of the uplink data
slot Cu . If the total number of users in the system is smaller
than the number of symbols in the uplink, i.e., KL ≤ Cu ,6
then with superimposed pilots, each user can be assigned a
unique orthogonal pilot p`,k ∈ CCu ×1 . The pilots are taken
from the columns of a unitary matrix P ∈ CCu ×Cu such
that PH P = Cu ICu , and therefore pH
`,k pn,p = Cu δ`,n δk,p .
If ρ`,k x`,k + λ`,k p`,k is the transmitted vector from user k in
6 For example, using the orthogonal frequency division multiplexing
(OFDM) parameters in long-term evolution (LTE) systems as in [1], i.e.,
Cu = 7 OFDM symbols, Nsmooth = 14 subcarriers, and assuming the pilots
are reused over L = 7 hexagonal cells, the maximum number of supported
users in the L cells is Cu Nsmooth = 98 users. Therefore, the number of
users per cell is Cu Nsmooth /L = 14 users. However, note that the value of
Cu = 7 has been chosen to allow user velocities of 350 km/h [27]. For lower
user speeds and with cell sectoring, larger number users can be supported and
the assumption KL ≤ Cu will easily be satisfied.
4
SINRSP−ul
=
j,m
2
ρ2`,k µ`,k βj,`,k
2
Cu λ2j,m ρ2j,m βj,j,m
`=0 k=0
+
1
M
L−1
X K−1
X
`=0 k=0
{`6=j,k6=m}
cell `, then the received signal at the j’th BS Yj ∈ CM ×Cu ,
when using the superimposed pilot scheme, can be written as
Yj =
L−1
X K−1
X
T
hj,`,k (ρ`,k x`,k + λ`,k p`,k ) + Wj
(7)
`=0 k=0
where λ2`,k and ρ2`,k are the fractions of the transmit power
reserved for the pilot and data symbols, respectively, and the
total transmitted power µ`,k is given as µ`,k = λ2`,k + ρ2`,k .
A. Non-Iterative Channel Estimation
Treating the data symbols of all users as additive noise, the
channel estimate of user k in cell ` can be obtained at the j’th
BS using the LS criterion [26]
b j,`,k , arg minkYj − λ`,k h pT k2 .
h
`,k F
(8)
h
Solving (8) yields
b j,`,k = Yj λ2 pH p`,k
h
`,k `,k
= hj,`,k +
+
−1
L−1
X K−1
X
−1
λ`,k p∗`,k =
1
Yj p∗`,k
Cu λ`,k
L−1
X K−1
X
1
ρm,n hj,m,n xTm,n p∗`,k
Cu λ`,k m=0 n=0
1
Wj p∗`,k .
Cu λ`,k
(9)
In order to estimate the data from the received observations, it
is necessary to remove the term corresponding to the transmitted superimposed pilot λj,m hj,j,m pTj,m from the observation
b j,j,m pT as an estimate for this
vector in (7). Using λj,m h
j,m
term, the estimate of xj,m can then be obtained from the
observation Yj using an MF and a decision operation as
follows
1
T
bH
b
eTj,j,m =
h
(10)
x
j,j,m Yj − λj,m hj,j,m pj,m
M ρj,m βj,j,m
bj,j,m = η (e
x
xj,j,m ) .
(11)
The SINR of user m in cell j, at the output of an MF that
employs the channel estimate in (9), is derived in Appendix A
and is given in (12) (shown at the top of the page). The SINR
in (12), when M → ∞, can be written as
SINRSP−ul
=
j,m
1
Cu
2
λ2j,m ρ2j,m βj,j,m
.
L−1
P K−1
P 2
2
ρ`,k µ`,k βj,`,k
(13)
`=0 k=0
The corresponding per-user rate in the uplink when using
Gaussian signaling is given as
Cu
SP−ul
Rj,m
=
log2 1 + SINRSP−ul
.
(14)
j,m
C
βj,`,k µ`,k
+
ρ2j,m βj,j,m
L−1
X K−1
X
L−1
X K−1
X
ρ2n,p βj,`,k βj,n,p µ`,k
2
Cu λ2j,m ρ2j,m βj,j,m
`=0 k=0 n=0 p=0
{`6=j,k6=m} {n6=`,p6=k}
(12)
B. Power Control and Choice of Parameters λj,m and ρj,m
From (13), it can be seen that the SINR of a user is
dependent on the product of the transmit powers and largescale fading coefficients of the remaining LK − 1 users in
addition to the product of its own transmit power and largescale fading coefficient. This dependence results in a situation
similar to the near-far problem in code division multiple access
(CDMA) systems, wherein users that have larger values of
large-scale fading coefficient β swamp users that have smaller
values of β. Therefore, it becomes necessary to use power
control to provide a uniform user experience.
While the parameters µ`,k , ρ`,k , and λ`,k can be optimized
by maximizing the sum-rate of all the users, i.e.,
(L−1 K−1
)
XX
SP−ul
max
R`,k
(15)
µ`,k ,ρ`,k ,λ`,k
`=0 k=0
the optimization problem is in general non-convex and requires
coordination between the BSs. As an alternative, a suboptimal
solution that does not involve coordination between the BSs
is obtained here for the parameters µ`,k , ρ`,k , and λ`,k . This
suboptimal solution will be shown to maximize a lower bound
on the sum-rate, and it is as follows.
The received signal in (7) can be equivalently written as
L−1
X K−1
X√
ρ`,k T
λ`,k T
µ`,k hj,`,k √
x`,k + √
p`,k +Wj
Yj =
µ`,k
µ`,k
=
`=0 k=0
L−1
X K−1
X
h̄j,`,k ρ̄`,k x`,k + λ̄`,k p`,k
T
+ Wj
(16)
(17)
`=0 k=0
where
h̄j,`,k ,
√
µ`,k hj,`,k ∼ CN 0, β̄j,`,k IM
β̄j,`,k , βj,`,k . µ`,k
s
ρ2`,k
ρ̄`,k ,
>0
µ`,k
s
λ2`,k
λ̄`,k ,
>0
µ`,k
(18)
λ̄2`,k + ρ̄2`,k = 1 .
(21)
(19)
(20)
From (16), it can be seen that a system having arbitrary
values of βj,`,k , µ`,k , ρ`,k , and λ`,k , can be reduced into an
equivalent system with parameters β̄j,`,k , ρ̄`,k , and λ̄`,k , such
that 0 ≤ ρ̄`,k , λ̄`,k ≤ 1. Substituting (18) – (21) into (12), an
equivalent expression for the SINR, as shown in (22) (shown
at the top of the next page) can be obtained.
To obtain the parameter µ`,k , we propose using the
statistics-aware power-control approach detailed in [18],
wherein user m in cell j transmits at a power µj,m = ω/βj,j,m
5
2
ρ̄2`,k β̄j,`,k
SINRSP−ul
=
j,m
C λ̄2 ρ̄2 β̄ 2
`=0 k=0 u j,m j,m j,j,m
+
1
M
L−1
X K−1
X
`=0 k=0
{`6=j,k6=m}
β̄j,`,k
+
ρ̄2j,m β̄j,j,m
L−1
X K−1
X
L−1
X K−1
X
SINRSP−ul
≥
j,m
ρ̄2`,k
Cu λ̄2j,m ρ̄2j,m
`=0 k=0
+
LK − 1
+
M ρ̄2j,m
L−1
X K−1
X
`=0 k=0 m=0 n=0
{`6=j,k6=m} {m6=`,n6=k}
where ω is a design parameter. The parameter ω is chosen such
that the transmitted power from a user satisfies a maximum
power constraint, and users with severely low SINRs that
would need a transmit power larger than this constraint would
be denied service. This power control policy results in an
identical received power of ω at the j’th BS for all the
users in cell j. In addition, as mentioned in [18], the ratio
0 ≤ βj,`,k /β`,`,k ≤ 1 is the relative strength of the interference
received at BS j from a user in cell `. This ratio is at most
1, when the user is at the edge of the j’th cell, and reduces
to zero as its distance from BS j increases. Therefore, setting
µ`,k = ω/β`,`,k and using the definitions of β̄j,`,k , ρ̄`,k , and
λ̄`,k , and the inequality 0 ≤ βj,`,k /β`,`,k ≤ 1, the following
equations can be obtained
β̄j,j,m = βj,j,m . µj,m = ω
(23)
= βj,`,k . µ`,k ≤ β`,`,k µ`,k = ω
=
=
≤
≤
ρ̄2j,m ω
λ̄2j,m ω
ρ2`,k β`,`,k
λ2`,k β`,`,k
∀ ` 6= j
(24)
(25)
(26)
= ρ̄2`,k ω
=
λ̄2`,k ω
∀ ` 6= j
∀ ` 6= j .
L−1
X K−1
X
(27)
(28)
Substituting the above equations into (22), a lower bound on
the SINR, as shown in (29) (shown at the top of the page),
can be obtained.
However, the maximization of the lower bound on the
SINR and hence, a lower bound on the sum rate, is still
a non-convex problem in the parameters ρ̄`,k and λ̄`,k and
requires coordination between the BSs. To circumvent this
problem, we restrict the parameters ρ̄`,k and λ̄`,k such that
ρ̄`,k = ρ̄, ∀ `, k and λ̄`,k = λ̄, ∀ `, k. The choice of this
restriction is motivated by the observation from (23) that
the statistics-aware power control scheme results in the same
large-scale path loss coefficient for all the desired users in the
cell, irrespective of their locations. As a result, from the BS’s
perspective, each of its users are identical, and therefore, there
is no benefit in assigning different values of ρ̄`,k to different
users. More importantly, such a restriction renders the choice
of ρ̄opt to depend only on L, K, Cu , and M as will be shown
next. Setting ρ̄`,k = ρ̄, ∀ `, k and λ̄`,k = λ̄, ∀ `, k in (29), we
obtain
!!
2 −1
1 LK − 1 (LK − 1)
LK
SP−ul
+
+
.
SINRj,m ≥
Cu (1 − ρ̄2 ) M
ρ̄2
Cu (1 − ρ̄2 )
(30)
L−1
X K−1
X
`=0 k=0 m=0 n=0
{`6=j,k6=m} {m6=`,n6=k}
ρ̄2m,n β̄j,`,k β̄j,m,n
2
Cu λ̄2j,m ρ̄2j,m β̄j,j,m
(22)
−1
β̄j,`,k
2
ρj,m βj,j,m
λ2j,m βj,j,m
ρ2`,k βj,`,k
λ2`,k βj,`,k
−1
L−1
X K−1
X
ρ̄2m,n
2
2
M Cu λ̄j,m ρ̄j,m
(29)
Differentiating the right hand side of (30) with respect to ρ̄2
and setting the resulting expression to zero, the value of ρ̄2
that maximizes the lower bound on SINRSP−ul
and the UL
j,m
sum rate can be obtained as
v
!−1
u LK (LK−1)2 −1
r
u
M + LK
t Cu + M C u
2
(31)
ρ̄opt=1+
≈ 1+
LK−1
Cu
M
and the optimal value of λ̄2 can be obtained as
!−1
r
Cu
2
2
λ̄opt = 1 − ρ̄opt ≈ 1 +
M + LK
(32)
where the approximations in (31) and (32) have been made
assuming LK 1 in order to obtain simpler expressions.
Based on the fact established in this subsection that systems using ρ, λ, β, h, and ρ̄, λ̄, β̄, h̄ are equivalent, we drop
the overbar for ease of notation and adopt the former set
of symbols in the rest of the paper. In addition, we set
µ`,k = ρ̄2`,k + λ̄2`,k = 1, ∀`, k.
C. Impact of Cu on the Performance of Superimposed Pilots
Using (13) and a fixed set of parameters r, τ , and K,
the following theorem presents an important condition that
guarantees the superiority of methods based on superimposed
pilots over the LS estimator that is based on time-multiplexed
pilots.
Theorem 1. With fixed values of K, r, and τ and if M → ∞,
there exists a UL duration κj,m beyond which a channel
estimator based on superimposed pilots outperforms the LS
based channel estimator that utilizes time-multiplexed pilots,
in terms of the SINR performance, in any channel scenario
{βj,`,m 0 ≤ j, ` ≤ L − 1, 0 ≤ m ≤ K − 1}.
Proof. If κj,m is defined as the number of symbols in the
uplink such that (5) and (13) are equal, i.e.,
1
κj,m λ2j,m ρ2j,m
2
βj,j,m
L−1
P K−1
P
`=0 k=0
=
2
ρ2`,k βj,`,k
β2
P j,j,m2
βj,`,m
(33)
`6=j
`∈Lj (r)
then it is evident from (13) and (33) that Cu > κj,m is a
sufficient condition for a method that is based on superim-
6
posed pilots to outperform the LS method that employs timemultiplexed pilots. In addition, κj,m is given as
1
κj,m ,
L−1
P K−1
P
2
ρ2`,k βj,`,k
`=0 k=0
P
2
βj,`,m
`6=j
`∈Lj (r)
λ2j,m ρ2j,m
.
(34)
This completes the proof.
Remark 1: An important consequence of the above theorem is that in scenarios with negligible pilot contamination, the LS method based on superimposed pilots requires
a large value of Cu to outperform the LS method based
on time-multiplexed pilots. As an example, consider the
case when r = 1, βj,j,m = 1, ∀m, βj,`,m = β, ∀` 6= j, m, and
ρ2j,m = λ2j,m , ∀ j, m. For such a scenario, κj,m is given as
1
.
(35)
κj,m = 2K 1 +
(L − 1)β 2
Then, if the LS estimator based on superimposed pilots is
required to maintain superiority over the LS estimator employing time-multiplexed pilots, Cu must scale inversely with
β 2 . This dependence on Cu is evident from the expression for
the channel estimation error, which is given as
b j,j,m = − 1
∆hj,j,m , hj,j,m − h
Cu λj,m
!
L−1
K−1
XX
T
×
ρ`,k hj,`,k x`,k + Wj p∗j,m . (36)
`=0 k=0
Remark 2: We build upon the discussion in [1] on grouping
users based on their coherence times. While such a grouping
does not offer any performance benefits to users when employing the approach in [1], the use of superimposed pilots offers
low-mobility users an increase in throughput, by minimizing
the channel estimation error resulting from transmitting the
data alongside the pilots. This improvement in performance is
a direct consequence of Theorem 1.
Remark 3: The type of pilot transmitted by a user can also be
chosen based on the coherence time. While users with highmobility or low pilot contamination would find it sufficient
to use time-multiplexed pilots, users with low-mobility who
suffer from significant pilot contamination due to their proximity to the cell-edge or due to shadowing would significantly
benefit from employing superimposed pilots.
Remark 4: Superimposed pilots require coordination between
BSs when assigning pilot sequences and synchronizing transmissions. In practical cellular networks, the cells are fairly
large and therefore, it can be assumed that the interference is
restricted to the first tier of cells and the interference from the
second and higher tiers of cells can be neglected. Therefore,
it is reasonable to assume that practical deployments of
superimposed pilots will require pilot assignment only over
the first tier of cells, implying that coordination is limited to
only this first tier. This overhead is not very different from that
required by time-multiplexed pilots in the presence of pilot
reuse. The coordination and synchronization requirements of
superimposed pilot-based systems that allocate pilots over the
first tier of cells are similar to that of time-multiplexed pilotbased systems that have a pilot reuse factor of r = 3 [18].
From (36), it can be seen that the error in the channel
estimate includes interference resulting from transmitting data
alongside the pilots. Hence, the quality of the channel estimate
can be improved by eliminating the interference from the
transmitted data through iterative data-aided schemes, thereby
increasing the robustness of the proposed method with respect
to Cu .
V. I TERATIVE DATA -A IDED C HANNEL E STIMATION
In the iterative approach to channel estimation developed in
this section, the estimated channel and data vectors of both the
desired and interfering users are used in feedback in order to
eliminate the first term in (36). In addition, to minimize error
propagation between the channel estimates of different users,
the iteration is started from the user with the highest SINR
and is progressed in the decreasing order of the SINRs of the
users. It has to be noted that the objective of this section is
to demonstrate that iterative methods for channel estimation
with superimposed pilots provide a significantly better SINR
performance than their non-iterative counterparts, and hence
we restrict ourselves to a simple iterative algorithm. However,
there is scope for developing improved iterative algorithms in
the future.
A. Algorithm
For the sake of clarity and without loss of generality, we
replace the two indices k, ` with a single index m that lies in
the range 0 ≤ m ≤ N − 1, where N , KL. The index m is
used to index the users in all the L cells. In addition, we drop
the index j and implicitly assume that the channel estimation
is performed at the j’th BS. Then, (7) can be rewritten as
Y=
N
−1
X
T
hm (ρm xm + λm pm ) + W .
(37)
m=0
Since for large M , the SINRs of the users are proportional to the users’ path-loss coefficients, the users are arranged in the decreasing order of their path-loss coefficients,
i.e., β0 > β1 > . . . > βN −1 .7 Then, using an estimate of
ρm hm xTm p∗m for each user as a correction factor to minimize
the interference from other users, the corresponding channel
estimate of user m can be written as
b (i) =
h
m
m−1
T
X
1
(i)
b (i) x
Y −
b
ρ
h
k
k
k
Cu λm
k=0
(i)
k∈Um
−
N
−1
X
k=m
(i)
k∈Um
T
(i−1)
b (i−1) x
p∗m
b
ρk h
k
k
(38)
7 It is assumed that the BSs have access to the exact values of the pathloss coefficients βm and that there is no false-ordering. This assumption is
reasonable since for large M , the path-loss coefficients can be computed at the
BS with negligible error by averaging the power of the channel coefficients
over the entire array.
7
(0)
(i)
(i)
(p)
b m = 0, ∀ m and Um is the set of users whose (S3) ∆xm is independent of ∆x , ∀p 6= i, m 6= k and the
where h
k
(i)
estimated data is used in feedback in the i’th iteration to
elements of ∆xm are i.i.d.
(p)
estimate the channel vector of user m. The approach to obtain (S4) ∆h(i)
m is independent of xk , W, and ∆xk , ∀ k, p.
(i)
Um has been detailed in Appendix C, and involves selecting
In scenarios with low interference and with large M , only
users such that the interference power, described in the next
a few of the received symbols will be erroneous. As a
subsection, does not increase with each iteration. The channel
(i)
result, the elements of ∆xm are sparse with the few nonestimate in the above equation is a modified version of the
zero elements restricted to locations that correspond to the
LS estimator defined in (9) with an added correction factor.
(i)
erroneous symbols. Moreover, the vector em represents the
Utilizing the resulting channel estimate in an MF and decision
error in the estimated data and in such low-interference
operation, similar to (10) and (11), the estimate of the data is
(i)
scenarios, the elements of em take small values. Therefore, the
obtained as follows
simplifications (S1), (S2), and (S3) are reasonably accurate for
H
T
1
(i)
(i) T
(i)
(i)
b
b
em
(39) these scenarios. Although the expression for ∆hm , (given in
Y − λm hm pm
hm
x
=
(i)
M ρ m βm
(82) in Appendix B) is explicitly dependent on xk and ∆xm ,
(i)
(i)
bm = η x
em
(40) we neglect the correlation between these terms since ∆h(i)
x
m
is inversely proportional to Cu , and the simplification (S4) is
(0)
bm = 0, ∀ m = 0, . . . , N − 1.
where x
fairly accurate when Cu is large with respect to N and when
(i)
Remark 5: If the unitary matrix P, whose columns are the su- scenarios with low interference are considered. Since em is
(i)
perimposed pilots, is chosen as P = blkdiag{P0 , . . . , PL−1 }, assumed to be a zero-mean random variable, ∆xk is also a
where the `’th block P` ∈ CK×K is comprised of the orthog- zero-mean random variable, provided the constellation points
onal pilot sequences used by the K users in cell `, then the in χ and their probability density functions are symmetric
(i)
latency introduced when the non-iterative method is employed about the origin. This is true since by definition, ∆xk and
(i)
is the same as that for time-multiplexed pilots. However, when em are related to each other through the following equation
the iterative method is employed, the channel and the data
(i)
vectors of the users are required and therefore, the uplink data
∆xk = xk − η xk − e(i)
.
(43)
m
in the entire slot will have to be aggregated before estimating
the channel, which introduces a latency of Cu symbols.
From (43), an expression for the variance of the elements of
(i)
Remark 6: From (9), the non-iterative method for channel ∆x(i)
k , i.e., αk can be found as
estimation requires M Cu operations per user, whereas the MF
h
i 2 Z
and decision operations in (10) and (11) require M and Cu
(i)
(i)
2
αk , E
∆xk
= |∆x| p∆x(i) (∆x)d∆x
operations per user, respectively.
k
n
Z Z
For the iterative method with ν iterations, the channel esti2
=
|x − η (x − e)| pe(i) ,x (e, x) de dx
mator, matched filter, and decision operations have a combined
k
k
x∈χ
complexity of O(νM Cu ) + O(νM ) + O(νCu ).
Z Z
B. Interference Power at the BS
(i)
k=0
k6=m
(i)
where the expression for ψm is given in (42) on the top of
(0)
the next page and ψm = 0, ∀m. The detailed derivation of
(i)
Im can be found in Appendix B.
In deriving (41), the following simplifying assumptions have
been made in order to obtain a closed form expression:
(i)
(S1) em is independent of xk and W, ∀ k, i.
(i)
(S2) ∆xm is independent of xk , W, and hk , ∀ k, i.
|x − η (x − e)| pe(i) (e) pxk (x) de dx
(44)
k
x∈χ
(i)
em be the error in the estimate of the data
Let em , xm − x
symbols of user m obtained from the MF in the i’th iteration.
(i)
(i)
bm be the corresponding error vector
Let ∆xm , xm − x
(i)
(i)
bm
after the decision operation and let ∆hm , hm − h
be the
(i)
associated error in the channel estimate. If αn is the variance
(i)
of the elements of ∆xn and assuming that the elements of
(i)
em are i.i.d. circular complex-Gaussian random variables with
(i)
zero mean and variance Im , an approximate expression for the
(i)
interference power Im can be written as
N
−1
2
X
1 1
σ βm
1
(i)
(i)
Im
≈ 2 2
βk βm +
+ 2 2 ψm
(41)
βm M ρ m
M ρ2m
M ρm
2
=
where pe(i) (·), p∆x(i) (·), and pxk (·) are the probability density
k
k
(i)
(i)
functions of the elements of ek , ∆xk , and xk , respectively,
and pe(i) ,x (·) is the joint density function of the random
k
k
(i)
variables ek and xk . The latter has been written as the product
of their individual distributions in the final expression of (44),
thanks to (S1).
(i)
Important example of αm : When the elements of xm are
uniformly distributed and take values from a unit-power P quarternary amplitude modulation (QAM) constellation, then
(i)
under the assumption that the symbol errors in ∆xk are
dominated by the closest neighboring symbols, the expression
(i)
for αm can be written as
!
r
3
(P
−1)
√ √
24
, i≥1
Q
(i)
(i)
P ( P +1)
Im
(45)
αm
=
1,
i=0
where Q (·) is the Q-function. The detailed derivation of the
above expression can be found in Appendix D.
8
2
i≥1
X
(i)
k∈Um ,m≤k≤N
τ
X
(i)
ρ2k βk2 αk
UTP
Uplink Data
Pilots
1
+
M
N
−1
X
(i)
βn βk αk
Cd
Downlink Data
Cu
τ
No
Transmission
(i)
)
N
−1
X
1
ρ2k βk2 +
+
+
βn βk
2
M
M n=0
(i)
(i)
n=0
k∈Um ,k<m
k∈U
/ m
!
(i−1)
−1
N
−1
σ 2 NX
1 + αk
X
1
(i−1)
(i−1)
(i−1)
ψk
+
(42)
ρ2k βk2 αk
+
βn βk αk
+
βn .
2
M
M
M
n=0
n=0
M
=
Cu λ2m
(i)
ψm
+
USP
Uplink Data
+
Superimposed Pilots
Cd
Downlink Data
1 + αk
(
(i)
ψk
X
data phase of Cu − τ symbols. Since these users maintain
radio silence during the pilot training phase of τ symbols,
they do not affect the quality of the channel estimates of the
users in UTP . As a result, under the assumption of asymptotic
orthogonality of the channels, there is no interference from the
users in USP to those in UTP . Therefore, the per-cell sum-rate
in the UL for the users in UTP remains unchanged and can
be found from (6) to be
Rul
j (UTP ) =
K−1
(Cu − τ ) X
log2 1 +
C
k=0
k∈UTP
Cu
Fig. 1. Frame structure of a hybrid system with users employing timemultiplexed and superimposed pilots.
VI. H YBRID S YSTEM
One of the main advantages of superimposed pilots over
time-multiplexed pilots is that it does not require a separate
set of symbols for pilot transmission. This property can be
used to construct a hybrid system that contains two disjoint
sets of users, with the users in one of the sets employing timemultiplexed pilots, and the users in the other set employing
superimposed pilots. The following theorem shows that this
hybrid system has a higher throughput and supports a larger
number of users than a system that employs only timemultiplexed pilots.
Theorem 2. In a system that employs time-multiplexed pilots
and is designed to maximize the UL and DL sum-rate,8 let
K be the optimal number of users per cell, L be the total
number of cells in the system, τ > 0 be the optimal number
of symbols used for pilot training, r be the optimal pilot-reuse
factor, and Cu − τ and Cd be the number of data symbols in
the UL and DL slots, respectively. Then, with M → ∞, there
exists a hybrid system, that uses both time-multiplexed and
superimposed pilots, which is capable of supporting Cu − τ
additional users and offers a higher sum-rate in the UL than
the optimal system that only employs time-multiplexed pilots.
Proof. Consider the frame structure in Fig. 1, wherein there
are two sets of users UTP and USP . The users in the set
UTP employ time-multiplexed pilots, with parameters selected
using approaches such as in [18]. The users in the set USP
maintain radio silence during the pilot training phase of the
users in UTP , i.e., for τ symbols in the frame, and transmit
orthogonal pilots superimposed with data during the uplink
8 Such
as the scheme described in [18].
β2
P j,j,k2
.
βj,`,k
`6=j
`∈Lj (r)
(46)
Assuming, for the sake of simplicity, that all the users in USP
are located in the j’th cell, the sum-rate of the users in USP
can be found using (13) and (14) as
(Cu − τ ) X
log2 (1 + SINRm (USP )) (47)
Rul
j (USP ) =
C
SINRm (USP ) ,
P
m∈USP
2
βj,j,m
2
ρ2j,k βj,j,k
(Cu −τ )ρ2j,m λ2j,m
.
(48)
k∈USP
In obtaining the above expression, it has been assumed that
the transmit power pu of the users in UTP is small enough
such that the interference to the users in USP can be neglected.9 Therefore, from (46) and (48), the combined rate
ul
ul
Rul
j (USP ) + Rj (UTP ) is strictly greater than Rj (UTP ). In
addition, since the data slot is made up of Cu − τ symbols, it
is possible to allocate Cu − τ orthogonal pilots and therefore,
the set USP can contain a maximum of Cu − τ users. This
concludes the proof.
In the above theorem, given a system with users employing
time-multiplexed pilots, we have shown that additional users
employing superimposed pilots can always be added to the
system, resulting in a hybrid system that offers a higher
throughput.
In the following section, we utilize the concept of the
above theorem to partition a given set of users employing
time-multiplexed pilots into two disjoint subsets UTP and
USP that contain users transmitting time-multiplexed pilots
and superimposed pilots, respectively. There are two main
benefits of performing such a partition: (i) there is an overall
9 This assumption is valid since the SINR and the rate of the users in U
TP
are independent of the transmit power pu when M → ∞. It has to be noted
that this assumption has been made for the sake of simplicity and the theorem
is valid even if this assumption does not hold.
9
improvement in the throughput as a result of the reduced intercell interference; and (ii) there is a reduction in the number
of users that use time-multiplexed pilots, thereby allowing for
more aggressive pilot reuse since r is a function of the number
of users employing time-multiplexed pilots [18].
`=0 k=0
(`,k)∈USP
VII. A S IMPLE I MPLEMENTATION OF THE H YBRID
S YSTEM
Given a set of K users per cell in L cells with channel
gains βj,`,k , ∀j, ` = 1, . . . , L, and k = 1, . . . , K, the problem
of partitioning users into disjoint sets UTP and USP can be
accomplished by minimizing the overall UL inter-cell and
intra-cell interference. This choice of objective function is
motivated by Theorem 1, wherein it is observed that users at
the cell edge cause significant pilot contamination and benefit
from being assigned superimposed pilots, whereas users that
are close to the BS cause negligible interference and could
be assigned time-multiplexed pilots that are potentially shared
with users in neighboring cells.
A. Framework
If the users in UTP transmit pilots with unit power and data
at a power pu , then the received signal from the hybrid system
in the UL phase at BS j can be written as
Yj =
YjTP
+
YjSP
+ Wj
(49)
where YjTP and YjSP are the received signals from the users
in UTP and USP , respectively. From Fig. 1, YjTP and YjSP
can be written as
YjTP ,
L−1
X K−1
X
√
hj,`,k φ T`,k , pu xT`,k
(50)
`=0 k=0
(`,k)∈UTP
YjSP
,
L−1
X
X K−1
hj,`,k 01×τ , ρxT`,k + λpT`,k
(51)
`=0 k=0
(`,k)∈USP
where the tuple (`, k) is used to denote user k in cell `.
If user (j, m) is a member of UTP , then the LS estimate of
its channel can be written as [1]
X
1
b j,j,m = 1 Yj bTP = hj,j,m +
h
hj,`,m + Wj bTP
j,m
j,m (52)
τ
τ
`6=j
`∈Lj (r)
(`,m)∈UTP
H
T
where bTP
. If M K, the SINR
j,m , φ j,m , 0(1×(Cu −τ ))
in the UL when using the channel estimate in (52) can be
obtained similar to (5) as
SINRTP−ul
j,m
2
βj,j,m
≈ P 2
βj,`,m
If user (j, m) is a member of USP , then the LS estimate of
its channel can be written as
1
b j,j,m =
Yj bSP
h
j,m
(Cu − τ ) λ
L−1
X K−1
X
ρ
= hj,j,m +
hj,`,k xT`,k p∗j,m
(Cu − τ ) λ
(53)
`6=j
`∈Lj (r)
(`,m)∈UTP
where the approximations in (53) is made for the sake of
simplicity and is valid when M is sufficiently large.
+
√
pu
L−1
X K−1
X
`=0 k=0
(`,k)∈UTP
hj,`,k xT`,k p∗j,m
+ Wj bSP
j,m
(Cu − τ ) λ
(54)
T
H
. Since it can be seen from
where bSP
j,m , 0(1×τ ) , pj,m
(53) that the UL SINR of the users in UTP is independent
of the UL transmit power pu , we assume that pu is small
enough with respect to the transmit powers of the users in
USP . As a result, the users in USP do not experience significant
interference during the data transmission phase of the users in
UTP and result in the transmissions of USP and UTP becoming
independent of each other.10 Then (54) simplifies as
L−1
X hj,`,k xT`,k p∗j,m
X K−1
b j,j,m ≈ hj,j,m + ρ
+ Wj bSP
h
j,m .
λ
(Cu − τ )
`=0 k=0
(`,k)∈USP
(55)
Then the SINR in the UL for the users in USP can be obtained
from (13) as
SP−ul
≈
SINRj,m
2
βj,j,m
L−1
P K−1
P
1
(Cu −τ )λ2
`=0 k=0
(`,k)∈USP
(56)
2
βj,`,k
where, similar to (53), the approximation in (56) is made for
the sake of simplicity and is valid when M is sufficiently large.
B. Algorithm to Obtain UTP and USP .
The goal in this subsection is to obtain an algorithm for
partitioning users into the sets UTP and USP by minimizing
the total UL inter-cell and intra-cell interference. In order to
accomplish this, we quantify the amount of interference caused
by a user that is assigned to either of the sets UTP or USP .
TP−ul
SP−ul
Let Ij,m
or Ij,m
be the contributions of user (j, m) to
the total UL inter/intra-cell interference power when assigned
to UTP or USP , respectively. If users (j, m) and (`, k) are
members of UTP , then from the denominator of (53), the
amount of interference that user (j, m) causes to user (`, k)
2
in the UL is β`,j,k
δm,k . Likewise, from (56), if both users are
members of USP , then the amount of interference that user
2
(j, m) causes to user (`, k) in the UL is β`,j,m
/ (Cu − τ ) λ2 .
TP−ul
SP−ul
Therefore, Ij,m
and Ij,m
can be obtained as
X K−1
X
X
TP−ul
2
2
β`,j,k
Ij,m
=
δm,k =
β`,j,m
`6=j k=0
`∈Lj (r)
(`,k)∈UTP
(57)
`6=j
`∈Lj (r)
(`,m)∈UTP
10 This assumption is made for the sake of clarity and simplicity. In the
absence of this assumption, the BS will have to estimate and remove YjTP
from Yj before estimating the channels of the users in USP .
10
SP−ul
Ij,m
=
L−1
X K−1
X
1
2
β`,j,m
.
2
(Cu − τ ) λ
(58)
`=0 k=0
(`,k)∈USP
From the above equations, the total cost due to UL inter/intracell interference can be expressed as
I (UTP , USP ) =
L−1
X K−1
X
`=0 k=0
TP−ul
I`,k
1{(`,k)∈UTP }
SP−ul
+ I`,k
1{(`,k)∈USP }
(59)
Using (59) as the objective function, the sets UTP and USP
can be obtained as the solution of the following optimization
problem
(UTP , USP ) = arg min I (UTP , USP )
UTP ⊆U
USP ⊆U
subject to UTP ∪ USP = U
UTP ∩ USP = ∅
(60)
where U is the set of all users in the L cells. However,
the optimization problem in (60) is combinatorial in nature
with 2Card(U ) possible choices for UTP and USP , making
it computationally hard to obtain the optimal solution. A
workaround is to employ a greedy approach to partition U
into UTP and USP
. Ateach step of this algorithm, given UTP
˜ k̃ in UTP is chosen as
and USP , a user `,
˜ k̃ = arg
`,
max
(`,k)∈UTP
TP−ul
I`,k
.
(61)
0
˜ k̃ and U 0 = USP ∪ `,
˜ k̃ , user
Setting UTP
= UTP \ `,
SP
˜ k̃ is added to USP if
`,
0
0
I (UTP
, USP
) ≤ I (UTP , USP ) .
(62)
The algorithm is initialized with UTP = U and is terminated
when (62) is no longer satisfied or when UTP is empty. The
approach described above is summarized in Algorithm 1.
The complexity of the greedy algorithm used for designing
the hybrid system can be obtained as follows. The terms
I TP−ul and I SP−ul require a maximum of Card (U) operations to compute, and therefore, computing I (UTP , USP )
2
requires Card (U) operations. Assuming that the greedy
algorithm runs till the condition UTP = ∅ is satisfied, then an
upper bound on the computational complexity of the greedy
3
algorithm is Card (U) operations. Moreover, an overhead of
2Card (U) data transmissions is required for sending the largescale path-loss coefficients to a central node and receiving the
sets UTP and USP .
It has to be noted that Algorithm 1 is sub-optimal, but it is
useful for illustrating the concept of the hybrid system. Partitioning algorithms that offer superior performance compared
to Algorithm 1 with lower coordination overhead are left as
topics for future research.
Algorithm 1 Greedy algorithm to select UTP and USP
Data: βj,`,k , ∀j, ` = 0, . . . , L − 1, k = 0, . . . , K − 1
Initialize: UTP
← U, USP ← ∅
˜ k̃ as in (61)
1: Compute `,
0
˜ k̃ and U 0 ← USP ∪ `,
˜ k̃
2: Set UTP
← UTP \ `,
SP
0
0
3: if UTP 6= ∅ and if I (UTP
, USP
) ≤ I (UTP , USP ) then
0
0
4:
UTP := UTP , USP := USP
5:
Return to Step (1).
6: else
7:
STOP
8: end if
VIII. S IMULATION R ESULTS
We compare the UL SINR and UL bit-error rate (BER)
performance of the LS-based and eigenvalue decomposition (EVD)-based methods (referred to as ‘TP - EVD-based
method’ in the plots) that use time-multiplexed pilots to the
performance of the channel estimator that uses superimposed
pilots, at the output of a MF that employs these channel
estimates. Two scenarios are considered for this comparison.
Scenario 1: The users are uniformly distributed in hexagonal cells of radius 1km with the BS at the center. In
addition, users are located at a distance of at least 100m
from the BS.
Scenario 2: Users in both the reference and interfering
cells are in a fixed configuration and are equally spaced
on a circle of a given radius with the BS in the center. The
size of the hexagonal cell is 1km and unless otherwise
specified, the users are on a circle of radius 800m.
Unless otherwise specified, the following parameters are used
in both scenarios. The channel estimation methods are tested
with L = 7 cells and K = 5 users per cell. A P QAM constellation is employed and the path-loss coefficient
is assumed to be 3. The simulations for the superimposed
pilots-based iterative channel estimation scheme have been
performed for 4 iterations. The number of symbols in the
uplink time slot Cu is set to 100, and for computing the
rate, C is set to 200 symbols. The values of ρ and λ are
computed from (31) and (32), respectively, and ω is set to 1,
where ω is the design parameter in the statistics-aware power
control scheme. The signal-to-noise ratio (SNR), i.e., ω/σ 2 is
set to 10dB. The methods based on time-multiplexed pilots
have been simulated with r = 1 and pu = 1. In addition,
the chosen channel estimation methods have been observed to
perform better with the statistics aware-power control scheme,
and therefore, this power control scheme has been employed
for both time-multiplexed and superimposed pilots. The plots
in Scenario 1 are generated by averaging over 104 realizations
of user locations across the cell. For each realization of user
location, the channel vectors are generated and 200 bits are
transmitted per user. The BER is computed by counting the
bit errors for all the users in the reference cell. Similarly, the
plots in Scenario 2 are generated for a fixed user location
by averaging over 104 channel realizations with 200 bits
transmitted per user for each realization.
11
2
22
SP - Ufixed
Approximate Rate - bits/s/Hz
18
(i)
SINR
14
10
6
2
100
200
300
400
500
Number of antennas at the BS (M )
1.4
1.2
1
200
300
400
500
Number of antennas at the BS (M )
600
10−1
(i)
Um
SP - Optimal
SP - Non-iterative
TP
10−2
0.7
0.6
BER
Cumulative Probability
1.6
Fig. 3. Approximate per-user UL rate obtained using a 16-QAM constellation vs. M in the reference BS in Scenario 2. The maximum UL rate that
can be achieved with the 16-QAM constellation, with half the symbols in a
coherence blocked used for UL transmission, is 2 bps/Hz.
1
0.8
1.8
0.8
100
600
Fig. 2. The UL SINR of a user in the reference BS vs. M in Scenario 2.
The values of ρ and λ are computed from (31) and (32), respectively, and
since they are approximations, they result in a non-smooth SINR behavior
for the iterative methods. The solid and dashed lines represent simulated and
theoretical curves, respectively.
0.9
SP - Optimal Um
SP - Ufixed
SP - Non-iterative
TP
(i)
SP - Optimal Um
SP - Non-iterative
TP
0.5
10−3
0.4
(i)
0.3
SP - Optimal Um
SP - Ufixed
SP - Non-iterative
TP - EVD based method in [14]
TP - LS Estimate
−4
10
0.2
0.1
0
−5
0
5
10
15
20
UL SINR (dB)
10−5
1
2
3
4
5
6
7
8
Number of users per cell (K)
9
10
Fig. 4. Cumulative distribution of the UL SINR in dBs for users in Scenario
1 with M = 300 antennas. The black line indicates SINRs with probability
≥ 0.95.
Fig. 5. BER in the UL vs. K in Scenario 1 with M/K = 50 and Cu = 70
symbols.
Fig. 2 shows the variation of the UL SINR of an arbitrary
user with respect to M in Scenario 2, whereas in Fig. 3,
the approximate rate of an arbitrary user, calculated using
16-QAM constellation, is plotted for the same scenario. We
compute the achievable rate for 16-QAM signaling, modeling
a practical scenario where highly mobile users are requesting
moderate-to-high data rates. The SINR when the proposed
method is employed, is shown to linearly increase in the
number of antennas, whereas the SINR performance is observed to saturate for the LS-based method that uses timemultiplexed pilots. This trajectory of the proposed method
could be potentially maintained using techniques such as
adaptive modulation and coding, thereby implying that the
effects of pilot contamination can be eliminated.
over 100 channel and data realizations for each realization
of user location. While the LS-based method employing
time-multiplexed pilots offers a higher SINR than the LS
method employing superimposed pilots with a probability of
approximately 0.6, the latter method can be seen to offer a
significantly higher minimum SINR compared to the former
method. Moreover, the users employing superimposed pilots
have a smaller variation in their SINR than those employing
time-multiplexed pilots. This is because the SINR of a user
when superimposed pilots are employed is limited by the
interference from the other users in the same cell, and the
statistics-aware power control scheme renders the intra-cell
interference power independent of the user location within
the cell. The iterative method based on superimposed pilots
is observed to offer a remarkably higher SINR performance
with respect to its non-iterative counterpart and the LS-based
In Fig. 4, the cumulative distribution of the UL SINR in
Scenario 1 is plotted. The interference power is averaged
12
ω
.
SIRRx
j , P P 2
βj,`,k
(63)
`6=j k
We assume L = 19 hexagonal cells, i.e., a central cell with
two tiers of interfering cells. Each cell has M = 1000
antennas, K = 5 users, and the value of Cu is chosen as 40
symbols. Although L is set to 19, the optimization described in
Algorithm 1 and the computation of the performance metrics
is performed over 7 cells which consist of the central and the
first tier of cells. The value of ω for users in USP is set to
10 and pu for the users in UTP is set to 1. The data symbols
are Gaussian distributed and the sum rate in Fig. 6 is obtained
by averaging over 103 realizations of the channel and data
symbols.
In Fig. 6, high and low values of SIR correspond to users
located close to the BS and at the cell-edge, respectively. It
can be observed that channel estimation methods based only
on superimposed pilots (even the non-iterative formulation) are
better in high interference scenarios, i.e., when the interfering
users are at the cell-edge, whereas time-multiplexed pilots are
better in low-interference scenarios. This behavior is a direct
consequence of Theorem 1 since higher interference scenarios
have smaller values of κ, resulting in superimposed pilots
outperforming time-multiplexed pilots. However, at smaller
values of user radius, the impact of pilot contamination is low
but the self-interference in superimposed pilots resulting from
transmitting the data alongside the pilots leads to a poorer
performance compared to methods based on time-multiplexed
pilots. In addition, it can be seen that the hybrid system adapts
to the level of inter and intra-cell interference and offers a
performance that is resilient to the location of the user within
the cell.
140
Hybrid
TP
SP - Non-iterative
120
UL Sum Rate - bits/s/Hz
method employing time-multiplexed pilots.
In Fig. 5, the BER is plotted against the number of users
per cell in Scenario 1, with K ranging from 1 to 10 and
Cu = 70 symbols. Since L = 7 cells, K = 10 implies
that the superimposed pilot-based system cannot support any
new users without sharing pilots across cells. The ratio M/K
is set to 50. While the non-iterative channel estimator based
on superimposed pilots performs better in the UL at lower
values of K than the estimators based on time-multiplexed
pilots, the non-iterative estimator performs poorly at higher
values of K. This is because the data transmitted alongside
the pilots causes self-interference and this interference power
increases with the number of users in the system. Therefore,
it is necessary to resort to iterative techniques to mitigate this
additional interference and it can be seen that the iterative
methods offer a better performance than methods based on
time-multiplexed pilots when LK is close to Cu .
In Fig. 6, the users are distributed as in Scenario 2 and the
distance of the users from the BS is varied between 0.2 and
0.9 km. For the chosen range of user distance, the total rate
in the UL is plotted against the corresponding received signalto-interference ratio (SIR). The received SIR of an arbitrary
user m in cell j is defined as
100
80
60
40
20
0
−10
0
10
20
30
40
50
SIRRx
(dB)
j
Fig. 6. UL sum rate vs. SIRRx
in Scenario 2 with M = 1000 antennas.
j
IX. C ONCLUSION AND D ISCUSSION
We have proposed superimposed pilots as a superior alternative to time-multiplexed data and pilots for uplink channel
estimation in massive MIMO. In the limit of an infinite
number of antennas, a hybrid system using both superimposed
pilots and time-multiplexed data and pilots offers a higher
UL rate and supports larger number of users than the optimal
system that utilizes only time-multiplexed data and pilots. The
resilience to pilot contamination can be significantly improved
with superimposed pilots through the use of an iterative dataaided channel estimation scheme that utilizes the data symbols
of both the desired and interfering users in the feedback loop.
Computer simulations in both a realistic scenario, in which
users are distributed uniformly over the entire cell, and a highinterference scenario, in which users are concentrated at the
cell edge, show that channel estimation methods using superimposed pilots offer a significant performance improvement
over those that use time-multiplexed pilots.
The objective of this paper is to advocate superimposed
pilots for practical use in massive MIMO systems by showing
their superiority through theoretical and simulation based
investigations. In standard MIMO communications, superimposed pilots are typically argued to be useful only for the
scenario with high user mobility, and therefore, have not found
practical application. On the contrary, in massive MIMO,
superimposed pilots in a hybrid system provide superior
performance in general. Therefore, there is a strong reason
for superimposed pilots to make their way to practical use.
The proposed iterative data-aided channel estimation
scheme and the greedy algorithm for partitioning users are
suboptimal solutions to corresponding non-convex problems.
Algorithms that offer performance close to the optimal solution at low computational complexities and overheads are
of interest for future research. Moreover, the downlink performance of superimposed pilots is another topic of practical
importance, which we have partially addressed in [28].
13
N
−1 N
−1
X
X
ρ2k βn βk µn
E ki2 k2 + ki3 k2 + ki4 k2 ≈
2
M λ2m ρ2m βm
n=0
A PPENDIX A
Uplink SINR of the Non-Iterative Channel Estimation Method
Using the notation described in Section V-A, (7) can be
rewritten as
N
−1
X
T
Y=
hm (ρm xm + λm pm ) + W
(64)
k=0
n6=m k6=n
N
−1 N
−1
X
X
ρ2n ρ2p βp βn
ρ2n µn βn2
+
+
2
λ2 ρ2 β 2
Cu ρ2m λ2m βm
n=0 p=0
n=0 m m m
N
−1
X
n6=m p6=m
n6=p
m=0
N
−1 N
−1
X
X
ρ2n ρ2p βn βp
E ki5 k2 ≈
2
Cu λ2m ρ2m βm
n=0 p=0
From (9), the estimation error of the channel estimate can be
obtained as
!
N
−1
X
1
T
bm = −
∆hm , hm − h
ρk hk xk + W p∗m .
Cu λm
k=0
(65)
From (10) and (65), the estimate of the received data after MF
with the estimated channel can be written as
1
b m pT
b H Y − λm h
eTm =
h
x
m
m
M ρ m βm
1
=
hH − ∆hH
m
M ρ m βm m
N
−1
X
T
×
hk (ρk xk + λk pk ) + W
k=0
−λm (hm − ∆hm ) pTm = gT + iT
(66)
where g and i are the signal and interference components of
the matched filtered signal, respectively, which can be written
as
khm k2
g,
xm
(67)
M βm
5
X
in
(68)
i,
i1 ,
n=1
N
−1
X
n=0
n6=m
(69)
λm
hH ∆hm pm
(70)
M ρ m βm m
1
i3 , −
∆hH
(71)
m hm xm
M βm
T
N
−1
X
∆hH
1
mW
H
∆hm hn (λn pn + ρn xn ) −
i4 , −
M ρm βm n=0
M ρ m βm
i2 ,
(77)
n6=m p6=m
n6=p
E
N
−1 N
−1
X
X
ρ2n ρ2p βn βp
H
i
≈
−
iH
+
i
5
3
4
2
Cu λ2m ρ2m βm
p=0 n=0
(78)
p6=m n6=m
n6=p
where the approximation errors in (75) – (78) are proportional
to either N/M , N/Cu , or Cu /M . In addition, the remaining
terms of the form iH
n ip , ∀ n 6= p in the expansion of (74) are
proportional to N/M or N/Cu . If M is large with respect to
N and Cu , then the approximation errors and terms that are
proportional to N/M and N/Cu can be neglected. Similarly,
error terms that are proportional to N/Cu can also be dropped,
and if σ 2 Cu , then the effect of noise can also be neglected.
Then, substituting (75) – (78) into the expansion of (74), the
interference power is obtained as
N
−1
N
−1
X
X
ρ2n µn βn2
Cu βn µn
E kik2 ≈
+
2 ρ2 β 2
λ
M
ρ2m βm
n=0 m m m
n=0
n6=m
+
T
hH W
hH
m hn
(λn pn + ρn xn ) + m
M ρm β m
M ρm β m
(76)
−1
N
−1 N
X
X
ρ2k βn βk µn
2
M λ2m ρ2m βm
n=0 k=0
n6=m k6=n
.
(79)
Using (79), the SINR can be obtained as
E kgk2
SINRSP−ul
,
m
E {kik2 }
Cu
= N −1
.
−1 NP
−1 2
N
−1
P ρ2n µn βn2
P Cu βn µn NP
ρk βn βk µn
+
+
λ2 ρ2 β 2
M ρ 2 βm
M λ2 ρ2 β 2
n=0
m m
m
n=0
n6=m
m
n=0 k=0
n6=m k6=n
m m
m
(80)
n6=m
(72)
λm
i5 , −
k∆hm k2 pm .
M ρ m βm
The average interference power can be found as
2
5
X
2
E kik = E
in
.
(73)
(i)
(74)
Then, using the definitions of in , ∀ n in (69) – (73) and the
definition of ∆hm in (65), the following expressions can be
easily obtained
N
−1
X
Cu
βn µn
M ρ2m βm n=0
n6=m
A PPENDIX B
Interference Power of the Iterative Method
n=1
E ki1 k2 ≈
It completes the derivation of (12).
(75)
To derive the SINR, using the definition of ∆xm , xm −
the channel estimate in (38) can be simplified as
T
X
X
(i)
b (i) x
b (i) = hm + 1 ρk hk xT −
bk
ρk h
h
k
m
k
Cu λm
k
k∈Uj ,k<m
T
X
(i−1)
b (i−1) x
bk
−
ρk h
+ Wj p∗m
(81)
k
(i)
bm ,
x
k∈Uj ,m≤k≤N
14
where
T
1 X
(i)
(i)
ρk hk ∆xk
+ ∆hk xTk
Cu λm
k∈Uj ,k<m
T
T
X
(i)
(i)
(i−1)
(i−1) T
−∆hk ∆xk
+
ρk hk ∆xk
+∆hk
xk
∆h(i)
m =−
k∈Uj ,m≤k≤N
(i−1)
−∆hk
(i−1)
∆xk
T
!
+
X
ρk hk xTk + W
p∗m .
k∈U
/ j
(82)
The received symbols after MF in (39) are then given as
−1
H NX
1
T
H
(i)
hm− ∆hm
hk (ρk xk + λk pk )
=
M ρm
k=0
T
p
+W − λm hm − ∆h(i)
m
m
bTm
x
7
=
X
1
khm k2 xTm +
aTk
M
(83)
k=1
where
N −1
1 X H
a1 ,
hm hk (ρk xk + λk pk )
M ρm
T
1
hH
mW
M ρm
λm H
a3 ,
h ∆h(i)
m pm
M ρm m
H
1
∆h(i)
a4 , −
hm xm
m
M
N −1
1 X (i) H
∆hm
a5 , −
hk (ρk xk + λk pk )
M ρm
(84)
(85)
(86)
H T
1
∆h(i)
W
m
M ρm
2
λm
a7 , −
∆h(i)
pm .
m
M ρm
(87)
(88)
(89)
(90)
k=1
where the terms a6 , a7 , and aH
p aq , ∀p, q have been dropped.
Further, it can be shown straightforwardly that
N −1
Cu X
β k βm
E ka1 k2 =
M ρ2m
(92)
2
E ka2 k
Cu σ 2 βm
=
.
M ρ2m
E ka5 k
N
−1 N
−1
H
X
X
1
(i)
H
(i)
= 2 2
E
∆hm
h` hk ∆hm
M ρm
`=0 k=0
`6=m k6=m
n
o
H
× E (ρ` x` + λ` p` ) (ρk xk + λk pk )
N
−1
X
H
Cu
H
(i)
h
h
∆h
.
(96)
∆h(i)
= 2 2 E
k k
m
m
M ρm
k=0
Summing up (94), (95), and (96), we obtain
( 5
)
X
Cu
2
E
kak k
= 2 2
M ρm
k=3
(
!
)
−1
H NX
(i)
H
(i)
×E
∆hm
hk hk ∆hm
.
(97)
(i)
Now, let ψm be defined as the second term in (97), i.e.,
(
!
)
−1
H NX
(i)
(i)
ψm
,E
∆hm
hn hH
∆h(i)
. (98)
n
m
i≥1
Using (82) and the simplifications (S1) to (S4), (98) can be
simplified to obtain (42). Substituting (92), (93), (97), and (42)
(i)
into (91), Im can be obtained as
(i)
Im
≈
N −1
σ 2 βm
1
1 X
(i)
β
β
+
+ 2 2 ψm
.
k
m
M ρ2m
M ρ2m
M ρm
(99)
k=0
k6=m
It completes the derivation of (41).
A PPENDIX C
(i)
Choice of the Set of Users Um
Let S be a set of the KL users in the system and let P (S)
be its power set. In addition, for the sake of clarity, let the
(i)
(i)
additional argument Um be added to the functions Im and
(i)
(i)
ψm in this section. Now, the optimal set Um can be obtained
by solving the following optimization problem
n
o
(i)
(i)
Um
= arg min Im
(U) .
(100)
U ∈P(S)
k=0
k6=m
2
n=0
Under the assumption that the interference power at each of the
received symbols is the same, the average interference power
of the m’th user at the j’th cell is given as
" ( 5
)#
2
7
X
X
1
1
(i)
E
E
kak k2
(91)
Im
=
ak
≈
Cu
Cu
k=1
k=0
k=0
k6=m
a6 , −
and
k6=m
k=0
k6=m
a2 ,
Moreover, E ka3 k2 , E ka4 k2 , and E ka5 k2 can be
written as
H
λ2m
2
H
(i) T ∗
(i)
E ka3 k = 2 2 E hm ∆hm pm pm ∆hm
hm
M ρm
H
Cu λ2
(i)
(i)
E
∆h
hm hH
(94)
= 2 m
m
m ∆hm
M ρ2m
H
1
(i)
E ka4 k2 = 2 E
∆h(i)
hm xTm x∗m hH
m
m ∆hm
M
H
1
(i)
(i)
= 2 E xTm x∗m E
∆hm
hm hH
∆h
m
m
M
H
Cu
(i)
= 2E
∆h(i)
hm hH
(95)
m
m ∆hm
M
(93)
Substituting (41) into (100) yields
n
o
(i)
(i)
Um
= arg min ψm
(U) .
U ∈P(S)
(101)
15
(i)
Now, ψm (U) can be rewritten as
(i)
ψm
(U) = c +
N
−1 n
X
n=0
(i)
ξn 1{n∈U
/ } + n (U)1{n∈U ,n<m}
+(i−1)
(U)1{n∈U ,n≥m}
n
o
(102)
(i)
where c, ξn , and n (U) are defined as
!
N −1
M σ2 X
βk
c,
Cu λ2m
k=0
(
)
N −1
1 X
M 2 ρ2n
2
βn +
βk βn
ξn ,
Cu λ2m
M
k=0
(
N
−1
2 2
X
M
ρ
1
n
2 (i)
(i)
β
α
+
βk βn αn(i)
n (U) ,
n
n
Cu λ2m
M
k=0
(i)
1 + αn
(i)
+
ψ
(U)
.
n
M2
(103)
(i)
n∈Um
(i)
< ψm
(i)
n∈U
/ m
.
The decision rule in (112) results in a set Ufixed that is
computed at the beginning of the first iteration and is left
unchanged for the subsequent iterations.
(104)
A PPENDIX D
Derivation of
(105)
(i)
αm
for a P -QAM constellation
For P -QAM constellation and i ≥ 1, the integral over xm
in (44) reduces to a summation, which can be written as
XZ
2
(i)
αm =
|x − η (x − e)| pe(i) (e) pxm (x) de . (113)
m
x∈χ
It can be seen from (102) that the optimization problem (101)
is separable over the user indices, implying that the decision
(i)
to include user n in Um is independent of the other N − 1
users. Therefore, the channel and data estimates of user n are
used in the i’th iteration if the following condition is satisfied
(i)
(i)
n ∈ Um
iff ψm
However, since the decision rules are based on approximate
SINR expressions, it is worth commenting that the reliability
of the decision rule in (108) decreases with increasing user
and iteration indices. Alternatively, a fixed and conservative
decision rule can be used to obtain U as follows
n
o
(2)
(2)
(1)
Ufixed = m ∈ N Im
({m}) < Im
(∅) = Im
(∅) .
(112)
(106)
Since the P symbols are equally likely, pxm (x) = 1/P, ∀ x
and under the assumption that the errors x − η (x − e) are
dominated by the closest neighboring symbols, the above
equation reduces to
dx
X
1
(i)
d2 kx Q q 2 (i)
(114)
αm
=
P x∈χ x
Im
2
(i)
From (102) and (106), the set Um is obtained as
n
(i)
Um
= n ∈ N (i)
n (U) < ξn when n < m
o
and (i−1)
(U) < ξn when n ≥ m .
n
(107)
Equivalently, using (104) and (105), the above expression
simplifies to
n
(i)
Um
= n ∈ N αn(i) < γn(i) when n < m
o
and αn(i−1) < γn(i−1) when n ≥ m
(108)
where
(
βn2
γn(i)
+
1
M
,(
βn2
+
1
M
NP
−1
k=0
NP
−1
k=0
βn βk −
(i)
ψn
)
(i)
n∈Um
M2
βn βk +
(0)
).
(i)
ψn
where dx is the distance between the symbol x and its closest
neighbor and kx is the number of symbols at a distance of dx
from x. The Q-function in the above equation results from the
(i)
assumption on the statistics
of em . For a unit-power P -QAM
p
constellation, dx = 6/P − 1, ∀x [29]. In addition, it can be
easily verified
that kx = 2 for the 4 corner symbols, kx = 3
√
for the ( P − 2)4 symbols
√ on the outer edges, and kx = 4
for the remaining P − 4 P + 4 symbols. Substituting these
values into (114) yields
s
3
24
−1)
(i)
.
Q (P (i)
αm
= √ √
(115)
i≥1
Im
P
P +1
(109)
(i)
n∈Um
M2
R EFERENCES
If xm takes values from the P -QAM constellation, then
(i)
substituting (45) into (108), the set Um can be obtained as
n
(i)
Um
= n ∈ N In(i) < fn(i) when n < m
o
and In(i) < fn(i−1) when n ≥ m
(110)
(i)
where fn is defined as
√ √
(i)
P
P + 1 γn
3
.
fn(i) ,
Q2
P −1
24
(0)
Moreover, since ∆x`,m = x`,m , the value of αm is 1. It
completes the derivation of (45).
(111)
[1] T. Marzetta, “Noncooperative cellular wireless with unlimited numbers
of base station antennas,” IEEE Trans. Wireless Commun., vol. 9, no. 11,
pp. 3590–3600, Nov. 2010.
[2] F. Boccardi, R. Heath, A. Lozano, T. Marzetta, and P. Popovski, “Five
disruptive technology directions for 5G,” IEEE Commun. Mag., vol. 52,
no. 2, pp. 74–80, Feb. 2014.
[3] E. Larsson, O. Edfors, F. Tufvesson, and T. Marzetta, “Massive MIMO
for next generation wireless systems,” IEEE Commun. Mag., vol. 52,
no. 2, pp. 186–195, Feb. 2014.
[4] J. Andrews, S. Buzzi, W. Choi, S. Hanly, A. Lozano, A. Soong, and
J. Zhang, “What will 5G be?” IEEE J. Sel. Areas Commun., vol. 32,
no. 6, pp. 1065–1082, Jun. 2014.
[5] L. Lu, G. Li, A. Swindlehurst, A. Ashikhmin, and R. Zhang, “An
overview of massive MIMO: Benefits and challenges,” IEEE J. Sel.
Topics Signal Process., vol. 8, no. 5, pp. 742–758, Oct. 2014.
16
[6] F. Rusek, D. Persson, B. K. Lau, E. Larsson, T. Marzetta, O. Edfors,
and F. Tufvesson, “Scaling up MIMO: Opportunities and challenges with
very large arrays,” IEEE Signal Process. Mag., vol. 30, no. 1, pp. 40–60,
Jan. 2013.
[7] J. Hoydis, S. ten Brink, and M. Debbah, “Massive MIMO in the UL/DL
of cellular networks: How many antennas do we need?” IEEE J. Sel.
Areas Commun., vol. 31, no. 2, pp. 160–171, Feb. 2013.
[8] H. Yang and T. Marzetta, “Performance of conjugate and zero-forcing
beamforming in large-scale antenna systems,” IEEE J. Sel. Areas Commun., vol. 31, no. 2, pp. 172–179, Feb. 2013.
[9] H. Q. Ngo, E. Larsson, and T. Marzetta, “Energy and spectral efficiency
of very large multiuser MIMO systems,” IEEE Trans. Commun., vol. 61,
no. 4, pp. 1436–1449, Apr. 2013.
[10] E. Bjornson, J. Hoydis, M. Kountouris, and M. Debbah, “Massive
MIMO systems with non-ideal hardware: Energy efficiency, estimation,
and capacity limits,” IEEE Trans. Inf. Theory, vol. 60, no. 11, pp. 7112–
7139, Nov. 2014.
[11] E. Björnson, E. G. Larsson, and T. L. Marzetta, “Massive MIMO: ten
myths and one critical question,” IEEE Commun. Mag., vol. 54, no. 2,
pp. 114–123, Feb. 2016.
[12] J. Jose, A. Ashikhmin, T. Marzetta, and S. Vishwanath, “Pilot contamination and precoding in multi-cell TDD systems,” IEEE Trans. Wireless
Commun., vol. 10, no. 8, pp. 2640–2651, Aug. 2011.
[13] H. Q. Ngo and E. Larsson, “EVD-based channel estimation in multicell
multiuser MIMO systems with very large antenna arrays,” in Proc. IEEE
Int. Conf. on Acoustics, Speech and Signal Processing (ICASSP), Kyoto,
Mar. 2012, pp. 3249–3252.
[14] R. Muller, L. Cottatellucci, and M. Vehkapera, “Blind pilot decontamination,” IEEE J. Sel. Topics Signal Process., vol. 8, no. 5, pp. 773–786,
Oct. 2014.
[15] J. Vinogradova, E. Björnson, and E. G. Larsson, “On the separability
of signal and interference-plus-noise subspaces in blind pilot decontamination,” in Proc. IEEE Int. Conf. on Acoustics, Speech and Signal
Processing (ICASSP), Mar. 2016, pp. 3421–3425.
[16] H. Yin, D. Gesbert, M. Filippou, and Y. Liu, “A coordinated approach
to channel estimation in large-scale multiple-antenna systems,” IEEE J.
Sel. Areas Commun., vol. 31, no. 2, pp. 264–273, Feb. 2013.
[17] K. Upadhya and S. A. Vorobyov, “An array processing approach to pilot
decontamination for massive MIMO,” in Proc. IEEE 6th Int. Workshop
on Computational Advances in Multi-Sensor Adaptive Processing (CAMSAP), Cancun, Dec. 2015, pp. 453–456.
[18] E. Björnson, E. G. Larsson, and M. Debbah, “Massive MIMO for
maximal spectral efficiency: How many users and pilots should be
allocated?” IEEE Trans. Wireless Commun., vol. 15, no. 2, pp. 1293–
1308, Feb. 2016.
[19] K. Takeuchi, R. R. Müller, M. Vehkaperä, and T. Tanaka, “On an
achievable rate of large rayleigh block-fading MIMO channels with no
CSI,” IEEE Trans. Inf. Theory, vol. 59, no. 10, pp. 6517–6541, Oct.
2013.
[20] S. He, J. Tugnait, and X. Meng, “On superimposed training for MIMO
channel estimation and symbol detection,” IEEE Trans. Signal Process.,
vol. 55, no. 6, pp. 3007–3021, Jun. 2007.
[21] M. Coldrey and P. Bohlin, “Training-based MIMO systems – part I:
Performance comparison,” IEEE Trans. Signal Process., vol. 55, no. 11,
pp. 5464–5476, Nov. 2007.
[22] H. Zhu, B. Farhang-Boroujeny, and C. Schlegel, “Pilot embedding for
joint channel estimation and data detection in MIMO communication
systems,” IEEE Commun. Lett., vol. 7, no. 1, pp. 30–32, Jan. 2003.
[23] T. Cui and C. Tellambura, “Pilot symbols for channel estimation in
OFDM systems,” in Proc. IEEE Global Telecommunications Conf.
(GLOBECOM), St. Louis, vol. 4, Dec. 2005, pp. 5 pp.–2233.
[24] M. Ghogho, D. McLernon, E. Alameda-Hernandez, and A. Swami,
“Channel estimation and symbol detection for block transmission using
data-dependent superimposed training,” IEEE Signal Process. Lett.,
vol. 12, no. 3, pp. 226–229, Mar. 2005.
[25] K. Upadhya, S. A. Vorobyov, and M. Vehkapera, “Superimposed pilots:
An alternative pilot structure to mitigate pilot contamination in massive
MIMO,” in Proc. IEEE Int. Conf. on Acoustics, Speech and Signal
Processing (ICASSP), Shanghai, Mar. 2016, pp. 3366–3370.
[26] H. Zhang, S. Gao, D. Li, H. Chen, and L. Yang, “On superimposed
pilot for channel estimation in multi-cell multiuser MIMO uplink: Large
system analysis,” IEEE Trans. Veh. Technol., vol. 65, no. 3, pp. 1492–
1505, Mar. 2016.
[27] C. Cox, An introduction to LTE: LTE, LTE-advanced, SAE and 4G
mobile communications, 2nd ed. Chichester, UK: John Wiley & Sons,
2012.
[28] K. Upadhya, S. A. Vorobyov, and M. Vehkapera, “Downlink performance of superimposed pilots in massive MIMO systems in the presence
of pilot contamination,” in Proc. IEEE Global Conf. on Signal and
Information Processing (GlobalSIP), Washington D.C., Dec. 2016, pp.
665–669.
[29] J. Proakis and M. Salehi, Digital Communications. McGraw-Hill, 2008.
Karthik Upadhya received the B.E. degree in electronics and communication engineering from Visvesvaraya Technological University, India in 2007 and
the M.Tech degree in communication systems from
Indian Institute of Technology Madras, India in
2011. He is currently a Ph.D. student at Aalto
University, Finland. His research interests include
wireless communications and array signal processing. Before joining the Ph.D. program, he worked
as a Member of Technical Staff at Saankhya Labs,
India from 2011 to 2013 and as a research assistant
at the Indian Institute of Science (IISc), India from 2013 to 2014.
Sergiy Vorobyov (M’02-SM’05) received the M.Sc.
and Ph.D. degrees in systems and control from
Kharkiv National University of Radio Electronics,
Ukraine, in 1994 and 1997, respectively. He is a Professor with the Department of Signal Processing and
Acoustics, Aalto University, Finland. He has been
previously with the University of Alberta, Alberta,
Canada as an Assistant Professor from 2006 to 2010,
Associate Professor from 2010 to 2012, and became
Full Professor there in 2012. Since his graduation,
he also held various research and faculty positions
at Kharkiv National University of Radio Electronics, Ukraine; the Institute
of Physical and Chemical Research (RIKEN), Japan; McMaster University,
Canada; Duisburg-Essen University and Darmstadt University of Technology,
Germany; and the Joint Research Institute between Heriot-Watt University
and Edinburgh University, U.K. His research interests include optimization
and liner algebra methods in signal processing and communications; statistical
and array signal processing; sparse signal processing; estimation and detection
theory; sampling theory; and multi-antenna, very large, cooperative, and
cognitive systems.
Dr. Vorobyov is a recipient of the 2004 IEEE Signal Processing Society
Best Paper Award, the 2007 Alberta Ingenuity New Faculty Award, the 2011
Carl Zeiss Award (Germany), the 2012 NSERC Discovery Accelerator Award,
and other awards. He is serving as Area Editor for IEEE Signal Processing
Letters since 2016. He served as an Associate Editor for the IEEE Transactions
on Signal Processing from 2006 to 2010 and for the IEEE Transactions
on Signal Processing Letters from 2007 to 2009. He was a member of the
Sensor Array and Multi-Channel Signal Processing and Signal Processing for
Communications and Networking Technical Committees of the IEEE Signal
Processing Society from 2007 to 2012 and from 2010 to 2016, respectively.
He has served as the Track Chair for Asilomar 2011, Pacific Grove, CA, the
Technical Co-Chair for IEEE CAMSAP 2011, Puerto Rico, and the Tutorial
Chair for ISWCS 2013, Ilmenau, Germany.
Mikko Vehkapera (M’10) received the Ph.D. degree from Norwegian University of Science and
Technology (NTNU), Trondheim, Norway, in 2010.
Between 2010–2013 he was a post-doctoral researcher at School of Electrical Engineering, and the
ACCESS Linnaeus Center, KTH Royal Institute of
Technology, Sweden, and 2013–2015 an Academy
of Finland Postdoctoral Researcher at Aalto University School of Electrical Engineering, Finland.
He is now an assistant professor (lecturer) at University of Sheffield, Department of Electronic and
Electrical Engineering, United Kingdom. He held visiting appointments at
Massachusetts Institute of Technology (MIT), US, Kyoto University and
Tokyo Institute of Technology, Japan, and University of Erlangen-Nuremberg,
Germany. His research interests are in the field of wireless communications,
information theory and signal processing. Dr. Vehkapera was a co-recipient for
the Best Student Paper Award at IEEE International Conference on Networks
(ICON2011) and IEEE Sweden Joint VT-COM-IT Chapter Best Student
Conference Paper Award 2015.
| 7cs.IT
|
Communications in Information and Systems
Volume 0, Number 0, 1–9, 0000
arXiv:1803.09871v1 [cs.IT] 27 Mar 2018
Rate-distortion functions of non-stationary Markoff
chains and their block-independent approximations
Mukul Agarwal
It is proved that the limit of the normalized rate-distortion functions of block independent approximations of an irreducible, aperiodic Markoff chain is independent of the initial distribution of the
Markoff chain and thus, is also equal to the rate-distortion function
of the Markoff chain.
AMS 2000 subject classifications: Primary 00K00 *, 00K01; secondary 00K02 *.
Keywords and phrases: * .
1. Introduction
Consider a random source which evolves on a finite set. It follows from existing literature, see for example [1] and [2] (Pages 491-500, in particular, see
Definition (9.8.3) and Theorem 9.8.3 for achievability), that the limit of the
normalized rate-distortion functions of block-independent approximations of
a a stationary, ergodic source is equal to the rate-distortion function of the
source. Specializing this theorem to irreducible, aperiodic Markoff chains, it
follows that the limit of rate-distortion functions of block-independent approximations of an irreducible, aperiodic Markoff chain which starts in the
stationary distribution is equal to the rate-distortion function of this Markoff
chain. It is known that the rate-distortion function of an irreducible, aperiodic Markoff chain is independent of its initial distribution (follows from
[3]). In this paper, it will be proved that the limit of the normalized ratedistortion functions of block-independent approximations of an irreducible,
aperiodic Markoff chain is independent of its initial distribution. It follows,
then, that the rate-distortion function of an irreducible, aperiodic Markoff
chain and the limit of the normalized rate-distortion functions of its block
independent approximations are equal and these functions are independent
of the initial distribution of the Markoff chain.
Literature on rate-distortion theory is vast. The seminal works are [1] and [4].
A work for rate-distortion theory for random processes is [5]. Much of the
1
2
classical point-to-point literature on rate distortion theory gets subsumed
under the books [2] and [3]. Another reference is [6]. The reader is refered to
these three books and references therein for the literature on rate-distortion
theory. In particular, the reader is referred to [3] because non-stationary
sources are dealt with in great detail in this book, and the concern here is
with a non-stationary process, albeit, a non-stationary Markoff chain. For
understanding Markoff chains, the reader is referred to [7], [8], and [9].
2. Notation and definitions
X and Y denote the source input and source reproduction spaces respectively.
Both are assumed to be finite sets. Asume that X = Y. Assume that the
cardinality of X is greater than or equal to 2. d : X × Y → [0, ∞) is the
single-letter distortion measure. Assume that d(x, x) = 0 ∀x ∈ X and that
d(x, y) > 0 if x 6= y. Denote
(1)
Dmax , max d(x, y), Dmin ,
x∈X,y∈Y
min
{x∈X,y∈Y|d(x,y)>0}
d(x, y)
In what follows, the distortion levels will be assumed to be strictly greater
than 0. For xn ∈ Xn , y n ∈ Yn , the n-letter rate-distortion measure is defined
additively:
(2)
dn (xn , y n ) ,
n
X
d(xn (i), y n (i))
i=1
where xn (i) denotes the ith component of xn and likewise for y n .
Let X1 , X2 , . . ., be a Markoff chain with transition probability matrix P ,
where each Xi is a random-variable on X. For x, x′ ∈ X, pxx′ denotes the
probability that the Markoff chain is in state x′ at time t + 1 given that it
is in state x at time t. pxx′ is independent of t. Assume that the Markoff
chain is irreducible, aperiodic. This implies that it has a stationary distribution, henceforth denoted by π, which will be reserved exclusively for the
stationary distribution. In order to specify the Markoff chain completely, we
need to specify its initial distribution. If X1 ∼ π ′ denote the Markoff chain
(X1 , X2 , . . .) by X[π′ ,P ] . Recall that P is the transition probability matrix
n
of the Markoff chain. X[π′ ,P ] will be called the Markoff X[π′ ,P ] chain. X[π
′ ,P ]
will denote (X1 , X2 , . . . , Xn ).
3
The above mentioned assumptions that X = Y, d(x, x) = 0 and d(x, y) > 0
is x 6= y, that the distortion levels are strictly greater than zero, and that,
the Markoff chain is irreducible, aperiodic, will be made throughtout this
paper and will not be re-stated.
A rate R source-code is a sequence:
n
n
⌊nR⌋ } and f n : {1, 2, . . . , 2⌊nR⌋ } →
< en , f n >∞
1 , where e : X → {1, 2, . . . , 2
Yn .
We say that rate R is achievable for source-coding the Markoff X[π′ ,P ] source
within distortion-level D under the expected distortion criterion if there
exists a rate R source code < en , f n >∞
1 such that
(3)
lim sup E
n→∞
1 n n n n n
d (Xπ′ , f (e (X[π′ ,P ] ))) ≤ D
n
E
(D).
The infimum of all achievable rates is the rate-distortion function RX
[π ′ ,P ]
T
The block-independent approximation (henceforth shortened to BIA) X[π
′ ,P ]
source is a sequence of random vectors (S1 , S2 , . . . , Sn , . . .), where Si are
T
independent, and ∀i, Si ∼ X[π
′ ,P ] . To simplify notation, we will sometimes
n
denote (S1 , S2 , . . .) by S. S will denote (S1 , S2 , . . . , Sn ). Note that BIA
T
X[π
′ ,P ] source is an i.i.d. vector source and will also be called the vector
T
T
i.i.d. X[π
′ ,P ] source. Since the BIA X[π ′ ,P ] source is an i.i.d vector source,
the rate-distortion function for it is defined in exactly the same way as for an
i.i.d. source. The details are as follows: The source input space for the BIA
T
T and the source reproduction space is YT . Denote these
X[π
′ ,P ] source is X
by S and T respectively. A generic point in S is a T -length sequence s. The ith
component of s is denoted by s(i). A generic point in T is a T -length sequence
t. The ith component of t is denoted by t(i). The single letter distortion
P
measure is denoted by d′ and is defined as d′ (s, t) , Tj=1 d(s(j), t(j)). For
sn ∈ Sn , tn ∈ Tn , the n-letter distortion measure d′n is defined additively:
P
d′n (sn , tn ) , ni=1 d′ (sn (i), tn (i)). Note that s can be thought of as either
a scalar in S or a T dimensional vector in XT . With this identification,
d′ = dT and d′n can be thought of as dnT . A rate R source code is a sequence
n
n
⌊nR⌋ } and f n : {1, 2, . . . , 2⌊nR⌋ } →
< en , f n >∞
1 , where e : S → {1, 2, . . . , 2
T
Tn . We say that rate R is achievable for source-coding the BIA X[π
′ ,P ] source
within distortion-level D under the expected distortion criterion if there exist
4
a sequence of rate R source codes < en , f n >∞
1 such that
(4)
1 ′n n n n n
lim sup E
d (S , f (e (S ))) ≤ D
n
n→∞
The infimum of all achievable rates corresponding to a given distortion level
D is the operational rate-distortion function at that distortion level, henceE
(D). The normalized rate-distortion function at
forth denoted by RX
T
′
[π ,P ]
block-length T and distortion level D is defined as
(5)
1 E
R T (T D)
T X[π′ ,P ]
and the limit is
(6)
1 E
R T (T D)
T →∞ T X[π′ ,P ]
lim
E
(D) and (6), and
The theorems in this paper prove the equality of RX
[π ′ ,P ]
′
that these functions do not depend on π . The statements of these theorems
are stated in Section 3. Before that, we carry out a discussion on the ratedistortion function of a non-stationary Markoff chain.
2.1. Discussion
To be entirely correct, the rate-distortion function of a Markoff source should
in
be defined as follows: Let n be the block-length. Denote Ui , X(i−1)n+1
.
n
n
∞
Each Ui is thus, a random vector of length n. Let < e , f >1 be a source
to code the source X[P,π′ ] . When the block length is n, we would like to
use the source-code successively over all intervals of time of block-length n.
Thus, it is more logical to define the distortion as:
1 n
n n
lim sup sup E
(7)
d (Ui , f (e (Ui )))
n
n→∞ i∈N
and correspondingly define the rate-distortion function. This does not end
up making a difference, and hence, we stick to the originally given definition
for distortion. Note that if π ′ = π, the stationary distribution, the sup in
the above definition can be removed since the distribution of X(i−1)n+1 is
independent of i.
5
3. The theorems
E
E
(D) where π is the stationary distribution
(D) = RX
Theorem 1. RX
[π,P ]
[π ′ ,P ]
′
and π is an arbitrary probability distribution on X
Proof. Follows from [3] or see Appendix A for an independent proof tailored
for Markoff chains.
Theorem 2. For D > 0,
(8)
1 E
R T (T D) exists, and is independent of π ′
T →∞ T X[π′ ,P ]
lim
This theorem will be proved in Section 5.
Theorem 3.
(9)
E
E
RX
(D) = RX
(D) = lim
[π,P ]
[π ′ ,P ]
T →∞
1 E
1 E
R T (T D) = lim RX
(T D)
T
[π ′ ,P ]
T →∞ T
T X[π,P ]
where π is the stationary distribution and π ′ is an arbitrary distribution on
X.
Proof. Follows from Theorems 1, 2 and [2], Pages 490-500.
In order to prove Theorem 2, we need more notation and this is the subject
of the next section. The theorem is proved in the section following the next.
4. Further notation
T
The information-theoretic rate-distortion function of the vector i.i.d. X[π
′ ,P ]
source is denoted and defined as
I
RX
T
′
(10)
[π ,P ]
(D) , inf I(X T ; Y T )
W
T
where X T ∼ X[π
′ ,P ] and W is the set of W : S → P(T) defined as
(11)
W, W
X
s∈S,y∈T
T
(s)W (t|s)d′ (s, t) ≤ D
pX[π
′ ,P ]
6
T
T
denotes the distribution corresponding to X[π
where pX[π
′ ,P ] . Note that
′ ,P ]
this is the usual definition of the information-theoretic rate-distortion function for an i.i.d. source; just that the source under consideration is vector
i.i.d.
Let s ∈ S. Denote by Jτ the projection transformation. Jτ (s) , (s(τ +
1), s(τ + 2), . . . s(T )). Fix s. Denote A , {t ∈ S | Jτ (t) = Jτ (s)}. Under the
T
distribution induced by X[π
′ ,P ] , the probability of the set A is
(12)
π ′(τ ) (s(τ + 1))
TY
−1
ps(i)s(i+1)
i=τ +1
for some distribution π ′(τ ) on X which satisfies π ′(τ ) (x) → π(x) as τ → ∞
∀x ∈ X. Note further, that if π ′ = π, π ′(τ ) = π. For x ∈ X, denote π ′(τ ) (x) =
π(x) + δ(τ ) (x) where δ(τ ) (x) → 0 as τ → ∞. δ(τ ) (x) may be negative.
T
T −τ which causes
Denote by Jτ (X[π
′ ,P ] ), the probability distribution on X
the probability of a sequence r ∈ XT −τ to be
(13)
π
′(τ )
(r(1))
TY
−1
pr(i)r(i+1)
i=1
T
T
Note that Jτ (X[π
′ ,P ] ) is the marginal of X[π ′ ,P ] on the last T − τ dimensions.
T
An i.i.d. source can be formed from Jτ (X[π
′ ,P ] ) by taking a sequence of indeT
pendent random vectors, each distributed as Jτ (X[π
′ ,P ] ). This will be called
T
the vector i.i.d. Jτ (X[π′ ,P ] ) source. The rate-distortion function for the vecT
tor i.i.d. Jτ (X[π
′ ,P ] ) source in defined in the same way as the rate-distortion
T
function for the vector i.i.d. X[π
′ ,P ] source: For T − τ length sequences, the
P −τ
single-letter distortion measure is defined as d′′ (p, q) = Ti=1
d(p(i), q(i))
where p ∈ XT −τ , q ∈ YT −τ
.
The
n-letter
rate-distortion
measure
is defined
Pn
′′n
n
n
′′
n
n
n
T
−τ
n
additively: d (p , q ) = i=1 d (p (i), q (i)) where p ∈ (X
) and q n ∈
T
−τ
n
n
n
(Y
) . A sequence of rate R source codes is a sequence < e , f >∞
1 , where
n
e : (XT −τ )n → {1, 2, . . . , 2⌊nR⌋ } and f n : {1, 2, . . . , 2⌊nR⌋ } → (YT −τ )n . The
T
rate-distortion functions for i.i.d. Jτ (X[π
′ ,P ] ) source when the distortion meaT
sure is d” is defined analogously as for the i.i.d. X[π
′ ,P ] vector source; the
details are omitted. Denote the operational rate-distortion function for the
T
E
vector i.i.d. Jτ (X[π
′′ ,P ] ) source by RJ (X T
) (·) and denote the informationτ
′
[π ,P ]
theoretic rate-distortion function for the same source by RJI τ (X T ′
[π ,P ]
) (·).
7
For the same reason as that stated before regarding d′ , d′′ = dT −τ and d′′n
can be thought of as dn(T −τ ) .
5. Proof of the Theorem 2
Before we prove the theorem, note the following:
Lemma 1. Let f : [0, ∞) → [0, ∞) be a convex ∪ non-increasing function.
Let f (0) = K. Let 0 < a < a′ . Then,
f (a) − f (a′ ) ≤
(14)
K ′
(a − a)
a
Proof.
f (a) − f (a′ )
f (0) − f (a)
K ′
(a − a) ⇐=
≤
′
a
a −a
a
a
f (a′ ) − f (a)
f (a) − f (0)
a
⇐=
≥
⇐= 1 − ′ f (0) + ′ f (a′ ) ≥ f (a)
a′ − a
a−0
a
a
⇐= Definition of convexity, see for example [10]
(15)
f (a) − f (a′ ) ≤
This lemma is a direct result of the the definition of convexity and this
observation will be used crucially in the proof of the theorem , which follows
below.
Proof of Theorem 2:
E
Proof. By the rate-distortion theorem, RX
T
′
[π ,P ]
I
(T D) = RX
T
′
(T D). Com-
[π ,P ]
paring definitions with [2], Page 491,
(16)
1 I
R T (T D) (notation in this document) = RT (D) (notation in [2])
T X[π,P ]
By Theorem 9.8.1 in [2], it follows that
(17)
1 E
R T (T D) exists
T →∞ T X[π,P ]
lim
(17) will be used crucially towards the end of the proof.
8
The proof follows three steps:
1. Bound the difference between RJEτ (X T ′
[π ,P ]
2. Relate
RJEτ (X T ′ ) (·)
[π ,P ]
and
) (·)
and RJEτ (X T
[π,P ] )
(·).
E
(·).
RX
T
[π ′ ,P ]
3. Use these relations to prove the desired result by computing various
bounds.
The first step in the proof is to come up with a bound for the difference
between RJEτ (X T ′ ) (·) and RJEτ (X T ) (·). To this end, we first do the same
[π,P ]
[π ,P ]
for RJI τ (X T ′
[π ,P ]
I
) (·) and RJτ (X T
[π,P ]
) (·). To this end, denote the distribution cor-
T
T −τ by Q′ , and the distribution corresponding
responding to Jτ (X[π
′ ,P ] ) on X
T
1
′
to Jτ (X[π,P
] ) by Q. The l distance between Q and Q,
(18)
X
l1 (Q′ , Q) ,
Q′ (xT −τ ) − Q(xT −τ )
xt−τ ∈XT −τ
X
=
|π ′
(τ )
=
|δ
(τ )
T −τ
Y−1
pxt−τ (i)xt−τ (i+1)
i=1
xt−τ ∈Xt−τ
X
(xt−τ (1) − π (τ ) (xt−τ (1)|
(x)|
x∈X
(τ )
,δ
In the above calculation, we have used the fact that if π ′ = π, π ′(τ ) = π.
Condition (Z) stated in [11] holds based on the assumptions we have made,
Lemma 2 in [11] can be applied, and it follows that for τ sufficiently large
(reasoning stated below after a few lines) and any T > τ ,
(19)
1
1
1
RJI τ (X T ′ ) ((T − τ )D) −
RJI τ (X T ) ((T − τ )D) ≤ Kδ(τ ) log (τ )
[π,P
]
[π
,P
]
T −τ
T −τ
δ
where
(20)
K=
1 7d∗
log XT −τ YT −τ
T − τ d˜
9
In (19) , δ(τ ) log δ(τ1 ) is defined as zero if δ(τ ) is zero. |XT −τ | and |YT −τ | denote
the cardinalities of the input and output spaces on which the random source
T
∗ is defined as
Jτ (X[π
′ ,P ] ) is defined. d
(21)
d∗ ,
max
xT −τ ∈XT −τ ,yT −τ ∈YT −τ
d′′ (xT −τ , y T −τ = (T − τ )Dmax
and d˜ is defined as
(22)
d˜ ,
min
{xT −τ ∈XT −τ ,y T −τ ∈YT −τ
|
d′′ (xT −τ ,y T −τ )>0}
d′′ (xT −τ , y T −τ ) = (T − τ )Dmin
It follows that
K=7
(23)
Dmax
(log(|X|) + log(|Y|))
Dmin
Note that K is a constant independent of T, τ, D.
Also, we said above that (19) holds for τ sufficiently large: this is because
by Lemma 2 in [11], we need τ large enough so that
δ(τ ) ≤ 4
(24)
Dmin
Dmax
which is possible considering the fact that δ(τ ) → 0 as τ → ∞, and it is for
this reason that we require τ to be sufficiently large.
Note that the bound in (24) is independent of T, τ . It then follows from (19)
and the equality of information-theoretic and operational rate-distortion
functions for i.i.d. sources, that for τ sufficiently large and any T > τ ,
(25)
1
1
1
RE T
((T − τ )D) −
RE T ((T − τ )D) ≤ Kδ(τ ) log (τ )
T − τ Jτ (X[π′ ,P ] )
T − τ Jτ (X[π,P ] )
δ
The bound (25) will be used crucially later, towards the end of the proof.
E
(·). We will argue the followNext step is to relate RJEτ (X T ′ ) (·) and RX
T
[π ′ ,P ]
[π ,P ]
ing:
(26)
E
RX
T
′
[π ,P ]
((T − τ )D + τ Dmax ) ≤ RJEτ (X T ′
[π ,P ]
) ((T
− τ )D)
10
and
RJEτ (X T ′
(27)
[π ,P ]
) (T D)
E
(T D)
≤ RX
T
′
[π ,P ]
Very rough idea to prove (26) is the following: Given a sequence of rate
T
R source codes for the vector i.i.d. Jτ (X[π
′ ,P ] ) source, we can use the same
T
sequence of rate R source-codes for the vector i.i.d. X[π
′ ,P ] source by not codT
ing the time-slots which were not projected onto when defining Jτ (X[π
′ ,P ] ).
These banished slots will incur a maximum distortion of τ Dmax per symbol
T
of X[π
′ ,P ] . (26) follows. See Appendix B for precise argument.
Very rough idea to prove (27) is the following: Consider a two-dimensional
random vector (A, B) on some space and the i.i.d. source got by taking i.i.d.
copies of (A, B). Consider a distortion measure which is additive over the
two dimensions. Consider, also, the i.i.d. source formed by taking identical
copies of A. Then, for a given distortion level, the rate-distortion function of
the vector i.i.d. (A, B) source is greater than or equal to the rate-distortion
function of the i.i.d. A source. This is stated more rigorously in Appendix B.
T
T
Note that Jτ (X[π,P
] ) is a projection of X[π,P ] onto certain dimensions and
the distortion measure over these dimensions is additive. (27) follows from
this.
Next, we get to Step 3. Assuming T D > τ Dmax , by replacing D in (26) by
D=
(28)
T D − τ Dmax
T −τ
and by (27), it follows that
RJEτ (X T ′
(29)
[π ,P ]
) (T D)
E
(T D) ≤ RJEτ (X T ′
≤ RX
T
′
[π ,P ]
[π ,P ]
) (T D
− τ Dmax )
It follows from (29) by rearranging, that
(30)
E
0 ≤ RX
T
′
[π ,P ]
(T D) − RJEτ (X T ′
[π ,P ]
By noting that RJEτ (X T ′
[π ,P ]
) (T D)
) (D)
≤ RJEτ (X T ′
[π ,P ]
) (T D
− τ Dmax ) − RJEτ (X T ′
[π ,P ]
is a non-increasing, convex ∪ function of D
which is upper bounded by (T − τ ) log |X| at D = 0, it follows, assuming
that T D > τ Dmax , by Lemma 1 that
(31)
RJEτ (X T ′
[π ,P ]
) (T D
− τ Dmax ) − RJEτ (X T ′
[π ,P ]
) (T D)
≤ τ Dmax log |X|
) (T D)
11
From (31) and (30), it follows that
lim
(32)
T →∞
1 E
1 E
R T (T D) − RJτ (X T ′ ) (T D) = 0
[π ,P ]
T X[π′ ,P ]
T
E
Note further, by noting that RX
T
′
(T D) ≤ T log |X|, that
[π ,P ]
(33)
lim
T →∞
1
τ
1 E
(T D) ≤ lim
RE T (T D) − RX
T log |X| = 0
T
[π ′ ,P ]
T →∞ T (T − τ )
T − τ X[π′ ,P ]
T
Also, by noting that RJEτ (X T ′
[π ,P ]
) (D)
is a non-increasing, convex ∪ function
of D which is upper bounded by (T − τ )|X|, it follows by use of Lemma 1
that
(34)
lim
T →∞
1
1
((T − τ )D) ≤
RJEτ (X T ′ ) (T D) −
RE T
[π ,P ]
T −τ
T − τ Jτ (X[π′ ,P ] )
log |X|
lim
τ D → 0 as T → ∞
T →∞ (T − τ )D
It follows, then, from (32), (33), (34) and by noting that
(35)
lim an + lim bn + lim cn = lim (an + bn + cn )
n→∞
n→∞
n→∞
n→∞
if the three limits on the left hand side exist (follows from definitions, see
for example [10]), that
(36)
lim
T →∞
1
1 E
E
((T − τ )D) = 0
R T (T D) −
R
T
T X[π′ ,P ]
T − τ Jτ (X[π′ ,P ] )
From (25) and (36), it follows by the use of triangle inequality, that for τ
sufficiently large and T > τ ,
(37)
1
1
1 E
RX T ′ (T D) −
RJEτ (X T ) ((T − τ )D) ≤ Kδ(τ ) log (τ ) + κτ,T
[π,P ]
[π ,P ]
T
T −τ
δ
for some κτ,T → 0 as T → ∞.
12
The above equation holds for π ′ = π too, that is, for τ sufficiently large and
T > τ,
(38)
1 E
1
1
RX T (T D) −
RJEτ (X T ) ((T − τ )D) ≤ Kδ(τ ) log (τ ) + ητ,T
[π,P ]
[π,P ]
T
T −τ
δ
for some ητ,T → 0 as T → ∞.
From (37) and (38), by use of the triangle inequality, it follows, that for τ
sufficiently large and T > τ ,
(39)
1
1 E
1 E
(T D) ≤ 2Kδ(τ ) log (τ ) + ητ,T + κτ,T
RX T ′ (T D) − RX
T
[π,P ]
[π ,P ]
T
T
δ
From (39) and (17), and by noting that δ(τ ) log δ(τ1 ) → 0 as τ → ∞, ητ,T → 0
as T → ∞, and κτ,T → 0 as T → ∞, it follows that
(40)
lim
T →∞
1 E
R T (T D) exists and is independent of π ′
T X[π′ ,P ]
This finishes the proof
The assumptions X = Y, d(x, x) = 0, d(x, y) > 0 if x 6= y which have
been made are not necessary, and can be replaced by weaker assumptions.
Nothing is lost in terms of idea of the proof by making these assumptions,
and making these assumptions prevents one from thinking of pathological
cases; for these reasons they have been made.
6. ψ-mixing sources or a variant?
A set of sources to which this result may be generalizable with the proof
technique used is ψ-mixing sources or close variants, appropriately defined.
See [12], [13] and [14] for mixing of sources and [13], [14], in particular,
for results on ψ-mixing sources. The main property (among others) that
made ψ-mixing sources amenable to the result in [14] is the decomposition
in Lemma 1 in [14], wherein, a stationary ψ-mixing source is written as a
convex combination of an i.i.d. distribution and another general distribution
13
where the i.i.d. distribution dominates as memory is lost with time. Precisely,
the equation is Equation (19) in [14]:
(41)
t+τ +T
t
′
Pr(Xt+τ
+1 ∈ B|X1 ∈ A) = (1 − λτ )PT (B) + λτ Pt,τ,T,A (B)
where λτ → 0 as τ → ∞. This lemma, though, required stationarity. If a
variant of (41) would hold for non-stationary sources, then, there is a possibility that the result in this paper be generalized to such sources. Irreducible,
aperiodic Markoff chains statisfy this property, with PT (B) taken as the stationary distribution, and P ′ is some distribution depending on the initial
distribution of the Markoff chain. An important bound in proving Theorem
2 in this paper is the l1 distance between Q and Q′ , see (18). This result will
hold for sources which satisfy (41) or a variant. Similarly, proving (26) and
(27) in the proof of Theorem 2 or similar equations may also be possible.
The rest of the proof of Theorem 2 is bounding various differences of ‘close
by’ rate-distortion functions and this may be possible too. This is just an
idea at this point and needs to be studied carefully to see if any of this is at
all possible.
7. Recapitulation and research directions
In this paper, it was proved that the limit of the normalized rate-distortion
functions of block independent approximations of an irreducible, aperiodic
Markoff chain is independent of the initial distribution and is equal to the
rate-distortion function of the Markoff chain.
It would be worthwhile trying to generalize this theorem to ergodic sources
to the extent possible, not necessarily Markoff sources, in particular, to ψmixing sources; this would not only make the result general, but also shed
light on the ‘internal workings’ of rate-distortion theory. Further, it would be
worthwhile trying to prove this result using existing literature, in particular,
see if it follows directly from some result, for example, in [3]; this would help
with generalization and insight into the ‘internal workings’ of rate-distortion
theory, too.
8. Acknowledgements
The author is infinitely grateful to Prof. Robert Gray for ten’s of hours of
his time spent in insightful discussions with the author.
14
References
[1] C. E. Shannon, “Coding theorems for a discrete source with a fidelity
criterion,” Institute of Radio Engineers, National Convention Record,
vol. 7, part 4, pp. 142–163, March 1959.
[2] R. G. Gallager, Information theory and reliable communication. Wiley,
January 1968.
[3] R. M. Gray, Entropy and information theory. Springer-Verlag, February 2011.
[4] C. E. Shannon, “A mathematical theory of communication,” Bell System Technical Journal, vol. 27, pp. 379–423 (Part 1) and pp. 623–656
(Part 2), July (Part 1) and October (Part 2) 1948.
[5] A. N. Kolmogorov, “Theory of transmission of information,” Acad. R.
P. Romine An. Romino-Soviet, vol. 28, no. 1, pp. 5–33, 1959, translated
in American Mathematical Society Translations, Series 2, Volume 33,
1963.
[6] T. Berger, Rate-distortion theory: mathematical basis for data compression, ser. Prentice-Hall series in information and system sciences. Prentice Hall, October 1971.
[7] B. V. Gnedenko, The theory of probability.
pany, New York, N. Y., 1962.
Chelsea Publishing Com-
[8] A. N. Shiryaev, Probability, 2nd ed. Springer, 1984.
[9] W. Feller, An introduction to probability theory and its applications,
Volume 1, 3rd ed. Wiley, 1968.
[10] V. A. Zorich, Mathematical Analysis, I and II. Springer, March 2016.
[11] H. Palaiyanur and A. Sahai, “On the uniform continuity of the ratedistortion function,” in 2008 IEEE International Symposium on Information Theory.
[12] Y. V. Prohorov and Y. A. Rozanov, Probability theory: basic concepts,
limit theorems, random processes, 1st ed., ser. Die Grundlehren der
Mathematischen Wissenschaften in Einzeldarstellungen mit besonderer
Berucksichtigung der Anwendungsgebiete, Band 157. Springer-Verlag,
1969.
[13] R. C. Bradley, “Basic properties of strong mixing conditions. a survey
and some open questions,” Probability surveys, vol. 2, pp. 107–144, 2005.
15
[14] M. Agarwal, S. Mitter, and A. Sahai, “Layered black-box, behavioral interconnection perspective and applications to problems in communications, Part II: sources satisfying ψ-mixing criterion,” Communications
in Information and Systems, vol. 17, no. 4, 2017.
Appendix A. Proof of Theorem 1
Proof. Consider two Markoff chains X[π′ ,P ] = (X1′ , X2′ , . . .) and X[π′′ ,P ] =
(X1′′ , X2′′ , . . .), where π ′ and π ′′ probability distributions on X.
Denote (X1′ , X2′ , . . . , Xn′ ) by X ′ n and (X1′′ , X2′′ , . . . , Xn′′ ) by X ′′ n . Let τ be
an integer. Think of n as large and τ to be much smaller than n. Denote
the distribution
of Xτ′ by µ′ and the
of Xτ′′ by µ′′ . Denote,
P distribution
P
′′
′
′′
′
ǫ =
x∈X |µ (x) − π(x)| , where π is the
x∈X |µ (x) − π(x)| and ǫ =
stationary distribution of the Markoff chain (note that both Markoff chains
have the same transition probability matrix P ). For every ǫ > 0, ∃τǫ∗ such
that ǫ′ < 2ǫ and ǫ′′ < 2ǫ ∀τ ≥ τǫ∗ . Let < en , f n >∞
1 be a source-code. Let
∗
the block-length be n. Think of n large and τǫ << n. Use (en , f n ) to code
K ′ n , (Xτ′ ǫ∗ +1 , Xτ′ ǫ∗ +2 , . . . , Xτ′ ǫ∗ +n ) and K ′′ n , (Xτ′′ǫ∗ +1 , Xτ′′ǫ∗ +2 , . . . , Xτ′′ǫ∗ +n ).
Note that
(42)
1 n ′n n n ′n
1 n ′′ n n n ′ n
d (K , f (e (K ))) − E
d (K , f (e (K ))) ≤ ǫDmax
E
n
n
For δ > 0 (think of δ small), D > 0, let < e′ n , f ′ n >∞
1 be a source-code
E
with rate ≤ R[P,π′ ] (D) + δ to code the X[P,π′ ] source with distortion D.
Construct a source code < e′′ n , f ′′ n >∞
1 to code the X[P,π ′′ ] source as follows. When the block-length is n + τǫ∗ , code X1′′ , X2′′ , . . . , Xτ′′ǫ∗ arbitrarily. and
code (Xτ′′ǫ∗ +1 , Xτ′′ǫ∗ +2 , . . . , Xτ′′ǫ∗ +n ) using (e′ n , f ′ n ). It follows, by calculation of
the distortion achieved for (X1′′ , X2′′ , . . . , Xτ′′∗ +n ) by use of this code, in the
process, using (42)
τǫ∗
E
E
RX[π′′ ,P ] D + ∗
(43)
(D) + δ
Dmax + ǫDmax ≤ RX
[π ′ ,P ]
τǫ + n
ǫ can be made arbitrarily small, τǫ∗ will depend on ǫ and n can be made arbiE
(D + α) ≤
trarily large. It follows that for every α > 0, every δ > 0, RX
[π ′′ ,P ]
E
E
(D) in D, it follows that ∀δ > 0
(D) + δ. By the continuity of RX
RX
[π ′ ,P ]
[π ′ ,P ]
16
E
E
(D)+ δ. It follows, then, since α > 0 can be arbitrarily
(D) ≤ RX
RX
[π ′ ,P ]
[π ′′ ,P ]
E
E
(D). By interchanging π ′ and π ′′ , it follows
(D) ≤ RX
small, that RX
[π ′ ,P ]
[π ′′ ,P ]
E
E
E
E
(D).
(D) = RX
(D) Thus, RX
(D) ≤ RX
that, RX
[π ′′ ,P ]
[π ′ ,P ]
[π ′′ ,P ]
[π ′ ,P ]
Appendix B. Proofs of (26) and (27)
To prove (26):
T
Proof. Let < en , f n >∞
1 be the source code for the i.i.d. vector Jτ (X[π,P ] )
source.
Note that en : (XT −τ )n → {1, 2, . . . , 2⌊nR⌋ } and
f n : {1, 2, . . . , 2⌊nR⌋ } → (YT −τ )n .
T
Let sn ∈ Sn be a realization of S n , the n-blocklength vector i.i.d. X[π,P
]
source which needs to be coded.
sn = (sn (1), sn (2), . . . sn (n)) where each sn (i) ∈ S:
sn (i) = (sn (i)(1), sn (i)(2), . . . , sn (i)(T )).
Recall the projection operator, Jτ (sn (i)) = (sn (i)(τ + 1), . . . , sn (i)(T )).
Denote Jτn (sn ) = (Jτ (sn (1)), Jτ (sn (2)), . . . , Jτ (sn (n))).
Then, Jτn (sn ) is an element of (XT −τ )n . Denote f n (en (Jτn (sn ))) = t′n .
Note that t′n = (t′n (1), t′n (2), . . . , t′n (n)) where
t′n (i) = (t′n (i)(1), t′n (i)(2), . . . , t′n (i)(T − τ )).
Fix a random y ∈ Y. Define the extension transformation,
Eτ (t′n (i)) = (y, y, . . . , y, t′n (i)(1), t′n (i)(2), . . . , t′n (i)(T − τ )), where the initial y’s occur τ times.
Denote Eτn (t′n ) = (Eτ (t′n (1)), Eτ (t′n (2)), . . . , Eτ (t′n (n))).
Note that < en ◦ Jτn , Eτn ◦ f n >∞
1 is a rate R source code to code the i.i.d.
T
′n (sn , E n (f n (en (J n (sn ))))) ≤ d′′n (s′n , t′n ) +
vector X[π,P
source
and
that,
d
τ
τ
]
nτ Dmax . (26) follows.
To prove (27):
17
Proof. Let (A, B) be a random vector on A × B. Let (A1 , B1 ), (A2 , B2 ), . . .
be a sequence where (Ai , Bi ) are independent of each other and (Ai , Bi ) ∼
(A, B). This sequence is the vector i.i.d (A, B) source. A × B is the source
space. Let the source reproduction space be A′ × B′ . d1 : A × A′ → [0, ∞) is
a distortion measure. d2 : B′ × B′ → [0, ∞) is a distortion measure. Assume
that A, A′ , B, B′ are finite sets. Define: d0 ((a, b), (a′ , b′ )) , d1 (a, a′ )+d2 (b, b′ ).
dn1 and dn2 , dn0 are respectively defined additively from d1 , d2 and d0 . We can
then define the rate-distortion functions for the i.i.d. A source and the i.i.d.
E (·) and RE
(A, B) source, denoted, respectively, by RA
(A,B) (·). Then,
(44)
E
E
RA
(D) ≤ R(A,B)
(D)
(44) is proved as follows: Given a code to code the i.i.d. (A, B) source, think
of Bi s as a source of common randomness, and use the obvious variant of
the same code for coding the i.i.d. A source. Since the same code is used,
(44) follows. Existence of a random code with a certain distortion implies
the existence of a deterministic code with the same or lesser distortion. From
this, (44) follows for deterministic codes.
Mukul Agarwal
Department of Electrical and Computer Engineering
Boston University
E-mail address: [email protected]
Received December 7, 2012
| 7cs.IT
|
PROCEEDINGS OF BIOSIGNAL 2010, JULY 14-16, 2010, BERLIN, GERMANY
1
Pituitary Adenoma Segmentation
J. Egger, M. H. A. Bauer, D. Kuhnt, B. Freisleben and Ch. Nimsky
Abstract—Sellar tumors are approximately 10-15% among all
intracranial neoplasms. The most common sellar lesion is the
pituitary adenoma. Manual segmentation is a time-consuming
process that can be shortened by using adequate algorithms. In
this contribution, we present a segmentation method for pituitary
adenoma. The method is based on an algorithm we developed
recently in previous work where the novel segmentation scheme
was successfully used for segmentation of glioblastoma
multiforme and provided an average Dice Similarity Coefficient
(DSC) of 77%. This scheme is used for automatic adenoma
segmentation. In our experimental evaluation, neurosurgeons
with strong experiences in the treatment of pituitary adenoma
performed manual slice-by-slice segmentation of 10 magnetic
resonance imaging (MRI) cases. Afterwards, the segmentations
were compared with the segmentation results of the proposed
method via the DSC. The average DSC for all data sets was
77.49%±4.52%. Compared with a manual segmentation that
took, on the average, 3.91±0.54 minutes, the overall segmentation
in our implementation required less than 4 seconds.
Index Terms—pituitary adenoma, segmentation, spherical
graph, MRI, mincut
I. INTRODUCTION
Approximately 10-15% of all intracranial neoplasms are
sellar tumors. The most common sellar lesion is the pituitary
adenoma [1]. The lesions can be classified according to size or
hormone-secretion (hormone-active and horme-inactive).
Microadenomas are less than 1 cm in diameter, whereas
macroadenomas measure more than 1 cm. The rare giantadenomas have more than 4 cm in diametric size. Secreted
hormones can be cortisol (cushing's disease), human growthhormone (hGH; acromegaly), follicle-stimulating hormone
(FSH), luteinising hormone (LH), thyroid-stimulating hormone
(TSH), prolactine or a combination of these. Only for the
prolactine-expressing tumors, a pharmacological treatment is
the initial treatment of choice in form of dopamine-agonists.
Treatment is most commonly followed by a decrease of
prolactine-levels and tumor volume. For acromegaly and
cushing's disease, surgery remains the first-line treatment,
although somatostatin receptor analogues or combined
dopamine/somatostatin receptor analogues are a useful secondline therapeutical option for hGH-expressing tumors. Current
medical therapies for Cushing's disease primarily focus on
adrenal blockade of cortisol production, although pasireotide
J.E., M.H.A.B., D.K., Ch.N. are with the Department of Neurosurgery,
University of Marburg, Marburg, Germany e-mail: [email protected]
J.E., M.H.A.B., B.F. are with the Department of Mathematics and
Computer Science, University of Marburg, Marburg, Germany
and cabergoline show promise as pituitary-directed medical
therapy for Cushing's disease [2].
Thus, not only for the most hormone-active, but also for
homone-inactive macroadenomas with mass-effect, surgery is
the treatment of choice, most possibly via a transsphenoidal
approach [3]. By contrast, for hormone-inactive
mircroadenomas (<1cm) there is no need for immediate
surgical resection. The follow-up contains endocrine and
ophthalmological evaluation as well as magnetic resonance
imaging (MRI). In case of continuous tumor volume progress,
microsurgical excision becomes the treatment of choice. Thus,
the tumor volume should be rigidly registered over the time of
follow-up.
In this contribution, we present a segmentation method for
pituitary adenoma. The method is based on an algorithm we
developed recently in a previous work where the novel
segmentation scheme was successfully used for segmentation
of glioblastoma multiforme and provided an average Dice
Similarity Coefficient (DSC) of 77%. For automatic and
adequate adenoma segmentation, the original scheme is used,
creating a directed 3D-graph within two steps: sending rays
through the surface points of a polyhedron and sampling the
graph’s nodes along every ray. The center of the polyhedron is
user-defined and located inside the adenoma. Then, the
minimal cost closed set on the graph is computed via a
polynomial time s-t-cut, creating an optimal segmentation of
the adenoma’s boundary and volume. For evaluation,
neurosurgeons with strong experiences in the treatment of
pituitary adenoma performed manual slice-by-slice
segmentation of 10 cases. Afterwards, the segmentations were
compared with the segmentation results of the proposed
method via the Dice Similarity Coefficient (DSC). The
average DSC for all data sets was 77.49%±4.52%. Compared
with a manual segmentation that took, on the average,
3.91±0.54 minutes, the overall segmentation in our
implementation required less than 4 seconds.
The paper is organized as follows. Section 3 presents the
details of the proposed approach. In Section 4, experimental
results are discussed. Section 5 concludes the paper and
outlines areas for future work.
II. RELATED WORK
Based on magnetic resonance imaging, several algorithms
have already been proposed for glioma segmentation. An
extensive overview of some deterministic and statistical
approaches is given by Angelini et al. [4]. The majority of
them are region-based approaches; more recent ones are based
on deformable models and include edge-information.
PROCEEDINGS OF BIOSIGNAL 2010, JULY 14-16, 2010, BERLIN, GERMANY
Neubauer et al. [5] and Wolfsberger et al. [6] introduce
STEPS, a virtual endoscopy system designed to aid surgeons
performing pituitary surgery. STEPS uses a semi-automatic
segmentation method that is based on the so-called watershedfrom-markers technique. The watershed-from-markers
technique uses user-defined markers in the object of interest
and the background. A memory efficient and fast
implementation of the watershed-from-markers algorithm –
also extended to 3D – has been developed by Felkel et al. [7].
Descoteau et al. [8] have proposed a novel multi-scale sheet
enhancement measure and apply it to paranasal sinus bone
segmentation. The measure has the essential properties to be
incorporated in the computation of anatomical models for the
simulation of pituitary surgery.
III. METHODS
Our overall method starts by setting up a directed 3D graph
from a user-defined seed point that is located inside the
pituitary adenoma. To set up the graph, the method samples
along rays that are sent through the surface points of a
polyhedron with the seed point as the center. The sampled
points are the nodes n ∈ V of the graph G(V,E) and e ∈ E is the
corresponding set of edges. There are edges between the nodes
and edges that connect the nodes to a source s and a sink t to
allow the computation of an s-t cut (the source and the sink s, t
∈ V are virtual nodes).
The idea of setting up the graph with a polyhedron goes
back to a catheter simulation algorithm where several
polyhedra were used to align the catheter inside the vessel [9].
In the segmentation scheme, this idea is combined with a
graph-based method that has been introduced for the semiautomatic segmentation of the aorta [10], [11], [12] and
diffusion tensor imaging (DTI) fiber bundle segmentation [13].
However, in this case, setting up the graph was performed by
sampling the nodes in several 2D planes and therefore is not
useful for the segmentation of spherical or elliptical 3D objects
(see Figure 1).
2
In a recent publication [14], the introduced approach for
segmenting spherical or elliptical 3D objects was enhanced by
an arbitrary number of additional seed points to a semiautomatic method. Thereby, the algorithm is supported with
grey value information and geometrical constraints that are
used to segment World Health Organization (WHO) grade IV
gliomas.
The arcs <vi, vj> ∈ E of the graph G connect two nodes vi,
vj. There are two types of ∞-weighted arcs: z-arcs Az and r-arcs
Ar (Z is the number of sampled points along one ray z=(0,…,Z1) and R is the number of rays sent out to the surface points of
a polyhedron r=(0,…,R-1)), where node V(xn,yn,zn) is one
neighbor of node V(x,y,z).
The arcs Az between two nodes along a ray ensure that all
nodes below the polyhedron surface in the graph are included
to form a closed set (correspondingly, the interior of the
spherical object is separated from the exterior in the data). The
arcs Ar between the nodes of different rays constrain the set of
possible segmentations and enforce smoothness via the
parameter ∆r. The larger this parameter is, the larger the
number of possible segmentations is (see Figure 2).
Fig. 2. Principle of a cut of edges between two rays for ∆r=1. Left and middle:
Same cost for a cut (2·∞). Right: Higher cost for a cut (4·∞).
After graph construction, the minimal cost closed set on the
graph is computed via a polynomial time s-t cut [15]. A
Markov Random Field (MRF) approach where each voxel of
the image is a node is definitely too time-consuming for the
data we used (512x512xX). A MRF approach in a recent
publication needed already several minutes for the cut of one
small 2D image [16]. We also thought about using an Active
Contour Method (ACM) [17] and [18] approach where the
initial contour is a polyhedron with an initial radius definitely
smaller than the object (pituitary adenoma). However, ACMs
can get stuck in local minima during the iterative segmentation
(expansion) process. In contrast, a graph cut approach
provides an optimal segmentation for the constructed graph.
IV. RESULTS
Fig. 1. Upper image: unfolding several 2D slices to segment a cylindrical
object. Lower image: unfolding a polyhedron to segment a spherical object.
Left: principal of sending rays through the surface of a polyhedron with 32
vertices. Right: Sampling nodes for the graph along the rays (polyhedron with
92 vertices).
The presented methods were implemented in C++ within the
MeVisLab platform [19]. Using 2432 and 7292 polyhedra
surface points, the overall segmentation (sending rays, graph
construction and mincut computation) in our implementation
took less that 4 seconds on an Intel Core i5-750 CPU, 4x2.66
GHz, 8 GB RAM, Windows XP Professional x64 Version,
Version 2003, Service Pack 2. The ray length is a fixed
parameter (5 cm), determined via the largest pituitary adenoma
of the 10 cases (all pituitary adenoma had a diameter of less
that 5 cm).
PROCEEDINGS OF BIOSIGNAL 2010, JULY 14-16, 2010, BERLIN, GERMANY
3
Fig. 3. Segmentation results for a pituitary adenoma data set with user-defined seed point (blue).
To evaluate the approach, neurological surgeons with
several years of experience in resection of brain tumors
performed manual slice-by-slice segmentation of 10 pituitary
adenomas. Afterwards, the manual segmentations were
compared with the one click segmentation results of the
proposed method via the Dice Similarity Coefficient (DSC)
[20]. The Dice Similarity Coefficient is the relative volume
overlap between A and R, where A and R are the binary masks
from the automatic (A) and the reference (R) segmentation.
V(•) is the volume (in cm3) of voxels inside the binary mask,
by means of counting the number of voxels, then multiplying
with the voxel size:
DSC =
2 ⋅ V ( A ∩ R)
V ( A) + V ( R)
(1)
In Figure 3, several axial slices of a pituitary adenoma are
shown. In every slice, the result of the automatic segmentation
is drawn (red). Additionally, the user-defined seed point is
shown in the middle slice (blue). The average Dice Similarity
Coefficient for all data sets was 77.49%±4.52% (see Table I
for details). Figure 4 shows an axial slice of a pituitary
adenoma (left) and a 3D mask of an automatic segmented
pituitary adenoma (right). Figure 5 shows different views of
sagittal slices with segmented pituitary adenoma performed
with the proposed algorithm.
Fig. 4. Axial slice of a pituitary adenoma (left). 3D mask of an automatic
segmented pituitary adenoma (right).
Fig. 5. Different views of sagittal slices with segmented pituitary adenoma.
V. CONCLUSION
In this paper, a method for pituitary adenoma segmentation
was presented. The method is based on an algorithm we
developed recently in a previous work where the novel
segmentation scheme was successfully used for segmenting
glioblastoma multiforme and provided an average Dice
Similarity Coefficient (DSC) of 77%. For automatic and
adequate adenoma segmentation, the original scheme was
used, creating a directed 3D-graph within two steps: sending
rays through the surface points of a polyhedron and sampling
the graph’s nodes along every ray. The center of the
polyhedron is user-defined and located inside the adenoma.
Then, the minimal cost closed set on the graph is computed via
a polynomial time s-t-cut, creating an optimal segmentation of
the adenoma’s boundary and volume.
For evaluation, neurosurgeons with strong experiences in
TABLE I
SUMMARY OF RESULTS: MIN., MAX., MEAN AND STANDARD DEVIATION FOR 10 PITUITARY ADENOMAS.
min
max
µ ±σ
Volume of tumor (cm3)
manual
algorithm
0.84
1.18
15.57
14.94
6.30 ± 4.07
6.22 ± 4.08
Number of voxels
manual
algorithm
4492
3461
106151
101902
47462.7
47700.6
DSC (%)
manual segmentation
time (min)
71.07
84.67
3
5
77.49 ± 4.52
3.91 ± 0.54
PROCEEDINGS OF BIOSIGNAL 2010, JULY 14-16, 2010, BERLIN, GERMANY
the treatment of pituitary adenoma performed manual slice-byslice segmentation of 10 cases. Afterwards, the segmentations
were compared with the segmentation results of the proposed
method via the Dice Similarity Coefficient. The average Dice
Similarity Coefficient for all data sets was 77.49%±4.52%.
Compared with a manual segmentation that took, on the
average, 3.91±0.54 minutes, the overall segmentation in our
implementation needed less than 4 seconds.
There are several areas of future work. For example, we
want to enhance the presented segmentation scheme with
statistical information about shape – e.g. Active Shape Models
(ASM) [21] – ort texture – e.g. Active Blobs [22] – or shape
and texture – e.g. Active Appearance Models [23] and [24] –
of the desired object. Furthermore, we want use the approach
to segment spherically or elliptically shaped organs like
kidneys [25] and pathologies like cerebral aneurysms [26] and
[27].
VI. ACKNOWLEDGEMENTS
The authors would like to thank Fraunhofer MeVis in Bremen,
Germany, for their collaboration and especially Horst K. Hahn
for his support.
REFERENCES
M.- A. Weber, S. Zoubaa, M. Politi, I. Grunwald, and M. Essig, Sellar
tumors (in German), Der Radiologe; 47(6):492-500. Springer Berlin /
Heidelberg, 2007.
[2] B. M. K. Biller, A. Colao, S. Petersenn, V. S. Bonert, and M. Boscaro,
Prolactinomas, Cushing's disease and acromegaly: debating the role of
medical therapy for secretory pituitary adenomas, BMC Endocrine
Disorders; 10(1):10, 2010.
[3] M. Buchfelder and S. Schlaffer, Surgical treatment of pituitary tumours,
Best Practice & Research Clinical Endocrinology & Metabolism;
23(5):677-92, 2009.
[4] E. D. Angelini, O. Clatz, E. Mandonnet, E. Konukoglu, L. Capelle, and
H. Duffau, Glioma Dynamics and Computational Models: A Review of
Segmentation, Registration, and In Silico Growth Algorithms and their
Clinical Applications, Current Medical Imaging Reviews, 3, pp. 262276, 2007.
[5] A. Neubauer, S. Wolfsberger, M.-T. Forster, L. Mroz, R. Wegenkittl,
and K. Bühler, Advanced virtual endoscopic pituitary surgery, IEEE
Transactions on Visualization and Computer Graphics (TVCG), 11(5);
pp. 497-507, 2005.
[6] S. Wolfsberger, A. Neubauer, K. Bühler, R. Wegenkittl, T. Czech, S.
Gentzsch, H.-G. Böcher-Schwarz, and E. Knosp, Advanced Virtual
Endoscopy for Endoscopic Transsphenoidal Pituitary Surgery,
Neurosurgery 59(5), pp. 1001-1010, 2006.
[7] P. Felkel, R. Wegenkittl, and M. Bruckschwaiger, Implementation and
Complexity of the Watershed-from-Markers Algorithm Computed as a
Minimal Cost Forest, In: Proceedings of Eurographics ’01, pp. 26-35,
2001.
[8] M. Descoteaux, M. Audette, K. Chinzei, and K. Siddiqi, Bone
enhancement filtering: application to sinus bone segmentation and
simulation of pituitary surgery, Computer Aided Surgery, 11(5):247-55,
2006.
[9] J. Egger, Z. Mostarkic, S. Großkopf, and B. Freisleben, A Fast Vessel
Centerline Extraction Algorithm for Catheter Simulation, In:
Proceedings of the 20th IEEE International Symposium on ComputerBased Medical Systems, Maribor, Slovenia, pp. 177-182, IEEE Press,
2007.
[10] J. Egger, Pre- and postoperative segmentation and virtual stenting of
aneurysms and stenosis (in German), Dissertation, Department of
4
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[1]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
[27]
Mathematics and Computer Science, University of Marburg, 215 pages,
Marburg, Germany, 2009.
J. Egger, T. O’Donnell, C. Hopfgartner, and B. Freisleben, Graph-Based
Tracking Method for Aortic Thrombus Segmentation, In: Proceedings of
the 4th European Congress for Medical and Biomedical Engineering,
Engineering for Health, Antwerp, Belgium, Springer-Press, 584-587,
2008.
J. Egger, B. Freisleben, R. Setser, R. Renapuraar, C. Biermann, and T.
O’Donnell, Aorta Segmentation for Stent Simulation, In: 12th
International Conference on Medical Image Computing and Computer
Assisted Intervention (MICCAI), Cardiovascular Interventional Imaging
and Biophysical Modelling Workshop, 10 pages, London, United
Kingdom (UK), 2009.
M. H. A. Bauer, J. Egger, T. O’Donnell, B. Freisleben, S. Barbieri, J.
Klein, H. K. Hahn, and Ch. Nimsky, A Fast and Robust Graph-based
Approach for Boundary Estimation of Fiber Bundles Relying on
Fractional Anisotropy Maps, In: Proceedings of the 20th International
Conference on Pattern Recognition (ICPR), Istanbul, Turkey, IEEE
Computer Society, 2010.
J. Egger, M. H. A. Bauer, D. Kuhnt, C. Kappus, B. Carl, B. Freisleben
and Ch. Nimsky, Flexible Semi-Automatic Approach for Glioblastoma
multiforme Segmentation,
In: Proceedings of the International
Biosignal Processing Conference, DGBMT, IEEE, Berlin, Germany,
2010.
Y. Boykov and V. Kolmogorov, An Experimental Comparison of MinCut/Max-Flow Algorithms for Energy Minimization in Vision, IEEE
Transactions on Pattern Analysis and Machine Intelligence, 26(9), pp.
1124-1137, 2004.
A. Shabou, F. Tupin, and J. Darbon, A graph-cut based algorithm for
approximate MRF optimization, International Conference on Image
Processing (ICIP), pp. 2413-2416, 2009.
M. Kass, A. Witkin, and D. Terzopoulos, Constraints on Deformable
Models: Recovering 3D Shape and Nongrid Motion, Artificial
Intelligence, 36:91-123, 1988.
D. Terzopoulos, A. Witkin, and M. Kass, Constraints on deformable
models: Recovering 3D shape and nongrid motion, Artificial
Intelligence, 36: 91-123, 1988.
MeVisLab – development environment for medical image processing
and visualization. MeVis Medical Solutions AG and Fraunhofer
MEVIS, Bremen, Germany.
http://www.mevislab.de
K. H. Zou, S. K. Warfield, A. Bharatha, C. M. C. Tempany, M. R. Kaus,
S. J. Haker, W. M. Wells, F. A. Jolesz, and R. Kikinis, Statistical
Validation of Image Segmentation Quality Based on a Spatial Overlap
Index: Scientific Reports, Academic Radiology, 11(2), pp. 178-189,
2004.
T. F. Cootes and C. J. Taylor, Active Shape Models - ‘Smart Snakes’,
In: Proceedings of the British Machine Vision Conference, pages 266275, 1992.
S. Sclaroff and J. Isidoro, Active Blobs, In: Proceedings of the Sixth
International Conference on Computer Vision, IEEE Computer Society,
pp. 1146-1153, Washington, DC, USA, 1998.
T. F. Cootes, G. J. Edwards, and C. J. Taylor, Active Appearance
Models, In: Proceedings of the European Conference on Computer
Vision (ECCV), volume 2, pp. 484-498. Springer-Press Berlin /
Heidelberg, 1998.
K. Greiner, J. Egger, S. Großkopf, J. N. Kaftan, R. Dörner, and B.
Freisleben, Segmentation of Aortic Aneurysms in CTA-images with the
Statistic Method of the Active Appearance Models (in German),
Proceedings of Bildverarbeitung für die Medizin (BVM), Berlin,
Germany, Springer-Press, 2008.
M. Spiegel, D. A. Hahn, V. Daum, J. Wasza, and J. Hornegger,
Segmentation of kidneys using a new active shape model generation
technique based on non-rigid image registration, Computerized Medical
Imaging and Graphics 33, pp. 29-39, 2009.
B. Ma, R. E. Harbaugh, and M. L. Raghavan, Three-Dimensional
Geometrical Characterization of Cerebral Aneurysms. Annals of
Biomedical Engineering. 32(2): 264-273, 2004.
D.-Y. Kim and J.-W. Park, Computer-Aided Detection of Kidney Tumor
on Abdominal Computed Tomography Scans, Acta Radiologica, 45(7):
791-795, 2004.
| 5cs.CE
|
1
Optimal Control of Large-Scale Networks using
Clustering Based Projections
arXiv:1609.05265v3 [cs.SY] 4 Oct 2017
Nan Xue, Student Member, IEEE, Aranya Chakrabortty, Senior Member, IEEE
Abstract—In this paper we present a set of projection-based designs for constructing simplified linear quadratic regulator (LQR)
controllers for large-scale network systems. When such systems
have tens of thousands of states, the design of conventional
LQR controllers becomes numerically challenging, and their
implementation requires a large number of communication links.
Our proposed algorithms bypass these difficulties by clustering
the system states using structural properties of its closed-loop
transfer matrix. The assignment of clusters is defined through
a structured projection matrix P , which leads to a significantly
lower-dimensional LQR design. The reduced-order controller is
finally projected back to the original coordinates via an inverse
projection. The problem is, therefore, posed as a model matching
problem of finding the optimal set of clusters or P that minimizes
the H2 -norm of the error between the transfer matrix of the
full-order network with the full-order LQR and that with the
projected LQR. We derive a tractable relaxation for this model
matching problem, and design a P that solves the relaxation. The
design is shown to be implementable by a convenient, hierarchical
two-layer control architecture, requiring far less number of
communication links than full-order LQR.
Index Terms—Clustering, Large-scale networks, Projection,
LQR, H2 performance.
I. I NTRODUCTION
A vast majority of practical networked dynamic systems
(NDS), ranging from power system networks to wireless
networks to social or biological networks, consist of several
hundreds to thousands of nodes that are spatially distributed
over wide geographical spans. Developing tractable control
designs for such large complex networks, and implementing
those designs through affordable communication, continue to
be a challenge for network designers. Traditionally, control
theorists have addressed the problem of controlling largedimensional systems by imposing structure on controllers. The
most promising approach, for example, started with the idea
of decentralized control [1], followed by techniques such as
singular perturbation theory [2], [3], balanced truncation [4],
[5], [6], and ν-gap reduction [7] among others. These methods
aim to simplify the design of controllers for large systems by
exploiting weak coupling between their state variables, and by
ignoring states that are ‘less important’ than others. The tradeoff, however, is that the resulting controllers are often agnostic
of the natural coupling between the states, especially the
coupling between the closed-loop states, since many of these
couplings are forcibly eliminated to facilitate the design itself.
N. Xue and A. Chakrabortty are with the Department of Electrical and
Computer Engineering, North Carolina State University, Raleigh, NC, 27695
USA, e-mail: [email protected], [email protected]
The work is supported partly by the US National Science Foundation (NSF)
under grant ECCS 1054394.
Therefore, extending these methods to facilitate controller
designs for networks, especially to NDS whose states may be
defined over highly structured topologies such as clustering, is
quite difficult. The literature for developing tangible and yet
simple low-dimensional controllers that satisfy global stability
and dynamic performance requirements of very large NDS
is still unfortunately sparse. Ideas on aggregate control [8],
glocal control [9] and hierarchical control [10], [11] have
recently been proposed to address this challenge. The goal of
these designs, however, is to guarantee stability by modular
tuning of local controller gains; their degrees of freedom for
guaranteeing performance can be limited.
To bridge this gap, in this paper we propose a design
method called control inversion. The approach is to cluster
the states of an n-dimensional network into r > 0 distinct,
non-overlapping groups. We assume n to be a large positive
integer, and r ≤ n to be a given design parameter. The
grouping is defined by a (r × n) structured projection matrix
P whose elements denote the identity of states in the clusters,
weighted by certain projection weights. The design thereafter
consists of three steps. First, for the full-order network an ndimensional LQR controller is defined for any given choice of
Q and R. We refer to this controller as the benchmark LQR.
Second, the projection matrix P is used to construct an rdimensional reference model for which an r-dimensional LQR
controller is designed. The design matrices for this reducedorder controller, however, are not free; they are constrained
by being related to Q and R through P . The important point,
however, is that the design dimension reduces to r from n.
Finally, this reduced-order controller is projected back to the
full-order network by the inverse projection P T . The problem
is then to find a projection matrix P that minimizes the H2 norm of the error between the transfer function matrices of
the full-order network with the benchmark LQR controller and
with the projected LQR controller.
This problem by itself, however, is non-convex even without
any structural constraint on P . To bypass this intractability, we
introduce a relaxation, which is done in two stages. The first
stage relaxes the error minimization to the minimization of
its upper bound, while the second stage applies a low-rank
approximation. We finally design a P that solves this relaxed
problem. Three distinct variants of the design are proposed.
In the first case, we optimize over cluster assignment while
keeping the projection weights fixed, and establish that this
minimization can be posed as an unsupervised clustering problem. We use weighted k-means [12] to solve this minimization.
In the second case, we fix the cluster identities, and optimize
over the projection weights. Depending on the stability of
2
the open-loop system, we show that this minimization can be
posed as finding the dominant eigenvector of the controllability
Gramian or as finding the Z-eigenvector of a tensor [13]. In
the third case, we propose an iterative method to optimize
over both cluster assignment and projection weights. The
controllers resulting from all three algorithms are shown to be
implementable by a convenient, hierarchical two-layer control
architecture, requiring far less number of communication links
than full-order LQR as well as sparsity-promoting LQR [14].
Recently, [15] and [16] have used structural projectionbased ideas for model reduction of large networks, but not
for control designs. Attention has also been drawn to designing LQR controllers for large systems by finding low-rank
solutions of algebraic Riccati equations [17]. However, like
most Krylov subspace-based reduction methods such as in
[6], the controller in [17] is unstructured, and hence demands
as many communication links as the full-order LQR itself.
Distributed controllers using model matching [18], sparsitypromoting LQR in [14] and structured LQR in [19], [20]
promise to reduce the communication density, but their designs
inherit the same dimensionality as the full-order design. Unlike
all of these methods, the novelty of our algorithms is in the
facilitation of closed-loop control from the perspective of both
design and implementation. The recent papers [21], [22] also
address both goals, but the dimensionality of their controllers
is subject to the sparsity structure of the open-loop network
while our design does not necessarily require any such sparsity.
Preliminary results on this design have been presented in the
recent conference paper [23], but only for a consensus model
with specific Q and R matrices.
The remainder of the paper is organized as follows. Section
II formulates the problem of clustering-based optimal control.
The relaxation for the original problem is derived in Section
III, which is then solved by a clustering algorithm based on
weighted k-means optimization in Section IV. The design for
cluster weights as well as that for the weights and clusters
taken together are discussed in Section V. All three algorithms
are illustrated via simulations in Section VI. Section VII
concludes the paper.
Notation We will use the following notations throughout
this paper:
|m|
absolute value of a scalar m
|S|c
cardinality of a set S
1n
column vector of size n with all 1 entries
Ik
identity matrix of size k
Mi,j
the (i, j)th entry of a matrix M
diag(m)
diagonal matrix with vector m on its
principal diagonal
M ⊗N
Kronecker product of M and N
M ◦N
Hadamard product of M and N
tr(M )
trace operation on a matrix M
kM kF
Frobenius pnorm of a matrix M , i.e.
kM kF = tr(M M T )
ker(M )
kernel of a matrix M
σ̄(M ), λ̄(M ) largest singular value, or eigenvalue with
largest real part of a matrix M
σ(M ), λ(M )
smallest singular value, or eigenvalue
with smallest real part of a matrix M
v̄(M )
right eigenvector of λ̄(M )
Given a matrix M = [m1 , ..., mn ] ∈ Rn×n , its vector form is defined by vec(M ) = [mT1 , ..., mTn ]T , with
the inverse operation defined by unvec(vec(M )) = M . A
−1
transfer matrix is defined as g(s) = C(sI
− A) B + D,
A B
. We rewith a realization form of g(s) =
C D
fer to g(s) as stable if A is Hurwitz, and unstable otherwise. Furthermore, the H2 and H∞ norms of a stable
qR transfer matrix g(s)qare R defined by kg(s)kH2 =
∞
∞
1
∗
∗
−∞ tr[g (t)g(t)]dt =
2π −∞ tr[g (jω)g(jω)]dω and
kg(s)kH∞ = supω σ̄[g(jω)].
From graph theory, a graph G = (V, E) is defined over a
node (vertex) set V = {1, ..., n} and an edge set E ⊂ V × V,
which contains two-element subsets of V. If {i, j} ∈ E, we
call nodes i and j adjacent, and denote the relation by i ∼ j,
or simply ij. The set of nodes adjacent to i ∈ V is noted
by Ni = {j ∈ V|i ∼ j}. In this paper, G is assumed to be
undirected, which implies ij is equivalent to ji, and there are
no loops or multiple edges between nodes.
II. P ROBLEM F ORMULATION
Consider a general LTI system of the form
(
ẋ(t) = Ax(t) + Bu(t) + Bd d(t), x(0) = x0
y(t) = Cx(t)
,
(1)
where x(t) ∈ Rn , u(t) ∈ Rm and y(t) ∈ Rp represent the
vector of state, control and output variables respectively, and
d(t) ∈ Rnb is a disturbance entering into the system. We
assume (1) to be defined over a network of ns ≤ n interconnected subsystems, with their network topology represented by
a connected graph G = (V, E), V = {1, ..., ns }. The dynamics
of each subsystem can be written as
P
ẋi (t)=Aii xi (t)+
Aij xj (t)+Bi ui (t)+Bdi di (t)
j∈Ni
(2)
y (t) = C x (t)
i
i i
i = 1, ..., ns , where Aij , Bi , Bdi and Ci are submatrices
with compatible dimensions from A, B = diag(B1 , ..., Bns ),
Bd = diag(Bd1 , ..., Bdns ) and C = diag(C1 , ..., Cns ). Notice
that the dimension of ui (t) can be zero, meaning that the ith
subsystem can have no input.
In this paper, we consider an LQR design for (1), and
assume C = In for full-state feedback. Given two real-valued
matrices Q = QT 0 and R = RT ≻ 0, the LQR problem
is posed as finding a feedback law u(t) = −Kx(t) such that
the cost function
Z ∞
J :=
[xT (t)Qx(t) + uT (t)Ru(t)]dt
(3)
0
is minimized. The expression (3), also known as the infinitehorizon continuous-time LQR, can be solved by the following
algebraic Riccati equation (ARE)
AT X + XA + Q − XGX = 0,
(4)
3
where G = BR−1 B T . The feedback matrix can be found
through K = R−1 B T X. For such a solution X to exist, we
will adhere to the following assumption throughout this paper.
T
1
Assumption 2.1: (Q 2 , A) is observable, and (A, BR− 2 ) is
stabilizable.
According to [7], the assumption above guarantees a unique
stabilizing solution X = X T ≻ 0. However, finding this
solution from (4) in practice is subject to O(n3 ) computational complexity, which can become unscalable for largescale systems. Moreover, the resulting matrix X is usually an
unstructured dense matrix, which demands every subsystem in
the network to communicate with every other subsystem for
implementing the feedback. These two factors together make
both the design and implementation of u = Kx very difficult,
especially when G consists of thousands to tens of thousands
of nodes. Therefore, we propose a design strategy, which we
refer to as control inversion, to repose this LQR problem using
a clustering-based projection.
A. Control Inversion
Definition 2.2: Given an integer r, where 0 < r ≤ n, and
a non-zero vector w ∈ Rn , define r non-empty, distinct, and
non-overlapping subsets of the state index set Vs = {1, ..., n},
respectively denoted as I = {I1 , ..., Ir }, such that I1 ∪ ... ∪
Ir = Vs . A clustering-based projection matrix P ∈ Rr×n is
defined as
( w
j
j ∈ Ii
,
(5)
Pi,j := kwIi k2
0
otherwise
where wIi = [wIi {1} , ..., wIi {|Ii |c } ]T is non-zero, and Ii {j}
denotes the j th element in the set Ii . The matrix P has the
following three properties:
2) Reduced-order LQR design: We similarly project the
LQR parameters by Q̃ = P QP T ∈ Rr×r , and let R̃ = R
such that G̃ := B̃ R̃−1 B̃ = P GP T ∈ Rr×r . An LQR problem
for the reduced-order model (6) is then posed as to minimize
Z ∞
[x̃T (t)Q̃x̃(t) + ũT (t)R̃ũ(t)]dt
(7)
J˜ :=
0
with respect to ũ(t) = −K̃ x̃(t). Here the feedback matrix
K̃ = R̃−1 B̃ T X̃ corresponds to the solution X̃ ∈ Rr×r of the
reduced-order ARE of (7), which is written as
ÃT X̃ + X̃ Ã + Q̃ − X̃ G̃X̃ = 0.
(8)
3) Inverse projection to original coordinates: The solution
X̃ from (8) is projected back to the original coordinates
through the inverse projection
X̂ = P T X̃P.
(9)
This projected controller can then be implemented in the fullorder model (1) using u = −R−1 B T X̂x, which implies that
the effective feedback gain matrix is
K̂ = R−1 B T X̂.
(10)
B. Problem Statement
The controller K̂ is dependent on the projection P through
equations (6), (7), and (9). The choice of P is guided in the
following way. Consider
g(s) := (sIn − A + BK)−1 Bd ,
(11)
which is the closed-loop transfer matrix from d to x for (1)
with full-order LQR. Similarly, consider
ĝ(s) := (sIn − A + B K̂)−1 Bd ,
(12)
which is the closed-loop transfer matrix from d to x for (1)
with the projected controller (10). Using (11) and (12), we
•
next state our main problem of interest.
•
Main problem: Given system (1) and an integer r > 0, the
problem
addressed in this paper is to find a clustering set I
The construction of P is shown by the following example.
T and a non-zero vector w such that the corresponding projection
Example 1: Let w = 1 1 1 2 1 1 1 1 1 1 ,
matrix P solves the model matching problem1
I1 = {1, 2}, I2 = {3, 4, 5} and I3 = {6, 7, 8, 9, 10}. Then,
minimize kg(s) − ĝ(s)kH2 .
1
(13)
P
√
√1
0
0
0
0
0
0
0
0
2
2
√1
√2
√1
However, finding an exact solution for this optimization
0
0
0
0
0
0
P= 0
.
6
6
6
problem is intractable given that the objective function is an
√1
√1
√1
√1
√1
0
0
0
0
0
5
5
5
5
5
implicit and non-convex function of P , and also because P
Given P defined over any clustering set I and weight vector is defined over a combinatorial structure. Our main contribuw, the control inversion strategy for the LQR problem (3) is tion, therefore, is finding a tractable relaxation for (13) as a
quadratic function of P , and thereafter designing P to solve
then composed of the following three steps.
1) Projection to reduced-order system: Using projection P , the relaxed problem. We make the following assumption so
that g(s) and ĝ(s) both have minimal realization.
we first construct a reduced-order model
Assumption 2.3: The pair (A, Bd ) is controllable.
(
˙
x̃(t)
= Ãx̃(t) + B̃ ũ(t) + B̃d d(t), x̃(0) = x̃0
Solution Strategy: The outline of our solution strategy is as
, (6)
follows. In Section III, we derive an upper bound relaxation for
ỹ(t) = C̃ x̃(t)
(13) such that its objective function is quadratic in P . Ideally
where à := P AP T ∈ Rr×r , B̃ := P B ∈ Rr×m , B̃d :=
1 The initial condition for the reduced-order model (6) does not need to be
P Bd ∈ Rr×nb and C̃ := P CP T = Ir . For this system x̃ ∈ Rr related to that of the full-order model (1). Our goal is to compare (11) and
is the state and ũ ∈ Rm is the control input.
(12), both of which have zero initial conditions.
•
It is row orthonormal, i.e. P P T = Ir ;
Image of P T P lies in the span of w, i.e. P T P w = w;
Given v 6= 0, P v = 0 only if wITi vIi = 0, i = 1, ..., r.
4
speaking, one can solve for P from this relaxed problem.
The computational complexity for constructing the objective
function is, however, O(n3 ) since it requires the computation
of the controllability Gramian of g(s). To bypass this difficulty,
a second round of relaxation is applied by exploiting the lowrank (denoted as κ) structure of the controllability Gramian.
After these two relaxations, the final objective function, still
quadratic in P , can be constructed in O(nκ2 ) complexity,
which is near linear if κ ≪ n. The solution to this optimization
is then addressed in two ways - first by finding the clustering
set I with a fixed weight vector w (Section IV), and second, by
finding w while keeping I fixed (Section V). We also propose
to combine these two approaches by an iterative algorithm.
The overall design flow and the numerical complexities for
each step are previewed in Fig. 1. Detailed explanations of
these complexities will be provided in the respective sections
to follow.
E"F&'()2?G
A#@4&"=23"4'0)
>#092?
E"F&'()2???
!"#$%$&'()
*!+,-.
ङ ࣄ
ङ ࣄ࢘
this process no coordinator will be able to infer individual
state measurements from other clusters. Finally, the coordinator of Ii computes the control vector ui by taking linear
combinations of the elements of X̃P x. The linear combination
follows from u = −R−1 B T P T X̃P x. The individual elements
of ui are broadcast to the respective input actuators in Ii .
Note that since the aggregation is applied to the state x(t),
and not on the subsystems, the state vector of any subsystem
can be partitioned among different clusters. Thus, in practice,
a subsystem may need to transmit its states to more than one
coordinator, and also receive control inputs from more than
one coordinator. We illustrate the three implementation steps
by an example in Fig. 2.
In the worst-case scenario when every subsystem has a
scalar state xi and a scalar control input ui , the two-layer
control implementation will result
in a much sparser commu
nication topology with n + r2 bidirectionallinks compared to
an LQR controller which would require n2 number of links.
This reduction, combined with standard networking protocols
such as multi-casting [24], makes the implementation of our
proposed controller convenient and cheap.
E"F&'()2??
A()&=(#2?)B"=4'()
?&"=$&'B"
>#092???
*8. C *:D.
E"F&'()2G
/"'01&23"4'0)5
617 89:2*4&$;#"2(<"),#((<.
(=2>#092??2*@)4&$;#"2(<"),#((<.
ङ
ङ
ङ ࢘ ङ ࢘
Figure 1: Step-by-step execution of the proposed designs
C. Benefits of Control Inversion
An important point to note is that the physical meaning of
the state x̃(t) of the reduced-order model (6) has no relation
to that of the state x(t) of our full-order model (1). This is a
key difference of the control inversion design from traditional
model-reduction based designs where the reduced-order state
vector is typically a direct projection of the full-order state
vector. The projection in our design is rather applied on the
controller X̃ instead of x(t). Two natural benefits of this
approach are as follow:
1. Tractability of design: The computational complexity
for constructing the reduced-order ARE in (8) is O(rn2 ),
while that for solving this ARE is O(r3 ). The computational
complexity required to design P through Sections III, IV and
V will be shown to be simpler than the O(n3 ) complexity of
a full-order LQR design. Thus, if r ≪ n, the overall control
inversion design becomes numerically more tractable than fullorder LQR.
2. Simplicity in implementation: The projected matrix X̂ =
P T X̃P is a structured r-ranked matrix, which results in a
sequential two-layer hierarchical control architecture. The implementation of the feedback u(t) = −R−1 B T X̂x(t) follows
three steps. First, a coordinator is assigned to each cluster Ii ,
which collects the measurements of all the states belonging
to that cluster. Each coordinator then computes the weighted
averaged state Pi,: x for its cluster, i = 1, 2, .., r. Next, the
coordinators exchange these weighted averages, and each of
them compute the r-dimensional vector X̃P x. Note that in
III. R ELAXATIONS
FOR
M ODEL M ATCHING
In this section we describe the theoretical derivation of the
relaxation for the model matching problem (13). We start by
discussing the well-posedness of the projected controller K̂
in Section III.A. The final optimization to be solved is then
obtained from two stages of relaxations as detailed in Section
III.B and III.C respectively. All proofs are presented in the
Appendix B.
A. Well-Posedness Conditions
The well-posedness of K̂ is equivalent to two conditions namely, if the reduced-order ARE (8) admits a solution X̃,
and if ĝ(s) is stable. We discuss these two factors as follows.
1) Existence condition: Similar to the full-order ARE (4),
the reduced-order ARE (8) is guaranteed with a unique soT
1
lution X̃ ≻ 0 if (Ã, G̃ 2 ) is stabilizable and (Q̃ 2 , Ã) is
observable [7]. However, unlike conventional model reduction
techniques that can utilize unstructured projections to preserve
the exact stabilizability and observability properties from
1
T
(A, G 2 ) and (Q 2 , A), the structured projection P in (5) does
not guarantee that. As a result, for a general system the ARE
(8) may not admit a unique solution X̃, and therefore K̂ may
not exist. To bypass this problem, we modify the definitions
of Q̃ and G̃ by using a constant shift. That is, in case the pair
T
1
(P AP T , P G 2 ) is not stabilizable and/or (Q 2 P T , P AP T ) is
not observable, we let
Q̃ = P QP T + αIr
and/or G̃ = P GP T + αIr
(14)
for a small constant shift α > 0. Using (14), Q̃ and G̃ become
1
positive-definite matrices, which yield (Ã, G̃ 2 ) controllable
T
and/or (Q̃ 2 , Ã) observable. The existence of X̃ and K̂ can
thus be guaranteed. The matching error introduced by this shift
will be discussed in the next subsection.
5
C1
ݔଷ ʹݔସ ݔହ
ݔଵ ݔଶ
ʹ
ݔ ݔ ଼ݔ ݔଽ ݔଵ
2
ݔଵ
10
ݔଽ
9
ݔଵ
1
ݔଷ
ݔସ
4
3
ݔ
6
༠
C1
څ
ଶ
C3
ݔହ
1
5
ݔ
ל
C2
C1
ͷ
C3
ݔଶ
څ
C2
2
༠
ହ
C
2
5
10
9
(a) Step 1 - state averaging P x
ל
7
10
8
ଶל
6
3
7
7
଼ݔ
C2
4
5
6
ל
3 ༠
༠ ହ
༠
༠
ହ
ହ
ହ
1
4
3
څ
ଶ
8
9
(b) Step 2 - reduced-order control X̃P x =
[⋆ ◦ ⋄]T
8
(c) Step 3 - control inversion u = −X̂x
Figure 2: Cyber-physical architecture for implementing the feedback controller K̂. Here I1 = {1, 2}, I2 = {3, 4, 5} and I3 = {6, 7, 8, 9, 10}.
Solid lines represent physical connections, while dashed lines represent communication links. C1 , C2 , C3 are coordinators for clusters 1, 2
and 3. For the simplicity of illustration, for this example we assume R−1 B T = In .
2) Bound of ARE solution: We first recall a lemma from
[27] about the eigenvalue bound of a general ARE solution.
Lemma 3.1: (Lemma 1.5 in [27]) The solution X from the
ARE (4) satisfies the following upper bound:
λ̄(X) ≤ β(A, G, Q) = λ̄(Dt )
λ̄[(Q + KtT Kt )Dt ]
,
λ(F Dt )
(15)
1
where Kt is any matrix stabilizing A + G 2 Kt , and Dt and F
are positive-definite matrices that satisfy
1
1
(A+G 2 Kt )T Dt + Dt (A+G 2 Kt ) ≤ −F.
(16)
From this lemma, given any stabilizing controller Kt one
can always find an upper bound for λ̄(X) using the function
β(A, G, Q). This function will be used next to verify stability
of ĝ(s).
3) Stability condition: Before stating the main stability
criterion for ĝ(s), we first make the following assumption on
the projection weight w from Definition 2.2.
Assumption 3.2: The weight vector w satisfies wITi vIi 6= 0,
i = 1, ..., r for any Av = λv, Re(λ) ≥ 0.
Note that from Definition 2.2, if wITi vIi = 0, i = 1, ..., r
then P v = 0. This would imply (A − B K̂)v = Av = λv,
Re(λ) ≥ 0, which means that the unstable eigenvalues of A
are retained in the closed-loop. Assumption 3.2 is made to
avoid this situation. With Assumption 3.2, we next state a
sufficient condition for the stability of ĝ(s).
Theorem 3.3: Given Assumption 2.1 and 3.2, the TFM ĝ(s)
is asymptotically stable if
σ(Q) − 2σ̄(X)σ̄(G)σ̄(E) + σ(X)2 σ(G) > 0,
(17)
where E = X − X̂ denotes the error between the full-order
ARE solution from (4) and the projected solution from (9).
Theorem 3.3 provides two options for achieving stability
of ĝ(s). In order to meet the condition (17), one can select
Q and R such that σ(Q) ≫ σ̄(G) > 0. In practice, this
choice of (Q, R) will make the design more robust towards
the uncertainties in A and B [29]. As a drawback, a larger
Q will also result in a high feedback gain, making the
system vulnerable to noise. An alternative to satisfy (17) is to
minimize σ̄(E). As will be shown in the next subsection, σ̄(E)
is proportional to the objective function of our proposed upper
bound relaxation for (13). Therefore, solving the relaxation
problem will also assist in enlarging the inequality gap in (17).
The following lemma provides a sufficient condition for (17)
by which stability of ĝ(s) can be verified without knowing X.
Lemma 3.4: The stability condition (17) holds if
σ̄(X̃) <
σ(Q)
− β(A, G, Q),
2σ̄(G)β(A, G, Q)
(18)
where function β(A, G, Q) is defined in Lemma 3.1.
Remark 1: For the case when A is unstable, computing
β(A, G, Q) would require the knowledge of Kt according
to Lemma 3.1. Similarly for w to meet Assumption 3.2, one
also needs to know the eigenvectors of all unstable modes of
A. Verifying stability of ĝ(s) for an unstable A is, therefore,
admittedly more computationally expensive than for a stable
A. This computational burden obviously does not exist when
A is stable, and may not also exist when A, despite having
zero eigenvalues, has specific structural properties. One such
example is when A is a weighted Laplacian matrix (i.e., when
(1) is a consensus network). We will illustrate this special case
of consensus network in Appendix A.
B. Relaxation I: Upper Bound Minimization
We present our first stage of relaxation assuming that the
sufficient condition for closed-loop stability from Theorem
3.3 holds. As mentioned before, finding a P that exactly
minimizes kg(s)− ĝ(s)kH2 is an intractable problem. To relax
this problem, we find an upper bound for kg(s) − ĝ(s)kH2 .
We denote the error system by ge (s), which can be written as
Ae Be
ge (s) : = g(s) − ĝ(s) =
Ce 0
A − GX
Bd
(19)
=
A − GX̂ −Bd .
In
In
0
6
I
I
From a similarity transformation of T = n n and T −1 =
In 0
0
In
, (19) yields
In −In
A−GX̂ G(X̂−X) 0
T Ae T −1 T Be
ge (s)=
=
A−GX
Bd
Ce T −1
0
In
0
0
−1
= −(sIn − A + GX̂)
−1
GE(sIn − A + GX)
Bd ,
(20)
where E = X − X̂. By taking norms on both sides of (20),
we get
kge (s)kH2 ≤ k(sIn − A + GX̂)−1 GkH∞ kEg(s)kH2 ,
which from the bounded real lemma [7] and the definition of
H2 norm reduces to
1
kge (s)kH2 ≤ γkEΦ 2 kF ,
(21)
where γ is any positive real number such that a real-valued
matrix Γ = ΓT 0 exists and satisfies
Γ(A − GX̂) + (A − GX̂)T Γ ΓG
In
GΓ
−γIn
0 ≺ 0,
In
0
−γIn
R ∞ (A−GX)τ
T
T
1
Bd BdT e(A−GX) τ dτ ≻ 0
and Φ := Φ 2 Φ 2 = 0 e
is the solution of the Lyapunov equation
(A − GX)Φ + Φ(A − GX)T + Bd BdT = 0.
(22)
Inequality (21) shows that kge (s)kH2 is linearly bounded by
1
kEΦ 2 kF . Therefore, one way to solve the original model
matching problem (13) will be to find a P that minimizes
1
kEΦ 2 kF . By doing so, σ̄(E) can also be minimized to some
1
1
extent since σ̄(E) ≤ kEkF ≤ σ̄(Φ− 2 )kEΦ 2 kF , which
will help in meeting the stability condition (17). This type
of bound minimization is common in model and controller
reduction, and has been attempted (see [17] and the references
therein) under the assumption that P is unstructured, or
more specifically P is an r-dimensional Krylov subspace
1
1
1
1
from K(A, Q 2 , r) = Q 2 AQ 2 · · · Ar−1 Q 2 . By this
assumption, E can be found as an explicit function of P
associated with a Householder transformation. In our case,
however, P has a structure as in (5), due to which this explicit
functional relationship does not hold anymore. We, therefore,
apply perturbation theory of ARE to further relax the bound in
1
(21), and derive a new upper bound on kEΦ 2 kF as an explicit
function of P in the following theorem.
1
Theorem 3.5: Denote ξ = k(In − P T P )Φ 2 kF . The norm
1
of the weighted error EΦ 2 satisfies the inequality
1
2
kEΦ kF ≤ f (ξ) = ǫ1 σ̄(Q)ξ 2 + 2ǫ1 ǫ2 ξ + αǫ1 ǫ3 ,
(23)
1
where ǫ1 =
σ̄(Φ− 2 )
1
σ[Φ− 2
1
1
(A−GX)Φ 2 ]
1
, ǫ2 = β̃ σ̄(A)σ̄(Φ 2 )+ σ̄(QΦ 2 ),
ǫ3 = (β̃ 2 + 1)σ̄(Φ), and β̃ = supP β(Ã, G̃, Q̃) are positive
scalars that are independent of P , and α is defined in (14).
The constant αǫ1 ǫ3 in (23) represents the matching error
introduced by the constant shift α from (14). This error can
be disregarded if the reduced-order ARE (8) admits a solution.
From (21) and (23), it then follows that kge (s)kH2 ≤ γf (ξ),
due to which we approach the minimization of kge (s)kH2 by
minimizing f (ξ) with respect to P . Since f (ξ) is a monotonic
function of ξ, the minimization of f (ξ) is equivalent to
minimizing the value of ξ as
minimize
P
1
1
ξ = kΦ 2 − P T P Φ 2 kF .
(RL-1)
The optimization (RL-1), therefore, serves as an upper bound
relaxation for the original model matching problem (13).
Note that, in general, it is impossible to exactly quantify the
optimality gap between (13) and (RL-1) since (13) is nonconvex even without posing any combinatorial constraints on
P . The optimality gap will be small if the minimum value of
ξ is close to zero, in which case the error kge (s)kH2 in (13)
will be nearly zero as well following (21) and (23).
1
1
The ideal case ξ = 0 will happen when Φ 2 = P T P Φ 2 ,
1
meaning that Φ 2 is invariant to the mapping P T P . This holds
for the trivial case where r = n and P = In . To achieve a
sufficiently small minimum for (RL-1) for r < n, we will
develop two designs in Sections IV and V.
C. Relaxation II: Low-Rank Approximation
We next discuss the numerical complexity in constructing
the optimization problem (RL-1), and how this complexity
can be simplified by making appropriate approximations on
Φ. In the most general case, Φ required for (RL-1) needs to
be computed through the following procedures. First, recall
the definition of the Hamiltonian matrix
A
−G
H :=
.
(24)
−Q −AT
The eigenvalues of H are symmetric about the imaginary
axis. Suppose
H is diagonalizable and that the columns of
Y
the matrix
span the stable invariant subspace of H,
Z 2n×n
i.e.
Y
Y
H
=
Λ− ,
(25)
Z
Z
−
where Λ− = diag([λ−
1 , ..., λn ]) consists of all the eigenvalues
−
of H in the left-half plane, i.e. 0 > λ−
1 > ... > λn . The
stabilizing solution of ARE can thus be found by X = ZY −1
[7]. The first n rows of (25) are expanded as A − GZY −1 =
Y Λ− Y −1 , which means Λ− and Y are the eigenvalues and
right eigenspace of the closed-loop state matrix A−GX. Then
from the Lyapunov equation (22), we can write Φ directly in
terms of Y and Λ− as [6]
Φ = Y (Y −1 Bd BdT Y −T ◦ C)Y T ,
where C ∈ Rn×n is a Cauchy matrix with
i
h
1
−
,
Ci,j = − λ−
+λ
i
j
1
(26)
and subsequently obtain Φ 2 from the Cholesky decomposition.
1
Therefore, to compute Φ and then Φ 2 , one will need to
compute the full stable eigenspace Y from H following (25).
This computation is as expensive as solving a full-order LQR
with O(n3 ) complexity for both computation and memory
[28]. This may defeat the purpose of our design since we
7
want our controller to be numerically much simpler than the
full-order LQR. To bypass this difficulty, we next show that
Φ can be approximated by a matrix Φκ that follows from a
κ-dimensional (κ < n, not necessarily equal to r) invariant
subspace of Y . Ideally κ should be at most r to justify
the computational benefit of our design while preserving an
acceptable accuracy in the error norm ξ. This matrix Φκ is
constructed as follows.
Definition 3.6: Define Φκ ∈ Rn×n as
1
T
Φκ := Φκ2 Φκ2 = Y1 (Ω1 Bd BdT ΩT1 ◦ C1,1 )Y1T ,
(27)
−1
where Y1 =Y:,1:κ , Ω1 =Y1:κ,:
and C1,1 =C1:κ,1:κ are respectively
the κ-dimensional partitions of Y , Y −1 and C.
By definition of Φκ , one only needs to compute the first κ
−
eigenvalues λ−
1 , ..., λκ of H, and the Y1 component of the first
κ eigenvectors. Ω1 can be approximated by the pseudo-inverse
of Y1 . These κ smallest eigenvalues and eigenvectors can be
solved by Krylov subspace-based techniques such as Arnoldi
algorithm in O(nκ2 ) time [28]. Therefore, instead of (RL-1),
we consider solving a computationally simpler approximation
of (RL-1) as
minimize
P
1
2
1
2
ξκ = kΦκ − P T P Φκ kF ,
(RL-2)
1
where Φκ2 ∈ Rn×κ from Definition 3.6 can be computed as
1
1
Φκ2 = Y1 (Ω1 Bd BdT ΩT1 ◦ C1,1 ) 2 . The optimality gap between
optimizations (RL-1) and (RL-2) can be quantified by the
following lemma.
Lemma 3.7: Assume Y −1 has a moderate condition number
η, and each column of Bd has a unitary norm. The minimum
ξ ∗ from (RL-1) and ξκ∗ from (RL-2) satisfy
v
u
n
X
u
1
∗
∗
(28)
− −.
ξ − ξκ ≤ tη 2 nb
2λi
i=κ+1
From Lemma 3.7, the optimality gap between
(RL-1) and (RL-2) will be negligible when the error
q
Pn
η 2 nb i=κ+1 − 2λ1− is kept small. In practice, this
i
situation happens when there exists κ (κ ≪ n) dominant
eigenvalues in the Hamiltonian matrix, i.e., H has the
following spectral gap
0<
|λ−
1|
< ... <
|λ−
κ|
≪
|λ−
κ+1 |
< ... <
|λ−
n |.
This gap can exist if the open-loop network (1) exhibits
coherent behavior [3]. As a result of this spectral gap, the
RHS of (28) can become sufficiently small, in which case the
optimal value of (RL-2) will closely resemble that of (RL-1)
while the computation of the objective function requiring a
much tractable complexity of O(nκ2 ), κ < n. The idea of
utilizing a κ-dimensional subspace for computing Φκ is similar
in spirit to finding an unstructured approximate ARE solution
as proposed in [17]. However, it should be noted that unlike
[17] where the selection of the κ eigenvectors is undetermined,
for our problem the error bound in Lemma 3.7 clearly guides
the choice of the κ eigenvectors in terms of tightening the
optimality gap between (RL-1) and (RL-2).
IV. D ESIGN I: C LUSTER D ESIGN
In this section we present an algorithm to design P by
solving (RL-2). Note that P has two degrees of freedom I and w. For the design in this section, we keep w fixed, and
minimize (RL-2) over I. Although inherently this is an NPhard problem, fortunately the specific structure of the objective
function ξκ , together with the structure imposed on P in (5),
enables (RL-2) to be solved by efficient numerical algorithms
such as weighted k-means [12]. We show these results as
follows.
To establish the equivalency of (RL-2) to the weighted kmeans optimization, it is useful to borrow a nominal projection
matrix P̄ as
(
1
j ∈ Ii
, i = 1, ..., r.
(29)
P̄i,j := kwIi k2
0
otherwise
From Definition 2.2, P̄ satisfies P = P̄ W , where W =
diag(w). With this notation, we have
1
1
Φκ2 − P T P Φκ2 = W (Ψ − P̄ T P̄ W 2 Ψ),
(30)
1
where Ψ = [ψ1 , ..., ψn ]T denotes the matrix W −1 Φκ2 . Therefore, the entries of the matrix P̄ T P̄ W 2 can be found by
( w2
k
j ∈ Ii & k ∈ Ii
2
T
2
[P̄ P̄ W ]j,k = kwIi k2
0
otherwise
for i = 1, ..., r. Thus the j th row of the matrix P̄ T P̄ W 2 Ψ
can be written as
P
2 T
k∈Ii wk ψk
T
2
T
[P̄ P̄ W Ψ]j,: = ci = P
(31)
2 ,
k∈Ii wk
for j ∈ Ii . It is clear from above that for all the index j that
are assigned to the same cluster Ii , cTi is a weighted average
(or a weighted centroid) of the row vectors ψjT . Moreover, the
matrix P̄ T P̄ W 2 Ψ will have identical rows for those whose
indices are inside the same cluster. Therefore, (RL-2) can be
posed as an unsupervised clustering problem
minimize
I1 ,...,Ir
ξκ2 =
n
X
j=1
wj2 kψj − ci k22 .
(32)
The optimization problem in (32) is in the same form as
a weighted k-means optimization, which minimizes the Euclidean distance weighted by wj2 between each data point
ψj and its centroid ci . Thus, data points which are close to
each other in the weighted distance are assigned to the same
cluster. A standard method for solving this problem is Lloyd’s
algorithm [12], using which we present the weighted k-means
1
clustering for (32) in Algorithm 1. With Φκ2 ∈ Rn×κ as the
input, the running time of Lloyd’s algorithm is O(nκrk),
where k is the total number of iterations. As in any heuristic
algorithm, Algorithm 1 does not guarantee convergence to a
global minimum. Hence, if one is not satisfied by the k-means
solution one can apply any state-of-art clustering algorithms
for solving (32). Using the resulting clustering set I and
the fixed weight w, we can then construct the projection
P and execute the control inversion design to get K̂. The
8
clustering weight w can be selected as any vector that satisfies
Assumption 3.2. When A is Hurwitz, a simple choice of w
can be the vector of all ones.
We conclude this section by summarizing the total numerical complexity for our design based on Algorithm 1.
The chain of approximations involved in this design till the
output of Algorithm 1 follows the sequence of equations: (21),
(23), (RL-1), (RL-2) and (32). The total complexity amounts
to O(nκ2 ) + O(nκrk) + O(n2 r) + O(r3 ), which includes
construction of the objective function for (RL-2), execution
of Algorithm 1, computation of reduced-order matrices triple
(Ã, Q̃, G̃), and solving the reduced-order LQR (7), respectively. This complexity can be at most O(n2 r) if κ ≤ r,
which is more tractable compared to the O(n3 ) complexity
of full-order LQR, especially when r and κ are sufficiently
small.
Algorithm 1: H2 closed-loop clustering
1
1
Input : Φκ2 , w and r
Initialization: Assign r random rows from
1
2
Ψ = W −1 Φκ2 = [ψ1 , ..., ψn ]T as the initial centroids
c01 , ..., c0r ;
Find initial clustering sets
I 0 = {j → Ii0 | argmin wj2 kψj − c0i k22 , j = 1, ..., n};
i=1,...,r
3
4
5
6
Update the centroids: c0i =
P
j∈I 0
i
P
8
9
wj2
j∈I 0
i
, i = 1, ..., r ;
k = 1;
while I k−1 6= I k or within maximum iterations do
Update clustering sets I k = {j →
Iik | argmin wj2 kψj − cik−1 k22 , j = 1, ..., n};
i=1,...,r
7
wj2 ψj
Update the centroids: cki =
k =k+1 ;
end
Output: I = I k
P
j∈I k
i
P
wj2 ψj
j∈I k
i
wj2
, i = 1, ..., r ;
V. D ESIGN II: W EIGHT D ESIGN
We next state a variant of our proposed controller where
we solve (RL-2) by fixing the clustering set I and varying
the projection weights in w instead. This type of a controller
may be needed when a subsystem prefers to have all of its
states assigned to the same cluster. Multiple subsystems in
the network may also exhibit spatial clustering based on their
geographical proximities, in which case they may prefer to
have a dedicated coordinator for themselves. This scenario
commonly arises in power system networks. Utility companies
always prefer to send the state information of their generators
to only their own local control centers instead of sharing
that information with any other company. Therefore, in this
section we develop a new set of algorithms where we fix I to
represent the identities of the desired clusters, and minimize
ξκ in (RL-2) over w. A visual interpretation of this approach
and its comparison to the design in Section IV are shown in
Fig. 3. Note that the optimal values of w so obtained denote
the relative importance of the network nodes in the closedloop system with the projected controller. In Fig. 3c we show
this relative importance by shrinking or expanding the size of
the nodes. We describe the design for two cases depending on
whether (1) is stable or not.
A. Case I: Stable Open-Loop
We consider the same optimization objective as in (RL-2)
but now minimize it with respect to w as
1
minimize
w
1
ξκ = kΦκ2 − P T P Φκ2 kF .
(33)
To solve (33), we consider a binary projection matrix P̂ as
(
1
j ∈ Ii
P̂ (i, j) :=
.
0
otherwise
As can be verified from Definition 2.2, P̂ here also satisfies
P = P̂ Ŵ with Ŵ = diag(ŵ), where this ŵ is defined by
wj
, j ∈ Ii ,
(34)
ŵj =
kwIi k2
such that ŵITi ŵIi = 1, for i = 1, ..., r. Using these notations,
we can rewrite the objective function in (33) as
ξκ2 = tr(Φκ − P T P Φκ ) = tr(Φκ ) − tr(Ŵ P̂ T P̂ Ŵ Φκ )
= tr(Φκ ) − ŵT (P̂ T P̂ ◦ Φκ )ŵ.
Since tr(Φκ ) is a constant number, an equivalent form of (33)
follows as
maximize
ŵT (P̂ T P̂ ◦ Φκ )ŵ
subject to
kŵIi k2 = 1, i = 1, ..., r.
ŵ
(35)
The Hadamard product P̂ T P̂ ◦Φκ preserves the structure from
P̂ T P̂ , or equivalently from the clustering set I in the objective
function. As a result, the optimization problem (35) boils down
to r decoupled optimizations
maximize
ŵITi Φκ[Ii ,Ii ] ŵIi
subject to
kŵIi k2 = 1,
ŵIi
(opt1)
for i = 1, ..., r, where Φκ[Ii ,Ii ] denotes the submatrix of Φ
corresponding to the indices in Ii . This decoupling can be
illustrated by the same example we used before.
Example 2: (Continued) Consider the sets I1 , I2 and I3 ,
and the matrix
1 1 0 0 0 0 0 0 0 0
P̂ = 0 0 1 1 1 0 0 0 0 0 .
0 0 0 0 0 1 1 1 1 1
The objective function in (35) can be block diagonalized as
Φκ[1:2,1:2]
0
0
ŵ=
0
Φκ[3:5,3:5]
0
ŵ (P̂ P̂ ◦Φκ )ŵ=ŵ
0
0
Φκ[6:10,6:10]
T
ŵ1:2 Φκ[1:2,1:2] ŵ1:2
0
0
T
.
0
ŵ3:5 Φκ[3:5,3:5] ŵ3:5
0
T
0
0
ŵ6:10
Φκ[6:10,6:10] ŵ6:10
T
T
T
Since Assumption 3.2 holds trivial for this case, we state
the following theorem.
9
(a) Spatial groups partitioned by the closeness in terms of geometric distances
(b) Node clusters resulting from cluster
design as in Section IV
(c) Imposing projection weights while fixing the clustering sets to the spatial clusters
Figure 3: Illustration of cluster design and weight design.
Theorem 5.1: The global optimum for (opt1) is obtained at
ŵIi = v̄(Φκ[Ii ,Ii ]), for i = 1, .., r.
Proof: Given P̂ T P̂ 0 and Φ 0, P̂ T P̂ ◦Φκ is positivesemidefinite according to the Schur product theorem. Also, the
objective function in (opt1) is a standard Rayleigh quotient for
symmetric eigenvalue problem. Therefore, the maximum of
objective function in (opt1) is obtained at the largest singular
value of Φκ[Ii ,Ii ] . Since Φκ[Ii ,Ii ] 0, its largest singular
value is the same as its largest eigenvalue λ̄(Φκ[Ii ,Ii ] ), and
hence the optimum is obtained at its dominant eigenvector
v̄(Φκ[Ii ,Ii ] ).
Given that Φκ[Ii ,Ii ] is symmetric, the eigenvector of its
largest eigenvalue can be computed efficiently by Krylov
subspace-based techniques, e.g. Lanczos algorithm [28] in this
case. Solving (opt1) for i = 1, ..., r, therefore, requires a worst
case complexity of O(n) in total. Once ŵ is solved from
Theorem 5.1, one can then choose any w that satisfies (34),
and this w would serve as global optimum for (33).
B. Case II: Unstable Open-Loop
For an unstable open-loop system (1), w solved from
Theorem 5.1 is not guaranteed to conform to Assumption
3.2, as a result of which ĝ(s) can become unstable. To avoid
such a hazardous situation, we add an extra penalty on (opt1)
to restrict w under Assumption 3.2. This penalty term is
formulated as follows.
Recall the reduced-order ARE (8). By pre- and postmultiplying it with P T and P , and after a few calculations,
we get AT X̂ + X̂A + Q̂ − X̂GX̂ = 0, with
Q̂ := P T Q̃P − (U T U AT X̂ + X̂AU T U ),
(36)
where U is the complement of P . The ARE above implies
−1 T
that our proposed controller
R ∞ K̂T = R BT X̂ is equivalent to
an LQR problem min 0 (x Q̂x + u Ru)dt. Denote the
eigenvalue decomposition of A as
Λ
,
(37)
A V V̄ = V V̄
Λ̄
where Λ̄ 0 contains all the unstable eigenvalues of A.
Applying the transformation z = [V V̄ ]−1 x, one can write
T
V T Q̂V̄
T
T V Q̂V
z.
x Q̂x = z
V̄ T Q̂V V̄ T Q̂V̄
The matrix V̄ T Q̂V̄ corresponds to the LQR weight for the
unstable dynamics of the state trajectories. Therefore, to assure
that the unstable modes of A are indeed penalized by Q̂, we
consider finding w that maximizes tr(V̄ T Q̂V̄ ). Given that
the matrix U T U AT X̂ + X̂AU T U in (36) is bounded but not
sign-definite, we consider maximizing tr(V̄ T Q̂V̄ ) by simply
maximizing its first part as
maximize
fV̄ = tr(V̄ T P T Q̃P V̄ )
subject to
kŵIi k2 = 1, i = 1, ..., r.
ŵ
(opt2)
Combining (opt2) with the main optimization (opt1), one can
formulate the weight design for the unstable case as
maximize
[ŵT (P̂ T P̂ ◦ Φκ )ŵ]2 + ρ · fV̄
subject to
kŵIi k2 = 1, i = 1, ..., r,
ŵ
(opt3)
where ρ > 0 is a penalty factor. In (opt3), the objective
function from (opt1) is squared to match with the order of
fV̄ . This optimization problem is in the form of a fourthorder sum of squares (SOS) over r > 1 sphere constraints, for
which finding even a local optimal is very difficult. One way
to bypass this can be to approximate matrix P T Q̃P in fV̄
as P T Q̃P ◦ P̂ T P̂ , meaning only to retain the block-diagonal
component of Q only. In this way, (opt3) is reduced to r SOS
sub-problems with one sphere constraint for each as,
maximize
ŵ
subject to
(ŵIi ⊗ ŵIi )T (Φκ[Ii ,Ii ] ⊗ Φκ[Ii ,Ii ]
+ ρQIi ,Ii ⊗ V̄Ii ,: V̄ITi ,: )(ŵIi ⊗ ŵIi )
(opt*)
kŵIi k2 = 1,
for i = 1, ..., r. While this approximation will follow naturally
if Q is block-diagonal, the upshot is that the closed-loop
performance of the projected system may suffer if Q has
dominant off-block-diagonals. In practical networks, however,
it is quite common to simply minimize the energy of a node
itself, or the energy within a cluster, which implies that Q
is very commonly a diagonal or block diagonal matrix. In
fact, Q would indeed be preferred as block-diagonal for the
scenario described in this section since network operators
will always try to discourage closed-loop coupling of their
own cluster nodes with other clusters. In those cases, (opt3)
and (opt*) become equivalent problems, yielding the same
solution. However, irrespective of whether Q is block-diagonal
or not, the following theorem shows that the solution of (opt*)
will satisfy Assumption 3.2.
Theorem 5.2: Given ŵIi , i = 1, ..., r solved from (opt*)
with ρ > 0, Assumption 3.2 holds for any w satisfying (34).
10
Proof: Note that V̄:,j , j = 1, ..., nv represents the
eigenvector of the j th unstable eigenvalue. The second part
ofPthe objective function in (opt*) can be rewritten as
v
ρ nj=1
(V̄ITi ,j ŵIi )2 (ŵITi QIi ,Ii ŵIi ). Once (opt*) is solved,
each individual scalar (V̄ITi ,j ŵIi )2 has to be maximized away
from 0, meaning V̄ITi ,j ŵIi 6= 0, i = 1, ..., r, j = 1, ..., nv for
AV̄:,j = λV̄:,j , Re(λ) ≥ 0. This satisfies Assumption 3.2.
The following lemma shows the performance trade-off due
to the approximation in (opt*) when Q is not block-diagonal.
Lemma 5.3: Let the maximum of (opt3) be J1 . The maximum J2 of (opt*) satisfies
where Je =
J 1 − J e ≤ J2 ≤ J 1 + J e ,
(38)
Pr
T
max
k=1, k6=i ρkQIi (j),Ik k1 kV̄l,: V̄Ik ,: k1 .
i=1,..,r
j,l=1,...,n
Proof: The proof follows directly from the Gershgorin
circle theorem [28], and is shown in the Appendix.
We next present the solution for (opt*). Since (opt*) is a set
of r decoupled problems, we illustrate the solution for just one
cluster I = {I1 }. This will also allow us to drop the subscripts
in all the variables used in (opt*), making the notations easier
to follow. We define a fourth-order tensor F ∈ Rn×n×n×n as
Fi,j,k,l = Φκ[i,j] Φκ[k,l] + ρQi,j Sk,l , i, j, k, l=1, ..., n. (39)
The proof is omitted as the equations above can be easily
verified by matching the coefficients of the polynomials on
both sides.
In summary, the optimization problem (opt**) can be approached by substituting F with a super-symmetric tensor F s .
One can, thereafter, solve (opt**) using techniques developed
for Z-eigenvalue problems. We solve (opt**) using the tensor
power iteration method [31] in Algorithm 2. The convergence
properties of this algorithm can be found in [31]. Due to the
super symmetry of F s , the worst case (only one cluster) complexity for each iteration of Algorithm 2 is O(n4 ). Although
this computation cost is expensive, the algorithm can be easily
parallelized, and is easier to implement than O(n3 ) full-order
LQR as the memory required is only O(n2 ). Moreover, the
value of n for Algorithm 2 scales down as the number of
clusters increases.
Algorithm 2: Power iteration for projection weight design
1
2
3
4
T
where S denotes the product matrix S = V̄ V̄ . After a few
manipulations, it can be shown that (opt*) is equivalent to the
following problem
maximize
F ⊙ (ŵ ⊗ ŵ ⊗ ŵ ⊗ ŵ)
subject to
kŵk2 = 1,
ŵ
5
6
(opt**)
7
where ⊙ denotes the element-wise product. It has been studied
in [13] that such a polynomial optimization is equivalent to
finding the largest Z-eigenvalue of F , if F is super-symmetric.
From the definition in [13], a super-symmetric tensor is one
whose entries are invariant to any permutation to the index,
i.e. Fi,j,k,l = ... = Fl,k,j,i , which fails for (39) as Fi,j,k,l 6=
Fi,k,j,l . However, note that although F is not super symmetric,
F ⊙ (ŵ ⊗ ŵ ⊗ ŵ ⊗ ŵ) is a one-dimensional polynomial which
is invariant to any index permutations.2 Following this logic,
we rewrite the objective function in (opt**) as follows.
Proposition 5.4: Given the fourth-order tensor F specified
by (39), the polynomial F ⊙ (ŵ ⊗ ŵ ⊗ ŵ ⊗ ŵ) is identical to
8
9
10
11
12
13
Input : Φκ , Q, V̄ , I, ρ and δ
Partition Φκ[Ii ,Ii ] , QIi ,Ii and V̄Ii based on I;
for i = 1, ..., r do
Find F s corresponding to Ii by Proposition 5.4;
Initialization: Compute the dominant eigenvector of
F s as v̄(F s ), then choose the initial vector v 0 as
v 0 = v̄(unvec(v̄(F s )));
k = 1;
k
k T
) F s (v k ⊗v k )
while (vk−1(v⊗v⊗v
k−1 )T F s (v k−1 ⊗v k−1 ) − 1 > δ or within
maximum iterations do
v k = unvec(F s (v k−1 ⊗ v k−1 ))v k−1 ;
k
v k = kvvk k2 ;
k = k + 1;
end
ŵIi = v k ;
end
Construct w and then P by (34);
Output: P
C. Optimizing (RL-2) with respect to both I and w
The designs proposed in Section IV and this section can
be combined to optimize (RL-2) as a function of both I and
F ⊙ (ŵ ⊗ ŵ ⊗ ŵ ⊗ ŵ) = (ŵ ⊗ ŵ) F (ŵ ⊗ ŵ),
w iteratively. In this case, one would start with an arbitrarily
chosen w, and minimize ξκ with respect to I using Algorithm
where F s is a super-symmetric tensor specified by
1. Say, the optimal cluster set is given as I ∗ . Thereafter,
1
s
one would fix I to I ∗ , and minimize ξ with respect to w
Fi,j,k,l = (Φκ[i,j] Φκ[k,l] + Φκ[i,k] Φκ[j,l] + Φκ[i,l] Φκ[j,k] )+
3
using Theorem 5.1 or Algorithm 2 depending on whether
1
ρ(Qi,j Sk,l +Qi,k Sj,l +Qi,l Sj,k +Qj,k Si,l +Qj,l Si,k +Qk,l Si,j ) (1) is stable, and so on. The resulting algorithm is shown in
6
Algorithm 3.
2
2
for i, j, k, l = 1, ..., n, and F s ∈ Rn ×n , the matrix unfolding
of F s , can be obtained from
VI. N UMERICAL E XAMPLES
s
s
s
Fn(i−1)+k,n(j−1)+l
= Fi,j,k,l
,
T
s
i, j, k, l = 1, ..., n.
2 This is analogous to an unsymmetric matrix whose quadratic form is inT
variant to the transpose operation, i.e. z T F 2+F z = z T F z = z T F T z ∈ R.
To verify our proposed algorithms, we use a consensus
network model defined over a 500-node graph G500 . The
graph is randomly generated with 0.5 overall probability for
edge attachment, and 6 spatial clusters with a proportion
11
(a) Clusters by coherency clustering
(b) Clusters by H2 open-loop clustering
(c) Clusters by H2 closed-loop clustering with Q1
(d) Clusters by H2 closed-loop clustering with Q2
Figure 4: Clustering of the 500-node network with r = 6 clusters. Nodes assigned to the same cluster are marked by the same color in one
figure. Note that only the node identities are comparable between figures, but not the cluster identities.
Table I: Input specifications for Algorithm 1
Algorithm 3: Iterative algorithm for finding P
Input : A, B, Bd , Q, R and r
1
2
by Definition 3.6;
1 Compute Φκ
0
0
0
2 Choose w = 1n , and compute W = diag(w ) and the
Algorithm
H2 closed-loop clustering
H2 open-loop clustering [16]
Coherency clustering [3]
1
3
4
5
6
7
8
k-means input Ψ0 = (W 0 )−1 Φκ2 ;
k = 1;
while I k−1 6= I k or within maximum iterations do
Solve I k from Algorithm 1 by (Ψk−1 , wk−1 , r) ;
Update wk from Theorem 5.1 or Algorithm 2 by I k ;
k =k+1 ;
end
Output: P
of 100 for the number of edges within clusters versus the
number of edges across clusters. We also apply a random
weight 1 ≤ Mi,i ≤ 2 on each node. The resulting state
matrix A follows the expression (40) presented in Appendix
A. The disturbance is assumed to enter from the 364th node,
i.e. Bd equals to the 364th column of I500 . We assume
B = R = I500 and two choices of Q as: a scaled identity
matrix Q1 = 1000 × I500 , and Q2 = [L(G500 )]2 , which is the
square of the unweighted Laplacian matrix of G500 . Both Q1
and Q2 satisfy Assumption 2.1. For this simulation example,
solvability of (8) and stability of ĝ(s) are guaranteed by
Theorem A.2 and Theorem A.5 in Appendix A, respectively.
Data
Φ
Φo a
Φc b
Weight
w
w
1n
Number of Clusters
r
r
r
1
T
Let vcT be the complement of v̄T , and then Φo := Φo2 Φo2 =
R
T T
T
vc [ 0∞ e(vc Avc )τ vcT Bd BdT vc e(vc A vc )τ dτ ]vcT .
b Let the eigenvalues of −L(G) be 0 = λ > λ ≥ ... ≥ λ , Ψ =
n
c
1
2
[v1 , ..., vr ] where vi is the right eigenvector of λi .
a
A. Cluster Design
We start by finding the closed-loop clustering set I with
respect to a fixed weight w = v̄ = v̄(A) with r = 6.
For comparison, we also apply two other popular clustering
algorithms, namely, H2 open-loop clustering proposed in [16],
and coherency-based clustering proposed in [3]. Both of these
clustering algorithms can be transformed into Algorithm 1,
with their equivalent inputs as summarized in Table I. Note that
these two algorithms capture only the open-loop characteristics
of the network, and hence do not depend on the choice of
Q and R. Fig. 4a shows that the clusters identified by coherency based clustering closely resemble the spatial clusters
of the open-loop network except for a few discrepancies.
For example, two distant groups of nodes are assigned to
the same cluster shown in yellow. These discrepancies arise
from the fact that the spatial clusters are only based on the
edge-weights (that model geographical distance between two
12
0.45
0.35
Closed−loop clustering
Open−loop clustering
Coherency clustering
0.4
Closed−loop clustering
Open−loop clustering
Coherency clustering
0.3
0.35
0.25
kg(s)−ĝ(s)k H2
kg(s)k H2
kg(s)−ĝ(s)k H2
kg(s)k H2
0.3
0.25
0.2
0.2
0.15
0.15
0.1
0.1
0.05
0.05
0
0
100
200
300
400
0
0
500
100
r: number of clusters
(a) Design with Q1
300
400
500
(b) Design with Q2
kg(s)−ĝ(s)kH2
kg(s)kH2
nodes), while coherent clusters are decided by both edgeweights and node-weights. The H2 open-loop clusters are
shown in Fig. 4b. As such, they do not follow any definite
pattern with respect to the spatial clusters as they are based
on node aggregation following from the H2 -norm distance of
their output responses. Figures 4c and 4d, on the other hand,
show the clusters identified by our H2 closed-loop algorithm
(Algorithm 1) for Q = Q1 and Q = Q2 , respectively. Both
of these clusters are different from each other for obvious
reasons. They are also different from the spatial clusters,
the coherent clusters as well as the H2 open-loop clusters
as Algorithm 1 is related to the closed-loop controllability
subspace.
We also illustrate the effectiveness of H2 closed-loop clustering with respect to the number of clusters r. As evident
from the design, the error between the transfer matrices in
(11) and (12) will be minimal when r = n, and will degrade
with decreasing r while improving tractability of the design.
kg(s)−ĝ(s)k
We vary r from 1 to 500, and calculate the ratio kg(s)kH H2
2
resulting from the three clustering algorithms. The results are
shown in Fig. 5. For both Q1 and Q2 the closed-loop clustering
outperforms the other two methods in approaching the H2
performance of g(s). Therefore, even for very small values
of r, the projected controller achieves significantly close H2
performance as the full-order LQR controller. In terms of
implementation, the projected controller needs far less number
of communication links than a full-order standard LQR as well
as a full-order L1 sparsity-promoting LQR [14]. For
example,
for this system a standard LQR would require 500
= 124750
2
links. Meanwhile as shown in Fig. 6, a sparsity promoting
LQR requires from 3104 to as many as 21325 links to retain
a performance loss under 5%. By choosing r ≤ 9, the similar
performance loss can be maintained by our design using at
most 536 links.
It is also noted that the H2 closed-loop clusters do not
need to strictly follow the spatial geometric clustering of
the network. For example, in both Figures (4c) and (4d), a
cluster can be one single node as shown by the red, or can
be scattered over the network such as yellow. In practice, this
means that to implement the proposed control law, nodes from
with respect to the number of clusters
0.05
Sparsity-promoting LQR
Reduced-order LQR
0.04
kg(s)−ĝ(s)kH2
kg(s)kH2
Figure 5: Performance loss
200
r: number of clusters
0.03
0.02
0.01
0
506
515
528 536 3104
11661
11989
15003
21325
Number of communication links
Figure 6: Projected LQR controller from clustering vs. LQR controller from sparsity-promoting algorithm with Q1 .
different geographical locations may need to be part of the
same cluster for the closed-loop model, i.e., nodes that belong
to two different spatial clusters in open-loop may need to
collaborate and send their states to a common coordinator. The
assignment, therefore, encourages system-wide participation
from nodes at various corners of G500 for implementing the
controller.
B. Weight Design
We next apply Algorithm 2 on G500 to find the optimal
projection weight w while fixing the clusters to those obtained
from coherency. These clusters as previously shown in Fig. 4a
closely resemble the spatial clusters, and their clustering sets
are given by I1 = {1, ..., 167}, I2 = {168, ..., 178}, I3 =
{179, ..., 344}, I4 = {345, ..., 379}, I5 = {380, ..., 463} and
I6 = {464, ..., 500}. After running the algorithm with both
Q1 and Q2 , we plot the two weight vectors compared with
v̄ = v̄(A) in Fig. 7. It can be seen that the weight vectors from
Q1 and Q2 are very different than v̄ or between themselves.
On the other hand, both of these weights at the 364th node,
i.e., the node where the disturbance enters, show a sudden
jump in magnitude from the rest of the nodes. To verify the
closed-loop performance, we construct the P matrices using
these two vectors, and summarize the error ratios with some
design parameters in Table II. As expected, by applying the
weight design, the closed-loop errors as shown in Table II are
significantly reduced from w = v̄. Despite the fact that these
13
Table II: Results of weight design
Case
Penalty factor ρ
Q1
2
0.011 v̄T Qv̄
Q2
2
0.007 v̄kΦk
T Qv̄
Relative error kg(s) − ĝ(s)kH2 /kg(s)kH2
with w = v̄
with w from Algorithm 2
29.14%
7.35%
kΦk
35.69%
21.79%
1
The convergence threshold for power iteration is chosen as δ = 0.05.
is included in the penalty factor to normalize the two objective
functions in (opt*) to the same scale.
2 kΦk2
v̄ T Qv̄
Table III. The table verifies the O(n3 ) complexity for fullorder LQR vs. the O(n2 r) complexity for reduced-order LQR.
When n = 8000, the full-order LQR is already beyond the
capability of our computation facility. The reduced-order LQR
design, however, requires remarkably less computation time,
while still providing a close performance match to the fullorder LQR controller.
Table III: Scalability Results
w from Q1
w from Q2
v̄
n
Node 364
1000
2000
4000
6000
8000
10000
Computation time
Full-order
Reduced-order
16.34 sec
0.79 sec
134.91 sec
2.64 sec
20.36 min
11.61 sec
71.06 min
24.73 sec
Out of memory
48.39 sec
Out of memory
75.90 sec
Perf. loss
kg(s)−ĝ(s)kH2
kg(s)kH2
3.74%
1.06%
0.25%
0.13%
-
VII. C ONCLUSION
ï
Cluster 1
ï
Cluster 2
Cluster 3
Cluster 4 Cluster 5 Cluster 6
Figure 7: Weight designs from Q1 and Q2
two errors are much larger than what we get from closedloop clustering (which yields an error of 0.68%), the weight
design still grants us with significant improvement over the
hard constraint on I.
Finally, we compare the closed-loop performance of the
iterative Algorithm 3 (where both w and I are free) with
Algorithm 1 in Fig. 8. The comparison is shown for Q1 and
only r ≤ 6 as the error ratio already becomes under 1% after
r = 6. For this example, it is worth mentioning that Algorithm
3 turns out to be surprisingly efficient as it converges right after
the first iteration. In this sense, the iterative process reduces
to a single weight design after the clustering. Fig. 8 verifies
that Algorithm 3 achieves better matching between g(s) and
ĝ(s) than Algorithm 1.
C. Scalability Results
To verify the scalability of Algorithm 1, we increase the
size of the network, and compare the computation time with
that of solving a full-order LQR controller. We let r = 5
1
and κ = 5 for computing Φκ2 , and present the results in
0.25
v̄
w
0.2
In this paper we developed a set of projection-based algorithms that improve the dynamic response of large-scale
network systems with reduced-order LQR controllers. The
advantage of these reduced-order controllers is that they are
structured, and significantly easier to design and implement
compared to regular full-order LQR controllers. Our future
work will be to address the robustness of this approach
to communication delays, to exploit additional input-output
properties such as passivity to further improve performance,
and to inspect the influence of network heterogeneity on
clustering.
A PPENDIX A: S PECIAL C ASE
The well-posedness conditions from Section III.A can be
relaxed if system (1) is a consensus network. Consensus is
commonly used to model the dynamic behavior of many
practical networks such as social networks, power networks,
and wireless networks. For the same network graph G = (V, E)
as in (1), we suppose n = ns . Each node has a real-valued
node-weight mi > 0, and each edge ij ∈ E has a real-valued
edge-weight aij = aji > 0. A consensus network can then be
modeled in its
simplest form by letting Aij = m−1
i aij and
−1 P
a
in
(2).
The
overall
system
writes as
Aii = − mi
ij
j∈Ni
ẋM (t) = M −1 L(G)xM (t) + BM u(t) + BdM d(t),
where L(G) is the edge-weighted graph Laplacian matrix, and
M = diag([m1 , ..., mn ]) is the matrix of node weights. For
ease of analysis, let us consider a coordinate transformation
1
x = M 2 xM , in which case the state matrices in (1) become
1
kg(s)−ĝ(s)k H2
kg(s)k H2
C ONSENSUS N ETWORK
1
1
1
A = M − 2 L(G)M − 2 , B = M 2 BM , Bd = M 2 BdM . (40)
0.15
Notice that A here is a negative-semidefinite matrix, with only
1
one zero eigenvalue at Av̄(A) = 0, v̄(A) = √ 1 M 2 1n ,
0.1
tr(M)
0.05
0
OF
2
3
4
r: number of clusters
5
Figure 8: Performance loss by weight refinement
6
which we will simply denote as v̄. The state matrix à from
the reduced-order system (6) satisfies the following property.
Proposition A.1: Ã is negative-semidefinite (or -definite) if
w ∈ span(v̄) (or w 6∈ span(v̄)).
14
Proof: Denote the complement of the projection matrix
P by U , such that D = [P T U T ]T is unitary, i.e. DDT = In .
Then, Ã = P AP T is a leading principal of the matrix
P AP T P AU T
T
DAD =
,
U AP T U AU T
which is similar to A. Given the leading principal of a
symmetric positive-definite (or-semidefinite) matrix still being
positive-definite (or-semidefinite), it follows that à 0.
Moreover, from Definition 2.2, P is defined over w such that
P T P w = w. Then if w ∈ span(v̄), Ã preserves the zero
eigenvalue from A since ÃP v̄ = P Av̄ = 0. Therefore, Ã
is negative-semidefinite if w ∈ span(v̄), or negative-definite
otherwise.
1) Existence Condition: The next theorem shows that the
existence of X̃ for the reduced-order ARE (8) is guaranteed
for any w and I, and matrices Q̃ = P QP T and G̃ = P GP T .
1
T
Theorem A.2: If (Q 2 , A) is detectable and (A, G 2 ) is
stabilizable, then (8) is guaranteed with a unique stabilizing
solution X̃ = X̃ T 0.
Proof: For Q 0 and G 0, similar to à in the proof
of Proposition A.1, we have both Q̃ 0 and G̃ 0. Recall
T
that (Q̃ 2 , Ã) is detectable if and only if for all λ and v that
T
Ãv = λv and λ ≥ 0, Q̃ 2 v 6= 0. If w ∈ span(v̄), Ã 0 has
only one zero eigenvalue, with the corresponding eigenvector
P v̄. Under this situation, v̄ T P T Q̃P v̄ = v̄ T Qv̄ 6= 0 since
T
(Q 2 , A) is detectable. If w 6∈ span(v̄), Ã is negative-definite,
T
which means (Q̃ 2 , Ã) is trivially detectable. The same rational
1
applies to showing stabilizability of (Ã, G̃ 2 ). Therefore, both
stabilizability and detectability are satisfied, and thus (8)
guarantees a unique positive-semidefinite solution.
2) Bound of ARE solution: The following lemma provides
a specific value for β(A, G, Q) in Lemma 3.1.
Lemma A.3: The bound λ̄(X) ≤ β(A, G, Q) holds for
β(A, G, Q) =
σ̄(Q + G)
.
2σ(G − A)
(41)
Proof: The expression for β(A, G, Q) above can be
T
simply obtained by letting Kt = −G 2 , Dt = In and
F = 2G − 2A ≻ 0 in Lemma 3.1.
Note that the value specified by this Lemma only involves
finding two extreme singular values, which can be computed
very efficiently by Lanczos algorithm in O(n) complexity.
3) Stability condition: We next state two sufficient conditions that guarantee the stability of ĝ(s) for consensus
networks.
Theorem A.4: The system ĝ(s) is stable if w = v̄, and I
is an almost equitable partition [25] of the graph G, which
means for k 6= l, the edge weight aij is equal for all i ∈ Ik
and j ∈ Il .
Proof: A similarity transformation of D = [P T U T ]T
and DT on A − GX̂ yields
à − G̃X̃
P AU T
. (42)
D(A − GX̂)DT =
U AP T − U GP T X̃ U AU T
From [25], if G admits an almost equitable partition I, the
corresponding matrix P with w = v̄ will satisfy AT P T =
P T N for some N ∈ Rr×r . As a result, P AU T = 0, and thus
ĝ(s) is stable since à − G̃X̃ and U AU T are both Hurwitz
given w = v̄.
Theorem A.5: Assume that B is a square invertible matrix.
Then, ĝ(s) is stable if G is similar to αIn for some α > 0.
Proof: Denote the right eigenspace of G by V . If G is
similar to αIn , it follows that V T (GX̂ + X̂G)V = 2αV T X̂V .
This means that GX̂ + X̂G is positive-semidefinite. From the
matrix majorization property [26], we also have
2λ̄(A − GX̂) ≤ z T (A + AT − GX̂ − X̂G)z
hold for any non-zero vector z, with the RHS being nonpositive given A + AT 0 and −GX̂ − X̂G 0. Hence
for λ̄(A − GX̂) to be strictly negative, z T Az = 0 should
contradict z T GX̂z = 0. To prove the contradiction, we assume
z T Az = 0, namely z ∈ span(v̄). Note that z T GX̂z = 0 holds
if and only if either one of the following three conditions
is satisfied: (1) z ∈ ker(P ), (2) z ∈ ker(G), or (3)
P z ∈ ker(X̃). First of all, Assumption 3.2 requires wITi v̄Ii ,
i = 1, ..., r, which means P v̄ 6= 0 and thus z 6∈ ker(P ).
1
Given that (A, BR− 2 ) is stabilizable, we have v̄ T Gv̄ 6= 0,
i.e. z 6∈ ker(G). To this end, the last condition remaining is
z ∈ span(v̄) 6⇔ P z ∈ ker(X̃) to complete the contradiction.
By assuming a non-zero vector v ∈ ker(X̃), pre- and postmultiplying (8) with v yields
v T ÃT X̃v + v T X̃ Ãv + v T Q̃v − v T X̃ G̃X̃v = 0.
It can be easily verified that ker(X̃) is an Ã-invariant subspace
contained in the null-space of Q̃. Given z ∈ span(v̄), from
the proof of Proposition A.1 we know that P z is Ã-invariant,
i.e., P z is an eigenvector of Ã, when w ∈ span(v̄). As a
T
result, z T P T Q̃P z = z T Qz 6= 0 since (Q 2 , A) is detectable.
This verifies that P z 6∈ ker(Q̃), which proves P z 6∈ ker(X̃).
Therefore, we conclude that z T Az = 0 contradicts z T (GX̂ +
X̂G)z = 0, and thus ĝ(s) is stable.
A PPENDIX B: P ROOFS
A. Proof of Theorem 3.3
Consider a Lyapunov function V (x) = xT Xx > 0, where
X ≻ 0 is the solution of ARE (4). For ĝ(s) to be asymptotically stable, V̇ (x) needs to be negative, or equivalently
(A − GX̂)T X + X(A − GX̂) ≺ 0.
Using the ARE (4), the LMI above reduces to
Q ≻ XGE + EGX − XGX,
which holds if λ(Q) > λ̄(XGE + EGX − XGX). Notice
that the RHS of the eigenvalue inequality follows
λ̄(XGE +EGX −XGX) ≤ λ̄(XGE +EGX)+ λ̄(−XGX),
where we have respectively λ̄(XGE + EGX)
≤
σ̄(XGE + EGX) ≤ 2σ̄(XGE) ≤ 2σ̄(X)σ̄(G)σ̄(E),
and −λ̄(−XGX) = σ(XGX) ≥ σ(X)2 σ(G). Incorporating
these two upper bounds yields the condition in (17).
15
B. Proof of Lemma 3.4
From Lemma 3.1, the ARE solution X is bounded by
σ̄(X) = λ̄(X) ≤ β(A, G, Q). Note that (17) is satisfied if
σ(Q) > 2σ̄(X)σ̄(G)σ̄(E),
(43)
The eigenvalues of B ⊗ A are counted by λi (A)λj (B) with
i, j = 1, ...n, and according to the Weyl’s inequality of
eigenvalues [26], we have the lower bound for σ 2 (L) as
σ 2 (L) ≥ λ(AAT ) + λ(B T B) + 2λ(B ⊗ A)
= σ 2 (A) + σ 2 (B) + 2λ̄(A)λ̄(B) ≥ σ 2 (B).
where the RHS is further bounded by
2σ̄(X)σ̄(G)σ̄(E) ≤ 2β(A, G, Q)σ̄(G)[β(A, G, Q)+σ̄(X̃)]. (44)
Therefore, (43), and then (17) will hold if σ(Q) is greater than
the RHS of (44), which yields (18) in Lemma 3.4.
Combining (52) with (53) then yields the following bound
where ǫ1 =
T
The bound in (23) is derived assuming the worst case from
(14), where Q̃=P QP T +αIr and G̃=P QP T +αIr for α > 0.
We divide the proof into three steps.
1) : We derive an analytical expression for E by recovering
the reduced-order ARE (8) to the full dimension as
P T (ÃT X̃ + X̃ Ã + Q̃ − X̃ G̃X̃)P = 0.
(45)
AT X̂ + X̂A + Q − X̂GX̂ = R,
(47)
with the residue of the approximate ARE denoted by
T
T
T
R := αX̂ + U U A X̂ + X̂AU U + Q − P Q̃P. (48)
By subtracting (4) from (47), we get the Sylvester equation
(A − GX̂)T E + E(A − GX) = −R.
(49)
From (49), we are able to explicitly write E as a function of
A − GX, A − GX̂ and R, and hence obtain an initial bound
1
for kEΦ 2 kF in the next step.
T
1
2) : Pre- and post-multiplying (49) with Φ 2 and Φ 2
respectively, the Sylvester equation takes the form
T
T
1
1
A1 Φ 2 EΦ 2 + Φ 2 EΦ 2 B1 = −Φ 2 RΦ 2 ,
T
(50)
T
where we use the notations A1 = Φ 2 (A − GX̂)T Φ− 2 and
1
1
B1 = Φ− 2 (A−GX)Φ 2 for brevity. It can be easily shown that
λ(A1 ) < 0 and λ(B1 ) < 0, which implies λi (A) + λj (B) 6= 0
for any i, j = 1, ..., n so that (50) is solvable. Therefore, the
1
weighted error EΦ 2 can be expressed as
T
1
T
1
EΦ 2 = Φ− 2 · unvec[L−1 · vec(−Φ 2 RΦ 2 )],
(51)
where L = In ⊗ A + B T ⊗ In is an n2 × n2 matrix. Since
the Frobenius norm is unitary invariant, taking norm on both
1
sides of (51) provides an upper bound on kEΦ 2 kF as
1
T
1
1
]
(54)
> 0 is independent of P . In
T
1
T
1
T
1
T
1
kEΦ 2 kF ≤ σ̄(L−1 )σ̄(Φ− 2 )kΦ 2 RΦ 2 kF .
Note that σ̄(L−1 ) follows σ̄(L−1 ) =
calculated by σ 2 (L) = λ(LLT ) with
1
σ(L) ,
(52)
where σ(L) is
LLT = In ⊗ AAT + B T B ⊗ In + B T ⊗ AT + B ⊗ A.
1
T
1
+ Φ 2 (QU T U + U T U Q − U T U QU T U − αP T P )Φ 2 , (55)
contains the inexplicit functional X̂. We then bypass this term
in the final step.
3) : Taking norm on both sides of (55), and then isolating
the norm of X̂, we can form the bound
1
1
1
+ σ̄(Q)ξ 2 + ασ̄(Φ)[σ̄ 2 (X̂) + 1],
(46)
where U is the complement of P . Thereby substituting ÃP
and P T ÃT in terms of (46), and after a few calculations, (45)
yields the approximated ARE (for details, please see [30])
1
1
σ̄(Φ− 2 )
1
−1
σ[Φ 2 (A−GX)Φ 2
Φ 2 RΦ 2 = αΦ 2 X̂ 2 Φ 2 + Φ 2 U T U AT X̂Φ 2 + Φ 2 X̂AU T U Φ 2
T
ÃP = P A − P AU T U,
T
1
kΦ 2 RΦ 2 kF ≤2[σ̄(A)σ̄(Φ 2 )σ̄(X̂) + σ̄(QΦ 2 )]ξ
Notice that A and à are related by
T
T
1
kEΦ 2 kF ≤ ǫ1 kΦ 2 RΦ 2 kF ,
(54), the norm of the weighted residue Φ 2 RΦ 2 , written by
C. Proof of Theorem 3.5
2
(53)
(56)
1
2
with ξ = kU T U Φ kF . Recall that σ̄(X̂) = σ̄(P T X̃P ) =
σ̄(X̃), where X̃ is the solution of the reduced-order ARE
(8). The norm σ̄(X̃) can be further bounded by β(Ã, G̃, Q̃)
through Lemma 3.1. Theorem 3.5, therefore, follows from
(54), (56) and σ̄(X̃) ≤ β(Ã, G̃, Q̃) ≤ supP β(Ã, G̃, Q̃).
D. Proof of Lemma 3.7
To prove the error bound (28), we define a matrix Φ̄ as
Φ̄ = Φ̄1 · · · Φ̄nb ,
(57)
1
where Φ̄i = Y diag(Y −1 bi )C 2 , i = 1, ..., nb and bi is the ith
1
T
column of Bd . The matrix Φ̄ satisfies Φ = Φ̄Φ̄T = Φ 2 Φ 2 .
Besides Y1 and Ω1 defined in Definition 3.6, we further denote
−1
Y2 = Y:,κ+1,n and Ω2 = Yκ+1:n,:
, and partition the Cholesky
1
Decomposition C 2 as
# " 1
#
" 1
2
2
1
0
0
C1:κ,1:κ
C1,1
2
=
.
C =
1
1
1
1
2
2
2
2
C2,1
Cκ+1:n,1:κ
C2,2
Cκ+1:n,κ+1:n
With these notations, Φ̄i in (57)
h can be decomposed
i into
1
2
Φ̄i = Φ̄i,s + Φ̄i,f , where Φ̄i,s = Y1 diag(Ω1 bi )C1,1
0 and
h
i
1
1
2
2
Φ̄i,f = Y2 diag(Ω2 bi )C2,1
, and thus Φκ
Y2 diag(Ω2 bi )C2,2
Pnb
can be rewritten as Φκ = i=1
Φ̄i,s Φ̄Ti,s . Notice that ξ =
1
k(In − P T P )Φ 2 kF = k(In − P T P )Φ̄kF satisfies
1
ξκ∗ ≤ ξ ≤ k(In − P T P )Φκ2 kF + k(In − P T P )Φ̄f kF , (58)
where Φ̄f = Φ̄1,f · · · Φ̄nb ,f . The second norm on the
RHS of (58) is further bounded by k(In − P T P )Φ̄f kF ≤
kΦ̄f kF with
v
v
u nb
u
n
X
u X
1
1
u
1
2
2
2
2
t
2
− −.
kΦ̄f kF ≤ η
(kC2,1 kF + kC2,2 kF ) = tη 2 nb
2λ
i
i=κ+1
i=1
Inserting this along with P = argminP ξκ to the RHS of (58)
yields the error bound (28).
16
E. Proof of Lemma 5.3
Denote ni = |Ii |c , i = 1, ..., r, the objective function fV̄
in (opt2) can be expanded as
fV̄ = tr(V̄ T P T P QP T P V̄ ) = (w ∗ w)T (Q ∗ V̄ V̄ T )(w ∗ w)
where ∗ is the Khatri-Rao product defined by
QI1 ,I1 ⊗ V̄I1 ,: V̄IT1 ,:
..
Q ∗ V̄ V̄ T =
.
QIr ,I1 ⊗ V̄Ir ,: V̄IT1 ,:
w ∗ w = wIT1 ⊗ wIT1
···
..
.
···
···
QI1 ,Ir ⊗ V̄I1 ,: V̄ITr ,:
..
,
.
QIr ,Ir ⊗ V̄Ir ,: V̄ITr ,:
wITr ⊗ wITr
T
.
Denote the block-diagonal submatrix of Q by Qd , i.e. Qd =
diag(QI1 ,I1 , · · · , QIr ,Ir ), and the off-diagonal by Qo = Q −
Qd . Therefore, we can find the difference between objective
functions of (opt3) and (opt*) as (w∗w)T (ρQo ∗ V̄ V̄ T )(w∗w).
According to the Gershgorin circle theorem, the eigenvalues
of ρQo ∗ V̄ V̄ T are all bounded inside the range of (−kρQo ∗
V̄ V̄ T k1 , kρQo ∗V̄ V̄ T k1 ). Therefore, the theorem follows when
Je = kρQo ∗ V̄ V̄ T k1 .
R EFERENCES
[1] D. D. Siljak, Decentralized Control of Complex Systems, Courier Corporation, 2011.
[2] P. Kokotovic, H. K. Khalil, and J. O’reilly, Singular Perturbation
Methods in Control: Analysis and Design, SIAM, 1999.
[3] J. Chow and P. Kokotovic, “Time Scale Modeling of Sparse Dynamic
Networks,” IEEE Transactions on Automatic Control, vol. 30(8), pp.
714-722, 1985.
[4] B. Moore, “Principal Component Analysis in Linear Systems: Controllability, Observability, and Model Reduction,” IEEE Transactions on
Automatic Control, vol. 26(1), pp. 17-32, 1981.
[5] G. Obinata and B. D. Anderson, Model Reduction for Control System
Design, Springer Science & Business Media, 2012.
[6] A. C. Antoulas, Approximation of Large-Scale Dynamical Systems,
SIAM, 2005.
[7] K. Zhou and J. C. Doyle, Essentials of Robust Control, vol. 180, Prentice
Hall Upper Saddle River, NJ, 1998.
[8] A. M. Boker, T. R. Nudell, and A. Chakrabortty, “On Aggregate Control
of Clustered Consensus Networks,” 2015 American Control Conference
(ACC), pp. 5527-5532, Chicago, IL, USA, 2015.
[9] S. Hara, J. I. Imura, K. Tsumura, T. Ishizaki, and T. Sadamoto, “Glocal
(Global/Local) Control Synthesis for Hierarchical Networked Systems,”
2015 IEEE Conference on Control Applications (CCA), pp. 107-112,
Sydney, NSW, 2015.
[10] T. Sadamoto, T. Ishizaki, and J. I. Imura, “Hierarchical Distributed
Control for Networked Linear Systems,” 53rd IEEE Conference on
Decision and Control, pp. 2447-2452, 2014.
[11] D. Madjidian and L. Mirkin, “Distributed Control with Low-Rank
Coordination,” IEEE Transactions on Control of Network Systems, vol.
1(1), pp. 53-63, March 2014.
[12] S. Lloyd, “Least Squares Quantization in PCM,” IEEE Transactions on
Information Theory, vol. 28(2), pp. 129-137, Mar 1982.
[13] L. Qi, F. Wang, and Y. Wang, “Z-Eigenvalue Methods for a Global
Polynomial Optimization Problem,” Mathematical Programming, vol.
118, pp. 301-316, 2009.
[14] F. Lin, M. Fardad, and M. R. Jovanovic, “Design of Optimal Sparse
Feedback Gains via the Alternating Direction Method of Multipliers,”
IEEE Transactions on Automatic Control, vol. 58(9), pp. 2426-2431,
Sep 2013.
[15] N. Monshizadeh, H. L. Trentelman, and M. K. Camlibel, “ProjectionBased Model Reduction of Multi-Agent Systems using Graph Partitions,” IEEE Transactions on Control of Network Systems, vol. 1(2),
pp. 145-154, Jun 2014.
[16] T. Ishizaki, K. Kashima, A. Girard, J. Imura, L. Chen, and K. Aihara,
“Clustered Model Reduction of Positive Directed Networks,” Automatica, vol. 59, pp. 238-247, 2015.
[17] P. Benner and Z. Bujanovi, “On the Solution of Large-Scale Algebraic
Riccati Equations by Using Low-Dimensional Invariant Subspaces,”
Linear Algebra and its Applications, vol. 488, pp. 430-459, 2016.
[18] M. Rotkowitz and S. Lall, “A Characterization of Convex Problems in
Decentralized Control,” IEEE Transactions on Automatic Control, vol.
51(2), pp. 274-286, Feb. 2006.
[19] X. Wu and M. R. Jovanovic, “Augmented Lagrangian Approach to
Design of Structured Optimal State Feedback Gains,” IEEE Transactions
on Automatic Control, vol. 56, pp. 2923-2929, Dec 2011.
[20] S. Fattahi, G. Fazelnia and J. Lavaei, “Transformation of Optimal
Centralized Controllers into Near-Global Static Distributed Controllers,”
53rd IEEE Conference on Decision and Control, pp. 4915-4922, 2015.
[21] Y. S. Wang, N. Matni, and J. C. Doyle, “Localized LQR Optimal
Control,” 53rd IEEE Conference on Decision and Control, pp. 16611668, 2014.
[22] Y. S. Wang and N. Matni, “Localized LQG Optimal Control for LargeScale Systems,” 2016 American Control Conference (ACC), pp. 19541961, Boston, MA, USA, 2016.
[23] N. Xue and A. Chakrabortty, “H2 -Clustering of Closed-Loop Consensus
Networks under Generalized LQR Designs,” 55th IEEE Conference on
Decision and Control, pp. 5116-5121, Las Vegas, NV, Dec 2016.
[24] A. Helmy, “Architectural Framework for Large-Scale Multicast in Mobile Ad Hoc Networks,” IEEE International Conference on Communications, pp. 2036-2042, vol.4, 2002.
[25] D. M. Cardoso, C. Delorme, and P. Rama, “Laplacian Eigenvectors
and Eigenvalues and Almost Equitable Partitions,” European Journal
of Combinatorics, vol. 28, pp. 665-673, 2007.
[26] J. N. Franklin, Matrix Theory, Courier Corporation, 2012.
[27] R. Davies, P. Shi, and R. Wiltshire, “New Lower Solution Bounds of
the Continuous Algebraic Riccati Matrix Equation,” Linear Algebra and
Its Applications, vol. 427(2-3), pp. 242-255, 2007.
[28] G. H. Golub and C. F. Van Loan, Matrix Computations, JHU Press, vol.
3, 2012.
[29] M. S. Traiantafyllou and F. S. Hover, Maneuvering and Control of
Marine Vehicles, Cambridge, MA, 2002.
[30] J. Sun, “Perturbation Theory for Algebraic Riccati Equations,” SIAM
Journal on Matrix Analysis and Applications, vol. 19, pp. 39-65, 1998.
[31] E. Kofidis and P. A. Regalia, “On the Best Rank-1 Approximation
of Higher-Order Supersymmetric Tensors,” SIAM Journal on Matrix
Analysis and Applications, vol. 23, pp. 863-884, 2002.
Nan Xue (S’15) received his B.E. degree in Electrical Engineering from Xi’an
Jiaotong University, China in 2013. He is currently pursuing his PhD degree
in Electrical Engineering at North Carolina State University, Raleigh, NC. His
research interests include analysis, control and model reduction of large-scale
networked dynamic systems and power systems.
Aranya Chakrabortty (M’08, SM’15) received his PhD degree in Electrical
Engineering from Rensselaer Polytechnic Institute, Troy, NY in 2008. From
2008 to 2009 he was a postdoctoral research associate at the Aeronautics and
Astronautics department of University of Washington, Seattle. He is currently
an Associate Professor in the Electrical and Computer Engineering department
of North Carolina State University, Raleigh, NC, where he is also affiliated
to the FREEDM Systems Center. His research interests are in all branches
of control theory with applications to power systems, especially in wide-area
monitoring and control of large power systems using Synchrophasors. He
received the NSF CAREER award in 2011.
| 3cs.SY
|
1
Diagnose like a Radiologist: Attention Guided
Convolutional Neural Network for Thorax Disease
Classification
arXiv:1801.09927v1 [cs.CV] 30 Jan 2018
Qingji Guan, Yaping Huang, Zhun Zhong, Zhedong Zheng, Liang Zheng and Yi Yang
Abstract—This paper considers the task of thorax disease classification on chest X-ray images. Existing methods generally use
the global image as input for network learning. Such a strategy
is limited in two aspects. 1) A thorax disease usually happens
in (small) localized areas which are disease specific. Training
CNNs using global image may be affected by the (excessive)
irrelevant noisy areas. 2) Due to the poor alignment of some CXR
images, the existence of irregular borders hinders the network
performance. In this paper, we address the above problems by
proposing a three-branch attention guided convolution neural
network (AG-CNN). AG-CNN 1) learns from disease-specific
regions to avoid noise and improve alignment, 2) also integrates
a global branch to compensate the lost discriminative cues by
local branch. Specifically, we first learn a global CNN branch
using global images. Then, guided by the attention heat map
generated from the global branch, we inference a mask to crop
a discriminative region from the global image. The local region
is used for training a local CNN branch. Lastly, we concatenate
the last pooling layers of both the global and local branches for
fine-tuning the fusion branch. The comprehensive experiment is
conducted on the ChestX-ray14 dataset. We first report a strong
global baseline producing an average AUC of 0.841 with ResNet50 as backbone. After combining the local cues with the global
information, AG-CNN improves the average AUC to 0.868. While
DenseNet-121 is used, the average AUC achieves 0.871, which is
a new state of the art in the community.
Index Terms—chest X-ray, convolutional neural network, thorax disease classification, visual attention
I. I NTRODUCTION
T
HE chest X-ray (CXR) has been one of the most
common radiological examinations in lung and heart
disease diagnosis. Currently, reading CXRs mainly relies on
professional knowledge and careful manual observation. Due
to the complex pathologies and subtle texture changes of different lung lesion in images, radiologists may make mistakes
even when they have experienced long-term clinical training
and professional guidance. Therefore, it is of importance to
develop the CXR image classification methods to support
Q. Guan is with the Beijing Key Laboratory of Traffic Data Analysis
and Mining, Beijing Jiaotong University, Beijing, 100044, China and Center
for Artificial Intelligence, University of Technology Sydney, NSW, 2007,
Australia (E-mail: [email protected]).
Y. Huang is with the Beijing Key Laboratory of Traffic Data Analysis
and Mining, Beijing Jiaotong University, Beijing, 100044, China (E-mail:
[email protected]).
Z. Zhong is with Cognitive Science Department, Xiamen University and
Center for Artificial Intelligence, University of Technology Sydney, NSW,
2007, Australia (E-mail: [email protected]).
Z. Zheng, L. Zheng, Y. Yang are with Center for Artificial Intelligence,
University of Technology Sydney, NSW, 2007, Australia (E-mail: {zdzheng12,
liangzheng06, yee.i.yang}@gmail.com).
clinical practitioners. The noticeable progress in deep learning
has benefited many trials in medical image analysis, such as
lesion segmentation or detection [1], [2], [3], [4], [5], diseases
classification [6], [7], [8], [9], noise induction [10], image
annotation [11], [12], registration [13], regression [14] and so
on. In this paper, we investigate the CXR classification task
using deep learning.
Several existing works on CXR classification typically employ the global image for training. For example,
Wang et al. [9] evaluate four classic CNN architectures,
i.e., AlexNet [15], VGGNet [16], GoogLeNet [17], ResNet
[18], to tell the presence of multiple pathologies using a
global CXR image. In addition, using the same network,
the disease lesion areas are located in a weakly supervised
manner. Viewing CXR classification as a multi-label recognition problem, Yao et al. [19] explore the correlation among
the 14 pathologic labels with global images in ChestX-ray14
[9]. Using a variant of DenseNet [20] as an image encoder,
they adopt the Long-short Term Memory Networks (LSTM)
[21] to capture the dependencies. Kumar et al. [7] investigate
that which loss function is more suitable for training CNNs
from scratch and present a boosted cascaded CNN for global
image classification. The recent effective method consists in
CheXNet [8]. It fine-tunes a 121-layer DenseNet on the global
chest X-ray images, which has a modified last fully-connected
layer.
However, the global learning strategy can be compromised
by two problems. On the one hand, as shown in Fig. 1 (the
first row), the lesion area can be very small (red bounding box)
and position unpredictable (e.g. , “Atelectasis”) compared with
the global image, so using the global image for classification
may include a considerable level of noise outside the lesion
area. This problem is rather different from generic image
classification [22], [23] where the object of interest is usually
positioned in the image center. Considering this fact, it is
beneficial to induce the network to focus on the lesion regions
when making predictions. On the other hand, due to the
variations of capturing condition, e.g. , the posture of the
patient and the small size of children body, the CXR images
may undergo distortion or misalignment. Fig. 1 (the second
row) presents a misalignment example. The irregular image
borders may exist an non-negligible effect on the classification
accuracy. Therefore, it is desirable to discover the salient lesion
regions and thus alleviate the impact of such misalignment.
To address the problems caused by merely relying on
the global CXR image, this paper introduces a three-branch
2
our method achieves superior performance over the stateof-the-art approaches.
II. R ELATED W ORKS
(a) original global image
(b) heatmap
(c) cropped local image
Fig. 1. Two training images from the ChestX-ray14 dataset. (a) The global
images. (b) Heat maps extracted from a specific convolutional layer. (c) The
cropped images from (a) guided by (b). In this paper, we consider both
the original global image and the cropped local image for classification, so
that 1) the noise contained in non-lesion area is less influencing, and 2) the
misalignment can be reduced. Note that there are some differences between
the global images and their heat maps. The reason is that the global images
are randomly cropped from 256×255 to 224×224 during training.
attention guided convolutional neural network (AG-CNN) to
classify the lung or heart diseases. AG-CNN is featured in two
aspects. First, it has a focus on the local lesion regions which
are disease specific. Generally, such a strategy is particularly
effective for diseases such as ”Nodule”, which has a small
lesion region. In this manner, the impact of the noise in nondisease regions and misalignment can be alleviated. Second,
AG-CNN has three branches, i.e., a global branch, a local
branch and a fusion branch. While the local branch exhibits the
attention mechanism, it may lead to information loss in cases
where the lesion areas are distributed in the whole images,
such as Pneumonia. Therefore, a global branch is needed to
compensate for this error. We show that the global and local
branches are complementary to each other and, once fused,
yield favorable accuracy to the state of the art.
The working mechanism of AG-CNN is similar to that of a
radiologist. We first learn a global branch that takes the global
image as input: a radiologist may first browse the whole CXR
image. Then, we discover and crop a local lesion region and
train a local branch: a radiologist will concentrate on the local
lesion area after the overall browse. Finally, the global and
local branches are fused to fine-tune the whole network: a
radiologist will comprehensively consider the global and local
information before making decisions.
Our contributions are summarized as follows.
•
•
•
We propose an attention guided convolutional neural
network (AG-CNN) which diagnoses thorax diseases by
combining the global and local information. AG-CNN
improves the recognition performance by correcting image alignment and reducing the impact of noise.
We introduce a CNN training baseline, which produces
competitive results to the state-of-the-art methods by
itself.
We present comprehensive experiment on the ChestXray14 dataset. The experiment results demonstrate that
Chest X-ray datasets. The problem of Chest X-ray image
classification has been extensively explored in the field of
medical image analysis. Several datasets have been released
in this context. For example, the JSRT dataset [24], [25]
contains 247 chest X-ray images including 154 lung nodules.
It also provides masks of the lung area for segmentation
performance evaluation. The Shenzhen chest X-ray set [26]
has a total of 662 images belonging to two categories (normal
and tuberculosis (TB)). Among them, 326 are normal cases
and 336 are cases with TB. The Montgomery County chest
X-ray set (MC) [26] collects 138 frontal chest X-ray images
from Montgomery Country’s Tuberculosis screen program, of
which 80 are normal and 58 are cases with manifestations of
TB. These three datasets are generally small for deep model
training. In comparison, the Indiana University Chest X-ray
Collection dataset [27] has of 3,955 radiology reports and
the corresponding 7,470 chest X-ray images. It is publicly
available through Open-I [28]. However, this dataset does not
provide explicit disease class labels, so we do not use it in this
paper. Recently, Wang et al. [9] released the ChestX-ray14
dataset, which is the largest chest X-ray dataset by far. ChestXray14 collects 112,120 frontal-view chest X-ray images of
30,805 unique patients. Each radiography is labeled with one
or more types of 14 common thorax diseases. This dataset
poses a multi-label classification problem and is large enough
for deep learning, so we adopt this dataset for performance
evaluation in this paper.
Deep learning for chest X-ray image analysis. Recent
surveys [29], [30], [31], [32] have demonstrated that deep
learning technologies have been extensively applied to the
field of chest X-ray image annotation [33], classification [6],
[34], [8], [9], and detection (localization) [35], [36]. Islam
et al. [34] explore different CNN architectures and find that a
single CNN does not perform well across all abnormalities.
Therefore, they leverage model ensemble to improve the
classification accuracy, at the cost of increased training and
testing time. Yao et al. [19] and Kumar et al. [7] classify the
chest X-ray images by investigating the potential dependencies
among the labels from the aspect of multi-label problems.
Rajpurkar et al. [8] train a convolutional neural network to
address the multi-label classification problem. This paper
departs from the previous methods in that we make use of the
attention mechanism and fuse the local and global information
to improve the classification performance.
Attention models in medical image analysis. The CXR
classification problem needs to tell the relatively subtle differences between different diseases. Usually, a disease is often
characterized by a lesion region, which contains critical dues
for classification. Ypsilantis et al. [37] explore where to look
in chest X-rays with recurrent attention model (RAM) [38].
The RAM learns to sample the entire X-ray image sequentially
and focus on informative areas. Only one disease enlarged
heart is considered in their work. Recently, Pesce et al. [39]
3
Input/224*224*3
112*112*64
56*56*256
pooling
Global Branch
28*28*512
14*14*1024
fc
7*7*2048
sigmoid loss
ResBlock
ResBlock
ResBlock
ResBlock
conv
B
C
E
Mask Inference
crop
X
[xmin, ymin, xmax, ymax]
mask
concat
heatmap
B
C
E
+
Fusion Branch
ResBlock
ResBlock
ResBlock
ResBlock
conv
B
C
E
Local Branch
Fig. 2. Overall framework of the attention guided convolutional neural network (AG-CNN). We show an example with ResNet-50 as backbone. AG-CNN
consists of three branches. Global and local branches consist of five convolutional blocks with batch normalization and ReLU. Each of them is then connected
to a max pooling layer (Pool5), a fully connected (FC) layer, and a sigmoid layer. Different from the global branch, the input of the local branch is a local
lesion patch which is cropped by the mask generated from global branch. Then, Pool5 layers of the these two branches are concatenated into the fusion
branch. ”BCE” represents binary cross entropy loss. The input image is added to the heat map for visualization.
explore a soft attention mechanism from the saliency map of
CNN features to locate lung nodule position in radiographies.
And a localization loss is calculated by comparing the predicted position with the annotated position.
In this paper, AG-CNN locates the salient regions with
an attention guided mask inference process, and learns the
discriminative feature for classification. Compared with the
method which relies on bounding box annotations, Our method
only need image-level labels without any extra information.
III. T HE P ROPOSED A PPROACH
In this section, we describe the proposed attention guided
convolutional neural network (AG-CNN) for thorax disease
classification. We will first illustrate the architecture of AGCNN in Section III-A. Second, we describe the mask inference
process for lesion region discovery in Section III-B. We then
present the training process of AG-CNN in Section III-C.
Finally, a brief discussion of the AG-CNN is provided.
A. Structure of AG-CNN
The architecture of AG-CNN is presented in Fig. 2. Basically, it has two major branches, i.e., the global and local
branches, and a fusion branch. Both the global and local
branches are classification networks that predict whether the
pathologies are present or not in the image. Given an image,
the global branch is first fine-tuned from a classification CNN
using the global image. Then, we crop an attended region
from the global image and train it for classification on the
local branch. Finally, the last pooling layers of both the global
and local branches are concatenated for fine-tuning the fusion
branch.
Multi-label setup. We label each image with a 15-dim
vector L = [l1 , l2 , ..., lC ] in which lc ∈ {0, 1}, C = 15.
lc represents whether the there is any pathology, i.e., 1 for
presence and 0 for absence. The last element of L represents
the label with ”No Finding”.
Global and local branches. The global branch informs the
underlying CXR information derived from the global image
as input. In the global branch, we train a variant of ResNet-50
[18] as the backbone model. It consists of five down-sampling
blocks, followed by a global max pooling layer and a 15dimensional fully connected (FC) layer for classification. At
last, a sigmoid layer is added to normalize the output vector
pg (c|I) of FC layer by
peg (c|I) = 1/(1 + exp(−p(c|I))),
(1)
where I is the global image. peg (c|I) represents the probability
score of I belonging to the cth class, c ∈ {1, 2, ..., C}. We
optimize the parameter Wg of global branch by minimizing
the binary cross-entropy (BCE) loss:
C
1 X
lc log(peg (c|I)) + (1 − lc )log(1 − peg (c|I)),
C c=1
(2)
where lc is the groundtruth label of the cth class, C is the
number of pathologies.
On the other hand, the local branch focuses on the lesion
area and is expected to alleviate the drawbacks of only using
the global image. In more details, the local branch possesses
the same convolutional network structure with the global
branch. Note that, these two branches do not share weights
since they have distinct purposes. We denote the probability
score of local branch as pel (c|Ic ), Wl as the parameters of
local branch. Here, Ic is the input image of local branch. We
perform the same normalization and optimization as the global
branch.
L(Wg ) = −
4
Fig. 3. The process of lesion area generation. (Top:) global CXR images of various thorax diseases for the global branch. The manually annotated legion
areas provided by [9] are annotated with red bounding boxes. Note that we do not use the bounding boxes for training or testing. (Middle:) corresponding
visual examples of the output of the mask inference process. The lesion areas are denoted by green bounding boxes. Higher response is denoted with red,
and lower blue. Note that the heat maps are resized to the same size as the input images. (Bottom:) cropped and resized images from the green bounding
boxes which are fed to the local branch.
Algorithm 1: Attention Guided CNN Procedure
Input: Input image I; Label vector L; Threshold τ .
Output: Probability score pf
f (c|[I, Ic ]).
Initialization: the global and local branch weights.
1 Learning Wg with I, computing peg (c|I), optimizing by
Eq. 2 (Stage I);
2 Computing mask M and the bounding box coordinates
[xmin , ymin , xmax , ymax ], cropping out Ic from I;
3 Learning Wl with Ic , computing p
el (c|Ic ), optimizing by
Eq. 2 (Stage II);
4 Concentrating P oolg and P ooll , learning Wf , computing
pf
f (c|[I, Ic ]), optimizing by Eq. 2.
Fusion branch. The fusion branch first concatenates the
Pool5 outputs of the global and local branches. The concatenated layer is connected to a 15-dimensional FC layer for final
classification. The probability score is pf
f (c|[I, Ic ]). We denote
Wf as the parameters of fusion branch and optimize Wf by
Eq. 2.
B. Attention Guided Mask Inference
In this paper, we construct a binary mask to locate the
discriminative regions for classification in the global image.
It is produced by performing thresholding operations on the
feature maps, which can be regarded as an attention process.
This process is described below.
Given a global image, let fgk (x, y) represent the activation
of spatial location (x, y) in the kth channel of the output of
the last convolutional layer, where k ∈ {1, ..., K}, K = 2, 048
in ResNet-50. g denotes the global branch. We first take the
absolute value of the activation values fgk (x, y) at position
(x, y). Then the attention heat map Hg is generated by
counting the maximum values along channels,
Hg (x, y) = max(|fgk (x, y)|), k ∈ {1, ..., K}.
k
(3)
The values in Hg directly indicate the importance of the
activations for classification. In Fig. 1(b) and Fig. 3 (the
second row), some examples of the heat maps are shown.
We observe that the discriminative regions (lesion areas)
of the images are activated. Heat map can be constructed
by computing different statistical values
the channel
PK across
1
k
dimensions,qsuch as L1 distance K
(x,
y)| or L2
|f
g
k=1
PK
1
k
2
distance K
k=1 (fg (x, y)) . Different statistics result in
subtle numerical differences in heat map, but may not effect
the classification significantly. Therefore, we computing heat
map with Eq. 3 in our experiment. The comparison of these
statistics is presented in Section IV-C.
We design a binary mask M to locate the regions with
large activation values. If the value of a certain spatial position
(x, y) in the heat map is larger than a threshold τ , the value
at corresponding position in the mask is assigned with 1.
Specifically,
(
1, Hg (x, y) > τ
M (x, y) =
(4)
0, otherwise
where τ is the threshold that controls the size of attended region. A larger τ leads to a smaller region, and vice versa. With
the mask M , we draw a maximum connected region that covers the discriminative points in M . The maximum connected
region is denoted as the minimum and maximum coordinates
in horizontal and vertical axis [xmin , ymin , xmax , ymax ]. At
last, the local discriminative region Ic is cropped from the
input image I and is resized to the same size as I. We visualize
the bounding boxes and cropped patches with τ = 0.7 in
Fig. 3. The attention informed mask inference method is
able to locate the regions (green bounding boxes) which are
reasonably close to the groundtruth (red bounding boxes).
C. Training Strategy of AG-CNN
This paper adopts a three-stage training scheme for AGCNN.
5
Atelectasis
Cardiomegaly
Effusion
Infiltrate
Mass
Nodule
Pneumonia
Pneumothorax
Fig. 4. Examples of 8 pathologies in ChestX-ray14. The lesion regions are annotated with the red bounding boxes provided by [9]. Note that these groundtruth
bounding boxes are only used for demonstration: they are neither used in training nor testing.
TABLE I
C OMPARISON RESULTS OF VARIOUS METHODS ON C HEST X- RAY 14.
Method
Wang et al. [9]
Yao et al. [19]
CNN
R-50
D-/
Rajpurkar et al. [8]∗ D-121
Kumar et al. [7]∗
D-161
Global branch (baseline) R-50
Local branch
R-50
AG-CNN
R-50
Global branch (baseline) D-121
Local branch
D-121
AG-CNN
D-121
*
Atel
0.716
0.772
0.821
0.762
0.818
0.798
0.844
0.832
0.797
0.853
Card
0.807
0.904
0.905
0.913
0.904
0.881
0.937
0.906
0.865
0.939
Effu
0.784
0.859
0.883
0.864
0.881
0.862
0.904
0.887
0.851
0.903
Infi
Mass
0.609 0.706
0.695 0.792
0.720 0.862
0.692 0.750
0.728 0.863
0.707 0.826
0.753 0.893
0.717 0.870
0.704 0.829
0.754 0.902
Nodu
0.671
0.717
0.777
0.666
0.780
0.736
0.827
0.791
0.733
0.828
Pne1
0.633
0.713
0.763
0.715
0.783
0.716
0.776
0.732
0.710
0.774
Pne2
0.806
0.841
0.893
0.859
0.897
0.872
0.919
0.891
0.850
0.921
Cons
0.708
0.788
0.794
0.784
0.807
0.805
0.842
0.808
0.802
0.842
Edem
0.835
0.882
0.893
0.888
0.892
0.874
0.919
0.905
0.882
0.924
Emph
0.815
0.829
0.926
0.898
0.918
0.898
0.941
0.912
0.874
0.932
Fibr
0.769
0.767
0.804
0.756
0.815
0.808
0.857
0.823
0.801
0.864
PT
0.708
0.765
0.814
0.774
0.800
0.770
0.836
0.802
0.769
0.837
Hern
0.767
0.914
0.939
0.802
0.889
0.887
0.903
0.883
0.872
0.921
Mean
0.738
0.803
0.842
0.795
0.841
0.817
0.868
0.840
0.810
0.871
We compute the AUC of each class and the average AUC across the 14 diseases. ∗ denotes that a different train/test split is used: 80% for training and the rest 20% for
testing. All the Other methods split the dataset with 70% for training, 10% for validation and 20% for testing. Each pathology is denoted with its first four characteristics,
e.g., Atelectasis with Atel. Pneumonia and Pneumothorax are denoted as Pneu1 and Pneu2, respectively. PT represents Pleural Thickening. We report the performance
with parameter τ = 0.7. ResNet-50 (R-50) and Desnet-121 (D-121) are used as backbones in our approach. For each column, the best and second best results are
highlighted in red and blue, respectively.
Stage I. Using the global images, we fine-tune the global
branch network pretrained by ImageNet. peg (c|I) is normalized
by Eq. 1.
Stage II. Once the local image Ic is obtained by mask
inference with threshold τ , we feed it into the local branch
for fine-tuning. pel (c|Ic ) is also normalized by Eq. 1. When
we fine-tune the local branch, the weights in the global branch
are fixed.
Stage III. Let P oolg and P ooll represent the Pool5 layer
outputs of the global and local branches, respectively. We
concatenate them for a final stage of fine-tuning and normalize
the probability score pf
f (c|[I, Ic ]) by Eq. 1. Similarly, the
weights of previous two branches are fixed when we fine-tune
the weights of fusion branch.
In each stage, we use the model with the highest AUC on
the validation set for testing. The overall AG-CNN training
procedure is presented in Algorithm 1. Variants of training
strategy may influence the performance of AG-CNN. We
discussed it in Section IV-C.
IV. E XPERIMENT
This section evaluates the performance of the proposed
AG-CNN. The experimental dataset, evaluation protocol and
the experimental settings are introduced first. Section IV-C
demonstrates the performance of global and local branches
and the effectiveness of fusing them. Furthermore, comparison
of AG-CNN and the state of the art is presented in Table. I.
In Section. IV-D, we analyze the parameter impact in mask
inference.
A. Dataset and Evaluation Protocol
Dataset. We evaluate the AG-CNN framework using the
ChestX-ray141 dataset [9]. ChestX-ray14 collects 112,120
frontal-view images of 30,805 unique patients. 51,708 images
of them are labeled with up to 14 pathologies, while the others
are labeled as “No Finding”. Fig. 4 presents some examples
of 8 out of 14 thorax diseases and the ground-truth bounding
boxes of the lesion regions provided by [9]. We observe that
the size of the lesion area varies a lot for different pathologies.
Evaluation protocol. In our experiment, we randomly
shuffle the dataset into three subsets: 70% for training, 10%
for validation and 20% for testing. Each image is labeled with
a 15-dim vector L = [l1 , l2 , ..., lC ] in which yc ∈ {0, 1}, C =
15. l15 represents the label with ”No Finding”.
B. Experimental Settings
For training (any of the three stages), we perform data
augmentation by resizing the original images to 256 × 256,
1 https://nihcc.app.box.com/v/ChestXray-NIHCC
6
Fig. 5. ROC curves of the global, local and fusion branches (DenseNet-121 as backbone) over the 14 pathologies. The corresponding AUC values are given
in Table. I. We observe that fusing global and local information yields clear improvement.
C. Evaluation
Fig. 6. ROC curves of AG-CNN on the 14 diseases (ResNet-50 and DenseNet121 as backbones, respectively).
randomly resized cropping to 224×224, and random horizontal
flipping. The ImageNet mean value is subtracted from the
image. When using ResNet-50 as backbone, we optimize
the network using SGD with a mini-batch size of 126, 64,
64 for global, local and fusion branch, respectively. But for
DenseNet-121, the network is optimized with a mini-batch
of 64, 32, and 32, respectively. We train each branch for 50
epochs. The learning rate starts from 0.01 and is divided by
10 after 20 epochs. We use a weight decay of 0.0001 and
a momentum of 0.9. During validation and testing, we also
resize the image to 256 × 256, and then perform center cropping to obtain an image of size 224 × 224. Except in Section
IV-D, we set τ to 0.7 which yields the best performance on
the validation set. We implement AG-CNN with the Pytorch
framework [40].
We evaluate our method on the ChestX-ray14 dataset.
Mostly, ResNet-50 [18] is used as backbone, but the AUC and
ROC curve obtained by DenseNet-121 [20] are also presented.
Global branch (baseline) performance. We first report the
performance of the baseline, i.e., the global branch. Results are
summarized in Table. I, Fig. 5 and Fig. 9.
The average AUC across the 14 thorax diseases arrives
at 0.841 and 0.840, using ResNet-50 and DenseNet-121,
respectively. For both backbone networks, this is a competitive
accuracy compared with the previous state of the art. Except
Herina, the AUC scores of the other 13 pathologies are very
close to or even higher than [8]. Moreover, we observe that
Infiltration has the lower recognition accuracy (0.728 and
0.717 for ResNet-50 and DenseNet-121). This is because the
diagnosis of Infiltration mainly relies on the texture change
among the lung area, which is challenging to recognize. The
disease Cardiomegaly achieves higher recognition accuracy
(0.904 and 0.912 for ResNet-50 and DenseNet-121, respectively), which is characterized by the relative solid region
(heart).
Performance of the local branch. The local branch is
trained on the cropped and resized lesion patches, which is
supposed to provide attention mechanisms complementary to
the global branch. The performance of the local branch is
demonstrated in Table. I, Fig. 5 and Fig. 9 as well.
Using ResNet-50 and DenseNet-121, the average AUC
score is 0.817 and 0.810, respectively, which is higher than [9],
[7]. Despite of being competitive, the local branch yields lower
accuracy than the global branch. For example, when using
Consolidation
ation
iltration
ysema
Classification results
Images
7
Effusion
Atelectasis
Infiltration
Consolidation
No Finding
Pneumonia
Mass
Nodule
Edema
Cardiomegaly
0.770
0.732
0.352
0.205
0.127
0.017
0.014
0.014
0.014
0.013
Emphysema
Pneumothorax
Effusion
Infiltration
Mass
No Finding
Atelectasis
Nodule
PT
Consolidation
0.831
0.754
0.106
0.101
0.087
0.082
0.075
0.030
0.027
0.024
Effusion
Atelectasis
Consolidation
Infiltration
No Finding
Pneumothorax
Emphysema
PT
Mass
Cardiomegaly
0.902
0.727
0.207
0.193
0.074
0.058
0.017
0.016
0.012
0.010
Effusion
Mass
Atelectasis
Infiltration
Nodule
No Finding
Consolidation
PT
Pneumothorax
Edema
0.820
0.780
0.201
0.130
0.115
0.065
0.051
0.046
0.028
0.011
Cardiomegaly
No Finding
Effusion
Infiltration
Atelectasis
Hernia
Nodule
Fibrosis
PT
Mass
0.752
0.304
0.133
0.108
0.068
0.054
0.048
0.037
0.035
0.022
Emphysema
Pneumothorax
Atelectasis
Effusion
No Finding
Infiltration
PT
Nodule
Mass
Fibrosis
0.854
0.810
0.264
0.139
0.138
0.085
0.054
0.034
0.018
0.016
Effusion
Cardiomegaly
Infiltration
Edema
Atelectasis
PT
Consolidation
Pneumonia
Mass
Nodule
0.915
0.807
0.415
0.144
0.089
0.078
0.052
0.037
0.029
0.029
Fig. 7. Examples of classification results. We present the top-10 predicted categories and the corresponding probability scores. The ground-truth labels are
highlighted in blue.
0.872
Average AUC
0.870
0.868
0.866
0.864
0.862
0.1
0.2
0.3
0.4
0.5
Threshold
0.6
0.7
0.8
0.9
Fig. 8. Average AUC scores of AG-CNN with different settings of τ on the
validation set (ResNet-50 as backbone).
ResNet-50, the performance gap is 2.4% (0.841 to 0.817). The
probable reason for this observation is that the lesion region
estimation and cropping process may lead to information loss
which is critical for recognition. So the local branch may suffer
from inaccurate estimation of the attention area.
Among the 14 classes, the largest performance drop is
observed at “Pneumonia” (0.067). The reason for the inferior performance at “Pneumonia” is probably that lots of
information are lost. Generally, the area where the lung is
inflamed is relative large and its corresponding attention heat
map shows a scattered distribution. With a higher value of τ ,
only a very small patch is cropped in original image. For the
classes “Hernia” and “Consolidation”, the local branch and
global branch yield very similar accuracy. We speculate that
the cropped local patch is consist with the lesion area in the
global image.
Effectiveness of fusing global and local branches. In
Table. I, Fig. 5, and Fig. 6, we illustrate the effectiveness of the
fusion branch, which yields the final classification results of
our model. Table. I shows AUC of AG-CNN over 14 classes.
The observations are consistent across different categories and
the two backbones. Fig. 5 presents the ROC curve of three
branches for each pathologies which illustrates that fusing
global and local branches can improve both of them obviously.
We presents the ROC curves of 14 pathologies with these
two backbones in Fig. 6. It shows the highly consistency
which demonstrate that AG-CNN is not sensitive to network
architecture of backbone.
For both ResNet-50 and DenseNet-121, the fusion branch,
i.e., AG-CNN, outperforms both the global branch and local
branch. For example, when using ResNet-50, the performance
gap from AG-CNN to the global and local branches is
0.027 and 0.051, respectively. Specifically AG-CNN (with
DenseNet-121 as backbone) surpasses the global and local
branches for all 14 pathologies.
The advantage of AG-CNN is consistent across the categories. Using ResNet-50 for example, the largest improvement
(0.047) is observed at the class “Nodule”, the disease of
which is featured by small lesion areas (see Fig. 4). In
fact, under such circumstances, the global branch can be
largely affected by the noise within the non-disease areas.
By paying attention on the small yet focused lesion areas,
our method effectively improves the classification performance
of Nodule. On the other hand, we also notice that under the
class Pneumonia, AG-CNN is inferior to the global branch, a
consistent observation made with the local branch: the local
branch is the least effective at this class. Some classification
results are presented in Fig. 7.
Another experiment, inputing a global image into both
global and local branch, is conducted to verify the effectiveness of fusing global and local cues. The same experimental
settings with Section IV-B are performed expect that the
mini-batchsize is 64 in training. Three branches are trained
together with ResNet-50 as backbone. The average AUC of
global, local and fusion branches achieve to 0.845, 0.846 and
0.851, respectively. The performance is lower 0.017 compared
with inputing a local patch into local branch. The results
show that AG-CNN is superior than both global and local
branch. In particular, the improvement is benefit from the
local discriminative region instead of increasing the number
of parameters.
Comparison with the state of the art. We compare our results with the state-of-the-art methods [9], [19], [7], [8] on the
ChestX-ray14 dataset. Wang et al. [9] classify and localize the
thorax disease in a unified weakly supervised framework. This
localization method actually compromises the classification
8
local branch
0.90
Average AUC
0.865
0.861
0.86
0.828
0.831
0.82
0.859
0.822
0.819
global branch (baseline)
0.868
0.867
0.858
0.858
0.822
fusion branch
0.825
0.866
0.863
0.841
0.817
0.804
0.78
0.751
0.74
0.1
0.2
0.3
0.4
0.5
Threshold
0.6
0.7
0.8
0.9
Fig. 9. Average AUCs for different settings of τ on the test set (ResNet-50 as backbone). Note that the results from global branch are our baseline.
accuracy. The reported results from Yao et al. [19] are based
on the model in which labels are considered independent.
Kumar et al. [7] try different boosting methods and cascade
the previous classification results for multi-label classification.
The accuracy of the previous step directly influences the result
of the following pathologies.
Comparing with these methods, this paper contributes
new state of the art to the community: average AUC =
0.871. AG-CNN exceeds the previous state of the art [8] by
2.9%. AUC scores of pathologies such as Cardiomegaly and
Infltration are higher than [8] by about 0.03. AUC scores of
Mass, Fibrosis and Consolidation surpass [8] by about 0.05.
Furthermore, we train AG-CNN with 70% of the dataset, but
80% are used in [7], [8]. In nearly all the 14 classes, our
method yields best performance. Only Rajpurkar et al. [8]
report higher accuracy on Hernia. In all, the classification
accuracy reported in this paper compares favorably against
previous art.
Variant of training strategy analysis. Training three
branches with different orders influences the performance of
AG-CNN. We perform 4 orders to train AG-CNN: 1) train
global branch first, and then local and fusion branch together
(G LF); 2) train global and local branch together, and then
fusion branch (GL F); 3) train three branches together (GLF);
4) train global, local and fusion branch sequentially (G L F).
Note that G L F is our three-stage training strategy. We limit
the batchsize to 64 for training two or three branches together,
such as GL F and GLF. And if the global branch is trained
first, the batchsize of each branch is set to 128, 64 and
64, respectively. The other experimental settings are same as
Section IV-B. We present the classification performance of
these training strategies in Table. II.
AG-CNN yields better performance (0.868 and 0.854) with
strategy of training three branches sequentially (G L F and
G L F∗ ). When global branch is trained first, we perform
the same model as the baseline in Table. I. Training with
G L F, AG-CNN obviously improves the baseline from 0.841
to 0.868. AG-CNN (G L F∗ ) performs a overall fine-tuning
when we train the fusion branch. It improves the global branch
TABLE II
R ESULTS OF DIFFERENT TRAINING STRATEGIES .
Strategy
GL F
GLF
G LF
G L F∗
G L F
*
Batchsize
64/64/64
64/64/64
128/64/64
128/64/64
128/64/64
Global
0.831
0.847
0.841
0.852
0.841
Local
0.800
0.815
0.809
0.819
0.817
Fusion
0.833
0.849
0.843
0.854
0.868
∗ represents that the parameters in global and local branch
are fine-tuned when we train the fusion branch. ResNet-50
is used as backbone.)
TABLE III
R ESULTS CORRESPONDING DIFFERENT STATISTICS .
Statistic
Max
L1
L2
*
Global
0.8412
0.8412
0.8412
Local
0.8171
0.8210
0.8213
Fusion
0.8680
0.8681
0.8672
ResNet-50 is used as backbone.
performance to 0.852, but not the local and fusion branches.
Compared with G L F and G L F∗ , performance of AGCNN (G LF) is much lower because its the inaccuracy of
local branch. When AG-CNN is trained with GL F and GLF,
it is inferior to G L F or G L F∗ . We infer that local branch
is essential to enhance AG-CNN performance.
Variant of heat map analysis. In Table. III, we report the
performance of using different heat map computing methods.
Based on the same baseline, the local branch produce a gap
of 0.0042 between Max and L2, but only 0.008 in fusion
branch. Max and L1 achieve very close performance on both
the local and fusion branch. It illustrates that different statistics
result in subtle differences in local branch, but not effect the
classification performance significantly.
D. Parameter Analysis
We analyze the sensitivity of AG-CNN to parameter variations. The key parameter of AG-CNN consists in τ in Eq. 4,
which defines the local regions and affects the classification
accuracy. Fig. 8 shows the average AUC of AG-CNN over
9
different τ on validation set. AG-CNN achieves the best
performance when τ is setting as 0.7. Therefore, we report
the results on test set with τ = 0.7.
Fig. 9 compares the average AUC of the global, local branch
and fusion branch on the test dataset when ResNet-50 is used
as basic network. τ changes from 0.1 to 0.9. When τ is small
(e.g. , close to 0), the local region is close to the global image.
For example, when τ = 0.1, the average AUC of the local
branch (0.828) is close to the result of the global branch
(0.841). In such cases, most of the entries in the attention
heat map are preserved, indicating that the cropped image
patches are close to the original input. On the other hand,
while τ reaches to 1, e.g., 0.9, the local branch is inferior
to the global branch by a large margin (0.9%). Under this
circumstance, most of the information in the global image is
discarded but only the top 10% largest values in the attention
heat map are retained. The cropped image patches reflect very
small regions.
Unlike the local branch, AG-CNN is relative stable to
changes of the threshold τ . When concentrating the global
and local branches, AG-CNN outperforms both branches by
at least 1.7% at τ = 0.4 and 0.5. AG-CNN exhibits the highest
AUC (>0.866) when τ ranges between [0.6, 0.8].
V. C ONCLUSION
In this paper, we propose an attention guided two-branch
convolutional neural network for thorax disease classification.
The proposed network is trained by considering both the global
and local cues informed in the global and local branches,
respectively. Departing from previous works which merely
rely on the global information, it uses attention heat maps to
mask the important regions which are used to train the local
branch. Extensive experiments demonstrate that combining
both global and local cues yields state-of-the-art accuracy
on the ChestX-ray14 dataset. We also demonstrate that our
method is relatively insensitive to parameter changes.
In the future research, we will continue the study from two
directions. First, we will investigate more accurate localization
of the lesion areas. Second, to tackle with the difficulties
in sample collection and annotation, semi-supervised learning
methods will be explored.
R EFERENCES
[1] P. Liskowski and K. Krawiec, “Segmenting retinal blood vessels with¡?
pub newline?¿ deep neural networks,” IEEE transactions on medical
imaging, vol. 35, no. 11, pp. 2369–2380, 2016. 1
[2] Y. Yuan, M. Chao, and Y.-C. Lo, “Automatic skin lesion segmentation
using deep fully convolutional networks with jaccard distance,” IEEE
Transactions on Medical Imaging, 2017. 1
[3] H. R. Roth, L. Lu, A. Farag, H.-C. Shin, J. Liu, E. B. Turkbey, and R. M.
Summers, “Deeporgan: Multi-level deep convolutional networks for automated pancreas segmentation,” in International Conference on Medical
Image Computing and Computer-Assisted Intervention. Springer, 2015,
pp. 556–564. 1
[4] H. Fu, J. Cheng, Y. Xu, D. W. K. Wong, J. Liu, and X. Cao, “Joint
optic disc and cup segmentation based on multi-label deep network and
polar transformation,” IEEE Transactions on Medical Imaging, 2018. 1
[5] H. Fu, Y. Xu, S. Lin, X. Zhang, D. W. K. Wong, J. Liu, A. F. Frangi,
M. Baskaran, and T. Aung, “Segmentation and quantification for angleclosure glaucoma assessment in anterior segment oct,” IEEE transactions
on medical imaging, vol. 36, no. 9, pp. 1930–1938, 2017. 1
[6] M. Anthimopoulos, S. Christodoulidis, L. Ebner, A. Christe, and
S. Mougiakakou, “Lung pattern classification for interstitial lung diseases using a deep convolutional neural network,” IEEE transactions on
medical imaging, vol. 35, no. 5, pp. 1207–1216, 2016. 1, 2
[7] P. Kumar, M. Grewal, and M. M. Srivastava, “Boosted cascaded convnets
for multilabel classification of thoracic diseases in chest radiographs,”
arXiv preprint arXiv:1711.08760, 2017. 1, 2, 5, 6, 7, 8
[8] P. Rajpurkar, J. Irvin, K. Zhu, B. Yang, H. Mehta, T. Duan, D. Ding,
A. Bagul, C. Langlotz, K. Shpanskaya et al., “Chexnet: Radiologistlevel pneumonia detection on chest x-rays with deep learning,” arXiv
preprint arXiv:1711.05225, 2017. 1, 2, 5, 6, 7, 8
[9] X. Wang, Y. Peng, L. Lu, Z. Lu, M. Bagheri, and R. M. Summers,
“Chestx-ray8: Hospital-scale chest x-ray database and benchmarks on
weakly-supervised classification and localization of common thorax
diseases,” arXiv preprint arXiv:1705.02315, 2017. 1, 2, 4, 5, 6, 7
[10] J. M. Wolterink, T. Leiner, M. A. Viergever, and I. Isgum, “Generative
adversarial networks for noise reduction in low-dose ct,” IEEE Transactions on Medical Imaging, 2017. 1
[11] S. Albarqouni, C. Baur, F. Achilles, V. Belagiannis, S. Demirci, and
N. Navab, “Aggnet: deep learning from crowds for mitosis detection in
breast cancer histology images,” IEEE transactions on medical imaging,
vol. 35, no. 5, pp. 1313–1321, 2016. 1
[12] Y. Xu, T. Mo, Q. Feng, P. Zhong, M. Lai, I. Eric, and C. Chang, “Deep
learning of feature representation with multiple instance learning for
medical image analysis,” in Acoustics, Speech and Signal Processing
(ICASSP), 2014 IEEE International Conference on. IEEE, 2014, pp.
1626–1630. 1
[13] R. Liao, S. Miao, P. de Tournemire, S. Grbic, A. Kamen, T. Mansi,
and D. Comaniciu, “An artificial agent for robust image registration.” in
AAAI, 2017, pp. 4168–4175. 1
[14] D. Yang, S. Zhang, Z. Yan, C. Tan, K. Li, and D. Metaxas, “Automated
anatomical landmark detection ondistal femur surface using convolutional neural network,” in Biomedical Imaging (ISBI), 2015 IEEE 12th
International Symposium on. IEEE, 2015, pp. 17–21. 1
[15] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification
with deep convolutional neural networks,” in Advances in neural information processing systems, 2012, pp. 1097–1105. 1
[16] K. Simonyan and A. Zisserman, “Very deep convolutional networks for
large-scale image recognition,” arXiv preprint arXiv:1409.1556, 2014.
1
[17] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan,
V. Vanhoucke, and A. Rabinovich, “Going deeper with convolutions,”
in Proceedings of the IEEE conference on computer vision and pattern
recognition, 2015, pp. 1–9. 1
[18] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image
recognition,” in Proceedings of the IEEE conference on computer vision
and pattern recognition, 2016, pp. 770–778. 1, 3, 6
[19] L. Yao, E. Poblenz, D. Dagunts, B. Covington, D. Bernard, and K. Lyman, “Learning to diagnose from scratch by exploiting dependencies
among labels,” arXiv preprint arXiv:1710.10501, 2017. 1, 2, 5, 7
[20] G. Huang, Z. Liu, K. Q. Weinberger, and L. van der Maaten, “Densely
connected convolutional networks,” arXiv preprint arXiv:1608.06993,
2016. 1, 6
[21] S. Hochreiter and J. Schmidhuber, “Long short-term memory,” Neural
computation, vol. 9, no. 8, pp. 1735–1780, 1997. 1
[22] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “Imagenet:
A large-scale hierarchical image database,” in Computer Vision and
Pattern Recognition, 2009. CVPR 2009. IEEE Conference on. IEEE,
2009, pp. 248–255. 1
[23] A. Krizhevsky and G. Hinton, “Learning multiple layers of features from
tiny images,” 2009. 1
[24] J. Shiraishi, S. Katsuragawa, J. Ikezoe, T. Matsumoto, T. Kobayashi, K.i. Komatsu, M. Matsui, H. Fujita, Y. Kodera, and K. Doi, “Development
of a digital image database for chest radiographs with and without a
lung nodule: receiver operating characteristic analysis of radiologists’
detection of pulmonary nodules,” American Journal of Roentgenology,
vol. 174, no. 1, pp. 71–74, 2000. 2
[25] B. Van Ginneken, M. B. Stegmann, and M. Loog, “Segmentation of
anatomical structures in chest radiographs using supervised methods:
a comparative study on a public database,” Medical image analysis,
vol. 10, no. 1, pp. 19–40, 2006. 2
[26] S. Jaeger, S. Candemir, S. Antani, Y.-X. J. Wáng, P.-X. Lu, and
G. Thoma, “Two public chest x-ray datasets for computer-aided screening of pulmonary diseases,” Quantitative imaging in medicine and
surgery, vol. 4, no. 6, p. 475, 2014. 2
[27] D. Demner-Fushman, M. D. Kohli, M. B. Rosenman, S. E. Shooshan,
L. Rodriguez, S. Antani, G. R. Thoma, and C. J. McDonald, “Preparing
10
[28]
[29]
[30]
[31]
[32]
[33]
[34]
[35]
[36]
[37]
[38]
[39]
[40]
a collection of radiology examinations for distribution and retrieval,”
Journal of the American Medical Informatics Association, vol. 23, no. 2,
pp. 304–310, 2015. 2
“Open-i: An open access biomedical search engine.” https://openi.nlm.
nih.gov. 2
G. Litjens, T. Kooi, B. E. Bejnordi, A. A. A. Setio, F. Ciompi,
M. Ghafoorian, J. A. van der Laak, B. van Ginneken, and C. I. Sánchez,
“A survey on deep learning in medical image analysis,” arXiv preprint
arXiv:1702.05747, 2017. 2
A. Qayyum, S. M. Anwar, M. Majid, M. Awais, and M. Alnowami,
“Medical image analysis using convolutional neural networks: A review,” arXiv preprint arXiv:1709.02250, 2017. 2
D. Shen, G. Wu, and H.-I. Suk, “Deep learning in medical image
analysis,” Annual Review of Biomedical Engineering, no. 0, 2017. 2
H.-C. Shin, H. R. Roth, M. Gao, L. Lu, Z. Xu, I. Nogues, J. Yao,
D. Mollura, and R. M. Summers, “Deep convolutional neural networks
for computer-aided detection: Cnn architectures, dataset characteristics
and transfer learning,” IEEE transactions on medical imaging, vol. 35,
no. 5, pp. 1285–1298, 2016. 2
H.-C. Shin, K. Roberts, L. Lu, D. Demner-Fushman, J. Yao, and R. M.
Summers, “Learning to read chest x-rays: recurrent neural cascade model
for automated image annotation,” in Proceedings of the IEEE Conference
on Computer Vision and Pattern Recognition, 2016, pp. 2497–2506. 2
M. T. Islam, M. A. Aowal, A. T. Minhaz, and K. Ashraf, “Abnormality
detection and localization in chest x-rays using deep convolutional neural
networks,” arXiv preprint arXiv:1705.09850, 2017. 2
S. Hwang and H.-E. Kim, “Self-transfer learning for weakly supervised
lesion localization,” in International Conference on Medical Image
Computing and Computer-Assisted Intervention. Springer, 2016, pp.
239–246. 2
C. Payer, D. Štern, H. Bischof, and M. Urschler, “Regressing heatmaps
for multiple landmark localization using cnns,” in International Conference on Medical Image Computing and Computer-Assisted Intervention.
Springer, 2016, pp. 230–238. 2
P.-P. Ypsilantis and G. Montana, “Learning what to look in chest
x-rays with a recurrent visual attention model,” arXiv preprint
arXiv:1701.06452, 2017. 2
V. Mnih, N. Heess, A. Graves et al., “Recurrent models of visual
attention,” in Advances in neural information processing systems, 2014,
pp. 2204–2212. 2
E. Pesce, P.-P. Ypsilantis, S. Withey, R. Bakewell, V. Goh, and G. Montana, “Learning to detect chest radiographs containing lung nodules using visual attention networks,” arXiv preprint arXiv:1712.00996, 2017.
2
A. Paszke, S. Gross, S. Chintala, and G. Chanan, “Pytorch,” 2017. 6
| 1cs.CV
|
arXiv:1403.3369v3 [cs.NE] 22 Apr 2017
Herbert Jaeger
Controlling Recurrent Neural Networks by
Conceptors
Technical Report No. 31
March 2014
School of Engineering and Science
Controlling Recurrent Neural Networks by Conceptors (Revision 2)
Herbert Jaeger
Jacobs University Bremen
School of Engineering and Science
Campus Ring
28759 Bremen
Germany
E-Mail: [email protected]
http: // minds. jacobs-university. de
Abstract
The human brain is a dynamical system whose extremely complex sensordriven neural processes give rise to conceptual, logical cognition. Understanding the interplay between nonlinear neural dynamics and concept-level
cognition remains a major scientific challenge. Here I propose a mechanism
of neurodynamical organization, called conceptors, which unites nonlinear
dynamics with basic principles of conceptual abstraction and logic. It becomes possible to learn, store, abstract, focus, morph, generalize, de-noise
and recognize a large number of dynamical patterns within a single neural
system; novel patterns can be added without interfering with previously
acquired ones; neural noise is automatically filtered. Conceptors help explaining how conceptual-level information processing emerges naturally and
robustly in neural systems, and remove a number of roadblocks in the theory
and applications of recurrent neural networks.
Changelog for Revision 1: Besides correcting trivial typos, the following edits
/ corrections have been made after the first publication on arXiv, leading to the
revision 1 version on arXiv:
• Apr 7, 2014: updated algorithm description for incremental memory management by documenting W out component.
• Jun 15, 2014: added missing parentheses in C update formula given at
beginning of Section 3.14.
• Oct 11, 2014: added missing Frobenious norms in similarity formulas (3)
and (10) (pointed out by Dennis Hamester)
• May 2016: re-wrote and extended the section on incremental memory management (Section 3.11)
• July 2016: deleted the subsubsection “To be or not to be an attractor” from
Section 3.14.3 because this material is repeated later in the text in more
detail.
Changelog for Revision 2:
• April 22, 2017: Proposition 12 was removed - it was wrong.
3
Notes on the Structure of this Report. This report introduces several novel
analytical concepts describing neural dynamics; develops the corresponding mathematical theory under aspects of linear algebra, dynamical systems theory, and
formal logic; introduces a number of novel learning, adaptation and control algorithms for recurrent neural networks; demonstrates these in a number of case
studies; proposes biologically (not too im-)plausible realizations of the dynamical
mechanisms; and discusses relationships to other work. Said shortly, it’s long.
Not all parts will be of interest to all readers. In order to facilitate navigation
through the text and selection of relevant components, I start with an overview
section which gives an intuitive explanation of the novel concepts and informal
sketches of the main results and demonstrations (Section 1). After this overview,
the material is presented in detail, starting with an introduction (Section 2) which
relates this contribution to other research. The main part is Section 3, where I
systematically develop the theory and algorithms, interspersed with simulation
demos. A graphical dependency map for this section is given at the beginning of
Section 3. The technical documentation of the computer simulations is provided
in Section 4, and mathematical proofs are collected in Section 5. The detailed
presentation in Sections 2 – 5 is self-contained. Reading the overview in Section
1 may be helpful but is not necessary for reading these sections. For convenience
some figures from the overview section are repeated in Section 3.
Acknowledgements. The work described in this report was partly funded through
the European FP7 project AMARSi (www.amarsi-project.eu). The author is indebted to Dr. Mathieu Galtier and Dr. Manjunath Ghandi for careful proofreading
(not an easy task).
4
Contents
1 Overview
7
2 Introduction
25
2.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.2 Mathematical Preliminaries . . . . . . . . . . . . . . . . . . . . . . 27
3 Theory and Demonstrations
3.1 Networks and Signals . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2 Driving a Reservoir with Different Patterns . . . . . . . . . . . . . .
3.3 Storing Patterns in a Reservoir, and Training the Readout . . . . .
3.4 Conceptors: Introduction and Basic Usage in Retrieval . . . . . . .
3.5 A Similarity Measure for Excited Network Dynamics . . . . . . . .
3.6 Online Learning of Conceptor Matrices . . . . . . . . . . . . . . . .
3.7 Morphing Patterns . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.8 Understanding Aperture . . . . . . . . . . . . . . . . . . . . . . . .
3.8.1 The Semantics of α as “Aperture” . . . . . . . . . . . . . . .
3.8.2 Aperture Adaptation and Final Definition of Conceptor Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.8.3 Aperture Adaptation: Example . . . . . . . . . . . . . . . .
3.8.4 Guides for Aperture Adjustment . . . . . . . . . . . . . . .
3.9 Boolean Operations on Conceptors . . . . . . . . . . . . . . . . . .
3.9.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.9.2 Preliminary Definition of Boolean Operations . . . . . . . .
3.9.3 Final Definition of Boolean Operations . . . . . . . . . . . .
3.9.4 Facts Concerning Subspaces . . . . . . . . . . . . . . . . . .
3.9.5 Boolean Operators and Aperture Adaptation . . . . . . . . .
3.9.6 Logic Laws . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.10 An Abstraction Relationship between Conceptors . . . . . . . . . .
3.11 Example: Memory Management in RNNs . . . . . . . . . . . . . . .
3.11.1 General Principle and Algorithm . . . . . . . . . . . . . . .
3.11.2 Demonstration 1: Incremental Loading of Integer-Periodic
Patterns . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.12 Incremental Loading of Integer-Periodic Patterns: Detailed Analysis
3.12.1 Demonstration 2: Incremental Loading of Irrational-Period
Patterns . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.12.2 Integer-Periodic Versus Parametrized Patterns: Close-up Inspection . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.12.3 Incremental Loading of Arbitrary Patterns . . . . . . . . . .
3.13 Example: Dynamical Pattern Recognition . . . . . . . . . . . . . .
3.14 Autoconceptors . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.14.1 Motivation and Overview . . . . . . . . . . . . . . . . . . .
3.14.2 Basic Equations . . . . . . . . . . . . . . . . . . . . . . . . .
5
29
29
30
34
34
38
38
39
43
43
44
46
46
50
50
51
53
55
56
56
58
59
60
63
65
67
69
72
74
80
80
82
3.15
3.16
3.17
3.18
3.19
3.14.3 Example: Autoconceptive Reservoirs as Content-Addressable
Memories . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.14.4 Analysis of Autoconceptor Adaptation Dynamics . . . . . .
Toward Biologically Plausible Neural Circuits: Random Feature
Conceptors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
A Hierarchical Filtering and Classification Architecture . . . . . . .
Toward a Formal Marriage of Dynamics with Logic . . . . . . . . .
Conceptor Logic as Institutions: Category-Theoretical Detail . . . .
Final Summary and Outlook . . . . . . . . . . . . . . . . . . . . . .
84
93
104
121
136
147
158
4 Documentation of Experiments and Methods
161
4.1 General Set-Up, Initial Demonstrations (Section 1 and Section 3.2
- 3.4) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
4.2 Aperture Adaptation (Sections 3.8.3 and 3.8.4) . . . . . . . . . . . 162
4.3 Memory Management, Demo 1 (Section 3.11.2) . . . . . . . . . . . 164
4.4 Memory Management, Demo 2 (Section 3.12.1) . . . . . . . . . . . 165
4.5 Memory Management, Close-Up Inspection (Section 3.12.2) . . . . . 165
4.6 Memory Management, Arbitrary Patterns (Section 3.12.3) . . . . . 165
4.7 Content-Addressable Memory (Section 3.14.3) . . . . . . . . . . . . 166
4.8 The Japanese Vowels Classification (Section 3.13) . . . . . . . . . . 166
4.9 Conceptor Dynamics Based on RFC Conceptors (Section 3.15) . . . 167
4.10 Hierarchical Classification and Filtering Architecture (Section 3.16) 168
5 Proofs and Algorithms
5.1 Proof of Proposition 1 (Section 3.4) . .
5.2 Proof of Proposition 6 (Section 3.9.3) .
5.3 Proof of Proposition 7 (Section 3.9.3) .
5.4 Proof of Proposition 8 (Section 3.9.3) .
5.5 Proof of Proposition 9 (Section 3.9.4) .
5.6 Proof of Proposition 10 (Section 3.9.5)
5.7 Proof of Proposition 11 (Section 3.9.6)
5.8 Proof of Proposition 12 (Section 3.9.6)
5.9 Proof of Proposition 13 (Section 3.10) .
5.10 Proof of Proposition 15 (Section 3.14.4)
5.11 Proof of Proposition 17 (Section 3.18) .
References
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
171
. 171
. 172
. 175
. 175
. 176
. 178
. 181
. 181
. 182
. 186
. 191
192
6
1
Overview
Scientific context. Research on brains and cognition unfolds in two directions.
Top-down oriented research starts from the “higher” levels of cognitive performance, like rational reasoning, conceptual knowledge representation, command of
language. These phenomena are typically described in symbolic formalisms developed in mathematical logic, artificial intelligence (AI), computer science and
linguistics. In the bottom-up direction, one departs from “low-level” sensor data
processing and motor control, using the analytical tools offered by dynamical systems theory, signal processing and control theory, statistics and information theory. The human brain obviously has found a way to implement high-level logical
reasoning on the basis of low-level neuro-dynamical processes. How this is possible, and how the top-down and bottom-up research directions can be united, has
largely remained an open question despite long-standing efforts in neural networks
research and computational neuroscience [82, 89, 33, 1, 36, 43], machine learning
[35, 47], robotics [10, 83], artificial intelligence [85, 106, 7, 9], dynamical systems
modeling of cognitive processes [96, 100, 107], cognitive science and linguistics
[22, 98], or cognitive neuroscience [4, 26].
Summary of contribution. Here I establish a fresh view on the neuro-symbolic
integration problem. I show how dynamical neural activation patterns can be
characterized by certain neural filters which I call conceptors. Conceptors derive
naturally from the following key observation. When a recurrent neural network
(RNN) is actively generating, or is passively being driven by different dynamical
patterns (say a, b, c, . . .), its neural states populate different regions Ra , Rb , Rc , . . .
of neural state space. These regions are characteristic of the respective patterns.
For these regions, neural filters Ca , Cb , Cc , . . . (the conceptors) can be incrementally learnt. A conceptor Cx representing a pattern x can then be invoked after
learning to constrain the neural dynamics to the state region Rx , and the network
will select and re-generate pattern x. Learnt conceptors can be blended, combined by Boolean operations, specialized or abstracted in various ways, yielding
novel patterns on the fly. The logical operations on conceptors admit a rigorous
semantical interpretation; conceptors can be arranged in conceptual hierarchies
which are structured like semantic networks known from artificial intelligence.
Conceptors can be economically implemented by single neurons (addressing patterns by neurons, leading to explicit command over pattern generation), or they
may self-organize spontaneously and quickly upon the presentation of cue patterns (content-addressing, leading to pattern imitation). Conceptors can also be
employed to “allocate free memory space” when new patterns are learnt and stored
in long-term memory, enabling incremental life-long learning without the danger
of freshly learnt patterns disrupting already acquired ones. Conceptors are robust
against neural noise and parameter variations. The basic mechanisms are generic
and can be realized in any kind of dynamical neural network. All taken together,
7
conceptors offer a principled, transparent, and computationally efficient account
of how neural dynamics can self-organize in conceptual structures.
Going bottom-up: from neural dynamics to conceptors. The neural
model system in this report are standard recurrent neural networks (RNNs, Figure
1 A) whose dynamics is mathematically described be the state update equations
x(n + 1) = tanh(W ∗ x(n) + W in p(n)),
y(n) = W out x(n).
Time here progresses in unit steps n = 1, 2, . . .. The network consists of N
neurons (typically in the order of a hundred in this report), whose activations
x1 (n), . . . , xN (n) at time n are collected in an N -dimensional state vector x(n).
The neurons are linked by random synaptic connections, whose strengths are collected in a weight matrix W ∗ of size N × N . An input signal p(n) is fed to the
network through synaptic input connections assembled in the input weight matrix W in . The “S-shaped” function tanh squashes the neuronal activation values
into a range between −1 and 1. The second equation specifies that an ouput signal y(n) can be read from the network activation state x(n) by means of output
weights W out . These weights are pre-computed such that the output signal y(n)
just repeats the input signal p(n). The output signal plays no functional role
in what follows; it merely serves as a convenient 1-dimensional observer of the
high-dimensional network dynamics.
The network-internal neuron-to-neuron connections W ∗ are created at random.
This will lead to the existence of cyclic (“recurrent”) connection pathways inside
the network. Neural activation can reverberate inside the network along these
cyclic pathways. The network therefore can autonomously generate complex neurodynamical patterns even when it receives no input. Following the terminology of
the reservoir computing [56, 5], I refer to such randomly connected neural networks
as reservoirs.
For the sake of introducing conceptors by way of an example, consider a reservoir with N = 100 neurons. I drive this system with a simple sinewave input p(n)
(first panel in first row in Fig. 1 B). The reservoir becomes entrained to this input,
each neuron showing individual variations thereof (Fig. 1 B second panel). The
resulting reservoir state sequence x(1), x(2), . . . can be represented as a cloud of
points in the 100-dimensional reservoir state space. The dots in the first panel
of Fig. 1 C show a 2-dimensional projection of this point cloud. By a statistical
method known as principal component analysis, the shape of this point cloud can
be captured by an N -dimensional ellipsoid whose main axes point in the main
scattering directions of the point cloud. This ellipsoid is a geometrical representation of the correlation matrix R of the state points. The lengths σ1 , . . . , σN of the
ellipsoid axes are known as the singular values of R. The directions and lengths of
these axes provide a succinct characterization of the geometry of the state point
8
&'('&)"*&!
W in
W out
W
p!!
*+%#$!!
!"
#$%&'$(
#"
y!
x!
0+1!"$!
)*+$&,-.+&/
/
!
"
!!
!
"
!!
"
!!"
!
!
"
!!
!
"
!!
"
!!"
!
!
"
!!
!
"
!!
"
!!"
!
"
!!"
!
!
"
!!
"
!
"
!!
2"
"
!"
!"
2"
!
"
$"
"#$%#$!!
!!
!!
"
"
"
"
3"
"
"
!""
!
!1
!2
"
"
!!
!!
!
3"
!""
!
s1
s2
"
"
!
!!
!!
"
!
Figure 1: Deriving conceptors from network dynamics. A. Network layout. Arrows indicate synaptic links. B. Driving the reservoir with four different input
patterns. Left panels: 20 timesteps of input pattern p(n) (black thin line) and
conceptor-controlled output y(n) (bold light gray). Second column: 20 timesteps
of traces xi (n), xj (n) of two randomly picked reservoir neurons. Third column: the
singular values σi of the reservoir state correlation matrix R in logarithmic scale.
Last column: the singular values si of the conceptors C in linear plotting scale. C.
From pattern to conceptor. Left: plots of value pairs xi (n), xj (n) (dots) of the two
neurons shown in first row of B and the resulting ellipse with axis lengths σ1 , σ2 .
Right: from R (thin light gray) to conceptor C (bold dark gray) by normalizing
axis lengths σ1 , σ2 to s1 , s2 .
9
cloud. The N = 100 lengths σi resulting in this example are log-plotted in Fig. 1
B, third column, revealing an exponential fall-off in this case.
As a next step, these lengths σi are normalized to become si = σi /(σi + α−2 ),
where α ≥ 0 is a design parameter that I call aperture. This normalization ensures
that all si are not larger than 1 (last column in Fig. 1 B). A new ellipsoid is
obtained (Fig. 1 C right) which is located inside the unit sphere. The normalized
ellipsoid can be described by a N -dimensional matrix C, which I call a conceptor
matrix. C can be directly expressed in terms of R by C = R(R + α−2 I)−1 , where
I is the identity matrix.
When a different driving pattern p is used, the shape of the state point cloud,
and subsequently the conceptor matrix C, will be characteristically different. In
the example, I drove the reservoir with four patterns p1 – p4 (rows in Fig. 1B). The
first two patterns were sines of slightly different frequencies, the last two patterns
were minor variations of a 5-periodic random pattern. The conceptors derived
from the two sine patterns differ considerably from the conceptors induced by the
two 5-periodic patterns (last column in Fig. 1B). Within each of these two pairs,
the conceptor differences are too small to become visible in the plots.
There is an instructive alternative way to define conceptors. Given a sequence of reservoir states x(1), . . . , x(L), the conceptor C which characterizes
this
matrix which minimizes the cost function
P state point cloud is 2the unique
−2
2
2
n=1,...,L kx(n) − Cx(n)k /L + α kCk , where kCk is the sum of all squared
matrix entries. The first term in this cost would become minimal if C were the
identity map, the second term would become minimal if C would be the all-zero
map. The aperture α strikes a balance between these two competing cost components. For increasing apertures, C will tend toward the identity matrix I; for
shrinking apertures it will come out closer to the zero matrix. In the terminology of machine learning, C is hereby defined as a regularized identity map.
The explicit solution to this minimization problem is again given by the formula
C = R (R + α−2 I)−1 .
Summing up: if a reservoir is driven by a pattern p(n), a conceptor matrix C
can be obtained from the driven reservoir states x(n) as the regularized identity
map on these states. C can be likewise seen as a normalized ellipsoid characterization of the shape of the x(n) point cloud. I write C(p, α) to denote a conceptor
derived from a pattern p using aperture α, or C(R, α) to denote that C was obtained from a state correlation matrix R.
Loading a reservoir. With the aid of conceptors a reservoir can re-generate a
number of different patterns p1 , . . . , pK that it has previously been driven with.
For this to work, these patterns have to be learnt by the reservoir in a special
sense, which I call loading a reservoir with patterns. The loading procedure works
as follows. First, drive the reservoir with the patterns p1 , . . . , pK in turn, collecting reservoir states xj (n) (where j = 1, . . . , K). Then, recompute the reservoir
connection weights W ∗ into W such that W optimally balances between the fol10
lowing two goals. First, W should be such that W xj (n) ≈ W ∗ xj (n) + W in pj (n)
for all times n and patterns j. That is, W should allow the reservoir to “simulate” the driving input in the absence of the same. Second, W should be such
that the weights collected in this matrix become as small as possible. Technically
this compromise-seeking learning task amounts to computing what is known as a
regularized linear regression, a standard and simple computational task. This idea
of “internalizing” a driven dynamics into a reservoir has been independently (re)introduced under different names and for a variety of purposes (self-prediction
[72], equilibration [55], reservoir regularization [92], self-sensing networks [102],
innate training [61]) and appears to be a fundamental RNN adaptation principle.
Going top-down: from conceptors to neural dynamics. Assume that conceptors C j = C(pj , α) have been derived for patterns p1 , . . . , pK , and that these
patterns have been loaded into the reservoir, replacing the original random weights
W ∗ by W . Intuitively, the loaded reservoir, when it is run using x(n + 1) =
tanh(W x(n)) (no input!) should behave exactly as when it was driven with input
earlier, because W has been trained such that W x(n) ≈ W ∗ x(n) + W in pj (n).
In fact, if only a single pattern had been loaded, the loaded reservoir would
readily re-generate it. But if more than one patter had been loaded, the autonomous (input-free) update x(n + 1) = tanh(W x(n)) will lead to an entirely
unpredictable dynamics: the network can’t “decide” which of the loaded patterns
it should re-generate! This is where conceptors come in. The reservoir dynamics is filtered through C j . This is effected by using the augmented update rule
x(n + 1) = C j tanh(W x(n)). By virtue of inserting C j into the feedback loop,
the reservoir states become clipped to fall within the ellipsoid associated with C j .
As a result, the pattern pj will be re-generated: when the reservoir is observed
through the previously trained output weights, one gets y(n) = W out x(n) ≈ pj (n).
The first column of panels in Fig. 1 B shows an overlay of the four autonomously
re-generated patterns y(n) with the original drivers pj used in that example. The
recovery of the originals is quite accurate (mean square errors 3.3e-05, 1.4e-05,
0.0040, 0.0019 for the four loaded patterns). Note that the first two and the last
two patterns are rather similar to each other. The filtering afforded by the respective conceptors is “sharp” enough to separate these twin pairs. I will later
demonstrate that in this way a remarkably large number of patterns can be faithfully re-generated by a single reservoir.
Morphing and generalization. Given a reservoir loaded with K patterns pj ,
the associated conceptors C j can be linearly combined by creating mixture conceptors M = µ1 C 1 + . . . + µK C K , where the mixing coefficients µj must sum to 1.
When the reservoir is run under the control of such a morphed conceptor M , the
resulting generated pattern is a morph between the original “pure” patterns pj . If
all µj are non-negative, the morph can be considered an interpolation between the
pure patterns; if some µj are negative, the morph extrapolates beyond the loaded
11
pure patterns. I demonstrate this with the four patterns used in the example
above, setting µ1 = (1 − a)b, µ2 = ab, µ3 = (1 − a)(1 − b), µ4 = a(1 − b), and letting
a, b vary from −0.5 to 1.5 in increments of 0.25. Fig. 2 shows plots of observer
signals y(n) obtained when the reservoir is generating patterns under the control
of these morphed conceptors. The innermost 5 by 5 panels show interpolations
between the four pure patterns, all other panels show extrapolations.
In machine learning terms, both interpolation and extrapolation are cases of
generalization. A standard opinion in the field states that generalization by interpolation is what one may expect from learning algorithms, while extrapolation
beyond the training data is hard to achieve.
Morphing and generalizing dynamical patterns is a common but nontrivial task
for training motor patterns in robots. It typically requires training demonstrations
of numerous interpolating patterns [90, 16, 68]. Conceptor-based pattern morphing
appears promising for flexible robot motor pattern learning from a very small
number of demonstrations.
b = –0.5
0.0
0.5
1.0
1.5
a = –0.5
0.0
0.5
1.0
1.5!
Figure 2: Morphing between, and generalizing beyond, four loaded patterns. Each
panel shows a 15-step autonomously generated pattern (plot range between −1
and +1). Panels with bold frames: the four loaded prototype patterns (same
patterns as in Fig. 1 B.)
12
Aperture adaptation. Choosing the aperture α appropriately is crucial for regenerating patterns in a stable and accurate way. To demonstrate this, I loaded
a 500-neuron reservoir with signals p1 – p4 derived from four classical chaotic
attractors: the Lorenz, Rössler, Mackey-Glass, and Hénon attractors. Note that
it used to be a challenging task to make an RNN learn any single of these attractors
[56]; to my knowledge, training a single RNN to generate several different chaotic
attractors has not been attempted before. After loading the reservoir, the regeneration was tested using conceptors C(pj , α) where for each attractor pattern
pj a number of different values for α were tried. Fig. 3 A shows the resulting regenerated patterns for five apertures for the Lorenz attractor. When the aperture
is too small, the reservoir-conceptor feedback loop becomes too constrained and the
produced patterns de-differentiate. When the aperture is too large, the feedback
loop becomes over-excited.
11
66
4e+02
2.4e+03
1.4e+04
A
1e+03
1.3e+03
6.3e+02
B
Lorenz
C
Roessler
MackeyïGlass
Hénon
0
0
0
0
ï5
ï5
ï5
ï5
1 2 3 4 5
log10 aperture
1
2
3
4
5
1
2
3
4
5
1
2
3
4
5
Figure 3: Aperture adaptation for re-generating four chaotic attractors. A Lorenz
attractor. Five versions re-generated with different apertures (values inserted in
panels) and original attractor (green). B Best re-generations of the other three
attractors (from left to right: Rössler, Mackey-Glass, and Hénon, originals in
green). C Log10 of the attenuation criterion plotted against the log10 of aperture.
Dots mark the apertures used for plots in A and B.
An optimal aperture can be found by experimentation, but this will not be
an option in many engineering applications or in biological neural systems. An
intrinsic criterion for optimizing α is afforded by a quantity that I call attenuation:
the damping ratio which the conceptor imposes on the reservoir signal. Fig. 3
C plots the attenuation against the aperture for the four chaotic signals. The
13
minimum of this curve marks a good aperture value: when the conceptor dampens
out a minimal fraction of the reservoir signal, conceptor and reservoir are in good
“resonance”. The chaotic attractor re-generations shown in Fig. 3 B were obtained
by using this minimum-attenuation criterion.
The aperture range which yields visibly good attractor re-generations in this
demonstration spans about one order of magnitude. With further refinements
(zeroing small singular values in conceptors is particularly effective), the viable
aperture range can be expanded to about three orders of magnitude. While setting
the aperture right is generally important, fine-tuning is unnecessary.
Boolean operations and conceptor abstraction. Assume that a reservoir
is driven by a pattern r which consists of randomly alternating epochs of two
patterns p and q. If one doesn’t know which of the two patterns is active at a
given time, all one can say is that the pattern r currently is p OR it is q. Let
C(Rp , 1), C(Rq , 1), C(Rr , 1) be conceptors derived from the two partial patterns
p, q and the “OR” pattern r, respectively. Then it holds that C(Rr , 1) = C((Rp +
Rq )/2, 1). Dropping the division by 2, this motivates to define an OR (mathematical notation: ∨) operation on conceptors C1 (R1 , 1), C2 (R2 , 1) by putting
C1 ∨ C2 := (R1 + R2 )(R1 + R2 + I)−1 . The logical operations NOT (¬) and AND
(∧) can be defined along similar lines. Fig. 4 shows two-dimensional examples of
applying the three operations.
1
1
1
0
0
0
ï1
ï1
0
ï1
1 ï1
0
ï1
1 ï1
0
1
Figure 4: Boolean operations on conceptors. Red/blue (thin) ellipses represent
source conceptors C1 , C2 . Magenta (thick) ellipses show C1 ∨ C2 , C1 ∧ C2 , ¬C1
(from left to right).
Boolean logic is the mathematical theory of ∨, ∧, ¬. Many laws of Boolean
logic also hold for the ∨, ∧, ¬ operations on conceptors: the laws of associativity, commutativity, double negation, de Morgan’s rules, some absorption rules.
Furthermore, numerous simple laws connect aperture adaptation to Boolean operations. Last but not least, by defining C1 ≤ C2 if and only if there exists a
conceptor B such that C2 = C1 ∨ B, an abstraction ordering is created on the set
of all conceptors of dimension N .
14
Neural memory management. Boolean conceptor operations afford unprecedented flexibility of organizing and controlling the nonlinear dynamics of recurrent
neural networks. Here I demonstrate how a sequence of patterns p1 , p2 , . . . can be
incrementally loaded into a reservoir, such that (i) loading a new pattern pj+1
does not interfere with previously loaded p1 , . . . , pj ; (ii) if a new pattern pj+1 is
similar to already loaded ones, the redundancies are automatically detected and
exploited, saving memory capacity; (iii) the amount of still “free” memory space
can be logged.
Let C j be the conceptor associated with pattern pj . Three ideas are combined
to implement the memory management scheme. First, keep track of the “already
used” memory space by maintaining a conceptor Aj = C 1 ∨ . . . ∨ C j . The sum of
all singular values of Aj , divided by the reservoir size, gives a number that ranges
between 0 and 1. It is an indicator of the portion of reservoir “space” which has
been used up by loading C 1 , . . . , C j , and I call it the quota claimed by C 1 , . . . , C j .
Second, characterize what is “new” about C j+1 (not being already represented by
previously loaded patterns) by considering the conceptor N j+1 = C j+1 \ Aj . The
logical difference operator \ can be re-written as A \ B = A ∧ ¬B. Third, load
only that which is new about C j+1 into the still unclaimed reservoir space, that is,
into ¬Aj . These three ideas can be straightforwardly turned into a modification
of the basic pattern loading algorithm.
For a demonstration, I created a series of periodic patterns p1 , p2 , . . . whose
integer period lengths were picked randomly between 3 and 15, some of these
patterns being sines, others random patterns. These patterns were incrementally
loaded in a 100-neuron reservoir, one by one. Fig. 5 shows the result. Since patterns j = 5, 6, 7 were identical replicas of patterns j = 1, 2, 3, no additional quota
space was consumed when these patterns were (re-)loaded. Accuracy was measured by the normalized root mean square error (NRMSE). The NRMSE jumps
from very small values to a high value when the last pattern is loaded; the quota of
0.99 at this point indicates that the reservoir is “full”. The re-generation testing
and NRMSE computation was done after all patterns had been loaded. An attempt to load further patterns would be unsuccessful, but it also would not harm
the re-generation quality of the already loaded ones.
This ability to load patterns incrementally suggests a solution to a notorious
problem in neural network training, known as catastrophic forgetting, which manifests itself in a disruption of previously learnt functionality when learning new
functionality. Although a number of proposals have been made which partially
alleviate the problem in special circumstances [32, 42], catastrophic forgetting was
still listed as an open challenge in an expert’s report solicited by the NSF in 2007
[21] which collected the main future challenges in learning theory.
Recognizing dynamical patterns. Boolean conceptor operations enable the
combination of positive and negative evidence in a neural architecture for dynamical pattern recognition. For a demonstration I use a common benchmark, the
15
1
j=1
j=5
j=9
j = 13
0
ï1
0.1
0.0014
1
0.36
0.0014
j=2
0.54
0.019
j=6
0.8
0.075
j = 10
j = 14
0
ï1
0.25
0.008
1
0.36
0.011
j=3
0.59
0.015
j=7
0.83
0.013
j = 11
j = 15
0
ï1
0.29
0.00071
1
0.36
0.00071
j=4
0.65
0.036
j=8
0.94
0.05
j = 12
j = 16
0
ï1
1
0.36
0.003
10
0.42
20
1
0.004
10
0.73
20
1
0.048
10
0.99
20
1
0.95
10
20
Figure 5: Incremental pattern storing in a neural memory. Each panel shows
a 20-timestep sample of the correct training pattern pj (black line) overlaid on
its reproduction (green line). The memory fraction used up until pattern j is
indicated by the panel fraction filled in red; the quota value is printed in the left
bottom corner of each panel.
Japanese vowel recognition task [60]. The data of this benchmark consist in preprocessed audiorecordings of nine male native speakers pronouncing the Japanese
di-vowel /ae/. The training data consist of 30 recordings per speaker, the test data
consist of altogether 370 recordings, and the task is to train a recognizer which
has to recognize the speakers of the test recordings. This kind of data differs from
the periodic or chaotic patterns that I have been using so far, in that the patterns
are non-stationary (changing in their structure from beginning to end), multidimensional (each recording consisting of 12 frequency band signals), stochastic,
and of finite duration. This example thus also demonstrates that conceptors can
be put to work with data other than single-channel stationary patterns.
A small (10 neurons) reservoir was created. It was driven with all training
recordings from each speaker j in turn (j = 1, . . . , 9), collecting reservoir response
signals, from which a conceptor C j characteristic of speaker j was computed. In
addition, for each speaker j, a conceptor N j = ¬ (C 1 ∨ . . . ∨ C j−1 ∨ C j+1 ∨ . . . C 9 )
was computed. N j characterizes the condition “this speaker is not any of the
other eight speakers”. Patterns need not to be loaded into the reservoir for this
application, because they need not be re-generated.
In testing, a recording p from the test set was fed to the reservoir, collecting a reservoir response signal x. For each of the conceptors, a positive evidence
E + (p, j) = x0 C j x was computed. E + (p, j) is a non-negative number indicating
how well the signal x fits into the ellipsoid of C j . Likewise, the negative evidence E − (p, j) = x0 N j x that the sample p was not uttered by any of the eight
speakers other than speaker j was computed. Finally, the combined evidence
E(p, j) = E + (p, i) + E − (p, i) was computed. This gave nine combined evidences
16
E(p, 1), . . . , E(p, 9). The pattern p was then classified as speaker j by choosing
the speaker index j whose combined evidence E(p, j) was the greatest among the
nine collected evidences.
In order to check for the impact of the random selection of the underlying
reservoir, this whole procedure was repeated 50 times, using a freshly created
random reservoir in each trial. Averaged over these 50 trials, the number of test
misclassifications was 3.4. If the classification would have been based solely on the
positive or negative evidences, the average test misclassification numbers would
have been 8.4 and 5.9 respectively. The combination of positive and negative
evidence, which was enabled by Boolean operations, was crucial.
State-of-the-art machine learning methods achieve between 4 and 10 misclassifications on the test set (for instance [93, 99, 81, 14]). The Boolean-logic-conceptorbased classifier thus compares favorably with existing methods in terms of classification performance. The method is computationally cheap, with the entire
learning procedure taking a fraction of a second only on a standard notebook
computer. The most distinctive benefit however is incremental extensibility. If
new training data become available, or if a new speaker would be incorporated
into the recognition repertoire, the additional training can be done using only the
new data without having to re-run previous training data. This feature is highly
relevant in engineering applications and in cognitive modeling and missing from
almost all state-of-the-art classification methods.
Autoconceptors and content-addressable memories. So far I have been
describing examples where conceptors C j associated with patterns pj were computed at training time, to be later plugged in to re-generate or classify patterns.
A conceptor C matrix has the same size as the reservoir connection matrix W .
Storing conceptor matrices means to store network-sized objects. This is implausible under aspects of biological modeling. Here I describe how conceptors can be
created on the fly, without having to store them, leading to content-addressable
neural memories.
If the system has no pre-computed conceptors at its disposal, loaded patterns
can still be re-generated in a two-stage process. First, the target pattern p is
selected by driving the system with a brief initial “cueing” presentation of the
pattern (possibly in a noisy version). During this phase, a preliminary conceptor
C cue is created by an online adaptation process. This preliminary C cue already enables the system to re-generate an imperfect version of the pattern p. Second, after
the cueing phase has ended, the system continues to run in an autonomous mode
(no external cue signal), initially using C cue , to continuously generate a pattern.
While this process is running, the conceptor in the loop is continuously adapted
by a simple online adaptation rule. This rule can be described in geometrical
terms as “adapt the current conceptor C(n) such that its ellipsoid matches better
the shape of the point cloud of the current reservoir state dynamics”. Under this
rule one obtains a reliable convergence of the generated pattern toward a highly
17
accurate replica of the target pattern p that was given as a cue.
Singular Values
1
0
!1
1
0
1
A
6+:3-;*,<:*/3-.9,,3,
0
23410.56789
0
0
0
y and p
1
0
!1
1
10
1
0
!1
20 0
!0.5
!1
!1.5
5
10
B
1
2 # 4 5
()**+,-./-0+1
;./55+<1/*4=46+:--1>3
.,610+)789:
!0(#
!1
!1(#
C
2 " # $ 1&2# #0 100
)*+,-+.,/010+2/331*45
Figure 6: Content-addressable memory. A First three of five loaded patterns.
Left panels show the leading 20 singular values of C cue (black) and C auto (gray).
Right panels show an overlay of the original driver pattern (black, thin) and the
reconstruction at the end of auto-adaptation (gray, thick). B Pattern reconstruction errors directly after cueing (black squares) and at end of auto-adaptation (gray
crosses). C Reconstruction error of loaded patterns (black) and novel patterns
drawn from the same parametric family (gray) versus the number of loaded patterns, averaged over 5 repetitions of the entire experiment and 10 patterns per
plotting point. Error bars indicate standard deviations.
Results of a demonstration are illustrated in Figure 6. A 200-neuron reservoir
was loaded with 5 patterns consisting of a weighted sum of two irrational-period
sines, sampled at integer timesteps. The weight ratio and the phaseshift were
chosen at random; the patterns thus came from a family of patterns parametrized
by two parameters. The cueing time was 30 timesteps, the free-running autoadaptation time was 10,000 timesteps, leading to an auto-adapted conceptor C auto
at the end of this process. On average, the reconstruction error improved from
about -0.4 (log10 NRMSE measured directly after the cueing) to -1.1 (at the
end of auto-adaptation). It can be shown analytically that the auto-adaptation
process pulls many singular values down to zero. This effect renders the combined
reservoir-conceptor loop very robust against noise, because all noise components in
the directions of the nulled singular values become completely suppressed. In fact,
18
all results shown in Figure 6 were obtained with strong state noise (signal-to-noise
ratio equal to 1) inserted into the reservoir during the post-cue auto-adaptation.
The system functions as a content-addressable memory (CAM): loaded items
can be recalled by cueing them. The paradigmatic example of a neural CAM are
auto-associative neural networks (AANNs), pioneered by Palm [82] and Hopfield
[48]. In contrast to conceptor-based CAM, which store and re-generate dynamical
patterns, AANNs store and cue-recall static patterns. Furthermore, AANNs do not
admit an incremental storing of new patterns, which is possible in conceptor-based
CAMs. The latter thus represent an advance in neural CAMs in two fundamental
aspects.
To further elucidate the properties of conceptor CAMs, I ran a suite of simulations where the same reservoir was loaded with increasing numbers of patterns,
chosen at random from the same 2-parametric family (Figure 6 C). After loading
with k = 2, 3, 5, . . . , 100 patterns, the reconstruction accuracy was measured at
the end of the auto-adaptation. Not surprisingly, it deteriorated with increasing
memory load k (black line). In addition, I also cued the loaded reservoir with
patterns that were not loaded, but were drawn from the same family. As one
would expect, the re-construction accuracy of these novel patterns was worse than
for the loaded patterns – but only for small k. When the number of loaded patterns exceeded a certain threshold, recall accuracy became essentially equal for
loaded and novel patterns. These findings can be explained in intuitive terms as
follows. When few patterns are loaded, the network memorizes individual patterns
by “rote learning”, and subsequently can recall these patterns better than other
patterns from the family. When more patterns are loaded, the network learns
a representation of the entire parametric class of patterns. I call this the class
learning effect.
Toward biological feasibility: random feature conceptors. Several computations involved in adapting conceptor matrices are non-local and therefore biologically infeasible. It is however possible to approximate matrix conceptors with
another mechanism which only requires local computations. The idea is to project
(via random projection weights F ) the reservoir state into a random feature space
which is populated by a large number of neurons zi ; execute the conceptor operations individually on each of these neurons by multiplying a conception weight ci
into its state; and finally to project back to the reservoir by another set of random
projection weights G (Figure 7).
The original reservoir-internal random connection weigths W are replaced by a
dyade of two random projections of first F , then G, and the original reservoir state
x segregates into a reservoir state r and a random feature state z. The conception
weights ci assume the role of conceptors. They can be learnt and adapted by
procedures which are directly analog to the matrix conceptor case. What had to
be non-local matrix computations before now turns into local, one-dimensional
(scalar) operations. These operations are biologically feasible in the modest sense
19
that any information needed to adapt a synaptic weight is locally available at that
synapse. All laws and constructions concerning Boolean operations and aperture
carry over.
G
p
ci
W in
r
reservoir
F'
z
feature
space
Figure 7: Random feature conceptors. This neural architecture has two pools of
neurons, the reservoir and the feature space.
A set of conception weights ci corresponding to a particular pattern can be neurally represented and “stored” in the form of the connections of a single neuron to
the feature space. A dynamical pattern thus can be represented by a single neuron. This enables a highly compact neural representation of dynamical patterns.
A machine learning application is presented below.
I re-ran with such random feature conceptors a choice of the simulations that I
did with matrix conceptors, using a number of random features that was two to five
times as large as the reservoir. The outcome of these simulations: the accuracy
of pattern re-generation is essentially the same as with matrix conceptors, but
setting the aperture is more sensitive.
A hierarchical classification and de-noising architecture. Here I present
a system which combines in a multi-layer neural architecture many of the items
introduced so far. The input to this system is a (very) noisy signal which at a
given time is being generated by one out of a number of possible candidate pattern
generators. The task is to recognize the current generator, and simultaneously to
re-generate a clean version of the noisy input pattern.
I explain the architecture with an example. It uses three processing layers to
de-noise an input signal u[1] (n) = pj (n) + noise, with pj being one of the four
patterns p1 , . . . , p4 used before in this report (shown for instance in Figure 1 B).
The architecture implements the following design principles (Figure 8 A). (i) Each
layer is a random feature based conceptor system (as in Figure 7 B). The four
patterns p1 , . . . , p4 are initially loaded into each of the layers, and four prototype
conceptor weight vectors c1 , . . . , c4 corresponding to the patterns are computed
and stored. (ii) In a bottom-up processing pathway, the noisy external input
signal u[1] (n) = pj (n) + noise is stagewise de-noised, leading to signals y[1] , y[2] , y[3]
on layers l = 1, 2, 3, where y[3] should be a highly cleaned-up version of the input
20
%&'()*+*,'-./,(0(0
c1,
c2,
y[3]auto
![12]
c[3]
!
y[2]
y[2]auto
u[1]
%&'()*1*,'-./,(0(0
$
!"#
G
!
![23]
c[2]auto
%&'()*$*,'-./,(0(0
$
c[2]
!"#
!
F'
/)20/*3&)4&5%(0
$
!"#
G
![12]
c[1]auto
y[1]
A
$
F'
u[3]
u[2]
c4
!"#
y[3]
![23]
c3,
!
c[1]
%.8$!*9:;<=0
!
!$
F'
!1
!
6!!!
-&//()>*0&?-%(0
G
$
!
!$
!"#$%&#!'()&
B
!
$!
1!
!
7!!!
$!
1!
Figure 8: Simultaneous signal de-noising and classification. A. Schema of architecture. B. Simulation results. Panels from above: first three panels: hypothesis
j
vectors γ[l]
(n) in the three layers. Color coding: p1 blue, p2 green, p3 red, p4 cyan.
Fourth panel: trust variables τ[1,2] (n) (blue) and τ[2,3] (n) (green). Fifth panel: signal reconstruction errors (log10 NRMSE) of y[1] (blue), y[2] (green) and y[3] (red)
versus clean signal pj . Black line: linear baseline filter. Bottom panels: 20-step
samples from the end of the two presentation periods. Red: noisy input; black:
clean input; thick gray: cleaned output signal y[3] .
(subscripts [l] refer to layers, bottom layer is l = 1). (iii) The top layer autoadapts a conceptor c[3] which is constrained to be a weighted OR combination of
the four prototype conceptors. In a suggestive notation this can be written as
j
1
4
c[3] (n) = γ[3]
(n) c1 ∨ . . . ∨ γ[3]
(n) c4 . The four weights γ[3]
sum to one and represent
a hypothesis vector expressing the system’s current belief about the current driver
j
pj . If one of these γ[3]
approaches 1, the system has settled on a firm classification
of the current driving pattern. (iv) In a top-down pathway, conceptors c[l] from
layers l are passed down to the respective layers l − 1 below. Because higher layers
should have a clearer conception of the current noisy driver pattern than lower
layers, this passing-down of conceptors “primes” the processing in layer l − 1 with
valuable contextual information. (v) Between each pair of layers l, l + 1, a trust
variable τ[l,l+1] (n) is adapted by an online procedure. These trust variables range
between 0 and 1. A value of τ[l,l+1] (n) = 1 indicates maximal confidence that the
signal y[l+1] (n) comes closer to the clean driver pj (n) than the signal y[l] (n) does,
21
that is, the stage-wise denoising actually functions well when progressing from
layer l to l + 1. The trust τ[l,l+1] (n) evolves by comparing certain noise ratios that
are observable locally in layers l and l + 1. (vi) Within layer l, an internal autoauto
and a candidate
adaptation process generates a candidate de-noised signal y[l]
auto
local autoconceptor cauto
.
The
local
estimate
y
is
linearly
mixed
with the signal
[l]
[l]
auto
+
y[l−1] , where the trust τ[l−1,l] sets the mixing rate. The mixture u[l] = τ[l−1,l] y[l]
(1−τ[l−1,l] ) y[l−1] is the effective signal input to layer l. If the trust τ[l−1,l] reaches its
maximal value of 1, layer l will ignore the signal from below and work entirely by
self-generating a pattern. (vii) In a similar way, the effective conceptor in layer l
is a trust-negotiated mixture c[l] = (1 − τ[l,l+1] ) cauto
[l] + τ[l,l+1] c[l+1] . Thus if the trust
τ[l,l+1] is maximal, layer l will be governed entirely by the passed-down conceptor
c[l+1] .
Summarizing, the higher the trusts inside the hierarchy, the more will the
system be auto-generating conceptor-shaped signals, or conversely, at low trust
values the system will be strongly permeated from below by the outside driver. If
the trust variables reach their maximum value of 1, the system will run in a pure
“confabulation” mode and generate an entirely noise-free signal y[3] – at the risk
of doing this under an entirely misguided hypothesis c[3] . The key to make this
architecture work thus lies in the trust variables. It seems to me that maintaining
a measure of trust (or call it confidence, certainty, etc.) is an intrinsically necessary
component in any signal processing architecture which hosts a top-down pathway
of guiding hypotheses (or call them context, priors, bias, etc.).
Figure 8 B shows an excerpt from a simulation run. The system was driven first
by an initial 4000 step period of p1 + noise, followed by 4000 steps of p3 + noise.
The signal-to-noise ratio was 0.5 (noise twice as strong as signal). The system
successfully settles on the right hypothesis (top panel) and generates very clean
de-noised signal versions (bottom panel). The crucial item in this figure is the
development of the trust variable τ[2,3] . At the beginning of each 4000 step period
it briefly drops, allowing the external signal to permeate upwards through the
layers, thus informing the local auto-adaptation loops about “what is going on
outside”. After these initial drops the trust rises to almost 1, indicating that the
system firmly “believes” to have detected the right pattern. It then generates
pattern versions that have almost no mix-in from the noisy external driver.
As a baseline comparison I also trained a standard linear transversal filter which
computed a de-noised input pattern point based on the preceding K = 2600 input
values. The filter length K was set equal to the number of trainable parameters in
the neural architecture. The performance of this linear de-noising filter (black line
in Figure 8) is inferior to the architecture’s performance both in terms of accuracy
and response time.
It is widely believed that top-down hypothesis-passing through a processing hierarchy plays a fundamental role in biological cognitive systems [33, 15]. However,
the current best artificial pattern recognition systems [39, 59] use purely bottom-up
processing – leaving room for further improvement by including top-down guid22
ance. A few hierarchical architectures which exploit top-down hypothesis-passing
have been proposed [33, 47, 43, 35]. All of these are designed for recognizing static
patterns, especially images. The conceptor-based architecture presented here appears to be the first hierarchical system which targets dynamical patterns and uses
top-down hypothesis-passing. Furthermore, in contrast to state-of-the-art pattern
recognizers, it admits an incremental extension of the pattern repertoire.
Intrinsic conceptor logic. In mathematical logics the semantics (“meaning”)
of a symbol or operator is formalized as its extension. For instance, the symbol cow
in a logic-based knowledge representation system in AI is semantically interpreted
by the set of all (physical) cows, and the OR-operator ∨ is interpreted as set union:
cow ∨ horse would refer to the set comprising all cows and horses. Similarly, in
cognitive science, concepts are semantically referring to their extensions, usually
called categories in this context [75]. Both in mathematical logic and cognitive
science, extensions need not be confined to physical objects; the modeler may also
define extensions in terms of mathematical structures, sensory perceptions, hypothetical worlds, ideas or facts. But at any rate, there is an ontological difference
between the two ends of the semantic relationship.
This ontological gap dissolves in the case of conceptors. The natural account
of the “meaning” of a matrix conceptor C is the shape of the neural state cloud
it is derived from. This shape is given by the correlation matrix R of neural
states. Both C and R have the same mathematical format: positive semi-definite
matrices of identical dimension. Figure 9 visualizes the difference between classical extensional semantics of logics and the system-internal conceptor semantics.
The symbol |= is the standard mathematical notation for the semantical meaning
relationship.
I have cast these intuitions into a formal specification of an intrinsic conceptor
logic (ICL), where the semantic relationship outlined above is formalized within
the framework of institutions [37]. This framework has been developed in mathematics and computer science to provide a unified view on the multitude of existing
“logics”. By formalizing ICL as an institution, conceptor logic can be rigorously
compared to other existing logics. I highlight two findings. First, an ICL cast as
an institution is a dynamcial system in its own right: the symbols used in this
logic evolve over time. This is very much different from traditional views on logic,
where symbols are static tokens. Second, it turns out that ICL is a logic which
is decidable. Stated in intuitive terms, in a decidable logic it can be calculated
whether a “concept” ψ subsumes a concept ϕ (as in “a cow is an animal”). Deciding concept subsumption is a core task in AI systems and human cognition.
In most logic-based AI systems, deciding concept subsumption can become computationally expensive or even impossible. In ICL it boils down to determining
whether all components of a certain conception weight vector ci are smaller or
equal to the corresponding components c0i of another such vector, which can be
done in a single processing step. This may help explaining why humans can make
23
Figure 9: Contrasting the extensional semantics of classical knowledge representation formalisms (upper half of graphics) with conceptor semantics (lower half).
classification judgements almost instantaneously.
Discussion. The human brain is a neurodynamical system which evidently supports logico-rational reasoning [49]. Since long this has challenged scientists to
find computational models which connect neural dynamics with logic. Very different solutions have been suggested. At the dawn of computational neuroscience,
McCulloch and Pitts have already interpreted networks of binary-state neurons as
carrying out Boolean operations [74]. Logical inferences of various kinds have been
realized in localist connectionist networks where neurons are labelled by concept
names [84, 98]. In neurofuzzy modeling, feedforward neural networks are trained
to carry out operations of fuzzy logic on their inputs [11]. In a field known as
neuro-symbolic computation, deduction rules of certain formal logic systems are
coded or trained into neural networks [7, 65, 9]. The combinatorial/compositional
structure of symbolic knowledge has been captured by dedicated neural circuits to
enable tree-structured representations [85] or variable-binding functionality [106].
All of these approaches require interface mechanisms. These interface mechanisms are non-neural and code symbolic knowledge representations into the numerical activation values of neurons and/or the topological structure of networks. One
could say, previous approaches code logic into specialized neural networks, while
conceptors instantiate the logic of generic recurrent neural networks. This novel,
simple, versatile, computationally efficient, neurally not infeasible, bi-directional
connection between logic and neural dynamics opens new perspectives for computational neuroscience and machine learning.
24
2
Introduction
In this section I expand on the brief characterization of the scientific context given
in Section 1, and introduce mathematical notation.
2.1
Motivation
Intelligent behavior is desired for robots, demonstrated by humans, and studied
in a wide array of scientific disciplines. This research unfolds in two directions. In
“top-down” oriented research, one starts from the “higher” levels of cognitive performance, like rational reasoning, conceptual knowledge representation, planning
and decision-making, command of language. These phenomena are described in
symbolic formalisms developed in mathematical logic, artificial intelligence (AI),
computer science and linguistics. In the “bottom-up” direction, one departs from
“low-level” sensor data processing and motor control, using the analytical tools
offered by dynamical systems theory, signal processing and control theory, statistics and information theory. For brevity I will refer to these two directions as
the conceptual-symbolic and the data-dynamical sets of phenomena, and levels
of description. The two interact bi-directionally. Higher-level symbolic concepts
arise from low-level sensorimotor data streams in short-term pattern recognition
and long-term learning processes. Conversely, low-level processing is modulated,
filtered and steered by processes of attention, expectations, and goal-setting in a
top-down fashion.
Several schools of thought (and strands of dispute) have evolved in a decadeslong quest for a unification of the conceptual-symbolic and the data-dynamical
approaches to intelligent behavior. The nature of symbols in cognitive processes
has been cast as a philosophical issue [97, 29, 44]. In localist connectionistic models, symbolically labelled abstract processing units interact by nonlinear spreading
activation dynamics [22, 98]. A basic tenet of behavior-based AI is that higher
cognitive functions emerge from low-level sensori-motor processing loops which
couple a behaving agent into its environment [10, 83]. Within cognitive science, a
number of cognitive pheneomena have been described in terms of self-organization
in nonlinear dynamical systems [96, 100, 107]. A pervasive idea in theoretical neuroscience is to interpret attractors in nonlinear neural dynamics as the carriers of
conceptual-symbolic representations. This idea can be traced back at least to the
notion of cell assemblies formulated by Hebb [45], reached a first culmination in
the formal analysis of associative memories [82, 48, 3], and has since then diversified into a range of increasingly complex models of interacting (partial) neural
attractors [111, 105, 89, 103]. Another pervasive idea in theoretical neuroscience
and machine learning is to consider hierarchical neural architectures, which are
driven by external data at the bottom layer and transform this raw signal into
increasingly abstract feature representations, arriving at conceptual representations at the top layer of the hierarchy. Such hierarchical architectures mark the
state of the art in pattern recognition technology [66, 38]. Many of these systems
25
process their input data in a uni-directional, bottom-to-top fashion. Two notable
exceptions are systems where each processing layer is designed according to statistical principles from Bayes’ rule [33, 47, 15], and models based on the iterative
linear maps of map seeking circuits [35, 113], both of which enable top-down guidance of recognition by expectation generation. More generally, leading actors in
theoretical neuroscience have characterized large parts of their field as an effort
to understand how cognitive phenomena arise from neural dynamics [1, 36]. Finally, I point out two singular scientific efforts to design comprehensive cognitive
brain models, the ACT-R architectures developed by Anderson et al. [4] and the
Spaun model of Eliasmith et al. [26]. Both systems can simulate a broad selection of cognitive behaviors. They integrate numerous subsystems and processing
mechanisms, where ACT-R is inspired by a top-down modeling approach, starting
from cognitive operations, and Spaun from a bottom-up strategy, starting from
neurodynamical processing principles.
Despite this extensive research, the problem of integrating the conceptualsymbolic with the data-dynamical aspects of cognitive behavior cannot be considered solved. Quite to the contrary, two of the largest current research initiatives
worldwide, the Human Brain Project [18] and the NIH BRAIN initiative [51], are
ultimately driven by this problem. There are many reasons why this question is
hard, ranging from experimental challenges of gathering relevant brain data to fundamental oppositions of philosophical paradigms. An obstinate stumbling block
is the different mathematical nature of the fundamental formalisms which appear
most natural for describing conceptual-symbolic versus data-dynamical phenomena: symbolic logic versus nonlinear dynamics. Logic-oriented formalisms can
easily capture all that is combinatorially constructive and hierarchically organized
in cognition: building new concepts by logical definitions, describing nested plans
for action, organizing conceptual knowledge in large and easily extensible abstraction hierarchies. But logic is inherently non-temporal, and in order to capture
cognitive processes, additional, heuristic “scheduling” routines have to be introduced which control the order in which logical rules are executed. This is how
ACT-R architectures cope with the integration problem. Conversely, dynamical
systems formalisms are predestined for modeling all that is continuously changing in the sensori-motor interface layers of a cognitive system, driven by sensor
data streams. But when dynamical processing modules have to be combined into
compounds that can solve complex tasks, again additional design elements have
to be inserted, usually by manually coupling dynamical modules in ways that
are informed by biological or engineering insight on the side of the researcher.
This is how the Spaun model has been designed to realize its repertoire of cognitive functions. Two important modeling approaches venture to escape from
the logic-dynamics integration problem by taking resort to an altogether different
mathematical framework which can accomodate both sensor data processing and
concept-level representations: the framework of Bayesian statistics and the framework of iterated linear maps mentioned above. Both approaches lead to a unified
26
formal description across processing and representation levels, but at the price of
a double weakness in accounting for the embodiment of an agent in a dynamical
environment, and for the combinatorial aspects of cognitive operations. It appears
that current mathematical methods can instantiate only one of the three: continuous dynamics, combinatorial productivity, or a unified level-crossing description
format.
The conceptor mechanisms introduced in this report bi-directionally connect
the data-dynamical workings of a recurrent neural network (RNN) with a conceptualsymbolic representation of different functional modes of the RNN. Mathematically,
conceptors are linear operators which characterize classes of signals that are being processed in the RNN. Conceptors can be represented as matrices (convenient
in machine learning applications) or as neural subnetworks (appropriate from a
computational neuroscience viewpoint). In a bottom-up way, starting from an operating RNN, conceptors can be learnt and stored, or quickly generated on-the-fly,
by what may be considered the simplest of all adaptation rules: learning a regularized identity map. Conceptors can be combined by elementary logical operations
(AND, OR, NOT), and can be ordered by a natural abstraction relationship.
These logical operations and relations are defined via a formal semantics. Thus,
an RNN engaged in a variety of tasks leads to a learnable representation of these
operations in a logic formalism which can be neurally implemented. Conversely,
in a top-down direction, conceptors can be inserted into the RNN’s feedback loop,
where they robustly steer the RNN’s processing mode. Due to their linear algebra
nature, conceptors can be continuously morphed and “sharpened” or “defocussed”,
which extends the discrete operations that are customary in logics into the domain
of continuous “mental” transformations. I highlight the versatility of conceptors
in a series of demonstrations: generating and morphing many different dynamical
patterns with a single RNN; managing and monitoring the storing of patterns
in a memory RNN; learning a class of dynamical patterns from presentations of
a small number of examples (with extrapolation far beyond the training examples); classification of temporal patterns; de-noising of temporal patterns; and
content-addressable memory systems. The logical conceptor operations enable an
incremental extension of a trained system by incorporating new patterns without
interfering with already learnt ones. Conceptors also suggest a novel answer to a
perennial problem of attractor-based models of concept representations, namely
the question of how a cognitive trajectory can leave an attractor (which is at odds
with the very nature of an attractor). Finally, I outline a version of conceptors
which is biologically plausible in the modest sense that only local computations
and no information copying are needed.
2.2
Mathematical Preliminaries
I assume that the reader is familiar with properties of positive semidefinite matrices, the singular value decomposition, and (in some of the analysis of adaptation
27
dynamics) the usage of the Jacobian of a dynamical system for analysing stability
properties.
[a, b], (a, b), (a, b], [a, b) denote the closed (open, half-open) interval between real
numbers a and b.
A0 or x0 denotes the transpose of a matrix A or vector x. I is the identity
matrix (the size will be clear from the context or be expressed as In×n ). The ith
unit vector is denoted by ei (dimension will be clear in context). The trace of a
square matrix A is denoted by tr A. The singular value decomposition of a matrix
A is written as U SV 0 = A, where U, V are orthonormal and S is the diagonal
matrix containing the singular values of A, assumed to be in descending order
unless stated otherwise. A† is the pseudoinverse of A. All matrices and vectors
will be real and this will not be explicitly mentioned.
I use the Matlab notation to address parts of vectors and matrices, for instance
M (:, 3) is the third column of a matrix M and M (2 : 4, :) picks from M the
submatrix consisting of rows 2 to 4. Furthermore, again like in Matlab, I use the
operator diag in a “toggling” mode: diag A returns the diagonal vector of a square
matrix A, and diag d constructs a diagonal matrix from a vector d of diagonal
elements. Another Matlab notation that will be used is “.∗” for the element-wise
multiplication of vectors and matrices of the same size, and “.∧ ” for element-wise
exponentation of vectors and matrices.
R(A) and N (A) denote the range and null space of a matrix A. For linear
subspaces S, T of Rn , S ⊥ is the orthogonal complement space of S and S +
T is the direct sum {x + y | x ∈ S, y ∈ T } of S and T . PS is the n × n
dimensional projection matrix on a linear subspace S of Rn . For a k-dimensional
linear subspace S of Rn , BS denotes any n × k dimensional matrix whose columns
form an orthonormal basis of S. Such matrices BS will occur only in contexts
where the choice of basis can be arbitrary. It holds that PS = BS (BS )0 .
E[x(n)] denotes the expectation (temporal average) of a stationary signal x(n)
(assuming it is well-defined, for instance, coming from an ergodic source).
For a matrix M , kM kfro is the Frobenius norm of M . For real M , it is the
square root of the summed squared elements of M . If M is positive semidefinite
with SVD M = U SU 0 , kM kfro is the same as the 2-norm of the diagonal vector
of S, i.e. kM kfro = ((diagS)0 (diagS))1/2 . Since in this report I will exclusively use
the Frobenius norm for matrices, I sometimes omit the subscript and write kM k
for simplicity.
In a number of simulation experiments, a network-generated signal y(n) will be
matched against a target pattern p(n). The accuracy of the
p match will be quantified by the normalized root mean square error (NRMSE), [(y(n) − p(n))2 ]/[(p(n)2 ],
where [·] is the mean operator over data points n.
The symbol N is reserved for the size of a reservoir (= number of neurons)
throughout.
28
3
Theory and Demonstrations
This is the main section of this report. Here I develop in detail the concepts,
mathematical analysis, and algorithms, and I illustrate various aspects in computer
simulations.
Figure 10 gives a navigation guide through the dependency tree of the components of this section.
Basic theory and
usage of conceptors:
3.1 – 3.6
Morphing
patterns demo:
3.7
How conceptors relate
to neural data; aperture;
chaotic attractor demo:
3.8
Boolean conceptor
logic and
abstraction:
3.9, 3.10
Memory
management
demo:
3.11
Pattern
recognition
demo:
3.12
Autoconceptors:
3.13
random feature
conceptors, biological
plausibility:
3.14
Hierarchical filtering
architecture:
3.15
Conceptor logic:
3.16, 3.17
Figure 10: Dependency tree of subsections in Section 3.
The program code (Matlab) for all simulations can be retrieved from
http://minds.jacobs-university.de/sites/default/files/uploads/...\..
.SW/ConceptorsTechrepR1Matlab.zip.
3.1
Networks and Signals
Throughout this report, I will be using discrete-time recurrent neural networks
made of simple tanh neurons, which will be driven by an input time series p(n).
In the case of 1-dimensional input, these networks consist of (i) a “reservoir”
of N recurrently connected neurons whose activations form a state vector x =
(x1 , . . . , xN )0 , (ii) one external input neuron that serves to drive the reservoir with
training or cueing signals p(n) and (iii) another external neuron which serves to
read out a scalar target signal y(n) from the reservoir (Fig. 11). The system
29
operates in discrete timesteps n = 0, 1, 2, . . . according to the update equations
x(n + 1) = tanh(W x(n) + W in p(n + 1) + b)
y(n) = W out x(n),
(1)
(2)
where W is the N × N matrix of reservoir-internal connection weights, W in is the
N × 1 sized vector of input connection weights, W out is the 1× N vector of readout
weights, and b is a bias. The tanh is a sigmoidal function that is applied to the
network state x component-wise. Due to the tanh, the reservoir state space or
simply state space is (−1, 1)N .
The input weights and the bias are fixed at random values and are not subject
to modification through training. The output weights W out are learnt. The reservoir weights W are learnt in some of the case studies below, in others they remain
fixed at their initial random values. If they are learnt, they are adapted from a
random initialization denoted by W ∗ . Figure 11 A illustrates the basic setup.
I will call the driving signals p(n) patterns. In most parts of this report,
patterns will be periodic. Periodicity comes in two variants. First, integer-periodic
patterns have the property that p(n) = p(n+k) for some positive integer k. Second,
irrational-periodic patterns are discretely sampled from continuous-time periodic
signals, where the sampling interval and the period length of the continuous-time
√
signal have an irrational ratio. An example is p(n) = sin(2 π n/(10 2)). These
two sorts of drivers will eventually lead to different kinds of attractors trained into
reservoirs: integer-periodic signals with period length P yield attractors consisting
of P points in reservoir state space, while irrational-periodic signals give rise to
attracting sets which can be topologically characterized as one-dimensional cycles
that are homeomorphic to the unit cycle in R2 .
3.2
Driving a Reservoir with Different Patterns
A basic theme in this report is to develop methods by which a collection of different
patterns can be loaded in, and retrieved from, a single reservoir. The key for
these methods is an elementary dynamical phenomenon: if a reservoir is driven
by a pattern, the entrained network states are confined to a linear subspace of
network state space which is characteristic of the pattern. In this subsection I
illuminate this phenomenon by a concrete example. This example will be re-used
and extended on several occasions throughout this report.
I use four patterns. The first two are irrational periodic and the last two
are integer-periodic: (1) a sinewave of period ≈ 8.83 sampled at integer times
(pattern p1 (n)) (2) a sinewave p2 (n) of period ≈ 9.83 (period of p1 (n) plus 1), (3)
a random 5-periodic pattern p3 (n) and (4) a slight variation p4 (n) thereof (Fig.
12 left column).
A reservoir with N = 100 neurons is randomly created. At creation time the
input weights W in and the bias b are fixed at random values; these will never be
modified thereafter. The reservoir weights are initialized to random values W ∗ ;
30
$"
p j (n) x ij (n)
W
p W
!
!
!
!"
x lj (n "1)
in
Wi
!
W out y
!
x
!
!
!
!
!
#"
!
!
x kj (n "1)
W i*
x lj (n "1)
j
x mj (n "1)
!
x kj (n "1)
x i (n)
!
!
in
x mj (n "1)
Wi
!
! 11: A. Basic system setup. Through input connections
Figure
W in , an input neuron feeds a driving signal p to a “reservoir” of N = 100 neurons which are recur! W . From the N -dimensional
rently connected to each other through connections
neuronal activation state x, an output signal y is read out by connections W out .
All broken connections are trainable. B. During the initial driving of the reservoir
with driver pj , using initial random weights W ∗ , neuron xi produces its signal
(thick gray line) based on external driving input p and feeds from other neurons x
from within the reservoir (three shown). C. After training new reservoir weights
W , the same neuron should produce the same signal based only on the feeds from
other reservoir neurons.
in this first demonstration they will not be subsequently modified either. The
readout weights are initially undefined (details in Section 4.1).
In four successive and independent runs, the network is driven by feeding the
respective pattern pj (n) as input (j = 1, . . . , 4), using the update rule
xj (n + 1) = tanh(W ∗ xj (n) + W in pj (n + 1) + b).
After an initial washout time, the reservoir dynamics becomes entrained to the
driver and the reservoir state xj (n) exhibits an involved nonlinear response to the
driver pj . After this washout, the reservoir run is continued for L = 1000 steps,
and the obtained states xj (n) are collected into N × L = 100 × 1000 sized state
collection matrices X j for subsequent use.
The second column in Fig. 12 shows traces of three randomly chosen reservoir
neurons in the four driving conditions. It is apparent that the reservoir has become
entrained to the driving input. Mathematically, this entrainment is captured by
the concept of the echo state property: any random initial state of a reservoir
is “forgotten”, such that after a washout period the current network state is a
31
function of the driver. The echo state property is a fundamental condition for
RNNs to be useful in learning tasks [54, 12, 46, 112, 101, 71]. It can be ensured
by an appropriate scaling of the reservoir weight matrix. All networks employed
in this report possess the echo state property.
driver and y
reservoir states
1
0
!1
0.0064
1
0
!1
0.0059
1
0
!1
0.091
1
0
!1
0
0.088
10
20
0
10
log10 PC energy
10
0
!10
!20
0
50
100
10
0
!10
!20
0
50
100
10
0
!10
!20
0
50
100
10
0
!10
!20
20
0
50
100
leading PC energy
40
20
0
0
40
5
10
5
10
5
10
5
10
20
0
0
40
20
0
0
40
20
0
0
Figure 12: The subspace phenomenon. Each row of panels documents situation
when the reservoir is driven by a particular input pattern. “Driver and y”: the
driving pattern (thin black line) and the signals retrieved with conceptors (broad
light gray line). Number inset is the NRMSE between original driver and retrieved
signal. “Reservoir states”: activations of three randomly picked reservoir neurons.
“Log10 PC energy”: log10 of reservoir signal energies in the principal component
directions. “Leading PC energy”: close-up on first ten signal energies in linear
scale. Notice that the first two panels in each row show discrete-time signals;
points are connected by lines only for better visual appearance.
A principal component analysis (PCA) of the 100 reservoir signals reveals
that the driven reservoir signals are concentrated on a few principal directions.
Concretely, for each of the four driving conditions, the reservoir state correlation
matrix was estimated by Rj = X j (X j )0 /L, and its SVD U j Σj (U j )0 = Rj was computed, where the columns of U j are orthonormal eigenvectors of Rj (the principal
component (PC) vectors), and the diagonal of Σj contains the singular values of
Rj , i.e. the energies (mean squared amplitudes) of the principal signal components.
Figure 12 (third and last column) shows a plot of these principal component energies. The energy spectra induced by the two irrational-period sines look markedly
different from the spectra obtained from the two 5-periodic signals. The latter lead
to nonzero energies in exactly 5 principal directions because the driven reservoir
dynamics periodically visits 5 states (the small but nonzero values in the log10
32
plots in Figure 12 are artefacts earned from rounding errors in the SVD computation). In contrast, the irrational-periodic drivers lead to reservoir states which
linearly span all of RN (Figure 12, upper two log10 plots). All four drivers however
share a relevant characteristic (Figure 12, right column): the total reservoir energy
is concentrated in a quite small number of leading principal directions.
When one inspects the excited reservoir dynamics in these four driving conditions, there is little surprise that the neuronal activation traces look similar to
each other for the first two and in the second two cases (Figure 12, second column). This “similarity” can be quantified in a number of ways. Noting that the
geometry of the “reservoir excitation space” in driving condition j is characterized
by a hyperellipsoid with main axes U j and axis lengths diag Σj , a natural way to
define a similarity between two such ellipsoids i, j is to put
simR
i,j =
k(Σi )1/2 (U i )0 U j (Σj )1/2 k2
.
kdiagΣi k kdiagΣj k
(3)
The measure simR
i,j ranges in [0, 1]. It is 0 if and only if the reservoir signals
i
j
x , x populate orthogonal linear subspaces, and it is 1 if and only if Ri = a Rj
for some scaling factor a. The measure simR
i,j can be understood as a generalized
squared cosine between Ri and Rj . Figure 13 A shows the similarity matrix
(simR
i,j )i,j obtained from (3). The similarity values contained in this matrix appear
somewhat counter-intuitive, inasmuch as the reservoir responses to the sinewave
patterns come out as having similarities of about 0.6 with the 5-periodic driven
reservoir signals; this does not agree with the strong visual dissimilarity apparent
in the state plots in Figure 12. In Section 3.5 I will introduce another similarity
measure which agrees better with intuitive judgement.
A
B
C
Figure 13: Matrix plots of pairwise similarity between the subspaces excited
in the four driving conditions. Grayscale coding: 0 = black, 1 = white. A:
i
similarity simR
i,j based on the data correlation matrices R . B,C: similarities based
i
on conceptors C(R , α) for two different values of aperture α. For explanation see
text.
33
3.3
Storing Patterns in a Reservoir, and Training the Readout
One of the objectives of this report is a method for storing several driving patterns
in a single reservoir, such that these stored patterns can later be retrieved and
otherwise be controlled or manipulated. In this subsection I explain how the
initial “raw” reservoir weights W ∗ are adapted in order to “store” or “memorize”
the drivers, leading to a new reservoir weight matrix W . I continue with the
four-pattern-example used above.
The guiding idea is to enable the reservoir to re-generate the driven responses
j
x (n) in the absence of the driving input. Consider any neuron xi (Fig. 11B).
During the driven runs j = 1, . . . , 4, it has been updated per
xji (n + 1) = tanh(Wi∗ xj (n) + Wiin pj (n + 1) + bi ),
where Wi∗ is the i-th row in W ∗ , Wiin is the i-th element of W in , and bi is the
ith bias component. The objective for determining new reservoir weights W is
that the trained reservoir should be able to oscillate in the same four ways as in
the external driving conditions, but without the driving input. That is, the new
weights Wi leading to neuron i should approximate
tanh(Wi∗ xj (n) + Wiin pj (n + 1) + bi ) ≈ tanh(Wi xj (n) + bi )
as accurately as possible, for j = 1, . . . , 4. Concretely, we optimize a mean square
error criterion and compute
X
X
(Wi∗ xj (n) + Wiin pj (n + 1) − W̃i xj (n))2 , (4)
Wi = argminW̃i
j=1,...,K n=1,...,L
where K is the number of patterns to be stored (in this example K = 4). This is
a linear regression task, for which a number of standard algorithms are available.
I employ ridge regression (details in Section 4.1).
The readout neuron y serves as passive observer of the reservoir dynamics.
The objective to determine its connection weights W out is simply to replicate the
driving input, that is, W out is computed (again by ridge regression) such that it
minimizes the squared error (pj (n) − W out xj (n))2 , averaged over time and the four
driving conditions.
I will refer to this preparatory training as storing patterns pj in a reservoir,
and call a reservoir loaded after patterns have been stored.
3.4
Conceptors: Introduction and Basic Usage in Retrieval
How can these stored patterns be individually retrieved again? After all, the
storing process has superimposed impressions of all patterns on all of the recomputed connection weights W of the network – very much like the pixel-wise
34
addition of different images would yield a mixture image in which the individual
original images are hard to discern. One would need some sort of filter which can
disentangle again the superimposed components in the connection weights. In this
section I explain how such filters can be obtained.
The guiding idea is that for retrieving pattern j from a loaded reservoir, the
reservoir dynamics should be restricted to the linear subspace which is characteristic for that pattern. For didactic reasons I start with a simplifying assumption (to
be dropped later). Assume that there exists a (low-dimensional) linear subspace
S j ⊂ RN such that all state vectors contained in the driven state collection X j lie
in S j . In our example, this is actually the case for the two 5-periodic patterns. Let
PS j be the projector matrix which projects RN on S j . We may then hope that
if we run the loaded reservoir autonomously (no input), constraining its states to
S j using the update rule
x(n + 1) = PS j tanh(W x(n) + b),
(5)
it will oscillate in a way that is closely related to the way how it oscillated when
it was originally driven by pj .
However, it is not typically the case that the states obtained in the original
driving conditions are confined to a proper linear subspace of the reservoir state
space. Consider the sine driver p1 in our example. The linear span of the reservoir
response state is all of RN (compare the log10 PC energy plots in Figure 12).
The associated projector would be the identity, which would not help to single
out an individual pattern in retrieval. But actually we are not interested in those
principal directions of reservoir state space whose excitation energies are negligibly
small (inspect again the quick drop of these energies in the third column, top panel
in Figure 12 – it is roughly exponential over most of the spectrum, except for an
even faster decrease for the very first few singular values). Still considering the
sinewave pattern p1 : instead of PRN we would want a projector that projects on
the subspace spanned by a “small” number of leading principal components of
the “excitation ellipsoid” described by the sine-driver-induced correlation matrix
R1 . What qualifies as a “small” number is, however, essentially arbitrary. So we
want a method to shape projector-like matrices from reservoir state correlation
matrices Rj in a way that we can adjust, with a control parameter, how many of
the leading principal components should become registered in the projector-like
matrix.
At this point I give names to the projector-like matrices and the adjustment
parameter. I call the latter the aperture parameter, denoted by α. The projectorlike matrices will be called conceptors and generally be denoted by the symbol
C. Since conceptors are derived from the ellipsoid characterized by a reservoir
state corrlation matrix Rj , and parametrized by the aperture parameter, I also
sometimes write C(Rj , α) to make this dependency transparent.
There is a natural and convenient solution to meet all the intuitive objectives
for conceptors that I discussed up to this point. Consider a reservoir driven by
35
a pattern pj (n), leading to driven states xj (n) collected (as columns) in a state
collection matrix X j , which in turn yields a reservoir state correlation matrix
Rj = X j (X j )0 /L. We define a conceptor C(Rj , α) with the aid of a cost function L(C | Rj , α), whose minimization yields C(Rj , α). The cost function has two
components. The first component reflects the objective that C should behave as
a projector matrix for the states that occur in the pattern-driven run of the reservoir. This component is En [kxj (n) − Cxj (n)k2 ], the time-averaged deviation of
projections Cxj from the state vectors xj . The second component of L adjusts how
many of the leading directions of Rj should become effective for the projection.
This component is α−2 kCk2fro . This leads to the following definition.
Definition 1 Let R = E[xx0 ] be an N × N correlation matrix and α ∈ (0, ∞).
The conceptor matrix C = C(R, α) associated with R and α is
C(R, α) = argminC E[kx − Cxk2 ] + α−2 kCk2fro .
(6)
The minimization criterion (6) uniquely specifies C(R, α). The conceptor matrix can be effectively computed from R and α. This is spelled out in the following
proposition, which also lists elementary algebraic properties of conceptor matrices:
Proposition 1 Let R = E[x x0 ] be a correlation matrix and α ∈ (0, ∞). Then,
1. C(R, α) can be directly computed from R and α by
C(R, α) = R (R + α−2 I)−1 = (R + α−2 I)−1 R,
(7)
2. if R = U ΣU 0 is the SVD of R, then the SVD of C(R, α) can be written as
C = U SU 0 , i.e. C has the same principal component vector orientation as
R,
3. the singular values si of C relate to the singular values σi of R by si =
σi /(σi + α−2 ),
4. the singular values of C range in [0, 1),
5. R can be recovered from C and α by
R = α−2 (I − C)−1 C = α−2 C (I − C)−1 .
(8)
The proof is given in Section 5.1. Notice that all inverses appearing in this
proposition are well-defined because α > 0 is assumed, which implies that all
singular values of C(R, α) are properly smaller than 1. I will later generalize
conceptors to include the limiting cases α = 0 and α = ∞ (Section 3.8.1).
In practice, the correlation matrix R = E[xx0 ] is estimated from a finite sample
X, which leads to the approximation R̂ = XX 0 /L, where X = (x(1), . . . , x(L)) is
a matrix containing reservoir states x(n) collected during a learning run.
36
Figure 14 shows the singular value spectra of C(R, α) for various values of α,
for our example cases of R = R1 (irrational-period sine driver) and R = R3 (5periodic driver). We find that the nonlinearity inherent in (7) makes the conceptor
matrices come out “almost” as projector matrices: the singular values of C are
mostly close to 1 or close to 0. In the case of the 5-periodic driver, where the
excited network states populate a 5-dimensional subspace of RN , increasing α lets
C(R, α) converge to a projector onto that subspace.
sine (pattern 1)
10ïperiodic random (pattern 3)
1
0
0
1
50
0
0
100
_=1
_ = 10
_ = 100
_ = 1000
_ = 10000
50
100
Figure 14: How the singular values of a conceptor depend on α. Singular value
spectra are shown for the first sinewave pattern and the first 5-periodic random
pattern. For explanation see text.
If one has a conceptor matrix C j = C(Rj , α) derived from a pattern pj through
the reservoir state correlation matrix Rj associated with that pattern, the conceptor matrix can be used in an autonomous run (no external input) using the update
rule
x(n + 1) = C j tanh(W x(n) + b),
(9)
where the weight matrix W has been shaped by storing patterns among which
there was pj . Returning to our example, four conceptors C 1 , . . . , C 4 were computed
with α = 10 and the loaded reservoir was run under rule (9) from a random initial
state x(0). After a short washout period, the network settled on stable periodic
dynamics which were closely related to the original driving patterns. The network
dynamics was observed through the previously trained output neuron. The left
column in Figure 12 shows the autonomous network output as a light bold gray line
underneath the original driver. To measure the achieved accuracy, the autonomous
output signal was phase-aligned with the driver (details in Section 4.1) and then
the NRMSE was computed (insets in Figure panels). The NRMSEs indicate that
the conceptor-constrained autonomous runs could successfully separate from each
other even the closely related pattern pairs p1 versus p2 and p3 versus p4 .
A note on terminology. Equation (9) shows a main usage of conceptor matrices: they are inserted into the reservoir state feedback loop and cancel (respectively, dampen) those reservoir state components which correspond to directions
in state space associated with zero (or small, respectively) singular values in the
37
conceptor matrix. In most of this report, such a direction-selective damping in
the reservoir feedback loop will be effected by way of inserting matrices C like
in Equation (9). However, inserting a matrix is not the only way by which such
a direction-selective damping can be achieved. In Section 3.15, which deals with
biological plausibility issues, I will propose a neural circuit which achieves a similar functionality of direction-specific damping of reservoir state components by
other means and with slightly differing mathematical properties. I understand the
concept of a “conceptor” as comprising any mechanism which effects a patternspecific damping of reservoir signal components. Since in most parts of this report
this will be achieved with conceptor matrices, as in (9), I will often refer to these C
matrices as “conceptors” for simplicity. The reader should however bear in mind
that the notion of a conceptor is more comprehensive than the notion of a conceptor matrix. I will not spell out a formal definition of a “conceptor”, deliberately
leaving this concept open to become instantiated by a variety of computational
mechanisms of which only two are formally defined in this report (via conceptor
matrices, and via the neural circuit given in Section 3.15).
3.5
A Similarity Measure for Excited Network Dynamics
In Figure 13 A a similarity matrix is presented which compares the excitation
ellipsoids represented by the correlation matrices Rj by the similarity metric (3).
I remarked at that time that this is not a fully satisfactory metric, because it
does not agree well with intuition. We obtain a more intuitively adequate similiarity metric if conceptor matrices are used as descriptors of “subspace ellipsoid
geometry” instead of the raw correlation matrices, i.e. if we employ the metric
simαi,j
k(S i )1/2 (U i )0 U j (S j )1/2 k2
,
=
kdiagS i k kdiagS j k
(10)
where U S j U 0 is the SVD of C(Rj , α). Figure 13 B,C shows the similarity matrices arising in our standard example for α = 10 and α = 10, 000. The intuitive
dissimilarity between the sinewave and the 5-periodic patterns, and the intuitive
similarity between the two sines (and the two 5-periodic pattern versions, respectively) is revealed much more clearly than on the basis of simR
i,j .
R
α
When interpreting similarities simi,j or simi,j one should bear in mind that one
is not comparing the original driving patterns but the excited reservoir responses.
3.6
Online Learning of Conceptor Matrices
The minimization criterion (6) immediately leads to a stochastic gradient online
method for adapting C:
Proposition 2 Assume that a stationary source x(n) of N -dimensional reservoir
states is available. Let C(1) be any N × N matrix, and λ > 0 a learning rate.
38
Then the stochastic gradient adaptation
C(n + 1) = C(n) + λ (x(n) − C(n) x(n)) x0 (n) − α−2 C(n)
(11)
will lead to limλ↓0 limn→∞ C(n) = C(E[x x0 ], α).
The proof is straightforward if one employs generally known facts about stochastic
gradient descent and the fact that E[kx − Cxk2 ] + α−2 kCk2fro is positive definite
quadratic in the N 2 -dimensional space of elements of C (shown in the proof of
Proposition 1), and hence provides a Lyapunov function for the gradient descent
(11). The gradient of E[kx − Cxk2 ] + α−2 kCk2fro with respect to C is
∂
E[kx − Cxk2 ] + α−2 kCk2fro = (I − C) E[xx0 ] − α−2 C,
∂C
(12)
which immediately yields (11).
The stochastic update rule (11) is very elementary. It is driven by two components, (i) an error signal x(n) − C(n) x(n) which simply compares the current
state with its C-mapped value, and (ii) a linear decay term. We will make heavy
use of this adaptive mechanism in Sections 3.14.1 ff. This observation is also illuminating the intuitions behind the definition of conceptors. The two components
strike a compromise (balanced by α) between (i) the objective that C should leave
reservoir states from the target pattern unchanged, and (ii) C should have small
weights. In the terminology of machine learning one could say, “a conceptor is a
regularized identity map”.
3.7
Morphing Patterns
Conceptor matrices offer a way to morph RNN dynamics. Suppose that a reservoir has been loaded with some patterns, among which there are pi and pj with
corresponding conceptors C i , C j . Patterns that are intermediate between pi and
pj can be obtained by running the reservoir via (9), using a linear mixture between
C i and C j :
x(n + 1) = (1 − µ)C i + µC j tanh(W x(n) + b).
(13)
Still using our four-pattern example, I demonstrate how this morphing works out
for morphing (i) between the two sines, (ii) between the two 5-periodic patterns,
(iii) between a sine and a 5-periodic pattern.
Frequency Morphing of Sines In this demonstration, the morphing was done
for the two sinewave conceptors C 1 = C(R1 , 10) and C 2 = C(R2 , 10). The morphing parameter µ was allowed to range from −2 to +3 (!). The four-pattern-loaded
reservoir was run from a random initial state for 500 washout steps, using (13)
with µ = −2. Then recording was started. First, the run was continued with the
intial µ = −2 for 50 steps. Then, µ was linearly ramped up from µ = −2 to µ = 3
during 200 steps. Finally, another 50 steps were run with the final setting µ = 3.
39
Note that morph values µ = 0 and µ = 1 correspond to situations where the
reservoir is constrained by the original conceptors C 1 and C 2 , respectively. Values
0 ≤ µ ≤ 1 correspond to interpolation. Values −2 ≤ µ < 0 and 1 < µ ≤ 3
correspond to extrapolation. The extrapolation range on either side is twice as
long as the interpolation range.
In addition, for eight equidistant values µk in −2 ≤ µ < 3, the reservoir was
run with a mixed conceptor C = (1−µk )C 1 +µk C 2 for 500 steps, and the obtained
observation signal y(n) was plotted in a delay-embedded representation, yielding
“snapshots” of the reservoir dynamics at these µ values (a delay-embedding plot
of a 1-dimensional signal y(n) creates a 2-dimensional plot by plotting value pairs
(y(n), y(n − d)) with a delay d chosen to yield an appealing visual appearance).
#
!
!#
!
#"
"!
#!!
#"!
$!!
$"!
%!!
#!
"
Figure 15: Morphing between (and beyond) two sines. The morphing range was
−2 ≤ µ ≤ 3. Black circular dots in the two bottom panels mark the points
µ = 0 and µ = 1, corresponding to situations where the two original conceptors
C 1 , C 2 were active in unadulterated form. Top: Delay-embedding plots of network
observation signal y(n) (delay = 1 step). Thick points show 25 plotted points,
thin points show 500 points (appearing as connected line). The eight panels have
a plot range of [−1.4, 1.4] × [−1.4, 1.4]. Triangles in center panel mark the morph
positions corresponding to the delay embedding “snapshots”. Center: the network
observation signal y(n) of a morph run. Bottom: Thin black line: the period length
obtained from morphing between (and extrapolating beyond) the original period
lengths. Bold gray line: period lengths measured from the observation signal y(n).
Figure 15 shows the findings. The reservoir oscillates over the entire inter/extrapolation range with a waveform that is approximately equal to a sampled
sine. At the morph values µ = 0 and µ = 1 (indicated by dots in the Figure), the
system is in exactly the same modes as they were plotted earlier in the first two
panels of the left column in Figure 12. Accordingly the fit between the original
40
driver’s period lenghtes and the autonomously re-played oscillations is as good as
it was reported there (i.e. corresponding to a steady-state NRMSE of about 0.01).
In the extrapolation range, while the linear morphing of the mixing parameter µ
does not lead to an exact linear morphing of the observed period lengths, still the
obtained period lengths steadily continue to decrease (going left from µ = 0) and
to increase (going right from µ = 1).
In sum, it is possible to use conceptor-morphing to extend sine-oscillatory
reservoir dynamics from two learnt oscillations of periods ≈ 8.83, 9.83 to a range
between ≈ 7.5 − 11.9 (minimal and maximal values of period lengths shown in
the Figure). The post-training sinewave generation thus extrapolated beyond the
period range spanned by the two training samples by a factor of about 4.4. From
a perspective of machine learning this extrapolation is remarkable. Generally
speaking, when neural pattern generators are trained from demonstration data
(often done in robotics, e.g. [52, 91]), interpolation of recallable patterns is what
one expects to achieve, while extrapolation is deemed hard.
From a perspective of neurodynamics, it is furthermore remarkable that the
dimension of interpolation/extrapolation was the speed of the oscillation. Among
the infinity of potential generalization dimensions of patterns, speedup/slowdown
of pattern generation has a singular role and is particularly difficult to achieve.
The reason is that speed cannot be modulated by postprocessing of some underlying generator’s output – the prime generator itself must be modulated [110].
Frequency adaptation of neural oscillators is an important theme in research on
biological pattern generators (CPGs) (reviews: [40, 50]). Frequency adaptation
has been modeled in a number of ways, among which (i) to use a highly abstracted
CPG model in the form of an ODE, and regulate speed by changing the ODE’s
time constant; (ii) to use a CPG model which includes a pacemaker neuron whose
pace is adaptive; (iii) to use complex, biologically quite detailed, modular neural architectures in which frequency adapatation arises from interactions between
modules, sensor-motoric feedback cycles, and tonic top-down input. However,
the fact that humans can execute essentially arbitrary motor patterns at different
speeds is not explained by these models. Presumably this requires a generic speed
control mechanism which takes effect already at higher (cortical, planning) layers
in the motor control hierarchy. Conceptor-controlled frequency adaptation might
be of interest as a candidate mechanism for such a “cognitive-level” generic speed
control mechanism.
Shape Morphing of an Integer-Periodic Pattern In this demonstration,
the conceptors C(R3 , 1000) and C(R4 , 1000) from the 5-periodic patterns p3 and
p4 were morphed, again with −2 ≤ µ ≤ 3. Figure 16 depicts the network observer
y(n) for a morph run of 95 steps which was started with µ = −2 and ended with
µ = 3, with a linear µ ramping in between. It can be seen that the differences
between the two reference patterns (located at the points marked by dots) become
increasingly magnified in both extrapolation segments. At each of the different
41
points in each 5-cycle, the “sweep” induced by the morphing is however neither
linear nor of the same type across all 5 points of the period (right panel). A simple
algebraic rule that would describe the geometric characteristics of such morphings
cannot be given. I would like to say, it is “up to the discretion of the network’s
nonlinear dynamics” how the morphing command is interpreted; this is especially
true for the extrapolation range. If reservoirs with a different initial random W ∗
are used, different morphing geometries arise, especially at the far ends of the
extrapolation range (not shown).
!
"
!!
!"
#"
$"
%"
&"
'"
("
)"
*"
Figure 16: Morphing between, and extrapolating beyond, two versions of a 5periodic random pattern. The morphing range was −2 ≤ µ ≤ 3. Bottom: Network
observation from a morphing run. Dots mark the points µ = 0 and µ = 1,
corresponding to situations where the two original conceptors C 1 , C 2 were active
in unadulterated form. The network observation signal y(n) is shown. Top: Delayembedding “snapshots”. Figure layout similar to Figure 15.
The snapshots displayed in Figure 16 reveal that the morphing sweep takes
the reservoir through two bifurcations (apparent in the transition from snapshot 2
to 3, and from 7 to 8). In the intermediate morphing range (snapshots 3 – 7), we
observe a discrete periodic attractor of 5 points. In the ranges beyond, on both
sides the attracting set becomes topologically homomorphic to a continuous cycle.
From a visual inspection, it appears that these bifurcations “smoothly” preserve
some geometrical characteristics of the observed signal y(n). A mathematical
characterisation of this phenomenological continuity across bifurcations remains
for future investigations.
Heterogeneous Pattern Morphing Figure 17 shows a morph from the 5periodic pattern p3 to the irrational-periodic sine p2 (period length ≈ 9.83). This
time the morphing range was 0 ≤ µ ≤ 1, (no extrapolation). The Figure shows
a run with an initial 25 steps of µ = 0, followed by a 50-step ramp to µ = 1 and
a tail of 25 steps at the same µ level. One observes a gradual change of signal
shape and period along the morph. From a dynamical systems point of view
this gradual change is unexpected. The reservoir is, mathematically speaking, an
42
autonomous system under the influence of a slowly changing control parameter µ.
On both ends of the morph, the system is in an attractor. The topological nature
of the attractors (seen as subsets of state space) is different (5 isolated points vs. a
homolog of a 1-dim circle), so there must be a at least one bifurcation taking place
along the morphing route. Such a bifurcations would usually be accompanied by
a sudden change of some qualitative characteristic of the system trajectory. We
find however no trace of a dynamic rupture, at least not by visual inspection of
the output trajectory. Again, a more in-depth formal characterization of what
geometric properties are “smoothly” carried through these bifurcations is left for
future work.
Figure 2 in Section 1 is a compound demonstration of the three types of pattern
morphing that I here discussed individually.
A possible application for the pattern morphing by conceptors is to effect
smooth gait changes in walking robots, a problem that is receiving some attention
in that field.
"
!
!"
!
"!
#!
$!
%!
&!
'!
(!
)!
*!
"!!
"!
&
Figure 17: Morphing from a 5-periodic random pattern to an irrational-periodic
sine. The morphing range was 0 ≤ µ ≤ 1. Figure layout otherwise is as in Figure
15.
3.8
3.8.1
Understanding Aperture
The Semantics of α as “Aperture”
Here I show how the parameter α can be interpreted as a scaling of signal energy,
and motivate why I call it “aperture”.
We can rewrite C(R, α) = C(E[xx0 ], α) as follows:
43
C(E[xx0 ], α) = E[xx0 ](E[xx0 ] + α−2 I)−1 = E[(αx)(αx)0 ](E[(αx)(αx)0 ] + I)−1
= C(E[(αx)(αx)0 ], 1) = C(α2 E[xx0 ], 1).
(14)
Thus, changing from C(R, 1) to C(R, α) can be interpreted as scaling the
reservoir data by a factor of α, or expressed in another way, as scaling the signal
energy of the reservoir signals by a factor of α2 . This is directly analog to what
adjusting the aperture effects in an optical camera. In optics, the term aperture
denotes the diameter of the effective lens opening, and the amount of light energy
that reaches the film is proportional to the squared aperture. This has motivated
the naming of the parameter α as aperture.
3.8.2
Aperture Adaptation and Final Definition of Conceptor Matrices
It is easy to verify that if Cα = C(R, α) and Cβ = C(R, β) are two versions of a
conceptor C differing in their apertures 0 < α, β < ∞, they are related to each
other by
!−1
2
α
Cβ = Cα Cα +
(I − Cα )
,
(15)
β
where we note that Cα + (α/β)2 (I − Cα ) is always invertible. Cβ is thus a function
of Cα and the ratio γ = β/α. This motivates to introduce an aperture adaptation
operation ϕ on conceptors C, as follows:
ϕ(C, γ) = C (C + γ −2 (I − C))−1 ,
(16)
where ϕ(C, γ) is the conceptor version obtained from C by adjusting the aperture
of C by a factor of γ. Specifically, it holds that C(R, α) = ϕ(C(R, 1), α).
We introduce the notation RC = C(I − C)−1 , which leads to the following
easily verified data-based version of (16):
Rϕ(C,γ) = γ 2 RC .
(17)
When we treat Boolean operations further below, it will turn out that the NOT
operation will flip zero singular values of C to unit singular values. Because of this
circumstance, we admit unit singular values in conceptors and formally define
Definition 2 A conceptor matrix is a positive semidefinite matrix whose singular
values range in [0, 1]. We denote the set of all N × N conceptor matrices by CN .
Note that Definition 1 defined the concept of a conceptor matrix associated
with a state correlation matrix R and an aperture α, while Definition 2 specifies
the more general class of conceptor matrices. Mathematically, conceptor matrices
(as in Definition 2) are more general than the conceptor matrices associated with
a state correlation matrix R, in that the former may contain unit singular values.
44
Furthermore, in the context of Boolean operations it will also become natural to
admit aperture adaptations of sizes γ = 0 and γ = ∞. The inversion in Equation
(16) is not in general well-defined for such γ and/or conceptors with unit singular
values, but we can generalize those relationships to the more general versions of
conceptors and aperture adaptations by a limit construction:
Definition 3 Let C be a conceptor and γ ∈ [0, ∞]. Then
for 0 < γ < ∞
C (C + γ −2 (I − C))−1
limδ↓0 C (C + δ −2 (I − C))−1
for γ = 0
ϕ(C, γ) =
−2
−1
limδ↑∞ C (C + δ (I − C))
for γ = ∞
(18)
It is a mechanical exercise to show that the limits in (18) exist, and to calculate the singular values for ϕ(C, γ). The results are collected in the following
proposition.
Proposition 3 Let C = U SU 0 be a conceptor and (s1 , . . . , sN )0 = diagS the
vector of its singular values. Let γ ∈ [0, ∞]. Then ϕ(C, γ) = U Sγ U 0 is the
conceptor with singular values (sγ,1 , . . . , sγ,N )0 , where
si /(si + γ −2 (1 − si ))
for 0 < si < 1, 0 < γ < ∞
0
for 0 < si < 1, γ = 0
1
for 0 < si < 1, γ = ∞
sγ,i =
(19)
0
for si = 0, 0 ≤ γ ≤ ∞
1
for si = 1, 0 ≤ γ ≤ ∞
Since aperture adaptation of C = U SU 0 only changes the singular values of C,
the following fact is obvious:
Proposition 4 If V is orthonormal, then ϕ(V CV 0 , γ) = V ϕ(C, γ) V 0 .
Iterated application of aperture adaptation corresponds to multiplying the
adaptation factors:
Proposition 5 Let C be a conceptor and γ, β ∈ [0, ∞]. Then ϕ(ϕ(C, γ), β) =
ϕ(C, γβ).
The proof is a straightforward algebraic verification using (19).
Borrowing again terminology from photography, I call a conceptor with SVD
C = U SU 0 hard if all singular values in S are 0 or 1 (in photography, a film
with an extremely “hard” gradation yields pure black-white images with no gray
tones.) Note that C is hard if and only if it is a projector matrix. If C is hard,
the following holds:
C = C † = C 0 = CC,
ϕ(C, γ) = C for γ ∈ [0, ∞].
45
(20)
(21)
The first claim amounts to stating that C is a projection operator, which is obviously the case; the second claim follows directly from (19).
Besides the aperture, another illuminating characteristic of a conceptor matrix is the mean value of its singular values, i.e. its normalized trace q(C) =
trace(C)/N . It ranges in [0, 1]. Intuitively, this quantity measures the fraction of
dimensions from the N -dimensional reservoir state space that is claimed by C. I
will call it the quota of C.
3.8.3
Aperture Adaptation: Example
In applications one will often need to adapt the aperture to optimize the quality of
C. What “quality” means depends on the task at hand. I present an illustrative
example, where the reservoir is loaded with very fragile patterns. Retrieving them
requires a prudent choice of α. Specifically, I loaded a reservoir of size N = 500
with four chaotic patterns, derived from the well-known Rössler, Lorenz, MackeyGlass, and Hénon attractors (details of this example are given in Section 4.2). Four
conceptors CR , CL , CM G , CH were computed, one for each attractor, using α = 1.
Then, in four retrieval experiments, the aperture of each of these was adapted using
(16) in a geometric succession of five different γ, yielding five versions of each of
the CR , CL , CM G , CH . Each of these was used in turn for a constrained run of the
reservoir according to the state update rule x(n+1) = C tanh(W x(n)+W in p(n+
1) + b), and the resulting output observation was plotted in a delay-embedding
format.
Figure 18 displays the findings. Per each attractor, the five apertures were
hand-selected such that the middle one (the third) best re-generated the original
chaotic signal, while the first failed to recover the original. One should mention
that it is not trivial in the first place to train an RNN to stably generate any single
chaotic attractor timeseries, but here we require the loaded network to be able to
generate any one out of four such signals, only by constraining the reservoir by a
conceptor with a suitably adapted aperture. Number insets in the panels of figure
18 indicate the apertures and quotas used per run.
3.8.4
Guides for Aperture Adjustment
The four chaotic attractors considered in the previous subsection were “best” (according to visual inspection) reconstructed with apertures between 630 and 1000.
A well-chosen aperture is clearly important for working with conceptors. In all
demonstrations reported so far I chose a “good” aperture based on experimentation and human judgement. In practice one will often need automated criteria for
optimizing the aperture which do not rely on human inspection. In this subsection
I propose two measures which can serve as such a guiding criterion.
A criterion based on reservoir-conceptor interaction. Introducing an
interim state variable z(n) by splitting the conceptor-constrained reservoir update
46
11
0.053
66
0.12
2.4e+03
0.38
1.4e+04
0.58
4e+02
0.23
A
20
0.033
1.4e+02
0.07
7e+03
0.21
4.9e+04
0.33
13
0.031
90
0.066
4.4e+03
0.24
3.1e+04
0.42
1e+03
0.13
B
26
0.083
1.8e+02
0.18
8.8e+03
0.56
6.2e+04
0.82
1.3e+03
0.34
C
6.3e+02
0.13
D
Figure 18: Invoking conceptors to retrieve four chaotic signals from a reservoir.
A Lorenz, B Rössler, C Mackey-Glass, and D Hénon attractor. All four are
represented by delay-embedding plots of the reservoir observation signal y(n).
The plot range is [0, 1] × [0, 1] in every panel. A – C are attractors derived from
differential equations, hence subsequent points are joined with lines; D derives
from an iterated map where joining lines has no meaning. Each 6-panel block
shows five patterns generated by the reservoir under the control of differently
aperture-adapted versions of a conceptor (blue, first five panels) and a plot of the
original chaotic reference signal (green, last panel). Empty panels indicate that
the y(n) signal was outside the [0, 1] range. The right upper panel in each block
shows a version which, judged by visual inspection, comes satisfactorily close to
the original. First number given in a panel: aperture α; second number: quota
q(C).
equation (9) into
z(n + 1) = tanh(W x(n) + b),
x(n + 1) = C(R, α) z(n + 1),
(22)
I define the attenuation measurable a as
aC,α = E[kz(n) − x(n)k2 ]/E[kz(n)k2 ],
(23)
where the states x(n), z(n) are understood to result from a reservoir constrained
by C(R, α). The attenuation is the fraction of the reservoir signal energy which
is suppressed by applying the conceptor. Another useful way to conceive of this
47
quantity is to view it as noise-to-signal ratio, where the “noise” is the component z(n) − x(n) which is filtered away from the unconstrained reservoir signal
z(n). It turns out in simulation experiments that when the aperture is varied, the
attenuation aC,α passes through a minimum, and at this minimum, the pattern
reconstruction performance peaks.
In Figure 19A the log10 of aC,α is plotted for a sweep through a range of
apertures α, for each of the four chaotic attractor conceptors (details in Section
4.2). As α grows, the attenuation aC,α first declines roughly linearly in the log-log
plots, that is, by a power law of the form aC,α ∼ α−K . Then it enters or passes
through a trough. The aperture values that yielded visually optimal reproductions
of the chaotic patterns coincide with the point where the bottom of the trough is
reached.
Lorenz
Roessler
0
ï2
ï2
ï4
ï4
ï6
ï6
2
3
4
5
ï8
1
log10 attenuation
A
0
ï2
ï4
ï4
ï6
ï6
ï8
1
ï8
1
2
3
4
log10 aperture
5
4
6
1
!2
0
0
2
4
6
0
!1 !5
!2
!2
2
3
4
!10
5
Hénon
0
2
0
MackeyïGlass
ï2
0
!5
2
3
log10 attenuation
ï8
1
!2
0
4
5
B
!2
0
!3 !10
0
2
4
6
1
!2
0
0
!5
!10
!4
0
2
4
6
1
0
!5
log10 aperture
!1
!1
!2 !10
!2
log10 NRMSE
0
Figure 19: Using attenuation to locate optimal apertures. A Dependancy of
attenuation on aperture for the four chaotic attractors. The blue dots mark the
apertures used to generate the plots in Figure 18. B Dependancy of attenuation on
aperture for the two sinewaves (top panels) and the two 5-point periodic patterns
(bottom) used in Sections 3.2ff. These plots also provide the NRMSEs for the
accuracy of the reconstructed patterns (gray). For explanation see text.
Figure 19B gives similar plots for the two irrational-period sines and the two 5point periodic patterns treated in earlier sections. The same reservoir and storing
procedures as described at that place were utilized again here. The dependence
of attenuation on aperture is qualitatively the same as in the chaotic attractor
example. The attenuation plots are overlaid with the NRMSEs of the original
drivers vs. the conceptor-constrained reservoir readout signals. Again, the “best”
aperture – here quantified by the NRMSE – coincides remarkably well with the
trough minimum of the attenuation.
Some peculiarities visible in the plots B deserve a short comment. (i) The initial constant plateaus in all four plots result from C(R, α) ≈ 0 for the very small
apertures in this region, which leads to x(n) ≈ 0, z(n) ≈ tanh(b). (ii) The jittery
climb of the attenuation towards the end of the plotting range in the two bottom
panels is an artefact due to roundoff errors in SVD computations which blows up
singular values in conceptors which in theory should be zero. Without rounding
48
error involved, the attenuation plots would remain at their bottom value once it
is reached. (iii) In the top two panels, some time after having passed through the
trough the attenuation value starts to decrease again. This is due to the fact that
for the irrational-period sinewave signals, all singular values of the conceptors are
nonzero. As a consequence, for increasingly large apertures the conceptors will
converge to the identity matrix, which would have zero attenuation.
A criterion based on conceptor matrix properties. A very simple criterion for aperture-related “goodness” of a conceptor can be obtained from monitoring the gradient of the squared Frobenius norm
∇(γ) =
d
kϕ(C, γ)k2
d log(γ)
(24)
with respect to the logarithm of γ. To get an intuition about the semantics of this
criterion, assume that C has been obtained from data with a correlation matrix
R with SVD R = U ΣU 0 . Then ϕ(C, γ) = R(R + γ −2 I)−1 and kϕ(C, γ)k2 =
kΣ(Σ + γ −2 I)−1 k2 = kγ 2 Σ(γ 2 Σ + I)−1 k2 . That is, ϕ(C, γ) can be seen as obtained
from data scaled by a factor of γ compared to ϕ(C, 1) = C. The criterion ∇(γ)
therefore measures the sensitivity of (the squared norm of) C on (expontential)
scalings of data. Using again photography as a metaphor: if the aperture of a
lens is set to the value where ∇(γ) is maximal, the sensitivity of the image (=
conceptor) to changes in brightness (= data scaling) is maximal.
Figure 20 shows the behavior of this criterion again for the standard example of
loading two irrational sines and two integer-periodic random patterns. Its maxima
coincide largely with the minima of the attenuation criterion, and both with what
was “best” performance of the respective pattern generator. The exception is the
two integer-periodic patterns (Figure 20 B bottom panels) where the ∇ criterion
would suggest a slightly too small aperture.
Comments on criteria for guiding aperture selection:
• The two presented criteria based on attenuation and norm gradient are
purely heuristic. A theoretical analysis would require a rigorous definition
of “goodness”. Since tasks vary in their objectives, such an analysis would
have to be carried out on a case-by-case basis for varying “goodness” characterizations. Other formal criteria besides the two presented here can easily
be construed (I experimented with dozens of alternatives (not documented),
some of which performed as well as the two instances reported here). Altogether this appears to be a wide field for experimentation.
• The attenuation-based criterion needs trial runs with the reservoir to be
calculated, while the norm-gradient criterion can be computed offline. The
former seems to be particularly suited for pattern-generation tasks where
the conceptor-reservoir feedback loop is critical (for instance, with respect
to stability). The latter may be more appropriate in machine learning tasks
49
0
2
4
6
!2
6
1
0
4
0
2
4
6
0
4
!1
2
norm2 gradient
0
!2
1.5
2
!2
!3
0
!4
1
!2
1.5
!2
0
2
4
1
6
4
6
1
0
0.5
!1
0
!0.5
2
1
0
0.5
0
log10 aperture
0
!2!0.5
!1
log10 NRMSE
!2
6
!2
Figure 20: The norm-gradient based criterion to determine “good” apertures for
the basic demo example from Sections 3.2 and 3.4. Plots show ∇(γ) against the
log10 of aperture γ. Figure layout similar as in Figure 19 B. For explanation see
text.
where conceptors are used for classifying reservoir dynamics in a “passive”
way without coupling the conceptors into the network updates. I will give
an example in Section 3.13.
3.9
3.9.1
Boolean Operations on Conceptors
Motivation
Conceptor matrices can be submitted to operations that can be meaningfully called
AND, OR, and NOT. There are two justifications for using these classical logical
terms:
Syntactical / algebraic: Many algebraic laws governing Boolean algebras are
preserved; for hard conceptor matrices the preservation is exact.
Semantical: These operations on conceptor matrices correspond dually to operations on the data that give rise to the conceptors via (7). Specifically, the
OR operation can be semantically interpreted on the data level by merging
two datasets, and the NOT operation by inverting the principal component
weights of a dataset. The AND operation can be interpreted on the data
level by combining de Morgan’s rule (which states that x ∧ y = ¬(¬x ∨ ¬y))
with the semantic interpretations of OR and NOT.
The mathematical structures over conceptors that arise from the Boolean operations are richer than standard Boolean logic, in that aperture adaptation operations can be included in the picture. One obtains a formal framework which one
might call “adaptive Boolean logic”.
50
There are two major ways how such a theory of conceptor logic may be useful:
A logic for information processing in RNNs (cognitive and neuroscience):
The dynamics of any N -dimensional RNN (of any kind, autonomously active or driven by external input), when monitored for some time period L,
yields an N × L sized state collection matrix X and its corresponding N × N
correlation matrix R, from which a conceptor matrix C = R(R + I)−1 can
be obtained which is a “fingerprint” of the activity of the network for this
period. The Boolean theory of conceptors can be employed to analyse the
relationships between such “activity fingerprints” obtained at different intervals, different durations, or from different driving input. An interesting
long-term research goal for cognitive neuroscience would be to map the logical structuring described on the network data level, to Boolean operations
carried out by task-performing subjects.
An algorithmical tool for RNN control (machine learning): By controlling
the ongoing activity of an RNN in a task through conceptors which are derived from logical operations, one can implement “logic control” strategies for
RNNs. Examples will be given in Section 3.11, where Boolean operations
on conceptors will be key for an efficient memory management in RNNs;
in Section 3.13, where Boolean operations will enable to combine positive
and negative evidences for finite-duration pattern recognition; and in Section 3.16, where Boolean operations will help to simultaneously de-noise and
classify signals.
3.9.2
Preliminary Definition of Boolean Operations
Defining Boolean operators through their data semantics is transparent and simple
when the concerned data correlation matrices are nonsingular. In this case, the resulting conceptor matrices are nonsingular too and have singular values ranging in
the open interval (0, 1). I treat this situation in this subsection. However, conceptor matrices with a singular value range of [0, 1] frequently arise in practice. This
leads to technical complications which will be treated in the next subsection. The
definitions given in the present subsection are preliminary and serve expository
purposes.
In the remainder of this subsection, conceptor matrices C, B are assumed to
derive from nonsingular correlation matrices.
I begin with OR. Recall that a conceptor matrix C (with aperture 1) derives
from a data source (network states) x through R = E[xx0 ], C = C(R, 1) = R(R +
I)−1 . Now consider a second conceptor B of the same dimension N as C, derived
from another data source y by Q = E[yy 0 ], B = B(Q, 1) = Q(Q + I)−1 . I define
C ∨ B := (R + Q)(R + Q + I)−1 ,
51
(25)
and name this the OR operation. Observe that R + Q = E[[x, y][x, y]0 ], where
[x, y] is the N × 2 matrix made of vectors x, y. C ∨ B is thus obtained by a merge
of the two data sources which previously went into C and B, respectively. This
provides a semantic interpretation of the OR operation.
Using (7), it is straightforward to verify that C ∨ B can be directly computed
from C and B by
−1 −1
C ∨ B = I + C(I − C)−1 + B(I − B)−1
,
(26)
where the assumption of nonsingular R, Q warrants that all inverses in this equation are well-defined.
I now turn to the NOT operation. For C = C(R, 1) = R(R + I)−1 with
nonsingular R I define it by
¬C := R−1 (R−1 + I)−1 .
(27)
Again this can be semantically interpreted on the data level. Consider the
SVDs R = U ΣU 0 and R−1 = U Σ−1 U 0 . R and R−1 have the same principal
components U , but the variances Σ, Σ−1 of data that would give rise to R and
R−1 are inverse to each other. In informal terms, ¬C can be seen as arising from
data which co-vary inversely compared to data giving rise to C.
Like in the case of OR, the negation of C can be computed directly from C.
It is easy to see that
¬C = I − C.
(28)
Finally, I consider AND. Again, we introduce it on the data level. Let again
C = R(R + I)−1 , B = Q(Q + I)−1 . The OR operation was introduced as addition
on data correlation matrices, and the NOT operation as inversion. Guided by
de Morgan’s law a ∧ b = ¬(¬a ∨ ¬b) from Boolean logic, we obtain a correlation
matrix (R−1 + Q−1 )−1 for C ∧ B. Via (7), from this correlation matrix we are led
to
C ∧ B := (R−1 + Q−1 )−1 (R−1 + Q−1 )−1 + I
−1
.
(29)
Re-expressing R, Q in terms of C, B in this equation, elementary transformations (using (7)) again allow us to compute AND directly:
C ∧ B = (C −1 + B −1 − I)−1 .
(30)
By a routine transformation of equations, it can be verified that the de Morgan’s laws C ∨ B = ¬(¬C ∧ ¬B) and C ∧ B = ¬(¬C ∨ ¬B) hold for the direct
computation expressions (26), (28) and (30).
52
3.9.3
Final Definition of Boolean Operations
We notice that the direct computations (26) and (30) for OR and AND are only
well-defined for conceptor matrices whose singular values range in (0, 1). I now
generalize the definitions for AND and OR to cases where the concerned conceptors
may contain singular values 0 or 1. Since the direct computation (30) of AND is
simpler than the direct computation (26) of OR, I carry out the generalization for
AND and then transfer it to OR through de Morgan’s rule.
Assume that C = U SU 0 , B = V T V 0 are the SVDs of conceptors C, B, where S
and/or T may contain zero singular values. The direct computation (30) is then
not well-defined.
Specifically, assume that diag(S) contains l ≤ N nonzero singular values and
that diag(T ) contains m ≤ N nonzero singular values, i.e. diag(S) = (s1 , . . . , sl , 0, . . . , 0)0
and diag(T ) = (t1 , . . . , tm , 0, . . . , 0)0 . Let δ be a positive real number. Define Sδ
to be the diagonal matrix which has a diagonal (s1 , . . . , sl , δ, . . . , δ)0 , and similarly
Tδ to have diagonal (t1 , . . . , tm , δ, . . . , δ)0 . Put Cδ = U Sδ U 0 , Bδ = V Tδ V 0 . Then
Cδ ∧ Bδ = (Cδ−1 + Bδ−1 − I)−1 is well-defined. We now define
C ∧ B = lim(Cδ−1 + Bδ−1 − I)−1 .
(31)
δ→0
The limit in this equation is well-defined and can be resolved into an efficient
algebraic computation:
Proposition 6 Let BR(C)∩R(B) be a matrix whose columns form an arbitrary
orthonormal basis of R(C) ∩ R(B). Then, the matrix B0R(C)∩R(B) (C † + B † −
I)BR(C)∩R(B) is invertible, and the limit (31) exists and is equal to
C ∧ B = lim(Cδ−1 + Bδ−1 − I)−1 =
δ→0
= BR(C)∩R(B) B0R(C)∩R(B) (C † + B † − I) BR(C)∩R(B)
−1
B0R(C)∩R(B) . (32)
Equivalently, let PR(C)∩R(B) = BR(C)∩R(B) B0R(C)∩R(B) be the projector matrix on
R(C) ∩ R(B). Then C ∧ B can also be written as
C ∧ B = PR(C)∩R(B) (C † + B † − I) PR(C)∩R(B)
†
.
(33)
The proof and an algorithm to compute a basis matrix BR(C)∩R(B) are given
in Section 5.2.
The formulas (32) and (33) not only extend the formula (30) to cases where C
or B are non-invertible, but also ensures numerical stability in cases where C or
B are ill-conditioned. In that situation, the pseudoinverses appearing in (32), (33)
should be computed with appropriate settings of the numerical tolerance which
one can specify in common implementations (for instance in Matlab) of the SVD.
One should generally favor (32) over (30) unless one can be sure that C and B are
well-conditioned.
53
The direct computation (28) of NOT is well-defined for C with a singular value
range [0, 1], thus nothing remains to be done here.
Having available the general and numerically robust computations of AND via
(32) or (33) and of NOT via (28), we invoke de Morgan’s rule C ∨B = ¬(¬C ∧¬B)
to obtain a general and robust computation for OR on the basis of (32) resp. (33)
and (28). Summarizing, we obtain the final definitions for Boolean operations on
conceptors:
Definition 4
¬ C := I − C,
C ∧ B := PR(C)∩R(B) (C † + B † − I) PR(C)∩R(B)
C ∨ B := ¬ (¬ C ∧ ¬ B),
†
,
where PR(C)∩R(B) is the projector matrix on R(C) ∩ R(B).
This definition is consistent with the preliminary definitions given in the previous subsection. For AND this is clear: if C and B are nonsingular, PR(C)∩R(B)
is the identity and the pseudoinverse is the inverse, hence (30) is recovered (fact
1). We noted in the previous subsection that de Morgan’s laws hold for conceptors derived from nonsingular correlation matrices (fact 2). Furthermore, if C is
derived from a nonsingular correlation matrix, then ¬C also corresponds to a nonsingular correlation matrix (fact 3). Combining facts 1 – 3 yields that the way of
defining OR via de Morgan’s rule from AND and NOT in Definition 4 generalises
(25)/(26).
For later use I state a technical result which gives a characterization of OR in
terms of a limit over correlation matrices:
Proposition 7 For a conceptor matrix C with SVD C = U SU 0 let S (δ) be a
version of S where all unit singular values (if any) have been replaced by 1 − δ,
(δ)
and let C (δ) = U S (δ) U 0 . Let RC = C (δ) (I − C (δ) )−1 . Similarly, for another
(δ)
conceptor B let RB = B (δ) (I − B (δ) )−1 . Then
(δ)
(δ)
(δ)
(δ)
(δ)
(δ)
C ∨ B = I − lim (RC + RB + I)−1 = lim (RC + RB ) (RC + RB + I)−1 . (34)
δ↓0
δ↓0
The proof is given in Section 5.3. Finally I note that de Morgan’s rule also
holds for AND (proof in Section 5.4):
Proposition 8
C ∧ B = ¬ (¬ C ∨ ¬ B).
54
3.9.4
Facts Concerning Subspaces
For an N × N matrix M , let I(M ) = {x ∈ RN | M x = x} be the identity space
of M . This is the eigenspace of M to the eigenvalue 1, a linear subspace of RN .
The identity spaces, null spaces, and ranges of conceptors are related to Boolean
operations in various ways. The facts collected here are technical, but will be
useful in deriving further results later.
Proposition 9 Let C, B be any conceptor matrices, and H, G hard conceptor matrices of the same dimension. Then the following facts hold:
1. I(C) ⊆ R(C).
2. I(C † ) = I(C) and R(C † ) = R(C) and N (C † ) = N (C).
3. R(¬C) = I(C)⊥ and I(¬C) = N (C) and N (¬C) = I(C).
4. R(C ∧ B) = R(C) ∩ R(B) and R(C ∨ B) = R(C) + R(B).
5. I(C ∧ B) = I(C) ∩ I(B) and I(C ∨ B) = I(C) + I(B).
6. N (C ∧ B) = N (C) + N (B) and N (C ∨ B) = N (C) ∩ N (B).
7. I(ϕ(C, γ)) = I(C) for γ ∈ [0, ∞) and R(ϕ(C, γ)) = R(C) for γ ∈ (0, ∞]
and N (ϕ(C, γ)) = N (C) for γ ∈ (0, ∞].
8. A = A ∧ C ⇐⇒ R(A) ⊆ I(C) and A = A ∨ C ⇐⇒ I(A)⊥ ⊆ N (C).
9. ϕ(C, 0) and ϕ(C, ∞) are hard.
10. ϕ(C, 0) = PI(C) and ϕ(C, ∞) = PR(C) .
11. H = H † = PI(H) .
12. I(H) = R(H) = N (H)⊥ .
13. ¬H = PN (H) = PI(H)⊥ .
14. H ∧ G = PI(H) ∩ I(G) .
15. H ∨ G = PI(H) + I(G) .
The proof is given in Section 5.5.
55
3.9.5
Boolean Operators and Aperture Adaptation
The Boolean operations are related to aperture adaptation in a number of ways:
Proposition 10 Let C, B be N × N sized conceptor matrices and γ, β ∈ [0, ∞].
We declare ∞−1 = ∞−2 = 0 and 0−1 = 0−2 = ∞. Then,
1. ¬ϕ(C, γ) = ϕ(¬C, γ −1 ),
2. ϕ(C, γ) ∨ ϕ(B, γ) = ϕ(C ∨ B, γ),
3. ϕ(C, γ) ∧ ϕ(B, γ) = ϕ(C ∧ B, γ),
p
4. ϕ(C, γ) ∨ ϕ(C, β) = ϕ(C, γ 2 + β 2 ),
5. ϕ(C, γ) ∧ ϕ(C, β) = ϕ(C, (γ −2 + β −2 )−1/2 ).
The proof can be found in Section 5.6. Furthermore, with the aid of aperture
adaptation and OR it is possible to implement an incremental model extension, as
follows. Assume that conceptor C has been obtained from a dataset X comprised
of m data vectors x, via R = XX 0 /m, C = R(R + α−2 I)−1 . Then, n new data
vectors y become available, collected as columns in a data matrix Y . One wishes
to update the original conceptor C such that it also incorporates the information
from Y , that is, one wishes to obtain
C̃ = R̃(R̃ + α−2 I)−1 ,
(35)
where R̃ is the updated correlation matrix obtained by Z = [XY ], R̃ = ZZ 0 /(m +
n). But now furthermore assume that the original training data X are no longer
available. This situation will not be uncommon in applications. The way to a
direct computation of (35) is barred. In this situation, the extended model C̃ can
be computed from C, Y, m, n as follows. Let CY = Y Y 0 (Y Y 0 + I)−1 . Then,
C̃ = ϕ ϕ(C, m1/2 α−1 ) ∨ CY , (m + n)1/2 α
(36)
−1
m
n
(I − C)−1 C +
α2 Y Y 0 + I
.
(37)
= I−
m+n
m+n
These formulas can be verified by elementary transformations using (7), (8),
(25) and (17), noting that C cannot have unit singular values because it is obtained
from a bounded correlation matrix R, thus (I − C) is invertible.
3.9.6
Logic Laws
Many laws from Boolean logic carry over to the operations AND, OR, NOT defined
for conceptors, sometimes with modifications.
56
Proposition 11 Let I be the N × N identity matrix, 0 the zero matrix, and
B, C, D any conceptor matrices of size N × N (including I or 0). Then the following laws hold:
1. De Morgan’s rules: C ∨ B = ¬ (¬ C ∧ ¬ B) and C ∧ B = ¬ (¬ C ∨ ¬ B).
2. Associativity: (B ∧ C) ∧ D = B ∧ (C ∧ D) and (B ∨ C) ∨ D = B ∨ (C ∨ D).
3. Commutativity: B ∧ C = C ∧ B and B ∨ C = C ∨ B.
4. Double negation: ¬(¬C) = C.
5. Neutrality of 0 and I: C ∨ 0 = C and C ∧ I = C.
6. Globality of 0 and I: C ∨ I = I and C ∧ 0 = 0.
7. Weighted
p self-absorption for OR: C ∨ C = ϕ(C,
ϕ(C, 1/2) = C.
p
√
2) and ϕ(C, 1/2) ∨
√
√
2)
and
ϕ(C,
2) ∧
8. Weighted
self-absorption
for
AND:
C
∧
C
=
ϕ(C,
1/
√
ϕ(C, 2) = C.
The proofs are given in Section 5.7. From among the classical laws of Boolean
logic, the general absorption rules A = A ∧ (A ∨ B) = A ∨ (A ∧ B) and the laws
of distributivity do not hold for conceptors.
While the absorption rules A = A ∧ (A ∨ B) = A ∨ (A ∧ B) are not valid for
conceptor matrices, it is possible to “invert” ∨ by ∧ and vice versa in a way that
is reminiscent of absorption rules:
Proposition 12 Let A, B be conceptor matrices of size N × N . Then,
†
1. C = PR(A) I + A† − (A ∨ B)† PR(A) is a conceptor matrix and
A = (A ∨ B) ∧ C.
(38)
†
2. C = I − PI(A)⊥ I + (I − A)† − (I − (A ∧ B))† PI(A)⊥ is a conceptor
matrix and
A = (A ∧ B) ∨ C.
(39)
The proof is given in Section 5.8.
57
3.10
An Abstraction Relationship between Conceptors
The existence of (almost) Boolean operations between conceptors suggests that
conceptors may be useful as models of concepts (extensive discussion in Section
3.17). In this subsection I add substance to this interpretation by introducing an
abstraction relationship between conceptors, which allows one to organize a set of
conceptors in an abstraction hierarchy.
In order to equip the set CN of N × N conceptors with an “abstraction” relationship, we need to identify a partial ordering on CN which meets our intuitive
expectations concerning the structure of “abstraction”. A natural candidate is the
partial order ≤ defined on the set of N × N real matrices by X ≤ Y if Y − X is
positive semidefinite. This ordering is often called the Löwner ordering. I will interpret and employ the Löwner ordering as an abstraction relation. The key facts
which connect this ordering to Boolean operations, and which justify to interpret
≤ as a form of logical abstraction, are collected in the following
Proposition 13 Let CN be the set of conceptor matrices of size N . Then the
following facts hold.
1. An N × N matrix A is a conceptor matrix if and only if 0 ≤ A ≤ IN ×N .
2. 0N ×N is the global minimal element and IN ×N the global maximal element of
(CN , ≤).
3. A ≤ B if and only if ¬A ≥ ¬B.
4. Let A, B ∈ CN and B ≤ A. Then
C = PR(B) (B † − PR(B) A† PR(B) + I)−1 PR(B)
is a conceptor matrix and
B = A ∧ C.
5. Let again A, B ∈ CN and A ≤ B. Then
C = I − PI(B)⊥ (I − B)† − PI(B)⊥ (I − A)† PI(B)⊥ + I
−1
PI(B)⊥
is a conceptor matrix and
B = A ∨ C.
6. If for A, B, C ∈ CN it holds that A ∧ C = B, then B ≤ A.
7. If for A, B, C ∈ CN it holds that A ∨ C = B, then A ≤ B.
8. For A ∈ CN and γ ≥ 1 it holds that A ≤ ϕ(A, γ); for γ ≤ 1 it holds that
ϕ(A, γ) ≤ A.
9. If A ≤ B, then ϕ(A, γ) ≤ ϕ(B, γ) for γ ∈ [0, ∞].
58
The proof is given in Section 5.9. The essence of this proposition can be reexpressed succinctly as follows:
Proposition 14 For conceptors A, B the following conditions are equivalent:
1. A ≤ B.
2. There exists a conceptor C such that A ∨ C = B.
3. There exists a conceptor C such that A = B ∧ C.
Thus, there is an equivalence between “going upwards” in the ≤ ordering on
the one hand, and merging conceptors by OR on the other hand. In standard
logic-based knowledge representation formalisms, a concept (or class) B is defined
to be more abstract than some other concept/class A exactly if there is some
concept/class C such that A ∨ C = B. This motivates me to interpret ≤ as an
abstraction ordering on CN .
3.11
Example: Memory Management in RNNs
In this subsection I demonstrate the usefulness of Boolean operations by introducing a memory management scheme for RNNs. I will show how it is possible
1. to store patterns in an RNN incrementally: if patterns p1 , . . . , pm have already been stored, a new pattern pm+1 can be stored in addition without
interfering with the previously stored patterns, and without having to know
them;
2. to maintain a measure of the remaining memory capacity of the RNN which
indicates how many more patterns can still be stored;
3. to exploit redundancies: if the new pattern is similar in a certain sense to
already stored ones, loading it consumes less memory capacity than when
the new pattern is dissimilar to the already stored ones.
Biological brains can learn new patterns during their lifetime. For artificial
neural networks (ANNs) “lifelong learning” presents a notorious difficulty. Training some task into an ANN typically means to adapt connection weights, often
by gradient descent optimization. When an ANN has been trained on some task
in the past and subsequently is trained on a new task, the new weight adaptations are prone to destroy the previously learnt competences. This catastrophic
forgetting (or catastrophic interference) phenomenon has been recognized since
long. Although a number of proposals have been made which partially alleviate
the problem in special circumstances ([32, 42, 73], skeptical overview: [77]), catastrophic forgetting was still considered a main challenge for neural learning theory
in an expert’s hearing solicited by the NSF in 2007 ([21]).
59
3.11.1
General Principle and Algorithm
Recall that in the original pattern storing procedure, the initial random weight
matrix W ∗ is recomputed to obtain the weight matrix W of the loaded reservoir,
such that
xj (n + 1) = tanh(W ∗ xj (n) + W in pj (n + 1) + b) ≈ tanh(W xj (n) + b),
where pj (n) is the j-th pattern signal and xj (n) is the reservoir state signal obtained when the reservoir is driven by the j-th pattern. For a transparent memory
management, it is more convenient to keep the original W ∗ and record the weight
changes into an input simulation matrix D, such that
xj (n + 1) = tanh(W ∗ xj (n) + W in pj (n + 1) + b) ≈ tanh(W ∗ xj (n) + D xj (n) + b).
(40)
In a non-incremental batch training mode, D would be computed by regularized
linear regression to minimize the following squared error:
X
X
D = argminD̃
kW in pj (n) − D̃ xj (n − 1)k2 ,
(41)
j=1,...,K n=n0 +1,...,L
where K is the number of patterns and L is the length of the training sequence
(subtracting an initial washout period of length n0 ). Trained in this way, the sum
W ∗ + D would be essentially identical (up to differences due to using another
regularization scheme) to the weight W matrix obtained in the original storing
procedure. In fact, the performance of loading a reservoir with patterns via an
input simulation matrix D as in (40) is indistinguishable from what is obtained in
the original procedure (not reported).
At recall time, the conceptor C j is inserted and the pattern pj is re-generated
by running
x(n + 1) = C j tanh (W ∗ x(n) + D x(n) + b) ,
y(n) = W out x(n).
(42)
Another variant of the loading procedure is even more minimalistic and aims
at replacing only the very input pj (n) by minimizing
X
X
kpj (n) − R xj (n)k2 .
(43)
j=1,...,K n=n0 +1,...,L
For d-dimensional input this yields a d × N matrix R of what I call input
recreation weights which at recall time is utilized by
x(n + 1) = C j tanh W ∗ x(n) + W in R x(n) + b , y(n) = W out x(n).
(44)
On the grounds of mathematical intuition, loading patterns by input integration weights W should be superior to the variant with input simulation weights
D or input recreation weights R, because of the greater number of re-computed
60
parameters (leading to greater accuracy) and the more comprehensive impact of
regularization (leading to greater dynamical stability, among other). However, in
all the various simulation experiments that I carried out so far I found only small
to negligable deteriorations in performance when using D or R instead of W , both
with respect to accuracy and with respect to dynamical stability.
With the aid of Boolean operations and input simulation weights it becomes
possible to incrementally load a collection of d-dimensional patterns p1 , p2 , . . . into
a reservoir, such that (i) loading pm+1 does not interfere with previously loaded
p1 , . . . , pm ; (ii) similarities between patterns are exploited to save memory space;
(iii) the amount of still free memory space can be monitored.
Here when I speak of “memory space”, I am referring to the N -dimensional
vector space M = RN spanned by reservoir states x. Furthermore, when I will
be speaking of “components” of a reservoir state x, I refer to projections of x on
some linear subspace of M, where the subspace will be clear from context.
The incremental loading procedure unfolds in loading cycles m = 1, 2, . . ..
After completion of the m-th cycle, input simulation weights Dm , output weights
W out,m , and conceptors C 1 , . . . , C m are obtained, such that when (42) is run with
Dm , W out,m and one of the C j (1 ≤ j ≤ m), the pattern pj is re-generated.
Dm+1 and W out,m+1 are computed from the previously obtained Dm and W out,m
by adding increments
m+1
Dm+1 = Dm + Dinc
,
out,m+1
W out,m+1 = W out,m + Winc
,
(45)
with zero-weight initialization in the first loading cycle
1
D1 = 0 + Dinc
,
out,1
W out,1 = 0 + Winc
.
(46)
Conceptors C 1 , C 2 , . . . are computed as usual from reservoir states collected in
runs of the native network x(n + 1) = tanh(W ∗ x(n) + W in pj (n) + b). The key
for incremental loading is a conceptor-based characterization of the memory space
claimed by the m previously loaded patterns. Concretely, in loading cycle m + 1
we make use of the conceptor
Am = C 1 ∨ . . . ∨ C m ,
(47)
which due to the associativity of conceptor-OR can be incrementally computed
by Am = Am−1 ∨ C m with initialization by the zero conceptor. In intuitive terms,
Am characterizes the geometry of the reservoir state cloud induced by all patterns
loaded up to cycle m. The complement F m = ¬Am in turn characterizes the
memory space that is still “free” for use in loading cycle m + 1.
The new input simulation weights Dm+1 and output weights W out,m+1 are
computed on the basis of F m and Dm as follows:
1. Drive the native reservoir with pattern pm+1 via xm+1 (n+1) = tanh(W ∗ x(n)m+1 +
W in pm+1 (n) + b) for L steps, collect reservoir states xm+1 (n0 ), . . . , xm+1 (L −
61
1) column-wise into a state collection matrix X m+1 of size N × (L − n0 ), furthermore collect time-shifted reservoir states xm+1 (n0 + 1), . . . , xm+1 (L) into
X+m+1 , collect pattern samples pm+1 (n0 + 1), . . . , pm+1 (L) into a d × (L − n0 )
matrix P m+1 , and compute C m+1 by (7).
m+1
2. Obtain the increment Dinc
by regularized linear regression from these states
xm+1 as the minimizer of the loss
X
m+1 m m+1
kW in pm+1 (n) − Dm xm+1 (n − 1) − Dinc
F x
(n − 1)k2 . (48)
n=n0 ,...,L
Concretely, this means to compute
0
0
m+1
Dinc
= (SS 0 /(L − n0 ) + α−2 I)−1 ST D /(L − n0 ) ,
(49)
where S = F m X m+1 , T D = W in P m+1 − Dm X m+1 contains the targets for
m+1
the linear regression leading to Dinc
, and α−2 is the Tychonov regularizer.
out,m+1
3. Obtain the increment Winc
by regularized linear regression as the minimizer of the loss
X
out,m+1 m m+1
kpm+1 (n) − W out,m xm+1 (n) − Winc
F x
(n)k2 .
(50)
n=n0 ,...,L
Concretely this means to compute
0
0
out m+1
Winc
= (S+ S+0 /(L − n0 ) + a2out I)−1 S+ T out /(L − n0 ) ,
(51)
where S+ = F m X+m+1 , T out = P m+1 − W out,m X+m+1 , and a2out is a regularization coefficient.
m+1
4. Update Am+1 = Am ∨C m+1 , F m+1 = ¬Am+1 , Dm+1 = Dm +Dinc
, W out,m+1 =
out,m+1
W out,m + Winc
.
It is interesting to note that it is intrinsically impossible to unlearn patterns selectively and “decrementally”. Assume that patterns p1 , . . . , pm have been trained,
resulting in Dm . Assume that one wishes to unlearn again pm . As a result of
this unlearning one would want to obtain Dm−1 . Thus one would have to comm
pute Dinc
from Dm , Am and pm (that is, from Dm and C m ), in order to recover
m−1
m
m
D
= Dm − Dinc
. However, the way to identify Dinc
from Dm , Am and C m is
barred because of the redundancy exploitation inherent in step 2. Given only Dm ,
and not knowing the patterns p1 , . . . , pm−1 which must be preserved, there is no
way to identify which directions of reservoir space must be retained to preserve
those other patterns. The best one can do is to put Ãm−1 = Am −C m = Am ∧¬C m
and re-run step 3 using Ãm−1 instead of Am−1 and putting T = W in P j in step 2.
m
m
This leads to a version D̃inc
which coincides with the true Dinc
only if there was
62
0
no directional overlap between C m and the earlier C m , i.e. if Dm−1 X m = 0 in
the original incremental learning procedure. To the extent that pm shared state
directions with the other patterns, i.e. to the extent that there was redundancy,
unlearning pm will degrade or destroy patterns that share state directions with
pm .
The incremental pattern learning method offers the commodity to measure
how much “memory space” has already been used after the first j patterns have
been stored. This quantity is the quota q(Aj ). When it approaches 1, the reservoir
is “full” and an attempt to store another pattern will fail because the F m matrix
will be close to zero.
Two demonstrations, detailed below, illustrate various aspects of the incremental storing procedure. In the first demonstration, the patterns were periodic with
integer period lengths. In the second demonstration, the patterns came from a
2-parametric family of created from weighted and phase-shifted sums of sinewaves
with irrational period lengths. These two kinds of patterns display interestingly
different characteristics in incremental storage.
3.11.2
Demonstration 1: Incremental Loading of Integer-Periodic Patterns
In the first demo I loaded K = 16 integer-periodic patterns into an N = 100 sized
reservoir, using the incremental loading procedure detailed above. The patterns
were sines of integer period length or random periodic signals. Period lengths
ranged between 3 and 15 (details documented in Section 4.3). Figure 21 displays
characteristic impressions.
Comments. When a reservoir is driven with a signal that has an integer period,
the reservoir states (after an initial washout time) will entrain to this period, i.e.
every neuron likewise will exhibit an integer-periodic activation signal. Thus, if
the period length of driver pj is Lj , the state correlation matrix as well as the
conceptor C j will be matrices of rank Lj . An aperture α = 1000 was used in
this demonstration. The large size of this aperture and the fact that the state
correlation matrix has rank Lj leads to a conceptor C j which comes close to a
projector matrix, i.e. it has Lj singular values that are close to one and N − Lj
zero singular values. Furthermore, if a new pattern pj+1 is presented, the periodic
reservoir state vectors arising from it will generically be linearly independent of
all state vectors that arose from earlier drivers. Both effects together (almost
projector C j and linear independence of nonzero principal directions of these C j )
imply that the sequence A1 , . . . , AK will essentially be a sequence of projectors,
where R(Aj+1 ) will comprise Lj+1 more dimensions than R(Aj ). This becomes
clearly apparent in Figure 21: the area under the singular value plot of Aj has an
almost rectangular shape, and the increments from one plot to the next match the
periods of the respective drivers, except for the last pattern, where the network
capacity is almost exhausted.
Patterns j = 5, 6, 7 were identical to j = 1, 2, 3. As a consequence, when the
63
1
j=1
j=5
j=9
j = 13
0
ï1
0.1
0.0014
1
0.36
0.0014
j=2
0.54
0.019
j=6
0.8
0.075
j = 10
j = 14
0
ï1
0.25
0.008
1
0.36
0.011
j=3
0.59
0.015
j=7
0.83
0.013
j = 11
j = 15
0
ï1
0.29
0.00071
1
0.36
0.00071
j=4
0.65
0.036
j=8
0.94
0.05
j = 12
j = 16
0
ï1
1
0.36
0.003
10
0.42
20
1
0.004
10
0.73
20
1
0.048
10
0.99
20
1
0.95
10
20
Figure 21: Incremental storing, first demonstration (figure repeated from Section
1 for convenience). 13 patterns with integer period lengths ranging between 3 and
15 were stored. Patterns were sinewaves with integer periods or random. Patterns
j = 5, 6, 7 are identical to j = 1, 2, 3. Each panel shows a 20-timestep sample of the
correct training pattern pj (black line) overlaid on its reproduction (green line).
The memory fraction used up until pattern j is indicated by the panel fraction
filled in red; this quota value is printed in the left bottom corner of each panel.
The red areas in each panel in fact show the singular value spectrum of Aj (100
values, x scale not shown). The NRMSE is inserted in the bottom right corners
of the panels. Note that the conceptor-controlled reproduction of all patterns was
carried out after the last pattern had been loaded.
storage procedure is run for j = 5, 6, 7, Aj remains essentially unchanged – no
further memory space is allocated.
When the network’s capacity is almost exhausted in the sense that the quota
q(Aj ) approaches 1, storing another pattern becomes inaccurate. In this demo,
this happens for that last pattern j = 16 (see Figure 21).
For a comparison I also loaded all 16 patterns simultaneously in a separate
simulation experiment, computing only a single input simulation weight matrix
D by minimizing (41). The dots in the left panel in Figure 22 indicate the recall accuracies based on this D. The mean of NRMSEs of this comparison was
0.063, while the mean NRMSE of recalled patterns in the incremental loading
condition after the final loading cycle was 0.078. These two mean values do not
differ much, but the recall NRMSEs scatter differently in the two conditions. In
the non-incremental loading condition, the recall accuracies are much closer to
each other than in the incremental condition, where we find many highly precise
accuracies and one poor outlier, namely the last pattern which spilled over the
remaining memory space. Apparently in simultaneous loading the common D
strikes a compromise for all loaded patterns on equal terms, whereas in incremental loading earlier-loaded patterns receive the benefit of competing only with the
64
1
sing. value
log10 NRMSE
0
ï2
ï4
4
0.5
0
0
8
12
16
loading cycle
50
dimension number
100
Figure 22: Incremental loading of integer-periodic patterns: detail. Left: Development of re-generation error (log10 of NRMSE) for the 16 patterns from Figure
21 during the incremental loading procedure. Each line corresponds to one pattern. The red diamond marks the NRMSE of the last-loaded pattern. The 16
dots at the right mark the recall errors obtained from non-incremental loading.
Right: The essentially rectangular singular value spectra of the conceptors Am ,
same color code as in left panel.
patterns that had been loaded before.
3.12
Incremental Loading of Integer-Periodic Patterns: Detailed Analysis
Here I provide a more detailed analysis of the incremental loading of integerperiodic patterns (demo in Section 3.11). I consider the case of incrementally
loading patterns p1 , p2 , . . ., which have integer period lengths L1 , L2 , . . ., into an
N -dimensional reservoir. I refer to M = RN as the memory space provided by
the reservoir.
The periodic patterns p1 , p2 , . . . give rise to a sequence of mutually orthogonal
linear “excitation” subspaces (E j )j=1,2,... of M, with dimensions Lj , as follows.
When the reservoir is driven by pattern p1 , after a washout period it will engage
in an L1 -periodic sequence of states, that is, x1 (n) = x1 (n + L1 − 1). Generically
(i.e. with probability 1 for random draws of network weights), these L1 states
x1 (n), . . . , x1 (n + L1 − 1) are linearly independent and span an L1 -dimensional
linear subspace E 1 of M. Define M1 = E 1 .
Now assume, by induction, that for the first m patterns p1 , . . . , pm we have
found m pairwise orthogonal linear subspaces E 1 , . . . , E m of dimensions L1 , . . . , Lm ,
with Mm = E 1 ⊕. . .⊕E m being the (L1 +. . . Lm )-dimensional subspace given by the
direct sum of these. Now let the native reservoir be driven with the next pattern
pm+1 . Provided that this pattern is not a duplicate of any of the previous ones, and
provided that L1 +. . .+Lm +Lm+1 ≤ N , the collection of the L1 +. . .+Lm+1 many
induced reservoir states x1 (n), . . . , x1 (n+L1 −1), . . . , xm+1 (n), . . . , xm+1 (n+Lm+1 −
1) will again generically be linearly independent, spanning an (L1 + . . . + Lm+1 )-
65
dimensional subspace Mm+1 of M. Define E m+1 to be the orthogonal complement
to Mm within Mm+1 . Then E m+1 has dimension Lm+1 . This concludes the induction step in the definition of the sequence (E j )j=1,2,... . Note that this construction
only works as long as L1 + . . . + Lm+1 ≤ N , and assumes that the patterns pj are
pairwise different.
The spaces Mj and E j can be characterized through the conceptors C j associated with the patterns pj . To see this, consider first the conceptor matrix C 1 .
It has exactly L1 nonzero eigenvalues corresponding to eigenvectors which span
M1 . For large aperture (I used α = 1000 in the demo), the eigenvalue spectrum of
C 1 will be approximately rectangular, with L1 eigenvalues close to 1 and N − L1
eigenvalues exactly zero (right panel in Figure 22). In linear algebra terms, C 1
approximately is the projector matrix which projects M on E 1 , and ¬C 1 as the
projector matrix that maps M on (E 1 )⊥ . Define A1 = C 1 , F 1 = ¬C 1 , as in the
algorithm given in the previous subsection.
Again by induction, assume that Am = C 1 ∨ . . . ∨ C m is approximately the
projector of M on Mm and F m = ¬Am the complement projector of M on
(Mm )⊥ . It holds that Am+1 = Am ∨ C m+1 is (approximately) the projector from
M on M m+1 (follows from Proposition 9), and F m+1 = ¬Am+1 the projector on
(M m+1 )⊥ .
Equipped with these interpretations of Am and F m we turn to the incremental
loading procedure.
m+1
According to the loss (48), Dinc
is optimized to map F m xm+1 (n − 1) to
W in pm+1 (n) − Dm xm+1 (n − 1), that is, to the pattern-input term W in pm+1 (n)
minus those reservoir state components which are already produced by the input
simulation weights Dm . As we have just seen, F m is approximately a projection of
reservoir states on the linear subspace of M which was not excited by any of the
previous driving patterns, that is, a projection on (Mm )⊥ . Basing the regularized
linear regression on arguments F m xm+1 (n − 1) which lie in (Mm )⊥ leads to a
m+1
m+1
nulls all reservoir state components that fall in
solution for Dinc
such that Dinc
m
the linear subspace M excited by previous patterns.
By an easy inductive argument it can be seen that Dm , in turn, nulls all vectors
in (Mm )⊥ .
Decomposing states excited by pm+1 as xm+1 (n) = u + v, with u ∈ Mm ,
v ∈ (Mm )⊥ , one obtains (approximately, to the extent that the C j are projectors)
Dm u
Dm v
m+1
Dinc
u
m+1
Dinc v
≈
≈
≈
≈
≈
Dm xm+1 (n),
0
0,
m+1 m m+1
m+1 m+1
F x
(n)
Dinc
x
(n) ≈ Dinc
in m+1
m m+1
W p
(n + 1) − D x
(n),
(52)
(53)
(54)
(55)
where the approximation (55) is accurate to the extent that the linear regression
66
solution to (48) is accurate. Hence,
m+1
Dm+1 xm+1 (n) = (Dm + Dinc
) (u + v)
m m+1
= D x
(n) + W in pm+1 (n + 1) − Dm xm+1 (n)
= W in pm+1 (n + 1),
(56)
as desired for input simulation weights. The reasoning for output weights computed on the basis of (50) is analog.
(56) explains why the input simulation weights Dm+1 can recover the pattern
m+1
p
at recall time. To see that furthermore Dm+1 also still recovers earlier loaded
patterns (avoiding catastrophic forgetting), observe that reservoir states associated
with earlier patterns fall in Mm and that by (52)–(54), Dm+1 restricted on Mm
operates identically as Dm , and Dm+1 restricted on Mm−1 as Dm−1 , etc.
I want to point out a possible misunderstanding of the working principle of
this incremental loading procedure. It does not function by splitting the memory
space M into orthogonal components E j , and loading the patterns pj separately
into these in some way. The states xm+1 (n) obtained by exciting the reservoir
with pm+1 are generically non-orthogonal to the excitation space Mm of previously
loaded patterns. The re-generation of pm “uses” reservoir state components from
all the subspaces E 1 , . . . , E m . In this respect, the incremental loading procedure
fundamentally differs from a common strategy adopted to counter catastrophic
forgetting by orthogonalizing internal representations of patterns in some way
(survey in [32]).
3.12.1
Demonstration 2: Incremental Loading of Irrational-Period Patterns
In the second demo I loaded 16 patterns that were randomly taken from the the
2-parametric family of patterns governed by
p(n) = a sin(2 π n/P ) + (1 − a) sin(4 π (b + n/P )).
(57)
These signals are weighted sums of two sines, the first with period length P and
the second with period length P/2. The weights of these two components are a and
(1−a), and the second component is phase-shifted relative to the first by a fraction
√
b of its period length P/2. The reference period length P was fixed to P = 30.
The parameters a, b were freshly sampled from the uniform distribution on [0, 1]
for each pj , where 16 patterns were used. Because P is irrational, these patterns
(when sampled at integer steps) are not integer-periodic but quasi-periodic. As a
consequence, the reservoir states xj (n) excited by a driver pj (n) span all of M,
and all singular values of C j are nonzero, with a reverse sigmoid-shaped spectrum
like we saw earlier for other irrational-period patterns (Figure 14 [left]). Network
parameters are reported in Section 4.4. Figures 23 and 24 display the findings.
Comments. When the driver has an irrational period length, the excited reservoir
states will span the available reservoir space RN . Each reservoir state correlation
67
1
j=1
j=5
j=9
j = 13
0
ï1
0.17
0.11
1
0.59
0.19
j=2
0.69
0.26
j=6
0.76
0.16
j = 10
j = 14
0
ï1
0.3
0.031
1
0.64
0.16
j=3
0.71
0.07
j=7
0.76
0.042
j = 11
j = 15
0
ï1
0.46
0.3
1
0.66
0.25
j=4
0.73
0.16
j=8
0.78
0.084
j = 12
j = 16
0
ï1
1
0.52
0.049
10
0.68
20
1
0.17
10
0.74
20
1
0.074
10
0.78
20
1
0.077
10
20
Figure 23: Incremental storing, second demonstration. 16 sinewave patterns with
irrational periods ranging between 4 and 20 were used. Plot layout is the same as
in Figure 21.
1
sing. value
log10 NRMSE
0
ï1
ï2
4
0.5
0
0
8
12
16
loading cycle
50
dimension number
100
Figure 24: Detail of incremental loading of irrational-periodic patterns. Figure
layout as in Fig. 22.
matrix will have only nonzero singular values, albeit of rapidly decreasing magnitude (these tails are so small in magnitude that they are not visible in the first few
plots of Aj in Figure 23). The fact that each driving pattern excites the reservoir
in all directions leads to the “reverse sigmoid” kind of shapes of the singular values
of the Aj visible in Figure 23.
As the iterated storing progresses, a redundancy exploitation effect becomes
apparent: while for the first 4 patterns altogether a quota q(A4 ) = 0.52 was
allocated, the remaining 12 patterns only needed an additional quota of q(A16 ) −
q(A4 ) = 0.26. Stated in suggestive terms, at later stages of the storing sequence
the network had already learnt how to oscillate in sinewave mixes in general, and
only needed to learn in addition how to oscillate at the particular newly presented
version taken from the parametrized pattern family. An aperture of size α = 1.5
was used in the second demonstration.
68
The mean recall NRMSE for the 16 patterns (testing after loading the last one)
was 0.136. The mean NRMSE for these patterns when loaded non-incrementally
was 0.131. Like for integer-periodic patterns, these values are not substantially
different from each other.
3.12.2
Integer-Periodic Versus Parametrized Patterns: Close-up Inspection
Notice the very different shape of the singular value spectra of the Aj in the integerperiodic versus the irrational-periodic patterns (right panels in Figures 22 and 24).
This suggests substantial differences in the underlying mechanisms, and it is also
obvious that the analysis offered in the previous subsection for integer-periodic
patterns does not transfer to the irrational-periodic patterns from a parametrized
family which we considered here.
To understand better the differences between the incremental loading of integerperiodic versus parametrized patterns, I ran a separate suite of simulations with
additional diagnostics, as follows (detail in Section 4.5). The simulation with
integer-periodic patterns was repeated 10 times, using the same global network
scalings and regularization coefficients as before, but using exclusively 6-periodic
patterns throughout. Loading sixteen such patterns into a 100-neuron reservoir
should claim a memory quota of about 6·16/100 = 0.96, just short of over-loading.
Network weights and patterns were randomly created for each of the 10 simulations. Similarly, the parametric-family simulation was repeated 10 times, using
the same pattern family (57) with freshly sampled pattern parameters and network weights in each simulation. Finally, I loaded 100 integer-periodic patterns of
period length 3 into the same reservoir that was used for the 6-periodic patterns,
again repeating the simulation 10 times. For each individual simulation, the following diagnostic quantities were computed (indicated colors refer to color code
in Figure 25):
1. The mean absolute values am of arguments am (n) = F m xm (n−1) that enter
m
(mean taken over n and vector components, orange).
the regression for Dinc
2. The mean absolute values tm of targets tm (n) = W in p(n) − Dm−1 xm (n − 1)
m
that enter the regression for Dinc
(red).
m
3. The mean absolute size of matrix entries in the Dinc
matrices, normalized by
m
the inverse of the scaling factor that Dinc has to realize. Concretely, define
m m m
m
Dinc
to be the average absolute size of matrix elements in Dinc
a /t (blue).
4. The mean absolute size of matrix entries in Dm (gray).
5. The memory quota used up to the current loading cycle (black ).
6. The condition number of the matrix inversion that has to be carried out in
m
. This is a number
the (regularized) linear regression when computing Dinc
69
integer-periodic patterns,
period length 6 ("IP6")
patterns from parametrized
family ("PF")
integer-periodic patterns,
period length 3 ("IP3")
5
5
5
4
4
4
3
3
3
2
2
2
1
1
1
0
0
0
ï1
ï1
ï1
ï2
ï2
ï2
ï3
4
8
12
loading cycle
16
ï3
4
8
12
loading cycle
16
ï3
20
40
60
80
loading cycle
100
• log10 of average absolute size
of Dinc arguments (x 0.5)
• normalized size of Dinc
(x 100)
• log10 of matrix inversion
condition
• log10 of average absolute size
of Dinc targets (x 0.5)
• size of D (x 5)
• log10 of final NRMSE
• memory quota (x 5)
• log10 of incremental NRMSE
Figure 25: Detailed diagnostics of the incremental loading procedure. Plots show
means of 10 simulations, error bars indicate standard deviations (omitted in third
panel for clarity). Some curves were scaled to fit the panel, legend indicates
scalings. The horizontal axis marks the loading cycles except for the final NRMSE
(cyan) where it marks the pattern number. For explanation see text.
≥ 1. Values close to 1 indicate that the linear regression faces a simple task
(arguments component signals are orthogonal and have same signal power;
regression amounts to linearly combine orthogonal projections) (green).
7. The recall NRMSEs for the 16 patterns p1 , . . . , p16 after the final loading
cycle. Gives 16 NRMSE values per simulation (cyan).
8. The recall NRMSEs computed for pattern pj after the j-th loading cycle.
Gives 16 NRMSE values per simulation (magenta).
For brevity I refer to the integer-periodic pattern simulations as the “IP6” and
“IP3” conditions and to the parametric family patterns as the “PF” condition, and
to the diagnostic quantities by their colors in Figure 25. Here are some observations
that help to gain intuitive insight—future mathematical analyses notwithstanding:
• In both IP6 and PF, the reconstruction accuracy for a pattern pj directly
after its loading is, by and large, about the same as the reconstruction accuracy after the last loading cycle (cyan, magenta). That is, subsequent loading
70
events do not degrade the representation of previously loaded patterns—this
is also manifest in the left panels of Figures 22 and 24.
• In IP6 the reconstruction accuracy (cyan, magenta) deteriorates roughly
linearly in terms of the log-NRMSE as loading cycles unroll. This might
be connected to the circumstance that on the one hand, each new pattern
is unrelated to the previous ones and should require the same amount of
“coding information” to be stored (apparent in the linear increase of the
quota, black ), but on the other hand, increasingly little “free memory space”
F m can be recruited for the coding and hence, increasingly less “degrees of
freedom for coding” are available.
• In PF the final reconstruction quality (cyan) is about the same for all patterns, and it is about the same as the incrementally assessed reconstruction
NRMSEs (magenta). In fact, if more patterns from this family were to be
added in further loading cycles, the reconstruction NRMSEs still would stay
on the same level (not shown). The memory system can learn the entire
pattern family.
• The condition number (green) of the matrix that has to be inverted in the
regression improves in all of IP6, FP and IP3, but to different degrees. Concretely, this matrix is meann {F m xm (n) (F m xm (n))0 } + % I, where % is the
ridge regression regularizing coefficient (set to 0.001, 0.02, 0.001 in IP6, FP
and IP3). In PF, the condition number swiftly approaches 1. This can be understood as an effect of dominance of regularization: given that the average
absolute value of arguments F m xm (n) decreases to an order of magnitude
of 0.01 (orange), the spectral radius of meann {F m xm (n) (F m xm (n))0 } will
drop to an order of 0.0001, hence the linear regression becomes dominated
by the regularization part with coefficient % = 0.02 0.0001. In intuitive
terms, in PF the iterated re-computations of Dm soon cease to “learn anything substantially new”. Because the recall accuracies do not degrade, this
means that nothing new has to be learned—after a few loading cycles, the
memory system has extracted from the examples stored so far almost all
that is necessary to represent the entire family.
• The situation is markedly different in IP6. Here the condition number never
falls below about 100, the spectral radius of meann {F m xm (n) (F m xm (n))0 }
is about the same size as % = 0.001, thus substantial novel information
becomes coded in each Dm until the end.
• The interpretation that in IP6 novel information becomes coded throughout
all loading cycles whereas in PF the rate of coding new information falls to
zero is further substantiated by the development of the (normalized) sizes
m
Dinc
of the increments that are added to Dm (blue). In IP6 these increments
retain approximately the same size through all loading cycles whereas in PF
71
they appear to decay toward zero (in simulations with much larger numbers
of loading cycles [not shown] this impression was numerically verified). In
the same vein, the mean absolute sizes of elements in the Dm matrices (gray)
grows superlinearly in IP6 and levels out in PF.
• The two different modes of operation that we find contrasted in the IP6 and
PF conditions can occur both together in a pattern loading sequence. In the
IP3 condition, the 3-periodic patterns can be viewed as being drawn from a 2parametric family (the random length-3 patterns were normalized to all have
the same minimal and maximal values, thus can be described up to period
shifts by the binary parameter “the maximal value point follows the minimal
value point directly yes/no” and the continuous parameter that gives the
non-extremal point value). When the first few (about 20) patterns from this
collection are loaded, they lead to a similar phenemenology as in the IP6
condition: each additional pattern is “novel” to the memory system. When
more loading cycles are executed, the phenomenology changes to the type
seen in the PF condition: the network (more precisely, the input simulation
matrix Dm ) has captured the “law” of the entire family and not much further
information has to be coded.
While these simulations help us to get basic intuitions about the mechanisms
that enable incremental pattern loading in different conditions, a formal analysis
would be quite complex and is left for future work. Questions that would need
to be addressed include the effects of regularization (which are strong), aperture
settings, and asymptotic behavior when the number of loaded patterns grows to
infinity.
3.12.3
Incremental Loading of Arbitrary Patterns
When one tries to store patterns (i) whose conceptor singular value spectra are
not close to rectangular and (ii) which do not come from a parametric family,
the incremental loading procedure breaks down (not shown). The reason for this
failure is clear on intuitive grounds. When the singular value spectra of conceptors C m are “sigmoid-shaped” as e.g. in Figure 24, the memory space claimed by
patterns loaded up to loading cycle m will have nonzero components in all directions of M—technically, the singular value spectrum of Am will nowhere be zero.
The argument vectors F m xm+1 (n) = ¬ Am xm+1 (n) which enter the regression for
m
Dinc
likewise will have signal components in all directions of M, and executing
the procedure for incremental loading will lead to a confounding of already stored
with newcoming patterns.
This diagnosis directly hints at a solution: make the singular value spectra of
the conceptors C m associated with (arbitrary) patterns pj rectangular or approximately rectangular. One possible procedure to achieve this runs like follows:
1. When loading pattern pm , compute the associated conceptor C m as before.
72
1
sing. value
log10 NRMSE
0
ï1
ï2
ï3
4
0.5
0
0
8
12
16
loading cycle
50
dimension number
100
Figure 26: Incremental loading of diverse patterns, using conceptors with singular
value spectra transformed to approximately rectangular shape. Figure layout as
in Figure 22.
2. Compute its SVD U S U 0 = C m . S is a diagonal matrix which contains the
singular values of C m on its diagonal. Denote the singular values by σi .
3. Change these σi in a way that yields new σ̃i which are either close to 0 or
close to 1. A drastic procedure is thresholding: σ̃i = 1 if σi > 0.5 else σ̃i = 0.
A softer procedure is to pass σi through some sufficiently steep sigmoid.
4. Use the “rectangularized” C̃ m = U S̃U 0 instead of the original C m in the
incremental loading procedure which otherwise remains unchanged (where
S̃ is the diagonal matrix with the σ̃i on its diagonal).
For a demonstration, I loaded a sequence of patterns where pm was integerperiodic with period 5 when m was uneven and where pm was one of the patterns from the 2-parametric family used above for even m (simulation detail in
Section 4.6). The original incremental loading procedure would fail, but when
the singular values of conceptors C m were passed through a steep sigmoid by
σ̃i = (tanh(50 (2 σi − 1))) /2, incremental loading functioned well, as revealed in
Figure 26. Interestingly, the incremental loading gave better recall accuracies
than simultaneous loading: the mean NRMSE for incrementally loaded patterns,
assessed after the final loading cycle, was 0.094, whereas the mean NRMSE for the
same patterns loaded simultaneously was 0.19 (spending a fair effort on optimizing
scaling and regularization parameters in the latter case). This is surprising and
further investigations are needed to understand why/when incremental loading
may give better results than simultaneous loading.
I conclude this section by remarking that when conceptors are computed by
the auto-adaptation rule introduced below in Section 3.14, approximately rectangular singular value spectra conducive for incremental loading are automatically
obtained.
73
3.13
Example: Dynamical Pattern Recognition
In this subsection I present another demonstration of the usefulness of Boolean
operations on conceptor matrices. I describe a training scheme for a pattern
recognition system which reaches (or surpasses) the classification test performance of state-of-the-art recognizers on a widely used benchmark task. Most
high-performing existing classifiers are trained in discriminative training schemes.
Discriminative classifier training exploits the contrasting differences between the
pattern classes. This implies that if the repertoire of to-be-distinguished patterns
becomes extended by a new pattern, the classifier has to be re-trained on the
entire dataset, re-visiting training data from the previous repertoire. In contrast,
the system that I present is trained in a “pattern-local” scheme which admits an
incremental extension of the recognizer if new pattern types were to be included in
its repertoire. Furthermore, the classifier can be improved in its exploitation phase
by incrementally incorporating novel information contained in a newly incoming
test pattern. The key to this local-incremental classifier training is agin Boolean
operations on conceptors.
Unlike in the rest of this report, where I restrict the presentation to stationary
and potentially infinite-duration signals, the patterns here are nonstationary and of
short duration. This subsection thus also serves as a demonstration how conceptors
function with short nonstationary patterns.
I use is the Japanese Vowels benchmark dataset. It has been donated by [60]
and is publicly available at the UCI Knowledge Discovery in Databases Archive
(http://kdd.ics.uci.edu/). This dataset has been used in dozens of articles
in machine learning as a reference demonstration and thus provides a quick first
orientation about the positioning of a new classification learning method. The
dataset consists of 640 recordings of utterances of two successive Japanese vowels
/ae/ from nine male speakers. It is grouped in a training set (30 recordings from
each of the speakers = 270 samples) and a test set (370 further recordings, with
different numbers of recordings per speaker). Each sample utterance is given in the
form of a 12-dimensional timeseries made from the 12 LPC cepstrum coefficients.
The durations of these recordings range between 7 and 29 sampling timesteps.
The task is to classify the speakers in the test data, using the training data to
learn a classifier. Figure 27 (top row) gives an impression of the original data.
I preprocessed the raw data into a standardized format by (1) shift-scaling each
of the twelve channels such that per channel, the minimum/maximum value across
all training samples was 0/1; (2) interpolating each channel trace in each sample
by a cubic polynomial; (3) subsampling these on four equidistant support points.
The same transformations were applied to the test data. Figure 27 (bottom row)
illustrates the normalized data format.
The results reported in the literature for this benchmark typically reach an
error rate (percentage of misclassifications on the test set) of about 5 – 10 test
errors (for instance, [93, 99, 81] report from 5 – 12 test misclassifications, all using
specialized versions of temporal support vector machines). The best result that I
74
speaker 4
speaker 6
speaker 8
2
2
2
1
1
1
0
0
0
!1
!1
!1
5
10 15 20 25
5
10 15 20 25
5
1
1
1
0.5
0.5
0.5
0
1
2
3
4
0
1
2
3
4
0
1
10 15 20 25
2
3
4
Figure 27: Three exemplary utterance samples from Japanese Vowels dataset.
Plots show values of twelve signal channels against discrete timesteps. Top row:
raw data as provided in benchmark repository, bottom row: standardized format
after preprocessing.
am aware of outside my own earlier attempts [57] is reported by [14] who reaches
about 4 errors, using refined hidden Markov models in a non-discriminative training scheme. It is however possible to reach zero errors, albeit with an extraordinary
effort: in own work [57] this was robustly achieved by combining the votes of 1,000
RNNs which were each independently trained in a discriminative scheme.
Here I present a “pocket-size” conceptor-based classification learning scheme
which can be outlined as follows:
1. A single, small (N = 10 units) random reservoir network is initially created.
2. This reservoir is driven, in nine independent sessions, with the 30 preprocessed training samples of each speaker j (j = 1, . . . , 9), and a conceptor Cj+
is created from the network response (no “loading” of patterns; the reservoir
remains unchanged throughout).
3. In exploitation, a preprocessed sample s from the test set is fed to the reservoir and the induced reservoir states x(n) are recorded and transformed
into a single vector z. For each conceptor then the positive evidence quantity z 0 Cj+ z is computed. This leads to a classification by deciding for
j = argmaxi z 0 Ci+ z as the speaker of s. The idea behind this procedure
is that if the reservoir is driven by a signal from speaker j, the resulting
response z signal will be located in a linear subspace of the (transformed,
75
see below) reservoir state space whose overlap with the ellipsoids given by
the Ci+ is largest for i = j.
4. In order to further improve
classification quality, for each speaker j
W the
−
+
+
+
also a conceptor Cj = ¬ {C1 , . . . , Cj−1
, Cj+1
, . . . , C9+ } is computed. This
conceptor can be understood as representing the event “not any of the other
speakers”. This leads to a negative evidence quantity z 0 Cj− z which can
likewise be used as a basis for classification.
5. By adding the positive and negative evidences, a combined evidence is obtained which can be paraphrased as “this test sample seems to be from
speaker j and seems not to be from any of the others”.
In more detail, the procedure was implemented as follows. A 10-unit reservoir
system with 12 input units and a constant bias term with the update equation
x(n + 1) = tanh(W x(n) + W in s(n) + b)
(58)
was created by randomly creating the 10 × 10 reservoir weight matrix W , the
10 × 12 input weight matrix W in and the bias vector b (full specification in Section
4.8). Furthermore, a random starting state xstart , to be used in every run in
training and testing, was created. Then, for each speaker j, the conceptor Cj+ was
learnt from the 30 preprocessed training samples skj (n) (where j = 1, . . . , 9; k =
1, . . . , 30; n = 1, . . . , 4) of this speaker, as follows:
1. For each training sample skj (k = 1, . . . , 30) of this speaker, the system
(58) was run with this input, starting from x(0) = xstart , yielding four
network states x(1), . . . , x(4). These states were concatenated with each
other and with the driver input into a 4 · (10 + 12) = 88 dimensional vector
zjk = [x(1); skj (1); . . . ; x(4); skj (4)]. This vector contains the entire network
response to the input skj and the input itself.
2. The 30 zjk were assembled as columns into a 88 × 30 matrix Z from which
a correlation matrix Rj = ZZ 0 /30 was obtained. A preliminary conceptor
C̃j+ = Rj (Rj + I)−1 was computed from Rj (preliminary because in a later
step the aperture is optimized). Note that C̃j+ has size 88 × 88.
After all “positive evidence” conceptors C̃j+ had been created, preliminary
“negative evidence” conceptors C̃j− were computed as
C̃j− = ¬
_
+
+
{C̃1+ , . . . , C̃j−1
, C̃j+1
, . . . , C̃9+ }.
(59)
An important factor for good classification performance is to find optimal apertures for the conceptors, that is, to find aperture adaptation factors γ + , γ − such
76
that final conceptors Cj+ = ϕ(C̃j+ , γ + ), Cj− = ϕ(C̃j− , γ − ) function well for classification. A common practice in machine learning would be to optimize γ by
cross-validation on the training data. This, however, is expensive, and more crucially, it would defy the purpose to design a learning procedure which can be
incrementally extended by novel pattern classes without having to re-inspect all
training data. Instead of cross-validation I used the ∇ criterion described in Section 3.8.4 to find a good aperture. Figure 28 shows how this criterion varies with
γ for an exemplary case of a ϕ(C̃j+ , γ + ) sweep. For each of the nine C̃j+ , the value
γ̃j+ which maximized ∇ was numerically computed, and the mean of these nine
values was taken as the common γ + to get the nine Cj+ = ϕ(C̃j+ , γ + ). A similar
procedure was carried out to arrive at Cj− = ϕ(C̃j− , γ − ).
10
5
0
0
1
2
3
4
5
log2 γ
6
7
8
Figure 28: The criterion ∇ from an exemplary conceptor plotted against the log
2 of candidate aperture adaptations γ.
The conceptors Cj+ , Cj− were then used for classification as follows. Assume z is
an 88-dimensional combined states-and-input vector as described above, obtained
from driving the reservoir with a preprocessed test sample. Three kinds of classification hypotheses were computed, the first only based on Cj+ , the second based
on Cj− , and one based on a combination of both. Each classification hypothesis is
a 9-dimensional vector with “evidences” for the nine speakers. Call these evidence
vectors h+ , h− , h+− for the three kinds of classifications. The first of these was
computed by setting h̃+ (j) = z 0 Cj+ z, then normalizing h̃+ to a range of [0, 1] to
obtain h+ . Similarly h− was obtained from using z 0 Cj− z, and h+− = (h+ + h− )/2
was simply the mean of the two former. Each hypothesis vector leads to a classification decision by opting for the speaker j corresponding to the largest component
in the hypothesis vector.
This classification procedure was carried out for all of the 370 test cases, giving
370 hypothesis vectors of each of the three kinds. Figure 29 gives an impression.
Results: The outlined classification experiment was repeated 50 times with random new reservoirs. On average across the 50 trials, the optimal apertures γ + / γ −
were found as 25.0 / 27.0 (standard deviations 0.48 / 0.75). The number of misclassifications for the three types of classification (positive, negative, combined
77
Figure 29: Collected evidence vectors h+ , h− , h+− obtained in a classification learning experiment. Grayscale coding: white = 0, black = 1. Each panel shows 370
evidence vectors. The (mostly) black segments along the diagonal correspond to
the correct classifications (test samples were sorted by speaker). For explanation
see text.
evidence) were 8.5 / 5.9 / 4.9 (standard deviations 1.0 / 0.91 / 0.85). The training errors for the combined classification (obtained from applying the classification
procedure on the training samples) was zero in all 50 trials. For comparison, a
carefully regularized linear classifier based on the same z vectors (detail in Section
4.8) reached 5.1 misclassifications across the 50 trials.
While these results are at the level of state-of-the-art classifiers on this benchmark, this basic procedure can be refined, yielding a significant improvement. The
idea is to compute the evidence for speaker j based on a conceptor C̄j+ which itself
is based on the assumption that the test sample s belongs to the class j, that is, the
computed evidence should reflect a quantity “if s belonged to class j, what evidence
can we collect under this assumption?”. Recall that Cj+ is obtained from the 30
training samples through Cj+ = R (R + (γ + )−2 I)−1 , where R = ZZ 0 /30 is the correlation matrix of the 30 training coding vectors belonging to speaker j. Now add
the test vector z to Z, obtaining Z̄ = [Zz], R̄ = Z̄ Z̄ 0 /31, C̄j+ = R̄(R̄ + (γ + )−2 I)−1 ,
and use C̄j+ in the procedure outlined above instead of Cj+ . Note that, in application scenarios where the original training data Z are no longer available at test
time, C̄j+ can be directly computed from Cj+ and z through the model update
formulas W
(36) or (37). The negative evidence conceptor is accordingly obtained by
−
+
+
C̄j = ¬ {C̄1+ , . . . , C̄j−1
, C̄j+1
, . . . , C̄9+ }.
Results of refined classification procedure: Averaged over 50 learn-test trials with independently sampled reservoir weights, the number of misclassifications
for the three types of classification (positive, negative, combined evidence) were 8.4
/ 5.9 / 3.4 (standard deviations 0.99 / 0.93 / 0.61). The training misclassification
errors for the combined classification was zero in all 50 trials.
The detection of good apertures through the ∇ criterion worked well. A manual
78
grid search through candidate apertures found that a minimum test misclassification rate of 3.0 (average over the 50 trials) from the combined classificator was obtained with an aperture α+ = 20, α− = 24 for both the positive and negative conceptors. The automated aperture detection yielded apertures α+ = 25, α− = 27
and a (combined classificator) misclassification rate of 3.4, close to the optimum.
Discussion. The following observations are worth noting.
Method also applies to static pattern classification. In the presented classification method, temporal input samples s (short preprocessed nonstationary timeseries) were transformed into static coding vectors z as a basis for
constructing conceptors. These z contained the original input signal s plus
the state response from a small reservoir driven by s. The reservoir was only
used to augment s by some random nonlinear interaction terms between the
entries in s. Conceptors were created and used in classification without referring back to the reservoir dynamics. This shows that conceptors can also
be useful in static pattern classification.
Extensibility. A classification model consisting of learnt conceptors Cj+ , Cj− for
k classes can be easily extended by new classes, because the computations
+
−
needed for new Ck+1
, Ck+1
only require positive training samples of the new
class. Similarly, an existing model Cj+ , Cj− can be extended by new training
samples without re-visiting original training data by an application of the
model extension formulae (36) or (37). In fact, the refined classification
procedure given above can be seen as an ad-hoc conditional model extension
by the test sample.
Including an “other” class. Given a learnt classification model Cj+ , Cj− for k
classes it appears
straightforward to include an “other” class by including
W
+
Cother
= ¬ {C1+ , . . . , Ck+ } and recomputing the negative evidence concep+
tors from the set {C1+ , . . . , Ck+ , Cother
} via (59). I have not tried this out
yet.
Discriminative nature of combined classification. The classification of the
combined type, paraphrased above as “sample seems to be from class j
and seems not to be from any of the others”, combines information from all
classes into an evidence vote for a candidate class j. Generally, in discriminative learning schemes for classifiers, too, contrasting information between the
classes is exploited. The difference is that in those schemes, these differences
are worked in at learning time, whereas in the presented conceptor-based
scheme they are evaluated at test time.
Benefits of Boolean operations. The three aforementioned points – extensibility, “other” class, discriminative classification – all hinge on the availability of the NOT and OR operations, in particular, on the associativity of the
latter.
79
Computational efficiency. The computational steps involved in learning and
applying conceptors are constructive. No iterative optimization steps are
involved (except that standard implementations of matrix inversion are iterative). This leads to short computation times. Learning conceptors from the
270 preprocessed data samples, including determining good apertures, took
650 ms and classifying a test sample took 0.7 ms for the basic and 64 ms for
the refined procedure (on a dual-core 2GHz Macintosh notebook computer,
using Matlab).
Competitiveness. The test misclassification rate of 3.4 is slightly better than
the best rate of about 4 that I am aware of in the literature outside own
work [57]. Given that the zero error performance in [57] was achieved with
an exceptionally expensive model (combining 1,000 independently sampled
classifiers), which furthermore is trained in a discriminative setup and thus is
not extensible, the attained performance level, the computational efficiency,
and the extensibility of the conceptor-base model render it a competitive
alternative to existing classification learning methods. It remains to be seen
though how it performs on other datasets.
Regularization by aperture adaptation? In supervised classification learning tasks, it is generally important to regularize models to find the best
balance between overfitting and under-exploiting training data. It appears
that the role of regularization is here played by the aperture adaptation,
though a theoretical analysis remains to be done.
Early stage of research. The proposed classifier learning scheme was based on
numerous ad-hoc design decisions, and quite different ways to exploit conceptors for classification are easily envisioned. Thus, in sum, the presented
study should be regarded as no more than a first demonstration of the basic
usefulness of conceptors for classification tasks.
3.14
Autoconceptors
3.14.1
Motivation and Overview
In the preceding sections I have defined conceptors as transforms C = R(R +
α−2 I)−1 of reservoir state correlation matrices R. In order to obtain some conceptor C j which captures a driving pattern pj , the network was driven by pj via
x(n + 1) = tanh(W ∗ x(n) + W in pj (n + 1) + b), the obtained reservoir states were
used to compute Rj , from which C j was computed. The conceptor C j could
then later be exploited via the conceptor-constrained update rule x(n + 1) =
C j tanh(W x(n) + b) or its variant x(n + 1) = C j tanh(W ∗ x(n) + D x(n) + b).
This way of using conceptors, however, requires that the conceptor matrices C j
are computed at learning time (when the original drivers are active), and they have
to be stored for later usage. Such a procedure is useful and feasible in engineering
80
or machine learning applications, where the conceptors C j may be written to
file for later use. It is also adequate for theoretical investigations of reservoir
dynamics, and logical analyses of relationships between reservoir dynamics induced
by different drivers, or constrained by different conceptors.
However, storing conceptor matrices is entirely implausible from a perspective
of neuroscience. A conceptor matrix has the same size as the original reservoir
weight matrix, that is, it is as large an entire network (up to a saving factor
of one half due to the symmetry of conceptor matrices). It is hard to envision
plausible models for computational neuroscience where learning a new pattern by
some RNN essentially would amount to creating an entire new network.
This motivates to look for ways of how conceptors can be used for constraining
reservoir dynamics without the necessity to store conceptors in the first place.
The network would have to create conceptors “on the fly” while it is performing
some relevant task. Specifically, we are interested in tasks or functionalities which
are relevant from a computational neuroscience point of view. This objective also
motivates to focus on algorithms which are not immediately biologically implausible. In my opinion, this largely excludes computations which explicitly exploit
the SVD of a matrix (although it has been tentatively argued that neural networks can perform principal component analysis [80] using biologically observed
mechanisms).
In the next subsections I investigate a version of conceptors with associated
modes of usage where there is no need to store conceptors and where computations are online adaptive and local in the sense that the information necessary for
adapting a synaptic weight is available at the concerned neuron. I will demonstrate the workings of these conceptors and algorithms in two functionalities, (i)
content-addressable memory (Section 3.14.3) and (ii) simultaneous de-noising and
classification of a signal (Section 3.16).
In this line of modeling, the conceptors are created by the reservoir itself at the
time of usage. There is no role for an external engineer or superordinate control
algorithm to “plug in” a conceptor. I will speak of autoconceptors to distinguish
these autonomously network-generated conceptors from the conceptors that are
externally stored and externally inserted into the reservoir dynamics. In discussions I will sometimes refer to those “externalistic” conceptors as alloconceptors.
Autoconceptors, like alloconceptors, are positive semidefinite matrices with
singular values in the unit interval. The semantic relationship to data, aperture
operations, and Boolean operations are identical for allo- and autoconceptors.
However, the way how autoconceptors are generated is different from alloconceptors, which leads to additional constraints on their algebraic characteristics. The
set of autoconceptor matrices is a proper subset of the conceptor matrices in general, as they were defined in Definition 2, i.e. the class of positive semidefinite
matrices with singular values ranging in the unit interval. The additional constraints arise from the circumstance that the reservoir states x(n) which shape an
autoconceptor C are themselves depending on C.
81
The treatment of autoconceptors will be structured as follows. I will first
introduce the basic equations of autoconceptors and their adaptation dynamics
(Section 3.14.2), demonstrate their working in a of content-addressable memory
task (Section 3.14.3) and mathematically analyse central properties of the adaptation dynamics (Section 3.14.4). The adaptation dynamics however has non-local
aspects which render it biologically implausible. In order to progress toward biologically feasible autoconceptor mechanisms, I will propose neural circuits which
implement autoconceptor dynamics in ways that require only local information for
synaptic weight changes (Section 3.15).
3.14.2
Basic Equations
The basic system equation for autoconceptor systems is
x(n + 1) = C(n) tanh(W ∗ x(n) + W in p(n + 1) + b)
(60)
or variants thereof, like
x(n + 1) = C(n) tanh(W x(n) + b)
(61)
x(n + 1) = C(n) tanh(W ∗ x(n) + Dx(n) + b),
(62)
or
the latter two for the situation after having patterns stored. The important novel
element in these equations is that C(n) is time-dependent. Its evolution will be
governed by adaptation rules that I will describe presently. C(n) need not be
positive semidefinite at all times; only when the adaptation of C(n) converges, the
resulting C matrices will have the algebraic properties of conceptors.
One can conceive of the system (60) as a two-layered neural network, where the
two layers have the same number of neurons, and where the layers are reciprocally
connected by the connection matrices C and W (Figure 30). The two layers have
states
r(n + 1) = tanh(W ∗ z(n) + W in p(n + 1) + b)
z(n + 1) = C r(n + 1).
(63)
(64)
The r layer has sigmoidal (here: tanh) units and the z layer has linear ones. The
customary reservoir state x becomes split into two states r and z, which can be
conceived of as states of two pools of neurons.
In order to determine an adaptation law for C(n), I replicate the line of reasoning that was employed to motivate the design of alloconceptors in Section 3.4.
Alloconceptors were designed to act as “regularized identity maps”, which led to
the defining criterion (6) in Definition 1:
C(R, α) = argminC E[kx − Cxk2 ] + α−2 kCk2fro .
82
W*
W in
p
r
C
!
z
! Network representation of a basic autoconceptor system. Bias b and
Figure 30:
optional readout mechanisms
are omitted.
!
! The broken arrow indicates that C
connections are online adaptive. For explanation see text.
The reservoir states x that appear in this criterion resulted from the update equation x(n + 1) = tanh(W ∗ x(n) + W in p(n + 1) + b). This led to the explicit solution
(7) stated in Proposition 1:
C(R, α) = R (R + α−2 I)−1 ,
where R was the reservoir state correlation matrix E[xx0 ]. I re-use this criterion
(6), which leads to an identical formula C = R (R + α−2 I)−1 for autoconceptors.
The crucial difference is that now the state correlation matrix R depends on C:
R = E[zz 0 ] = E[Cr(Cr)0 ] = C E[rr0 ] C =: CQC,
(65)
where we introduce Q = E[rr0 ]. This transforms the direct computation formula
(7) to a fixed-point equation:
C = CQC (CQC + α−2 I)−1 ,
which is equivalent to
(C − I)CQC − α−2 C = 0.
(66)
Since Q depends on r states, which in turn depend on z states, which in turn
depend on C again, Q depends on C and should be more appropriately be written
as QC . Analysing the fixed-point equation (C − I)CQC C − α−2 C = 0 is a little
inconvenient, and I defer this to Section 3.14.4. When one uses autoconceptors,
however, one does not need to explicitly solve (66). Instead, one can resort to a
version of the incremental adaptation rule (11) from Proposition 2:
C(n + 1) = C(n) + λ (z(n) − C(n) z(n)) z 0 (n) − α−2 C(n) ,
which implements a stochastic gradient descent with respect to the cost function
E[kz − Czk2 ] + α−2 kCk2fro . In the new sitation given by (60), the state z here
depends on C. This is, however, of no concern for using (11) in practice. We thus
complement the reservoir state update rule (60) with the conceptor update rule
(11) and comprise this in a definition:
83
Definition 5 An autoconceptive reservoir network is a two-layered RNN with
fixed weights W ∗ , W in and online adaptive weights C, whose dynamics are given
by
z(n + 1) = C(n) tanh(W ∗ z(n) + W in p(n + 1) + b)
C(n + 1) = C(n) + λ (z(n) − C(n) z(n)) z 0 (n) − α−2 C(n) ,
(67)
(68)
where λ is a learning rate and p(n) an input signal. Likewise, when the update
equation (67) is replaced by variants of the kind (61) or (62), we will speak of
autoconceptive networks.
I will derive in Section 3.14.4 that if the driver p is stationary and if C(n)
converges under these rules, then the limit C is positive semidefinite with singular values in the set (1/2, 1) ∪ {0}. Singular values asymptotically obtained
under the evolution (68) are either “large” (that is, greater than 1/2) or they are
zero, but they cannot be “small” but nonzero. If the aperture α is fixed at increasingly smaller values, increasingly many singular values will be forced to zero.
Furthermore, the analysis in Section 3.14.4 will also reveal that among the nonzero
singular values, the majority will be close to 1. Both effects together mean that autoconceptors are typically approximately hard conceptors, which can be regarded
as an intrinsic mechanism of contrast enhancement, or noise suppression.
3.14.3
Example: Autoconceptive Reservoirs as Content-Addressable
Memories
In previous sections I demonstrated how loaded patterns can be retrieved if the
associated conceptors are plugged into the network dynamics. These conceptors
must have been stored beforehand. The actual memory functionality thus resides
in whatever mechanism is used to store the conceptors; furthermore, a conceptor
is a heavyweight object with the size of the reservoir itself. It is biologically
implausible to create and “store” such a network-like object for every pattern that
is to be recalled.
In this section I describe how autoconceptor dynamics can be used to create
content-addressable memory systems. In such systems, recall is triggered by a cue
presentation of the item that is to be recalled. The memory system then should
in some way autonomously “lock into” a state or dynamics which autonomously
re-creates the cue item. In the model that will be described below, this “locking
into” spells out as running the reservoir in autoconceptive mode (using equations
(62) and (68)), by which process a conceptor corresponding to the cue pattern
shapes itself and enables the reservoir to autonomously re-generate the cue.
The archetype of content-addressable neural memories is the Hopfield network
[48]. In these networks, the cue is a static pattern (technically a vector, in demonstrations often an image), which typically is corrupted by noise or incomplete.
If the Hopfield network has been previously trained on the uncorrupted complete
84
pattern, its recurrent dynamics will converge to an attractor state which re-creates
the trained original from the corrupted cue. This pattern completion characteristic is the essence of the memory functionality in Hopfield networks. In the
autoconceptive model, the aspect of completion manifests itself in that the cue is
a brief presentation of a dynamic pattern, too short for a conceptor to be properly
adapted. After the cue is switched off, the autoconceptive dynamics continues to
shape the conceptor in an entirely autonomous way, until it is properly developed
and the reservoir re-creates the cue.
This autoconceptive adaptation is superficially analog to the convergence to
an attractor point in Hopfield networks. However, there are important conceptual
and mathematical differences between the two models. I will discuss them at the
end of this section.
Demonstration of basic architecture. To display the core idea of a contentaddressable memory, I ran simulations according to the following scheme:
1. Loading. A collection of k patterns pj (j = 1, . . . , k) was loaded in an N dimensional reservoir, yielding an input simulation matrix D as described in
Equation (40), and readout weights W out , as described in Section 3.3. No
conceptors are stored.
2. Recall. For each pattern pj , a recall run was executed which consisted of
three stages:
(a) Initial washout. Starting from a zero network state, the reservoir
was driven with pj for nwashout steps, in order to obtain a task-related
reservoir state.
(b) Cueing. The reservoir was continued to be driven with pj for another
ncue steps. During this cueing period, C j was adapted by using r(n +
1) = tanh(W r(n) + W in pj (n) + b), C j (n + 1) = C j (n) + λcue ((r(n) −
C j (n) r(n)) r0 (n) − α−2 C j (n)). At the beginning of this period, C j was
initialized to the zero matrix. At the end of this period, a conceptor
C j cue was obtained.
(c) Autonomous recall. The network run was continued for another
nrecall steps in a mode where the input was switched off and replaced
by the input simulation matrix D, and where the conceptor C j cue was
further adapted autonomously by the autoconceptive update mechanism, via z(n + 1) = C j (n) tanh(W z(n) + Dz(n) + b), C j (n + 1) =
C j (n) + λrecall ((z(n) − C j (n) z(n)) z 0 (n) − α−2 C j (n)). At the end of
this period, a conceptor C j recall was available.
3. Measuring quality of conceptors. The quality of the conceptors C j cue
and C j recall was measured by separate offline runs without conceptor adaptation using r(n) = tanh(W z(n) + Dz(n) + b); z(n + 1) = C j cue r(n) (or
85
z(n + 1) = C j recall r(n), respectively). A reconstructed pattern y(n) =
W out r(n) was obtained and its similarity with the original pattern pj was
quantified in terms of a NRMSE.
I carried out two instances of this experiment, using two different kinds of
patterns and parametrizations:
4-periodic pattern. The patterns were random integer-periodic patterns of period 4, where per pattern the four pattern points were sampled from a uniform distribution and then shift-scaled such that the range became [−1 1].
This normalization implies that the patterns are drawn from an essentially
3-parametric family (2 real-valued parameters for fixing the two pattern
values not equal to −1 or 1; one integer parameter for fixing the relative
temporal positioning of the −1 and 1 values). Experiment parameters:
k = 10, N = 100, α = 100, nwashout = 100, ncue = 15, nrecall = 300, γ cue =
0.02, γ recall = 0.01 (full detail in Section 4.7).
√
√
Mix of 2 irrational-period sines. Two sines of period lengths 30 and 30/2
were added with random phase angles and random amplitudes, where however the two amplitudes were constrained to sum to 1. This means that
patterns were drawn from a 2-parametric family. Parameters: k = 10, N =
200, α = 100, nwashout = 100, ncue = 30, nrecall = 10000, γ cue = γ recall = 0.01.
Furthermore, during the auto-adaptation period, strong Gaussian iid noise
was added to the reservoir state before applying the tanh, with a signal-tonoise rate of 1.
Figure 31 illustrates the outcomes of these two experiments. Main observations:
1. In all cases, the quality of the preliminary conceptor C j cue was very much
improved by the subsequent auto-adaptation (Panels B, D), leading to an
ultimate pattern reconstruction whose quality is similar to the one that
would be obtained from precomputed/stored conceptors.
2. The effects of the autoconceptive adaptation are reflected in the singular
value profiles of C j cue versus C j recall (Panels A, C). This is especially well
visible in the case of the sine mix patterns (for the period-4 patterns the effect
is too small to show up in the plotting resolution). During the short cueing
time, the online adaptation of the conceptor from a zero matrix to C j cue
only manages to build up a preliminary profile that could be intuitively
called “nascent”, which then “matures” in the ensuing network-conceptor
interaction during the autoconceptive recall period.
3. The conceptors C j recall have an almost rectangular singular value profile. In
the next section I will show that if autoconceptive adaptation converges, singular values are either exactly zero or greater than 0.5 (in fact, typically close
to 1), in agreement with what can be seen here. Autoconceptive adaptation
has a strong tendency to lead to almost hard conceptors.
86
Singular Values
y and p
0
1
1
0
!1
!1
log10 NRMSE
0
1
1
0
0
!1
!3
!4
1
1
!5
0
0
0
5
10
!1
0
5
!6
10
A.
B.
Singular Values
0
!0.2
0
!0.4
!1
!0.6
1
1
0
0
!1
1
0
0
10
20
1 2 3 4 5 6 7 8 9 10
Pattern index
y and p
1
log10 NRMSE
1
C.
!2
!0.8
!1
!1.2
1
!1.4
0
!1.6
!1
0
5
!1.8
10
D.
1 2 3 4 5 6 7 8 9 10
Pattern index
Figure 31: Basic content-addressable memory demos. A, B: 4-periodic pattern,
C, D: mix of sines pattern. Panels A, C show the first three of the 10 patterns.
The singular value plots show the first 10 (20, respectively) singular values of C j cue
(black) and C j recall (light gray). The “y and p” panels show the reconstructed pattern y obtained with C j recall (bold light gray) and the original training pattern pj
(broken black), after optimal phase-alignment. B, D plot the pattern reconstruction NRMSEs in log10 scale for the reconstructions obtained from C j cue (black
squares) and from C j recall (gray crosses). For explanation see text.
87
4. The fact that adapted autoconceptors typically have a close to rectangular
singular value spectrum renders the auto-adaptation process quite immune
against even strong state noise. Reservoir state noise components in directions of the nulled eigenvectors are entirely suppressed in the conceptorreservoir loop, and state noise components within the nonzero conceptor
eigenspace do not impede the development of a “clean” rectangular profile. In fact, state noise is even beneficial: it speeds up the auto-adaptation
process without a noticeable loss in final pattern reconstruction accuracy
(comparative simulations not documented here).
This noise robustness however depends on the existence of zero singular
values in the adapting autoconceptor C. In the simulations reported above,
such zeros were present from the start because the conceptor was initialized
as the zero matrix. If it had been initialized differently (for instance, as
identity matrix), the auto-adaptation would only asymptotically pull (the
majority of) singular values to zero, with noise robustness only gradually
increasing to the degree that the singular value spectrum of C becomes
increasingly rectangular. If noise robustness is desired, it can be reached
by additional adaptation mechanisms for C. In particular, it is helpful to
include a thresholding mechanism: all singular values of C(n) exceeding a
suitable threshold are set to 1, all singular values dropping below a certain
cutoff are zeroed (not shown).
Exploring the effects of increasing memory load – patterns from a
parametrized family. A central theme in neural memory research is the capacity of a neural storage system. In order to explore how recall accuracy depends
on the memory load, I carried out two further experiments, one for each pattern
type. Each of these experiments went along the following scheme:
1. Create a reservoir.
2. In separate trials, load this reservoir with an increasing number k of patterns
(ranging from k = 2 to k = 200 for the 4-period and from k = 2 to k = 100
for the mixed sines).
3. After loading, repeat the recall scheme described above, with the same parameters. Monitor the recall accuracy obtained from C j recall for the first 10
of the loaded patterns (if less than 10 were loaded, do it only for these).
4. In addition, per trial, also try to cue and “recall” 10 novel patterns that were
drawn randomly from the 4-periodic and mixed-sine family, respectively, and
which were not part of the collection loaded into the reservoir. Monitor the
“recall” accuracy of these novel patterns as well.
5. Repeat this entire scheme 5 times, with freshly created patterns, but re-using
always the same reservoir.
88
0
!0.2
!0.5
!0.4
!1
!0.6
log10 NRMSE
log10 NRMSE
!1.5
!2
!2.5
!3
!0.8
!1
!1.2
!3.5
!1.4
!4
!1.6
!4.5
!5
A.
2
3
5
8 12 16 25
50
Nr of loaded patterns
100
!1.8
200
B.
2
3
5
8 12 16 25
Nr of loaded patterns
50
100
Figure 32: Exploring the effects of memory load. A: 4-periodic patterns, B:
mix-of-sines patterns. Each diagram shows the log10 NRMSE of recalling loaded
patterns with C cue (black solid line) and with C recall (black broken line), as well
as of “recalling” patterns not contained in the loaded set, again obtained from
C cue (gray solid line) and with C recall (gray broken line). Error bars indicate 95
% confidence intervals. Both axes are in logarithmic scaling. For explanation see
text.
Figure 32 shows the results of these experiments. The plotted curves are the
summary averages over the 10 recall targets and the 5 experiment repetitions.
Each plot point in the diagrams thus reflects an average over 50 NRMSE values
(except in cases where k < 10 patterns were stored; then plotted values correspond
to averages over 5k NRMSE values for recalling of loaded patterns). I list the main
findings:
1. For all numbers of stored patterns, and for both the recall loaded patterns
and recall novel patterns conditions, the autoconceptive “maturation” from
C cue to C recall with an improvement of recall accuracy is found again.
2. The final C recall -based recall accuracy in the recall loaded pattern condition
has a sigmoid shape for both pattern types. The steepest ascent of the
sigmoid (fastest deterioration of recall accuracy with increase of memory
load) occurs at about the point where the summed quota of all C cue reaches
the reservoir size N – the point where the network is “full” according to
this criterion (a related effect was encountered in the incremental loading
study reported in Section 3.11). When the memory load is further increased
beyond this point (one might say the network becomes “overloaded”), the
recall accuracy does not break down but levels out on a plateau which still
translates into a recall performance where there is a strong similarity between
the target signal and the reconstructed one.
89
3. In the recall novel patterns conditions, one finds a steady improvement of
recall accuracy with increasing memory load. For large memory loads, the
accuracy in the recall novel patterns condition is virtually the same as in the
recall loaded patterns conditions.
Similar findings were obtained in other simulation studies (not documented
here) with other types of patterns, where in each study the patterns were drawn
from a parametrized family.
A crucial characteristic of these experiments is that the patterns were samples
from a parametrized family. They shared a family resemblance. This mutual
relatedness of patterns is exploited by the network: for large numbers k of stored
patterns, the storing/recall mechanism effectively acquires a model of the entire
parametric family, a circumstance revealed by the essentially equal recall accuracy
in the recall loaded patterns and recall novel patterns conditions. In contrast, for
small k, the recall loaded patterns condition enables a recall accuracy which is
superior to the recall novel patterns condition: the memory system stores/recalls
individual patterns. I find this worth a special emphasis:
• For small numbers of loaded patterns (before the point of network overloading) the system stores and recalls individual patterns. The input simulation
matrix D represents individual patterns.
• For large numbers of loaded patterns (overloading the network), the system
learns a representation of the parametrized pattern family and can be cued
with, and will “recall”, any pattern from that family. The input simulation
matrix D represents the class of patterns.
At around the point of overloading, the system, in a sense, changes its nature
from a mere storing-of-individuals device to a learning-of-class mechanism. I call
this the class learning effect.
Effects of increasing memory load – mutually unrelated patterns. A
precondition for the class learning effect is that the parametric pattern family is
simple enough to become represented by the network. If the pattern family is too
richly structured to be captured by a given network size, or if patterns do not have
a family resemblance at all, the effect cannot arise. If a network is loaded with
such patterns, and then cued with novel patterns, the “recall” accuracy will be on
chance level; furthermore, as k increases beyond the overloading region, the recall
accuracy of patterns contained in the loaded collection will decline to chance level
too.
In order to demonstrate this, I loaded the same 100-unit reservoir that was
used in the 4-periodic pattern experiments with random periodic patterns whose
periods ranged from 3 through 9. While technically this is still a parametric family,
the number of parameters needed to characterize a sample pattern is 8, which
90
renders this family far too complex for a 100-unit reservoir. Figure 33 illustrates
what, expectedly, happens when one loads increasingly large numbers of such
effectively unrelated patterns. The NRMSE for the recall novel patterns condition
is about 1 throughout, which corresponds to entirely uncorrelated pattern versus
reconstruction pairs; and this NRMSE is also approached for large k in the recall
loaded patterns condition.
0.5
0
log10 NRMSE
!0.5
!1
!1.5
!2
!2.5
!3
!3.5
4
6
8 10 14 1822 30 40 50 65 85110
Nr of loaded patterns
Figure 33: Effects of memory load on recall accuracy for unrelated patterns. Figure
layout as in Figure 32. For explanation see text.
Discussion. Neural memory mechanisms – how to store patterns in, and retrieve
from, neural networks – is obviously an important topic of research. Conceptorbased mechanisms bring novel aspects to this widely studied field.
The paradigmatic model for content-addressable storage of patterns in a neural
network is undoubtedly the family of auto-associative neural networks (AANNs)
whose analysis and design was pioneered by Palm [82] and Hopfield [48] (with a
rich history in theoretical neuroscience, referenced in [82]). Most of these models
are characterized by the following properties:
• AANNs with N units are used to store static patterns which are themselves
N -dimensional vectors. The activity profile of the entire network coincides
with the very patterns. In many demonstrations, these patterns are rendered
as 2-dimensional images.
• The networks are typically employed, after training, in pattern completion or
restauration tasks, where an incomplete or distorted N -dimensional pattern
is set as the initial N -dimensional network state. The network then should
evolve toward a completed or restored pattern state.
91
• AANNs have symmetric connections and (typically) binary neurons. Their
recurrent dynamics can be formalized as a descent along an energy function,
which leads to convergence to fixed points which are determined by the input
pattern.
• An auto-associative network is trained from a set of k reference patterns,
where the network weights are adapted such that the network state energy
associated with each training pattern is minimized. If successful, this leads
to an energy landscape over state space which assumes local minima at the
network states that are identical to the reference patterns.
The comprehensive and transparent mathematical theory available for AANNs
has left a strong imprint on our preconceptions of what are essential features of
a content-addressable neural memory. Specifically, AANN research has settled
the way how the task of storing items in an associative memory is framed in the
first place: “given k reference patterns, train a network such that in exploitation,
these patterns can be reconstructed from incomplete cues”. This leads naturally
to identifying stored memory items with attractors in the network dynamics. Importantly, memory items are seen as discrete, individual entities. For convenience
I will call this the “discrete items stored as attractors” (DISA) paradigm.
Beyond modeling memory functionality proper, the DISA paradigm is historically and conceptually connected to a wide range of models of neural representations of conceptual knowledge, where attractors are taken as the neural representatives of discrete concepts. To name only three kinds of such models: point
attractors (cell assemblies and bistable neurons) in the working memory literature
[23]; spatiotemporal attractors in neural field theories of cortical representation
[95, 30, 31]; (lobes of) chaotic attractors as richly structured object and percept
representations [111, 6].
Attractors, by definition, keep the system trajectory confined within them.
Since clearly cognitive processes do not become ultimately trapped in attractors,
it has been a long-standing modeling challenge to account for “attractors that
can be left again” – that is, to partly disengage from a strict DISA paradigm.
Many answers have been proposed. Neural noise is a plausible agent to “kick”
a trajectory out of an attractor, but a problem with noise is its unspecificity
which is not easily reconciled with systematic information processing. A number
of alternative “attractor-like” phenomena have been considered that may arise in
high-dimensional nonlinear dynamics and offer escapes from the trapping problem: saddle point dynamics or homoclinic cycles [89, 41]; chaotic itinerancy [105];
attractor relics, attractor ruins, or attractor ghosts [103]; transient attractors [53];
unstable attractors [104]; high-dimensional attractors (initially named partial attractors) [70]; attractor landscapes [78].
All of these lines of work revolve around a fundamental conundrum: on the
one hand, neural representations of conceptual entities need to have some kind of
stability – this renders them identifiable, noise-robust, and temporally persistent
92
when needed. On the other hand, there must be cognitively meaningful mechanisms for a fast switching between neural representational states or modes. This
riddle is not yet solved in a widely accepted way. Autoconceptive plane attractor
dynamics may lead to yet another answer. This kind of dynamics intrinsically
combines dynamical stability (in directions complementary to the plane of attraction) with dynamical neutrality (within the plane attractor). However, in the
next section we will see that this picture, while giving a good approximation, is
too simple.
3.14.4
Analysis of Autoconceptor Adaptation Dynamics
Here I present a formal analysis of some asymptotic properties of the conceptor
adaptation dynamics.
Problem Statement
We consider the system of the coupled fast network state updates and slow conceptor adaptation given by
z(n + 1) = C(n) tanh(W z(n))
(69)
and
C(n + 1) = C(n) + λ (z(n + 1) − C(n) z(n + 1)) z 0 (n + 1) − α−2 C(n)
= C(n) + λ (I − C(n)) z(n + 1)z 0 (n + 1) − α−2 C(n) ,
(70)
where λ is a learning rate. When λ is small enough, the instantaneous state
correlation z(n)z 0 (n) in (70) can be replaced by its expectation under C fixed at
C(n), that is, we consider the dynamical system in time k
z(k + 1) = C(n) tanh(W z(k))
and take the expectation of zz 0 under this dynamics,
En [zz 0 ] := Ek [C(n) tanh(W z(k)) tanh(W z(k))0 C(n)0 ]
= C(n) Ek [tanh(W z(k)) tanh(W z(k))0 ]C(n)0 =: C(n)Q(n)C(n)0 ,
where Q(n) is a positive semi-definite correlation matrix. Note that Q(n) is a
function of C and itself changes on the slow timescale of the C adaptation. For
further analysis it is convenient to change to continuous time and instead of (70)
consider
Ċ(t) = (I − C(t)) C(t)Q(t)C 0 (t) − α−2 C(t).
(71)
I now investigate the nature of potential fixed point solutions under this dynamics. If C is a fixed point of this dynamics, Q(t) is constant. In order to
93
investigate the nature of such fixed point solutions, we analyse solutions in C for
the general fixed point equation associated with (71), i.e. solutions in C of
0 = (I − C) CQC 0 − α−2 C,
(72)
where Q is some positive semidefinite matrix. We will denote the dimension of C
by N throughout the remainder of this section. Let V DV 0 = Q be the SVD of Q,
where D is a diagonal matrix containing the singular values of Q on its diagonal,
without loss of generality in descending order. Then (72) is equivalent to
0 = V 0 ((I − C) CQC 0 − α−2 C)V
= (I − V 0 CV ) V 0 CV D(V 0 CV )0 − α−2 V 0 CV.
(73)
We may therefore assume that Q is in descending diagonal form D, analyse
solutions of
0 = (I − C) CDC 0 − α−2 C,
(74)
and then transform these solutions C of (74) back to solutions of (72) by C →
V CV 0 . In the remainder we will only consider solutions of (74). I will characterize
the fixed points of this system and analyse their stability properties.
Characterizing the Fixed-point Solutions
The case α = 0. In this degenerate case, neither the discrete-time update rule
(70) nor the dynamical equation (71) is well-defined. The aperture cannot be set
to zero in practical applications where (70) is used for conceptor adaptation.
However, it is clear that (i) for any α > 0, C = 0 is a fixed point solution of (71),
and that (ii) if we define B(α) = sup{kCk | C is a fixed-point solution of (71)},
then limα→0 B(α) = 0. This justifies to set, by convention, C = 0 as the unique
fixed point of (71) in the case α = 0. In practical applications this could be implemented by a reset mechanism: whenever α = 0 is set by some superordinate
control mechanism, the online adaptation (70) is over-ruled and C(n) is immediately set to 0.
The case α = ∞. In the case α = ∞ (i.e., α−2 = 0) our task is to characterize
the solutions C of
0 = (I − C)CDC 0 .
(75)
We first assume that D has full rank. Fix some k ≤ N . We proceed to
characterize rank-k solutions C of (75). CDC 0 is positive semidefinite and has
rank k, thus it has an SVD CDC 0 = U ΣU 0 where Σ is diagonal nonnegative and
can be assumed to be in descending order, i.e. its diagonal is (σ1 , . . . , σk , 0, . . . , 0)0
with σi > 0. Any solution C of (75) must satisfy U ΣU 0 = CU ΣU 0 , or equivalently,
Σ = U 0 CU Σ. It is easy to see that this entails that U 0 CU is of the form
Ik×k 0
0
U CU =
0 A
94
for some arbitrary (n − k) × (n − k) submatrix A. Requesting rank(C) = k implies
A = 0 and hence
Ik×k 0
C=U
U 0.
(76)
0 0
Since conversely, if U is any orthonormal matrix, a matrix C of the form given
in (76) satisfies C 2 = C, any such C solves (75). Therefore, the rank-k solutions
of (75) are exactly the matrices of type (76).
If D has rank l < N , again we fix a desired rank k for solutions C. Again
let CDC 0 = U ΣU 0 , with Σ in descending order. Σ has a rank m which satisfies
m ≤ k, l. From considering Σ = U 0 CU Σ it follows that U 0 CU has the form
Im×m 0
0
U CU =
0
A
for some (N − m) × (N − m) submatrix A. Since we prescribed C to have rank k,
the rank of A is k − m. Let U>m be the n × (N − m) submatrix of U made from
the columns with indices greater than m. We rearrange U ΣU 0 = CDC 0 to
Im×m 0
Im×m 0
0
Σ=
U DU
,
0
A
0
A
from which it follows (since the diagonal of Σ is zero at positions greater than m)
0
that AU>m
DU>m A0 = 0. Since the diagonal of D is zero exactly on positions > l,
this is equivalent to
A (U (1 : l, m + 1 : N ))0 = A U (m + 1 : N, 1 : l) = 0.
(77)
We now find that (77) is already sufficient to make C = U (Im×m |0 / 0|A) U 0
solve (75), because a simple algebraic calculation yields
(U (1 : N, 1 : m))0
CD = CCD = U
D.
0
We thus have determined the rank-k solutions of (75) to be all matrices of
form C = U (Im×m |0 / 0|A) U 0 , subject to (i) m ≤ l, k, (ii) rank(A) = k − m, (iii)
A (U (1 : l, m + 1 : N ))0 = 0. Elementary considerations (omitted here) lead to the
following generative procedure to obtain all of these matrices:
1. Choose m satisfying l − N + k ≤ m ≤ k.
2. Choose a size N × l matrix Ũ 0 made from orthonormal columns which is zero
in the last k − m rows (this is possible due to the choice of m).
3. Choose an arbitrary (N − m) × (N − m) matrix A of SVD form A = V ∆W 0
where the diagonal matrix ∆ is in ascending order and is zero exactly on the
first N − k diagonal positions (hence rank(A) = k − m).
95
4. Put
Ũ˜ 0 =
Im×m 0
0
W
Ũ 0 .
This preserves orthonormality of columns, i.e. Ũ˜ 0 is still made of orthonormal
columns. Furthermore, it holds that (0|A) Ũ˜ 0 = 0.
5. Pad Ũ˜ 0 by adding arbitrary N − l further orthonormal colums to the right,
obtaining an N × N orthonormal U 0 .
6. We have now obtained a rank-k solution
Im×m 0
C=U
U 0,
0
A
(78)
where we have put U to be the transpose of the matrix U 0 that was previously
constructed.
The case 0 < α < ∞. We proceed under the assumption that α < ∞, that is,
∞ > α−2 > 0.
I first show that any solution C of (74) is a positive semidefinite matrix. The
matrix CDC 0 is positive semidefinite and therefore has a SVD of the form U ΣU 0 =
CDC 0 , where U is orthonormal and real and Σ is the diagonal matrix with the
singular values of CDC 0 on its diagonal, without loss of generality in descending
order. From (I − C)U ΣU 0 = α−2 C it follows that
U ΣU 0 = α−2 C + C U ΣU 0 = C (α−2 I + U ΣU 0 )
= C (U (α−2 I + Σ)U 0 ).
α−2 I + Σ and hence U (α−2 I + Σ)U 0 are nonsingular because α−2 > 0, therefore
C = U ΣU 0 (U (α−2 I + Σ)U 0 )−1 = U Σ(α−2 I + Σ)−1 U 0 =: U SU 0 ,
where S = Σ(α−2 I+Σ)−1 is a diagonal matrix, and in descending order since Σ was
in descending order. We therefore know that any solution C of (74) is of the form
C = U SU 0 , where U is the same as in CDC 0 = U ΣU 0 . From S = Σ(α−2 I + Σ)−1
it furthermore follows that si < 1 for all singular values si of C, that is, C is a
conceptor matrix.
We now want to obtain a complete overview of all solutions C = U SU 0 of
(74), expressed in terms of an orthonormal real matrix U and a nonnegative real
diagonal matrix S. This amounts to finding the solutions in S and U of
(S − S 2 )U 0 DU S = α−2 S,
(79)
subject to S being nonnegative real diagonal and U being real orthonormal. Without loss of generality we furthermore may assume that the entries in S are in
descending order.
96
Some observations are immediate. First, the rank of S is bounded by the rank
of D, that is, the number of nonzero diagonal elements in S cannot exceed the
number of nonzero elements in D. Second, if U, S is a solution, and S ∗ is the
same as S except that some nonzero elements in S are nulled, then U, S ∗ is also
a solution (to see this, left-right multiply both sides of (79) with a thinned-out
identity matrix that has zeros on the diagonal positions which one wishes to null).
Fix some k ≤ rank(D). We want to determine all rank-k solutions U, S, i.e.
where S has exactly k nonzero elements that appear in descending order in the
first k diagonal positions. We write Sk to denote diagonal real matrices of size
k × k whose diagonal entries are all positive. Furthermore, we write Uk to denote
any N × k matrix whose columns are real orthonormal.
It is clear that if S, U solve (79) and rank(S) = k (and S is in descending
order), and if U ∗ differs from U only in the last N − k columns, then also S, U ∗
solve (79). Thus, if we have all solutions Sk , Uk of
(Sk − Sk2 )Uk0 DUk Sk = α−2 Sk ,
(80)
then we get all rank-k solutions S, U to (74) by padding Sk with N − k zero
rows/columns, and extending Uk to full size N × n by appending any choice of
orthonormal columns from the orthogonal complement of Uk . We therefore only
have to characterize the solutions Sk , Uk of (80), or equivalently, of
Uk0 DUk = α−2 (Sk − Sk2 )−1 .
(81)
To find such Sk , Uk , we first consider solutions S̃k , Uk of
Uk0 DUk = S̃k ,
(82)
subject to S̃k being diagonal with positive diagonal elements. For this we employ
the Cauchy interlacing theorem and its converse. I restate, in a simple special case
adapted to the needs at hand, this result from [27] where it is presented in greater
generality.
Theorem 1 (Adapted from Theorem 1 in [27], see remark of author at the end
of the proof of that theorem for a justification of the version that I render here.)
Let A, B be two symmetric real matrices with dim(A) = n ≥ k = dim(B), and
singular values σ1 , . . . , σn and τ1 , . . . , τk (in descending order). Then there exists a
real n × k matrix U with U 0 U = Ik×k and U 0 AU = B if and only if for i = 1, . . . , k
it holds that σi ≥ τi ≥ σn−k+i .
This theorem implies that if Uk , S̃k is any solution of (82), with Uk made of
k orthonormal columns and S̃k diagonal with diagonal elements s̃i (where j =
1, . . . , k, and the enumeration is in descending order), then the latter “interlace”
with the diagonal entries d1 , . . . , dN of D per di ≥ s̃i ≥ dN −k+i . And conversely,
any diagonal matrix S̃k , whose elements interlace with the diagonal elements of
D, appears in a solution Uk , S̃k of (82).
97
Equipped with this overview of solutions to (82), we revert from (82) to (81).
Solving S̃k = α−2 (Sk − Sk2 )−1 for Sk we find that the diagonal elements si of Sk
relate to the s̃i by
s
!
si =
1
2
1±
1−
4α−2
s̃i
.
(83)
Since si must be positive real and smaller than 1, only such solutions S̃k to (82)
whose entries are all greater than 4α−2 yield admissible solutions to our original
problem (81). The interlacing condition then teaches us that the possible rank of
solutions C of (74) is bounded from above by the number of entries in D greater
than 4α−2 .
For each value s̃i > 4α−2 , (83) gives two solutions si,1 < 1/2 < si,2 . We will
show further below that the solutions smaller than 1/2 are unstable while the
solutions greater than 1/2 are stable in a certain sense.
Summarizing and adding algorithmic detail, we obtain all rank-k solutions
C = U SU 0 for (74) as follows:
1. Check whether D has at least k entries greater than 4α−2 . If not, there are
no rank-k solutions. If yes, proceed.
2. Find a solution in Uk , S̃k of Uk0 DUk = S̃k , with S̃k being diagonal with
diagonal elements greater than 4α−2 , and interlacing with the elements of
D. (Note: the proof of Theorem 1 in [27] is constructive and could be used
for finding Uk given S̃k .)
3. Compute Sk via (83), choosing between the ± options at will.
4. Pad Uk with any orthogonal complement and Sk with further zero rows and
columns to full n×n sized U, S, to finally obtain a rank-k solution C = U SU 0
for (74).
Stability Analysis of Fixed-point Solutions
The case α = ∞. Note again that α = ∞ is the same as α−2 = 0. We consider
the time evolution of the quantity kI −Ck2 as C evolves under the zero-α−2 version
of (71):
Ċ(t) = (I − C(t)) C(t)Q(t)C 0 (t).
(84)
We obtain
(kI − Ck2 )˙ =
=
=
=
trace((I − C)(I − C 0 ))˙
trace((C − C 2 )QC 0 C 0 + CCQ(C 0 − C 02 ) − (C − C 2 )QC 0 − CQ(C 0 − C 02 ))
2 trace((C − C 2 )Q(C 02 − C 0 ))
−2 trace((C − C 2 )Q(C 0 − C 02 )) ≤ 0,
(85)
98
where in the last line we use that the trace of a positive semidefinite matrix is
nonnegative. This finding instructs us that no other than the identity C = I
can be a stable solution of (84), in the sense that all eigenvalues of the associated
Jacobian are negative. If Q(t) has full rank for all t, then indeed this is the case (it
is easy to show that kI − C(t)k2 is strictly decreasing, hence a Lyapunov function
in a neighborhood of C = I).
The stability characteristics of other (not full-rank) fixed points of (84) are intricate. If one computes the eigenvalues of the Jacobian at rank-k fixed points C
(i.e. solutions of sort C = U (Ik×k |0 / 0|0)U 0 , see (76)), where k < N , one finds negative values and zeros, but no positive values. (The computation of the Jacobian
follows the pattern of the Jacobian for the case α < ∞, see below, but is simpler;
it is omitted here). Some of the zeros correspond to perturbation directions of C
which change only the coordinate transforming matrices U . These perturbations
are neutrally stable in the sense of leading from one fixed point solution to another
one, and satisfy C + ∆ = (C + ∆)2 . However, other perturbations C + ∆ with the
property that C +∆ 6= (C +∆)2 lead to (kI −Ck2 )˙ < 0. After such a perturbation,
the matrix C + ∆ will evolve toward I in the Frobenius norm. Since the Jacobian
of C has no positive eigenvalues, this instability is non-hyperbolic. In simulations
one accordingly finds that after a small perturbation ∆ is added, the divergence
away from C is initially extremely slow, and prone to be numerically misjudged
to be zero.
For rank-deficient Q, which leads to fixed points of sort C = U (Im×m |0 / 0|A)U 0 ,
the computation of Jacobians becomes involved (mainly because A may be nonsymmetric) and I did not construct them. In our context, where Q derives from a
random RNN, Q can be expected to have full rank, so a detailed investigation of
the rank-deficient case would be an academic exercise.
The case 0 < α < ∞. This is the case of greatest practical relevance, and I
spent a considerable effort on elucidating it.
Note that α < ∞ is equivalent to α−2 > 0. Let C0 = U SU 0 be a rank-k
fixed point of Ċ = (I − C)CDC 0 − α−2 C, where α−2 > 0, U SU 0 is the SVD of
C and without loss of generality the singular values s1 , . . . , sN in the diagonal
matrix S are in descending order, with s1 , . . . , sk > 0 and si = 0 for i > k (where
1 ≤ k ≤ N ). In order to understand the stability properties of the dynamics Ċ in
a neighborhood of C0 , we compute the eigenvalues of the Jacobian JC = ∂ Ċ/∂C
at point C0 . Notice that C is an N × N matrix whose entries must be rearranged
into a vector of size N 2 × 1 in order to arrive at the customary representation of a
Jacobian. JC is thus an N 2 × N 2 matrix which should be more correctly written
as JC (µ, ν) = ∂ vec Ċ(µ)/∂ vec C(ν), where vec is the rearrangement operator
(1 ≤ µ, ν ≤ N 2 are the indices of the matrix JC ). Details are given in Section 5.10
within the proof of the following central proposition:
Proposition 15 The Jacobian JC (µ, ν) = ∂ vec Ċ(µ)/∂ vec C(ν) of a rank-k fixed
point of (71) has the following multiset of eigenvalues:
99
1. k(N − k) instances of 0,
2. N (N − k) instances of −α−2 ,
3. k eigenvalues α−2 (1 − 2sl )/(1 − sl ), where l = 1, . . . , k,
4. k(k − 1) eigenvalues which come in pairs of the form
s
2
−2
sm
sm
α sl
sl
+
±
−
+ 4 ,
λ1,2 =
2
sm − 1 sl − 1
sm − 1 sl − 1
where m < l ≤ k.
An inspection of sort 3. eigenvalues reveals that whenever one of the sl is
smaller than 1/2, this eigenvalue is positive and hence the fixed point C0 is unstable.
If some sl is exactly equal to 1/2, one obtains additional zero eigenvalues by
3. I will exclude such cases in the following discussion, considering them to be
non-generic.
If all sl are greater than 1/2, it is straightforward to show that the values of
sorts 3. and 4. are negative. Altogether, JP thus has k(N −k) times the eigenvalue
0 and otherwise negative ones. I will call such solutions 1/2-generic. All solutions
that one will effectively obtain when conceptor auto-adaptation converges are of
this kind.
This characterization of the eigenvalue spectrum of 1/2-generic solutions does
not yet allow us to draw firm conclusions about how such a solution will react
to perturbations. There are two reasons why Proposition 15 affords but a partial
insight in the stability of 1/2-generic solutions. (A) The directions connected to
zero eigenvalues span a k(N − k)-dimensional center manifold whose dynamics
remains un-analysed. It may be stable, unstable, or neutral. (B) When a 1/2generic solution is perturbed, the matrix D which reflects the conceptor-reservoir
interaction will change: D is in fact a function of C and should be more correctly
written D = D(C). In our linearization around fixed point solutions we implicitly
considered D to be constant. It is unclear whether a full treatment using D =
D(C) would lead to a different qualitative picture. Furthermore, (A) and (B) are
liable to combine their effects. This is especially relevant for the dynamics on the
center manifold, because its qualitative dynamics is determined by components
from higher-order approximations to (71) which are more susceptible to become
qualitatively changed by non-constant D than the dynamical components of (71)
orthogonal to the center manifold.
Taking (A) and (B) into account, I now outline a hypothetical picture of the
dynamics of (71) in the vicinity of 1/2-generic fixed-point solutions. This picture
is based only on plausibility considerations, but it is in agreement with what I
observe in simulations.
100
First, a dimensional argument sheds more light on the nature of the dynamics in
the k(N − k)-dimensional center manifold. Consider a 1/2-generic rank-k solution
C = U SU 0 of (79). Recall that the singular values si in S were derived from s̃i
which interlace with the diagonal elements d1 , . . . , dN of D by di ≥ s̃i ≥ dN −k+i ,
and where Uk0 DUk = S̃k (Equation (82)). I call C a 1/2&interlacing-generic
solution if the interlacing is proper, i.e. if di > s̃i > dN −k+i . Assume furthermore
that D(C) is constant in a neighborhood of C. In this case, differential changes to
Uk in (82) will lead to differential changes in S̃k . If these changes to Uk respect the
conditions (i) that Uk remains orthonormal and (ii) that S̃k remains diagonal, the
changes to Uk lead to new fixed point solutions. The first constraint (i) allows us
to change Uk with (N − 1) + (N − 2) + . . . + (N − k) = kN − k(k + 1)/2 degrees of
freedom. The second constraint (ii) reduces this by (k − 1)k/2 degrees of freedom.
Altogether we have kN − k(k + 1)/2 − (k − 1)k/2 = k(N − k) differential directions
of change of C that lead to new fixed points. This coincides with the dimension of
the center manifold associated with C. We can conclude that the center manifold
of a 1/2&interlacing-generic C extends exactly in the directions of neighboring
fixed point solutions. This picture is based however on the assumption of constant
D. If the dependency of D on C is included in the picture, we would not expect
to find any other fixed point solutions at all in a small enough neighborhood of
C. Generically, fixed point solutions of an ODE are isolated. Therefore, in the
light of the considerations made so far, we would expect to find isolated fixed
point solutions C, corresponding to close approximations of stored patterns. In a
local vicinity of such solutions, the autoconceptor adaptation would presumably
progress on two timescales: a fast convergence toward the center manifold K
associated with the fixed point C, superimposed on a slow convergence toward C
within K (Figure 34 A).
The situation becomes particularly interesting when many patterns from a dparametric class have been stored. Taking into account what the stability analysis
above has revealed about center manifolds of fixed points C, I propose the following
picture as a working hypothesis for the geometry of conceptor adaptation dynamics
that arises when a d-parametric pattern class has been stored by overloading:
• The storing procedure leads to a number of stable fixed point solutions Ci
for the autoconceptor adaptation (blue dots in Figure 34 B). These Ci are
associated with patterns from the pattern family, but need not coincide with
the sample patterns that were loaded.
• The k(N − k)-dimensional center manifolds of the Ci merge into a comprehensive manifold K of the same dimension. In the vicinity of K, the
autoadaptive C evolution leads to a convergence toward K.
• Within K a d-dimensional submanifold M is embedded, representing the
learnt class of patterns. Notice that we would typically expect d << k(N −k)
(examples in the previous section had d = 2 or d = 3, but k(N − k) in
101
A
B
"
"
!
Figure 34: Hypothetical phase portraits of C autoadaptation in the parameter
space of C (schematic). Blue points show stable fixed point solutions C. The gray
plane in A represents the center manifold K and in B the merged center manifolds
of neighboring fixed point C. Green arrows represent sample trajectories of C
adaptation. Green crosses mark the starting points of the adaptation trajectories
set by the cueing procedure. A. When a small number of patterns has been loaded,
individual stable fixed point conceptors C are created. B. In the case of learning
a d-parametric pattern class, fixed point solutions Ci become located within a
d-dimensional pattern manifold M (bold magenta line). For explanation see text.
the order of several 100). Conceptor matrices located on M correspond to
patterns from the learnt class.
• The convergence of C adaptation trajectories toward K is superimposed with
a slower contractive dynamics within K toward the class submanifold M.
• The combined effects of the attraction toward K and furthermore toward M
appear in simulations as if M were acting as a plane attractor.
• On an even slower timescale, within M there is an attraction toward the
isolated fixed point solutions Ci . This timescale is so slow that the motion
within M toward the fixed points Ci will be hardly observed in simulations.
In order to corroborate this refined picture, and especially to confirm the last
point from the list above, I carried out a long-duration content-addressable memory simulation along the lines described in Section 3.14.3. Ten 5-periodic patterns
were loaded into a small (50 units) reservoir. These patterns represented ten stages
of a linear morph between two similar patterns p1 and p10 , resulting in a morph
sequence p1 , p2 , . . . , p10 where pi = (1 − (i − 1)/9) p1 + ((i − 1)/9) p10 , thus representing instances from a 1-parametric family. Considering what was found in
Section 3.14.3, loading these ten patterns should enable the system to re-generate
102
by auto-adaptation any linear morph ptest between p1 and p10 after being cued
with ptest .
Figure 35: Numerical exploration of fixed point solutions under C auto-adaptation.
Each panel shows pairwise distances of 20 conceptor matrices obtained after n
auto-adaptation steps, after being cued along a 20-step morph sequence of cue
signals. Color coding: blue – zero distance; red – maximum distance. For explanation see text.
After loading, the system was cued with 20 different cues. In each of these
j = 1, . . . , 20 conditions, the cueing pattern pjtest was the j-th linear interpolation
between the stored p1 and p10 . The cueing was done for 20 steps, following the
procedure given at the beginning of Section 3.14.3. At the end of the cueing, the
system will be securely driven into a state z that is very accurately connected to
re-generating the pattern pjtest , and the conceptor matrix that has developed by
the end of the cueing would enable the system to re-generate a close simile of pjtest
(a post-cue log10 NRMSE of about −2.7 was obtained in this simulation).
After cueing, the system was left running in conceptor auto-adaptation mode
using (68) for 1 Mio timesteps, with an adaptation rate of λ = 0.01.
At times n = 1, 1000, 10000, 1e6 the situation of convergence was assessed
as follows. The pairwise distances between the current twenty autoconceptors
C j (n) were compared, resulting in a 20 × 20 distance matrix D(n) = (kC k (n) −
C l (n)kfro )k,l=1,...,20 . Figure 35 shows color plots of these distance matrices. The
outcome: at the beginning of autoadaptation (n = 1), the 20 autoconceptors are
spaced almost equally widely from each other. In terms of the schematic in Figure
34 B, they would all be almost equi-distantly lined up close to M. Then, as the
adaptation time n grows, they contract toward three point attractors within M
(which would correspond to a version of 34 B with three blue dots). These three
point attractors correspond to the three dark blue squares on the diagonal of the
last distance matrix shown in Figure 35.
This singular simulation cannot, of course, provide conclusive evidence that
the qualitative picture proposed in Figure 34 is correct. A rigorous mathematical
characterization of the hypothetical manifold M and its relation to the center
manifolds of fixed point solutions of the adaptation dynamics needs to be worked
out.
Plane attractors have been proposed as models for a number of biological neural
103
adaptation processes (summarized in [24]). A classical example is gaze direction
control. The fact that animals can fix their gaze in arbitrary (continuously many)
directions has been modelled by plane attractors in the oculomotoric neural control
system. Each gaze direction corresponds to a (controlled) constant neural activation profile. In contrast to and beyond such models, conceptor auto-adaptation
organized along a manifold M leads not to a continuum of constant neural activity profiles, but explains how a continuum of dynamical patterns connected by
continuous morphs can be generated and controlled.
In sum, the first steps toward an analysis of autoconceptor adaptation have
revealed that this adaptation dynamics is more involved than either the classical
fixed-point dynamics in autoassociative memories or the plane attractor models
suggested in computational neuroscience. For small numbers of stored patterns,
the picture bears some analogies with autoassociative memories in that stable fixed
points of the autonomous adaptation correspond to stored patterns. For larger
numbers of stored patterns (class learning), the plane attractor metaphor captures
essential aspects of phenomena seen in simulations of not too long duration.
3.15
Toward Biologically Plausible Neural Circuits: Random Feature Conceptors
The autoconceptive update equations
z(n + 1) = C(n) tanh(W z(n) + Dz(n) + b)
C(n + 1) = C(n) + λ (z(n) − C(n)z(n)) z 0 (n) − α−2 C(n)
could hardly be realized in biological neural systems. One problem is that the
C update needs to evaluate C(n)z(n), but z(n) is not an input to C(n) in the
z update but the outcome of applying C. The input to C is instead the state
r(n) = tanh(W z(n) + Dz(n) + b). In order to have both computations carried out
by the same C, it seems that biologically hardly feasible schemes of installing two
weight-sharing copies of C would be required. Another problem is that the update
of C is nonlocal: the information needed for updating a “synapse” Cij (that is, an
element of C) is not entirely contained in the presynaptic or postsynaptic signals
available at this synapse.
Here I propose an architecture which solves these problems, and which I think
has a natural biological “feel”. The basic idea is to (i) randomly expand the
reservoir state r into a (much) higher-dimensional random feature space, (ii) carry
out the conceptor operations in that random feature space, but in a simplified
version that only uses scalar operations on individual state components, and (iii)
project the conceptor-modulated high-dimensional feature space state back to the
reservoir by another random projection. The reservoir-conceptor loop is replaced
by a two-stage loop, which first leads from the reservoir to the feature space
(through connection weight vectors fi , collected column-wise in a random neural
projection matrix F ), and then back to the reservoir through a likewise random
104
set of backprojection weights G (Figure 36 A). The reservoir-internal connection
weights W are replaced by the combination of F and G, and the original reservoir
state x known from the basic matrix conceptor framework is split into a reservoir
state vector r and a feature space state vector z with components zi = ci fi0 r. The
conception weights ci take over the role of conceptors. In full detail,
1. expand the N -dimensional reservoir state r = tanh(W z + W in p + b) into
the M -dimensional random feature space by a random feature map F 0 =
(f1 , . . . , fM )0 (a synaptic connection weight matrix of size M × N ) by computing the M -dimensional feature vector F 0 r,
2. multiply each of the M feature projections fi0 r with an adaptive conception
weight ci to get a conceptor-weighted feature state z = diag(c) F 0 r, where
the conception vector c = (c1 , . . . , cM )0 is made of the conception weights,
3. project z back to the reservoir by a random N × M backprojection matrix
G̃, closing the loop.
Since both W and G̃ are random, they can be joined in a single random map
G = W G̃. This leads to the following consolidated state update cycle of a random
feature conception (RFC) architecture:
r(n + 1) = tanh(G z(n) + W in p(n) + b),
z(n + 1) = diag(c(n)) F 0 r(n + 1),
(86)
(87)
where r(n) ∈ RN and z(n), c(n) ∈ RM .
From a biological modeling perspective there exist a number of concrete candidate mechanisms by which the mathematical operation of multiplying-in the
conception weights could conceivably be realized. I will discuss these later and for
the time being remain on this abstract mathematical level of description.
The conception vector c(n) is adapted online and element-wise in a way that
is analog to the adaptation of matrix autoconceptors given in Definition 5. Per
each element ci of c, the adaptation aims at minimizing the objective function
E[(zi − ci zi )2 ] + α−2 c2i ,
(88)
which leads to fixed point solutions satisfying
ci = E[zi2 ](E[zi2 ] + α−2 )−1
(89)
and a stochastic gradient descent online adaptation rule
ci (n + 1) = ci (n) + λi zi2 (n) − ci (n) zi2 (n) − α−2 ci (n) ,
(90)
where i = 1, . . . , M , λi is an adaptation rate, and zi is the i-the component of z.
In computer simulations one will implement this adaptation not element-wise but
in an obvious vectorized fashion.
105
fi
!
G
D
p
u
!
!
A
"c i
W in
r
reservoir
F'
!
z
ci fi
!i fi
"
!!
feature space
B
!!
"
!
!Figure 36: An alternative conceptor architecture aiming at greater biological plau!
sibility. A Schematic of random feature space architecture. The reservoir state
!
r is projected by a feature map F 0 into a higher-dimensional feature space with
states z, from where it is back-projected by G into the reservoir. The conceptor
dynamics is realized by unit-wise multiplying conception weights ci into fi0 r to
obtain the z state. The input unit u is fed by external input p or by learnt input
simulation weights D. B Basic idea (schematic). Black ellipse: reservoir state r
correlation matrix R. Magenta dumbbell: scaling sample points fi from the unit
sphere by their mean squared projection on reservoir states. Green dumbbell:
feature vectors fi scaled by auto-adapted conception weights ci . Red ellipse: the
resulting virtual conceptor CF . For detail see text.
If (90) converges, the converged fixed point is either ci = 0, which always is a
possible and stable solution, or it is of the form
p
(91)
ci = 1/2 + (α2 φi − 4)/4α2 φi ,
which is another possible stable solution provided that α2 φi − 4 > 0. In this
formula, φi denotes the expectation φi = Er [(fi0 r)2 ], the mean energy of the feature
signal fi0 r. These possible values of stable solutions can be derived in a similar
way as was done for the singular values of autoconceptive matrix C in Section
3.14.4, but the derivation is by far simpler (because it can be done element-wise
for each ci and thus entails only scalars, not matrices) and is left as an exercise.
Like the singular values of stable autoconceptors C, the possible stable value range
for conception weights obtainable through (90) is thus {0} ∪ (1/2, 1).
Some geometric properties of random feature conceptors are illustrated in Figure 36 B. The black ellipse represents the state correlation matrix R = E[rr0 ] of a
hypothetical 2-dimensional reservoir. The random feature vectors fi are assumed
to have unit norm in this schematic and therefore sample from the surface of the
unit sphere. The magenta-colored dumbbell-shaped surface represents the weigthing of the random feature vectors fi by the mean energies φi = E[(fi0 r)2 ] of the
feature signals fi0 r. Under the autoconception adaptation they give rise to conception weights ci according to (91) (green dumbbell surface). For values α2 φi −4 < 0
one obtains ci = 0, which shows up in the illustration as the wedge-shaped indentation in the green curve. The red ellipse renders the virtual conceptor CF (see
106
below) which results from the random feature conception weights.
Two properties of this RFC architecture are worth pointing out.
First, conceptor matrices C for an N -dimensional reservoir have N (N + 1)/2
degrees of freedom. If, using conception vectors c instead, one wishes to attain a
performance level of pattern reconstruction accuracy that is comparable to what
can be achieved with conceptor matrices C, one would expect that M should be
in the order of N (N + 1)/2. At any rate, this is an indication that M should be
significantly larger than N . In the simulations below I used N = 100, M = 500,
which worked robustly well. In contrast, trying M = 100 (not documented), while
likewise yielding good accuracies, resulted in systems that were rather sensitive to
parameter settings.
Second, the individual adaptation rates λi can be chosen much larger than the
global adaptation rate λ used for matrix conceptors, without putting stability at
risk. The reason is that the original adpatation rate λ in the stochastic gradient
descent formula for matrix conceptors given in Definition 5 is constrained by the
highest local curvature in the gradient landscape, which leads to slow convergence
in the directions of lower curvature. This is a notorious general characteristic of
multidimensional gradient descent optimization, see for instance [28]. This problem becomes irrelevant for the individual ci updates in (90). In the simulations
presented below, I could safely select the λi as large as 0.5, whereas when I was
using the original conceptor matrix autoadaption rules, λ = 0.01 was often the
fastest rate possible. If adaptive individual adaptation rates λi would be implemented (not explored), very fast convergence of (90) should become feasible.
Geometry of feature-based conceptors. Before I report on simulation experiments, it may be helpful to contrast geometrical properties of the RFC architecture and with the geometry of matrix autoconceptors.
For the sake of discussion, I split the backprojection N × M matrix G in (86)
into a product G = W F where the “virtual” reservoir weight matrix W := GF †
has size N ×N . That is, I consider a system z(n+1) = F diag(c(n)) F 0 tanh(W z(n))
equivalent to (86) and (87), where c is updated according to (90). For the
sake of simplicity I omit input terms and bias in this discussion. The map
F ◦ diag(c) ◦ F 0 : RN → RN then plugs into the place that the conceptor matrix C held in the conceptor systems z(n + 1) = C tanh(W z(n)) discussed in
previous sections. The question I want to explore is how F ◦ diag(c) ◦ F 0 compares
to C in geometrical terms. A conceptor matrix C has an SVD C = U SU 0 , where
U is orthonormal. In order to make the two systems directly comparable, I assume
0
that all feature vectors fi in F have unit norm. Then CF = kF k−2
2 F ◦ diag(c) ◦ F
is positive semidefinite with 2-norm less or equal to 1, in other words it is an N ×N
conceptor matrix.
Now furthermore assume that the adaptation (90) has converged. The adaptation loop (86, 87, 90 ) is then a stationary process and the expectations φi =
Er [(fi0 r)2 ] are well-defined. Note that these expectations can equivalently be writ107
ten as φi = fi0 R fi , where R = E[rr0 ]. According to what I remarked earlier, after
convergence to a stable fixed point solution we have, for all 1 ≤ i ≤ M ,
{0},
if α2 φi − 4 ≤ 0,
p
ci ∈
(92)
{0, 1/2 + (α2 φi − 4)/4α2 φi }, if α2 φi − 4 > 0.
Again for the sake of discussion I restrict my considerations to converged solutions where all ci that can be nonzero (that is, α2 φi − 4 > 0) are indeed nonzero.
It would be desirable to have an analytical result which gives the SVD of the
0
N × N conceptor CF = kF k−2
2 F ◦ diag(c) ◦ F under these assumptions. Unfortunately this analysis appears to be involved and at this point I cannot deliver it. In
order to still obtain some insight into the geometry of CF , I computed a number of
such matrices numerically and compared them to matrix-based autoconceptors C
that were derived from the same assumed stationary reservoir state process. The
outcome is displayed in Figure 37.
Concretely, these numerical investigations were set up as follows. The reservoir dimension was chosen as N = 2 to admit plotting. The number of features
was M = 200. The feature vectors fi were chosen as (cos(i 2 π/M ), sin(i 2 π/M ))0
(where i = 1, . . . , M ), that is, the unit vector (1 0)0 rotated in increments of
(i/M ) 2 π. This choice mirrors a situation where a very large number of fi would
be randomly sampled; this would likewise result in an essentially uniform coverage
of the unit circle. The conception weights ci (and hence CF ) are determined by the
reservoir state correlation matrix R = E[rr0 ]. The same holds for autoconceptor
matrices C. For an exploration of the CF versus C geometries, I thus systematically varied R = U ΣU 0 . The principal directions U were randomly chosen and
remained the same through all variations. The singular values Σ = diag(σ1 σ2 )
were chosen as σ1 ≡ 10, σ2 ∈ {0, 1, 5}, which gave three versions of R. The aperture α was selected in three variants as α ∈ {1, 2, 3}, which altogether resulted in
nine (R, α) combinations.
For each of these combinations, conception weights ci were computed via (92),
from which CF were obtained. Each of these maps the unit circle on an ellipse,
plotted in Figure 37 in red. The values of the ci are represented in the figure as
the dumbbell-shaped curve (green) connecting the vectors ci fi . The wedge-shaped
constriction to zero in some of these curves corresponds to angular values of fi
where ci = 0.
For comparison, for each of the same (R, α) combinations also an autoconceptor matrix C was computed using the results from Section 3.14.4. We saw
on that occasion that nonzero singular values of C are not uniquely determined
by R; they are merely constrained by certain interlacing bounds. To break this
indeterminacy, I selected those C that had the maximal admissible singular values. According
p to (83), this means that the singular values of C were set to
si = 1/2 + (α2 σi − 4)/4α2 σi (where i = 1, 2) provided the root argument was
positive, else si = 0.
Here are the main findings that can be collected from Figure 37:
108
!2 = 0
!2 = 1
aperture = 1
1
!2 = 5
0
!1
aperture = 2
1
0
!1
aperture = 3
1
0
!1
!1
0
1
!1
0
1
!1
0
1
Figure 37: Comparing matrix-based autoconceptors (bold blue ellipses) with
feature-based autoconceptors CF (red ellipses). Broken lines mark principal directions of the reservoir state correlation matrix R. Each panel corresponds
to a particular combination of aperture and the second singular value σ2 of R.
The dumbbell-shaped surfaces (green line) represent the values of the conception
weights ci . For explanation see text.
1. The principal directions of CF , C and R coincide. The fact that CF and R
have the same orientation can also be shown analytically, but the argument
that I have found is (too) involved and not given here. This orientation of
CF hinges on the circumstance that the fi were chosen to uniformly sample
the unit sphere.
2. The CF ellipses are all non-degenerate, that is, CF has no zero singular
values (although many of the ci may be zero as becomes apparent in the
wedge constrictions in the dumbbell-shaped representation of these values).
In particular, the CF are also non-degenerate in cases where the matrix autoconceptors C are (panels in left column and center top panel). The finding
that CF has no zero singular values can be regarded as a disadvantage compared to matrix autoconceptors, because it implies that no signal direction
109
in the N -dimensional reservoir signal space can be completely suppressed
by CF . However, in the M -dimensional feature signal space, we do have
nulled directions. Since the experiments reported below exhibit good stability properties in pattern reconstruction, it appears that this “purging” of
signals in the feature space segment of the complete reservoir-feature loop is
effective enough.
3. Call the ratio of the largest over the smallest singular value of CF or C
the sharpness of a conceptor (also known as eigenvalue spread in the signal
processing literature). Then sometimes CF is sharper than C, and sometimes
the reverse is true. If sharpness is considered a desirable feature of concepors
(which I think it often is), then there is no universal advantage of C over CF
or vice versa.
System initialization and loading patterns: generic description. Returning from this inspection of geometrical properties to the system (86) – (90), I
proceed to describe the initial network creation and pattern loading procedure in
generic terms. Like with the matrix conceptor systems considered earlier in this
report, there are two variants which are the analogs of (i) recomputing the reservoir weight matrix W ∗ , as in Section 3.3, as opposed to (ii) training an additional
input simulation matrix D, as in Section 3.11. In the basic experiments reported
below I found that both work equally well. Here I document the second option.
A readout weight vector W out is likewise computed during loading. Let K target
patterns pj be given (j = 1, . . . , K), which are to be loaded. Here is an outline:
Network creation. A random feature map F , random input weights W in , a random bias vector b, and a random backprojection matrix G∗ are generated. F
and G∗ are suitably scaled such that the combined N × N map G∗ F 0 attains
a prescribed spectral radius. This spectral radius is a crucial system parameter and plays the same role as the spectral radius in reservoir computing in
general (see for instance [108, 112]). All conception weights are initialized
to cji = 1, that is, diag(cj ) = IM ×M .
Conception weight adaptation. The system is driven with each pattern pj in
turn for nadapt steps (discarding an initial washout), while cj is being adapted
per
z j (n + 1) = diag(cj (n)) F 0 tanh(G∗ z j (n) + W in pj (n) + b),
cji (n + 1) = cji (n) + λi zij (n)2 − cji (n) zij (n)2 − α−2 cji (n) ,
leading to conception vectors cj at the end of this period.
State harvesting for computing D and W out , and for recomputing G. The
conception vectors cj obtained from the previous step are kept fixed, and for
each pattern pj the input-driven system rj (n) = tanh(G∗ z j (n) + W in pj (n) +
110
b); z j (n + 1) = diag(cj ) F 0 rj (n) is run for nharvest time steps, collecting states
rj (n) and z j (n).
Computing weights. The N × M input simulation matrix D is computed by
solving the regularized linear regression
X
2
kW in pj (n) − D̃z j (n − 1)k2 + βD
kD̃k2fro
(93)
D = argminD̃
n,j
where βD is a suitably chosen Tychonov regularizer. This means that the autonomous system update z j (n+1) = diag(cj ) F 0 tanh(G∗ z j (n)+W in D z j (n)+
b) should be able to simulate input-driven updates z j (n) =
diag(cj ) F 0 tanh(G∗ z j (n) + W in pj (n) + b). W out is similarly computed by
solving
X
2
2
kpj (n) − W̃ rj (n)k2 + βW
W out = argminW̃
out kW̃ k .
n,j
Optionally one may also recompute G∗ by solving the trivial regularized
linear regression
X
kG∗ z j (n) − G̃z j (n)k2 + βG2 kG̃k2fro .
G = argminG̃
n,j
for a suitably chosen Tychonov regularizer βG . While G∗ and G should
behave virtually identically on the training inputs, the average absolute size
of entries in G will be (typically much) smaller than the original weights in
G∗ as a result of the regularization. Such regularized auto-adaptations have
been found to be beneficial in pattern-generating recurrent neural networks
[92], and in the experiments to be reported presently I took advantage of
this scheme.
The feature vectors fi that make up F can optionally be normalized such that
they all have the same norm. In my experiments this was not found to have a
noticeable effect.
If a stored pattern pj is to be retrieved, the only item that needs to be changed
is the conception vector cj . This vector can either be obtained by re-activating
that cj which was adapted during the loading (which implies that it needs to
be stored in some way). Alternatively, it can be obtained by autoadaptation
without being previously stored, as in Sections 3.14.1 – 3.14.4. I now describe
two simulation studies which demonstrate how this scheme functions (simulation
detail documented in Section 4.9). The first study uses stored conception vectors,
the second demonstrates autoconceptive adaptation.
111
Figure 38: Using stored random feature coded conceptors in a replication of the
basic pattern retrieval experiment from Section 3.4, with M = 500 random feature
vectors fi . First column: sorted conception vectors cj . Second column: spectra
of virtual conceptors CF . Third column: reconstructed patterns (bold light gray)
and original patterns (thin black) after phase alignment. NRMSEs are given in
insets. Last column: The adaptation of cj during the 2000 step runs carried out
in parallel to the loading process. 50 of 500 traces are shown. For explanation see
text.
Example 1: pattern retrieval with stored conception vectors cj . This
simulation re-used the N = 100 reservoir from Sections 3.2 ff. and the four driver
patterns (two irrational-period sines, two very similar 5-periodic random patterns).
The results are displayed in Figure 38.
The loading procedure followed the generic scheme described above (details
2
= 0.01 and
in Section 4), with nadapt = 2000, nharvest = 400, λi = 0.5, βG2 = βD
2
βW out = 1. The aperture was set to α = 8. The left column in Figure 38 shows
the resulting cj spectra, and the right column shows the evolution of cj during this
adaptation. Notice that a considerable portion of the conception weights evolved
toward zero, and that none ended in the range (0 1/2), in agreement with theory.
For additional insight into the dynamics of this system I also computed “virtual” matrix conceptors CFj by Rj = E[(rj )0 rj ], CFj = Rj (Rj + α−2 )−1 (second
column). The singular value spectrum of CFj reveals that the autocorrelation spectra of rj signals in RFC systems is almost identical to the singular value spectra
obtained with matrix conceptors on earlier occasions (compare Figure 14).
The settings of matrix scalings and aperture were quite robust; variations in a
range of about ±50% about the chosen values preserved stability and accuracy of
pattern recall (detail in Section 4.9).
For testing the recall of pattern pj , the loaded system was run using the update
112
routine
rj (n) = tanh(G z j (n) + W in D z j (n) + b),
y j (n) = W out rj (n),
z j (n + 1) = diag(cj ) F 0 rj (n),
starting from a random starting state z j (0) which was sampled from the normal
distribution, scaled by 1/2. After a washout of 200 steps, the reconstructed pattern
y j was recorded for 500 steps and compared to a 20-step segment of the target
pattern pj . The second column in Figure 38 shows an overlay of y j with pj and
gives the NRMSEs. The reconstruction is of a similar quality as was found in
Section 3.4 where full conceptor matrices C were used.
Example 2: content-addressed pattern retrieval. For a demonstration of
content-addressed recall similar to the studies reported in Section 3.14.3, I reused the M = 500 system described above. Reservoir scaling parameters and the
loading procedure were identical except that conception vectors cj were not stored.
Results are collected in Figure 39. The cue and recall procedure for a pattern pj
was carried out as follows:
1. Starting from a random reservoir state, the loaded reservoir was driven with
the cue pattern for a washout time of 200 steps by z j (n+1) = F 0 tanh(G z j (n)+
W in pj (n) + b).
2. Then, for a cue period of 800 steps, the system was updated with cj adaptation by
z j (n + 1) = diag(cj (n)) F 0 tanh(G z j (n) + W in pj (n) + b),
cji (n + 1) = cji (n) + λi zij (n)2 − cji (n) zij (n)2 − α−2 cji (n) , (1 ≤ i ≤ M )
starting from an all-ones cj , with an adaptation rate λi = 0.5 for all i. At
the end of this period, a conception vector cj,cue was obtained.
3. To measure the quality of cj,cue , a separate run of 500 steps without c adapation was done using
rj (n) = tanh(G z j (n) + W in D z j (n) + b),
y j (n) = W out rj (n),
z j (n + 1) = diag(cj,cue ) F 0 rj (n)
obtaining a pattern reconstruction y j (n). This was phase-aligned with the
original pattern pj and an NRMSE was computed (Figure 39, third column).
113
4. The recall run was resumed after the cueing period and continued for another
10,000 steps in auto-adaptation mode, using
z j (n + 1) = diag(cj (n)) F 0 tanh(G z j (n) + W in D z j (n) + b),
cji (n + 1) = cji (n) + λi zij (n)2 − cji (n) zij (n)2 − α−2 cji (n) , (1 ≤ i ≤ K)
leading to a final cj,adapted at the end of this period.
5. Another quality measurement run was done identical to the post-cue measurement run, using cj,adapted (NRMSE results in Figure 39, third column).
Figure 39: Content-addressed recall using RFC conceptors with M = 500 feature vectors fi . First column: sorted feature projection weight vectors cj after
the cue phase (black) and after 10,000 steps of autoadaptation (gray). Second
column: spectra of virtual conceptors CF after cue (black) and at the end of autonomous adaptation (gray). Both spectra are almost identical. Third column:
reconstructed patterns (bold light gray: after cue, bold dark gray: after autoadaptation; the latter are mostly covered by the former) and original patterns (thin
black). NRMSEs are given in insets (top: after cue, bottom: after autoadaptation). Fourth column: The adaptation of cj during the cueing period. Last column:
same, during the 10000 autoadaptation steps. 50 of 500 traces are shown. Note
the different timescales in column 4 versus column 5. For explanation see text.
Like in the matrix-C-based content-addressing experiments from Section 3.14.3,
the recall quality directly after the cue further improved during the autoconceptive adaption afterwards, except for the first pattern. Pending a more detailed
investigation, this may be attributed to the “maturation” of the cj during autoadaption which reveals itself in the convergence of a number of cji to zero during
autoadaptation (first and last column in Figure 39). We have seen similar effects
in Section 3.14.3.
114
An obvious difference to those earlier experiments is that the cueing period is
much longer now (800 versus 15 – 30 steps). This is owed to the circumstance that
now the conceptor adaptation during cueing started from an all-ones cj , whereas
in Section 3.14.3 it was started from a zero C. In the latter case, singular values
of C had to grow away from zero toward one during cueing, whereas here they
had to sink away from one toward zero. The effects of this mirror situation are
not symmetrical. In an “immature” post-cue conceptor matrix C started from
a zero C, all the singular values which eventually should converge to zero are
already at zero at start time and remain there. Conversely, the post-cue feature
projection weights cj,cue , which should eventually become zero, have not come
close to this destination even after the 800 cue steps that were allotted here (left
panels in Figure 39). This tail of “immature” nonzero elements in cj,cue leads to
an insufficient filtering-out of reservoir state components which do not belong to
the target pattern dynamics.
The development of the cji during the autonomous post-cue adapation is not
monotonous (right panels). Some of these weights meander for a while before they
settle to what appear stable final values. This is due to the transient nonlinear
reservoir–cj interactions which remain to be mathematically analyzed.
A potentially important advantage of using random feature conceptors c rather
than matrix conceptors C in machine learning applications is the faster convergence of the former in online adaptation scenarios. While an dedicated comparison
of convergence properties between c and C conceptors remains to be done, one
may naturally expect that stochastic gradient descent works more efficiently for
random feature conceptors than for matrix conceptors, because the gradient can
be followed individually for each coordinate ci , unencumbered by the second-order
curvature interactions which notoriously slow down simple gradient descent in
multidimensional systems. This is one of the reasons why in the complex hierarchical signal filtering architecture to be presented below in Section 3.16 I opted
for random feature conceptors.
Algebraic and logical rules for conception weights. The various definitions
and rules for aperture adaptation, Boolean operations, and abstraction introduced
previously for matrix conceptors directly carry over to random feature conceptor.
The new definitions and rules are simpler than for conceptor matrices because they
all apply to the individual, scalar conception weights. I present these items without
detailed derivations (easy exercises). In the following, let c = (c1 , . . . , cM )0 , b =
(b1 , . . . , bM )0 be two conception weight vectors.
Aperture adaptation (compare Definition 3) becomes
115
Definition 6
ϕ(ci , γ) := ci /(ci + γ −2 (1 − ci )) for 0 < γ < ∞,
0 if ci < 1,
ϕ(ci , 0) :=
1 if ci = 1,
1 if ci > 0,
ϕ(ci , ∞) :=
0 if ci = 0.
Transferring the matrix-based definition of Boolean operations (Definition 4)
to conception weight vectors leads to the following laws:
Definition 7
¬ ci := 1 − ci ,
ci bi /(ci + bi − ci bi ) if not ci = bi = 0,
ci ∧ bi :=
0
if
ci = bi = 0,
(ci + bi − 2ci bi )/(1 − ci bi ) if not ci = bi = 1,
ci ∨ bi :=
1
if
ci = bi = 1.
The matrix-conceptor properties connecting aperture adaptation with Boolen
operations (Proposition 10) and the logic laws (Proposition 11) remain valid after
the obvious modifications of notation.
We define c ≤ b if for all i = 1, . . . , M it holds that ci ≤ bi . The main elements
of Proposition 13 turn into
Proposition 16 Let a = (a1 , . . . , aM )0 , b = (b1 , . . . , bM )0 be conception weight
vectors. Then the following facts hold.
1. If b ≤ a, then b = a ∧ c, where c is the conception weight vector with entries
0
if bi = 0,
ci =
−1
−1
−1
(bi − ai + 1)
if bi > 0.
2. If a ≤ b, then b = a ∨ c, where c is the conception weight vector with entries
1
if bi = 1,
ci =
−1
−1
−1
1 − ((1 − bi ) − (1 − ai ) + 1)
if bi < 1.
3. If a ∧ c = b, then b ≤ a.
4. If a ∨ c = b, then a ≤ b.
Note that all of these definitions and rules can be considered as restrictions
of the matrix conceptor items on the special case of diagonal conceptor matrices.
The diagonal elements of such diagonal conceptor matrices can be identified with
conception weights.
116
Aspects of biological plausibility. “Biological plausibility” is a vague term
inviting abuse. Theoretical neuroscientists develop mathematical or computational models of neural systems which range from fine-grained compartment models of single neurons to abstract flowchart models of cognitive processes. Assessing
the methodological role of formal models in neuroscience is a complex and sometimes controversial issue [1, 36]. When I speak of biological plausibility in connection with conceptor models, I do not claim to offer a blueprint that can be directly
mapped to biological systems. All that I want to achieve in this section is to show
that conceptor systems may be conceived which do not have characteristics that
are decidedly not biologically feasible. In particular, (all) I wanted is a conceptor
system variant which can be implemented without state memorizing or weight
copying, and which only needs locally available information for its computational
operations. In the remainder of this section I explain how these design goals may
be satisfied by RFC conceptor architectures.
I will discuss only the adaptation of conception weights and the learning of
the input simulation weights D, leaving cueing mechanisms aside. The latter
was implemented in the second examples above in an ad-hoc way just to set the
stage and would require a separate treatment under the premises of biological
plausibility.
In my discussion I will continue to use the discrete-time update dynamics that
was used throughout this report. Biological systems are not updated according
to a globally clocked cycle, so this clearly departs from biology. Yet, even a
synchronous-update discrete-time model can offer relevant insight. The critical
issues that I want to illuminate – namely, no state/weight copying and locality of
computations – are independent of choosing a discrete or continuous time setting.
I first consider the adaptation of the input simulation weights D. In situated,
life-long learning systems one may assume that at given point in (life-)time, some
version of D is already present and active, reflecting the system’s learning history
up to that point. In the content-addressable memory example above, at the end of
the cueing period there was an abrupt switch from driving the system with external
input to an autonomous dynamics using the system’s own input simulation via
D. Such binary instantaneous switching can hardly be expected from biological
systems. It seems more adequate to consider a gradual blending between the
input-driven and the autonomous input simulation mode, as per
z j (n + 1) =
(94)
j
0
j
in
j
= diag(c (n)) F tanh G z (n) + W
τ (n) D z (n) + (1 − τ (n)) p(n) + b ,
where a mixing between the two modes is mediated by a “slide ruler” parameter
τ which may range between 0 and 1 (a blending of this kind will be used in the
architecture in Section 3.16).
As a side remark, I mention that when one considers comprehensive neural
architectures, the question of negotiating between an input-driven and an autonomous processing mode arises quite generically. A point in case are “Bayesian
117
brain” models of pattern recognition and control, which currently receive much
attention [33, 15]. In those models, a neural processing layer is driven both from
“lower” (input-related) layers and from “higher” layers which autonomously generate predictions. Both influences are merged in the target layer by some neural
implementation of Bayes’ rule. Other approaches that I would like to point out in
this context are layered restricted Boltzmann machines [47], which likewise can be
regarded as a neural implementation of Bayes’ rule; hierarchical neural field models of object recognition [113] which are based on Arathorn’s “map seeking circuit”
model of combining bottom-up and top-down inputs to a neural processing layer
[35]; and mixture of experts models for motor control (for example [109]) where a
“responsibility” signal comparable in its function to the τ parameter negotiates a
blending of different control signals.
zi from feature space
D
e
from external
_
d
+
!"
d'
1
1
u
!'
p
W in
to reservoir
1
p'
1
Figure 40: An abstract circuit which would implement the τ negotiation between
driving a reservoir with external input p versus with simulated input Dz. Abstract
neurons are marked by filled gray circles. Connections that solely copy a neural
state forward are marked with “1”. Connections marked −• refer to multiplicative
modulation. Connections that are inhibitory by their nature are represented by
a. Broken arrows indicate a controlling influence on the weight adaptation of D.
For explanation see text.
Returning to (94), the mathematical formula could be implemented in an abstract neural circuit as drawn in Figure 40. Explanation of this diagram: gray dots
represent abstract firing-rate neurons (biologically realized by individual neurons
or collectives). All neurons are linear. Activation of neuron d: simulated input
d(n) = Dz(n); of p: external driver p(n). Neuron e maintains the value of the
“error” p(n) − d(n). d and p project their activation values to d0 and p0 , whose
activations are multiplicatively modulated by the activations of neurons τ and τ 0 .
The latter maintain the values of τ and τ 0 = 1 − τ from (94). The activations
d0 (n) = τ (n) Dz(n) and p0 (n) = (1 − τ (n)) p(n) are additively combined in u,
which finally feeds to the reservoir through W in .
For a multiplicative modulation of neuronal activity a number of biological
118
mechanisms have been proposed, for example [94, 13]. The abstract model given
here is not committed to a specific such mechanism. Likewise I do not further
specify the biological mechanism which balances between τ and τ 0 , maintaining a
relationship τ 0 = 1 − τ ; it seems natural to see this as a suitable version of mutual
inhibition.
An in-depth discussion by which mechanisms and for which purposes τ is administered is beyond the scope of this report. Many scenarios are conceivable. For
the specific purpose of content-addressable memory recall, the setting considered
in this section, a natural option to regulate τ would be to identify it with the
(0-1-normalized and time-averaged) error signal e. In the architecture presented
in Section 3.16 below, regulating τ assumes a key role and will be guided by novel
principles.
The sole point that I want to make is that this abstract architecture (or similar
ones) requires only local information for the adaptation/learning of D. Consider
a synaptic connection Di from a feature neuron zi to d. The learning objective
(93) can be achieved, for instance, by the stochastic gradient descent mechanism
Di (n + 1) = Di (n) + λ (p(n) − Di (n)zi (n)) zi (n) − α−2 Di (n) ,
(95)
where the “error” p(n) − Di (n)zi (n) is available in the activity of the e neuron.
The learning rate λ could be fixed, but a more suggestive option would be to
scale it by τ 0 (n) = 1 − τ (n), as indicated in the diagram. That is, Di would be
adapted with an efficacy proportional to the degree that the system is currently
being externally driven.
I now turn to the action and the adaptation of the conception weights, stated in
mathematical terms in equations (87) and (90). There are a number of possibilities
to implement these formulae in a model expressed on the level of abstract firingrate neurons. I inspect three of them. They are sketched in Figure 41.
!i
Di
A
!i
Di
fi
B
ci
!i
"i
Di
fi
C
!i
1
1 "
i
fi
Figure 41: Three candidate neural implementations of conception weight mechanisms. In each diagram, ϕi is an abstract neuron whose activation is ϕ(n) =
zi (n) = ci (n) fi0 r(n). In B and C, ζi has activation fi0 r(n). In C, γi has activation
ci . For explanation see text.
The simplest model (Figure 41 A) represents the quantity zi (n) = ci (n) fi0 r(n)
by the activation of a single neuron ϕi . It receives synaptic input fi0 r(n) through
connections fi and feeds to the reservoir (or to an input gating circuit as discussed
above) through the single synaptic connection Di . The weighting of fi0 r(n) with
119
the factor ci is effected by some self-regulated modulation of synaptic gain. Taking
into account that ci changes on a slower timescale than fi0 r(n), the information
needed to adapt the strength ci of this modulation (90) is a moving average of the
neuron’s own activation energy zi2 (n) and the current synaptic gain ci (n), which
are characteristics of the neuron ϕi itself and thus are trivially locally available.
In the next model (Figure 41 B), there is a division of labor between a neuron
ϕi which again represents ci (n) fi0 r(n) and a preceding neuron ζi which represents
fi0 r(n). The latter feeds into the former through a single synaptic connection
weighted by ci . The adaptation of the synaptic strength ci here is based on the
(locally time-averaged) squared activity of the postsynaptic neuron ϕi , which again
is information locally available at the synaptic link ci .
Finally, the most involved circuit offered in Figure 41 C delegates the representation of ci to a separate neuron γi . Like in the second model, a neuron ζi
which represents fi0 r(n) feeds to ϕi , this time copying its own activation through
a unit connection. The γi neuron multiplicatively modulates ϕi by its activation
ci . Like in the D adaptation proposal described in Figure 40, I do not commit to
a specific biological mechanism for such a multiplicative modulation. The information needed to adapt the activation ci of neuron γi according to (90) is, besides
ci itself, the quantity zi = ci (n) fi0 r(n). The latter is represented in ϕi which is
postsynaptic from the perspective of γi and therefore not directly accessible. However, the input fi0 r(n) from neuron ζi is available at γi , from which the quantity
zi = ci fi0 r(n) can be inferred by neuron γi . The neuron γi thus needs to instantiate an intricate activation dynamics which combines local temporal averaging
of (fi0 r(n))2 with an execution of (90). A potential benefit of this third neural
circuit over the preceding two is that a representation of ci by a neural activation can presumably be biologically adapted on a faster timescale than the neuron
auto-modulation in system A or the synaptic strength adaptation in B.
When I first considered content-addressable memories in this report (Section
3.14.3), an important motivation for doing so was that storing entire conceptor
matrices C for later use in retrieval is hardly an option for biological systems. This
may be different for conception vectors: it indeed becomes possible to “store” conceptors without having to store network-sized objects. Staying with the notation
used in Figure 41: a single neuron γ j might suffice to represent and “store” a
conception vector cj associated with a pattern pj . The neuron γ j would project to
all ϕi neurons whose states correspond to the signals zi , with synaptic connection
weights cji , and effecting a multiplicative modulation of the activation of the ϕi
neurons proportional to these connection weights. I am not in a position to judge
whether this is really an option in natural brains. For applications in machine
learning however, using stored conception vectors cj in conjunction with RFC systems may be a relevant alternative to using stored matrix conceptors, because
vectors cj can be stored much more cheaply in computer systems than matrices.
A speculative outlook. I allow myself to indulge in a brief speculation of how
RFC conceptor systems might come to the surface – literally – in mammalian
120
brains. The idea is to interpret the activations of (groups of) neurons in the
neocortical sheet as representing conception factors ci or zi values, in one of the
versions shown in Figure 41 or some other concrete realization of RFC conceptors.
The “reservoir” part of RFC systems might be found in deeper brain structures.
When some patches of the neocortical sheet are activated and others not (revealed
for instance through fMRI imaging or electro-sensitive dyes), this may then be interpreted as a specific cj vector being active. In geometrical terms, the surface of
the hyperellipsoid of the “virtual” conceptor would be mapped to the neocortical
sheet. Since this implies a reduction of dimension from a hypothetical reservoir
dimension N to the 2-dimensional cortical surface, a dimension folding as in selforganizing feature maps [58, 79, 34] would be necessary. What a cognitive scientist
would call an “activation of a concept” would find its neural expression in such an
activation of a dimensionally folded ellipsoid pertaining to a “virtual” conceptor
CF in the cortical sheet. An intriguing further step down speculation road is to
think about Boolean operations on concepts as being neurally realized through
the conceptor operations described in Sections 3.9 – 3.13. All of this is still too
vague. Still, some aspects of this picture have already been explored in some detail in other contexts. Specifically, the series of neural models for processing serial
cognitive tasks in primate brains developed by Dominey et al. [19, 20] combine
a reservoir dynamics located in striatal nuclei with cortical context-providing activation patterns which shares some characteristics with the speculations offered
here.
3.16
A Hierarchical Filtering and Classification Architecture
A reservoir equipped with some conceptor mechanism does not by itself serve
a purpose. If this computational-dynamical principle is to be made useful for
practical purposes like prediction, classification, control or others, or if it is to
be used in cognitive systems modeling, conceptor-reservoir modules need to be
integrated into more comprehensive architectures. These architectures take care of
which data are fed to the conceptor modules, where their output is channelled, how
apertures are adapted, and everything else that is needed to manage a conceptor
module for the overall system purpose. In this section I present a particular
architecture for the purpose of combined signal denoising and classification as an
example. This (still simple) example introduces a number of features which may
be of more general use when conceptor modules are integrated into architectures:
Arranging conceptor systems in bidirectional hierarchies: a higher conceptor module is fed from a lower one by the output of the latter (bottom-up
data flow), while at the same time the higher module co-determines the conceptors associated with the lower one (top-down “conceptional bias” control).
Neural instantiations of individual conceptors: Using random feature con121
ceptors, it becomes possible to economically store and address individual
conceptors.
Self-regulating balance between perception and action modes: a conceptorreservoir module is made to run in any mixture of two fundamental processing modes, (i) being passively driven by external input and (ii) actively generating an output pattern. The balance between these modes is autonomously
steered by a criterion that arises naturally in hierarchical conceptor systems.
A personal remark: the first and last of these three items constituted the
original research questions which ultimately guided me to conceptors.
The task. The input to the system is a timeseries made of alternating sections
of the four patterns p1 , . . . , p4 used variously before in this report: two sines of
irrational period lengths, and two slightly differing 5-periodic patterns. This signal
is corrupted by strong Gaussian i.i.d. noise (signal-to-noise ratio = 0.5) – see
bottom panels in Fig. 43. The task is to classify which of the four patterns is
currently active in the input stream, and generate a clean version of it. This
task is a simple instance of the generic task “classify and clean a signal that
intermittently comes from different, but familiar, sources”.
c1, c2, c3, c4
(&.)2'1".%&#"1''
,)**"!.'
y[3]
u[3]
![23]
y[2]
3%!"'1".%&#"1''
,)**"!.'
u[2]
![12]
y[1]
&.&*&)22/'1".%&#"1''
,)**"!.'
u[1]
r[3]
Win
Wout
F'
0%1+2"'
Dz[3]
G
D
![23]
'''![3]'
c[2]
F'
Win
0%1+2"'
Dz[2]
G
D
y[1]
z[3]
Wout c[2]aut
r[2]
y[2]
c[3]
r[1]
Win
z[2]
Wout c aut
[1]
![12]
'''![2]'
c[1]
F'
0%1+2"'
p
G
.%&#/'&.,+*'
!"#"!$%&!'
z[1]
''''![1]'
(")*+!"'#,)-"'
Figure 42: Schematic of 3-layer architecture for signal filtering and classification.
For explanation see text.
122
Architecture. The basic idea is to stack copies of a reservoir-conceptor loop,
giving a hierarchy of such modules (compare Figure 42). Here I present an example
with three layers, having essentially identical modules M[1] on the lowest, M[2]
on the middle, and M[3] on the highest layer (I use subscript square brackets [l]
to denote levels in the hierarchy).
Each module is a reservoir-conceptor loop. The conceptor is implemented here
through the M -dimensional feature space expansion described in the previous
section, where a high-dimensional conception weight vector c is multiplied into the
feature state (as in Figure 36). At exploitation time the state update equations
are
u[l] (n + 1)
r[l] (n + 1)
z[l] (n + 1)
y[l] (n + 1)
=
=
=
=
(1 − τ[l−1,l] (n)) y[l−1] (n + 1) + τ[l−1,l] (n) D z[l] (n),
tanh(G z[l] (n) + W in u[l] (n + 1) + b),
c[l] (n) .∗ F 0 r[l] (n + 1),
W out r[l] (n + 1),
where u[l] is the effective signal input to module M[l] , y[l] is the output signal from
that module, and the τ are mixing parameters which play a crucial role here and
will be detailed later. In addition to these fast timescale state updates there are
several online adaptation processes, to be described later, which adapt τ ’s and c’s
on slower timescales. The weight matrices D, G, F, W in , W out are identical in all
modules. F, W in are created randomly at design time and remain unchanged. D
and W out are trained on samples of clean “prototype” patterns in an initial pattern
loading procedure. G is first created randomly as G∗ and then is regularized using
white noise (all detail in Section 4.10).
The effective input signal u[l] (n + 1) to M[l] is thus a mixture mediated by
a “trust” variable τ[l−1,l] of a module-external external input y[l−1] (n + 1) and a
module-internal input simulation signal D z[l] (n). On the bottom layer, τ[01] ≡ 0
and y[0] (n) = p(n), that is, this layer has no self-feedback input simulation and
is entirely driven by the external input signal p(n). Higher modules receive the
output y[l−1] (n + 1) of the respective lower module as their external input. Both
input mix components y[l−1] and D z[l] represent partially denoised versions of the
external pattern input. The component y[l−1] from the module below will typically
be noisier than the component Dz[l] that is cycled back within the module, because
each module is supposed to de-noise the signal further in its internal reservoirconceptor feedback loop. If τ[l−1,l] were to be 1, the module would be running in
an autonomous pattern generation mode and would be expected to re-generate a
very clean version of a stored pattern – which might however be a wrong one. If
τ[l−1,l] were to be 0, the module would be running in an entirely externally driven
mode, with no “cleaning” in effect. It is crucial for the success of this system that
these mixing weights τ[l−1,l] are appropriately set. They reflect a “trust” of the
system in its current hypothesis about the type of the driving pattern p, hence I
call them trust variables. I mention at this point that when the external input p
changes from one pattern type to another, the trust variables must quickly decrease
123
in order to temporarily admit the architecture to be in an altogether more inputdriven, and less self-generating, mode. All in all this constitutes a bottom-up flow
of information, whereby the raw input p is cleaned stage-wise with an amount of
cleaning determined by the current trust of the system that it is applying the right
conceptor to effect the cleaning.
The output signals y[l] of the three modules are computed from the reservoir
states r[l] by output weights W out , which are the same on all layers. These output
weights are initially trained in the standard supervised way of reservoir computing
to recover the input signal given to the reservoir from the reservoir state. The 3-rd
layer output y[3] also is the ultimate output of the entire architecture and should
give a largely denoised version of the external driver p.
Besides this bottom-up flow of information there is a top-down flow of information. This top-down pathway affects the conception weight vectors c[l] which
are applied in each module. The guiding idea here is that on the highest layer
(l = 3 in our example), the conceptor c[3] is of the form
_
c[3] (n) =
ϕ(cj , γ j (n)),
(96)
j=1,...,4
where c1 , . . . , c4 are prototype conception weight vectors corresponding to the four
training patterns. These prototype vectors are computed and stored at training
time. In words, at the highest layer the conception weight vector is constrained to
be a disjunction of aperture-adapted versions of the prototype conception weight
vectors. Imposing this constraint on the highest layer can be regarded as inserting
a qualitative bias in the ensuing classification and denoising process. Adapting
c[3] (n) amounts to adjusting the aperture adaptation factors γ j (n).
At any time during processing external input, the current composition of c[3] as
j
a γ -weighted disjunction of the four prototypes reflects the system’s current hypothesis about the type of the current external input. This hypothesis is stagewise
passed downwards through the lower layers, again mediated by the trust variables.
This top-down pathway is realized as follows. Assume that c[3] (n) has been
computed. In each of the two modules below (l = 1, 2), an (auto-)conception
weight vector caut
[l] (n) is computed by a module-internal execution of the standard
autoconception adaptation described in the previous section (Equation (90)). To
arrive at the effective conception weight vector c[l] (n), this caut
[l] (n) is then blended
with the current conception weight vector c[l+1] (n) from the next higher layer,
again using the respective trust variable as mixing coefficient:
c[l] (n) = (1 − τ[l,l+1] (n)) caut
[l] (n) + τ[l,l+1] (n) c[l+1] (n).
(97)
In the demo task reported here, the raw input p comes from either of four
sources p1 , . . . , p4 (the familiar two sines and 5-periodic patterns), with additive
noise. These four patterns are initially stored in each of the modules M[l] by
training input simulation weights D, as described in Section 3.15. The same D is
used in all layers.
124
In the exploitation phase (after patterns have been loaded into D, output
weights have been learnt, and prototype conceptors cj have been learnt), the architecture is driven with a long input sequence composed of intermittent periods
where the current input is chosen from the patterns pj in turn. While being driven
with pj , the system must autonomously determine which of the four stored pattern is currently driving it, assign trusts to this judgement, and accordingly tune
the degree of how strongly the overall processing mode is autonomously generative
(high degree of cleaning, high danger of “hallucinating”) vs. passively input-driven
(weak cleaning, reliable coupling to external driver).
Summing up, the overall functioning of the trained architecture is governed by
two pathways of information flow,
• a bottom-up pathway where the external noisy input p is successively denoised,
• a top-down pathway where hypotheses about the current pattern type, expressed in terms of conception weight vectors, are passed downwards,
and by two online adaptation processes,
• adjusting the trust variables τ[l−1,l] , and
• adjusting the conception weight vector c[l] in the top module.
I now describe the two adaptation processes in more detail.
Adapting the trust variables. Before I enter technicalities I want to emphasize
that here we are confronted with a fundamental problem of information processing
in situated intelligent agents (“SIA”: animals, humans, robots). A SIA continuously has to “make sense” of incoming sensor data, by matching them to the
agent’s learnt/stored concepts. This is a multi-faceted task, which appears in
many different instantiations which likely require specialized processing strategies. Examples include online speech understanding, navigation, or visual scene
interpretation. For the sake of this discussion I will lump them all together and
call them “online data interpretation” (ODI) tasks. ODI tasks variously will involve subtasks like de-noising, figure-ground separation, temporal segmentation,
attention control, or novelty detection. The demo architecture described in this
section only addresses de-noising and temporal segmentation. In many cognitive
architectures in the literature, ODI tasks are addressed by maintaining an online
representation of a “current best” interpretation of the input data. This representation is generated in “higher” levels of a processing hierarchy and is used in a
top-down fashion to assist lower levels, for instance by way of providing statistical
bias or predictions (discussion in [15]). This top-down information then tunes the
processing in lower levels in some way that enables them to extract from their
respective bottom-up input specific features while suppressing others – generally
speaking, by making them selective. An inherent problem in such architectures
is that the agent must not grow overly confident in its top-down pre-conditioning
125
of lower processing layers. In the extreme case of relying entirely on the current
interpretation of data (instead of on the input data themselves), the SIA will be
hallucinating. Conversely, the SIA will perform poorly when it relies entirely on
the input data: then it will not “understand” much of it, becoming unfocussed and
overwhelmed by noise. A good example is semantic speech or text understanding.
Linguistic research has suggested that fast forward inferences are involved which
predispose the SIA to interpret next input in terms of a current representation
of semantic context (for instance, [98]). As long as this current interpretation is
appropriate, it enables fast semantic processing of new input; but when it is inappropriate, it sends the SIA on erroneous tracks which linguists call “garden paths”.
Generally speaking, for robust ODI an agent should maintain a reliable measure
of the degree of trust that the SIA has in its current high-level interpretation.
When the trust level is high, the SIA will heavily tune lower levels by higher-level
interpretations (top-down dominance), while when trust levels are low, it should
operate in a bottom-up dominated mode.
Maintaining an adaptive measure of trust is thus a crucial objective for an
SIA. In Bayesian architectures (including Kalman filter observers in control engineering), such a measure of trust is directly provided by the posterior probability
p(interpretation | data). A drawback here is that a number of potentially complex
probability distributions have to be learnt beforehand and may need extensive
training data, especially when prior hyperdistributions have to be learnt instead
of being donated by an oracle. In mixture of predictive expert models (for instance
[109]), competing interpretation models are evaluated online in parallel and are
assigned relative trust levels according to how precisely they can predict the current input. A problem that I see here is computational cost, besides the biological
implausibility of executing numerous predictors in parallel. In adaptive resonance
theory [43], the role of a trust measure is filled by the ratio between the norm of a
top-down pattern interpretation over the norm of an input pattern; the functional
effects of this ratio for further processing depends on whether that ratio is less
or greater than a certain “vigilance” parameter. Adaptive resonance theory however is primarily a static pattern processing architecture not designed for online
processing of temporal data.
Returning to our demo architecture, here is how trust variables are computed.
They are based on auxiliary quantities δ[l] (n) which are computed within each module l. Intuitively, δ[l] (n) measures the (temporally smoothed) discrepancy between
the external input signal fed to the module and the self-generated, conceptorcleaned version of it. For layers l > 1 the external input signal is the bottom-up
passed output y[l−1] of the lower layer. The conceptor-cleaned, module-generated
version is the signal Dz[l] (n) extracted from the conception-weighted feature space
signal z[l] (n) = c[l] (n) .∗ F 0 r[l] (n) by the input simulation weights D, where r[l] (n)
is the reservoir state in layer l. Applying exponential smoothing with smoothing
rate σ < 1, and normalizing by the likewise smoothed variance of y[l−1] , gives
126
update equations
ȳ[l−1] (n + 1) = σ ȳ[l−1] (n) + (1 − σ) y[l−1] (n + 1), (running average) (98)
var y[l−1] (n + 1) =
σ var y[l−1] (n + 1) + (1 − σ) (y[l−1] (n + 1) − ȳ[l−1] (n + 1))2 , (99)
(y[l−1] (n + 1) − Dz[l] (n + 1))2
δ[l] (n + 1) = σ δ[l] (n) + (1 − σ)
(100)
var y[l−1] (n + 1)
for the module-internal detected discrepancies δ[l] . In the bottom module M[1] ,
the same procedure is applied to obtain δ[1] except that the module input is here
the external driver p(n) instead the output y[l−1] (n) from the level below.
From these three discrepancy signals δ[l] (n) two trust variables τ[12] , τ[23] are
derived. The intended semantics of τ[l,l+1] can be stated as “measuring the degree
by which the discrepancy is reduced when going upwards from level l to level l+1”.
The rationale behind this is that when the currently active conception weights in
modules l and l + 1 are appropriate for the current drive entering module l from
below (or from the outside when l = 1), the discrepancy should decrease when
going from level l to level l+1, while if the the currently applied conception weights
are the wrong ones, the discrepancy should increase when going upwards. The
core of measuring trust is thus the difference δ[l] (n) − δ[l+1] (n), or rather (since we
want the same sensitivity across all levels of absolute values of δ) the difference
log(δ[l] (n)) − log(δ[l+1] (n)). Normalizing this to a range of (0, 1) by applying a
logistic sigmoid with steepness d[l,l+1] finally gives
τ[l,l+1] (n) =
1+
δ[l+1] (n)
δ[l] (n)
d[l,l+1] !−1
.
(101)
The steepness d[l,l+1] of the trust sigmoid is an important design parameter, which
currently I set manually. Stated in intuitive terms it determines how “decisively”
the system follows its own trust judgement. It could be rightfully called a “metatrust” variable, and should itself be adaptive. Large values of this decisiveness
leads the system to make fast decisions regarding the type of the current driving
input, at an increased risk of settling down prematurely on a wrong decision.
Low values of d[l,l+1] allow the system to take more time for making a decision,
consolidating information acquired over longer periods of possibly very noisy and
only weakly pattern-characteristic input. My current view on the regulation of
decisiveness is that it cannot be regulated on the sole basis of the information
contained in input data, but reflects higher cognitive capacities (connected to
mental attitudes like “doubt”, “confidence”, or even “stubbornness”...) which are
intrinsically not entirely data-dependent.
Adapting the top-level conception weight vectors c[l] . For clarity of notation
I will omit the level index [l] in what follows, assuming throughout l = 3. By
equation (96), the effective conception weight vector used in the top module will
127
W
be constrained to be a disjunction c(n) = j=1,...,4 ϕ(cj , γ j (n)), where the cj are
prototype conception weight vectors, computed at training time. Adapting c(n)
amounts to adjusting the apertures of the disjunctive components cj via γ j (n).
This is done indirectly.
The training of the prototype conception weights (and of the input simulation
matrix D and of the readout weights W out ) is done with a single module that is
driven by the clean patterns pj . Details of the training procedure are given in the
Experiments and Methods Section 4.10. The prototype conception weight vectors
can be written as
cj = E[(z j ).∧2 ] .∗ (E[(z j ).∧2 ] + α−2 ).∧−1 ,
where z j (n) = cj .∗ F 0 rj (n) is the M -dimensional signal fed back from the feature
space to the reservoir while the module is being driven with pattern j during
training, and the aperture α is a design parameter. Technically, we do not actually
store the cj but their constituents α and the corresponding mean signal energy
vectors E[(z j ).∧2 ], the latter of which are collected in an M × 4 prototype matrix
P = (E[(zij )2 ])i=1,...,M ; j=1,...,4 .
(102)
I return to the conceptor adaptation dynamics in the top module at exploitation
time. Using results from previous sections, equation (96) can be re-written as
!.∧−1
!
c(n) =
X
(γ j (n)).∧2 .∗ E[(z j ).∧2 ]
.∗
X
(γ j (n)).∧2 .∗ E[(z j ).∧2 ] + α−2
j
,
j
(103)
where the +α operation is applied component-wise to its argument vector. The
strategy for adapting the factors γ j (n) is to minimize the loss function
X
L{γ 1 , . . . , γ 4 } = k
(γ j ).∧2 E[(z j ).∧2 ] − E[z .∧2 ]k2 ,
(104)
−2
j
where z is the feature space output signal z(n) = c(n) .∗ F 0 r(n) available during
exploitation time in the top module. In words, the adaptation of c aims at finding
a weighted disjunction of prototype vectors which optimally matches the currently
observed mean energies of the z signal.
It is straightforward to derive a stochastic gradient descent adaptation rule for
minimizing the loss (104). Let γ = (γ 1 , . . . , γ 4 ) be the row vector made from the
γ j , and let ·.2 denote element-wise squaring of a vector. Then
0
γ(n + 1) = γ(n) + λγ z(n + 1).∧2 − P (γ 0 (n)).∧2 P diag(γ(n))
(105)
implements the stochastic gradient of L with respect to γ, where λγ is an adaptation rate. In fact I do not use this formula as is, but add two helper mechanisms,
128
effectively carrying out
γ ∗ (n + 1) = γ(n) +
0
λγ z(n + 1).∧2 − P (γ 0 (n)).∧2 P diag(γ(n)) + d (1/2 − γ(n)) (106)
γ(n + 1) = γ ∗ (n + 1)/sum(γ ∗ (n + 1)).
(107)
The addition of the term d (1/2−γ(n)) pulls the γ j away from the possible extremes
0 and 1 toward 1/2 with a drift force d, which is a design parameter. This is helpful
to escape from extreme values (notice that if γ j (n) = 0, then γ j would forever
remain trapped at that value in the absence of the drift force). The normalization
(107) to a unit sum of the γ j greatly reduces adaptation jitter. I found both
amendments crucial for a reliable performance of the γ adaptation.
Given the γ(n) vector, the top-module c(n) is obtained by
.∧−1
c(n) = P (γ 0 (n)).∧2 .∗ P (γ 0 (n)).∧2 + α−2
.
Simulation Experiment 1: Online Classification and De-Noising. Please consult Figure 43 for a graphical display of this experiment. Details (training, initalization, parameter settings) are provided in the Experiments and Methods section
4.10.
The trained 3-layer architecture was driven with a 16,000 step input signal
composed of four blocks of 4000 steps each. In these blocks, the input was generated from the patterns p1 , p3 , p2 , p4 in turn (black lines in bottom row of Figure
43), with additive Gaussian noise scaled such that a signal-to-noise ratio of 1/2
was obtained (red lines in bottom row of Figure 43). The 3-layer architecture was
run for the 16,000 steps without external intervention.
The evolution of the four γ j weights in the top layer represent the system’s
classification hypotheses concerning the type of the current driver (top row in
Figure 43). In all four blocks, the correct decision is reached after an initial “rethinking” episode. The trust variable τ[23] quickly approaches 1 after taking a drop
at the block beginnings (green line in fourth row of Figure). This drop allows the
system to partially pass through the external driver signal up to the top module,
de-stabilizing the hypothesis established in the preceding block. The trust variable
τ[12] (blue line in fourth row) oscillates more irregularly but also takes its steepest
drops at the block beginnings.
j
For diagnostic purposes, γ[l]
weights were also computed on the lower layers
l = 2, 3, using (106) and (107). These quantities, which were not entering the
system’s processing, are indicators of the “classification belief states” in lower
modules (second and third row in the Figure). A stagewise consolidation of these
hypotheses can be observed as one passes upwards through the layers.
The four patterns fall in two natural classes, “sinewaves” and “period-5”. Inspecting the top-layer γ j it can be seen that within each block, the two hypothesis
indicators associated with the “wrong” class are quickly suppressed to almost zero,
129
1
0.5
0
1
0.5
0
1
0.5
0
1
0.5
0
0
!1
!2
0
4000
8000
12000
16000
2
1
0
!1
!2
0
10
20
0
10
20
0
10
20
0
10
20
Figure 43: Denoising and classification of prototype patterns. Noisy patterns were
given as input in the order p1 , p3 , p2 , p4 for 4000 timesteps each. Top row: evolution
of γ 1 (blue), γ 2 (green), γ 3 (red), γ 4 (cyan) in top layer module. Rows 2 and 3:
same in modules 2 and 1. Fourth row: trust variables τ[12] (blue) and τ[23] (green).
Fifth row: NRMSEs for reconstructed signals y[1] (blue), y[2] (green) and y[3] (red).
Black line shows the linear filter reference NRMSE. Thin red line: NRMSE of
phase-aligned y[3] . The plotting scale is logarithmic base 10. Bottom: pattern
reconstruction snapshots from the last 20 timesteps in each pattern presentation
block, showing the noisy input (red), the layer-3 output y[3] (thick gray) and the
clean signal (thin black). For explanation see text.
130
while the two indicators of the current driver’s class quickly dominate the picture
(summing to close to 1) but take a while to level out to their relative final values.
Since the top-level c(n) is formally a γ j -weighted disjunction of the four prototype
cj conception vectors, what happens in the first block (for instance) can also be
rephrased as, “after the initial re-thinking, the system is confident that the current
driver is p1 OR p3 , while it is quite sure that it is NOT (p2 OR p4 )”. Another
way to look at the same phenomena is to say, “it is easy for the system to quickly
decide between the two classes, but it takes more time to distinguish the rather
similar patterns within each class”.
The fifth row in Figure 43 shows the log10 NRMSE (running smoothed average)
of the three module outputs y[l] (n) with respect to a clean version of the driver
(thick lines; blue = y[1] , green = y[2] , red = y[3] ). For the 5-periodic patterns
(blocks 2 and 4) there is a large increase in accuracy from y[1] to y[2] to y[3] . For
the sinewave patterns this is not the case, especially not in the first block. The
reason is that the re-generated sines y[2] and y[3] are not perfectly phase-aligned
to the clean version of the driver. This has to be expected because such relative
phase shifts are typical for coupled oscillator systems; each module can be regarded
as an oscillator. After optimal phase-alignment (details in the Experiments and
Methods section), the top-level sine re-generation matches the clean driver very
accurately (thin red line). The 5-periodic signal behaves differently in this respect.
Mathematically, an 5-periodic discrete-time dynamical system attractor is not an
oscillation but a fixed point of the 5-fold iterated map, not admitting anything
like a gradual phase shift.
As a baseline reference, I also trained a linear transversal filter (Wiener filter,
see [28] for a textbook treatment) on the task of predicting the next clean input
value (details in the Experiments and Methods section). The length of this filter
was 2600, matching the number of trained parameters in the conceptor architecture
(P has 500 ∗ 4 learnt parameters, D has 500, W out has 100). The smoothed log10
NRMSE of this linear predictor is plotted as a black line. It naturally can reach
its best prediction levels only after 2600 steps in each block, much more slowly
than the conceptor architecture. Furthermore, the ultimate accuracy is inferior
for all four patterns.
Simulation Experiment 2: Tracking Signal Morphs. Our architecture can be
characterized as a signal cleaning-and-interpretation systems which guides itselft
by allowing top-down hypotheses to make lower processing layers selective. An
inherent problem in such systems is that that they may erroneously lock themselves
on false hypotheses. Top-down hypotheses are self-reinforcing to a certain degree
because they cause lower layers to filter out data components that do not agree
with the hypothesis – which is the essence of de-noising after all.
In order to test how our architecture fares with respect to this “self-locking
fallacy”, I re-ran the simulation with an input sequence that was organized as a
linear morph from p1 to p2 in the first 4000 steps (linearly ramping up the sine
frequency), then in the next block back to p1 ; this was followed by a morph from
131
p3 to p4 and back again. The task now is to keep track of the morph mixture in the
top-level γ j . This is a greatly more difficult task than the previous one because
the system does not have to just decide between 4 patterns, but has to keep track
of minute changes in relative mixtures. The signal-to-noise ratio of the external
input was kept at 0.5. The outcome reveals an interesting qualitative difference
in how the system copes with the sine morph as opposed to the 5-periodic morph.
As can be seen in Figure 44, the highest-layer hypothesis indicators γ j can track
the frequency morph of the sines (albeit with a lag), but get caught in a constant
hypothesis for the 5-period morph.
This once again illustrates that irrational-period sines are treated qualitatively
differently from integer-periodic signals in conceptor systems. I cannot offer a
mathematical analysis, only an intuition. In the sinewave tracking, the overall
architecture can be described as a chain of three coupled oscillators, where the
bottom oscillator is externally driven by a frequency-ramping sine. In such a driven
chain of coupled oscillators, one can expect either chaos, the occurrence of natural
harmonics, or frequency-locking across the elements of the chain. Chaos and
harmonics are ruled out in our architecture because the prototype conceptors and
the loading of two related basic sines prevent it. Only frequency-locking remains
as an option, which is indeed what we find. The 5-periodic morph cannot benefit
from this oscillation entrainment. The minute differences in shape between the two
involved prototypes do not stand out strongly enough from the noise background
to induce a noticable decline in the trust variable τ[23] : once established, a single
hypothesis persists.
On a side note it is interesting to notice that the linear filter that was used
as a baseline cannot at all cope with the frequency sweep, but for the 5-periodic
morph it performs as well as in the previous simulation. Both effects can easily
be deduced from the nature of such filters.
Only when I used a much cleaner input signal (signal-to-noise ratio of 10), and
after the decisiveness d was reduced to 0.5, it became possible for the system to
also track the 5-period pattern morph, albeit less precisely than it could track the
sines (not shown).
Variants and Extensions. When I first experimented with architectures of the
kind proposed above, I computed the module-internal conception weight vectors
caut
[l] (compare Equation (97)) on the two lower levels not via the autoconception
mechanism, but in a way that was similar to how I computed the top-level conception weight vector c[3] , that is, optimizing a fit to a disjunction of the four
prototypes. Abstractly speaking, this meant that a powerful piece of prior information, namely of knowing that the driver was one of the four prototypes, was
inserted in all processing layers. This led to a better system performance than
what I reported above (especially, faster decisions in the sense of faster convergence of the γ[3] ). However I subsequently renounced this “trick” because the
differences in performance were only slight, and from a cognitive modelling perspective I found it more appealing to insert such a valuable prior only in the top
132
1
0.5
0
1
0.5
0
1
0.5
0
1
0.5
0
0
!1
!2
0
4000
8000
12000
16000
2
1
0
!1
!2
0
10
20
0
10
20
0
10
20
0
10
20
Figure 44: Morph-tracking. First 4000 steps: morphing from p1 to p2 , back again
in next 4000 steps; steps 8000 – 16,000: morphing from p3 to p4 and back again.
Figure layout same as in Figure 43.
layer (motto: the retina does not conceptually understand what it sees).
Inspecting again the top row in Figure 43, one finds fast initial decision between
the alternatives “pattern 1 or 2” versus “pattern 3 or 4”, followed by a much slower
differentation within these two classes. This suggests architecture variants where
all layers are informed by priors of the kind as in Equation (96), that is, the local
conceptor on a layer is constrained to an aperture-weighted disjunction of a finite
number of prototype conceptors. However, the number of prototype conceptors
would shrink as one goes upwards in the hierarchy. The reduction in number would
be effected by merging several distinct prototype
conception vectors cj1 , . . . , cjk in
W
layer l into a single prototype vector cj = {cj1 , . . . , cjk }. In terms of classical AI
knowledge representation formalisms this would mean to implement an abstraction
133
hierarchy. A further refinement that suggests itself would be to install a top-down
processing pathway by which the current hypothesis on layer l + 1 selects which
finer-grained disjunction
on layer l is chosen. For instance, when
W ofj1prototypes
jk
j
j
c[l+1] (n) = c and c = {c , . . . , c }, then the conception weight vector c[l] (n) is
W
ji
constrained to be of the form i=1,...,k ϕ(cji , γ[l]
(n)). This remains for future work.
The architecture presented above is replete with ad-hoc design decisions. Numerous details could have been realized differently. There is no unified theory
which could inform a system designer what are the “right” design decisions. A
complete SIA architecture must provide a plethora of dynamical mechanisms for
learning, adaptation, stabilization, control, attention and so forth, and each of
them in multiple versions tailored to different subtasks and temporal scales. I do
not see even a theoretical possibility for an overarching, principled theory which
could afford us with rigorous design principles for all of these. The hierarchical
conceptor architecture presented here is far from realizing a complete SIA system,
but repercussions of that under-constrainedness of design already show.
Discussion. Hierarchical neural learning architectures for pattern recognition
have been proposed in many variants (examples: [66, 38, 33, 47, 35, 113]), albeit almost always for static patterns. The only example of hierarchical neural
architectures for temporal pattern recognition that I am aware of are the localistconnectionistic SHRUTI networks for text understanding [98]. Inherently temporal hierarchical pattern classification is however realized in standard hiddenMarkov-model (HMM) based models for speech recognition.
There is one common characteristic across all of these hierarchical recognition
systems (neural or otherwise, static or temporal). This shared trait is that when
one goes upward through the processing layers, increasingly “global” or “coarsegrained” or “compound” features are extracted (for instance, local edge detection
in early visual processing leading through several stages to object recognition in
the highest layer). While the concrete nature of this layer-wise integration of
information differs between approaches, at any rate there is change of represented
categories across layers. For the sake of discussion, let me refer to this as the
feature integration principle.
From the point of view of logic-based knowldege representation, another important trait is shared by hierarchical pattern recognition systems: abstraction. The
desired highest-level output is a class labelling of the input data. The recognition
architecture has to be able to generalize from the particular input instance. This
abstraction function is not explicitly implemented in the layers of standard neural
(or HMM) recognizers. In rule-based decision-tree classification systems (textbook: [76]), which can also be regarded as hierarchical recognition systems, the
hierarchical levels however directly implement a series of class abstractions. I will
refer to the abstraction aspect of pattern recognition as the categorical abstraction
principle.
The conceptor-based architecture presented in this section implements categorical abstraction through the γ variables in the highest layer. They yield (graded)
134
class judgements similar to what is delivered by the class indicator variables in
the top layers of typical neural pattern recognizers.
The conceptor-based architecture is different from the typical neural recognition systems in that it does not implement the feature integration principle.
As one progresses upwards through the layers, always the same dynamic item is
represented, namely, the current periodic pattern, albeit in increasingly denoised
versions. I will call this the pattern integrity principle. The pattern integrity
principle is inherently conflicting with the feature integration principle.
By decades of theoretical research and successful pattern recognition applications, we have become accustomed to the feature integration principle. I want
to argue that the pattern integrity principle has some cognitive plausibility and
should be considered when one designs SIA architectures.
Consider the example of listening to a familiar piece of music from a CD player
in a noisy party environment. The listener is capable of two things. Firstly, s/he
can classify the piece of music, for instance by naming the title or the band.
This corresponds to performing categorical abstraction, and this is what standard
pattern recognition architectures would aim for. But secondly, the listener can
also overtly sing (or whistle or hum) along with the melody, or s/he can mentally
entrain to the melody. This overt or covert accompaniment has strong de-noising
characteristics – party talk fragments are filtered out in the “mental tracing” of
the melody. Furthermore, the mental trace is temporally entrained to the source
signal, and captures much temporal and dynamical detail (single-note-level of
accuracy, stressed versus unstressed beats, etc). That is an indication of pattern
integrity.
Another example of pattern integrity: viewing the face of a conversation partner during a person-to-person meeting, say Anne meeting Tom. Throughout the
conversation, Anne knows that the visual impression of Tom’s face is indeed Tom’s
face: categorical abstraction is happening. But also, just like a listener to noiseoverlaid music can trace online the clean melody, Anne maintains a “clean video”
representation of Tom’s face as it undergoes a succession of facial expressions and
head motions. Anne experiences more of Tom’s face than just the top-level abstraction that it is Tom’s; and this online experience is entrained to Anne’s visual
input stream.
Pattern integrity could be said to be realized in standard hierarchical neural
architectures to the extent that they are generative. Generative models afford of
mechanisms by which example instances of a recognition class can be actively produced by the system. Prime examples are the architectures of adaptive resonance
theory [43], the Boltzmann Machine [2] and the Restricted Boltzmann Machine /
Deep Belief Networks [47]. In systems of this type, the reconstruction of pattern
instances occurs (only) in the input layer, which can be made to “confabulate” or
“hallucinate” (both terms are used as technical terms in the concerned literature)
pattern instances when primed with the right bias from higher layers.
Projecting such architectures to the human brain (a daring enterprise) and
135
returning to the two examples above, this would correspond to re-generating
melodies in the early auditory cortex or facial expressions in the early visual cortex
(or even in the retina). But I do not find this a convincing model of what happens
in a human brain. Certainly I am not a neuroscientist and not qualified to make
scientific claims here. My doubts rest on introspection (forbidden! I know) and
on a computational argument. Introspection: when I am mentally humming along
with a melody at a party, I still do hear the partytalk – I dare say my early auditory modules keep on being excited by the entire auditory input signal. I don’t feel
like I was hallucinating a clean version of the piece of music, making up an auditory reality that consists only of clean music. But I do not listen to the talk noise,
I listen only to the music components of the auditory signal. The reconstruction
of a clean version of the music happens – as far as I can trust my introspection
– “higher up” in my brain’s hierarchy, closer to the quarters where consciously
controllable cognition resides. The computational argument: generative models,
such as the ones mentioned, cannot (in their current versions at least) generate
clean versions of noisy input patterns while the input is presented. They either
produce a high-level classification response while being exposed to input (bottomup processing mode), or they generate patterns in their lowest layer while being
primed to a particular class in their highest layer (top-down mode). They can’t
do both at the same time. But humans can: while being exposed to input, a
cleaned-up version of the input is being maintained. Furthermore, humans (and
the conceptor architecture) can operate in an online-entrained mode when driven
by temporal data, while almost all existing recognition architectures in machine
learning are designed for static patterns.
Unfortunately I cannot offer a clear definition of “pattern integrity”. An aspect
of pattern integrity that I find important if not defining is that some temporal
and spatial detail of a recognized pattern is preserved across processing layers.
Even at the highest layers, a “complete” representation of the pattern should
be available. This seems to agree with cognitive theories positing that humans
represent concepts by prototypes, and more specifically, by exemplars (critical
discussion [63]). However, these cognitive theories relate to empirical findings on
human classification of static, not temporal, patterns. I am aware that I am vague.
One reason for this is that we lack a scientific terminology, mathematical models,
and standard sets of examples for discussing phenomena connected with pattern
integrity. All I can bring to the table at this point is just a new architecture that
has some extravagant processing characteristics. This is, I hope, relevant, but it
is premature to connect this in any detail to empirical cognitive phenomena.
3.17
Toward a Formal Marriage of Dynamics with Logic
In this subsection I assume a basic acquaintance of the reader with Boolean and
first-order predicate logic.
So far, I have established that conceptor matrices can be combined with (al-
136
most) Boolean operations, and can be ordered by (a version of) abstraction. In this
subsection I explore a way to extend these observations into a formal “conceptor
logic”.
Before I describe the formal apparatus, I will comment on how I will be understanding the notions of “concept” and “logic”. Such a preliminary clarification
is necessary because these two terms are interpreted quite differently in different
contexts.
“Concepts” in the cognitive sciences. I start with a quote from a recent survey
on research on concepts and categories in the cognitive sciences [75]: “The concept
of concepts is difficult to define, but no one doubts that concepts are fundamental
to mental life and human communication. Cognitive scientists generally agree that
a concept is a mental representation that picks out a set of entities, or a category.
That is, concepts refer, and what they refer to are categories. It is also commonly
assumed that category membership is not arbitrary but rather a principled matter.
What goes into a category belongs there by virtue of some law-like regularities.
But beyond these sparse facts, the concept CONCEPT is up for grabs.” Within
this research tradition, one early strand [87, 17] posited that the overall organization of a human’s conceptual representations, his/her semantic memory, can be
formally well captured by AI representation formalisms called semantic networks
in later years. In semantic network formalisms, concepts are ordered in abstraction hierarchies, where a more abstract concept refers to a more comprehensive
category. In subsequent research this formally clear-cut way of defining and organizing concepts largely dissolved under the impact of multi-faceted empirical
findings. Among other things, it turned out that human concepts are graded,
adaptive, and depend on features which evolve by learning. Such findings led
to a diversity of enriched models of human concepts and their organization (my
favourites: [62, 22, 64]), and many fundamental questions remain controversial.
Still, across all diversity and dispute, the basic conception of concepts spelled
out in the initial quote remains largely intact, namely that concepts are mental
representations of categories, and categories are defined extensionally as a set of
“entities”. The nature of these entities is however “up to grabs”. For instance,
the concept named “Blue” might be referring to the set of blue physical objects,
to a set of wavelengths of light, or to a set of sensory experiences, depending on
the epistemic approach that is taken.
“Concepts” in logic formalisms. I first note that the word “concept” is not
commonly used in logics. However, it is quite clear what elements of logical systems
are equated with concepts when such systems are employed as models of semantic
memory in cognitive science, or as knowledge representation frameworks in AI.
There is a large variety of logic formalisms, but almost all of them employ typed
symbols, specifically unary predicate symbols, relation symbols of higher arity,
constant symbols, and function symbols. In the model-theoretic view on logic,
such symbols become extensionally interpreted by sets of elements defined over the
domain set of a set-theoretic model. Unary predicate symbols become interpreted
137
by sets of elements; n-ary relation symbols become interpreted by n-tuples of such
elements; function symbols by sets of argument-value pairs; constant symbols by
individual elements. A logic theory uses a fixed set of such symbols called the
theory’s signature. Within a theory, the interpretation of the signature symbols
becomes constrained by the axioms of the theory. In AI knowledge representation
systems, this set of axioms can be very large, forming a world model and situation
model (sometimes called “T-Box” and “A-Box”). In the parlance of logic-oriented
AI, the extension of unary predicate symbols are often called classes instead of
“categories”.
In AI applications, the world model is often implemented in the structure of a
semantic network [67], where the classes are represented by nodes labelled by predicate symbols. These nodes are arranged in a hierarchy with more abstract class
nodes in higher levels. This allows the computer program to exploit inheritance
of properties and relations down the hierarchy, reducing storage requirements and
directly enabling many elementary inferences. Class nodes in semantic networks
can be laterally linked by relation links, which are labelled by relation symbols.
At the bottom of such a hierarchy one may locate individual nodes labelled by
constant symbols. A cognitive scientist employing such a semantic network representation would consider the class nodes, individual nodes, and relation links
as computer implementations or formal models of class concepts, individual concepts, and relational concepts, respectively. Also, semantic network specification
languages are sometimes called concept description languages in AI programming.
On this background, I will understand the symbols contained in a logic signature
as names of concepts.
Furthermore, a logical expression ϕ[x1 , . . . , xn ] containing n free (first-order)
variables can be interpreted by the set of all n-tuples satisfying this expression.
ϕ[x1 , . . . , xn ] thus defines an n-ary relation. For example, ϕ[x] = Fruit(x) ∧
Yellow(x) ∧ Longish(x) ∧ Curved(x) would represent a class (seems to be the
class of bananas). Quite generally, logical expressions formed according to the
syntax of a logic formalism can build representations of new concepts from given
ones.
There is an important difference between how “concepts” are viewed in cognitive modeling versus logic-based AI. In the latter field, concepts are typically
named by symbols, and the formal treatment of semantics is based on a reference relationship between the symbols of a signature and their interpretations.
However, even in logic-based knowledge representation formalisms there can be
un-named concepts which are formally represented as logic expressions with free
variables, as for instance the banana formula above. In cognitive science, concepts
are not primarily or necessarily named, although a concept can be optionally labelled with a name. Cognitive modeling can deal with conceptual systems that
have not a single symbol, for instance when modeling animal cognition. By contrast, AI-style logic modeling typically is strongly relying on symbols (the only
exception being mathematical theories built on the empty signature; this is of
138
interest only for intra-mathematical investigations).
Remarks on “Logics”. In writing this paragraph, I follow the leads of the PhD
thesis [88] of Florian Rabe which gives a comprehensive and illuminating account
of today’s world of formal logic research. The field of mathematical logics has
grown and diversified enormously in the last three decades. While formal logic
historically has been developed within and for pure mathematics, much of this
recent boost was driven by demands from theoretical computer science, AI, and
semantic web technologies. This has led to a cosmos populated by a multitude of
“logics” which sometimes differ from each other even in basic premises of what,
actually, qualifies a formal system as a “logic”. In turn, this situation has led to
meta-logical research, where one develops formal logical frameworks in order to
systematically categorize and compare different logics.
Among the existing such logical frameworks, I choose the framework of institutions [37], because it has been devised as an abstraction of model-theoretic
accounts of logics, which allows me to connect quite directly to concepts, categories, and the semantic reference link between these two. Put briefly, a formal
system qualifies as a logic within this framework if it can be formulated as an
institution. The framework of institutions is quite general: all logics used in AI,
linguistics and theoretical cognitive sciences can be characterized as institutions.
The framework of institutions uses tools from category theory. In this section
I do not assume that the reader is familiar with category theory, and therefore
will give only an intuitive account of how “conceptor logic” can be cast as an
institution. A full categorical treatment is given in Section 3.18.
An institution is made of three main components, familiar from the model
theory of standard logics like first-order predicate logic:
1. a collection Sign of signatures, where each signature Σ is a set of symbols,
2. for each signature Σ, a set Sen(Σ) of Σ-sentences that can be formed using
the symbols of Σ,
3. again for each signature Σ, a collection Mod(Σ) of Σ-models, where a Σ-model
is a mathematical structure in which the symbols from Σ are interpreted.
Furthermore, for every signature Σ there is a model relation |=Σ ⊆ Mod(Σ) ×
Sen(Σ). For a Σ-model m and a Σ-sentence χ, we write infix notation m |=Σ χ
for (m, χ) ∈ |=Σ , and say “m is a model of χ”, with the understanding that the
sentence χ makes a true statement about m.
The relationsships between the main elements of an institution can be visualized as in Figure 45.
The full definition of an institution includes a mechanism for symbol re-naming.
The intuitive picture is the following. If a mathematician or an AI engineer writes
down a set of axioms, expressed as sentences in a logic, the choice of symbols
should be of no concern whatsoever. As Hilbert allegedly put it, the mathematical theory of geometry should remain intact if instead of “points, lines, surfaces”
139
Sen
Sen(Σ)
Σ
Mod
Mod(Σ)
Figure 45: How the elements of an institution relate to each other. For explanation
see text.
one would speak of “tables, chairs, beer mugs”. In the framework of institutions
this is reflected by formalizing how a signature Σ may be transformed into another
signature Σ0 by a signature morphism φ : Σ → Σ0 , and how signature morphisms
are extended to sentences (by re-naming symbols in a sentence according to the
signature morphism) and to models (by interpreting the re-named symbols by the
same elements of a model that were previously used for interpreting the original symbols). Then, if m0 , χ0 denote the re-named model m and sentence χ, an
institution essentially demands that m |=Σ χ if and only if m0 |=Σ0 χ0 .
For example, first-order logic (FOL) can be cast as an institution by taking
for Sign the class of all FOL signatures, that is the class of all sets containing
typed predicate, relation, function and constant symbols; Sen maps a signature
Σ to the set of all closed (that is, having no free variables) Σ-expressions (usually
called sentences); Mod assigns to each signature the class of all set-theoretic Σstructures; and |= is the satisfaction relation of FOL (also called model relation).
For another example, Boolean logic can be interpreted as an institution in several
ways, for instance by declaring Sign as the class of all totally ordered countable
sets (the elements of which would be seen as Boolean variables); for each signature
Σ of Boolean variables, Sen(Σ) is the set of all Boolean expressions ϕ[Xi1 , . . . , Xin ]
over Σ and Mod(Σ) is the set of all truth value assignments τ : Σ → {T, F } to the
Boolean variables in Σ; and τ |=Σ ϕ if ϕ evaluates to T under the assignment τ .
In an institution, one can define logical entailment between Σ-sentences in the
familiar way, by declaring that χ logically entails χ0 (where χ, χ0 are Σ-sentences)
if and only if for all Σ-models m it holds that m |=Σ χ implies m |=Σ χ0 . By a
standard abuse of notation, this is also written as χ |=Σ χ0 or χ |= χ0 .
I will sketch two entirely different approaches to define a “conceptor logic”.
The first follows in the footsteps of familiar logics. Conceptors can be named by
arbitrary symbols, sentences are built by an inductive procedure which specifies
how more complex sentences can be constructed from simpler ones by similar syntax rules as in first-order logic, and models are designated as certain mathematical
structures built up from named conceptors. This leads to a logic that essentially
represents a version of first-order logic constrained to conceptor domains. It would
be a logic useful for mathematicians to investigate “logical” characteristics of conceptor mathematics, especially whether there are complete calculi that allow one
to systematically prove all true facts concerning conceptors. I call such logics ex140
trinsic conceptor logics. Extrinsic conceptor logics are tools for mathematicians to
reason about conceptors. A particular extrinsic conceptor logic as an institution
is detailed in Section 3.18.
The other approach aims at a conceptor logic that, instead of being a tool for
mathematicians to reason about conceptors, is a model of how a situated intelligent
agent does “logical reasoning” with conceptors. I call this intrinsic conceptor logic
(ICL). An ICL has a number of unconventional properties:
• An ICL should function as a model of a situated agent’s conceptor-based
information processing. Agents are bound to differ widely in their structure
and their concrete lifetime learning histories. Therefore I do not attempt
to design a general “fits-all-agents” ICL. Instead, for every single, concrete
agent life history there will be an ICL, the private ICL of that agent life.
• An agent with a personal learning history is bound to develop its private
“logic” over time. The ICL of an agent life thus becomes a dynamical system
in its own right. The framework of institutions was not intended by its
designers to model temporally evolving objects. Specifying an institution
such that it can be considered a dynamical system leads to some particularly
unconventional characteristics of an agent life ICL. Specifically, signatures
become time-varying objects, and signature morphisms (recall that these
model the “renaming” of symbols) are used to capture the temporal evolution
of signatures.
An agent’s lifetime ICL is formalized differently according to whether the agent
is based on matrix conceptors or random feature conceptors. Here I work out only
the second case.
In the following outline I use the concrete three-layer de-noising and classification architecture from Section 3.16 as a reference example to fill the abstract
components of ICL with life. Even more concretely, I use the specific “lifetime history” of the 16000-step adaptation run illustrated in Figure 43 as demonstration
example. For simplicity I will refer to that particular de-noising and classification
architecture run as “DCA”.
Here is a simplified sketch of the main components of an agent’s lifetime ICL
(full treatment in Section 3.18):
1. An ICL is designed to model a particular agent lifetime history. A specification of such an ICL requires that a formal model of such an agent life is
available beforehand. The core part of an agent life model AL is a set of
m conceptor adaptation sequences {a1 (n), . . . , am (n)}, where each ai (n) is
an M -dimensional conception weight vector. It is up to the modeler’s discretion which conceptors in a modeled agent become included in the agent
life model AL. In the DCA example I choose the four prototype conception
weight vectors c1 , . . . , c4 and the two auto-adapted caut
[l] on layers l = 1, 2. In
141
this example, the core constituent of the agent life model AL is thus the set
aut
of m = 6 conceptor adaptation trajectories c1 (n), . . . , c4 (n), caut
[1] (n), c[2] (n),
where 1 ≤ n ≤ 16000. The first four trajectories c1 (n), . . . , c4 (n) are constant over time because these prototype conceptors are not adapted; the last
two evolve over time. Another part of an agent life model is the lifetime T ,
which is just the interval of timepoints n for which the adaptation sequences
ai (n) are defined. In the DCA example, T = (1, 2, . . . , 16000).
(n)
(n)
2. A signature is a finite non-empty set Σ(n) = {A1 , . . . , Am } of m timeindexed symbols Ai . For every n ∈ T there is a signature Σ(n) .
DCA example: In the ICL of this example agent life, the collection Sign of
(n)
(n)
(n)
(n)
signatures is made of 16000 signatures {C1 , . . . , C4 , A1 , A2 } containing
six symbols each, with the understanding that the first four symbols refer
to the prototype conceptors c1 (n), . . . , c4 (n) and the last two refer to the
aut
auto-adapted conceptors caut
[1] (n), c[2] (n).
3. For every pair Σ(n+k) , Σ(n) of signatures, where k ≥ 0, there is a signature
(n)
(n+k)
to Ai . These signamorphism φ(n+k,n) : Σ(n+k) → Σ(n) which maps Ai
ture morphisms introduce a time arrow into Sign. This time arrow “points
backwards”, leading from later times n + k to earlier times n. There is a
good reason for this backward direction. Logic is all about describing facts.
In a historically evolving system, facts χ(n+k) established at some later time
n + k can be explained in terms of facts ζ (n) at preceding times n, but not
vice versa. Motto: “the future can be explained in terms of the past, but the
past cannot be reduced to facts from the future”. Signature morphisms are a
technical vehicle to re-formulate descriptions of facts. They must point backwards in time in order to allow facts at later times to become re-expressed
in terms of facts stated for earlier times. Figure 46 illustrates the signatures
and their morphisms in an ICL.
(n)
(n)
4. Given a signature Σ(n) = {A1 , . . . , Am }, the set of sentences Sen(Σ(n) )
which can be expressed with the symbols of this signature is the set of
syntactic expressions defined inductively by the following rules (incomplete,
full treatment in next subsection):
(n)
(n)
(a) A1 , . . . , Am are sentences in Sen(Σ(n) ).
(n)
(b) For k ≥ 0 such that n, n + k ∈ T , for Ai
Sen(Σ(n) ).
(n)
(n)
∈ Σ(n) , δk Ai
is in
(c) If ζ, ξ ∈ Sen(Σ(n) ), then (ζ ∨ ξ), (ζ ∧ ξ), ¬ζ ∈ Sen(Σ(n) ).
(d) If ζ ∈ Sen(Σ(n) ), then ϕ(ζ, γ) ∈ Sen(Σ(n) ) for every γ ∈ [0, ∞] (this
captures aperture adaptation).
(e) If ζ, ξ ∈ Sen(Σ(n) ) and 0 ≤ b ≤ 1, then βb (ζ, ξ) ∈ Sen(Σ(n) ) (this will
take care of linear blends bζ + (1 − b)ξ).
142
" (n+1, n – 1)
(n+1, n)
" (n, n – 1) "
!(n)
!(n +1)
!(n – 1)
...
A1(n – 1)
A2(n – 1)
...
A1(n)
A2(n)
...
A1(n + 1)
...
A2(n + 1)
...
time!n
Figure 46: Signatures and their morphisms in an ICL (schematic). For explanation
see text.
In words, sentences express how new conceptors can be built from existing
ones by Boolean operations, aperture adaptation, and linear blends. The
“seed” set for these inductive constructions is provided by the conceptors
that can be directly identified by the symbols in Σ(n) .
(n)
(n)
The sentences of form δk Ai deserve a special comment. The operators
(n)
(n) (n)
δk are time evolution operators. A sentence δk Ai will be made to refer
to the conceptor version ai (n + k) at time n + k which has evolved from
ai (n).
5. For every time n, the set Mod(Σ(n) ) of Σ(n) -models is the set Z of M dimensional nonnegative vectors.
Remarks: (i) The idea for these models is that they represent mean energy
vectors E[z .∧.2 ] of feature space states. (ii) The set of models Mod(Σ(n) ) is
the same for every signature Σ(n) .
DCA example: Such feature space signal energy vectors occur at various
places in the DCA, for instance in Equations (102), (103), and conception
weight vectors which appear in the DCA evolution are all defined or adapted
in one way or other on the basis of such feature space signal energy vectors.
6. Every Σ(n) -sentence χ is associated with a concrete conception weight vector
ι(χ) by means of the following inductive definition:
(n)
(a) ι(Ai ) = ai (n).
(n)
(n)
(b) ι(δk Ai ) = ai (n + k).
(c) Case χ = (ζ ∨ ξ): ι(χ) = ι(ζ) ∨ ι(ξ) (compare Definition 7).
143
(d) Case χ = (ζ ∧ ξ): ι(χ) = ι(ζ) ∧ ι(ξ).
(e) Case χ = ¬ζ: ι(χ) = ¬ι(ζ).
(f) Case χ = ϕ(ζ, γ): ι(χ) = ϕ(ι(ζ), γ) (compare Definition 6).
(g) Case χ = βb (ζ, ξ): ι(χ) = b ι(ζ) + (1 − b)ι(ξ).
Remark: This statement of the interpretation operator ι is suggestive only.
The rigorous definition (given in the next section) involves additional non(n)
trivial mechanisms to establish the connection between the symbol Ai and
the concrete conceptor version ai (n) in the agent life. Here I simply appeal
to the reader’s understanding that symbol Ai refers to object ai .
7. For z .∧2 ∈ Z and χ ∈ Sen(Σ(n) ), the model relationship is defined by
z .∧2 |=Σ(n) χ iff z .∧2 .∗ (z .∧2 + 1).∧−1 ≤ ι(χ).
(108)
Remark: This definition in essence just repeats how a conception weight
vector is derived from a feature space signal energy vector.
When all category-theoretical details are filled in which I have omitted here,
one obtains a formal definition of an institution which represents the ICL of an
agent life AL. It can be shown that in an ICL, for all ζ, ξ ∈ Sen(Σ(n) ) it holds
that
ζ |=Σ(n) ξ iff ι(ζ) ≤ ι(ξ).
By virtue of this fact, logical entailment becomes decidable in an ICL: if one wishes
to determine whether ξ is implied by ζ, one can effectively compute the vectors
ι(ζ), ι(ξ) and then check in constant time whether ι(ζ) ≤ ι(ξ).
Returning to the DCA example (with lifetime history shown in Figure 43), its
ICL identiefies over time the four prototype conceptors c1 , . . . , c4 and the two auto(n)
(n)
auto
by temporally evolving symbols {C1 , . . . , C4 ,
adapted conceptors cauto
[1] , c[2]
(n)
(n)
A1 , A2 }. All other conceptors that are computed in this architecture can
be defined in terms of these six ones. For instance, the top-level conceptor
c[3] (n) can be
W expressed in terms of the identifiable four prototype conceptors
by c[3] (n) = j=1,...,4 ϕ(cj , γ j (n)) by combining the operations of disjunction and
aperture adaptation. In ICL syntax this construction would be expressible by a
Σ(n) sentence, for instance by
(n)
(n)
(n)
(n)
(((ϕ(C1 , γ 1 (n)) ∨ ϕ(C2 , γ 2 (n))) ∨ ϕ(C3 , γ 3 (n))) ∨ ϕ(C4 , γ 4 (n))).
A typical adaptation objective of a conception vector c(n) occurring in an agent
life is to minimize a loss of the form (see Definition 88)
Ez [kz − c(n) .∗ zk2 ] + α−2 kc(n)k2 ,
144
or equivalently, the objective is to converge to
c(n) = E[α2 z .∧2 ] .∗ (E[α2 z .∧2 ] + 1).∧−1 .
This can be re-expressed in ICL terminology as “adapt c(n) such that α2 E[z .∧2 ] |=Σ(n)
χc(n) , and such that not z .∧2 |=Σ(n) χc(n) for any z .∧2 > α2 E[z .∧2 ]” (here χc(n) is
an adhoc notation for an ICL sentence χc(n) ∈ Sen(Σ(n) ) specifying c(n)). In
more abstract terms, the typical adaptation of random feature conceptors can be
understood as an attempt to converge toward the conceptor that is maximally
|=-specific under a certain constraint.
Discussion. I started this section by a rehearsal of how the notion of “concept” is understood in cognitive science and logic-based AI formalisms. According
to this understanding, a concept refers to a category (terminology of cognitive
science); or a class symbol or logical expression with free variables is interpreted
by its set-theoretical extension (logic terminology). Usually, but not necessarily,
the concepts/logical expressions are regarded as belonging to an “ontological” domain that is different from the domain of their respective referents. For instance,
consider a human maintaining a concept named cow in his/her mind. Then many
cognitive scientists would identifiy the category that is referred to by this concept with the some set of physical cows. Similarly, an AI expert system set up
as a farm management system would contain a symbol cow in its signature, and
this symbol would be deemed to refer to a collection of physical cows. In both
cases, the concept / symbolic expression cow is ontologically different from a set of
physical cows. However, both in cognitive science and AI, concepts / symbolic expressions are sometimes brought together with their referents much more closely.
In some perspectives taken in cognitive science, concepts are posited to refer to
other mental items, for instance to sensory perceptions. In most current AI proof
calculi (“inference engines”), models of symbolic expressions are created which
are assembled not from external physical objects but from symbolic expressions
(“Herbrand universe” constructions). Symbols from a signature Σ then refer to
sets of Σ-terms. In sum, fixing the ontological nature of referents is ultimately left
to the modeling scientist in cognitive science or AI.
In contrast, ICL is committed to one particular view on the semantic relationship: Σ(n) -sentences are always describing conception weight vectors, and refer
to neural activation energy vectors z .∧2 . In the case of matrix conceptor based
agents, Σ(n) -sentences describe conceptor matrices and refer to neural activation
correlation matrices R by the following variant of (108):
R |=Σ(n) χ iff R(R + I)−1 ≤ ι(χ).
(109)
In Figure 47 I try to visualize this difference between the classical, extensional
view on symbols and their referents, and the view adopted by ICL. This figure
contrasts how classical logicians and cognitive scientists would usually model an
agent’s representation of farm livestock, as opposed to how ICL renders that situation. The semantic relation is here established between the physical world on
145
the one side, and symbols and logical expressions on the other side. The world is
idealized as a set of individuals (individual animals in this example), and symbols
for concepts (predicate symbols in logic) are semantically interpreted by sets of
individuals. In the farmlife example, a logician might introduce a symbol lifestock
which would denote the set of all economically relevant animals grown in farms,
and one might introduce another symbol poultry to denote the subset of all feathered such animals. The operator that creates “meaning” for concept symbols is
the grouping of individuals into sets (the bold “{ }” in Figure 47).
With conceptors, the semantic relation connects neural activity patterns triggered by perceiving animals on the one side, with conceptors acting on neural
dynamics on the other side. The core operator that creates meaning is the condensation of the incoming data into a neural activation energy pattern z .∧2 (or
correlation matrix R for matrix conceptors) from which conceptors are generated
via the fundamental construction c = E[z .∧2 ] .∗ (E[z .∧2 ]+1).∧−1 or C = R(R+I)−1
(Figure 47 depicts the latter case).
Figure 47: Contrasting the extensional semantics of classical knowledge representation formalisms (upper half of graphics) with the system-internal neurodynamical
semantics of conceptors (lower half). In both modeling approaches, abstraction
hierarchies of “concepts” arise. For explanation see text.
ICL, as presented here, cannot claim to be a model of all “logical reasoning”
in a neural agent. Specifically, humans sometimes engage in reasoning activities
which are very similar to how syntactic logic calculi are executed in automated
theorem proving. Such activities include the build-up and traversal of search trees,
146
creating and testing hypotheses, variable binding and renaming, and more. A
standard example is the step-by-step exploration of move options done by a human
chess novice. ICL is not designed to capture such conscious combinatorial logical
reasoning. Rather, ICL is intended to capture the automated aspects of neural
information processing of a situated agent, where incoming (sensor) information
is immediately transformed into perceptions and maybe situation representations
in a tight dynamical coupling with the external driving signals.
The material presented in this and the next section is purely theoretical and
offers no computational add-on benefits over the material presented in earlier
sections. There are three reasons why nonetheless I invested the effort of defininig
ICLs:
• By casting conceptor logic rigorously as an institution, I wanted to substantiate my claim that conceptors are “logical” in nature, beyond a mere appeal
to the intuition that anything admitting Boolean operations is logic.
• The institutional definition given here provides a consistent formal picture
of the semantics of conceptors. A conceptor c identified by an ICL sentence
χc “means” neural activation energy vectors z .∧2 . Conceptors and their
meanings are both neural objects of the same mathematical format, M dimensional nonnegative vectors. Having a clear view on this circumstance
helps to relate conceptors to the notions of concepts and their referents,
which are so far from being fully understood in the cognitive sciences.
• Some of the design ideas that went into casting ICLs as institutions may
be of more general interest for mathematical logic research. Specifically,
making signatures to evolve over time – and hence, turn an institution into
a dynamical system – might be found a mechanism worth considering in
scenarios, unconnected with conceptor theory or neural networks, where one
wants to analyse complex dynamical systems by means of formal logics.
3.18
Conceptor Logic as Institutions: Category-Theoretical
Detail
In this section I provide a formal specification of conceptor logic as an institution.
This section addresses only readers with a dedicated interest in formal logic. I
assume that the reader is familiar with the institution framework for representing
logics (introduced in [37] and explained in much more detail in Section 2 in [88])
and with basic elements of category theory. I first repeat almost verbatim the
categorical definition of an institution from [37].
Definition 8 An institution I consists of
1. a category Sign, whose objects Σ are called signatures and whose arrows are
called signature morphisms,
147
2. a functor Sen : Sign → Set, giving for each signature a set whose elements
are called sentences over that signature,
3. a functor Mod : Sign → Catop , giving for each signature Σ a category
Mod(Σ) whose objects are called Σ-models, and whose arrows are called Σ(model) morphisms, and
4. a relation |=Σ ⊆ Mod(Σ) × Sen(Σ) for each Σ ∈ Sign, called Σ-satisfaction,
such that for each morphism φ : Σ1 → Σ2 in Sign, the Satisfaction Condition
m2 |=Σ2 Sen(φ)(χ1 )
iff
Mod(φ)(m2 ) |=Σ1 χ1
(110)
holds for each m2 ∈ Mod(Σ2 ) and each χ1 ∈ Sen(Σ1 ).
The interrelations of these items are visualized in Figure 48.
Set
Sen
Sign
Mod
!1"
Catop
Mod(!1)
!1
Sen(!)
Mod(!)
!
!2"
Sen(!1)
Mod(!2)
!2
Sen(!2)
Figure 48: Relationships between the constituents of an institution (redrawn from
[37]).
Remarks:
1. The morphisms in Sign are the categorical model of re-naming the symbols
in a logic. The essence of the entire apparatus given in Definition 8 is to
capture the condition that the model-theoretic semantics of a logic is invariant to renamings of symbols, or, as Goguen and Burstall state it, “Truth is
invariant under change of notation”.
2. The intuition behind Σ-model-morphisms, that is, maps µ : I1Σ → I2Σ , where
I1Σ , I2Σ are two Σ-models, is that µ is an embedding of I1Σ in I2Σ . If we take
first-order logic as an example, with I1Σ , I2Σ being two Σ-structures, then
148
µ : I1Σ → I2Σ would be a map from the domain of I1Σ to the domain of I2Σ
which preserves functional and relational relationships specified under the
interpretations of I1Σ and I2Σ .
3. In their original 1992 paper [37], the authors show how a number of standard logics can be represented as institutions. In the time that has passed
since then, institutions have become an important “workhorse” for software
specification in computer science and for semantic knowledge management
systems in AI, especially for managing mathematical knowledge, and several
families of programming toolboxes have been built on institutions (overview
in [88]). Alongside with the model-theoretic spirit of institutions, this proven
usefulness of institutions has motivated me to adopt them as a logical framework for conceptor logic.
Logical entailment between sentences is defined in institutions in the traditional
way:
Definition 9 Let χ1 , χ2 ∈ Sen(Σ). Then χ1 entails χ2 , written χ1 |=Σ χ2 , if for
all m ∈ Ob(Mod(Σ)) it holds that m |=Σ χ1 → m |=Σ χ2 .
Institutions are flexible and offering many ways for defining logics. I will frame
two entirely different kinds of conceptor logics. The first kind follows the intuitions
behind the familiar first-order predicate logic, and should function as a formal
tool for mathematicians to reason about (and with) conceptors. Since it looks
at conceptors “from the outside” I will call it extrinsic conceptor logic (ECL).
Although ECL follows the footsteps of familiar logics in many respects, in some
aspects it deviates from tradition. The other kind aims at modeling the “logical”
operations that an intelligent neural agent can perform whose “brain” implements
conceptors. I find this the more interesting formalization; certainly it is the more
exotic one. I will call it intrinsic conceptor logic (ICL).
Extrinsic conceptor logic. I first give an intuitive outline. I treat only the case
of matrix-based conceptors. An ECL concerns conceptors of a fixed dimension
N and their logical interrelationships, so one should more precisely speak of N dimensional ECL. I assume some N is fixed. Sentences of ECL should enable
a mathematician to talk about conceptors in a similar way as familiar predicate
logics allow a mathematician to describe facts about other mathematical objects.
For example, “for all conceptors X, Y it holds that X ∧ Y ≤ X and X ∧ Y ≤
Y ” should be formalizable as an ECL sentence. A little notational hurdle arises
because Boolean operations appear in two roles: as operators acting on conceptors
(the “∧” in the sentence above), and as constituents of the logic language (the
“and” in that sentence). To keep these two roles notationally apart, I will use
AND, OR, NOT (allowing infix notation) for the role as operators, and ∧, ∨, ¬
for the logic language. The above sentence would then be formally written as
“∀x∀y (x AND y ≤ x) ∧ (x AND y ≤ y)”.
149
The definition of signatures and ECL-sentences in many respects follows standard customs (with significant simplifications to be explained after the definition)
and is the same for any conceptor dimension N :
Definition 10 Let Var = {x1 , x2 , ...} be a fixed countable indexed set of variables.
1. (ECL-signatures) The objects (signatures) of Sign are all countable sets,
whose elements are called symbols. For signatures Σ1 , Σ2 , the set of morphisms hom (Σ1 , Σ2 ) is the set of all functions φ : Σ1 → Σ2 .
2. (ECL-terms) Given a signature Σ, the set of Σ-terms is defined inductively
by
(a) Every variable xi , every symbol a ∈ Σ, and I is a Σ-term.
(b) For Σ-terms t1 , t2 and γ ∈ [0, ∞], the following are Σ-terms: NOT t1 ,
(t1 AND t2 ), (t1 OR t2 ), and ϕ(t1 , γ).
3. (ECL-expressions) Given a signature Σ, the set Exp(Σ) of Σ-expressions is
defined inductively by
(a) If t1 , t2 are Σ-terms, then t1 ≤ t2 is a Σ-expression.
(b) If e1 , e2 are Σ-expressions, and xi a variable, then the following are
Σ-expressions: ¬e1 , (e1 ∧ e2 ), (e1 ∨ e2 ), ∀xi e1 .
4. (ECL-sentences) A Σ-expression that contains no free variables is a Σsentence (free occurrence of variables to be defined as usual, omitted here.)
Given a Σ-morphism φ : Σ1 → Σ2 , its image Sen(φ) under the functor Sen is
the map which sends every Σ1 -sentence χ1 to the Σ2 -sentence χ2 obtained from
χ1 by replacing all occurrences of Σ1 symbols in χ1 by their images under φ. I
omit the obvious inductive definition of this replacement construction.
Notes:
• ECL only has a single sort of symbols with arity 0, namely constant symbols
(which will be made to refer to conceptors later). This renders the categorical
treatment of ECL much simpler than it is for logics with sorted symbols of
varying arities.
• The operator symbols NOT, AND, OR, the parametrized operation symbol ϕ(·, γ) and the relation symbol ≤ are not made part of signatures, but
become universal elements in the construction of sentences.
The models of ECL are quite simple. For a signature Σ, the objects of Mod (Σ)
are the sets of Σ-indexed N -dimensional conceptor matrices
Ob(Mod (Σ)) = {m ⊂ CN ×N × Σ | ∀σ ∈ Σ ∃=1 C ∈ CN ×N : (C, σ) ∈ m}
150
where CN ×N is the set of all N -dimensional conceptor matrices. The objects of
Mod (Σ) are thus the graph sets of the functions from Σ to the set of N -dimensional
conceptor matrices. The model morphisms of Mod (Σ) are canonically given by
the index-preserving maps
hom({(C1 , σ)}, {(C2 , σ)}) = {µ : {(C1 , σ)} → {(C2 , σ)} | µ : (C1 , σ) 7→ (C2 , σ)}.
Clearly, hom({(C, σ)}, {(C 0 , σ)}) contains exactly one element.
φ
Given a signature morphism Σ1 → Σ2 , then Mod (φ) is defined to be a map
from Mod (Σ2 ) to Mod (Σ1 ) as follows. For a Σ2 -model m2 = {(C2 , σ2 )} let [[σ2 ]]m2
denote the interpretation of σ2 in m2 , that is, [[σ2 ]]m2 is the conceptor matrix C2
for which (C2 , σ2 ) ∈ m2 . Then Mod (φ) assigns to to m2 the Σ1 -model m1 =
Mod (φ)(m2 ) = {([[φ(σ1 )]]m2 , σ1 )} ∈ Mod (Σ1 ).
The model relations |=Σ are defined in the same way as in the familiar firstorder logic. Omitting some detail, here is how:
Definition 11 Preliminaries: A map β : Var → CN ×N is called a variable assignment. B is the set of all variable assignments. We denote by β xCi the variable
assignment that is identical to β except that xi is mapped to C. A Σ-interpretation
is a pair I = (m, β) consisting of a Σ-model m and a variable assignment β. By
I xCi we denote the interpretation (m, β xCi ). For a Σ-term t, the interpretation
I(t) ∈ CN ×N is defined in the obvious way. Then |=∗Σ ⊆ (Mod(Σ) × B) × Exp(Σ)
is defined inductively by
1. I |=∗Σ t1 ≤ t2
2. I |=∗Σ ¬e
iff
iff
I(t1 ) ≤ I(t2 ),
not I |=∗Σ e,
3. I |=∗Σ (e1 ∧ e2 )
iff
I |=∗Σ e1 and I |=∗Σ e2 ,
4. I |=∗Σ (e1 ∨ e2 )
iff
I |=∗Σ e1 or I |=∗Σ e2 ,
5. I |=∗Σ ∀xi e
iff
for all C ∈ CN ×N it holds that I xCi |=∗Σ e.
|=Σ then is the restriction of |=∗Σ on sentences.
This completes the definition of ECL as an institution. The satisfaction condition obviously holds. While in many respects ECL follows the role model of
first-order logic, the associated model theory is much more restricted in that only
N -dimensional conceptors are admitted as interpretations of symbols. The natural
next step would be to design calculi for ECL and investigate whether this logic
is complete or even decidable. Clarity on this point would amount to an insight
in the computational tractability of knowledge representation based on matrix
conceptors with Boolean and aperture adaptation opertors.
Intrinsic conceptor logic. I want to present ICL as a model of the “logics”
which might unfold inside a neural agent. All constituents of ICL should be
151
realizable in terms of neurodynamical processes, giving a logic not for reasoning
about conceptors, but with conceptors.
Taking the idea of placing “logics” inside an agent seriously has a number of
consequences which lead quite far away from traditional intuitions about “logics”:
• Different agents may have different logics. I will therefore not try to define
a general ICL that would fit any neural agent. Instead every concrete agent
with a concrete lifetime learning history will need his/her/its own individual
conceptor logic. I will use the signal de-noising and classification architecture
from Section 3.16 as an example “agent” and describe how an ICL can be
formulated as an institution for this particular case. Some general design
principles will however become clear from this case study.
• Conceptors are all about temporal processes, learning and adaptation. An
agent’s private ICL will have to possess an eminently dynamical character.
Concepts will change their meaning over time in an agent. This “personal
history dynamics” is quintessential for modeling an agent and should become
reflected in making an ICL a dynamical object itself – as opposed to introducing time through descriptive syntactical elements in an otherwise static
logic, like it is traditionally done by means of modal operators or axioms describing a timeline. In my proposal of ICLs, time enters the picture through
the central constituent of an institution, signature morphisms. These maps
between signatures (all commanded by the same agent) will model time, and
an agent’s lifetime history of adaptation will be modeled by an evolution of
signatures. Where the original core motif for casting logics as institutions
was that “truth is invariant under change of notation” ([37]), the main point
of ICLs could be contrasted as “concepts and their meaning change with
time”. The role of signature morphisms in ICLs is fundamentally different
in ICLs compared to customary formalizations of logics. In the latter, signature changes should leave meaning invariant; in the former, adaptive changes
in conceptors are reflected by temporally indexed changes in signature.
• Making an ICL private to an agent implies that the model relation |= becomes agent-specific. An ICL cannot be specified as an abstract object in
isolation. Before it can be defined, one first needs to have a formal model of
a particular agent with a particular lifelong adaptation history.
In sum, an ICL (formalized as institution) itself becomes a dynamical system,
defined relative to an existing (conceptor-based) neural agent with a particular
adaptation history. The “state space” of an ICL will be the set of signatures. A
“trajectory” of the temporal evolution of an ICL will essentially be a sequence of
signatures, enriched with information pertaining to forming sentences and models.
For an illustration, assume that a neural agent adapts two random feature conceptors a(n), b(n). These are named by two temporally indexed symbols A(n) , B (n) .
A signature will be a timeslice of these, Σ(n) = {A(n) , B (n) }. For every pair of
152
integer timepoints (n + k, n) (where k ≥ 0) there will be a signature morphism
φ(n+k,n) : Σ(n+k) → Σ(n) . The (strong) reason why signature morphisms point
backwards in time will become clear later. Figure 46 visualizes the components of
(n)
this example. The dotted lines connecting the Ai are suggestive graphical hints
(n)
that the symbols Ai all name the “same” conceptor ai . How this “sameness of
identity over time” can be captured in the institution formalism will become clear
presently.
Formal definitions of ICLs will vary depending on what kind of conceptors are
used (for instance, matrix or random feature based), or whether time is taken to
be discrete or continuous. I give a definition for discrete-time, random feature
conceptor based ICLs.
Because ICLs will be models of an agent’s private logic which evolves over the
agent’s lifetime, the definition of an ICL is stated relative to an agent’s lifetime
conceptor adaptation history. The only property of such an agent that is needed
for defining an ICL is the existence of temporally adapted conceptors owned by
the agent. Putting this into a formal definition:
Definition 12 An agent life (here: M -dimensional random feature conceptor
based, discrete time) is a structure AL = (T, Σ, ιAL ), where
1. T ⊆ Z is an interval (finite or infinite) of the integers, the lifetime of AL,
2. Σ = {A1 , . . . , Am } is a finite nonempty set of conceptor identifiers,
3. ιAL : Σ × T → [0, 1]M , (Ai , n) 7→ ai (n) assigns to every time point and
conceptor identifier an adaptation version ai (n) of the conceptor identified
by the symbol Ai .
As an example of an agent consider the signal de-noising and classification
architecture (DCA) presented in Section 3.16, with a “life” being the concrete
16000-step adaptation run illustrated in Figure 43. In this example, the lifetime
is T = {1, . . . , 16000}. I will identify by symbols the four prototype conceptors
auto
c1 , . . . , c4 and the two auto-adapted conceptors cauto
[1] , c[2] . Accordingly I choose Σ
to be {C1 , . . . , C4 , A1 , A2 }. The map ιAL is constant in time for the four protype
conceptors: ιAL (n, Cj ) = cj for all n ∈ T, j = 1, . . . , 4. For the remaining two
conceptors, ιAL (n, Ai ) = cauto
[i] (n).
The stage is now prepared to spell out the definition of an agent’s lifetime ICL
(for the case of an agent based on M -dimensional random feature conceptor and
discrete time):
Definition 13 The intrinsic conceptor logic (ICL) of an agent lifeAL = (T, Σ, ιAL )
is an institution whose components obey the following conditions:
(n)
(n)
1. The objects (signatures) of Sign are the pairs Σ(n) = ({A1 , . . . , Am }, σ (n) ),
(n)
(n)
where n ∈ T , and σ (n) : Σ → {A1 , . . . , Am } is a bijection.
153
DCA example: The lifetime of this example is T = {1, . . . , 16000}. A
(n)
(n)
(n)
(n)
signature Σ(n) = ({C1 , . . . , C4 , A1 , A2 }, σ (n) ) at time n ∈ T will later
be employed to denote some of the conceptors in the DCA in their adapted
versions at time n. These conceptor adaptation versions will thus become
identifiable by symbols from Σ(n) . For σ (n) I take the natural projection
(n)
(n)
Cj 7→ Cj , Ai 7→ Ai .
(n+k)
2. For every n, n + k ∈ T (where k ≥ 0), φ(n+k,n) : Σ(n+k) → Σ(n) , Ai
7→
(n+k)
(n)
(n+k) −1
(σ ◦ (σ
) )(Ai
) is a morphism in Sign. There are no other
morphisms in Sign besides these. Remark: At first sight this might seem
(n+k)
(n)
unneccessarily complicated. Why not simply require φ(n+k,n) : Ai
7→ Ai ?
(n)
(n)
The reason is that the set of symbols {A1 , . . . Am } of Σ(n) is just that, a
(n)
(n+k)
set of symbols. That over time Ai should correspond to Ai
is only
visually suggested to us, the mathematicians, by the chosen notation for
these symbols, but by no means does it actually follow from that notation.
3. Sen(Σ(n) ) is inductively defined as follows:
(n)
(n)
(a) A1 , . . . , Am and I and 0 are sentences in Sen(Σ(n) ).
(n)
(n)
(n)
(b) For k ≥ 0 such that n, n + k ∈ T , for Ai ∈ Σ(n) , δk Ai is in
Sen(Σ(n) ). Remark: the δ operators capture the temporal adaptation of
(n)
conceptors. The symbol Ai will be used to denote a conceptor ai (n)
(n) (n)
in its adaptation version at time n, and the sentence δk Ai will be
made to refer to ai (n + k).
(c) If ζ, ξ ∈ Sen(Σ(n) ), then (ζ ∨ ξ), (ζ ∧ ξ), ¬ζ ∈ Sen(Σ(n) ). Remark:
unlike in ECL there is no need for a notational distinction between ∧
and AND etc.
(d) If ζ ∈ Sen(Σ(n) ), then ϕ(ζ, γ) ∈ Sen(Σ(n) ) for every γ ∈ [0, ∞] (this
captures aperture adaptation).
(e) If ζ, ξ ∈ Sen(Σ(n) ) and 0 ≤ b ≤ 1, then βb (ζ, ξ) ∈ Sen(Σ(n) ) (this will
take care of linear blends bζ + (1 − b)ξ).
Remark: Including I and 0 in the sentence syntax is a convenience item. 0
∧
(n)
(n)
(n)
could be defined in terms of any Ai by 0 = (ϕ(Ai , ∞) ∧ ¬ϕ(Ai , ∞)),
∧
and I by I = ¬0. Likewise, ∨ (or ∧) could be dismissed because it can be
expressed in terms of ∧ and ¬ (∨ and ¬, respectively).
4. For a signature morphism φ(n+k,n) : Σ(n+k) → Σ(n) , Sen(φ(n+k,n) ) : Sen(Σ(n+k) ) →
Sen(Σ(n) ) is the map defined inductively as follows:
(a) Sen(φ(n+k,n) ) : I 7→ I, 0 7→ 0.
154
(n+k)
(n)
(n+k)
(b) Sen(φ(n+k,n) ) : Ai
7→ δk φ(n+k,n) (Ai
). Remark 1: When we use
(n)
(n)
the natural projections σ : Ai 7→ Ai , this rule could be more simply
(n+k)
(n) (n)
written as Sen(φ(n+k,n) ) : Ai
7→ δk Ai . Remark 2: This is the
pivotal point in this entire definition, and the point where the difference
to customary views on logics comes to the surface most conspicuously.
Usually signature morphisms act on sentences by simply re-naming all
signature symbols that occur in a sentence. The structure of a sentence
remains unaffected, in agreement with the motto “truth is invariant un(n+k)
der change of notation”. By contrast, here a signature symbol Ai
(n) (n)
is replaced by an temporal change operator term δk Ai , reflecting the
new motto “meaning changes with time”. The fact that φ(n+k,n) leads
(n+k)
(n)
from Ai
to Ai establishes “sameness of identity over time” be(n+k)
(n)
tween Ai
and Ai . Usually one would formally express sameness
of identity of some mathematical entity by using the same symbol to
name that entity at different time points. Here different symbols are
used, and thus another mechanism has to be found in order to establish
that an entity named by different symbols at different times remains
“the same”. The dotted “identity” lines in Figure 46 are fixed by the
signature morphisms φ(n+k,n) , not by using the same symbol over time.
Remark 3: At this point it also becomes clear why the signature morphisms φ(n+k,n) : Σ(n+k) → Σ(n) lead backwards in time. A conceptor
ai (n + k) in its time-(n + k) version can be expressed in terms of the
earlier version ai (n) with the aid of the temporal evolution operator δ,
but in general an earlier version ai (n) cannot be expressed in terms of
a later ai (n + k). This reflects the fact that, seen as a trajectory of an
input-driven dynamical system, an agent life is (typically) irreversible.
To put it into everyday language, “the future can be explained from the
past, but not vice versa”.
(n+k)
(c) Sen(φ(n+k,n) ) : δl
(n+k)
Ai
(n)
(n+k)
7→ δ(k+l) φ(n+k,n) (Ai
).
(d) For ζ, ξ ∈ Sen(Σ(n+k) ), put
Sen(φ(n+k,n) ) :
(ζ ∨ ξ)
(ζ ∧ ξ)
¬ζ
ϕ(ζ, γ)
βb (ζ, ξ)
7→
7
→
7→
7
→
7→
(Sen(φ(n+k,n) )(ζ) ∨ Sen(φ(n+k,n) )(ξ)),
(Sen(φ(n+k,n) )(ζ) ∧ Sen(φ(n+k,n) )(ξ)),
¬Sen(φ(n+k,n) )(ζ),
ϕ(Sen(φ(n+k,n) )(ζ), γ),
βb (Sen(φ(n+k,n) )(ζ), Sen(φ(n+k,n) )(ξ)).
5. For every signature Σ(n) ∈ Sign, Mod(Σ(n) ) is always the same category
Z with objects all non-negative M -dimensional vectors z .∧2 . There are no
id
model morphisms except the identity morphisms z .∧2 → z .∧2 .
155
6. For every morphism φ(n+k,n) ∈ Sign, Mod(φ(n+k,n) ) is the identity morphism
of Z.
7. As a preparation for defining the model relationships |=Σ(n) we assign by
induction to every sentence χ ∈ Sen(Σ(n) ) an M -dimensional conception
weight vector ι(χ) as follows:
(a) ι(I) = (1, . . . , 1)0 and ι(0) = (0, . . . , 0)0 .
(n)
(n)
(b) ι(Ai ) = ιAL (n, (σ (n) )−1 Ai ).
(n)
(n)
(n)
(c) ι(δk Ai ) = ιAL (n + k, (σ (n) )−1 Ai ).
(d) Case χ = (ζ ∨ ξ): ι(χ) = ι(ζ) ∨ ι(ξ) (compare Definition 7).
(e) Case χ = (ζ ∧ ξ): ι(χ) = ι(ζ) ∧ ι(ξ).
(f ) Case χ = ¬ζ: ι(χ) = ¬ι(ζ).
(g) Case χ = ϕ(ζ, γ): ι(χ) = ϕ(ι(ζ), γ) (compare Definition 6).
(h) Case χ = βb (ζ, ξ): ι(χ) = b ι(ζ) + (1 − b)ι(ξ).
8. For z .∧2 ∈ Z = Mod(Σ(n) ) and χ ∈ Sen(Σ(n) ), the model relationship is
defined by
z .∧2 |=Σ(n) χ
z
iff
z .∧2 .∗ (z .∧2 + 1).∧−1 ≤ ι(χ).
The satisfaction condition (110) requires that for χ(n+k) ∈ Sen(Σ(n+k) ) and
∈ Mod(Σ(n) ) it holds that
.∧2
z .∧2 |=Σ(n) Sen(φ(n+k,n) )(χ(n+k) ) iff Mod(φ(n+k,n) )(z .∧2 ) |=Σ(n+k) χ(n+k) ,
which is equivalent to
z .∧2 |=Σ(n) Sen(φ(n+k,n) )(χ(n+k) ) iff z .∧2 |=Σ(n+k) χ(n+k)
(111)
because Mod(φ(n+k,n) ) is the identity morphism on Z. This follows directly from
the following fact:
Lemma 1 ι Sen(φ(n+k,n) )(χ(n+k) ) = ι(χ(n+k) ),
which in turn can be established by an obvious induction on the structure of
(n+k)
sentences, where the crucial steps are the cases (i) χ(n+k) = Ai
and (ii) χ(n+k) =
(n+k) (n+k)
δl
Ai
.
(n+k)
(n)
(n+k)
(n)
In case (i), ι(Sen(φ(n+k,n) )(Ai
)) = ι(δk φ(n+k,n) (Ai
)) = ι(δk (σ (n) ◦
(n+k)
(n+k)
(σ (n+k) )−1 )(Ai
)) = ιAL (n + k, (σ (n) ))−1 (σ (n) ◦ (σ (n+k) )−1 )(Ai
)) = ιAL (n +
(n+k)
(n+k)
(n+k) −1
k, (σ
) (Ai
)) = ι(Ai
).
156
(n+k)
In case (ii), conclude ι(Sen(φ(n+k,n) )(δl
(n)
ι(δ(k+l)
(σ
(n+k)
Ai
(n)
(n+k)
)) = ι(δ(k+l) φ(n+k,n) (Ai
)) =
(n+k)
(n+k)
◦(σ
) )(Ai
)) = ιAL (n+k+l, (σ (n) ))−1 (σ (n) ◦(σ (n+k) )−1 )(Ai
))
(n+k)
(n+k) (n+k)
+ l, (σ (n+k) )−1 (Ai
)) = ι(δl
Ai
).
(n)
(n+k) −1
ιAL (n + k
An important difference between “traditional” logics and the ICL of an agent
AL concerns different intuitions about semantics. Taking first-order logic as an
example of a traditional logic, the “meaning” of a first-order sentence χ with
signature Σ is the class of all of its models. Whether a set is a model of χ depends
on how the symbols from Σ are extensionally interpreted over that set. Firstorder logic by itself does not prescribe how the symbols of a signature have to be
interpreted over some domain. In contrast, an ICL is defined with respect to a
concrete agent AL, which in turn uniquely fixes how the symbols from an ICL
signature Σ(n) must be interpreted – this is the essence of points 7. (a – c) in
Definition 13.
Logical entailment in an ICL coincides with abstraction of conceptors:
Proposition 17 In an ICL, for all ζ, ξ ∈ Sen(Σ(n) ) it holds that
ζ |=Σ(n) ξ
iff
ι(ζ) ≤ ι(ξ).
(112)
The simple proof is given in Section 5.11. In an agent’s lifetime ICL, for
any sentence χ ∈ Sen(Σ(n) ) the concrete interpretation ι(χ) ∈ [0, 1]M can be
effectively computed via the rules stated in Nr. 7 in Definition 13, provided one
has access to the identifiable conceptors ai (n) in the agent’s life. Since for two
vectors a, b ∈ [0, 1]M it can be effectively checked whether a ≤ b, it is decidable
whether ι(ζ) ≤ ι(ξ). An ICL is therefore decidable.
Seen from a categorical point of view, an ICL is a particularly small-size institution. Since the (only) category in the image of Mod is a set, we can regard
the functor Mod as having codomain Setop instead of Catop . Also the category
Sign is small, that is, a set. Altogether, an ICL institution nowhere needs proper
classes.
The ICL definition I gave here is very elementary. It could easily be augmented
in various natural ways, for instance by admitting permutations and/or projections
of conceptor vector components as model morphisms in Z, or allowing conceptors
of different dimensions in the makeup of an agent life. Likewise it is straightforward
to spell out definitions for an agent life and its ICL for matrix-based conceptors.
In the latter case, the referents of sentences are correlation matrices R, and the
defining equation for the model relationship appears as
R |=Σ(n) χ iff R(R + I)−1 ≤ ι(χ).
In traditional logics and their applications, an important role is played by calculi. A calculus is a set of syntactic transformation rules operating on sentences
(and expressions containing free variables) which allows one to derive purely syntactical proofs of logical entailment statements ζ |= ξ. Fundamental properties of
157
=
familiar logics, completeness and decidability in particular, are defined in terms
of calculi. While it may be possible and mathematically interesting to design
syntactical calculi for ICLs, they are not needed because ζ |= ξ is decidable via
the semantic equivalent ι(ζ) ≤ ι(ξ). Furthermore, the ICL decision procedure
is computationally very effective: only time O(1) is needed (admitting a parallel
execution) to determine whether some conception vector is at most as large as
another in all components. This may help to explain why humans can so quickly
carry out many concept subsumption judgements (“this looks like a cow to me”).
3.19
Final Summary and Outlook
Abstracting from all technical detail, here is a summary account of the conceptor
approach:
From neural dynamics to conceptors. Conceptors capture the shape of a neural state cloud by a positive semi-definite operator.
From conceptors to neural dynamics. Inserting a conceptor into a neurodynamical state update loop allows to select and stabilize a previously stored
neural activation pattern.
Matrix conceptors are useful in machine learning applications and as mathematical tools for analysing patterns emerging in nonlinear neural dynamics.
Random feature conceptors are not biologically apriori implausible and can
be neurally coded by single neurons.
Autoconceptor adaptation dynamics leads to content-addressable neural memories of dynamical patterns and to signal filtering and classification systems.
Boolean operations and the abstraction ordering on conceptors establish
a bi-directional connection between logic and neural dynamics.
From static to dynamic models. Conceptors allow to understand and control
dynamical patterns in scenarios that previously have been mostly restricted
to static patterns. Specifically this concerns content-addressable memories,
morphing, ordering concepts in logically structured abstraction hierarchies,
and top-down hypothesis control in hierarchical architectures.
The study of conceptors is at an early stage. There are numerous natural
directions for next steps in conceptor research:
Affine conceptor maps. Conceptors constrain reservoir states by applying a
positive semi-definite map. This map is adapted to the “soft-bounded” linear
subspace visited by a reservoir when it is driven through a pattern. If the
pattern-driven reservoir states do not have zero mean, it seems natural to
158
first subtract the mean before applying the conceptor. If the mean is µ, this
would result in an update loop of the form x(n + 1) = µ + C(tanh(...) − µ).
While this may be expected to improve control characteristics of pattern
re-generation, it is not immediately clear how Boolean operations transfer
to such affine conceptors which are characterized by pairs (C, µ).
Nonlinear conceptor filters. Even more generally, one might conceive of conceptors which take the form of nonlinear filters, for instance instantiated as
feedforward neural networks. Such filters F could be trained on the same objective function as I used for conceptors, namely minimizing E[kz−F (z)k2 ]+
α−2 kF k2 , where kF k is a suitably chosen norm defined for the filter. Like
with affine conceptor maps, the pattern specificity of such nonlinear conceptor filters would be greater than for our standard matrix conceptors, but
again it is not clear how logical operations would extend to such filters.
Basic mathematical properties. From a mathematics viewpoint, there are some
elementary questions about conceptors which should be better understood,
for instance:
• What is the relationship between Boolean operations, aperture adaptation, and linear blending? in particular, can the latter be expressed
in terms of the two former?
• Given a dimension N , what is the minimial number of “basis” conceptors such that the transitive closure under Boolean operations, aperture
adaptation, and/or linear blending is the set of all N -dimensional conceptors?
• Are there normal forms for expressions which compose conceptors by
Boolean operations, aperture adaptation, and/or linear blending?
• Find conceptor analogs of standard structure-building mathematical
operations, especially products. These would be needed to design architectures with several conceptor modules (likely of different dimension)
where the “soft subspace constraining” of the overall dynamics works
on the total architecture state space. Presumably this leads to tensor
variants of conceptors. This may turn out to be a challenge because
a mathematical theory of “semi positive-definite tensors” seems to be
only in its infancy (compare [86]).
Neural realization of Boolean operations. How can Boolean operations be
implemented in biologically not impossible neural circuits?
Complete analysis of autoconceptor adaptation. The analysis of autoconceptor adaptation in Section 3.14.4 is preliminary and incomplete. It only
characterizes certain aspects of fixed-point solutions of this adaptation dynamics but remains ignorant about the effects that the combined, nonlinear
159
conceptor-reservoir update dynamics may have when such a fixed-point solution is perturbed. Specifically, this reservoir-conceptor interaction will have
to be taken into account in order to understand the dynamics in the center
manifold of fixed-point solutions.
Applications. The usefulness of conceptors as a practical tool for machine learning and as a modeling tool in the cognitive and computational neurosciences
will only be established by a suite of successful applications.
160
4
Documentation of Experiments and Methods
In this section I provide details of all simulation experiments reported in Sections
1 and 3.
4.1
General Set-Up, Initial Demonstrations (Section 1 and
Section 3.2 - 3.4)
A reservoir with N = 100 neurons, plus one input unit and one output unit was
created with a random input weight vector W in , a random bias b and preliminary
reservoir weights W ∗ , to be run according to the update equations
x(n + 1) = tanh(W ∗ x(n) + W in p(n + 1) + b),
y(n) = W out x(n).
(113)
Initially W out was left undefined. The input weights were sampled from a
normal distribution N (0, 1) and then rescaled by a factor of 1.5. The bias was
likewise sampled from N (0, 1) and then rescaled by 0.2. The reservoir weight matrix W ∗ was first created as a sparse random matrix with an approximate density
of 10%, then scaled to obtain a spectral radius (largest absolute eigenvalue) of 1.5.
These scalings are typical in the field of reservoir computing [69] for networks to
be employed in signal-generation tasks.
For each of the four driving signals pj a training time series of length 1500
was generated. The reservoir was driven with these pj (n) in turn, starting each
run from a zero initial reservoir state. This resulted in reservoir state responses
xj (n). The first 500 steps were discarded in order to exclude data influenced by
the arbitrary starting condition, leading to four 100-dimensional reservoir state
time series of length L = 1000, which were recorded into four 100 × 1000 state
collection matrices X j , where X j (:, n) = xj (n + 500) (j = 1, . . . , 4). Likewise,
the corresponding driver signals were recorded into four pattern collection (row)
vectors P j of size 1 × 1000. In addition to this, a version X̃ j of X j was built,
identical to X j except that it was delayed by one step: X̃ j (:, n) = xj (n + 499).
These collections were then concatenated to obtain X = [X 1 |X 2 |X 3 |X 4 ], X̃ =
[X̃ 1 |X̃ 2 |X̃ 3 |X̃ 4 ], P = [P 1 |P 2 |P 3 |P 4 ].
The “PC energy” plots in Figure 12 render the singular values of the correlation
matrices X j (X j )0 /L.
The output weights W out were computed as the regularized Wiener-Hopf solution (also known as ridge regression, or Tychonov regularization)
W out = ((XX 0 + %out IN ×N )−1 X P 0 )0 ,
(114)
where the regularizer %out was set to 0.01.
Loading: After loading, the reservoir weights W should lead to the approximate
equality W xj (n) ≈ W ∗ xj (n) + W in pj (n + 1), across all patterns j, which leads to
161
the objective of minimizing the squared error j (n + 1) = ((tanh−1 (xj (n + 1)) −
b) − W xj (n))2 , averaged over all four j and training time points. Writing B for
the 100 × (4 ∗ 1000) matrix whose columns are all identical equal to b, this has the
ridge regression solution
W = ((X̃ X̃ 0 + %W IN ×N )−1 X̃ (tanh−1 (X) − B))0 ,
(115)
where the regularizer %W was set to 0.0001. To assess the accuracy of the weight
computations, the training normalized root mean square error (NRMSE) was computed. For the readout weights, the NRMSE between y(n) = W out x(n) and the
target P was 0.00068. For the reservoir weights, the average (over reservoir neurons i, times n and patterns j) NRMSE between W (i, :)xj (n) and the target
W ∗ (i, :) xj (n) + W in (i) pj (n + 1) was 0.0011.
In order to determine the accuracy of fit between the original driving signals
pj and the network observation outputs y j (n) = W out C j tanh(W x(n − 1) + b) in
the conceptor-constrained autonomous runs, the driver signals and the y j signals
were first interpolated with cubic splines (oversampling by a factor of 20). Then
a segment length 400 of the oversampled driver (corresponding to 20 timesteps
before interpolation) was shifted over the oversampled y j in search of a position of
best fit. This is necessary to compensate for the indeterminate phaseshift between
the driver data and the network outputs. The NRMSEs given in Figure 12 were
calculated from the best-fit phaseshift position, and the optimally phase-shifted
version of y j was also used for the plot.
4.2
Aperture Adaptation (Sections 3.8.3 and 3.8.4)
Data generation. For the Rössler attractor, training time series were obtained
from running simple Euler approximations of the following ODEs:
ẋ = −(y + z)
ẏ = x + a y
ż = b + x z − c z,
using parameters a = b = 0.2, c = 8. The evolution of this system was Euler
approximated with stepsize 1/200 and the resulting discrete time series was then
subsampled by 150. The x and y coordinates were assembled in a 2-dimensional
driving sequence, where each of the two channels was shifted/scaled to a range of
[0, 1]. For the Lorenz attractor, the ODE
ẋ = σ(y − x)
ẏ = r x − y − x z
ż = x y − b z
162
with σ = 10, r = 28, b = 8/3 was Euler-approximated with stepsize 1/200 and
subsequent subsampling by 15. The x and z coordinates were collected in a 2dimensional driving sequence, again each channel normalized to a range of [0, 1].
The Mackey Glass timeseries was obtained from the delay differential equation
ẋ(t) =
β x(t − τ )
− γ x(t)
1 + x(t − τ )n
with β = 0.2, n = 10, τ = 17, γ = 0.1, a customary setting when this attractor
is used in neural network demonstrations. An Euler approximation with stepsize
1/10 was used. To obtain a 2-dim timeseries that could be fed to the reservoir
through the same two input channels as the other attractor data, pairs x(t), x(t−τ )
were combined into 2-dim vectors. Again, these two signals were normalized to
the [0, 1] range. The Hénon attractor is governed by the iterated map
x(n + 1) = y(n) + 1 − a x(n)
y(n + 1) = b x(n),
where I used a = 1.4, b = 0.3. The two components were filed into a 2-dim
timeseries (x(n), y(n))0 with no further subsampling, and again normalization to
a range of [0, 1] in each component.
Reservoir setup. A 500-unit reservoir RNN was created with a normal distributed, 10%-sparse weight matrix W ∗ scaled to a spectral radius of 0.6. The
bias vector b and input weights W in (sized 400 × 2 for two input channels) were
sampled from standard normal distribution and then scaled by 0.4 and 1.2, respectively. These scaling parameters were found by a (very coarse) manual optimization of the performance of the pattern storing process. The network size was
chosen large enough to warrant a robust trainability of the four chaotic patterns.
Repeated executions of the experiment with different randomly initialized weights
(not documented) showed no significant differences.
Pattern storing. The W ∗ reservoir was driven, in turn, by 2500 timesteps of
each of the four chaotic timeseries. The first 500 steps were discarded to account
for initial reservoir state washout, and the remaining 4 × 2000 reservoir states
were collected in a 500 × 8000 matrix X. From this, the new reservoir weights W
were computed as in (115), with a regularizer %W = 0.0001. The readout weights
were computed as in (114) with regularizer %out = 0.01. The average NRMSEs
obtained for the reservoir and readout weights were 0.0082 and 0.013, respectively.
Computing conceptors. From each of the four n = 2000 step reservoir state
sequences X recorded in the storing procedure, obtained from driving the reservoir with one of the four chaotic signals, a preliminary correlation matrix R̃ =
XX 0 /2000 and its SVD U S̃U 0 = R̃ were computed. This correlation matrix was
then used to obtain a conceptor associated with the respective chaotic signal, using
an aperture α = 1. From these unit-aperture conceptors, versions with differing
α were obtained through aperture adaptation per (16).
163
In passing I note that the overall stability and parameter robustness of this simulation can be much improved if small singular values in S̃ (for instance, with values smaller than 1e-06) are zeroed, obtaining a clipped S, from which a “cleanedup” correlation matrix R = U SU 0 would be computed. This would lead to a range
of well-working apertures spanning three orders of magnitude (not shown). I did
not do this in the reported simulation in order to illustrate the effects of too large
apertures; these effects would be partly suppressed when the spectrum of R is
clipped.
Pattern retrieval and plotting. The loaded network was run using the conceptorconstrained update rule x(n+1) = C tanh(W x(n)+W in p(n+1)+b) with various
C = ϕ(C# , γ#,i ) (# = R, L, MG, H, i = 1, . . . , 5) for 800 steps each time, of which
the first 100 were discarded to account for initial state washout. The delay embedding plots in Figure 18 were generated from the remaining 700 steps. Embedding
delays of 2, 2, 3, 1 respectively were used for plotting the four attractors.
For each of the four 6-panel blocks in Figure 18, the five aperture adaptation
factors γ#,i were determined in the following way. First, by visual inspection, the
middle γ#,3 was determined to fall in the trough center of the attenuation plot of
Fig. 19 A. Then the remaining γ#,1 , γ#,2 , γ#,4 , γ#,5 were set in a way that (i) the
entire γ sequence was a geometrical progression, and (ii) that the plot obtained
from the first γ#,1 was visually strongly corrupted.
4.3
Memory Management, Demo 1 (Section 3.11.2)
Data: The patterns were either sines sampled at integer fractions of 2π or random
periodic patterns, always scaled to a range of [−0.9, 0.9]. Period lengths were
picked randomly between 3 and 15. The length of pattern signals used for loading
was L = 100, with an additional washout of the same length.
Parameters: Reservoir size N = 100, W ∗ sparse with density approximately
10%, weights sampled from a standard normal distribution, rescaled to a spectral
radius of 1.5. The input weights W in were non-sparse, sampled from a standard
normal distribution, scaled by 1.5. The bias b was sampled from a standard normal
distribution and scaled by 0.25. Ridge regression coefficients 0.01 and 0.001 for
computing the increments to W out and D, respectively. Data from 100 reservoir
update steps were used for computing conceptors and regressions.
Computing test errors: The conceptors C j obtained during the loading procedure were used with an aperture α = 1000 by running the reservoir via x(n + 1) =
C j tanh(W ∗ x(n) + D x(n) + b), starting from random initial states. After a
washout period of 200 steps, the network outputs y j (n) were recorded for 50
steps. A 20-step portion of the original driver pattern was interpolated (supersampling rate 10) and shifted over a likewise interpolation version of these outputs.
The best matching shift position was used for plotting and NRMSE computation.
Such shift-search for a good fit is necessary because the autonomous runs are not
phase-synchronized with the original drivers.
164
Non-incremental loading: For the standard non-incremental loading experiment that served as comparison and yielded the results marked by dots in the left
panel of Figure 22, the same parameters (scalings, aperture) were used as for the
incremental loading procedure.
4.4
Memory Management, Demo 2 (Section 3.12.1)
The set-up is the same as for the integer-periodic basic demo. Scalings: spectral
radius of W ∗ : 1.5; W in : 1.5; b: 1.0. Aperture: 10. Ridge regression coefficients
0.02 both for increments to W out and D. Data collection runlengths in loading:
500 (plus 200 washout). Test runlength: 800 (plus 200 washout). Test output was
compared to 20-step sample from original driver after interpolation and shift as
described for the integer-periodic basic demo.
4.5
Memory Management, Close-Up Inspection (Section
3.12.2)
Integer-periodic patterns, period length 6 (condition IP6): Scalings of reservoirs:
spectral radius of W ∗ : 1.5; W in : 1.5; b: 0.5. Aperture: 1000. Ridge regression
coefficients 0.01 for W out and 0.001 for D. Data collection runlengths in loading:
100 (plus 100 washout) per pattern. Pattern recall tests were started from random
network states and recorded for 50 steps after a washout of 200. Recall NRMSEs
were computed as described in Section 4.3.
Irrational-period patterns from a parametric family (condition PF): Scalings
of reservoirs: spectral radius of W ∗ : 1.5; W in : 1.5; b: 1.0. Aperture: 10. Ridge
regression coefficients 0.02 both for W out and D. Data collection runlengths in
loading: 100 (plus 100 washout) per pattern. Pattern recall tests were started
from random network states and recorded for 80 steps after a washout of 200.
Integer-periodic patterns, period 3 (condition IP3): Same as in condition IP6.
4.6
Memory Management, Arbitrary Patterns (Section 3.12.3)
Spectral radius of W ∗ : 1.5; W in : 1.5; b: 0.25. Apertures of “raw” conceptors C m
(before “rectangularization”) were set to 10 for integer-periodic patterns and to 1.2
for the quasiperiodic patterns. Ridge regression coefficients: 0.001 for increments
to W out and 0.005 for increments to D. Data collection runlengths in loading: 500
(plus 100 washout). Test runlength: 200 (plus 100 washout), compared to 20-step
sample from original driver as described in Section 4.3. — For the standard nonincremental loading procedure that served as comparison (dots in Figure 26, left
panel), manually optimized conceptor apertures of 1000 — 10 were used for the
integer-periodic — quasi-periodic patterns.
165
4.7
Content-Addressable Memory (Section 3.14.3)
Network setup. Reservoir network matrices W were sampled sparsely (10% nonzero
weights) from a normal distribution, then scaled to a spectral radius of 1.5 in all
experiments of this section. Reservoir size was N = 100 for all period-4 experiments and the unrelated patterns experiment, and N = 200 for all experiments
that used mixtures-of-sines patterns. For all experiments in the section, input
weights W in were randomly sampled from the standard normal distribution and
rescaled by a factor of 1.5. The bias vector b was likewise sampled from the standard normal distribution and rescaled by 0.5. These scaling parameters had been
determined by a coarse manual search for a well-working configuration when this
suite experiments was set up. The experiments are remarkably insensitive to these
parameters.
Storing patterns. The storage procedure was set up identically for all experiments in this section. The reservoir was driven by the k loading patterns in turn
for l = 50 steps (period-4 and unrelated patterns) or l = 500 steps (mix-of-sines)
time steps, plus a preceding 100 step initial washout. The observed network states
x(n) were concatenated into a N × kl sized state collection matrix X, and the onestep earlier states x(n−1) into a matrix X̃ of same size. The driver pattern signals
were concatenated into a kl sized row vector P . The readout weights W out were
then obtained by ridge regression via W out = ((XX 0 + 0.01 I)−1 XP 0 )0 , and D by
D = ((X̃ X̃ 0 + 0.001 I)−1 X̃(W in P )0 )0 .
Quality measurements. After the cueing, and at the end of the recall period (or
at the ends of the three interim intervals for some of the experiments), the current
conceptor C was tested for retrieval accuracy as follows. Starting from the current
network state x(n), the reservoir network was run for 550 steps, constrained by C.
The first 50 steps served as washout and were discarded. The states x from the
last 500 steps were transformed to patterns by applying W out , yielding a 500-step
pattern reconstruction. This was interpolated with cubic splines and then sampled
at double resolution, leading to a 999-step pattern reconstruction ỹ.
A 20-step template sample of the original pattern was similarly interpolatedresampled and then passed over ỹ, detecting the best-fitting position where the
the NRMSE between the target template and ỹ was minimal (this shift-search
accomodated for unknown phase shifts between the target template and ỹ). This
minimal NRMSE was returned as measurement result.
Irrational-period sines. This simulation was done exactly as the one before,
using twelve irrational-period sines as reference patterns.
4.8
The Japanese Vowels Classification (Section 3.13)
Network setup. In each of the 50 trials, the weights in a fully connected, 10 × 10
reservoir weight matrix, a 12-dimensional input weight vector, a 10-dimensional
bias vector, and a 10-dimensional start state were first sampled from a normal
distribution, then rescaled to a spectral radius of 1.2 for W , and by factors of 0.2,
166
1, 1 for W in , b, xstart respectively.
Numerical determination of best aperture. To determine γi+ , the quantities
kϕ(C̃i+ , 2g )k2fro were computed for g = 0, 1, . . . , 8. These values were interpolated
on a 0.01 raster with cubic splines, the support point gmax of the maximum of the
interpolation curve was detected, returning γi+ = 2gmax .
Linear classifier training. The linear classifier that serves as a baseline comparison was designed in essentially the same way as the Echo State Networks based
classifiers which in [57] yielded zero test misclassifications (when combining 1,000
such classifiers made from 4-unit networks) and 2 test misclassifications (when a
single such classifier was based on a 1,000 unit reservoir), respectively. Thus, linear classifiers based on reservoir responses outperform all other reported methods
on this benchmark and therefore provide a substantive baseline information.
In detail, the linear classifier was learnt from 270 training data pairs of the form
(z, yteacher ), where the z were the same 88-dimensional vectors used for constructing
conceptors, and the yteacher were 9-dimensional, binary speaker indicator vectors
with a “1” in the position of the speaker of z. The classifier consists in a 9 ×
88 sized weight matrix V , and the cost function was the quadratic error kV z −
yteacher k2 . The classifier weight matrix V which minimized this cost function on
average over all training samples was computed by linear regression with Tychonov
regularization, also known as ridge regression [108]. The Tychonov parameter
which determines the degree of regularization was determined by a grid search over
a 5-fold cross-validation on the training data. Across the 50 trials it was found
to vary quite widely in a range between 0.0001 and 0.25; in a separate auxiliary
investigation it was also found that the effect of variation of the regularizer within
this range was very small and the training of the linear classifier can therefore be
considered robust.
In testing, V was used to determine classification decisions by computing ytest =
V ztest and opting for the index of the largest entry in ytest as the speaker.
4.9
Conceptor Dynamics Based on RFC Conceptors (Section 3.15)
Network setup. In both experiments reported in this section, the same reservoir
made from N = 100 units was used. F and G were full matrices with entries
first sampled from the standard normal distribution. Then they were both scaled
by an identical factor a such that the product a2 G F attained a spectral radius
of 1.4. The input weight vector W in was sampled from the standard normal
distribution and then scaled by 1.2. The bias b was likewise sampled from the
normal distribution and then scaled by 0.2.
These values were determined by coarse manual search, where the main guiding
criterion was recall accuracy. The settings were rather robust in the first experiment which used stored cj . The spectral radius could be individually varied from
0.6 to 1.45, the input weight scaling from 0.3 to 1.5, the bias scaling from 0.1 to
167
0.4, and the aperture from 3 to 8.5, while always keeping the final recall NRMSEs
for all four patterns below 0.1. Furthermore, much larger combined variations of
these scalings were also possible (not documented).
In the second experiment with content-addressed recall, the functional parameter range was much narrower. Individual parameter variation beyond ±5% was
disruptive. Specifically, I observed a close inverse coupling between spectral radius
and aperture: if one of the two was raised, the other had to be lowered.
Loading procedure. The loading procedure is described in some detail in the
report text. The mean NRMSEs on training data was 0.00081 for recomputing G,
0.0011 for D, and 0.0029 for W out . The mean absolute size of matrix elements in
G was 0.021, about a third of the mean absolute size of elements of G∗ .
Computing NRMSEs. The NRMSE comparison between the re-generated patterns at the end of the c adaptation and the original drivers was done in the
same way as reported on earlier occasions (Section 4.7), that is, invoking spline
interpolation of the comparison patterns and optimal phase-alignment.
4.10
Hierarchical Classification and Filtering Architecture
(Section 3.16)
Module setup. The three modules are identical copies of each other. The reservoir
had N = 100 units and the feature space had a dimension of M = 500. The
input weight matrix W in was sampled from the standard normal distribution and
rescaled by a factor of 1.2. The reservoir-featurespace projection and backprojection matrices F, G∗ (sized N × M ) were first sampled from the standard normal
distribution, then linearly rescaled by a common factor such that the N × N matrix G∗ F 0 (which functionally corresponds to an internal reservoir weight matrix)
had a spectral radius of 1.4. The bias b was likewise sampled from the standard
normal distribution and then scaled by 0.2.
A regularization procedure was then applied to G∗ to give G as follows. The
preliminary module was driven per
z(n + 1) = F 0 r(n),
r(n + 1) = tanh(G∗ z(n + 1) + W in u(n) + b),
with an i.i.d. input signal u(n) sampled uniformly from [−1, 1], for 1600 steps (after
discarding an initial washout). The values obtained for z(n + 1) were collected
as columns in a M × 1600 matrix Z. The final G was then computed by a ridge
regression with a regularizer a = 0.1 by
G = ((ZZ 0 + aI)−1 Z (G∗ Z)0 )0 .
In words, G should behave as the initially sampled G∗ in a randomly driven module, but do so with minimized weight sizes. This regularization was found to be
important for a stable working of the final architecture.
Training. The input simulation weights D (size 1 × M ) and W out (size 1 × N )
were trained by driving a single module with the four target patterns, as follows.
168
The module was driven with clean p1 , . . . , p4 in turn, with auto-adaptation of
conception weights c activated:
z(n + 1) = c(n) .∗ F 0 r(n),
r(n + 1) = tanh(G z(n + 1) + W in pj (n) + b),
c(n + 1) = c(n) + λc (z(n + 1) − c(n) .∗ z(n + 1)) .∗ z(n + 1) − α−2 c(n) ,
where the c adaptation rate was set to λc = 0.5, and an aperture α = 8 was used.
After discarding initial washouts in each of the four driving conditions (long enough
for c(n) to stabilize), 400 reservoir state vectors r(n + 1), 400 z vectors z(n) and
400 input values pj (n) were collected for j = 1, . . . , 4, and collected column-wise in
matrices R (size N × 1600), Z (size M × 1600) and Q (size 1 × 1600), respectively.
In addition, the 400-step submatrices Z j of Z containing the z-responses of the
module when driven with pj were registered separately.
The output weights were then computed by ridge regression on the objective
to recover pj (n) from r(n + 1) by
0
W out = (RR0 + aI)−1 RQ0 ,
using a regularizer a = 0.1. In a similar way, the input simulation weights were
obtained as
0
D = (ZZ 0 + aI)−1 ZQ0
with a regularizer a = 0.1 again. The training NRMSEs for W out and D were
0.0018 and 0.0042, respectively.
The M × 4 prototype matrix P was computed as follows. First, a preliminary version P ∗ was constructed whose j-the column vector was the mean of the
element-wise squared column vectors in Z j . The four column vectors of P ∗ were
then normalized such that the norm of each of them was the mean of the norms
of columns in P ∗ . This gave P . This normalization is important for the performance of the architecture, because without it the optimization criterion (104)
would systematically lead to smaller values for those γ j that are associated with
smaller-norm columns in P .
Baseline linear filter. The transversal filter that served as a baseline was a row
vector w of size 2600. It was computed to minimize the loss function
4
1 X
E[pj (n + 1) − (pj (n − 2600 + 1), . . . , pj (n))2 ] + a2 kwk2 ,
L(w) =
4 j=1
where pj (n) were clean versions of the four patterns. 400 timesteps per pattern
were used for training, and a was set to 1.0. The setting of a was very robust.
Changing a in either direction by factors of 100 changed the resulting test NRMSEs
at levels below the plotting accuracy in Figure 43.
Parameter settings in testing. The adaptation rate λγ was set to 0.002 for the
classification simulation and to 0.004 for the morph-tracking case study. The other
169
global control parameters were identical in both simulations: trust smoothing rate
σ = 0.99, decisiveness d[12] = d[23] = 8, drift d = 0.01, caut
[l] adaptation rate λ = 0.5
(compare Equation (90); I used the same adaptation rate λi ≡ λ for all of the 500
feature units).
Computing and plotting running NRMSE estimates. For the NRMSE plots in
the fifth rows in Figures 43 and 44, a running estimate of the NRMSE between
the module outputs y[l] and the clean input patterns p (unknown to the system)
was computed as follows. A running estimate var p(n) of the variance of the clean
pattern was maintained like it was done for var y[l] (n) in (98) and (99), using an
exponential smoothing rate of σ = 0.95. Then the running NRMSE was computed
by another exponential smoothing per
nrmse y[l] (n + 1) = σ nrmse y[l] (n) + (1 − σ)
(p(n + 1) − y[l] (n + 1))2
var p(n + 1)
1/2
.
The running NRMSE for the baseline transversal filter were obtained in a similar
fashion.
170
5
5.1
Proofs and Algorithms
Proof of Proposition 1 (Section 3.4)
Claim 1. We first re-write the minimization quantity, using R = E[xx0 ]:
E[kx − Cxk2 ] + α−2 kCk2fro =
= E[tr (x0 (I − C 0 )(I − C)x)] + α−2 kCk2fro
= tr ((I − C 0 )(I − C)R + α−2 C 0 C)
= tr (R − C 0 R − CR + C 0 C(R + α−2 I))
X
=
e0i (R − C 0 R − CR + C 0 C(R + α−2 I))ei .
i=1,...,N
This quantity is quadratic in the parameters of C and non-negative. Because
tr C 0 C(R + α−2 I) = tr C(R + α−2 I)C 0 and R + α−2 I is positive definite, tr C(R +
α−2 I)C 0 is positive definite in the N 2 -dimensional space of C elements. Therefore
E[kx − Cxk2 ] + α−2 kCk2fro has a unique minimum in C space. To locate it we
compute the derivative of the i-th component of this sum with respect to the entry
C(k, l) = Ckl :
∂
e0 (R − C 0 R − CR + C 0 C(R + α−2 I))ei =
∂Ckl i
X
= −2 Rkl + ∂/∂Ckl
Cij Cja Aai
j,a=1,...,N
= −2 Rkl +
X
∂/∂Ckl Ckj Cka Aai
a=1,...,N
= −2 Rkl +
X
(Cka Aal + Cki Ali )
a=1,...,N
= −2 Rkl + (CA)kl + N Cki Ail .
where we used the abbreviation A = R + α−2 I, observing in the last line that
A = A0 . Summing over i yields
∂
E[kx − Cxk2 ] + α−2 kCk2fro = −2 N Rkl + 2N (CA)kl ,
∂Ckl
which in matrix form is
∂
E[kx − Cxk2 ] + α−2 kCk2fro = −2 N R + 2N C(R + α−2 I),
∂C
where we re-inserted the expression for A. Setting this to zero yields the claim 1.
stated in the proposition.
The subclaim that R(R + α−2 I)−1 = (R + α−2 I)−1 R can be easily seen when
R is written by its SVD R = U ΣU 0 : U ΣU 0 (U ΣU 0 + α−2 I)−1 = U ΣU 0 (U (Σ +
α−2 I)U 0 )−1 = U ΣU 0 U (Σ+α−2 I)−1 U 0 = U Σ(Σ+α−2 I)−1 U 0 = U (Σ+α−2 I)−1 ΣU 0 =
... = (R + α−2 I)−1 R. This also proves claim 2.
Claims 3.–5. can be derived from the first claim by elementary arguments.
171
5.2
Proof of Proposition 6 (Section 3.9.3)
−1 −1
−1 0
Cδ−1 can be written as U Sδ−1 U 0 , where the diagonal of Sδ−1 is (s−1
1 , . . . , sl , δ , . . . , δ ) .
−1
−1
0
−1
Putting e = δ −1 and Se := (s−1
1 , . . . , sl , e, . . . , e) , and similarly Te = (t1 , . . . , tm ,
−1
−1
0
−1
e, . . . , e) , we can express the limit limδ→0 (Cδ + Bδ − I) equivalently as
lime→∞ (U Se U 0 + V Te V 0 − I)−1 . Note that U Se U 0 + V Te V 0 − I is invertible for
sufficiently large e. Let U>l be the N × (N − l) submatrix of U made from
the last N − l columns of U (spanning the null space of C), and let V>m be
N × (N − m) submatrix of V made from the last N − m columns of V . The
N × N matrix U>l (U>l )0 + V>m (V>m )0 is positive semidefinite. Let W ΣW 0 be
its SVD, with singular values in Σ in descending order. Noting that C † =
−1
−1
0 0
0 0
†
−1
U diag(s−1
1 , . . . , sl , 0, . . . , 0) U , and B = V diag(t1 , . . . , tm , 0, . . . , 0) V , we can
rewrite
lim (U Se U 0 + V Te V 0 − I)−1
lim Cδ ∧ Bδ =
e→∞
δ→0
lim (C † + B † + e W ΣW 0 − I)−1
= W lim (W 0 C † W + W 0 B † W + eΣ − I)−1 W 0 .
=
e→∞
e→∞
(116)
If Σ is invertible, clearly (116) evaluates to the zero matrix. We proceed to
consider the case of non-invertible Σ = diag(σ1 , . . . , σk , 0, . . . , 0), where 0 ≤ k <
N.
We derive two auxiliary claims. Let W>k be the N × (N − k) submatrix of
W made from the last N − k columns. Claim 1: the (N − k) × (N − k) matrix
A = (W>k )0 (C † + B † − I) W>k is invertible. We rewrite
A = (W>k )0 C † W>k + (W>k )0 B † W>k − I(N −k)×(N −k) ,
(117)
and analyse (W>k )0 C † W>k . It holds that (W>k )0 U>l = 0(N −k)×(N −l) , because
σ1
W
..
.
σk
0
0
W =
...
U>l (U>l )0 + V>m (V>m )0
0
σ1
=⇒
..
.
σk
0
..
.
=
W 0 U>l (U>l )0 W + W 0 V>m (V>m )0 W
=
=
(W>k )0 U>l (U>l )0 W>k + (W>k )0 V>m (V>m )0 W>k
0(N −k)×(N −l) .
(118)
0
=⇒
=⇒
0(N −k)×(N −k)
(W>k )0 U>l
172
Let U≤l be the N × l submatrix of U made of the first l columns. Because of
(118) it follows that
(W>k )0 C † W>k
s−1
1
= (W>k )0 U
= (W>k )0 U≤l
...
s−1
l
s−1
1
0
..
.
0
...
s−1
l
0
U W>k
(U≤l )0 W>k
(119)
The rows of the (N −k)×l sized matrix (W>k )0 U≤l are orthonormal, which also
follows from (118). The Cauchy interlacing theorem (stated in Section 3.14.4 in
another context) then implies that all singular values of (W>k )0 C † W>k are greater
−1
or equal to min{s−1
1 , . . . , sl }. Since all si are smaller or equal to 1, all singular
values of (W>k )0 C † W>k are greater or equal to 1. The singular values of the
matrix (W>k )0 C † W>k − 1/2 I(N −k)×(N −k) are therefore greater or equal to 1/2.
Specifically, (W>k )0 C † W>k − 1/2 I(N −k)×(N −k) is positive definite. By a similar
argument, (W>k )0 B † W>k − 1/2 I(N −k)×(N −k) is positive definite. The matrix A
from Claim 1 is therefore revealed as the sum of two positive definite matrices,
and hence is invertible.
Claim 2: If M is a symmetric N × N matrix, and the right lower principal submatrix M>k>k = M (k + 1 : N, k + 1 : N ) is invertible, and Σ =
diag(σ1 , . . . , σk , 0, . . . , 0) with all σi > 0, then
0
0
−1
lim (M + eΣ) = 0 M −1
.
(120)
e→∞
>k>k
We exploit the following elementary block representation of the inverse of a
symmetric matrix (e.g. [8], fact 2.17.3):
X Y −1
V −1
−V −1 Y Z −1
=
,
Y0 Z
−Z −1 Y 0 V −1 Z −1 Y 0 V −1 Y Z −1 + Z −1
where Z is assumed to be invertible and V = (X − Y Z −1 Y 0 ) is assumed to be
invertible. Block-structuring M + eΣ analogously to this representation, where
M>k>k is identified with Z, then easily leads to the claim (120).
Applying Claims 1 and 2 to the limit expression lime→∞ (W 0 C † W + W 0 B † W +
eΣ − I)−1 in (116), where the matrix M in Claim 2 is identified with W 0 C † W +
W 0 B † W − I and the matrix Z from Claim 2 is identified with the matrix A =
(W>k )0 (C † + B † − I) W>k from Claim 1, yields
0
0
0 †
0 †
−1
−1
lim (W C W + W B W + eΣ − I) =
0 (W>k )0 (C † + B † − I) W>k
e→∞
173
which, combined with (116), leads to
lim Cδ ∧ Bδ = W>k (W>k )0 (C † + B † − I) W>k
δ→0
−1
(W>k )0 .
(121)
W>k is an N × (N − k) size matrix whose columns are orthonormal. Its range is
R(W>k ) =
=
=
=
N (W ΣW 0 ) = N (U>l (U>l )0 + V>m (V>m )0 )
N ((U>l (U>l )0 ) ∩ N (V>m (V>m )0 ) = N ((U>l )0 ) ∩ N ((V>m )0 )
R(U>l )⊥ ∩ R(V>m )⊥ = N (C)⊥ ∩ N (B)⊥
R(C 0 ) ∩ R(B 0 ) = R(C) ∩ R(B),
(122)
that is, W>k is a matrix whose columns form an orthonormal basis of R(C)∩R(B).
Let BR(C)∩R(B) be any matrix whose columns form an orthonormal basis of
R(C) ∩ R(B). Then there exists a unique orthonormal matrix T of size (N − k) ×
(N − k) such that BR(C)∩R(B) = W>k T . It holds that
−1
W>k (W>k )0 (C † + B † − I) W>k
(W>k )0 =
−1
(W>k T )0
= W>k T (W>k T )0 (C † + B † − I) W>k T
−1 0
= BR(C)∩R(B) B0R(C)∩R(B) (C † + B † − I) BR(C)∩R(B)
BR(C)∩R(B) ,(123)
which gives the final form of the claim in the proposition.
For showing equivalence of (33) with (123), I exploit a fact known in matrix theory ([8], Fact 6.4.16): for two real matrices X, Y of sizes n × m, m × l the following
two condition are equivalent: (i) (X Y )† = Y † X † , and (ii) R(X 0 X Y ) ⊆ R(Y ) and
R(Y Y 0 X 0 ) ⊆ R(X 0 ). Observing that PR(C)∩R(B) = BR(C)∩R(B) B0R(C)∩R(B) and
that B†R(C)∩R(B) = B0R(C)∩R(B) , setting X = BR(C)∩R(B) and Y = B0R(C)∩R(B) (C † +
B † − I)PR(C)∩R(B) in (33) yields
−1
PR(C)∩R(B) (C † + B † − I) PR(C)∩R(B)
=
†
B0R(C)∩R(B) (C † + B † − I) PR(C)∩R(B) B0R(C)∩R(B) ,
where condition (ii) from the abovementioned fact is easily verified. In a second,
†
entirely analog step one can pull apart B0R(C)∩R(B) (C † + B † − I) PR(C)∩R(B)
into
†
BR(C)∩R(B) B0R(C)∩R(B) (C † + B † − I) BR(C)∩R(B) .
Algorithm for Computing BR(C)∩R(B) . Re-using ideas from this proof, a
basis matrix BR(C)∩R(B) can be computed as follows:
1. Compute the SVDs C = U diag(s1 , . . . , sl , 0, . . . , 0)U 0 and B = V diag(t1 , . . . , tm ,
0, . . . , 0)V 0 .
2. Let U>l be the submatrix of U made from the last N − l columns in U , and
similarly let V>m consist of the last N − m columns in V .
174
3. Compute the SVD U>l (U>l )0 + V>m (V>m )0 = W ΣW 0 , where
Σ = diag(σ1 , . . . , σk , 0, . . . , 0).
4. Let W>k be the submatrix of W consisting of the last N − k columns of W .
Then BR(C)∩R(B) = W>k .
5.3
Proof of Proposition 7 (Section 3.9.3)
.
By Definition 4, Proposition 6, and Equation (7),
−1
−1
C ∨ B = ¬ (¬ C ∧ ¬ B) = I − lim (¬C)−1
+
(¬B)
−
I
δ
δ
δ↓0
−1
= I − lim (I − C (δ) )−1 + (I − B (δ) )−1 − I
δ↓0
−1
(δ)
(δ)
(δ)
(δ)
= I − lim (I − RC (RC + I)−1 )−1 + (I − RB (RB + I)−1 − I
.
δ↓0
It is easy to check that (I−A(A+I)−1 )−1 = I+A holds for any positive semidefinite
matrix A. Therefore,
−1
(δ)
(δ)
.
C ∨ B = I − lim RC + RB + I
δ↓0
Furthermore, for positive semidefinite A, B it generally holds that I − (A + B +
I)−1 = (A + B)(A + B + I)−1 , and hence
(δ)
(δ)
(δ)
(δ)
C ∨ B = lim(RC + RB ) (RC + RB + I)−1 .
δ↓0
5.4
Proof of Proposition 8 (Section 3.9.3)
(δ)
(δ)
Using (32), Proposition 7 and that fact A(δ) = RA (RA + I)−1 holds for any
(δ)
conceptor A (which entails I − A(δ) = (RA + I)−1 ), we derive the claim as follows:
−1
C ∧ B = lim Cδ−1 + Bδ−1 − I
δ↓0
−1
−1
= lim (¬¬C)−1
δ + (¬¬B)δ − I
δ↓0
−1
= lim (¬(¬C)(δ) )−1 + (¬(¬B)(δ) )−1 − I
δ↓0
−1
= lim (I − (¬C)(δ) )−1 + (I − (¬B)(δ) )−1 − I
δ↓0
−1
(δ)
(δ)
= lim (R¬C + I) + (R¬B ) + I) − I
δ↓0
(δ)
(δ)
= lim(R¬C + R¬B + I)−1
δ↓0
(δ)
(δ)
= I − lim(I − (R¬C + R¬B + I)−1 )
δ↓0
= ¬(¬C ∨ ¬B).
175
5.5
Proof of Proposition 9 (Section 3.9.4)
Claims 1. – 3. are elementary.
Claim 4a: R(C ∧ B) = R(C) ∩ R(B). By definition we have
C ∧ B = BR(C)∩R(B) B0R(C)∩R(B) (C † + B † − I)BR(C)∩R(B)
−1
B0R(C)∩R(B) .
Since B0R(C)∩R(B) (C † + B † − I)BR(C)∩R(B) is invertible and R(B0R(C)∩R(B) ) =
dim(R(C)∩R(B))
0
†
†
−1
0
R
, we have R (BR(C)∩R(B) (C + B − I)BR(C)∩R(B) ) BR(C)∩R(B) =
Rdim(R(C)∩R(B)) . From this it follows that R(C ∧ B) = R(BR(C)∩R(B) ) = R(C) ∩
R(B).
Claim 5a: I(C ∧ B) = I(C) ∩ I(B). We have, by definition,
C ∧ B = BR(C)∩R(B) B0R(C)∩R(B) (C † + B † − I)BR(C)∩R(B)
−1
B0R(C)∩R(B) .
For shorter notation put B = BR(C)∩R(B) and X = B0R(C)∩R(B) (C † +B † −I)BR(C)∩R(B) .
Note (from proof of Proposition 6) that X is invertible. We characterize the unit
eigenvectors of C ∧ B. It holds that BX −1 B0 x = x if and only if BXB0 x = x.
We need to show that the conjunction Cx = x and Bx = x is equivalent to
(C ∧ B) x = BX −1 B x = x.
First assume that Cx = x and Bx = x. This implies x ∈ R(C) ∩ R(B) and
C † x = x and B † x = x, and hence PR(C)∩R(B) (C † + B † − I)PR(C)∩R(B) x = x. But
PR(C)∩R(B) (C † + B † − I)PR(C)∩R(B) = BXB0 , thus (C ∧ B) x = x.
Now assume conversely that not Cx = x or not Bx = x.
Case 1: x ∈
/ R(C) ∩ R(B). Then PR(C)∩R(B) (C † + B † − I)PR(C)∩R(B) x 6= x
and hence BXB0 x 6= x, which implies (C ∧ B) x 6= x.
Case 2: x ∈ R(C) ∩ R(B). We first show an auxiliary claim: k(C † + B † −
I) xk > kxk. Let C0 = C † − CC † , B0 = B † − BB † . C0 and B0 are positive
semidefinite because the nonzero singular values of C † , B † are greater or equal to
1. Furthermore, CC † x = BB † x = x. Thus, (C † + B † ) x = 2Ix + C0 x + B0 x, i.e.
(C † + B † − I) x = Ix + C0 x + B0 x. From not Cx = x or not Bx = x it follows
that C0 x 6= 0 or B0 x 6= 0. We infer
C0 x 6= 0 or B0 x 6= 0 =⇒ x0 C0 x > 0 or x0 B0 x > 0
=⇒ x0 (C0 + B0 )x > 0 =⇒
x0 (C0 + B0 )2 x > 0.
This implies kIx+C0 x+B0 xk2 = kxk2 +2x0 (C0 +B0 )x+x0 (C0 +B0 )2 x > kxk2 ,
or equivalently, k(C † + B † − I) xk > kxk, the auxiliary claim.
Since PR(C)∩R(B) preserves vector norm on R(C) ∩ R(B) and R(C † + B † −
I)PR(C)∩R(B) ⊆ R(C)∩R(B), it follows that kPR(C)∩R(B) (C † +B † −I)PR(C)∩R(B) xk >
kxk, hence (C ∧ B) x 6= x.
Altogether we have that the conjunction Cx = x and Bx = x is equivalent to
(C ∧ B) x = x, which is equivalent to the claim.
176
Claim 6a: N (C ∧ B) = N (C) + N (B). This follows from 4a by N (C ∧ B) =
(R(C ∧ B))⊥ = (R(C) ∩ R(B))⊥ = (N (C)⊥ ∩ N (B)⊥ )⊥ = N (C) + N (B).
Claims 4b, 5b, 6b: The second statements in 4., 5., 6. follow from the first
statements and 3., exploiting de Morgan’s rule.
Claim 7 follows from Equation (19).
Claim 8: Let A = A ∧ C. Then claim 4. implies R(A) ∩ R(C) = R(A). By
Proposition 6 we can write
A ∧ C = BR(A) B0R(A) (C † + B † − I) BR(A)
−1
B0R(A) .
Let A = U SU 0 be the SVD of A, and assume A has rank k ≤ N , that is,
exactly the first k singular values in S are nonzero. Let Uk be the N × k matrix
consisting of the first k columns of U . It holds that Uk = BR(A) . We obtain
S = U 0 AU =
−1 0
= U 0 Uk Uk0 (A† + C † − I) Uk
Uk U
−1
= Ik Uk0 (A† + C † − I) Uk
Ik0 ,
(124)
where Ik is the N × k matrix consisting of the first k columns of I.
−1Let Sk be the
0
†
†
k × k upper left submatrix of S. Then Sk = Uk (A + C − I) Uk
and
Sk−1 = Uk0 (A† + C † − I) Uk
= Uk0 A† Uk + Uk0 (C † − I) Uk
= Sk−1 + Uk0 (C † − I) Uk ,
hence Uk0 (C † −I) Uk = 0k×k or equivalently, Uk0 C † Uk = Ik×k . This implies R(A) ⊆
I(C).
Conversely, assume R(A) ⊆ I(C). Going through the above line of arguments
−1 0
in reverse order establishes again S = Ik Uk0 (A† + C † − I) Uk
Ik which implies
A = BR(A) B0R(A) (C † + B † − I) BR(A)
−1
B0R(A) .
R(A) ⊆ I(C) implies R(A) ⊆ R(C), which leads to
A = BR(A)∩R(C) B0R(A)∩R(C) (C † + B † − I) BR(A)∩R(C)
= A ∧ C.
−1
B0R(A)∩R(C)
The dual A = A ∨ C ⇔ I(A)⊥ ⊆ N (C) is easily obtained from A = A ∧ C ⇔
R(A) ⊆ I(C) by applying de Morgan’s rules and claim 3..
Claims 9 – 13 follow from Equation (19). For Claim 14. use 11. and 4. and 6.
to first establish that R(H ∧ G) = I(H) ∩ I(G) and N (H ∧ G) = (I(H) ∩ I(G))⊥ ,
from which 14. follows. Claim 15 is analog.
177
5.6
Proof of Proposition 10 (Section 3.9.5)
Claim 1: ¬ϕ(C, γ) = ϕ(¬C, γ −1 ). Notation: All of the matrices C, ¬C, ϕ(C, γ),
¬ϕ(C, γ), ϕ(¬C, γ −1 ) are positive semidefinite and have SVDs with identical principal component matrix U . For any matrix X among these, let U S X U 0 be its
X
SVD. We write sX
i for the ith singular value in S . We have to show that
−1
¬ϕ(C,γ)
ϕ(¬C,γ )
si
= si
. In the derivations below, we use various facts from Proposition 9 and Equation (19) without explicit reference.
Case 0 < γ < ∞, 0 < sC
i < 1:
¬ϕ(C,γ)
si
sC
1 − sC
i
i
=
C
C
C
−2
si + γ (1 − si )
(1 − si ) + γ 2 sC
i
¬C
si
ϕ(¬C,γ −1 )
= ¬C
= si
.
¬C
−1
−2
si + (γ ) (1 − si )
= 1−
C
¬C
Case 0 < γ < ∞, sC
= 1 we have
i = 0: Using si = 0 ⇔ si
¬ϕ(C,γ)
si
ϕ(C,γ)
= 1 − si
ϕ(¬C,γ −1 )
= 1 = si
.
Case 0 < γ < ∞, sC
i = 1: dual of previous case.
Case γ = 0: We show ¬ϕ(C, γ) = ϕ(¬C, γ −1 ) directly.
¬ϕ(C, γ) = ¬ϕ(C, 0) = I − ϕ(C, 0) = I − PI(C) = PI(C)⊥
= PN (¬C)⊥ = ϕ(¬C, ∞) = ϕ(¬C, γ −1 ).
Case γ = ∞: the dual analog.
Claim 2: ϕ(C, γ) ∨ ϕ(B, γ) = ϕ(C ∨ B, γ).
Case 0 < γ < ∞: Using concepts and notation from Proposition 7, it is easy
to check that any conceptor A can be written as
(δ)
(δ)
A = lim RA (RA + I)−1 ,
(125)
δ↓0
and its aperture adapted versions as
(δ)
(δ)
ϕ(A, γ) = lim RA (RA + γ −2 I)−1 .
(126)
δ↓0
Using Proposition 7 and (125) we thus have
(δ)
(δ)
(δ)
(δ)
C ∨ B = lim (RC + RB ) (RC + RB + I)−1
δ↓0
(δ)
(δ)
= lim RC∨B (RC∨B + I)−1 .
(127)
(128)
δ↓0
Furthermore, again by Proposition 7 and by (126),
(δ)
(δ)
(δ)
(δ)
ϕ(C, γ) ∨ ϕ(B, γ) = lim (Rϕ(C,γ) + Rϕ(B,γ) ) (Rϕ(C,γ) + Rϕ(B,γ) + I)−1
δ↓0
178
(129)
and
(δ)
(δ)
ϕ(C ∨ B, γ) = lim (RC∨B )(RC∨B + γ −2 I)−1 .
(130)
δ↓0
Using (17), it follows for any conceptor A that
(δ/(δ+γ −2 (1−δ)))
(δ)
Rϕ(A,γ) = γ 2 RA
.
(131)
Applying this to (129) and observing that limδ↓0 δ = 0 = limδ↓0 δ/(δ + γ −2 (1 − δ))
yields
(δ)
(δ)
(δ)
(δ)
ϕ(C, γ) ∨ ϕ(B, γ) = lim (γ 2 RC + γ 2 RB ) (γ 2 RC + γ 2 RB + I)−1
δ↓0
(δ)
(δ)
(δ)
(δ)
= lim (RC + RB ) (RC + RB + γ −2 I)−1 .
δ↓0
(132)
We now exploit the following auxiliary fact which can be checked by elementary means: If (X (δ) )δ is a δ-indexed family of positive semidefinite matrices
whose eigenvectors are identical for different δ, and similarly the members of the
familiy (Y (δ) )δ have identical eigenvectors, and if the limits limδ↓0 X (δ) (X (δ) +I)−1 ,
limδ↓0 Y (δ) (Y (δ) + I)−1 exist and are equal, then the limits limδ↓0 X (δ) (X (δ) +
γ −2 I)−1 , limδ↓0 Y (δ) (Y (δ) + γ −2 I)−1 exist and are equal, too. Putting X (δ) =
(δ)
(δ)
(δ)
RC + RB and Y (δ) = RC∨B , combining (127), (128), (130) and (132) with this
auxiliary fact yields ϕ(C ∨ B, γ) = ϕ(C, γ) ∨ ϕ(B, γ).
Case γ = 0: Using various findings from Proposition 9 we have
ϕ(C, 0) ∨ ϕ(B, 0) = PI(C) ∨ PI(B) = PI(PI(C) ) + I(PI(B) )
= PI(C)+I(B) = PI(C∨B) = ϕ(C ∨ B, 0).
Case γ = ∞:
ϕ(C, ∞) ∨ ϕ(B, ∞) = PR(C) ∨ PR(B) = PI(PR(C) ) + I(PR(B) )
= PR(C) + R(B) = PR(C∨B) = ϕ(C ∨ B, ∞).
Claim 3: ϕ(C, γ) ∧ ϕ(B, γ) = ϕ(C ∧ B, γ): follows from Claims 1. and 2. with
de Morgan’s law.
p
Claim 4: ϕ(C, γ) ∨ ϕ(C, β) = ϕ(C, γ 2 + β 2 ):
Case 0 < γ, β < ∞: Using Proposition 9 and Equations (131), (127), (126),
179
we obtain
(δ)
(δ)
(δ)
(δ)
ϕ(C, γ) ∨ ϕ(C, β) = lim (Rϕ(C,γ) + Rϕ(C,β) ) (Rϕ(C,γ) + Rϕ(C,β) + I)−1
δ↓0
(δ/(δ+γ −2 (1−δ)))
= lim (γ 2 RC
δ↓0
(δ/(δ+γ −2 (1−δ)))
· (γ 2 RC
(∗)
(δ/(δ+β −2 (1−δ)))
+ β 2 RC
(δ)
)·
(δ/(δ+β −2 (1−δ)))
+ β 2 RC
+ I)−1
(δ)
= lim (γ 2 + β 2 )RC ((γ 2 + β 2 )RC + I)−1
δ↓0
(δ)
(δ)
= lim RC (RC + (γ 2 + β 2 )−1 I)−1
δ↓0
p
= ϕ(C, γ 2 + β 2 ),
(δ/(δ+γ −2 (1−δ)))
where in step (*) we exploit the fact that the singular values of RC
corresponding to eigenvectors whose eigenvalues in C are less than unity are identical
(δ)
to the singular values of RC at the analog positions.
Case γ = 0, 0 < β < ∞: Using Proposition 7, facts from Proposition 9, and
Equation (125), we obtain
ϕ(C, 0) ∨ ϕ(C, β) = PI(C) ∨ ϕ(C, β)
(δ)
(δ)
(δ)
(δ)
= lim (RPI(C) + Rϕ(C,β) ) (RPI(C) + Rϕ(C,β) + I)−1
δ↓0
(∗)
(δ/(2−δ))
(δ/(2−δ))
(δ)
(δ/)
(Rϕ(C,β) + I)−1 = lim Rϕ(C,β) (Rϕ(C,β) + I)−1
δ↓0
δ↓0
p
= ϕ(C, β) = ϕ(C, 02 + β 2 ),
= lim Rϕ(C,β)
where step (*) is obtained by observing I(C) = I(ϕ(C, β)) and applying the
(δ)
definition of RA given in the statement of Proposition 7.
Case γ = ∞, 0 < β < ∞: the dual analog to the previous case:
ϕ(C, ∞) ∨ ϕ(C, β) = PR(C) ∨ ϕ(C, β)
(δ)
(δ)
(δ)
(δ)
= lim (RPR(C) + Rϕ(C,β) ) (RPR(C) + Rϕ(C,β) + I)−1
δ↓0
(∗)
(δ)
(δ/)
= lim RPR(C) (RPR(C) + I)−1
δ↓0
= PR(C) = ϕ(C, ∞) = ϕ(C,
p
∞2 + β 2 ),
where in step (*) I have omitted obvious intermediate calculations.
The cases 0 < γ < ∞, β ∈ {0, ∞} are symmetric to cases already treated, and
the cases γ, β ∈ {0, ∞} are obvious.
Claim 5: ϕ(C, γ) ∧ ϕ(C, β) = ϕ(C, (γ −2 + β −2 )−2 ): an easy exercise of applying
de Morgan’s rule in conjunction with Claims 1. and 4.
180
5.7
Proof of Proposition 11 (Section 3.9.6)
1. De Morgan’s rules: By Definition 4 and Proposition 8. 2. Associativity: From
Equations (127) and (128) it follows that for any conceptors B, C it holds that
(δ)
(δ)
(δ)
limδ↓0 RB∨C = limδ↓0 RB + RC . Employing this fact and using Proposition 7
yields associativity of OR. Applying de Morgan’s law then transfers associativity
to AND. 3. Commutativity and 4. double negation are clear. 5. Neutrality:
Neutrality of I: Observing that R(C) ∩ R(I) = R(C) and I † = I, starting from
the definition of ∨ we obtain
C ∨ I = (PR(C) C † PR(C) )†
= (PR(C † ) C † PR(C † ) )†
(by Prop. 9 Nr. 2)
= (C † )†
= C.
Neutrality of 0 can be obtained from neutrality of I via de Morgan’s rules.
6. Globality: C ∧ 0 = 0 follows immediately from the Definition of ∧ given in
4, observing that PR(C)∩R(0) = 0. The dual C ∨ I = I is obtained by applying de
Morgan’s rule on C ∧ 0 = 0.
7. and 8. weighted absorptions follows from Proposition 10 items 4. and 5.
5.8
Proof of Proposition 12 (Section 3.9.6)
Let ≤ denote the well-known Löwner ordering on the set of real N × N matrices
defined by A ≤ B if B − A is positive semidefinite. Note that a matrix C is a
conceptor matrix if and only if 0 ≤ C ≤ I. I first show the following
Lemma 2 Let A ≤ B. Then A† ≥ PR(A) B † PR(A) .
Proof of Lemma. A† and B † can be written as
A† = lim PR(A) (A + δI)−1 PR(A) and B † = lim PR(B) (B + δI)−1 PR(B) .
δ→0
δ→0
(133)
From A ≤ B it follows that R(A) ⊆ R(B), that is, PR(A) PR(B) = PR(A) , which
in turn yields
PR(A) B † PR(A) = lim PR(A) (B + δI)−1 PR(A) .
(134)
δ→0
A ≤ B entails A + δI ≤ B + δI, which is equivalent to (A + δI)−1 ≥ (B + δI)−1
(see [8], fact 8.21.11), which implies
PR(A) (B + δI)−1 PR(A) ≤ PR(A) (A + δI)−1 PR(A) ,
see Proposition 8.1.2 (xii) in [8]. Taking the limits (133) and (134) leads to the
claim of the lemma (see fact 8.10.1 in [8]).
181
Proof of Claim 1. Let A, B be conceptor matrices of size N × N . According to
Proposition 14 (which is proven independently of the results stated in Proposition
12), it holds that A ≤ A ∨ B, which combined with the lemma above establishes
PR(A) (A† − (A ∨ B)† )PR(A) ≥ 0,
from which it follows that I + PR(A) (A† − (A ∨ B)† )PR(A) is positive semidefinite
with all singular values greater or equal to one. Therefore, PR(A) (I + A† − (A ∨
B)† )PR(A) is positive semidefinite with all nonzero singular values greater or equal
†
to one. Hence C = PR(A) I + A† − (A ∨ B)† PR(A) is a conceptor matrix. It
is furthermore obvious that R(C) = R(A).
From A ≤ A ∨ B it follows that R(A) ⊆ R(A ∨ B), which together with
R(C) = R(A) leads to R(A ∨ B) ∩ R(C) = R(A). Exploiting this fact, starting
from the definition of AND in Def. 4, we conclude
†
(A ∨ B) ∧ C = PR(A∨B)∩R(C) (A ∨ B)† + C † − I PR(A∨B)∩R(C)
†
= PR(A) (A ∨ B)† + C † − I PR(A)
†
= PR(A) (A ∨ B)† + PR(A) I + A† − (A ∨ B)† PR(A) − I PR(A)
= (PR(A) A† PR(A) )† = A.
Proof of Claim 2. This claim is the Boolean dual to claim 1 and can be
straightforwardly derived by transformation from claim 1, using de Morgan’s rules
and observing that R(¬A) = I(A)⊥ (see Prop. 9 item 3), and that ¬A = I − A.
5.9
Proof of Proposition 13 (Section 3.10)
Claim 1: 0 ≤ A is equivalent to A being positive semidefinite, and for a positive
semidefinite matrix A, the condition A ≤ I is equivalent to all singular values of
A being at most one. Both together yield the claim.
Claim 2: Follows from claim 1.
Claim 3: A ≤ B iff −A ≥ −B iff I − A ≥ I − B, which is the same as
¬A ≥ ¬B.
Claim 4: We first show an auxiliary, general fact:
Lemma 3 Let X be positive semidefinite, and P a projector matrix. Then
(P(PXP + I)−1 P)† = PXP + P.
Proof of Lemma. Let U = R(P) be the projection space of P. It is clear that
PXP + I : U → U and PXP + I : U ⊥ → U ⊥ , hence PXP + I is a bijection on U .
Also P is a bijection on U . We now call upon the following well-known property
of the pseudoinverse (see [8], fact 6.4.16):
For matrices K, L of compatible sizes it holds that (KL)† = L† K † if and only
if R(K 0 KL) ⊆ R(L) and R(LL0 K) ⊆ R(K 0 ).
182
Observing that P and PXP + I are bijections on U = R(P), a twofold application of the mentioned fact yields (P(PXP + I)−1 P)† = P† (PXP + I)P† =
PXP + P which completes the proof of the lemma.
Now let A, B ∈ CN and B ≤ A. By Lemma 2, B † − PR(B) A† PR(B) is positive semidefinite, hence B † − PR(B) A† PR(B) + I is positive definite with singular values greater or equal to one, hence invertible. The singular values of
(B † − PR(B) A† PR(B) + I)−1 are thus at most one, hence C = PR(B) (B † −
PR(B) A† PR(B) + I)−1 PR(B) is a conceptor matrix. Obviously R(C) = R(B).
Using these findings and Lemma 3 we can now infer
†
A ∧ C = PR(A)∩R(C) (A† + C † − I) PR(A)∩R(C)
†
†
†
†
†
−1
= PR(B) (A + PR(B) (B − PR(B) A PR(B) + I) PR(B) − I) PR(B)
†
= PR(B) A† + PR(B) (B † − PR(B) A† PR(B) ) PR(B) PR(B)
= B.
Claim 5: This claim is the Boolean dual to the previous claim. It follows by
a straightforward transformation applying de Morgan’s rules and observing that
R(¬B) = I(B)⊥ (see Prop. 9 item 3), and that ¬A = I − A.
Claim 6: Let A ∧ C = B. Using the notation and claim from Prop. 6 rewrite
−1
A ∧ C = limδ→0 (Cδ−1 + A−1
δ − I) . Similarly, obviously we can also rewrite
A = limδ→0 Aδ . Since Cδ−1 ≥ I, conclude
−1
Cδ−1 + A−1
δ − I ≥ Aδ
−1
⇐⇒ (Cδ−1 + A−1
≤ Aδ
δ − I)
−1
−1
=⇒ lim(Cδ + Aδ − I)−1 ≤ lim Aδ
δ→0
δ→0
⇐⇒ A ∧ C ≤ A,
where use is made of the fact that taking limits preserves ≤ (see [8] fact 8.10.1).
Claim 7: Using the result from the previous claim, infer A ∨ C = B =⇒
¬A ∧ ¬C = ¬B =⇒ ¬A
p≥ ¬B =⇒ A ≤ B.
Claim 8: Let γ = 1 + β 2 ≥ 1, where β ≥ 0. By Proposition 10 4. we get
ϕ(A, γ) = ϕ(A, 1) ∨ ϕ(A, β) = A ∨ ϕ(A, β), hence A ≤ ϕ(A, γ). The dual version
is obtained from this result by using Proposition 5: let γ ≤ 1, hence γ −1 ≥ 1.
Then ϕ(A, γ) ≤ ϕ(ϕ(A, γ), γ −1 ) = A.
Claim 9: If γ = 0, then from Proposition 3 it is clear that ϕ(A, 0) is the
projector matrix on I(A) and ϕ(B, 0) is the projector matrix on I(B). From
A ≤ B and the fact that A and B do not have singular values exceeding 1 it is
clear that I(A) ⊆ I(B), thus ϕ(A, 0) ≤ ϕ(B, 0).
If γ = ∞, proceed in an analog way and use Proposition 3 to conclude that
ϕ(A, ∞), ϕ(B, ∞) are the projectors on R(A), R(B) and apply that A ≤ B implies
R(A) ⊆ R(B).
It remains to treat the case 0 < γ < ∞. Assume A ≤ B, that is, there exists
a positive semidefinite matrix D such that A + D = B. Clearly D cannot have
183
singular values exceeding one, so D is a conceptor matrix. For (small) δ > 0, let
A(δ) = (1 − δ)A. Then A(δ) can be written as A(δ) = R(δ) (R(δ) + I)−1 for a positive
semidefinite R(δ) , and it holds that
A = lim A(δ) ,
δ→0
and furthermore
0
A(δ) ≤ A(δ ) ≤ A for δ ≥ δ 0 .
Similarly, let D(δ) = (1 − δ)D, with D(δ) = Q(δ) (Q(δ) + I)−1 , and observe again
D = lim D(δ)
δ→0
0
and D(δ) ≤ D(δ ) ≤ D for δ ≥ δ 0 .
Finally, define B (δ) = (1 − δ)B, where B (δ) = P (δ) (P (δ) + I)−1 . Then
B = lim B (δ)
δ→0
0
and B (δ) ≤ B (δ ) ≤ B for δ ≥ δ 0
and B (δ) = A(δ) + D(δ) .
Because of B (δ) = A(δ) + D(δ) we have
R(δ) (R(δ) + I)−1 ≤ P (δ) (P (δ) + I)−1 .
(135)
We next state a lemma which is of interest in its own right too.
Lemma 4 For correlation matrices R, P of same size it holds that
R(R + I)−1 ≤ P (P + I)−1
iff
R ≤ P.
(136)
Proof of Lemma. Assume R(R + I)−1 ≤ P (P + I)−1 . By claim 4. of this proposition, there is a conceptor matrix C such that P (P + I)−1 = R(R + I)−1 ∨ C.
Since P (P + I)−1 < I, C has no unit singular values and thus can be written as S(S + I)−1 , where S is a correlation matrix. Therefore, P (P + I)−1 =
R(R + I)−1 ∨ S(S + I)−1 = (R + S)(R + S + I)−1 , hence P = R + S, that is,
R ≤ P.
Next assume R ≤ P , that is, P = R + S for a correlation matrix S. This
implies P (P + I)−1 = (R + S)(R + S + I)−1 = R(R + I)−1 ∨ S(S + I)−1 . By claim
6. of this proposition, R(R + I)−1 ≤ P (P + I)−1 follows. This concludes the proof
of the lemma.
Combining this lemma with (135) and the obvious fact that R(δ) ≤ P (δ) if and
only if γ 2 R(δ) ≤ γ 2 P (δ) yields
γ 2 R(δ) (γ 2 R(δ) + I)−1 ≤ γ 2 P (δ) (γ 2 P (δ) + I)−1 .
(137)
Another requisite auxiliary fact is contained in the next
Lemma 5 Let 0 < γ < ∞. If A = U SU 0 = limδ→0 R(δ) (R(δ) + I)−1 and for all δ,
R(δ) has a SVD R(δ) = U Σ(δ) U 0 , then ϕ(A, γ) = limδ→0 γ 2 R(δ) (γ 2 R(δ) + I)−1 .
184
Proof of Lemma. Since all R(δ) (and hence, all R(δ) (R(δ) +I)−1 and γ 2 R(δ) (γ 2 R(δ) +
I)−1 ) have the same eigenvectors as A, it suffices to show the convergence claim on
(δ)
the level of individual singular values of the concerned matrices. Let s, sγ , σ (δ) , s(δ) , sγ
denote a singular value of A, ϕ(A, γ), R(δ) , R(δ) (R(δ) + I)−1 , γ 2 R(δ) (γ 2 R(δ) + I)−1 ,
respectively (all these versions referring to the same eigenvector in U ). For convenience I restate from Proposition 3 that
(
s/(s + γ −2 (1 − s))
for 0 < s < 1,
sγ = 0
for s = 0,
1
for s = 1.
(δ)
It holds that s(δ) = σ (δ) /(σ (δ) + 1) and limδ→0 s(δ) = s, and similarly sγ
(δ)
γ 2 σ (δ) /(γ 2 σ (δ) + 1). It needs to be shown that limδ→0 sγ = sγ .
Case s = 0:
=
s = 0 =⇒ lim s(δ) = 0 =⇒ lim σ (δ) = 0
δ→0
=⇒
δ→0
lim s(δ)
γ
δ→0
= 0.
Case s = 1:
s = 1 =⇒ lim s(δ) = 1 =⇒ lim σ (δ) = ∞
δ→0
=⇒
δ→0
lim s(δ)
γ
δ→0
= 1.
Case 0 < s < 1:
s = lim s(δ) =⇒ s = lim σ (δ) /(σ (δ) + 1) =⇒ lim σ (δ) = s/(1 − s)
δ→0
δ→0
δ→0
2
=⇒ lim s(δ)
γ =
δ→0
γ s/(1 − s)
= s/(s + γ −2 (1 − s)).
− s) + 1
γ 2 s/(1
This concludes the proof of the lemma.
After these preparations, we can finalize the proof of claim 9. as follows. From
Lemma 5 we know that
ϕ(A, γ) = lim γ 2 R(δ) (γ 2 R(δ) + I)−1
δ→0
and
ϕ(B, γ) = lim γ 2 P (δ) (γ 2 P (δ) + I)−1 .
δ→0
From (137) and the fact that ≤ is preserved under limits we obtain ϕ(A, γ) ≤
ϕ(B, γ).
185
5.10
Proof of Proposition 15 (Section 3.14.4)
We use the following notation for matrix-vector transforms. We sort the entries
of an N × N matrix M into an N 2 -dimensional vector vec M row-wise (!). That
is, vec M (µ) =
M (dµ/N e, mod1 (µ, N )), where the ceiling dxe of a real number x is the smallest
integer greater or equal to x, and mod1 (µ, N ) is the modulus function except for
arguments of the form (lk, k), where we replace the standard value mod(lm, m) = 0
by mod1 (lm, m) = m. Conversely, M (i, j) = vec M ((i − 1)N + j).
The Jacobian JC can thus be written as a N 2 ×N 2 matrix JC (µ, ν) = ∂ vec Ċ(µ)/∂ vec C(ν).
The natural parametrization of matrices C by their matrix elements does not
lend itself easily to an eigenvalue analysis. Assuming that a reference solution
C0 = U SU 0 is fixed, any N × N matrix C is uniquely represented by a parameter
matrix P through C = C(P ) = U (S + P ) U 0 , with C(P ) = C0 if and only if
P = 0. Conversely, any parameter matrix P yields a unique C(P ).
Now we consider the Jacobian JP (µ, ν) = ∂ vec ṖC (µ)/∂ vec PC (ν). By using
that (i) Ṗ = U 0 ĊU , (ii) vec (X 0 Y X) = (X ⊗ X)0 vec Y for square matrices X, Y ,
and (iii) ∂ Aẋ/∂ Ax = A(∂ ẋ/∂ x)A−1 for invertible A, (iv) (U ⊗ U )−1 = (U ⊗ U )0 ,
one obtains that JP = (U ⊗ U )0 JC (U ⊗ U ) and hence JP and JC have the same
eigenvalues.
Using C0 = U SU 0 and Ċ = (I − C)CDC 0 − α−2 C and the fact that diagonal
entries in S of index greater than k are zero, yields
∂ ṗlm
∂ pij
= e0l (Iij U 0 DU S + SUk0 DU Iji
P =0
−Iij SUk0 DU S − SIij U 0 DU S − S 2 Uk0 DU Iji − α−2 Iij ) em ,(138)
where el is the l-th unit vector and Iij = ei e0j . Depending on how l, m, i, j relate to
k and to each other, calculating (138) leads to numerous case distinctions. Each of
the cases concerns entries in a specific subarea of JP . These subareas are depicted
in Fig. 49, which shows JP in an instance with N = 5, k = 3.
I will demonstrate in detail only two of these cases (subareas A and B in Fig.
49) and summarize the results of the others (calculations are mechanical).
The case A concerns all entries JP (µ, ν) with µ < ν, ν ≤ kN, mod1 (ν, N ) ≤ k.
Translating indices µ, ν back to indices l, m, i, j via JP (µ, ν) = ∂ vec ṖC (µ) / ∂ vec PC (ν) =
∂ ṗ(dµ/N e,mod1 (µ,N )) / ∂ p(dν/N e,mod1 (ν,N )) = ∂ ṗlm / ∂pij yields conditions (i) i ≤ k
(from i = dν/N e and ν ≤ kN ), (ii) j ≤ k (from j = mod1 (ν, N ) ≤ k) and (iii.a)
l < i or (iii.b) l = i ∧ m < j (from µ < ν).
I first treat the subcase (i), (ii), (iii.a). Since l 6= i one has e0l Iij = 0 and eqn.
(138) reduces to the terms starting with S, leading to
186
n!
K
A
C
B
D
m!
E
F
I
J
G
H
L
Figure 49: Main case distinction areas for computing values in the matrix JP . An
instance with N = 5, k = 3 is shown. Areas are denoted by A, ..., L; same color =
same area. JP has size N 2 × N 2 . Its structure is largely organized by a kN × kN dimensional and a (N −k)N ×(N −k)N submatrix on the diagonal (areas ABEFK
and HIJL, respectively). Column/row indices are denoted by µ, ν. Area specifications: A: µ < ν, ν ≤ kN, mod1 (ν, N ) ≤ k. B: µ < ν, ν ≤ kN, mod1 (ν, N ) > k.
C: ν > kN, µ ≤ kN, mod1 (ν, N ) ≤ k. D: ν > kN, µ ≤ kN, mod1 (ν, N ) > k.
E: µ > ν, µ ≤ kN, mod1 (ν, N ) ≤ k. F: µ > ν, µ ≤ kN, mod1 (ν, N ) > k. G:
µ > kN, ν ≤ kN. H: ν > kN, ν < µ. I: µ > kN, µ < ν, mod1 (ν, N ) ≤ k. J:
µ > kN, µ < ν, mod1 (ν, N ) > k. K: µ = ν ≤ kN. L: µ = ν > kN.
∂ ṗlm
∂ pij
= e0l (SUk0 DU Iji − SIij U 0 DU S − S 2 Uk0 DU Iji ) em
P =0
= sl u0l Duj δim − sl e0l Iij U 0 DU Sem − s2l u0l Duj δim
= sl u0l Duj δim − s2l u0l Duj δim
(
0,
if i 6= m
(subcase A1)
0,
if
i
=
m,
j
=
6
l
(A2)
=
−2
α , if i = m, j = l (A3),
where ul is the l-th column in U and δim = 1 if and only if i = m (else 0) is
the Kronecker delta. The value α−2 noted for subcase A3 is obtained through
(sl − s2l )u0l Dul = (sl − s2l )d˜l = α−2 . Note that since l = i ≤ k in subcase A3 it
holds that sl > 1/2.
Next, in the subcase (i), (ii), (iii.b) one has
187
∂ ṗlm
∂ pij
= u0j Dum sm + sl u0l Duj δim − sj u0j Dum sm
P =0
−sl u0j Dum sm − s2l u0l Duj δim − α−2 e0j em
= (sl − s2l )u0l Duj δim
= 0, (A4)
(since u0j Dum = 0 and j 6= m)
(139)
because assuming i 6= m or j 6= l each null the last expression, and i = m ∧ j = l
is impossible because condition (iii.b) would imply m = j contrary to (iii.b).
The case B concerns all entries JP (µ, ν) with µ < ν, ν ≤ kN, mod1 (ν, N ) > k.
Like in case A above, this yields conditions on the P-matrix indices: (i) i ≤ k, (ii)
j > k, (iii.a) l < i or (iii.b) l = i ∧ m < j.
Again we first treat the subcase (i), (ii), (iii.a). Since l 6= i one has e0l Iij = 0
and eqn. (138) reduces to the terms starting with S, leading to
∂ ṗlm
∂ pij
= e0l (SUk0 DU Iji − SIij U 0 DU S − S 2 Uk0 DU Iji ) em
P =0
= sl u0l Duj δim − sl e0l Iij U 0 DU Sem − s2l u0l Duj δim
= sl u0l Duj δim − s2l u0l Duj δim
0,
if i 6= m (B1)
=
(sl − s2l )u0l Duj if i = m (B2).
where ul is the l-th column in U and δim = 1 if and only if i = m (else 0) is the
Kronecker delta. Note that since l < i ≤ k it holds that sl > 1/2.
In the subcase (i), (ii), (iii.b) from (138) one obtains
∂ ṗlm
∂ pij
= u0j Dum sm + sl u0l Duj δim − sj u0j Dum sm
P =0
−sl u0j Dum sm − s2l u0l Duj δim − α−2 e0j em
= sm (1 − sl )u0j Dum + (sl − s2l )u0l Duj δim
if i 6= m and m ≤ k (B3)
sm (1 − sl )u0j Dum
0
if i 6= m and m > k (B4)
=
s (1 − s )u0 Du + (s − s2 )u0 Du if i = m
(B5),
m
l j
m
l
j
l
l
where in the step from the first to the second line one exploits j > k, hence sj = 0;
and m < j, hence e0j em = 0. Note that since l = i ≤ k it holds that sl > 0; that
m > k implies sm = 0 and that m = i ≤ k implies sm > 0.
Most of the other cases C – L listed in Fig. 49 divide into subcases like A and
B. The calculations are similar to the ones above and involve no new ideas. Table
1 collects all findings. It only shows subcases for nonzero entries of JP . Fig. 50
depicts the locations of these nonzero areas.
188
Subcase Index range
Cell value
A3
l = j ≤ k; i = m ≤ k;
α−2
B2
l<i=m≤k<j
(sl − s2l ) u0l Duj
B3
l = i ≤ k < j; m ≤ j; m 6= i sm (1 − sl )u0j Dum
B5
l = i ≤ k < j; m ≤ j; m = i sl (1 − sl )u0j Dul + (sl − s2l )u0l Duj
C1
l = j ≤ k; i = m > k
α−2
D1
l ≤ k; i, j > k; i = m
(sl − s2l ) u0l Duj
E1
i=m<l=j≤k
α−2
F1
i=m<l≤k<j
(sl − s2l ) u0l Duj
J1
i = l > k; m ≤ k < j
sm u0j Dum
K1
i=j=l=m≤k
α−2 (1 − 2sl )/(1 − sl )
K2
i = l; j = m > k; l 6= m
−α−2
K3
i = l; j = m ≤ k; l 6= m
−α−2 sl /(1 − sm )
L1
i = l > k; m = j > k
−α−2
Table 1: Values in nonzero areas of the Jacobian JP .
The eigenvalues of JP are now readily obtained. First observe that the eigenvalues of a matrix with block structure
!
K L
,
0 M
where K and M are square, are the eigenvalues collected from the principal submatrices K and M . In JP we therefore only need to consider the leading kN × kN
and the trailing (N − k)N × (N − k)N submatrices; call them K and M .
M is upper triangular, its eigenvalues are therefore its diagonal values. We thus
can collect from M (N − k)k times eigenvalues 0 and (N − k)2 times eigenvalues
−α−2 .
By simultaneous permutations of rows and columns (which leave eigenvalues
unchanged) in K we can bring it to the form Kperm shown in Fig. 50B. A block
structure argument as before informs us that the eigenvalues of Kperm fall into
three groups. There are k(N − k) eigenvalues −α−2 (corresponding to the lower
right diagonal submatrix of Kperm , denoted as area K2), k eigenvalues α−2 (1 −
2sl )/(1 − sl ), where l = 1, . . . , k earned from the k leading diagonal elements of
Kperm (stemming from the K1 entries in JP ), plus there are the eigenvalues of
k(k − 1)/2 twodimensional submatrices, each of which is of the form
189
B5
K1
K3 B3
D1
B2
C1
A3
E1
K1
K3 A3
E1
F1
K2
K2
B
L1
J1
J1
A
C
L1
Figure 50: A. Nonzero areas in the Jacobian matrix JP . An instance with N =
5, k = 3 is shown. Areas denotations correspond to Table 1. Same color =
same area. Values: areas A3, C1, E1: α−2 ; B2, D1, F1: (sl − s2l ) u0l Duj ; B3:
sm (1−sl )u0j Dum ; B5: sl (1−sl )u0j Dul +(sl −s2l )u0l Duj ; J1: sm u0j Dum ; K1: α−2 (1−
2sl )/(1 − sl ); K2, L1: −α−2 ; K3: −α−2 sl /(1 − sm ). B. The left upper principal
submatrix re-arranged by simultaneous row/column permutations. C. One of the
N ×N submatrices Cr from the diagonal of the (N −k)N ×(N −k)N right bottom
submatrix of JP . For explanations see text.
Kl,m = −α−2
sl /(1 − sm )
1
1
sm /(1 − sl )
!
,
where l = 1, . . . , k; m < l. By solving the characteristic polynomial of Kl,m its
eigenvalues are obtained as
−2
λ1,2 =
sm
α sl
+
±
2
sm − 1 sl − 1
s
sl
sm
−
sm − 1 sl − 1
2
+ 4 .
(140)
Summarizing, the eigenvalues of JP are constituted by the following multiset:
1. k(N − k) instances of 0,
2. N (N − k) instances of −α−2 ,
3. k eigenvalues α−2 (1 − 2sl )/(1 − sl ), where l = 1, . . . , k,
4. k(k − 1) eigenvalues which come in pairs of the form given in Eqn. (140).
190
5.11
Proof of Proposition 17 (Section 3.18)
ζ |=Σ(n) ξ iff ∀z .∧2 : z .∧2 |=Σ(n) ζ → z .∧2 |=Σ(n) ξ
iff ∀z .∧2 : z .∧2 .∗ (z .∧2 + 1)−1 ≤ ι(ζ) → z .∧2 .∗ (z .∧2 + 1)−1 ≤ ι(ξ)
iff ι(ζ) ≤ ι(ξ),
where the last step rests on the fact that all vector components of ζ, ξ are at most
1 and that the set of vectors of the form z .∧2 .∗ (z .∧2 +1)−1 is the set of nonnegative
vectors with components less than 1.
191
References
[1] L. F. Abbott. Theoretical neuroscience rising. Neuron, 60(November 6):489–
495, 2008.
[2] D.H. Ackley, G.E. Hinton, and T.J. Sejnowski. A learning algorithm for
Boltzmann machines. Cognitive Science, 9:147–169, 1985.
[3] D. J. Amit, H. Gutfreund, and H. Sompolinsky. Spin-glass models of neural
networks. Phys. Rev. A, 32:1007–1018, 1985.
[4] J. R. Anderson, D. Bothell, M. D. Byrne, S. Douglass, C. Lebiere, and Y .
Qin. An integrated theory of the mind. Psychological Review, 111(4):1036–
1060, 2004.
[5] L. Appeltant, M. C. Soriano, G. Van der Sande, J. Danckaert, S. Massar, J. Dambre, B. Schrauwen, C. R. Mirasso, and I. Fischer. Information
processing using a single dynamical node as complex system. Nature Communications, 2(468), 2011. DOI: 10.1038/ncomms1476.
[6] A. Babloyantz and C. Lourenço. Computation with chaos: A paradigm for
cortical activity. Proceedings of the National Academy of Sciences of the
USA, 91:9027–9031, 1994.
[7] S. Bader, P. Hitzler, and S. Hölldobler. Connectionist model generation: A
first-order approach. Neurocomputing, 71(13):2420–2432, 2008.
[8] D. S. Bernstein. Matrix Mathematics, 2nd Edition. Princeton Univ. Press,
2009.
[9] R. V. Borges, A. Garcez, and L. C. Lamb. Learning and representing temporal knowledge in recurrent networks. IEEE Trans. on Neural Networks,
22(12):2409–2421, 2011.
[10] R.A. Brooks. The whole iguana. In M. Brady, editor, Robotics Science,
pages 432–456. MIT Press, Cambridge, Mass., 1989.
[11] M. Brown and C. Harris. Neurofuzzy Adaptive Modelling and Control. Prentice Hall, 1994.
[12] M. Buehner and P. Young. A tighter bound for the echo state property.
IEEE Transactions on Neural Networks, 17(3):820– 824, 2006.
[13] F. S. Chance and L. F. Abbott. Divisive inhibition in recurrent networks.
Network: Comput. Neural Syst., 11:119–129, 2000.
[14] S. P. Chatzis. Hidden Markov models with nonelliptically contoured state
densities. IEEE Trans. on Pattern Analysis and Machine Intelligence,
32(12):2297 – 2304, 2010.
192
[15] A. Clark. Whatever next? Predictive brains, situated agents, and the future
of cognitive science. Behavioral and Brain Sciences, pages 1–86, 2012.
[16] A. Coates, P. Abbeel, and A. Y. Ng. Learning for control from multiple
demonstrations. In Proc. 25th ICML, Helsinki, 2008.
[17] A. M. Collins and M. r. Quillian. Retrieval time from semantic memory.
Journal of verbal learning and verbal behavior, 8(2):240–247, 1969.
[18] HBP Consortium. The Human Brain Project. Report to the European Commission, EPFL Lausanne, 2012. URL: http://www.humanbrainproject.eu/.
[19] P. Dominey, M. Arbib, and J.-P. Joseph. A model of corticostriatal plasticity
for learning oculomotor associations and sequences. Journal of Cognitive
Neuroscience, 7(3):311–336, 1995.
[20] P. F. Dominey. From sensorimotor sequence to grammatical construction: Evidence from simulation and neurophysiology. Adaptive Behaviour,
13(4):347–361, 2005.
[21] R. Douglas and T. Sejnowski. Future challenges for the sciene and engineering of learning: Final workshop report. Technical report, National Science
Foundation, 2008.
[22] G.L. Drescher. Made-up Minds: A Constructivist Approach to Artificial
Intelligence. MIT Press, Cambridge, Mass., 1991.
[23] D. Durstewitz, J. K. Seamans, and T. J. Sejnowski. Neurocomputational
models of working memory. Nature Neuroscience, 3:1184–91, 2000.
[24] C. Eliasmith. A unified approach to building and controlling spiking attractor networks. Neural Computation, 17:1276–1314, 2005.
[25] C. Eliasmith. Attractor network. Scholarpedia, 2(10):1380, 2007.
[26] C. Eliasmith, Stewart T. C., Choo X., Bekolay T., Tang Y. DeWolf T.,
and D. Rasmussen. A large-scale model of the functioning brain. Science,
338(6111):1202–1205, 2012.
[27] K. Fan and G. Pall. Imbedding conditions for Hermitian and normal matrices. Canad. J. Math., 9:298–304, 1957.
[28] B. Farhang-Boroujeny. Adaptive Filters: Theory and Applications. Wiley,
1998.
[29] J.A. Fodor and Z.W. Pylyshin. Connectionism and cognitive architecture:
A critical analysis. Cognition, 28:3–71, 1988.
193
[30] W. J. Freeman. Definitions of state variables and state space for braincomputer interface. part 1: Multiple hierarchical levels of brain function.
Cognitive Neurodynamics, 1(1):3–14, 2007.
[31] W. J. Freeman. Definitions of state variables and state space for braincomputer interface. part 2. extraction and classification of feature vectors.
Cognitive Neurodynamics, 1(2):85–96, 2007.
[32] R. M. French. Catastrophic interference in connectionist networks. In
L. Nadel, editor, Encyclopedia of Cognitive Science, volume 1, pages 431–
435. Nature Publishing Group, 2003.
[33] K. Friston. A theory of cortical response. Phil. Trans. R. Soc. B, 360:815–
836, 2005.
[34] M. Galtier, O. D. Faugeras, and P. C. Bressloff. Hebbian learning of recurrent
connections: a geometrical perspective. Neural Computation, 24(9):2346–
2383, 2012.
[35] T. Gedeon and D. Arathorn. Convergence of map seeking circuits. J. Math.
Imaging Vis., 29:235–248, 2007.
[36] W. Gerstner, H. Sprekeler, and D. Deco. Theory and simulation in neuroscience. Science, 338(5 Oct):60–65, 2012.
[37] J. Goguen and R. Burstall. Institutions: Abstract model theory for specification and programming. J. of the ACM, 39(1):95–146, 1992.
[38] A. Graves, M. Liwicki, S. Fernandez, R. Bertolami, H. Bunke, and J. Schmidhuber. A novel connectionist system for unconstrained handwriting recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence,
31(5):855 – 868, 2009.
[39] A. Graves and J. Schmidhuber. Offline handwriting recognition with multidimensional recurrent neural networks. In Proc. NIPS 2008. MIT Press,
2008.
[40] S. Grillner. Biological pattern generation: The cellular and computational
logic of networks in motion. Neuron, 52:751–766, 2006.
[41] C. Gros and G. Kaczor. Semantic learning in autonomously active recurrent
neural networks. Logic Journal of the IGPL, 18(5):686–704, 2010.
[42] S. Grossberg. Linking attention to learning, expectation, competition, and
consciousness. In L. Itti, G. Rees, and J. Tsotsos, editors, Neurobiology of
attention, chapter 107, pages 652–662. San Diego: Elsevier, 2005.
[43] S. Grossberg. Adaptive resonance theory. Scholarpedia, 8(5):1569, 2013.
194
[44] S. Harnad. The symbol grounding problem. Physica, D42:335–346, 1990.
[45] D. O. Hebb. The Organization of Behavior. New York: Wiley & Sons, 1949.
[46] M. Hermans and B. Schrauwen. Recurrent kernel machines: Computing
with infinite echo state networks. Neural Computation, 24(1):104–133, 2012.
[47] G. E. Hinton and R. R. Salakuthdinov. Reducing the dimensionality of data
with neural networks. Science, 313(July 28):504–507, 2006.
[48] J. J. Hopfield. Neural networks and physical systems with emergent collective computational abilities. Proc. NatL Acad. Sci. USA, 79:2554–2558,
1982.
[49] O. Houdé and N. Tzourio-Mazoyer. Neural foundations of logical and mathematical cognition. Nature Reviews Neuroscience, 4(June 2003):507–514,
2003.
[50] A. J. Ijspeert. Central pattern generators for locomotion control in animals
and robots: A review. Neural Networks, 21:642–653, 2008.
[51] T. R. Insel, S. C. Landis, and F. S. Collins. The NIH BRAIN initiative.
Science, 340(6133):687–688, 2013.
[52] M. Ito and J. Tani. Generalization in learning multiple temporal patterns
using RNNPB. In Neural Information Processing, number 3316 in LNCS,
pages 592–598. Springer Verlag, 2004.
[53] H. Jaeger. Identification of behaviors in an agent’s phase space. Arbeitspapiere der GMD 951, GMD, St. Augustin, 1995.
[54] H. Jaeger. The ”echo state” approach to analysing and training recurrent
neural networks. GMD Report 148, GMD - German National Research
Institute for Computer Science, 2001.
[55] H. Jaeger. Reservoir self-control for achieving invariance against slow input
distortions. technical report 23, Jacobs University Bremen, 2010.
[56] H. Jaeger and H. Haas. Harnessing nonlinearity: Predicting chaotic systems
and saving energy in wireless communication. Science, 304:78–80, 2004.
[57] H. Jaeger, M. Lukosevicius, D. Popovici, and U. Siewert. Optimization and
applications of echo state networks with leaky integrator neurons. Neural
Networks, 20(3):335–352, 2007.
[58] T. Kohonen and T. Honkela. Kohonen network. In Scholarpedia, volume 2,
page 1568. 2007.
195
[59] A. Krizhevsky, I. Sutskever, and G. Hinton. Imagenet classification with
deep convolutional neural networks. In Advances in Neural Information
Processing Systems 25, pages 1106–1114, 2012.
[60] M. Kudo, J. Toyama, and M. Shimbo. Multidimensional curve classification
using passing-through regions. Pattern Recognition Letters, 20(11):1103–
1111, 1999.
[61] R. Laje and D. V. Buonomano. Robust timing and motor patterns by taming
chaos in recurrent neural networks. Nature Neuroscience, 16(7):925–933,
2013.
[62] G. Lakoff. Women, fire, and dangerous things: What categories reveal about
the mind. University of Chicago, 1987.
[63] G. Lakoff. Cognitive models and prototype theory. In I. Margolis and S. Laurence, editors, Concepts: Core Readings, chapter 18, pages 391–422. Bradford Books / MIT Press, 1999.
[64] G. Lakoff and R. E. Nunez. Where mathematics comes from: How the
embodied mind brings mathematics into being. Basic Books, 2000.
[65] L. C. Lamb. The grand challenges and myths of neural-symbolic computation. In L. De Raedt, B. Hammer, P. Hitzler, and W. Maass, editors,
Recurrent Neural Networks- Models, Capacities, and Applications, number
08041 in Dagstuhl Seminar Proceedings, Dagstuhl, Germany, 2008. Internationales Begegnungs- und Forschungszentrum für Informatik (IBFI), Schloss
Dagstuhl, Germany.
[66] Y. LeCun, L. Bottou, J. Bengio, and P. Haffner. Gradient-based learning
applied to document recognition. Proceedings of the IEEE, 86(11):2278–
2324, 1998. http://yann.lecun.com/exdb/publis/pdf/lecun-98.pdf.
[67] F. Lehmann. Semantic networks in artificial intelligence. Elsevier Science,
1992.
[68] L. Lukic, J. Santos-Victor, and A. Billard. Learning coupled dynamical
systems from human demonstration for robotic eye-arm-hand coordination.
In Proc. IEEE-RAS International Conference on Humanoid Robots, Osaka
2012, 2012.
[69] M. Lukoševičius and H. Jaeger. Reservoir computing approaches to recurrent
neural network training. Computer Science Review, 3(3):127–149, 2009.
[70] W. Maass, P. Joshi, and E. Sontag. Computational aspects of feedback in
neural circuits. PLOS Computational Biology, 3(1):1–20, 2007.
196
[71] G. Manjunath and H. Jaeger. Echo state property linked to an input: Exploring a fundamental characteristic of recurrent neural networks. Neural
Computation, 25(3):671–696, 2013.
[72] N. M. Mayer and M. Browne. Echo state networks and self-prediction. In Biologically Inspired Approaches to Advanced Information Technology, volume
3141 of LNCS, pages 40–48. Springer Verlag Berlin / Heidelberg, 2004.
[73] S. McCallum. Catastrophic Forgetting and the Pseudorehearsal Solution in
Hopfield Networks. Phd thesis, Univ. of Otago, NZ, 2007.
[74] W. S. McCulloch and W. Pitts. A logical calculus of the ideas immanent in
nervous activity. Bull. of Mathematical Biophysics, 5:115–133, 1943.
[75] D. L. Medin and L. J. Rips. Concepts and categories: Memory, meaning, and
metaphysics. In K. J. Holyoak and R. G. Morrison, editors, The Cambridge
Handbook of Thinking and Reasoning, chapter 3, pages 37–72. Cambridge
University Press, 2005.
[76] T. M. Mitchell. Machine Learning. McGraw-Hill, 1997.
[77] O-M. Moe-Helgensen and H. Stranden. Catastophic forgetting in neural
networks. Technical report, Norwegian University of Science and Technology,
2005.
[78] M. Negrello and F. Pasemann. Attractor landscapes and active tracking:
The neurodynamics of embodied action. Adaptive Behaviour, 16:196 – 216,
2008.
[79] K. Obermayer, H. Ritter, and K. Schulten. A principle for the formation of
the spatial structure of cortical feature maps. Proc. of the National Academy
of Sciences of the USA, 87:8345–8349, 1990.
[80] E. Oja. A simplified neuron model as a principal component analyzer. J.
Math. Biol., 15:267–273, 1982.
[81] C. Orsenigo and C. Vercellis. Combining discrete SVM and fixed cardinality warping distances for multivariate time series classification. Pattern
Recognition, 43(11):3787–3794, 2010.
[82] G. Palm. On associative memory. Biol. Cybernetics, 36(1):19–31, 1980.
[83] R. Pfeifer and Ch. Scheier. Understanding Intelligence. MIT Press, 1999.
[84] G. Pinkas. Propositional non-monotonic reasoning and inconsistency in symmetric neural networks. In Proc. 12th international joint conference on Artificial intelligence - Volume 1, pages 525–530, 1991.
197
[85] J. B. Pollack. Recursive distributed representations. Artificial Intelligence,
46(1-2):77–105, 1990.
[86] L. Qi.
Symmetric nonnegative tensors and copositive tensors.
arxiv.org/pdf/1211.5642, 2012.
[87] M. R. Quillain. Word concepts: A theory and simulation of some basic
semantic capabilities. Behavioral Science, 12(5):410–430, 1967.
[88] F. Rabe. Representing Logics and Logic Translations. Phd thesis, School of
Engineering and Science, Jacobs University Bremen, 2008.
[89] M. I. Rabinovich, R. Huerta, P. Varona, and V. S. Afraimovich. Transient
cognitive dynamics, metastability, and decision making. PLOS Computational Biology, 4(5):e1000072, 2008.
[90] F. R. Reinhart and J. J. Steil. Recurrent neural associative learning of forward and inverse kinematics for movement generation of the redundant pa-10
robot. In A. Stoica, E. Tunsel, T. Huntsberger, T. Arslan, S. Vijayakumar,
and A. O. El-Rayis, editors, Proc. LAB-RS 2008, vol. 1, pages 35–40, 2008.
[91] R. F. Reinhart, A. Lemme, and J. J. Steil. Representation and generalization of bi-manual skills from kinesthetic teaching. In Proc. of IEEE-RAS
International Conference on Humanoid Robots, Osaka, 2012, in press.
[92] R. F. Reinhart and J. J. Steil. A constrained regularization approach for
input-driven recurrent neural networks. Differential Equations and Dynamical Systems, 19(1–2):27–46, 2011. DOI 10.1007/s12591-010-0067-x is an 2010
online pre-publication.
[93] J. J. Rodriguez, C. J. Alonso, and J. A. Maestro. Support vector machines
of interval-based features for time series classification. Knowledge-Based
Systems, 18(4-5):171–178, 2005.
[94] J. S. Rothman, L. Cathala, V. Steuber, and R. A. Silver. Synaptic depression
enables neuronal gain control. Nature, 457(19 Feb):1015–1018, 2009.
[95] G. Schöner, M. Dose, and C. Engels. Dynamics of behavior: theory and
applications for autonomous robot architectures. Robotics and Autonomous
Systems, 16(2):213–246, 1995.
[96] G. Schöner and J. A. Kelso. Dynamic pattern generation in behavioral and
neural systems. Science, 239(4847):1513–1520, 1988.
[97] J.R. Searle. Minds, brains, and programs. The Behavioral and Brain Sciences, 3:417–457, 1980.
198
[98] L. Shastri. Advances in Shruti – a neurally motivated model of relational
knowledge representation and rapid inference using temporal synchrony. Artificial Intelligence, 11:79–108, 1999.
[99] K. R. Sivaramakrishnan, K. Karthik, and C. Bhattacharyya. Kernels for
large margin time-series classification. In Proc. IJCNN 2007, pages 2746 –
2751, 2007.
[100] L.B. Smith and E. Thelen, editors. A Dynamic Systems Approach to Development: Applications. Bradford/MIT Press, Cambridge, Mass., 1993.
[101] T. Strauss, W. Wustlich, and R. Labahn. Design strategies for weight matrices of echo state networks. Neural Computation, 24(12):3246–3276, 2012.
[102] D. Sussillo and L. Abbott. Transferring learning from external to internal weights in echo-state networks with sparse connectivity. PLoS ONE,
7(5):e37372, 2012.
[103] D. Sussillo and O Barak. Opening the black box: Low-dimensional dynamics in high-dimensional recurrent neural networks. Neural Computation,
25(3):626–649, 2013.
[104] M. Timme, F. Wolf, and Th. Geisel. Unstable attractors induce perpetual synchronization and desynchronization. Chaos, 13:377–387, 2003.
http://arxiv.org/abs/cond-mat/0209432.
[105] I. Tsuda. Towards an interpretation of dynamic neural activity in terms of
chaotic dynamical systems. Behavioural and Brain Sciences, 24(5):793–810,
2001.
[106] F. van der Velde and M. de Kamps. Neural blackboard architectures of combinatorial structures in cognition. Behavioural and Brain Sciences, 29(1):37–
70, 2006.
[107] T. van Gelder. The dynamical hypothesis in cognitive science. Behavioural
and Brain Sciences, 21(5):615–628, 1998.
[108] D. Verstraeten. Reservoir Computing: Computation with Dynamical Systems. PhD thesis, Electronics and Information Systems, University of Ghent,
2009.
[109] D.M. Wolpert and M. Kawato. Multiple paired forward and inverse models
for motor control. Neural Networks, 11(7-8):1317–1330, 1998.
[110] F. wyffels, J. Li, T. Waegeman, B. Schrauwen, and H. Jaeger. Frequency
modulation of large oscillatory neural networks. Biological Cybernetics,
108:145–157, 2014.
199
[111] Y. Yao and W.J. Freeman. A model of biological pattern recognition with
spatially chaotic dynamics. Neural Networks, 3(2):153–170, 1990.
[112] I. B. Yildiz, H. Jaeger, and S. J. Kiebel. Re-visiting the echo state property.
Neural Networks, 35:1–20, 2012.
[113] S. K. U. Zibner, C. Faubel, I. Iossifidis, and G. Schöner. Dynamic neural
fields as building blocks of a cortex-inspired architecture for robotic scene
representation. IEEE Trans. on Autonomous Mental Development, 3(1):74–
91, 2011.
200
| 9cs.NE
|
arXiv:1604.03857v1 [math.GR] 13 Apr 2016
THE TORSION-FREE RANK OF HOMOLOGY IN TOWERS OF
SOLUBLE PRO-p GROUPS
MARTIN R. BRIDSON AND DESSISLAVA H. KOCHLOUKOVA
Abstract. We show that for every finitely presented pro-p nilpotent-by-abelianby-finite group G there is an upper bound on dimQp (H1 (M, Zp ) ⊗Zp Qp ), as
M runs through all pro-p subgroups of finite index in G.
1. Introduction
In [4] we examined how betti numbers behave in towers of finitely presented
soluble groups and proved that if G is finitely presented and nilpotent-by-abelianby-finite, then there is an upper bound on dimQ H1 (M, Q), as M runs through all
subgroups of finite index in G. In this paper we shall prove the corresponding result
for pro-p groups.
By definition, a pro-p group G is finitely presented if G ∼
= F/R, where F is a
finitely generated free pro-p group and R is a pro-p subgroup of F generated (as a
topological group) by the union of finitely many conjugacy classes in F . For discrete
groups, one cannot detect finite presentability using homological conditions, but for
pro-p groups one can: G is finitely presented if and only if the homology groups
H1 (G, Zp ) and H2 (G, Zp ) are finitely generated. Moreover, the number of elements
required to generate G as a topological group is d(G) := dimFp H1 (G, Fp ).
In [12] John Wilson proved that the Golod-Shafarevich inequality holds for soluble pro-p groups. Using this, he proved [12, Corollary A,(i)] that for every finitely
presented soluble pro-p group G there is a constant k > 0 such that for any pro-p
subgroup U of finite index in G, the inequality d(U ) ≤ k[G : U ]1/2 holds. He also
proved that every normal pro-p subgroup with quotient Zp is finitely generated
[12, Corollary A, (ii)]. Note how this contrasts with the discrete case: there are
many finitely presented solvable groups N ⋊ Z with N not finitely generated. (In
the positive direction, Bieri, Neumann and Strebel [3, Thm. D] proved that if G is
finitely presented and soluble group with rk (G/G′ ) ≥ 2, then there exists a normal
finitely generated subgroup N such that G/N ∼
= Z.)
Our first result generalizes [12, Cor. A,(ii)].
Proposition A. Let G be a finitely presented soluble pro-p group. Then
sup
G/H ∼
=Zp
d(H) < ∞.
The following result forms the technical heart of this paper. It is proved using
methods from commutative algebra. It will enable us to retain the finiteness in
homology given by Proposition A as we extend scalars. We remind the reader that
Date: Oxford, 11 April 2016.
2000 Mathematics Subject Classification. Primary 20J05 Secondary 20E18 .
Key words and phrases. pro-p group, virtual first betti number.
1
2
MARTIN R. BRIDSON AND DESSISLAVA H. KOCHLOUKOVA
for a pro-p group G which is the inverse limit of its finite p-group quotients G/U ,
homology is defined in terms of modules over the completed group algebra Zp [[G]],
which is the inverse limit of the rings Zp [G/U ]. Note that Zp [[G]] is a local ring
whose unique maximal ideal is the kernel of the natural morphism Zp [[G]] ։ Fp .
The torsion-free rank of an abelian pro-p group B is
rk B := dimQp B⊗Zp Qp .
Proposition B. Let Q be a free abelian pro-p group of finite rank and let A be a
finitely generated pro-p Zp [[Q]]-module. Let B be the set of pro-p subgroups H ⊳ Q
with Q/H ∼
= Zp . If supH∈B dimFp (A⊗Zp [[H]] Fp ) < ∞, then
sup rk (A⊗Zp [[Qps ]] Zp ) < ∞.
s≥1
To prove the following theorem, we will reduce to the metabelian case and then
apply Proposition B.
Theorem C. Let G be a finitely generated pro-p group that is nilpotent-by-(torsionfree abelian). If supG/H ∼
=Zp d(H) < ∞, then
sup rk H1 (M, Zp ) < ∞,
M∈A
where A is the set of all pro-p subgroups of finite index in G.
By combining Theorem C with Proposition A and using the fact that a subgroup
of finite index in a finitely presented group is finitely presented, we obtain the result
stated in the abstract.
Corollary D. For every finitely presented pro-p nilpotent-by-abelian-by-finite group
G, there is an upper bound on rk H1 (M, Zp ), as M runs through all pro-p subgroups
of finite index in G.
We separated Theorem C from Corollary D because, as we shall explain in Section
6, there are interesting examples of metabelian pro-p groups, arising in the work
of Jeremy King [8], that are not finitely presented but still satisfy the conclusion
of Proposition A. In Section 6 we shall also describe examples which show that
Theorem C can fail when one changes the field of coefficients from Qp to Fp .
The analogue of Corollary D in the setting of discrete groups is this: if G is
finitely presented and nilpotent-by-abelian-by-finite, then there is an upper bound
on dimQ H1 (M, Q), where M runs through all subgroups of finite index in G. This
was our main result in [4]. It was proved independently by Andrei Jaikin Zapirain
[6] in the metabelian case under the weaker hypothesis dimQ H0 (Q, H2 (N, Q)) < ∞.
Acknowledgements: We thank Andrei Jaikin Zapirain for a suggestion that
allowed us to remove an unnecessary hypothesis from an earlier version of Theorem
C and we thank the referee for the helpful comments that improved the paper.
The work of the first author was supported by grants from the EPSRC and by a
Wolfson Merit Award from the Royal Society; the work of the second author was
supported by “bolsa de produtividade em pesquisa”, CNPq, Brazil: we thank all
of these organizations.
THE TORSION-FREE RANK OF HOMOLOGY IN TOWERS OF SOLUBLE PRO-p GROUPS 3
2. Preliminaries
2.1. Generating sets. For a pro-p group G with a subset X denote by hXi the
pro-p subgroup generated by X i.e. the smallest pro-p subgroup of G that contains
X. If G = hXi we say that X is a generating set for G.
2.2. Completed tensor products. Let R be a profinite ring, V1 a profinite right
b R V2
R-module and V2 a profinite left R-module. The profinite tensor product V1 ⊗
can be defined either by a universal property or as an inverse limit of finite abstract
tensor products [11]. In the case when at least one of V1 and V2 is finitely presented
b R V2 ∼
as a profinite R-module, there is a natural isomorphism V1 ⊗
= V1 ⊗R V2 (see
[1, Cor. 2.4]). Thus, whenever possible (e.g. Proposition B and Theorem C) we use
b to emphasize the need for completed tensor products
the notation ⊗, reserving ⊗
when we do not know whether either of the profinite modules is finitely presented.
In this paper, the only profinite modules we consider are pro-p modules over pro-p
rings.
2.3. Noetherian completed group algebras. Let Q be a finitely generated prop abelian group. Q ∼
=
= Q0 ⊕ K, where K is a finite abelian p-group and Q0 ∼
n
Zp is free-abelian pro-p. Note that Zp [[Q0 ]] ∼
= Zp [[t1 , . . . , tn ]] is the power series
ring in n commuting variables, where under the identification, Q0 is generated
by {1 + t1 , . . . , 1 + tn }. In particular, Zp [[Q0 ]] is abstractly Noetherian. Moroever,
since Zp [[Q]] is finitely generated as an abstract Zp [[Q0 ]]-module, it too is abstractly
Noetherian.
According to [13, Lemma 7.2.2], for every pro-p group G, every pro-p Zp [[G]]module V that is topologically generated by a finite set X is generated as an abstract
Zp [[G]]-module by the same set X. In this sense, the topology of V is hidden in
the topology of the ring Zp [[G]] that acts on V .
Thus we see that for a finitely generated abelian pro-p group Q, the ring Zp [[Q]] is
Noetherian in both the abstract and pro-p sense. In fact, every abstract ideal (topology ignored) of Zp [[Q]] is finitely generated (as an abstract Zp [[Q]]-submodule),
moreover it is closed ([13, Lemma 7.2.2]), and therefore is a finitely generated pro-p
Zp [[Q]]-submodule of Zp [[Q]]. Thus when we work with ideals in Zp [[Q]], we do not
need to specify that they are closed, because this condition is automatic. This freedom to ignore the topology allows us to apply the methods of commutative algebra
for Noetherian rings, exploiting in particular the fact that above every proper ideal
of Zp [[Q]] there are only finitely many minimal prime ideals.
2.4. Primary decomposition in Noetherian rings. We recall some basic ideas
from commutative algebra; see [5, Chapter 3] or [2]. Let R be a commutative
Noetherian ring with unit and M a finitely generated R-module. An associate
prime of M is a prime ideal P of R such that R/P arises as an R-submodule of
M (i.e. P = annR (m) for some m ∈ M ). The set of associated primes of M
is finite and is denote by Ass(M ). All of the minimal primes among the prime
ideals containing annR (M ) are associated primes of M . An R-module W is call
P -primary if Ass(W ) = {P }. In this case P acts nilpotently on W .
For any R-submodule V of M there is a primary decomposition V = V1 ∩ V2 ∩
. . . ∩ Vs , where for each Vi we have that M/Vi is Pi -primary and the prime ideals Pi
are pairwise distinct (although one can be included in another). The prime ideals
4
MARTIN R. BRIDSON AND DESSISLAVA H. KOCHLOUKOVA
P1 , . . . , Ps are the associated primes of M/V . The component Vi is unique if Pi is
minimal (with respect to inclusion) among the associated primes Ass(M/V ).
We will apply the primary decomposition for R = Zp [[Q]], Q = Znp and M a
finitely generated pro-p R-module. As explained in the previous subsection, we are
free to ignore the topological structure of M and view it as an abstract R-module.
All abstract R-submodules of M are finitely generated as abstract R-submodules
and are automatically pro-p R-submodules.
3. Proof of Proposition A
We have G = H ⋊ H0 , where H0 = hqi ∼
= Zp . As we noted in the introduction,
Wilson [12, Corollary A, (ii)] proved that H is finitely generated. On other hand,
using that H/H ′ H p is a finite Fp [[H0 ]]-module and Fp [[H0 ]] is a principal ideal
domain, we get a decomposition
H/H ′ H p ∼
= (Fp [[H0 ]]/I1 ) ⊕ . . . ⊕ (Fp [[H0 ]]/Is ),
where H ′ is the pro-p subgroup of H generated by commutators, H p is the pro-p
subgroup of H generated by all the set {hp | h ∈ H}, every Fp [[H0 ]]/Ij is finite,
and the isomorphism is of Fp [[H0 ]]-modules. There is an isomorphism of modules
Fp [[H0 ]]/Ij ∼
= Fp [[t]]/(tij ),
with q acting as multiplication by t + 1. (Here we have used the fact that every
ideal of Fp [[H0 ]] is of the form (ti ).) Let
m = max{i1 , . . . , is } ≥ 1.
By [12, Cor. B], there is a constant k such that for every pro-p subgroup U of finite
index in G we have
d(U ) ≤ k[G : U ]1/2 .
(3.1)
We apply this to
j
U = hH, q p i, where pj > m ≥ pj−1 ≥ 1,
noting that since U is a split extension of H by the pro-p subgroup generated by
j
qp ,
(3.2)
U/U ′ U p ∼
Fp ) ⊕ Fp ,
= (H/H ′ H p ⊗
pj
Fp [[hq
i]]
j
j
where the second summand Fp is generated by the image of q p . The action of q p
j
on Fp [[t]]/(tij ) is given by multiplication with the image of (1 + t)p which, since
j
j
we are in characteristic p, is 1 + tp . Since pj ≥ m ≥ ij we see that (1 + t)p acts
as the identity. Thus
(H/H ′ H p ) ⊗
Fp ∼
= H/H ′ H p ,
pj
Fp [[hq
i]]
and by (3.2) we have
d(U ) = d((H/H ′ H p ) ⊗Fp [[hqpj i]] Fp ) + 1 = dim(H/H ′ H p ) + 1 = i1 + . . . + is + 1.
Combining this with (3.1) we get
thus
i1 + . . . + is + 1 ≤ kpj/2 ,
(i1 + . . . + is )2 < k 2 pj = k 2 ppj−1 ≤ k 2 pm ≤ k 2 p(i1 + . . . + is ),
THE TORSION-FREE RANK OF HOMOLOGY IN TOWERS OF SOLUBLE PRO-p GROUPS 5
and
i1 + . . . + is < k 2 p.
Therefore,
d(H) = i1 + . . . + is < k 2 p.
4. Proof of Proposition B
Throughout this section we write R for the completed group algebra Zp [[Q]].
This is a local ring whose unique maximal ideal is the kernel P of the natural
morphism Zp [[Q]] ։ Fp .
We are given a free abelian pro-p group of finite rank Q and a finitely generated
R-module A, which we know to be Noetherian. We assume that as H < Q ranges
over the normal pro-p subgroups with Q/H ∼
= Zp , dimFp (A⊗Zp [[H]] Fp ) remains
bounded, and from this we must deduce that rk (A⊗Zp [[Qps ]] Zp ) remains bounded
as the positive integer s increases. Roughly speaking, the idea of the proof is to
reduce to the case of a cyclic module (with prime annihilator), and we do this by
examining the primary decomposition. The key gain in finiteness comes in Lemma
4.2, from Noetherian properties of R and A, and from Lazard’s bound [10] on the
order of finite subgroups in GLd (Zp ).
Consider the R-submodule Am = AΩm in A, where m is a power of p and Ωm
is the augmentation ideal ker(Zp [[Qm ]] → Zp ). Let
Am = Nm,1 ∩ Nm,2 ∩ . . . ∩ Nm,j
be a primary decomposition of Am in A. Then
Ass(A/Nm,i ) = {Pm,i },
where Pm,i is a prime ideal in R, and
Ass(A/Am ) = {Pm,1 , . . . , Pm,j }.
Note that, by the definition of an associate prime ideal, R/Pm,i embeds in A/Am ,
hence Pm,i contains Jm = annR (A/Am ). Thus
(4.1)
Qm − 1 ⊂ Jm = annR (A/Am ) ⊆ Pm,i .
Now, Zp [[Q/Qm ]] = Zp [Q/Qm ] has Krull dimension 1, so either Pm,i is the unique
maximal ideal P ⊳ R or else Pm,i is a minimal among the prime ideals containing
Jm and there is no prime ideal P0 lying strictly between Pm,i and P . Furthermore,
by [5, Thm. 3.1], every ideal in R that is minimal among the prime ideals containing
Jm is an associated prime ideal of A/Am .
The following lemma is used to limit the possibilities for the domain R/Pm,i .
Recall that B is the set of pro-p subgroups H ⊳ Q with Q/H ∼
= Zp .
Lemma 4.1. For each Pm,i there exists H0 ∈ B such that H0 − 1 ⊆ Pm,i .
Proof. Let Fm,i be the abstract field of fractions of the (abstract) domain R/Pm,i .
The lemma is equivalent to the assertion that the image of Q in Fm,i is a cyclic
group. Note that we already know that the image of Q in Fm,i is finite.
We consider two cases. First suppose that Fm,i has characteristic 0. Then the
subfield in Fm,i generated by Q and the image of Q is inside the spliting field Km,i
(over Q) of the polynomial xs −1, where s is the exponent of the image of Q in Fm,i .
By the uniqueness of the splitting field (up to isomophism), Km,i is isomorphic to
6
MARTIN R. BRIDSON AND DESSISLAVA H. KOCHLOUKOVA
the subfield of C generated by Q and a primitive s-th root of unity. In particular
the image of Q in Fm,i is cyclic, as claimed.
Suppose now that Fm,i has positive characteristic. Since the elements of Zp rpZp
are invertible in Zp , the characteristic of Fm,i is p. By (4.1) Fm,i is a finite field,
∗
hence Fm,i
is a cyclic group of order coprime to p. It follows that the image of Q
in Fm,i is trivial and Fm,i = Fp , so Pm,i = P and the lemma holds.
Lemma 4.2. Let Xm = {Pm,1 , . . . , Pm,j } and define X = ∪m≥1 Xm . Then X is
finite.
Proof. Consider Pm,i 6= P . Write Wm,i = A/Nm,i and note that by [5, Prop. 3.9]
sm,i
= 0.
Pm,i acts nilpotently on Wm,i i.e. there is some sm,i ∈ N such that Wm,i Pm,i
We claim that
annR (Vm,i ) = Pm,i where Vm,i = Wm,i /Wm,i Pm,i .
√
To see this, note that for any ideal S ⊳ R that strictly contains Pm,i , the radical S
is the intersection of the minimal elements
in the set of prime ideals containing
√
√ z S.
We are in Krull dimension one, so S = P . Since R is Noetherian, we have S ⊆
S for some z ∈ N, and therefore S has finite index in R (as an abelian subgroup).
Thus, if annR (Vm,i ) were to contain Pm,i as a proper ideal, then annR (Vm,i ) has
finite index in R, forcing Vm,i and Wm,i to be finite. But this would mean that the
domain R/Pm,i was finite, hence a field. Since P is the unique maximal ideal in R,
this would imply Pm,i = P , contrary to hypothesis.
By Lemma 4.1, there is H0 ∈ B such that H0 − 1 ⊆ Pm,i . Then H0 acts trivially
on Vm,i and so Vm,i ∼
= Vm,i ⊗Zp [[H0 ]] Zp , which is a quotient of A⊗Zp [[H0 ]] Zp . By
assumption
sup dimFp (A⊗Zp [[H]] Fp ) < ∞.
H∈B
Recall that the minimal number of generators of a pro-p group B is
d(B) = dimFp B/B ′ B p .
Since Pm,i is an associated prime for Vm,i we deduce that R/Pm,i embeds in Vm,i ,
so
d(R/Pm,i ) ≤ d(Vm,i ) = d(Vm,i ⊗Zp [[H0 ]] Zp ) ≤ d(A⊗Zp [[H0 ]] Zp ) = dimFp (A⊗Zp [[H0 ]] Fp ).
At this stage we have established the following weak form of the conclusion |X| < ∞
that we seek:
δ := sup d(R/Pm,i ) ≤ sup dimFp (A⊗Zp [[H]] Fp ) < ∞.
Pm,i ∈X
H∈B
As before, we write Fm,i for the abstract field of fractions of the domain R/Pm,i .
(Remember that R = Zp [[Q]].) Let Qm,i be the image of Q in R/Pm,i . From
Lemma 4.1 we know that Qm,i is finite cyclic, of p-power order.
If the characteristic of Fm,i was positive, it would be p. But then, as in the
proof of Lemma 4.1, we would get Pm,i = P , a contradiction. Therefore, Fm,i
d(j)
has characteristic 0, in which case R/Pm,i ∼
= Zp for some d(j) ≤ δ, and Qm,i
occurs as a finite cyclic subgroup of GLδ (Zp ). This is a virtually pro-p analytic,
and therefore it has finite virtual cohomological dimension [10]. (We refer to [14]
for background on pro-p analytic groups.) It follows that there is an upper bound
on the finite subgroups of GLd (Zp ), yielding an integer r (a power of p) such that
the order of each Qm,i always divides r.
THE TORSION-FREE RANK OF HOMOLOGY IN TOWERS OF SOLUBLE PRO-p GROUPS 7
We have proved
(Qr − 1) ⊆ Pm,i
for all
Pm,i ∈ X,
and the lemma now follows from the observation that in R = Zp [[Q]] there are only
finitely many prime ideals above (Qr − 1); equivalently, there are only finitely many
minimal prime ideals in the local ring Zp [[Q]]/(Qr − 1) ∼
= Zp [Q/Qr ]. To see that
r
this is the case, note that since Zp [Q/Q ] is finitely generated as a Zp -module, its
Krull dimension equals that of Zp , which is 1. Thus any prime ideal in Zp [Q/Qr ]
that is not minimal is the unique maximal prime ideal. And there are only finitely
many minimal prime ideals.
Lemma 4.3. For each p-power m, define Jm = annR (A/Am ) and Bm = Jm ⊗Zp Qp .
Then, there is a p-power m0 ≥ 1 such that Bm = Bm0 for all m ≥ m0 .
Proof. Note that if m1 ≥ m2 are powers of p, then A/Am2 is a quotient of A/Am1 ,
hence Jm1 ⊆ Jm2 . We claim that
(4.2)
Xm2 \ {P } ⊆ Xm1 \ {P },
where P is the unique maximal ideal of Zp [[Q]]. Indeed by the paragraph after
(4.1) any minimal prime ideal Pm2 ,i above Jm2 such that Pm2 ,i 6= P belongs to
Xm2 and any element of Xm2 \ {P } is of this type. Thus there is no prime ideal of
Zp [[Q]] that contains Jm1 and is strictly contained in Pm2 ,i (otherwise there would
be a nested sequence of three primes ideals above Jm1 , which would contradict the
paragraph after (4.1) with m = m1 ). Thus Pm2 ,i is a minimal prime above Jm1 , so
Pm2 ,i ∈ Xm1 and (4.2) holds.
By Lemma 4.2 X = ∪m Xm is a finite set of primes. This together with (4.2)
implies the existence of m0 , a power of p, such that
(4.3)
Xm0 \ {P } = Xm \ {P } for m ≥ m0 a power of p.
If m ≥ m0 is a power of p and X 6= {P }, then
p
p
Jm = ∩I∈Xm I = ∩I∈Xm \{P } I = ∩I∈Xm0 \{P } I = ∩I∈Xm0 I = Jm0
(4.4)
√
because, as we noted earlier, that for any ideal S ⊳ R, the radical S is the
intersection of the minimal elements in the set of prime ideals containing S. Note
that in the case X = {P } we have Xm = Xm0 = X, so
p
p
(4.5)
Jm = ∩I∈Xm I = ∩I∈Xm0 I = Jm0 for m ≥ m0 a power of p.
With equalities (4.4) and (4.5) in hand, the proof of Claim 2 in [4, Thm 3.1]
applies verbatim after replacing Q with Qp and Z with Zp . We repeat the argument
for the convenience of the reader. It suffices to show that Bm = Bmp for m ≥ m0 .
Fix q ∈ Q and write y for its image in R ⊗Zp Qp . Then y m − 1 ∈ Bm and by (4.4)
there is a positive integer s such that (y m − 1)s ∈ Bmp . The greatest common
divisor of xmp − 1 and (xm − 1)s in the polynomial ring Qp [x] is xm − 1, because
xmp − 1 has no repeated roots in a field of charateristic 0. Therefore, y m − 1 ∈ Bmp ,
by Bezout’s lemma, and hence Bm = Bmp .
Proof of Proposition B. We fix a finite set of generators a1 , . . . , as for A as an Rmodule. Then A/Am is generated as an R/Jm -module by the images of a1 , . . . , as
and we have an epimorphism of R/Jm -modules (R/Jm )s → A/Am inducing an
epimorphism of finite dimensional Qp -modules
(R/Jm )s ⊗Zp Qp → (A/Am ) ⊗Zp Qp .
8
MARTIN R. BRIDSON AND DESSISLAVA H. KOCHLOUKOVA
Thus
dimQp (A/Am ) ⊗Zp Qp ≤ s dimQp ((R/Jm ) ⊗Zp Qp ).
By Lemma 4.3, for p-powers m ≥ m0 the canonical projection R/Jmp → R/Jm
induces an isomorphism (R/Jmp ) ⊗Zp Qp → (R/Jm ) ⊗Zp Qp , so
dimQp ((R/Jm ) ⊗Zp Qp ) = dimQp ((R/Jm0 ) ⊗Zp Qp ).
Thus for C = {pi | i ≥ 1} we have
sup dimQp ((A/Am ) ⊗Zp Qp ) ≤ s. dimQp ((R/Jm0 ) ⊗Zp Qp ) < ∞.
m∈C
And, by definition, A/Am = A ⊗Zp [[Qm ]] Zp .
5. Proof of Theorem C
We shall see that in the metabelian case, Theorem C can be deduced from
Proposition B by means of straightforward calculations in homological algebra.
The following proposition reduces the general case to the metabelian case. For
a pro-p group H let {γi (H)}i≥1 be the lower central series of H. By definition
γ1 (H) = H, γi+1 (H) is the closed subgroup of H generated by [γi (H), H].
Proposition 5.1. Let N G ։ Q be a short exact sequence of pro-p groups,
where N is nilpotent, Q is abelian and G is finitely generated. Let Gn be a pro-p
subgroup of finite index in G and let Gn be the image of Gn in the metabelian group
G/N ′ . Then
H1 (Gn , Zp )⊗Zp Qp ∼
= H1 (Gn , Zp )⊗Zp Qp .
Proof. We write Qn = Gn N/N for the image of Gn in G/N . Set V = N/N ′ and
Vn = (Gn ∩ N )N ′ /N ′ . Let Vn,i be the image of Gn ∩ γi (N ) in γi (N )/γi+1 (N ). Note
that Vn,i ∼
= Gn ∩ γi (N )/Gn ∩ γi+1 (N ) and the filtration {Gn ∩ γi (N )}i≥1 of Gn ∩ N
has quotients Vn,i .
Let Wn,i denote the image of Gn ∩ γi (N ) in H1 (Gn , Zp ). Then, for i ≥ 2, we
have that Wn,i /Wn,i+1 is an image of Vn,i /γi (N ∩ Gn ), where γi (N ∩ Gn ) is the
image of γi (N ∩ Gn ) in γi (N )/γi+1 (N ).
In this paragraph all completed tensor products are over Zp . We think of
b i V via the map that sends g1 N ′ ⊗
b . . . . . . ⊗g
b i N ′ to
γi (N )/γi+1 (N ) as a quotient of ⊗
[g1 , . . . , gi ]γi+1 (N ), where the commutator is left-normed. Thus we get
(γi (N )/γi+1 (N ))/γi (N ∩ Gn )
b i V /⊗
b i Vn , where ⊗
b i Vn is the image of ⊗
b i Vn in ⊗
b i V . Since Vn
as an image of ⊗
b i V /⊗
b i Vn has finite expohas finite index in V , by induction on i we see that ⊗
nent as an abelian group, hence (γi (N )/γi+1 (N ))/γi (N ∩ Gn ) and its abelian subgroup Vn,i /γi (N ∩ Gn ) have finite exponent as abelian groups for i ≥ 2. Then
(Vn,i /γi (N ∩ Gn )) ⊗Zp Qp = 0, hence (Wn,i /Wn,i+1 ) ⊗Zp Qp = 0 for i ≥ 2. Then
Wn,2 ⊗Zp Qp = 0, so the image of Gn ∩γ2 (N ) in H1 (Gn , Zp )⊗Zp Qp is trivial. And for
the image Gn of Gn in G/N ′ we have H1 (Gn , Zp ) ⊗Zp Qp ∼
= H1 (Gn , Zp ) ⊗Zp Qp .
Lemma 5.2. Let N G ։ Q be a short exact sequence of pro-p groups with G
finitely generated. Then
(i) d(G) ≤ dimFp H0 (Q, H1 (N, Fp )) + dimFp H1 (Q, Fp ) ≤ d(G) + dimFp H2 (Q, Fp );
(ii) dimQp H1 (G, Zp )⊗ Qp ≤ dimQp H0 (Q, H1 (N, Zp ))⊗ Qp + dimQp H1 (Q, Zp )⊗ Qp ,
THE TORSION-FREE RANK OF HOMOLOGY IN TOWERS OF SOLUBLE PRO-p GROUPS 9
where all tensor products are over Zp .
Proof. Recall that for any finitely generated pro-p group, d(G) = dim H1 (G, Fp ).
Part (i) follows immediately from the following exact sequence (which is part of the
5-term exact sequence in homology)
H2 (Q, Fp ) → H0 (Q, H1 (N, Fp )) → H1 (G, Fp ) → H1 (Q, Fp ) → 0.
Part (ii) follows from the exact sequence (which is also part of a 5-term exact
sequence) H0 (Q, H1 (N, Zp )) → H1 (G, Zp ) → H1 (Q, Zp ) → 0.
Proof of Theorem C. We start with a short exact sequence of pro-p groups
M G ։ G/M with M nilpotent and G/M torsion-free and abelian. Since G/G′
is a finitely generated abelian pro-p group, it is the direct sum of a finite p-group
T and a free pro-p abelian group Znp . Let N be the full pre-image of T in G. Since
G/M is torsion-free abelian, N ⊆ M ; in particular N is nilpotent. Thus we have a
short exact sequence of pro-p groups N G ։ Znp . And by construction, N ⊂ H
for every normal H < G with G/H ∼
= Zp .
Let A = N/N ′ . If we write H for the image of H < G in G = G/N ′ , then
d(H) ≤ d(H), hence supG/H ∼
=Zp d(H) < ∞. This bound tells us that our hypothesis
on G is preserved if we replace G by G/N ′ , and Proposition 5.1 assures us that if
we can prove the theorem for G = G/N ′ , then the result will follow for G. Thus
we assume henceforth that A = N and G = G.
By applying Lemma 5.2(i) to the short exact sequence A H ։ H/A ∼
= Zpn−1
we get
n−1
sup dimFp (A⊗Zp [[H]] Fp ) ≤ sup [d(H) +
− n + 1] < ∞,
2
G/H ∼
G/H ∼
=Z p
= Zp
since H0 (H, H1 (A, Fp )) = A⊗Zp [[H]] Fp and dim H1 (H/A, Fp ) = n − 1, and the binomial coefficient is the dimension of H2 (H/A, Fp ), which (as with discrete groups)
is the exterior square of the abelian group (H/A)/p(H/A) ∼
= Fpn−1 .
From Proposition B we deduce that
(5.1)
sup rk (A⊗Zp [[Qpi ]] Zp ) < ∞.
i≥1
Let Gm be a pro-p subgroup of finite index in G and consider the following exact
sequence1
Z [[Gm /A∩Gm ]]
Tor1 p
(A/A ∩ Gm , Zp ) → (A ∩ Gm )⊗Zp [[Gm /A∩Gm ]] Zp →
A⊗Zp [[Gm /A∩Gm ]] Zp → (A/A ∩ Gm )⊗Zp [[Gm /A∩Gm ]] Zp → 0.
Because A/A ∩ Gm is finite, the first and last groups in this sequence are torsion,
so
rk ((A ∩ Gm )⊗Zp [[Gm /A∩Gm ]] Zp ) = rk (A⊗Zp [[Gm /A∩Gm ]] Zp ).
Then, by applying Lemma 5.2(ii) to A ∩ Gm Gm ։ Gm /(A ∩ Gm ), we get
dimQp (H1 (Gm , Zp ) ⊗Zp Qp ) ≤ n + rk ((A ∩ Gm )⊗Zp [[Gm /A∩Gm ]] Zp )
= n + rk (A⊗Zp [[Gm /A∩Gm ]] Zp ),
1The use of ⊗ rather than ⊗
ˆ is permitted because the trivial module Zp is finitely presented
over Zp [[G]] for every finitely generated pro-p group G.
10
MARTIN R. BRIDSON AND DESSISLAVA H. KOCHLOUKOVA
where n is the rank of Gm /(A ∩ Gm ), which has finite index in Q = G/A. Finally,
taking a p-power s such that Qs ⊆ Gm /(A ∩ Gm ) ⊆ Q, we have
rk (A⊗Zp [[Gm /A∩Gm ]] Zp ) ≤ rk (A⊗Zp [[Qs ]] Zp )
and hence
dimQp (H1 (Gm , Zp ) ⊗Zp Qp ) ≤ n + rk (A⊗Zp [[Qs ]] Zp ).
Then (5.1) completes the proof.
6. Examples
In this section we shall present two examples to show that Theorem C fails if
the coefficient field Qp is replaced by Fp . We shall then explain an example due to
Jeremy King [9] which shows that Theorem C is of interest beyond the setting of
finitely presented groups.
As we noted in the introduction, finite presentability for pro-p groups is essentially a homological condition. More precisely, a pro-p group G is finitely presented
if and only if is of type FP2 . In the pro-p setting, a group G is defined to be of type
FPm if, as a trivial Zp [[G]]-module, Zp has a projective resolution with the modules
finitely generated in dimensions ≤ m. Since Zp [[G]] is a local ring, it is easy to see
that G is of type FPm if and only if the pro-p homology groups Hi (G, Fp ) are finite
for i ≤ m.
In [8] King showed that if N is a normal pro-p subgroup of G, with G/N of
finite rank (in which case Zp [[G/N ]] is left and right Noetherian), then G is of type
FPm if and only if Hi (N, Zp ) is finitely generated as a pro-p Zp [[G/N ]]-module for
i ≤ m, where the action of G/N is induced by conjugation. In particular, if G is
a finitely generated metabelian pro-p group, i.e. there is a short exact sequence of
pro-p groups A → G → Q with A and Q abelian, then G will be finitely presented
b A is finitely generated as a Zp [[Q]]-module, where Q
if and only if H2 (A, Zp ) ∼
= A∧
b A.
acts diagonally on the completed exterior square A∧
This criterion for finitely presentability of metabelian pro-p groups plays a crucial
role in our examples: it is hidden in our appeals to King’s papers [7], [9].
6.1. The first example: switching to Fp when p is odd. Suppose p > 2 and
consider the pro-p group G = A ⋊ Q where A = Fp [[t]] and Q = hx, yi = Zp × Zp ,
with x acting as multiplication by 1 + t while y acts as multiplication by 1 + 2t. By
i
i
[9, Thm. A], G is finitely presented. And since (1 + τ )p = 1 + τ p in characteristic
p, we have
i
i
i
A ⊗Fp [[Qpi ]] Fp ∼
= Fp [[t]]/((1 + t)p − 1, (1 + 2t)p − 1) = Fp [[t]]/(tp ).
Hence
i
sup dimFp (H1 (M, Zp ) ⊗Zp Fp ) ≥ sup dimFp (H1 (A ⋊ Qp , Zp ) ⊗Zp Fp ) =
i≥1
M∈A
2 + sup dimFp (A ⊗Fp [[Qpi ]] Fp ) = 2 + sup pi = ∞,
i≥1
i≥1
where A is the set of all pro-p subgroups of finite index in G.
THE TORSION-FREE RANK OF HOMOLOGY IN TOWERS OF SOLUBLE PRO-p GROUPS11
6.2. The second example: switching to Fp when p = 2. Consider the pro-2
group G = A ⋊ Q, where A = F2 [[t]] and Q = hx, yi = Z2 × Z2 , with x acting as
multiplication by 1 + t while y acts as multiplication by 1 + t + t2 . Again, G is
finitely presented by [9, Thm. B] and
i
i
i
A ⊗F2 [[Q2i ]] F2 ∼
= F2 [[t]]/((1 + t)2 − 1, (1 + t + t2 )2 − 1) = F2 [t]]/(t2 ).
Hence
i
sup dimF2 (H1 (M, Z2 ) ⊗Z2 F2 ) ≥ sup dimF2 (H1 (A ⋊ Q2 , Z2 ) ⊗Z2 F2 ) =
M∈A
i≥1
2 + sup dimF2 (A ⊗F2 [[Q2i ]] F2 ) = 2 + sup 2i = ∞,
i≥1
i≥1
where A is the set of all pro-2 subgroups of finite index in G.
6.3. King’s example. Let p be an odd prime, let Q = hx, yi = Zp × Zp , let
A = Fp [[Q]]/(x + x−1 + y + y −1 − 4) and, following [7], consider the pro-p group
G = A ⋊ Q. By [7, Prop. 3.4] G is not finitely presented and by [7, Prop. 3.5]
for every normal pro-p subgroup N of G such that G/N ∼
= Zp we have that N is
finitely generated i.e. d(N ) < ∞. Our aim is to show that
sup
G/N ∼
=Zp
d(N ) < ∞.
Theorem C can then be applied to G, even though G is not finitely presented.
To this end, observe first that N = A ⋊ H, where H is a pro-p subgroup of Q
with Q/H ∼
= Zp . According to Lemma 5.2(i), we will be done if we can bound the
dimension of
b Fp [[H]] Fp .
H0 (H, H1 (A, Fp )) = A⊗
Since the relation that defines A is symmetric with respect to x and y, we can
assume that H = hxy −λ i for some λ ∈ Zp . Indeed we have that H = hxi y j i with
i, j ∈ Zp such that the ideal of Zp generated by i and j is the whole ring Zp . Then
i∈
/ pZp or j ∈
/ pZp . Since Zp \ pZp is the set of invertible elements in Zp we can
assume that i or j is invertible; say i is invertible. Then for λ = −ji−1
H = hxi y j i = h(xy ji
−1
)i i = hxy ji i = hxy −λ i.
−1
Then
b Fp [[H]] Fp ∼
A⊗
= Fp [[Q]]/(x+x−1 +y+y −1 −4, xy −λ −1) ∼
= Fp [[hyi]]/(y λ +y −λ +y+y −1−4)
where the last isomorphism sends x to y λ . Since Fp [[hyi]] is the power series ring
Fp [[t]] with y = 1 + t we deduce that
b F [[H]] Fp ∼
A⊗
= Fp [[t]]/(fλ )
p
where
fλ = (1 + t)λ + (1 + t)−λ + (1 + t) + (1 + t)−1 − 4.
We claim that as λ varies, the first non-zero coefficient of fλ appears in degree less
than p, in other words,
(6.1)
sup inf{i | fλ 6∈ ti+1 Fp [[t]]} < p.
λ
With this bound in hand, writing f ∈ zti + ti+1 Fp [[t]] with z ∈ Fp \ {0}, we have
(f ) = (ti ) and
b Fp [[H]] Fp = dimFp Fp [[t]]/(ti ) = i < p.
dimFp A⊗
12
MARTIN R. BRIDSON AND DESSISLAVA H. KOCHLOUKOVA
The rest of the proof is devoted to establishing the bound in (6.1). Given λ ∈ Zp ,
we write
λ = z0 + pλ1 and − λ = a0 + pλ2
where z0 , a0 ∈ {0, 1, . . . , p− 1} and λ1 , λ2 ∈ Zp . As (1 + t)pλj = (1 + tp )λj in Fp [[tp ]],
we have
(1 + t)λ = (1 + t)z0 (1 + t)pλ1 ∈ (1 + t)z0 + tp Fp [[t]]
and
(1 + t)−λ = (1 + t)a0 (1 + t)pλ2 ∈ (1 + t)a0 + tp Fp [[t]].
Note that if z0 > 0 then a0 + z0 = p. In particular, taking λ = z0 = 1 we have
(1 + t)−1 ∈ (1 + t)p−1 + tp Fp [[t]].
In general,
(6.2)
fλ ∈ gλ + tp Fp [[t]],
where gλ = (1 + t)z0 + (1 + t)a0 + (1 + t) + (1 + t)p−1 − 4. To establish (6.1)
we must argue that gλ is a non-zero polynomial. If z0 = 0 then a0 = 0 and
gλ = t + (1 + t)p−1 − 1, which is a non-zero element of Fp [t] because p > 2. If z0 > 0
then
gλ = (1 + t)z0 + (1 + t)p−z0 + (1 + t) + (1 + t)p−1 − 4,
which is again non-zero. This completes the proof.
By calculating with binomial coefficients, one can replace (6.1) by a bound that
is independent of p > 2.
References
[1] Corob Cook, Ged; Bieri-Eckmann Criteria for Profinite Groups, to appear in Israel Journal
of Mathematcs, arXiv 1412.1703v2
[2] Bourbaki, Nicolas; Commutative algebra. Chapters 1 - 7. Elements of Mathematics (Berlin).
Springer-Verlag, Berlin, 1998.
[3] Bieri, Robert; Neumann, Walter D.; Strebel, Ralph; A geometric invariant of discrete groups.
Invent. Math. 90 (1987), no. 3, 451–477.
[4] Bridson M.R.; Kochloukova D.; The virtual first Betti number of soluble groups, Pacific J.
Math., 274 (2015), 497–510.
[5] Eisenbud, D.; Commutative algebra: With a view toward algebraic geometry. Graduate Texts
in Mathematics, 150. Springer-Verlag, New York, 1995.
[6] Jaikin Zapirain, A. Unpublished note
[7] King, Jeremy D.; A geometric invariant for metabelian pro-p groups. J. London Math. Soc.
(2) 60 (1999), no. 1, 83–94.
[8] King, Jeremy D.; Homological finiteness conditions for pro-p groups. Comm. Algebra 27
(1999), no. 10, 4969–4991.
[9] King, Jeremy D.; Embedding theorems for pro-p groups. Math. Proc. Cambridge Philos. Soc.
123 (1998), no. 2, 217–226.
[10] Lazard, M.; Groupes analytiques p-adiques, Publ. Math. IHES, 26 (1965), 389–603.
[11] Ribes, L.; Zalesskii, P. Profinite groups. Second edition. 40. Springer-Verlag, Berlin, 2010
[12] Wilson, John S. Finite presentations of pro-p groups and discrete groups. Invent. Math. 105
(1991), no. 1, 177-183
[13] Wilson, John S. Profinite groups. London Mathematical Society Monographs. New Series,
19. The Clarendon Press, Oxford University Press, New York, 1998
[14] Dixon, J. D.; du Sautoy, M. P. F.; Mann, A.; Segal, D. Analytic pro-p-groups. London
Mathematical Society Lecture Note Series, 157. Cambridge University Press, Cambridge,
2003
THE TORSION-FREE RANK OF HOMOLOGY IN TOWERS OF SOLUBLE PRO-p GROUPS13
Mathematical Institute, Andrews Wiles Building, University of Oxford, Oxford
OX2 6GG, UK
E-mail address: [email protected]
Department of Mathematics, State University of Campinas (UNICAMP), 13083-859,
Campinas, SP, Brazil
E-mail address: [email protected]
| 4math.GR
|
Proceedings of Machine Learning Research vol 65:1–37, 2017
Algorithmic Chaining and the Role of Partial Feedback
in Online Nonparametric Learning
Nicolò Cesa-Bianchi
NICOLO . CESA - BIANCHI @ UNIMI . IT
arXiv:1702.08211v2 [stat.ML] 30 Jun 2017
Università degli Studi di Milano, Milano, Italy
Pierre Gaillard
PIERRE . GAILLARD @ INRIA . FR
INRIA - Sierra Project-team, Département d’Informatique de l’École Normale Supérieure, Paris, France
Claudio Gentile
CLAUDIO . GENTILE @ UNINSUBRIA . IT
Università degli Studi dell’Insubria, Varese, Italy
Sébastien Gerchinovitz
SEBASTIEN . GERCHINOVITZ @ MATH . UNIV- TOULOUSE . FR
Université Toulouse III - Paul Sabatier, Toulouse, France
Abstract
We investigate contextual online learning with nonparametric (Lipschitz) comparison classes
under different assumptions on losses and feedback information. For full information feedback and
Lipschitz losses, we design the first explicit algorithm achieving the minimax regret rate (up to log
factors). In a partial feedback model motivated by second-price auctions, we obtain algorithms
for Lipschitz and semi-Lipschitz losses with regret bounds improving on the known bounds for
standard bandit feedback. Our analysis combines novel results for contextual second-price auctions
with a novel algorithmic approach based on chaining. When the context space is Euclidean, our
chaining approach is efficient and delivers an even better regret bound.
Keywords: online learning, nonparametric, chaining, bandits.
1. Introduction
In online learning (Cesa-Bianchi and Lugosi, 2006; Shalev-Shwartz, 2011; Hazan, 2015) an agent
(or learner) interacts with an unknown and arbitrary environment in a sequence of rounds. At each
round, the learner chooses an action from a given action space and incurs the loss associated with the
chosen action. The loss functions, which are different in each round, are fixed by the environment
at the beginning of the interaction. After choosing an action, the learner observes some feedback,
which can be used to reduce his loss in subsequent rounds. A variety of different feedback models
are discussed in the literature. The most common feedback model is full information, also known as
prediction with expert advice, where the learner gets access to the entire loss function at the end of
each round. Another common feedback model is bandit information, where the learner just observes
the loss assigned to the action chosen in the current round. Feedback models in between full and
bandit information are also possible, and can be used to describe many interesting online learning
applications —see e.g., (Alon et al., 2014, 2015). The performance of an online learner is measured
using a notion of regret, which is typically defined as the amount by which the learner’s cumulative
loss exceeds the cumulative loss of the best fixed action in hindsight.
Accepted
.
for presentation at Conference on Learning Theory (COLT) 2017.
c 2017 N. Cesa-Bianchi, P. Gaillard, C. Gentile & S. Gerchinovitz.
C ESA -B IANCHI G AILLARD G ENTILE G ERCHINOVITZ
Online contextual learning is a generalization of online learning where the loss functions generated by the environment are paired with contexts from a given context space. On each round,
before choosing an action, the learner observes the current context. In the presence of contextual
information, the learner’s regret is no longer defined against the best action in hindsight, but rather
against the best policy (i.e., mapping from the context space to the action space) in a given reference class of policies. In agreement with the online learning framework, online contextual learning
is nonstochastic. Namely, regret bounds must hold for arbitrary sequences of contexts and losses.
In order to capture complex environments, the reference class of policies should be as large as
possible. In this work, we focus on nonparametric classes of policies, such as classes containing
policies that are Lipschitz with respect to metrics defined on the context and action spaces. The
best possible (minimax) growth rate of the regret, as a function of the number T of rounds, is
then determined by the interplay among the richness of the policy class, the constraints on the loss
functions (e.g., Lipschitz, convex, etc.), and the type of feedback information (full, bandit, or in
between). Whereas most of the previous works study online nonparametric learning with convex
losses, in this paper we investigate nonparametric regret rates for general Lipschitz losses (in fact,
some of our results apply to an even larger class of loss functions).
In the full information setting, a very general yet simple algorithmic approach to online nonparametric learning with convex and Lipschitz losses was introduced by Hazan
and Megiddo (2007). For
e T (d+1)/(d+2) upper bound1 on the regret
any reference class of Lipschitz policies, they proved a O
e notation hides logarithmic factors in T . In
for any context space of metric dimension d, where
the O
(d−1)/d
the same work, they also proved a Ω T
lower bound. The gap between the upper and lower
bound was closed by Rakhlin et al. (2015) for arbitrary Lipschitz (not necessarily convex) losses,
showing that T (d−1)/d is indeed the minimax rate for full information. Yet, since their approach is
nonconstructive, they did not give an explicit algorithm achieving this bound.
As noted elsewhere —see, e.g., (Slivkins, 2014)—
the approach of Hazan and Megiddo (2007)
e T (d+p+1)/(d+p+2) upper bound on the regret against any class
can be also adapted to prove a O
of Lipschitz policies in the bandit information setting with Lipschitz
losses, where p is the metric
dimension of the action space. The lower bound Ω T (p+1)/(p+2) proven for d = 0 (Bubeck et al.,
2011a; Kleinberg et al., 2008) rules out the possibility of improving the dependence on p in the
upper bound.
Our contributions. In the full information
model, we show the first explicit algorithm achieving
e T (d−1)/d for Lipschitz policies and Lipschitz losses (excluding logathe minimax regret rate O
rithmic factors in T and polynomial factors in the metric dimension of the action space). When
the context space is [0, 1]d , our algorithm can be implemented efficiently (i.e., with a running time
polynomial in T ).
Motivated by a problem in online advertising where the action space is the [0, 1] interval, we
also study a “one-sided” full information model in which the loss of each action greater than or
equal to the chosen action is available to the learner after each round. For this feedback model,
which lies between full and bandit information,
we prove a regret bound for Lipschitz policies and
e T d/(d+1) , which is larger than the minimax regret for full information
Lipschitz losses of order O
but smaller than the upper bound for bandit information when p = 1. For the special case when the
1. This bound has a polynomial dependence on the metric dimension of the action space, which is absorbed by the
asymptotic notation.
2
A LGORITHMIC C HAINING AND THE ROLE OF PARTIAL F EEDBACK IN O NLINE N ONPARAMETRIC L EARNING
context space is [0, 1]d , we use a specialized approach offering the double advantage of an improved
e T (d−1/3)/(d+2/3) regret bound which is also attained by a time-efficient algorithm.
O
We then study a concrete application for minimizing the seller’s regret in contextual secondprice auctions with reserve price, a setting where the loss function is not Lipschitz but only semiLipschitz. When the feedback after each auction is the
seller’s revenue together with the highest bid
e T (d+1)/(d+2) regret bound against Lipschitz policies (in this
for the current auction, we prove a O
setting,
√ a policy maps contexts to reserve prices for the seller). As a by-product, we show the first
e
O T regret bound on the seller’s revenue in context-free second-price auctions under the same
feedback model as above. Table 1 summarizes our results.
Feedback model
Bandit
One-sided full information
Loss functions
Lipschitz
Convex
T
d+1
T d+2
Semi-Lipschitz
T d+2
Lipschitz
Full information
Upper bound
d+2
d+3
Lipschitz
d+1
T
T
d−1/3
d+2/3
d−1
d
(Theorem 1)
(Corollary 2)
(Theorem 3)
(Theorem 6)
(Theorem 7)
Table 1: Some regret bounds obtained in this paper. The rates are up to logarithmic factors for
Lipschitz policies f : [0, 1]d → [0, 1] with d > 2. All upper bounds are constructive
(i.e., achieved by explict algorithms). The only matching lower bound is the one for full
information feedback due to Hazan and Megiddo (2007).
In order to prove our results, we approximate the action space using a finite covering (finite
coverability is a necessary condition for our results to hold). This allows us to use the many existing
algorithms for experts (full information feedback) and bandits when the action space is finite, such
as Hedge (Freund and Schapire, 1997) and Exp3/Exp4 (Auer et al., 2002). The simplest of our
algorithms, adapted from Hazan and Megiddo (2007), incrementally covers the context space with
balls of fixed radius. Each ball hosts an instance of an online learning algorithm which predicts in
all rounds when the context falls into the ball. New balls are adaptively created when new contexts
are observed which fall outside the existing balls (see Algorithm 1 for an example). We use this
simple construction to prove the regret bound for contextual second-price auctions, a setting where
losses are not Lipschitz. In order to exploit the additional structure provided by Lipschitz losses,
we resort to more sophisticated constructions based on chaining (Dudley, 1967). In particular,
inspired by previous works in this area (especially the work of Gaillard and Gerchinovitz, 2015),
we design a chaining-inspired algorithm applied to a hierarchical covering of the policy space.
Despite we are not the first ones to use chaining algorithmically in online learning, our idea of
constructing a hierarchy of online learners, where each node uses its children as experts, is novel in
this context as far as we know. Finally, the time-efficient algorithm achieving the improved regret
bound is derived from a different (and more involved) chaining algorithm based on wavelet-like
approximation techniques.
Setting and main definitions. We assume the context space X is a metric space (X , ρX ) of finite
metric dimension d and the action space Y is a metric space (Y, ρY ) of finite metric dimension p.
3
C ESA -B IANCHI G AILLARD G ENTILE G ERCHINOVITZ
Hence, there exist CX , CY > 0 such that, for all 0 < ε 6 1, X and Y can be covered, respectively,
with at most CX ε−d and at most CY ε−p balls of radius ε. For any 0 < ε 6 1, we use Yε to denote
any ε-covering of Y of size Kε 6 CY ε−p . Finally, we assume that Y has diameter bounded by 1
with respect to metric ρY .
We consider the following online learning protocol with oblivious adversary and loss functions
ℓt : Y → [0, 1]. Given an unknown sequence (x1 , ℓ1 ), (x2 , ℓ2 ), . . . of contexts xt ∈ X and loss
functions ℓt : Y → [0, 1], for every round t = 1, 2, . . . :
1. The environment reveals context xt ∈ X ;
2. The learner selects an action ybt ∈ Y and incurs loss ℓt ybt ;
3. The learner obtains feedback from the environment.
Loss functions ℓt satisfy the 1-Lipschitz2 condition ℓt (y) − ℓt (y ′ ) 6 ρY (y, y ′ ) for all y, y ′ ∈ Y.
However, we occasionally consider losses satisfying a weaker semi-Lipschitz condition.
We study three different types of feedback: bandit feedback (the learner only observes the loss
ℓt (b
yt ) of the selected action ybt ), full information feedback (the learner can compute ℓt (y) for any
y ∈ Y), and one-sided full information feedback (Y ≡ [0, 1], and the learner can compute ℓt (y) if
and only if y > ybt ). Given a reference class F ⊆ Y X of policies, the learner’s goal is to minimize
the regret against the best policy in the class,
" T
#
T
X
X
ℓt ybt − inf
RegT (F) , E
ℓt f (xt ) ,
f ∈F
t=1
t=1
where the expectation is with respect to the learner’s internal randomization. We derive regret
bounds for the competitor class F made up of all bounded functions f : X → Y that are 1Lipschitz3 w.r.t. ρX and ρY . Namely, ρY f (x), f (x′ ) 6 ρX (x, x′ ) for all f ∈ F and all x, x′ ∈ X .
We occasionally use the dot product notation pt · ℓt to indicate the expectation of ℓt according to
law pt . Finally, the set of all probability distributions over a finite set of K elements is denoted by
∆(K).
Organization of the paper. The rest of the paper is organized as follows. In Section 2, we give
an overview of the related literature. Then, starting from the subsequent section, our results are
presented in the order dictated by the amount of feedback available to the learner, from bandit
feedback (Section 3) to one-sided full information feedback (Section 4) to full information feedback
(Section 5).
2. Related Work
Contextual online learning generalizes online convex optimization (Hazan, 2015) to nonconvex
losses, nonparametric policies, and partial feedback. Papers about nonparametric online learning
in full information include (Vovk, 2007; Gaillard and Gerchinovitz, 2015) for the square loss, and
(Hazan and Megiddo, 2007; Rakhlin and Sridharan, 2015) for general convex losses. In the bandit
2. Assuming a unit Lipschitz constant is without loss of generality because our algorithms are oblivious to it.
3. Assuming a unit Lipschitz constant Lf for the functions in f ∈ F is also without loss of generality because our
algorithms are oblivious to it. The only exception is algorithm HierExp4⋆ of Section 4.3, which is only guaranteed to
work with unit Lipschitz constants. However, a similar regret bound can be achieved for arbitrary (yet known) values
of Lf via a simple modification—see also (Bubeck et al., 2011b) for regret bounds optimized for Lf in a stochastic
and non-contextual bandit setting when Lf is unknown.
4
A LGORITHMIC C HAINING AND THE ROLE OF PARTIAL F EEDBACK IN O NLINE N ONPARAMETRIC L EARNING
feedback model, earlier work on context-free bandits on metric spaces includes (Kleinberg, 2004;
Kleinberg et al., 2008). The paper (Auer et al., 2002) introduces the Exp4 algorithm for nonstochastic contextual bandits when both the action space and the policy space are finite, and policies are
maps from contexts to distributions over actions. Moreover, rather than observing the current context, the learner sees the output of each policy for that context. In the contextual bandit model of
Maillard and Munos (2011), context space and action space are finite, and the learner observes the
current context while competing against the best policy among all functions mapping contexts to
actions. Finally, a nonparametric bandit setting related to ours was studied by Slivkins (2014). We
refer the reader to the discussion after Theorem 1 for connections with our results.
Chaining (Dudley, 1967) is a powerful technique to obtain tail bounds on the suprema of
stochastic processes. In nonparametric online learning with full information feedback, chaining was
used constructively by Cesa-Bianchi and Lugosi (1999) to design an algorithm for linear losses, and
nonconstructively by Rakhlin et al. (2015) to derive minimax rates for Lipschitz losses. Other notable examples of chaining are the stochastic bandit algorithms of Contal et al. (2015) and Contal and Vayatis
(2016). The constructive algorithmic chaining technique developed in this work is inspired by the
nonparametric analysis of the full information setting of Gaillard and Gerchinovitz (2015). However, their multi-variable EG algorithm heavily relies on convexity of losses and requires access to
loss gradients. In order to cope with nonconvex losses and lack of gradient information, we develop
a novel chaining approach based on a tree of hierarchical coverings of the policy class, where each
internal tree node hosts a bandit algorithm.
In our nonstochastic online setting, chaining yields improved rates when the regret is decomposed into a sum of local regrets, each one scaling with the range of the local losses. However,
deriving regret bounds that scale with the effective range of the losses is not always possible, as
shown by Gerchinovitz and Lattimore (2016) in the nonstochastic K-armed bandit setting. This result suggests that chaining might not be useful in online nonparametric learning when the feedback
is bandit. However, as we show in this paper, algorithmic chaining does help improving the regret
when the feedback is one-sided full information or full information. In full information, chainingbased algorithms deliver regret bounds that match (up to log factors) the nonconstructive bounds
of (Rakhlin et al., 2015).
In a different but interesting research thread on contextual bandits, the learner is confronted with
the best within a finite (but large) class of policies over finitely-many actions, and is assumed to have
access to this policy class through an optimization oracle for the offline full information problem.
Relevant references include (Agarwal et al., 2014; Rakhlin and Sridharan, 2016; Syrgkanis et al.,
2016). The main concern is to devise (oracle-based) algorithms with small regret and requiring as
few calls to the optimization oracle as possible.
3. Warmup: Nonparametric Bandits
As a simple warmup exercise, we prove a known result —see e.g., (Slivkins, 2014). Namely, a
regret bound for contextual bandits with Lipschitz policies and Lipschitz losses. ContextualExp3
(Algorithm 1) is a bandit version of the algorithm by Hazan and Megiddo (2007) and maintains
a set of balls of fixed radius ε in the context space, where each ball hosts an instance of the Exp3
5
C ESA -B IANCHI G AILLARD G ENTILE G ERCHINOVITZ
algorithm of Auer et al. (2002).4 At each round t, if a new incoming context xt ∈ X is not contained
in any existing ball, then a new ball centered at xt is created, and a fresh instance of Exp3 is allocated
to handle xt . Otherwise, the Exp3 instance associated with the closest context so far w.r.t. ρX is
used to handle xt . Each allocated Exp3 instance operates on the discretized action space Yε whose
size Kε is at most CY ε−p . For completeness, the proof of the following theorem is provided in
Algorithm 1: ContextualExp3 (for bandit feedback)
Input: Ball radius ε > 0, ε-covering Yε of Y such that |Yε | ≤ CY ε−p .
for t = 1, 2, . . . do
1. Get context xt ∈ X ;
2. If xt does not belong to any existing ball, then create a new ball of radius ε centered on xt ,
and allocate a fresh instance of Exp3;
3. Let the active Exp3 instance be the instance allocated to the existing ball whose center xs is
closest to xt ;
4. Draw an action
ybt using the active Exp3 instance;
5. Get ℓt ybt and use it to update the active Exp3 instance.
end
Appendix B.
Theorem 1 Fix any any sequence (x1 , ℓ1 ), (x2 , ℓ2 ), . . . of contexts xt ∈ X and 1-Lipschitz loss
functions ℓt : Y → [0, 1]. If ContextualExp3 is run in the bandit feedback model with parameter5
2
1
ε = (ln T ) p+d+2 T − p+d+2 , then its regret RegT (F) with respect to the set F of 1-Lipschitz functions
p+d+1
e T p+d+2 , where the O
e notation hides factors polynomial in
f : X → Y satisfies RegT (F) = O
CX and CY , and ln T factors.
A lower bound matching up to log factors the upper bound of Theorem 1 is contained in (Slivkins,
2014) —see also (Lu et al., 2010) for earlier results in the same setting. However, our setting and
his are subtly different: the adversary of Slivkins (2014) uses more general Lipschitz losses which,
translated into our context, imply that the Lipschitz assumption is required to hold only for the
composite function ℓt (f (·)), rather than the two functions ℓt and f separately. Hence, being the
adversary less constrained (and the comparison class wider), the lower bound contained in (Slivkins,
2014) does not seem to apply to our setting.
While we are unaware of a lower bound matching the upper bound in Theorem 1 when F
is the class of (global) Lipschitz
functions and d > 1, in the noncontextual case (d = 0), the
lower bound Ω T (p+1)/(p+2) proven by Bubeck et al. (2011a); Kleinberg et al. (2008) shows that
improvements on the dependence on p are generally impossible. Yet, the dependence on p in the
bound of Theorem 1 can be greatly improved in the special case when the Lipschitz losses are also
convex. Assume Y is a convex and compact subset of Rp . Then we use the same approach as in
Theorem 1, where the Exp3 algorithm hosted at each ball is replaced by an instance of the algorithm
4. Instead of Exp3 we could use INF (Audibert and Bubeck, 2010), which enjoys a minimax optimal regret bound up
to constant factors. This would avoid a polylog factor in T in the bound. Since we do not optimize for polylog factors
anyway, we opted for the better known algorithm.
5. Here and throughout, T is assumed to be large enough so as to ensure ε 6 1.
6
A LGORITHMIC C HAINING AND THE ROLE OF PARTIAL F EEDBACK IN O NLINE N ONPARAMETRIC L EARNING
action space Y. The regret of the algorithm that
by Bubeck et al. (2016), run on the non-discretized
√
replaces Exp3 is bounded by poly(p, ln T ) T . This immediately gives the following corollary.
Corollary 2 Fix any any sequence (x1 , ℓ1 ), (x2 , ℓ2 ), . . . of contexts xt ∈ X and convex loss functions ℓt : Y → [0, 1], where Y is a convex and compact subset of Rp . Then there exists an algorithm
for the bandit feedback model whose regret with respect to the set F of 1-Lipschitz functions satisfies
RegT (F) 6 poly(p, ln T )T (d+1)/(d+2) , where poly is a polynomial function of its arguments.
4. One-Sided Full Information Feedback
In this section we show that better nonparametric rates can be achieved in the one-sided full information setting, where the feedback is larger than the standard bandit feedback but smaller than
the full information feedback. More precisely, we consider the same setting as in Section 3 in the
special case when the action space Y is [0, 1]. We further assume that, after each play ybt ∈ Y, the
learner can compute the loss ℓt (y) of any number of actions y > ybt . This in contrast to observing
only ℓt ybt , as in the standard bandit setting. We start with an important special case: maximizing the seller’s revenue in a sequence of repeated second-price auctions. In Section 4.2, we use
the chaining technique to design a general algorithm for arbitrary Lipschitz losses in the one-sided
full information model. An efficient variant of this algorithm is obtained using a more involved
construction in Section 4.3.
4.1. Nonparametric second-price auctions
In online advertising, publishers sell their online ad space to advertisers through second-price auctions managed by ad exchanges. For each impression (ad display) created on the publisher’s website, the ad exchange runs an auction on the fly. Empirical evidence (Ostrovsky and Schwarz, 2011)
shows that an informed choice of the seller’s reserve price, disqualifying any bid below it, can indeed have a significant impact on the revenue of the seller. Regret minimization in second-price
auctions was studied by Cesa-Bianchi et al. (2015) in a non-contextual setting. They showed that,
when buyers draw their bids i.i.d. from the same unknown distribution on [0, 1], there exists
√ an efe T with
ficient strategy for setting reserve prices such that the seller’s regret is bounded by O
high probability with respect to the bid distribution. Here we extend those results to a nonstochastic
and nonparametric contextual setting with nonstochastic bids, and prove a regret bound of order
T (d+1)/(d+2) where d is the context space dimension. This improves on the bound T (d+2)/(d+3)
of Theorem
1 when p = 1. As a byproduct, in Theorem 3 (Appendix C) we prove the first
√
e T regret bound for the seller in nonstochastic and noncontextual second-price auctions. UnO
like (Cesa-Bianchi et al., 2015), where the feedback after each auction was “strictly bandit” (i.e.,
just the seller’s revenue), here we assume the seller is also observing the highest bid together with
the revenue. This richer feedback, which is key to proving our results, is made available by some
ad exchanges such as AppNexus.
The seller’s revenue in a second-price auction is computed as follows: if the reserve price yb
is not larger than the second-highest bid b(2), then the item is sold to the highest bidder and the
seller’s revenue is b(2). If yb is between b(2) and the highest bid b(1), then the item is sold to the
highest bidder but the seller’s revenue is the reserve price. Finally, if yb is bigger than b(1), thenthe
item is not sold and the seller’s revenue is zero. Formally, the seller’s revenue is g yb, b(1), b(2) =
7
C ESA -B IANCHI G AILLARD G ENTILE G ERCHINOVITZ
max yb, b(2) Iyb≤b(1) . Note that the revenue only depends on the reserve price yb and on the two
highest bids b(1) ≥ b(2), which —by assumption— belong all to the unit interval [0, 1].
In the online contextual version of the problem, unknown sequences of contexts x1 , x2 , . . . ∈
X and bids are fixed beforehand (in the case of online advertising, the context could be public
information about the targeted customers). At the beginning of each auction t = 1, 2, . . . , the seller
observes context xt and computes a reserve price ybt ∈ [0, 1]. Then, bids bt (1), bt (2) are collected
by the auctioneer, and the seller (which is not the same as the auctioneer) observes his revenue
yt ) and bt (1)
gt ybt = g ybt , bt (1), bt (2) , together with the highest bid bt (1). Crucially, knowing gt (b
allows to compute gt (y) for all y > ybt . For technical reasons, we use losses ℓt ybt = 1 − gt ybt
instead of revenues, see Figure 1 for a pictorial representation.
1
0.8
0.6
1 − bt (2)
1 − bt (1)
0.4
0.2
0
0
0.2
0.4
0.6
bt (2)
0.8
1
bt (1)
Figure 1: The loss function ℓt (b
yt ) = 1−max{b
yt , bt (2)}Iybt ≤bt (1) when bt (1) = 0.7 and bt (2) = 0.5.
Remarkably, the loss functions ℓt are not Lipschitz and not even continuous, and so this problem
falls outside the scope of standard results for contextual bandits. Instead, the losses only satisfy the
semi-Lipschitz property ℓt (y + δ) ≥ ℓt (y) − δ for all 0 6 y 6 y + δ 6 1. We now state a bound
on the regret RegT (F) with respect to any class F of Lipschitz functions f : X → [0, 1]. The
algorithm that achieves this bound is ContextualRTB (where RTB stands for Real Time Bidding
—see Algorithm 3 in Appendix C), a variant of ContextualExp3 (Algorithm 1), where each ball
hosts an instance of Exp3-RTB, instead of Exp3. The proof is given in Appendix D.
Theorem 3 Fix any sequence of contexts xt ∈ X and1 bid pairs 0 6 bt (2) 6 bt (1) 6 1 for
t > 1. If ContextualRTB is run with parameter ε = T − d+2 and Exp3-RTB is tuned with parameter
γ = ε, then the regret
with respect to any class of 1-Lipschitz functions f : X → [0, 1] satisfies
d+1
e T d+2 , where d is the dimension of X and the O
e notation hides constants and ln T
RegT (F) = O
factors.
ContextualRTB and ContextualExp3 of Section 3 can be modified so to avoid knowing the horizon
T and so that the dimension d of the context space is replaced in the bound by the (unknown, and
possibly much smaller) dimension of the set of contexts actually occurring in the sequence chosen
by the adversary. This modification involves using a time-varying radius ε and a doubling trick to
check when the current guess for the dimension is violated by the current number of balls. The omitted proof of this statement goes along the lines of the proof in (De Rosa et al., 2015, Theorem 1).
8
A LGORITHMIC C HAINING AND THE ROLE OF PARTIAL F EEDBACK IN O NLINE N ONPARAMETRIC L EARNING
4.2. Chaining the bandits
We now show that whenever the richer feedback structure —i.e., the learner can compute the loss
ℓt (y) of any number of actions y > ybt — is combined with Lipschitz losses (rather than just semiLipschitz), then an improved regret bound T d/(d+1) can be derived. The key technical idea enabling
this improvement is the application of the chaining technique to a hierarchical covering of the policy
space (as opposed to the flat covering of the context space used in both Section 3 and Section 4.1).
We start with a computationally inefficient algorithm that works for arbitrary policy classes F (not
only Lipschitz) and is easier to understand. In Section 4.3 we derive an efficient variant for F that
are Lipschitz. In this case we obtain even better regret bounds via a penalization trick.
A way of understanding the chaining approach is to view the hierarchical covering of the policy
class F as a tree whose nodes are functions in F, and where the nodes at each depth m define a
(2−m )-covering of F. The tree represents any function f ∗ ∈ F (e.g., the function with the smallest
cumulative loss) by a unique path (or chain) f0 → f1 → · · · → fM → f ∗ , where f0 is the root and
fM is the function best approximating f ∗ in the cover at the largest available depth M . By relying
on this representation, we control the regret against any function in F by running an instance of an
online bandit algorithm A on each node of the tree. The instance Af at node f uses the predictions
of the instances running on the nodes that are children of f as expert advice. The action drawn
by instance A0 running on the root node is the output of the tree. For any given sequence of pairs
(xt , ℓt ) of contexts and losses, the regret against f ∗ with path f0 → f1 → · · · → fM → f ∗ can
then be written (ignoring some constants) as
#
" T
−1
T
M
X
X
X
+ 2−M T
ℓt Am (xt ) − ℓt Am+1 (xt )
E
E ℓt A0 (xt ) − ℓt f ∗ (xt ) 6
t=1
m=0
t=1
where Am is the instance running on node fm for m = 0, . . . , M − 1 and AM ≡ fM . The last
term 2−M T accounts for the cost of approximating f ∗ with the closest function fM in a (2−M )cover of F under suitable Lipschitz assumptions. The outer sum in the right-hand side of the above
display can be viewed as a sum of M regrets, where the m-th term in the sum is the regret of Am
against the instances running on the children of the node hosting Am . Since we face an expert
learning problem in a partial information setting, the Exp4 algorithm of Auer et al. (2002) is a
natural choice for the learner A. However, a first issue to consider is that we are using A0 to draw
actions in the bandit problem, and so the other Exp4 instances receive loss estimates that are based
on the distribution used by A0 rather than being based on their own distributions. A second issue
is that our regret decomposition crucially relies on the fact that each instance Am only competes
(in the sense of regret) against functions f at the leaves of the subtree rooted at the node where
Am runs. By construction, these functions at the leaves are roughly (2−m )-close to each other
and —by Lipschitzness— so are their losses. As a consequence, the regret of Am should scale
with the true loss range 2−m . Via an appropriate modification of the original Exp4 algorithm, we
manage to address both these issues. In particular, in order to make the regret dependent on the loss
range, we heavily rely on the one-sided full information model assumed in this section. Finally,
the hierarchical covering requires that losses be Lipschitz, rather than just semi-Lipschitz as in the
application of Subsection 4.1, which uses a simpler flat covering.
Fix any class F of functions f : X → [0, 1]. Let us introduce the sup norm
kf − gk∞ = sup f (x) − g(x) .
x∈X
9
(1)
C ESA -B IANCHI G AILLARD G ENTILE G ERCHINOVITZ
We denote by N∞ F, ε the cardinality of the smallest ε-cover
of F w.r.t. the sup norm. Throughout this section, our only assumption on F is that N∞ F, ε < +∞ for all ε > 0 (this
is
property
d
known as total boundedness). In Section 4.3 we will focus on the special case F = f : [0, 1] →
[0, 1] : f is 1-Lipschitz to derive an efficient version of our algorithm.
We now define a tree TF of depth M , whose nodes are labeled by functions in the class F,
so that functions corresponding to nodes with a close common ancestor are close to one another
according to the sup norm (1). For all m = 0, 1, . . . , M , let Fm be a (2−m )-covering of F in sup
norm with minimal cardinality Nm = N∞ (F, 2−m ). Since the diameter of (F, k·k∞ ) is bounded
by 1, we have N0 = 1 and F0 = {f0 } for some f0 ∈ F. For each m = 0, 1, . . . , M and for every
fv ∈ Fm we have a node v in TF at depth m. The parent of a node w at depth m + 1 is some node
v at depth m such that
v∈
arg min
v′ : depth(v′ )=m
kfv′ − fw k∞
(ties broken arbitrarily)
and we say that w is a child of v. Let L be the set of all the leaves of TF , Lv be the set of all the
leaves under v ∈ TF (i.e., the leaves of the subtree rooted at v), and Cv be the set of children of
v ∈ TF .
u
Level m
2−m covering of F
Level m + 1
2−(m+1) covering
Exp4
...
..
..
...
Exp4
.
v
.
.
Exp4
..
..
.
...
...
w
Level M (=leaves L)
2−M covering
Leaves under v = Lv
Figure 2: Hierarchical covering of the function space (used in Algorithm 2).
Our new bandit algorithm HierExp4 (Algorithm 2 below) is a hierarchical composition of
instances of Exp4 on the tree TF constructed above (see Figure 2). Let K = 2M and K =
{y1 , . . . , yK }, where yk = 2−M (k − 1) for k = 1, . . . , 2M , be our discretization of the action
space Y = [0, 1]. At every round t, after observing context xt ∈ X , each leaf v ∈ L recommends
the best approximation of fv (xt ) in K, it (v) ∈ arg mini=1,...,K yi − fv (xt ) . Therefore, the leaves
v ∈ L correspond to deterministic strategies t 7→ it (v), and we will find it convenient to view a set
of leaves L as the set of actions played by those leaves at time t. Each internal node v ∈ TF \ L runs
an instance of Exp4 using the children of v as experts. More precisely, we use a variant of Exp4
(see Algorithm 4 in Appendix E) which adapts to the effective range of the losses. Let Exp4v be
the instance of the Exp4 variant run on node v. At each time t, this instance updates a distribution
qt (v, ·) ∈ ∆(|C
Pv |) over experts in Cv and a distribution pt (v, ·) ∈ ∆(K) over actions in K defined
by pt (v, i) = w∈Cv qt (v, w)pt (w, i).
Let v0 be the root of TF . The prediction of HierExp4 at time t is ybt = yIt ∈ K, where It is
drawn according to a mixture of pt (v0 , ·) and a unit mass on the minimal action y1 ∈ K.
For each v ∈ TF \ L, let Kt (v) = {i : (∃w ∈ Cv ) pt (w, i) > 0} and jt (v) = max Kt (v).
Note that ℓbt (v, i) in (2) has to be explicitly computed only for those actions i such that i > It and
10
A LGORITHMIC C HAINING AND THE ROLE OF PARTIAL F EEDBACK IN O NLINE N ONPARAMETRIC L EARNING
i ∈ Kt (v). This is because ℓbt (v, i) is needed for the computation of ℓet (v, w) only when pt (w, i) > 0.
Therefore, whenever
ℓbt (v, i) has to be computed for some i, then It 6 i 6 max Kt (v) = jt (v), so
that ℓt yjt (v) is observed and ℓbt (v, i) is well defined.
Algorithm 2: HierExp4 (for one-sided full information feedback)
Input
: Tree TF with root v0 and leaves L, exploration parameter γ ∈ (0, 1), learning rate
sequences η1 (v) > η2 (v) > · · · > 0 for v ∈ TF \ L.
Initialization: Set q1 (v, ·) to the uniform distribution in ∆(|Cv |) for every v ∈ TF \ L.
for t = 1, 2, . . . do
1. Get context xt ∈ X ;
2. Set pt (v, i) = Ii=it (v) for all i ∈ K and for all v ∈ L;
3. Set pt (v, i) = qt (v, ·) · pt (·, i) for all i ∈ K and for all v ∈ TF \ L;
4. Draw It ∼ p∗t and play ybt = yIt , where p∗t (i) = (1 − γ)pt (v0 , i) + γIi=1 for all i ∈ K;
5. Observe ℓt (y) for all y > yIt ;
6. For every v ∈ TF \ L and for every i ∈ Kt (v) compute
ℓt (yi ) − ℓt yjt (v)
ℓbt (v, i) =
IIt 6i ,
Pi
∗
k=1 pt (k)
end
where Kt (v) = {i : (∃w ∈ Cv ) pt (w, i) > 0} and jt (v) = max Kt (v).
7. For each v ∈ TF \ L and for each w ∈ Cv compute the expert loss
ℓet (v, w) = pt (w, ·) · ℓbt (v, ·) and perform the update
P
exp −ηt+1 (v) ts=1 ℓes (v, w)
qt+1 (v, w) = P
Pt e
′)
exp
−η
(v)
ℓ
(v,
w
′
t+1
s=1 s
w ∈Cv
Next, we show that the regret of HierExp4 is at most of the order of T d/(d+1) , which improves
on the rate T (d+1)/(d+2) obtained in Section 4.1 without using chaining. The required proofs are
contained in Appendix F.
Theorem 4 Fix any class F of functions f : X → [0, 1] and any sequence (x1 , ℓ1 ), (x2 , ℓ2 ), . . .
of contexts xt ∈ X and 1-Lipschitz loss functions ℓt : [0, 1] → [0, 1]. Assume the HierExp4 (Algorithm 2) is run with one-sided full information feedback using tree TF of depth M = ⌊ln2 (1/γ)⌋.
Moreover, the learning rate ηt (v) used at each node v at depth m = 0, . . . , M − 1 is given by (11)
in Appendix E, where Vet−1 (v) is the cumulative variance of ℓes (v, ·) according to qs (v, ·) up to time
s = t − 1. Then for all T > 1 the regret satisfies
!
Z 1/2 s
1
T
7
RegT (F) 6 5γT + 2
ln N∞ (F, ε) + 1
dε .
ln N∞ (F, ε) +
γ
γ
γ/2
In particular, if X ≡ [0, 1]d is endowed with the sup norm ρX (x, x′ ) = kx − x′ k∞ , then the set F
of all 1-Lipschitz functions from X to [0, 1] satisfies ln N∞ (F, ε) . ε−d . Theorem 4 thus entails
the following corollary.
11
(2)
(3)
C ESA -B IANCHI G AILLARD G ENTILE G ERCHINOVITZ
Corollary 5 Under the assumptions of Theorem 4, if F is
f : [0, 1]d → [0, 1], then the regret of HierExp4 satisfies
O T 2/3
RegT (F) =
O T 2/3 (ln T )2/3
O T d/(d+1)
the set of all 1-Lipschitz functions
if d = 1
if d = 2
if d > 3
where the last inequality is obtained by optimizing the choice of γ for the different choices of d.
4.3. Efficient chaining
Though very general, HierExp4 (Algorithm 2) may be very inefficient. For example, when F is the
d
set of all 1-Lipschitz
functions from [0, 1] to [0, 1], a direct implementation of HierExp4 would require exp poly(T ) weight updates at every round. In this section we tackle the special case when
F is the class of all 1-Lipschitz functions f : [0, 1]d → [0, 1] w.r.t. the sup norm on [0, 1]d (for simplicity). We construct an ad-hoc hierarchical covering of F and define a variant of HierExp4 whose
running time at every round is polynomial in T . We rely on a well-known wavelet-like approximation technique which was used earlier —see, e.g., (Gaillard and Gerchinovitz, 2015)— for online
nonparametric regression with full information feedback. However, we replace their multi-variable
Exponentiated Gradient algorithm, which requires convex losses and gradient information, with a
more involved chaining algorithm that still enjoys a polynomial running time. The definitions of
our covering tree TF⋆ and of our algorithm HierExp4⋆ , as well as the proof of the following regret
bound, can be found in Appendix G. The exact value of cT (depending at most logarithmically on
T ) is also provided there.
Theorem 6 Let F be the set of all 1-Lipschitz functions f : [0, 1]d → [0, 1] w.r.t. the sup norm
on [0, 1]d . Consider T > 3 and any sequence (x1 , ℓ1 ), . . . , (xT , ℓT ) of contexts xt ∈ [0, 1]d and
1-Lipschitz loss functions ℓt : [0, 1] → [0, 1]. Assume HierExp4⋆ (Algorithm 5 in Appendix G) is
run with one-sided full information feedback using tree TF⋆ of depth M = ⌈ln2 (1/γ)⌉, exploration
1
1
d
parameter γ = T −1/2 (ln
)−1 Id=1 + T −1/(d+2/3) Id>1 , learning rate ηm = cT 2m( 4 +1) γ 2 T − 4 ,
PT
4−2j η for m = 0, . . . , M − 1. Then the regret satisfies
and penalization αm = M
j
j=m+1 2
(
√
O T ln T
if d = 1,
d−1/3
RegT (F) =
O T d+2/3 (ln T )3/2 if d > 2.
Moreover, the running time at every round is O T a with a = (1 + ln2 3)/(d + 2/3).
The above result improves on Corollary 5 in two ways. First, as we said, the running time is now
polynomial in T , contrary to what could be obtained via a direct implementation of HierExp4.
Second, when d > 2, the regret bound is of order T (d−1/3)/(d+2/3) , √
improving on the rate T d/(d+1)
e T ) for nonparametric bandits
from Corollary 5. Remarkably, Theorem 6 also yields a regret of O(
with one-sided full information feedback in dimension d = 1. The improvement on the rates
compared to HierExp4 is possible because we use a variant of Exp4 with penalized loss estimates.
This allows for a careful hierarchical control of the variance terms inspired by our analysis of Exp3RTB in Appendix C.
Note that the time complexity decreases as the dimension d increases. Indeed, when d increases
the regret gets worse but, at the same time, the size of the discretized action space and the number
of layers in our wavelet-like approximation can be both set to smaller values.
12
A LGORITHMIC C HAINING AND THE ROLE OF PARTIAL F EEDBACK IN O NLINE N ONPARAMETRIC L EARNING
5. A Tight Bound for Full Information through an Explicit Algorithm
In this section we apply the machinery developed in Section 4.2 to the full information setting,
where after each round t the learner can compute the loss ℓt (y) of any number of actions y ∈ Y.
We obtain the first explicit algorithm achieving, up to logarithmic factors, the minimax regret rate
T (d−1)/d for all classes of Lipschitz functions, where d is the dimension of the context space. This
achieves the same upper bound as the one proven by Rakhlin et al. (2015) in a nonconstructive
manner, and matches the lower bound of Hazan and Megiddo (2007). Our approach generalizes the
approach of Gaillard and Gerchinovitz (2015) to nonconvex Lipschitz losses. We consider a full
information variant of HierExp4 (Algorithm 2, Section 4.2), where —using the same notation as in
Section 4.2— the Exp4 instances running on the nodes of the tree TF are replaced by instances of
Hedge —e.g., (Bubeck and Cesa-Bianchi, 2012). Note that, due to the full information assumption,
the new algorithm, called HierHedge, observes losses at all leaves v ∈ L. As a consequence, no
exploration is needed and so we can set γ = 0. For the same reason, the estimated loss vectors
defined in (2) can be replaced with the true loss vectors, ℓt . See Algorithm 6 in Appendix H for a
definition of HierHedge. The same appendix also contains a proof of the next result.
Theorem 7 Fix any class F of functions f : X → Y and any sequence (x1 , ℓ1 ), (x2 , ℓ2 ), . . . of
contexts xt ∈ X and 1-Lipschitz loss functions ℓt : Y → [0, 1]. Assume HierHedge (Algorithm 6
in Appendix H) is run with full information feedback on the tree TF of depth M = ⌊ln2 (1/ε)⌋
with action set Yε for ε > 0. Moreover, the learning rate ηt (v) used at each node v at depth
m = 0, . . . , M − 1 is given by (11) in Appendix E, with E = 2−m+3 , N = |Cv |, and Vet−1 being the
cumulative variance of ℓes according to qs (v, ·) up to time s = t − 1. Then for all T > 1 the regret
satisfies
Z 1/2 p
2 T ln N∞ (F, x) + ln N∞ (F, x) dx .
RegT (F) 6 5εT + 27
ε/2
In particular, if d > 3 and F is the set of 1-Lipschitz functions f : [0, 1]d → [0, 1]p , where [0, 1]d and
1/d
e (d−1)/d ,
[0, 1]p are endowed with their sup norms, the
√ choice ε = (p/T ) yields RegT (F) = O T
while for 1 6 d 6 2 the regret is of order pT , ignoring logarithmic factors.
When using the sup norms, the dimension p of the action space only appears as a multiplicative factor p1/d in the regret bound for Lipschitz functions. Note also that an efficient version of HierHedge
for Lipschitz functions can be derived along the same lines as the construction in Section 4.3.
Acknowledgments
This work was partially supported by the CIMI (Centre International de Mathématiques et
d’Informatique) Excellence program. Pierre Gaillard and Sébastien Gerchinovitz acknowledge the
support of the French Agence Nationale de la Recherche (ANR), under grants ANR-13-BS01-0005
(project SPADRO) and ANR-13-CORD-0020 (project ALICIA).
References
Alekh Agarwal, Daniel Hsu, Satyen Kale, John Langford, Lihong Li, and Robert Schapire. Taming
the monster: A fast and simple algorithm for contextual bandits. In International Conference on
Machine Learning (ICML), 2014.
13
C ESA -B IANCHI G AILLARD G ENTILE G ERCHINOVITZ
Noga Alon, Nicolò Cesa-Bianchi, Claudio Gentile, Shie Mannor, Yishay Mansour, and
Ohad Shamir. Nonstochastic multi-armed bandits with graph-structured feedback. CoRR,
abs/1409.8428, 2014.
Noga Alon, Nicolò Cesa-Bianchi, Ofer Dekel, and Tomer Koren. Online learning with feedback
graphs: Beyond bandits. In COLT, pages 23–35, 2015.
Jean-Yves Audibert and Sébastien Bubeck. Regret bounds and minimax policies under partial
monitoring. Journal of Machine Learning Research, 11(Oct):2785–2836, 2010.
Peter Auer, Nicolò Cesa-Bianchi, Yoav Freund, and Robert E Schapire. The nonstochastic multiarmed bandit problem. SIAM journal on computing, 32(1):48–77, 2002.
Sébastien Bubeck and Nicolò Cesa-Bianchi. Regret analysis of stochastic and nonstochastic multiarmed bandit problems. Foundations and Trends in Machine Learning, 5(1):1–122, 2012.
Sébastien Bubeck, Rémi Munos, Gilles Stoltz, and Csaba Szepesvári. X-armed bandits. Journal of
Machine Learning Research, 12(May):1655–1695, 2011a.
Sébastien Bubeck, Gilles Stoltz, and Jia Yuan Yu. Lipschitz bandits without the Lipschitz constant.
In International Conference on Algorithmic Learning Theory, pages 144–158. Springer, 2011b.
Sébastien Bubeck, Ronen Eldan, and Yin Tat Lee. Kernel-based methods for bandit convex optimization. arXiv preprint arXiv:1607.03084, 2016.
Nicolò Cesa-Bianchi and Gábor Lugosi. On prediction of individual sequences. The Annals of
Statistics, 27(6):1865–1895, 1999.
Nicolò Cesa-Bianchi and Gábor Lugosi. Prediction, learning, and games. Cambridge university
press, 2006.
Nicolò Cesa-Bianchi, Yishay Mansour, and Gilles Stoltz. Improved second-order bounds for prediction with expert advice. Machine Learning, 66(2-3):321–352, 2007.
Nicolò Cesa-Bianchi, Claudio Gentile, and Yishay Mansour. Regret minimization for reserve prices
in second-price auctions. IEEE Transactions on Information Theory, 61(1):549–564, 2015.
Emile Contal and Nicolas Vayatis. Stochastic process bandits: Upper confidence bounds algorithms
via generic chaining. arXiv preprint arXiv:1602.04976, 2016.
Emile Contal, Cédric Malherbe, and Nicolas Vayatis. Optimization for gaussian processes via
chaining. arXiv preprint arXiv:1510.05576, 2015.
Rocco De Rosa, Francesco Orabona, and Nicolò Cesa-Bianchi. The ABACOC algorithm: A novel
approach for nonparametric classification of data streams. In Proceedings of the IEEE International Conference on Data Mining (ICDM), pages 733–738, 2015.
Richard M Dudley. The sizes of compact subsets of Hilbert space and continuity of Gaussian
processes. Journal of Functional Analysis, 1(3):290–330, 1967.
14
A LGORITHMIC C HAINING AND THE ROLE OF PARTIAL F EEDBACK IN O NLINE N ONPARAMETRIC L EARNING
Yoav Freund and Robert E Schapire. A decision-theoretic generalization of on-line learning and an
application to boosting. Journal of Computer and System Sciences, 55(1):119–139, 1997.
Pierre Gaillard and Sebastien Gerchinovitz. A chaining algorithm for online nonparametric regression. In Proceedings of COLT’15, volume 40, pages 764–796. JMLR: Workshop and Conference
Proceedings, 2015.
Sébastien Gerchinovitz and Tor Lattimore. Refined lower bounds for adversarial bandits. In Advances in Neural Information Processing Systems 29 (NIPS’16), pages 1198–1206, 2016.
Elad Hazan. Introduction to online convex optimization. Foundations and Trends R in Optimization,
2(3-4):157–325, 2015.
Elad Hazan and Nimrod Megiddo. Online learning with prior knowledge. In International Conference on Computational Learning Theory (COLT’07), pages 499–513. 2007.
Robert Kleinberg. Nearly tight bounds for the continuum-armed bandit problem. In NIPS, volume 17, pages 697–704, 2004.
Robert Kleinberg, Aleksandrs Slivkins, and Eli Upfal. Multi-armed bandits in metric spaces. In
Proceedings of the fortieth annual ACM symposium on Theory of computing, pages 681–690.
ACM, 2008.
Tyler Lu, Dávid Pál, and Martin Pál. Contextual multi-armed bandits. In AISTATS, pages 485–492,
2010.
Odalric-Ambrym Maillard and Rémi Munos. Adaptive bandits: Towards the best history-dependent
strategy. In AISTATS, pages 570–578, 2011.
Michael Ostrovsky and Michael Schwarz. Reserve prices in Internet advertising auctions: a field
experiment. In ACM Conference on Electronic Commerce, pages 59–60, 2011.
Alexander Rakhlin and Karthik Sridharan. Online nonparametric regression with general loss functions. CoRR, abs/1501.06598, 2015.
Alexander Rakhlin and Karthik Sridharan. Bistro: an efficient relaxation-based method for contextual bandits. In Proceedings of the Twentieth International Conference on Machine Learning
(ICML), 2016.
Alexander Rakhlin, Karthik Sridharan, and Ambuj Tewari. Online learning via sequential complexities. Journal of Machine Learning Research, 16:155–186, 2015.
Shai Shalev-Shwartz. Online learning and online convex optimization. Foundations and Trends in
Machine Learning, 4(2):107–194, 2011.
Aleksandrs Slivkins. Contextual bandits with similarity information. Journal of Machine Learning
Research, 15(1):2533–2568, 2014.
Vasilis Syrgkanis, Akshay Krishnamurthy, and Robert Schapire. Efficient algorithms for adversarial contextual learning. In Proceedings of the Twentieth International Conference on Machine
Learning (ICML), 2016.
15
C ESA -B IANCHI G AILLARD G ENTILE G ERCHINOVITZ
Vladimir Vovk. Competing with wild prediction rules. Machine Learning, 69(2-3):193–212, 2007.
Appendix A. A Useful Lemma
The following is a standard result in the analysis of Hedge with variable learning rate —see, e.g.,
the analysis of (Bubeck and Cesa-Bianchi, 2012, Theorem 3.2). Recall that, when run on a loss
sequence ℓ1 , ℓ2 , . . . Hedge
distributions p1 , p2 , . . . where p1 is uniform and pt (i) is proP computes
portional to exp − ηt t−1
ℓ
(i)
for all actions i.
s
s=1
Lemma 8 The sequence p1 , p2 , . . . of probability distributions computed by Hedge when run on
K actions with learning rates η1 > η2 > · · · > 0 and losses ℓt (k) > 0 satisfies
T X
K
X
t=1 i=1
pt (i)ℓt (i) − min
k=1,...,K
T
X
ℓt (k) 6
t=1
T
K
t=1
i=1
1X X
ln K
pt (i)ℓt (i)2 .
ηt
+
ηT
2
Appendix B. Proof of Theorem 1
Proof Let Nt be the number of balls created by the ContextualExp algorithm after t rounds, Bs be
the s-th ball so created, being x̄s its center (x̄s is some past context), and Ts be the subsequence of
rounds t such that xt is handled by the s-th ball. Notice that Nt and Ts are deterministic quantities,
since the xt ’s are generated obliviously. Since f is 1-Lipschitz
and ℓt is also 1-Lipschitz, for all xt
handled by Bs , we can write ℓt f (xt ) − ℓt f (x̄s ) ≤ ε . Now fix any 1-Lipschitz policy f . For
each s = 1, . . . , NT , there exists ȳs ∈ Yε such that ρY ȳs , f (x̄s ) ≤ ε. Then we can write
E
" T
X
t=1
#
ℓt (b
yt ) −
=
T
X
t=1
NT
X
ℓt f (xt )
(4)
X
E ℓt (b
yt ) − ℓt f (xt )
s=1 t∈Ts
=
NT X
X
E ℓt (b
yt ) − ℓt ȳs + ℓt ȳs − ℓt (f (x̄s )) + ℓt (f (x̄s )) − ℓt (f (xt ))
s=1 t∈Ts
≤
NT X
X
E ℓt (b
yt ) − ℓt ȳs + 2T ε .
(5)
s=1 t∈Ts
We now apply to each s = 1, . . . , NT the standard regret bound of Exp3 with learning rate η (e.g.,
(Bubeck and Cesa-Bianchi, 2012)) w.r.t. the best action ȳs . This yields
X
ln Kε η
+ |Ts | Kε ,
E ℓt (b
yt ) − ℓt ȳs 6
η
2
t∈Ts
implying
NT X
X
NT ln Kε η
E ℓt (b
yt ) − ℓt ȳs 6
+ T Kε .
η
2
s=1 t∈Ts
16
A LGORITHMIC C HAINING AND THE ROLE OF PARTIAL F EEDBACK IN O NLINE N ONPARAMETRIC L EARNING
q
Combining with (5), setting η = 2NTT KlnεKε , recalling that Kε 6 CY ε−p and observing that, by
the way balls are constructed, NT can never exceed the size of the smallest ε/2-cover of X (that is,
NT 6 CX (ε/2)−d ), we obtain
!
r
p
1
T ε−(d+p) ln + T ε .
RegT (F) 6 2 T NT Kε ln Kε + 2 T ε = O
ε
Setting ε =
p+d
2
ln T
T 1/2
2
p+d+2
p+d+1
e T p+d+2 as claimed.
gives RegT (F) = O
Appendix C. The Exp3-RTB algorithm for reserve-price optimization
√
e T can be achieved with high probCesa-Bianchi et al. (2015) showed that a regret of order O
ability for the problem of regret minimization in second-price auctions with i.i.d. bids, when the
feedback received after each auction is the seller’s revenue. In this appendix, we show that the same
regret rate can be obtained even when the sequence of bids is nonstochastic, provided the feedback
also includes the highest bid. We use this result in order to upper bound the contextual regret in
Section 4.1.
We consider a setting slightly more general than second-price auctions. Fix any unknown sequence ℓ1 , ℓ2 , . . . of loss functions ℓt : [0, 1] → [0, 1] satisfying the semi-Lipschitz condition,
ℓt (y + δ) ≥ ℓt (y) − δ
for all 0 6 y 6 y + δ 6 1.
(6)
In each auction instance t = 1, 2, . . . , the learner selects a reserve price ybt ∈ Y = [0, 1] and suffers
the loss ℓt (b
yt ). The learner’s feedback is ℓt (y) for all y > ybt (i.e., the one-sided full information
feedback). As explained in Section 4.1, this setting includes online revenue maximization in secondprice auctions as a special case when the learner’s feedback includes both revenue and highest bid.
The learner’s regret is defined by
#
" T
T
X
X
ℓt (b
yt ) − inf
ℓt (y) ,
RegT , E
06y61
t=1
t=1
where the expectation is with respect to the randomness in the predictions ybt . We introduce the
Exp3-RTBalgorithm (Algorithm 3 below), a variant of Exp3 (Auer et al., 2002) exploiting the richer
feedback ℓt (y) : y > ybt . The algorithm uses a discretization of the action space [0, 1] in
K = ⌈1/γ⌉ actions yk := (k − 1)γ for k = 1, . . . , K.
Theorem 9 In the one-sided full information feedback, the Exp3-RTB algorithm tuned with γ > 0
satisfies
2 ln⌈1/γ⌉
1 e
.
+
RegT 6 γT 2 + ln
4 γ
γ
√
e T .
In particular, γ = T −1/2 gives RegT = O
17
C ESA -B IANCHI G AILLARD G ENTILE G ERCHINOVITZ
Algorithm 3: Exp3-RTB (for one-sided full information feedback)
Input
: Exploration parameter γ > 0.
Initialization: Set learning rate η = γ/2 and uniform distribution p1 over {1, . . . , K} where
K = ⌈1/γ⌉;
for t = 1, 2, . . . do
1. compute distribution qt (k) = (1 − γ)pt (k) + γIk=1 for k = 1, . . . , K;
2. draw It ∼ qt and choose ybt = yIt = (It − 1)γ;
3. for each k = 1, . . . , K, compute the estimated loss
ℓt (yk )
IIt 6k
ℓbt (k) = Pk
j=1 qt (j)
end
4. for each k = 1, . . . , K, compute the new probability assignment
P
exp − η ts=1 ℓbs (k)
P
.
pt+1 (k) = P
K
t
bs (j)
exp
−η
ℓ
j=1
s=1
Proof The proof follows the same lines as the regret analysis of Exp3 in (Auer et al., 2002). The
key change is a tighter control of the variance term allowed by the richer feedback.
Pick any reserve price yk = (k − 1)γ. We first control the regret associated with actions drawn
from pt (the regret associated with qt will be studied as a direct consequence). More precisely, since
the estimated losses ℓbt (j) are nonnegative, we can apply Lemma 8 to get
T
X
t=1
pt · ℓbt −
T K
ln K
η XX
b
.
pt (j)ℓbt (j)2 +
ℓt (k) 6
2 t=1
η
t=1
T
X
(7)
j=1
Writing Et−1 [ · ] for the expectation conditioned on I1 , . . . , It−1 , we note that
Et−1 ℓbt (j) = ℓt (yj )
and
pt (j)ℓt (yj )2
qt (j)
Et−1 pt (j)ℓbt (j)2 = Pj
6
,
P
(1 − γ) ji=1 qt (i)
i=1 qt (i)
where we used the definition of qt and the fact that |ℓt (yj )| 6 1 by assumption. Therefore, taking
expectation on both sides of (7) entails, by the tower rule for expectations,
#
" T
K
T
T
X
X
X
X
q
(j)
η
t
+ ln K .
E
ℓt (yk ) 6
p t · ℓt −
E
Pj
2(1 − γ)
η
i=1 qt (i)
t=1
t=1
t=1
18
j=1
A LGORITHMIC C HAINING AND THE ROLE OF PARTIAL F EEDBACK IN O NLINE N ONPARAMETRIC L EARNING
Setting st (j) ,
K
X
j=1
Pj
i=1 qt (i),
qt (j)
Pj
i=1 qt (i)
we can upper bound the sum with an integral,
=1+
K
X
st (j) − st (j − 1)
st (j)
j=2
61+
K Z
X
j=2
st (j)
st (j−1)
=1+
K Z
X
j=2
dx
=1+
x
Z
1
qt (1)
st (j)
st (j−1)
dx
st (j)
dx
1
6 1 − ln qt (1) 6 1 + ln ,
x
γ
where we used qt (1) > γ. Therefore, substituting into the previous bound, we get
#
" T
T
X
X
ηT ln(e/γ) ln K
ℓt (yk ) 6
p t · ℓt −
E
+
.
2(1 − γ)
η
(8)
t=1
t=1
We now control the regret of the predictions ybt = yIt , where It is drawn from qt = (1 − γ)pt + γδ1 .
By the tower rule,
" T
#
#
" T
T
T
X
X
X
X
ℓt (yk )
ℓt (yk ) = E
(1 − γ)pt · ℓt + γℓt (y1 ) −
ℓt (b
yt ) −
E
t=1
t=1
t=1
6 (1 − γ) E
6
"
T
X
t=1
p t · ℓt −
T
X
t=1
#
t=1
ℓt (yk ) + γT
ηT ln(e/γ) ln K
+
+ γT ,
2
η
(9)
where the last inequality is by (8).
To conclude the proof, we upper bound the regret against any fixed y ∈ [0, 1]. Since there
exists k ∈ {1, . . . , K} such that y ∈ [yk , yk + γ], and since each ℓt satisfies the semi-Lipschitz
condition (6), we have ℓt (y) > ℓt (yk ) − γ. This gives
#
" T
T
X
X
ℓt (yk ) 6 min
min E
ℓt (y) + γ T .
k=1,...,K
06y61
t=1
t=1
Replacing the last inequality into (9), and recalling that K = ⌈1/γ⌉ and η = γ2 , finally yields
RegT 6
γT
e 2 ln⌈1/γ⌉
ln +
+ 2γT .
4
γ
γ
Choosing γ ≈ T −1/2 concludes the proof.
Appendix D. Proof of Theorem 3
Proof The proof is very similar to that of Theorem 1. We only highlight the main differences.
Let f ∈ F be any 1-Lipschitz function from X to [0, 1]. For all s = 1, . . . , NT , we define the
constant approximation of f in the s-th ball by ysmin = mint∈Ts f (xt ). Since f is 1-Lipschitz and
19
C ESA -B IANCHI G AILLARD G ENTILE G ERCHINOVITZ
the balls have radius ε, we have maxt,t′ ∈Ts f (xt ) − f (xt′ ) ≤ 2ε. Hence, for all t ∈ Ts , by the
semi-Lipschitz property (6),
ℓt f (xt ) ≥ ℓt (ysmin ) − 2ε .
(10)
Therefore,
T
X
t=1
E[ℓt (b
yt )] −
T
X
t=1
NT X
X
ℓt f (xt ) =
E[ℓt (b
yt )] − ℓt f (xt )
s=1 t∈Ts
=
NT X
NT X
X
X
ℓt (ysmin ) − ℓt f (xt )
E[ℓt (b
yt )] − ℓt (ysmin ) +
s=1 t∈Ts
≤ γT
|
{z
Rs
1 e
2 + ln
4 γ
+
}
s=1 t∈Ts
2NT ln⌈1/γ⌉
+ 2εT .
γ
Each term Rs is the regret suffered by the s-th instance of Exp3-RTB against the constantP
value ysmin ,
and so we bound it using Theorem 9 in Appendix C and then sum over s recalling that s Ts = T .
The other double sum is bounded by 2T ε using (10). We bound NT as in Theorem 1 obtaining
ε−d 1
1
ln + εT .
+
RegT (F) . γT 1 + ln
γ
γ
γ
Finally, choosing ε = γ = T −1/(d+2) 6 1 gives
concluding the proof.
d+1
e T d+2
RegT (F) = O
Appendix E. Exp4 regret bound scaling with the range of the losses
In this section we revisit the Exp4 algorithm (Bubeck and Cesa-Bianchi, 2012, Figure 4.1) and
prove a regret bound for the one-sided full information model that scales with the range of the
losses. In view of the application to chaining, we formulate this result in a setting similar to
(Maillard and Munos, 2011). Namely, when the action It played at time t is not necessarily drawn
from the distribution prescribed by Exp4.
Exp4 (see Algorithm 4) operates in a bandit setting with K actions and N experts. We assume
a total order 1 < · · · < K on the action set K = {1, . . . , K}. At each time step t = 1, 2, . . . the
advice ξt (j, ·) of expert j is a probability distribution over the K actions. The learner combines
the expert advice using convex coefficients qt ∈ ∆(N ). These coefficients are computed by Hedge
b
based on the expert losses ℓet (j) = ξt (j, ·) · ℓbt for j = 1, . . . , N , where
PNℓt is a vector of suitably
defined loss estimates. The distribution prescribed by Exp4 is pt = j=1 qt (j)ξt (j, ·) ∈ ∆(K),
but the learner’s play at time t is It ∼ p∗t for some other distribution p∗t ∈ ∆(K). The feedback at
time t is the vector of losses ℓt (i) for all i > It .
20
A LGORITHMIC C HAINING AND THE ROLE OF PARTIAL F EEDBACK IN O NLINE N ONPARAMETRIC L EARNING
Algorithm 4: Exp4 with unspecified sampling distributions p∗t (for one-sided full info feedback)
Input
: Learning rate sequence η1 > η2 > · · · > 0.
Initialization: Set q1 to the uniform distribution over {1, . . . , N };
for t = 1, 2, . . . do
1. get expert advice ξt (1, ·), . . . , ξt (N, ·) ∈ ∆(K);
2. compute distribution p∗t over {1, . . . , K};
3. draw It ∼ p∗t and observe ℓt (i) for all i > It ;
4. compute loss estimates ℓbt (i) for i = 1, . . . , K;
5. compute expert losses ℓet (j) = ξt (j, ·) · ℓbt for j = 1, . . . , N ;
6. compute the new probability assignment
P
exp −ηt+1 ts=1 ℓes (j)
for each j = 1, . . . , N .
qt+1 (j) = P
Pt e
N
k=1 exp −ηt+1
s=1 ℓs (k)
end
In this setting, the goal is to minimize the regret with respect to the performance of the best
expert,
#
" T
T
X
X
ξt (j, ·) · ℓt
p t · ℓt −
max E
j=1,...,N
t=1
t=1
in expectation with respect to the random draw of I1 , . . . , IT , where ξt (j, ·) · ℓt is the expected loss
of expert j at time t. In view of our chaining application, we defined the quantities ξt (j, i) as random
variables because the expert advice at time t might depend on the past realizations of I1 , . . . , It−1 .
For all k ∈ K let
k
X
p∗t (i) .
Pt∗ (k) =
i=1
Also, let Kt ≡ {k ∈ K : (∃j) ξt (j, k) > 0} (in chaining, Kt is typically small compared to K).
The next result shows that, when Exp4 is applied to arbitrary real-valued losses, there is a way of
choosing the learning rate sequence so that the regret scales with a quantity smaller than the largest
loss value. More specifically, the regret scales with a known bound E on the size of the effective
range of the losses
max max
ℓt (k) − ℓt (k′ ) 6 E .
′
t=1,...,T k,k ∈Kt
The rich feedback
structure (ℓt (k), k > It ) is crucial to get our result, since it enables us to use
ℓt max Kt in the definition of the loss estimates ℓbt (k) below. Indeed, as explained in Section 4.2,
ℓbt (k) has to be explicitly computed only for those actions k such that It 6 k and ξt (j, k) > 0 for
some j, i.e., k ∈ Kt . In this case, It 6 k 6 max Kt so that ℓt max Kt is observed.
Theorem 10 Let 1 < · · · < K be a total order on the action set K = {1, . . . , K}. Suppose
Exp4 (Algorithm 4) is run with one-sided full information feedback on an arbitrary loss sequence
21
C ESA -B IANCHI G AILLARD G ENTILE G ERCHINOVITZ
ℓ1 , ℓ2 , . . . with loss estimates
ℓt (k) − ℓt max Kt
b
ℓt (k) =
IIt 6k
Pt∗ (k)
for k = 1, . . . , K
and adaptive learning rate
ηt = min
(
γ
,
2E
s
2
)
2 − 1 (ln N )
(e − 2)Vet−1
√
for t > 2
(11)
P
PN
e
e 2 is the cumulative
for some parameter E > 0 and where Vet−1 = t−1
s=1
j=1 qs (j) ℓs (j) − qs · ℓs
variance of Hedge up to time t − 1. If p∗t (1) > γ and
max ℓt (k) − ℓt (k′ ) 6 E
max
t=1,...,T k,k ′ ∈Kt
almost surely for all t > 1, then, for all T > 1,
s
#
" T
T
X
X
E
T ln N
ξt (j, ·) · ℓt 6 4E
p t · ℓt −
max E
+ (4 ln N + 1) .
j=1,...,K
γ
γ
t=1
t=1
Note that the above regret bound does not depend on the number K of actions, but instead on a
lower bound γ on the probability of observing the smallest action.
Proof From the definition of Kt and because Pt∗ (k) > p∗t (1) > γ,
max
i,j=1,...,N
ℓet (i) − ℓet (j) =
max
i,j=1,...,N
ξt (i, ·) · ℓbt − ξt (j, ·) · ℓbt
6 max
ℓbt (k) − ℓbt (k′ )
′
6
k,k ∈Kt
2E
2E
6
.
∗
mink∈K Pt (k)
γ
e = (2E)/γ is an upper bound on the size of the range of the losses ℓet (j), we can use
Since E
the bound of (Cesa-Bianchi et al., 2007, Theorem 5), which applies to Hedge run on arbitrary realvalued losses with the learning rate (11). This gives us
q
T
T
X
X
e
e
e
e ln N + E .
(12)
qt · ℓt 6 min
ℓt (j) + 4 (ln N )VeT + 2E
j=1,...,N
2
t=1
t=1
Note that qt · ℓet = pt · ℓbt . Moreover, Kt = max Kt is measurable with respect to (I1 , . . . , It−1 ),
implying Et−1 ℓbt (k) = ℓt (k)−ℓt (Kt ). Therefore, taking the expectation on both sides of (12) with
P P
respect to the random draw of I1 , . . . , IT , and using Jensen together with VeT ≤ Tt=1 N
j=1 qt (j) ·
2
ℓet (j) , we get
#
" T
T
X
X
ξt (j, ·) · ℓt
p t · ℓt −
max E
j=1,...,N
t=1
t=1
v
u
T
N
u
X
X
2
4E
E
u
qt (j) ξt (j, ·) · ℓbt +
E
6 4t(ln N )
ln N + .
γ
γ
t=1
j=1
22
(13)
A LGORITHMIC C HAINING AND THE ROLE OF PARTIAL F EEDBACK IN O NLINE N ONPARAMETRIC L EARNING
The variance term inside the square root can be upper bounded as follows. Using Jensen again,
N
X
j=1
K
N
2 X
X
b
qt (j)
ξt (j, k)ℓbt (k)2
qt (j) ξt (j, ·) · ℓt 6
=
j=1
k=1
N
X
K
X
qt (j)
j=1
6
ξt (j, k)
k=1
ℓt (k) − ℓt (Kt )
Pt∗ (k)
2
IIt 6k
K
N
X
E2 X
ξt (j, k)
qt (j)
II 6k
γ
Pt∗ (k) t
j=1
k=1
where the last inequality is because ℓt (k) − ℓt (Kt ) 6 E when ξt (j, k) > 0 and because Pt∗ (k) >
p∗t (1) > γ. Therefore, recalling Pt∗ (k) = Pt−1 (It 6 k),
N
K
N
X
X
2
ξt (j, k)
E2 X
E2
Et−1
qt (j) ξt (j, k) · ℓbt 6
qt (j)
P
(I
6
k)
=
.
t−1 t
γ
Pt∗ (k)
γ
j=1
j=1
k=1
Substituting the last bound in (13) concludes the proof.
Next we extend the previous result to penalized loss estimates, which is useful in Section 4.3 to
control the variance terms all along the covering tree.
Theorem 11 Let 1 < · · · < K be a total order on the action set K = {1, . . . , K}. Let E, F > 0.
Consider any penalty pent ∈ RK measurable with respect to (I1 , . . . , It−1 ) at time t. Suppose
Exp4 (Algorithm 4) is run with one-sided full information feedback on an arbitrary loss sequence
ℓt ∈ RK , t > 1, with loss estimates
ℓt (k) − ℓt max Kt + E
b
IIt 6k + pent (k) + F
for k = 1, . . . , K
ℓt (k) =
Pt∗ (k)
and constant learning rate η > 0. If we have, for all t > 1, almost surely,
max
max ℓt (k) − ℓt (k′ ) 6 E
and
t=1,...,T k,k ′ ∈Kt
max max pent (k) 6 F ,
t=1,...,T k∈Kt
then, for all T > 1,
max E
j=1,...,K
"
T
X
t=1
pt · ℓt + pent −
T
X
ξt (j, ·) · ℓt + pent
t=1
T
X
2
ln N
6
+ 4ηT F 2 + 4ηE
η
t=1
"
K
X
pt (k)
E
Pt∗ (k)
k=1
#
#
.
Proof From the definition of Kt and because E and F are upper bounds on the losses and penalties
associated with actions in Kt ,
min ℓet (j) =
j=1,...,N
min ξt (j, ·) · ℓbt > min ℓbt (k) > 0 .
j=1,...,N
k∈Kt
23
C ESA -B IANCHI G AILLARD G ENTILE G ERCHINOVITZ
We can thus use the regret bound for Hedge with weight vectors qt , constant learning rate η, and
nonnegative losses ℓet (k) (see Lemma 8 in Appendix A):
T
X
t=1
qt · ℓet 6
min
j=1,...,N
T
X
t=1
T
N
η XX
ln N
+
qt (j)ℓet (j)2 .
ℓet (j) +
η
2
(14)
t=1 j=1
Note that qt · ℓet = pt · ℓbt . Moreover,
Kt = max Kt and pent are measurable with respect to
b
(I1 , . . . , It−1 ), implying Et−1 ℓt (k) = ℓt (k) − ℓt (Kt ) + E + pent (k) + F . Therefore, taking the
expectation on both sides of (14) with respect to the random draw of I1 , . . . , IT , we get
#
" T
T
X
X
ξt (j, ·) · ℓt + pent
pt · ℓt + pent −
max E
j=1,...,N
t=1
t=1
6
ln N
η
+
η
2
T
X
t=1
N
2
X
qt (j) ξt (j, ·) · ℓbt .
E
(15)
j=1
The variance term can be upper bounded as follows. Using Jensen’s inequality,
N
X
j=1
K
N
2 X
X
qt (j)
ξt (j, k)ℓbt (k)2
qt (j) ξt (j, ·) · ℓbt 6
j=1
=
N
K X
X
k=1
qt (j)ξt (j, k)
k=1 j=1
ℓt (k) − ℓt (Kt ) + E
IIt 6k + pent (k) + F
Pt∗ (k)
2
K
X
pt (k)
II 6k ,
6 8F + 8E
Pt∗ (k)2 t
2
2
k=1
where the last inequality is because ℓt (k) − ℓt (Kt ) 6 E and pent (k) 6 F when ξt (j, k) > 0
and because (a + b + c + d)2 6 4 a2 + b2 + c2 + d2 . Therefore, recalling Pt∗ (k) = Pt−1 (It 6 k),
N
K
X
X
2
pt (k)
Et−1
qt (j) ξt (j, k) · ℓbt 6 8F 2 + 8E 2
.
Pt∗ (k)
j=1
k=1
Substituting the last bound in (15) concludes the proof.
∗
Note that when
p pent ≡ 0 and pt (1) > γ almost surelypfor all t > 1, Theorem 11 above used with
−1
η = (2E)
γ ln(N )/T yields a regret bound of 4E T ln(N )/γ, similarly to Theorem 10.
We use yet another corollary in Section 4.3. It follows directly from the choice of pent (k) =
−α/Pt∗ (k) and F = α/γ in Theorem 11.
Corollary 12 Let 1 < · · · < K be a total order on the action set K = {1, . . . , K}. Let E, α, γ > 0
be three parameters. Suppose Exp4 (Algorithm 4) is run on an arbitrary loss sequence ℓt ∈ RK ,
t > 1, with loss estimates
ℓt (k) − ℓt max Kt + E
α
α
b
IIt 6k − ∗
+
for k = 1, . . . , K
ℓt (k) =
∗
Pt (k)
Pt (k) γ
24
A LGORITHMIC C HAINING AND THE ROLE OF PARTIAL F EEDBACK IN O NLINE N ONPARAMETRIC L EARNING
and constant learning rate η > 0. If we have, for all t > 1, almost surely,
max ℓt (k) − ℓt (k′ ) 6 E
max
t=1,...,T k,k ′ ∈Kt
and
p∗t (1) > γ ,
then, for all T > 1,
" T K
X
#
T X
K
XX
α
α + 4ηE 2
−
ξt (j, k) ℓt (k) − ∗
max E
pt (k) ℓt (k) −
j=1,...,K
Pt∗ (k)
Pt (k)
t=1 k=1
t=1 k=1
6
ln N
+
η
4ηT α2
γ2
.
Appendix F. Missing Proofs from Section 4.2
We prove Theorem 4 and Corollary 5. As we said in the main text, the key contribution of chaining
is that it allows us to sum up local regret bounds scaling as the range of the local losses. This is
possible because of the richer feedback structure induced by the total order on the actions. We first
state a lemma indicating that the losses associated with neighboring nodes are indeed close to one
another. Recall that M is the depth of TF .
Lemma 13 Let v ∈ T be any node at level m ∈ {0, 1, . . . , M − 1}. Then all leaves w, w′ ∈ Lv
satisfy fw − fw′ ∞ 6 2−m+2 . Therefore, ℓt (yit (w) ) − ℓt (yit (w′ ) ) 6 2−m+3 for all t > 1.
Proof Consider a path v = vm → vm+1 → · · · → vM = w joining v to leaf w in the tree. For each
k = m, . . . , M −1, since Fk is a (2−k )-covering of F in sup norm, we have kfvk −fvk+1 k∞ 6 2−k .
Therefore,
kfv − fw k∞ =
M
−1
X
k=m
(fvk − fvk+1 )
6
∞
M
−1
X
k=m
kfvk − fvk+1 k∞ 6
M
−1
X
2−k 6 2−m+1 .
k=m
Therefore, since
∈ Lv as well, kfw − fw′ k∞ = kfv − fw k∞ + kfv − fw′ k∞ 6 2−m+2 , which
proves the first inequality. Now, since ℓt is 1-Lipschitz,
ℓt (yit (w) ) − ℓt (yit (w′ ) ) 6 yit (w) , yit (w′ ) 6 2−M +1 + kfw − fw′ k∞
6 2−M +1 + 2−m+2 6 2−m+3
w′
where the second inequality uses the definition of it (v) in Section 4.2 and the fact that K is a (2−M )covering of [0, 1]. This concludes the proof.
We are now ready to prove Theorem 4 from the main text.
Proof (of Theorem 4) Each node v at level m = 0, . . . , M − 1 is running an instance of the
variant of Exp4 described in Algorithm 4 (Appendix E) over expert set Cv , where the advice of
w ∈ Cv is ξt (w, ·) = pt (w, ·), and effective action set Kt = {i : (∃w ∈ Cv ) pt (w, i) > 0} = Kt (v).
Note that, for any w ∈ Cv , the distribution pt (w, ·) is a mixture of actions in Lw ⊆ Lv , so that
Kt (v) ⊆ {it (w′ ) : w′ ∈ Lv }. By Lemma 13, the losses of these actions belong to a range of size
E = 2−m+3 . Since p∗t (1) > γ by definition, we are in position to apply Theorem 10 in Appendix E
with N = |Cv | ≤ Nm+1 , and obtain the bound
s
#
" T
T
X
X
T ln Nm+1 2−m+3
+
(4 ln Nm+1 + 1)
pt (w, ·) · ℓt 6 2−m+5
pt (v, ·) · ℓt −
max E
w∈Cv
γ
γ
t=1
t=1
25
C ESA -B IANCHI G AILLARD G ENTILE G ERCHINOVITZ
(for simplicity, we use ℓt to denote the vector of elements ℓt (yi ) for i = 1, . . . , K). Now consider
the path v0 → v1 → · · · → vM = w∗ from the root v0 to the leaf vM = w∗ minimizing ℓ1 (yi1 (w) )+
· · · + ℓT (yiT (w) ) over w ∈ L. Recalling that pt (w, i) = Ii=it (w) for any leaf w, we get
E
" T
X
t=1
#
pt (v0 , ·) · ℓt − min
ℓt (yit (w) )
t=1
" T
M
−1
X
X
w∈L
=
T
X
E
m=0
6 25
t=1
M
−1
X
pt (vm , ·) · ℓt −
s
2−m
m=0
T
X
t=1
pt (vm+1 , ·) · ℓt
#
!
1
T ln Nm+1
+ (ln Nm+1 + 1) .
γ
γ
(16)
Now, recalling that It ∼ p∗t and p∗t (i)ℓt (yi ) = (1 − γ)pt (v0 , i)ℓt (yi ) + γ Ii=1 ℓt (y1 ) we get
)
#
( T
" T
T
X
X
X
∗
ℓt (y1 ) − ℓt yit (w)
ℓt yit (w) + γ
E
pt · ℓt 6 min
w∈L
t=1
+ 25
M
−1
X
s
2−m
m=0
t=1
t=1
!
1
T ln Nm+1
+ (ln Nm+1 + 1) .
γ
γ
Now clearly ℓt (y1 ) − ℓt yit (w) 6 1. Moreover, because L is a (2−M )-covering of F and K is a
(2−M )-covering of [0, 1], for any f ∈ F there exists w ∈ L such that ℓt yit (w) − ℓt f (xt ) 6
yit (w) − f (xt ) 6 1 yit (w) − fw (xt ) + fw (xt ) − f (xt ) 6 21−M by definition of it (w). Hence,
E
" T
X
t=1
#
p∗t · ℓt − inf
f ∈F
T
X
t=1
ℓt f (xt )
6 (21−M + γ)T + 25
M
−1
X
2−m
m=0
s
!
T ln Nm+1
1
+ (ln Nm+1 + 1) .
γ
γ
(17)
We now use Nm+1 = N∞ (F, 2−(m+1) ), and follow the standard chaining approach approximating
the sums by integrals,
M
−1
X
m=0
−m
2
p
ln Nm+1 = 4
64
M
−1
X
−(m+1)
2
m=0
M
−1 Z 2−(m+1)
X
m=0
2−(m+2)
−(m+2)
−2
p
q
ln N∞ (F, 2−(m+1) )
ln N∞ (F, ε)dε 6 4
Z
1/2 p
γ/2
ln N∞ (F, ε)dε
(18)
where the second inequality is by monotonicity of ε 7→ ln N∞ (F, ε) and the last inequality follows
from γ 6 2−M due to M = ⌊ln2 (1/γ)⌋. Similarly,
M
−1
X
m=0
−m
2
ln Nm+1 6 4
Z
26
1/2
γ/2
ln N∞ (F, ε)dε .
(19)
A LGORITHMIC C HAINING AND THE ROLE OF PARTIAL F EEDBACK IN O NLINE N ONPARAMETRIC L EARNING
We conclude the proof by substituting (18) and (19) into (17), and using 2−M 6 2γ.
Proof (of Corollary 5) The metric entropy satisfies ln N∞ (F, ε) = O ε−d . Therefore, substituting into the regret bound of Theorem 4 and computing the integrals, the regret satisfies
q
1
1
T
+
ln
if d = 1
O
γ
γ
q γ
T
1
−2
RegT (F) 6 5T γ +
O
if d = 2
γ ln γ + γ
√
O T γ (1−d)/2 + γ −d if d > 3.
Optimizing γ for the different choices of d concludes the proof.
Appendix G. Algorithm HierExp4⋆ and proof of Theorem 6
G.1. Algorithm HierExp4⋆
We first construct the HierExp4⋆ algorithm used in Theorem 6. It is a variant of HierExp4 based
on a special hierarchical covering of F that we first define below. Recall that F is the set of all
1-Lipschitz functions from [0, 1]d to [0, 1].
In the sequel, we use a dyadic discretization of the input space [0, 1]d . For each depth m =
0, 1, . . . we define a partition of [0, 1]d with 2md equal cubes of width 2−m . Each cube Xm (σ1:m ) ⊆
[0, 1]d at depth m is indexed by σ1:m = (σ1 , . . . , σm ) ∈ {1, . . . , 2d }m . Note that the partitions at
different depths are nested. We can thus represent them via a tree P whose root ∅ is labeled by
[0, 1]d and whose each node σ1:m at depth m is labeled with the cube Xm (σ1:m ). The children of
node σ1:m are the nodes at depth m + 1 of the form (σ1 , . . . , σm , σm+1 ) with σm+1 ∈ {1, . . . , 2d }.
They correspond to sub-cubes of Xm (σ1:m ).
Wavelet-like approximation of F. Using the above dyadic discretization P, we approximate any
f ∈ F with piecewise-constant functions fM : [0, 1]d → [−1/2, 3/2] of the form
fM (x) =
M
1 X
+
2 m=1
X
2−m cm (σ1:m ) Ix∈Xm (σ1:m )
σ1:m ∈{1,...,2d }m
where cm (σ1:m ) ∈ −1, 0, 1 . (20)
As shown in the next lemma, the functions fM form a 2−M -covering of F in the sup norm
kgk∞ = supx∈[0,1]d g(x) . We use the following important property: for any given x ∈ [0, 1]d ,
there exists a unique σ1:M ∈ {1, . . . , 2d }M such that x ∈ Xm (σ1:m ) for all m = 1, . . . , M , so that
M
1 X −m
fM (x) = +
2 cm (σ1:m ) .
2 m=1
(21)
Lemma 14 Let f ∈ F and M > 1. There exist coefficients cm (σ1:m ) ∈ −1, 0, 1 such that fM
defined in (20) satisfies kfM − f k∞ 6 2−M .
27
C ESA -B IANCHI G AILLARD G ENTILE G ERCHINOVITZ
Proof We denote the center of each cube Xm (σ1:m ) by xm (σ1:m ) and prove by induction on M > 1
that there exist coefficients cm (σ1:m ), m = 0, . . . , M , such that
(22)
fM xM (σ1:M ) − f xM (σ1:M ) 6 2−(M +1) .
This directly yields the conclusion, since every x ∈ XM (σ1:M ) is 2−(M +1) -close to the
center
xM (σ1:M ) in sup norm and f is 1-Lipschitz, which by (22) and fM (x) = fM xM (σ1:M ) implies
fM (x) − f (x) 6 fM xM (σ1:M ) − f xM (σ1:M ) + f xM (σ1:M ) − f (x)
6 2−(M +1) + 2−(M +1) = 2−M .
We now carry out the induction. For M = 1 and σ1 ∈ {1, . . . , 2d }, we set
c1 (σ1 ) ∈ arg min
c∈{−1,0,1}
1 c
+ − f x1 (σ1 ) ,
2 2
which corresponds to projecting the value of f at the center x1 (σ1 ) onto the coarse grid {0, 1/2, 1}.
Therefore 1/2 + c1 (σ1 )/2 − f (x1 (σ1 )) 6 1/4, which implies (22) by (21).
Now, let M > 2 and assume that there exist coefficients cm (σ1:m ), m = 0, . . . , M − 1, such
that (22) holds true at level M − 1. We complete these coefficients at level M as follows: for all
σ1:M ∈ {1, . . . , 2d }M , we set
c
cM (σ1:M ) ∈ arg min fM −1 xM (σ1:M ) + M − f xM (σ1:M ) .
2
c∈{−1,0,1}
Note that by (21) and xM (σ1:M ) ∈ XM (σ1:M ) ⊂ XM −1 (σ1:(M −1) ), we have
cM (σ1:M )
.
fM xM (σ1:M ) = fM −1 xM (σ1:M ) +
2M
(23)
Therefore, the definition of cM (σ1:M ) above corresponds to making fM xM (σ1:M ) as close to
f xM (σ1:M ) as possible. Now, note that fM −1 is constant over XM −1 (σ1:(M −1) ), so that the
difference ∆M −1 , fM −1 xM (σ1:M ) − f xM (σ1:M ) satisfies
∆M −1 = fM −1 xM −1 (σ1:(M −1) ) − f xM (σ1:M )
6 fM −1 xM −1 (σ1:(M −1) ) − f xM −1 (σ1:(M −1) ) + f xM −1 (σ1:(M −1) ) − f xM (σ1:M )
(22)
6 2−M + xM −1 (σ1:(M −1) ) − xM (σ1:M )
∞
= 2−M + 2−(M +1) ,
where the last inequality is by (22) at level M − 1, and where the last equality is by comparison of
two cube centers at subsequent depths.
To conclude the proof, we note that the bound ∆M −1 6 2−M +2−(M +1) implies the existence
of cM (σ1:M ) ∈ {−1, 0, 1} such that fM xM (σ1:M ) − f xM (σ1:M ) 6 2−(M +1) , as required
to finish the induction. We can indeed consider three possible cases. If ∆M −1 6 2−(M +1) , then
setting cM (σ 1:M ) = 0 concludes the proof
(by (23)). If ∆M −1 > 2−(M +1) , then ∆M −1 lies in
the interval 2−(M +1) , 2−M + 2−(M +1) , so that setting cM (σ1:M ) = −1 also concludes the proof
(using (23) again). Similarly we set cM (σ1:M ) = 1 when ∆M −1 < −2−(M +1) .
We are now ready to define our hierarchical covering of F.
28
A LGORITHMIC C HAINING AND THE ROLE OF PARTIAL F EEDBACK IN O NLINE N ONPARAMETRIC L EARNING
Construction of TF⋆ . We build a tree TF⋆ such that the value of any function fM at any point
x ∈ [0, 1]d —as given by (21)— can be computed by following a path from the root to a leaf of
TF⋆ . Unlike Section 4.2, this tree is not labeled by functions but instead by either values in R or
cubes Xm (σ1:m ) ⊆ [0, 1]d , as illustrated in Figure 3. More precisely, our tree TF⋆ is composed of
two types of nodes:
• Nodes at depths m. The root v0 = ∅ (at depth m = 0) is labeled by 1/2. Each node at depth
m = 1, 2, . . . is indexed by a tuple (σ1 , c1 , σ2P
, c2 , . . . , σm , cm ) with σk ∈ {1, . . . , 2d } and
−k
ck ∈ {−1, 0, 1}; the node is labeled by 1/2 + m
k=1 2 ck , which corresponds to (21).
• Nodes at depths m + 1/2. Each node at depth m + 1/2, m > 0, is indexed by a tuple
(σ1 , c1 , σ2 , c2 , . . . , σm , cm , σm+1 ) with σk ∈ {1, . . . , 2d } and ck ∈ {−1, 0, 1}; the node is
labeled by the cube Xm (σ1:m+1 ).
We connect nodes in the following natural way. Nodes v = (σ1 , c1 , σ2 , c2 , . . . , σm , cm ) at depth m
are connected to all nodes at depth m + 1/2 of the form (v, σm+1 ), with σm+1 ∈ {1, . . . , 2d }.
Nodes w = (σ1 , c1 , σ2 , c2 , . . . , σm , cm , σm+1 ) at depth m + 1/2 are connected to all nodes at depth
m + 1 of the form (w, cm+1 ), with cm+1 ∈ {−1, 0, 1}.
We stop the tree at depth M > 1. As previously, we denote by L the set of all the leaves of TF⋆ ,
by Lv the set of all the leaves under v ∈ TF⋆ , and by Cv the set of children of v ∈ TF⋆ .
1
2
depth
0
depth
0+
depth
1
depth
1+
binning
Exp4 0, 12
[ 12 , 1[
1
2
0
1
Exp4
1
2
0
1
[0, 41 [
[ 41 , 12 [
[0, 14 [
[ 14 , 12 [
[0, 41 [
[ 14 , 21 [
[ 12 , 43 [
[ 43 , 1]
[ 12 , 43 [
[ 43 , 1]
[ 21 , 34 [
[ 34 , 1]
..
.
..
.
..
.
..
.
..
.
..
.
..
.
..
.
..
.
..
.
..
.
..
.
1
2
1
2
Figure 3: The structure of the efficient chaining tree in dimension d = 1.
Algorithm HierExp4⋆ . Our new bandit algorithm HierExp4⋆ (Algorithm 5) is an efficient modification of HierExp4 based on the tree TF⋆ defined above. HierExp4⋆ alternates between “binning
nodes" at depths m and “Exp4 nodes" at depths m + 1/2 (see Figure 3). We discretize the action
space Y = [0, 1] with
K = k2−M : k = 1, . . . , 2M } of cardinality K = 2M . Note that, af −M
ter projecting onto 2 , 1 , the set of labels of all the leaves v ∈ L exactly corresponds to this
discretization, i.e.,
! !
)
(
M
1
k
1 X −m
M
2 cm , M , 1 : cm ∈ {−1, 0, 1} =
+
: k = 1, . . . , 2
= K.
min max
2 m=1
2
2M
At every round t, after observing
context xt ∈ [0, 1]d , the algorithm activates the root v0 and all
nodes σ1 , c1 , . . . , σm (, cm ) such that xt ∈ Xm (σ1:m ); the other nodes are asleep during that
round. In particular, there are 3M activated leaves v = (σ1 , c1 , . . . , σM , cM ) ∈ L. Each such leaf
29
C ESA -B IANCHI G AILLARD G ENTILE G ERCHINOVITZ
recommends the action it (v) ∈ {1, . . . , K} matching its label after projection onto 2−M , 1 , i.e.,
such that
! !
M
it (v)
1
1 X −m
2 cm , M , 1 .
= min max
+
(24)
2M
2
2
m=1
TF⋆
Each node v ∈
maintains a probability distribution pt (v, ·) ∈ ∆(K) over actions in K, which is
only used and updated when v is activated. The activated leaves v ∈ L correspond to unit masses
pt (v, i) = Ii=it (v) . As mentioned earlier, the internal nodes v ∈ TF⋆ \ L are of two types:
• Each internal node v = (σ1 , c1 , σ2 , c2 , . . . , σm , cm ) at depth m = 0, . . . , M − 1 is a “binning
node". When activated, it identifies its only child w = (v, σm+1 ) handling xt (i.e., such that
xt ∈ Xm+1 (σ1:(m+1) )) and outputs pt (v, ·) = pt (w, ·). We denote by Tbin ⊂ TF⋆ \ L the set
of “binning nodes".
• Each internal node v = (σ1 , c1 , σ2 , c2 , . . . , σm , cm , σm+1 ) at depth m + 1/2 is an “Exp4
node". It runs an instance of Exp4 using the children of v as experts. The variant of Exp4
we use is Algorithm 4 (see Appendix E) applied to penalized loss estimates (cf. (2)), which
adapts to the range of the losses and allows for a careful control of the variance terms along
the tree. Let Exp4v be the instance of the Exp4 variant run at node v. On all the rounds t
when v is activated, thisP
instance updates a distribution qt (v, ·) ∈ ∆(|Cv |) over experts in Cv
and outputs pt (v, i) = w∈Cv qt (v, w)pt (w, i) for all actions i = 1, . . . , K. We denote by
TExp4 ⊂ TF⋆ \ L the set of “Exp4 nodes".
Note that the root v0 of TF⋆ is activated at all rounds t. The prediction of HierExp4⋆ at time t is
ybt = It 2−M ∈ K, where It is drawn according to a mixture of pt (v0 , ·) and a unit mass on the
minimal action i = 1. A pseudo-code for HierExp4⋆ can be found in Algorithm 5.
G.2. Proof of Theorem 6
As for the proof of Theorem 4, we start by stating a lemma indicating that the losses associated with
neighboring leaves are close to one another.
Lemma 15 Let v ∈ TExp4 , with depth m + 1/2, m ∈ {0, . . . , M − 1}. Then, all leaves w, w′ ∈ Lv
in the subtree rooted at v satisfy
it (w)
it (w′ )
ℓt
6 21−m
− ℓt
2M
2M
at all rounds t > 1 when w and w′ are both activated.
′ , c′ ). First, we remark that
Proof Denote w = (σ1 , c1 , . . . , σM , cM ) and w′ = (σ1′ , c′1 , . . . , σM
M
since w and w′ have the common ancestor v at level m, we have σi = σi′ and ci = c′i for all i 6 m,
′
as well as σm+1 = σm+1
(recall that v ∈ TExp4 ). Thus,
M
X
j=1
−j
2
cj −
M
X
j=1
2−j c′j
6
M
X
j=m+1
−j
2 |cj −
30
c′j |
6
M
X
j=m+1
2−j+1 6 21−m .
A LGORITHMIC C HAINING AND THE ROLE OF PARTIAL F EEDBACK IN O NLINE N ONPARAMETRIC L EARNING
Algorithm 5: HierExp4⋆ (for the one-sided full information feedback)
Input
: Tree TF⋆ with root v0 and leaves L, exploration parameter γ ∈ (0, 1), penalization
parameters α0 , . . . , αM −1 > 0, learning rates η0 , . . . , ηM −1 > 0.
Initialization: Set q1 (v, ·) to the uniform distribution in ∆(|Cv |) for every v ∈ TExp4 .
for t = 1, 2, . . . do
1. Get context xt ∈ X ;
2. Activate v0 and all nodes v = σ1 , c1 , . . . , σm (, cm ) ∈ TF⋆ such that xt ∈ Xm (σ1:m );
3. Set pt (v, i) = Ii=it (v) for all i ∈ K and all activated leaves v ∈ L;
4. Set pt (v, i) = qt (v, ·) · pt (·, i) for all i ∈ K and all activated nodes v ∈ TExp4 ;
5. Set pt (v, ·) = pt (w, ·) for all activated nodes v ∈ Tbin , where w ∈ Cv is the unique activated
child of v;
6. Draw It ∼ p∗t and play ybt = It 2−M , where p∗t (i) = (1 − γ)pt (v0 , i) + γIi=1 for all i ∈ K;
7. Observe ℓt (y) for all y > ybt ;
8. For all m = 0, . . . , M − 1 and all activated nodes v ∈ TExp4 at level m + 1/2, with
v = (σ1 , c1 , σ2 , c2 , . . . , σm , cm , σm+1 ), compute the loss estimate for each i ∈ Kt (v),
ℓt (i/2M ) − ℓt jt (v)/2M + 21−m
αm
αm
b
,
(25)
IIt 6i − Pi
+
ℓt (v, i) =
Pi
∗
∗
γ
k=1 pt (k)
k=1 pt (k)
where Kt (v) = {i : (∃w ∈ Cv ) pt (w, i) > 0} and jt (v) = max Kt (v).
Then, for each w ∈ Cv , compute the expert loss ℓet (v, w) = pt (w, ·) · ℓbt (v, ·) and perform the
update
qt+1 (v, w) =
exp −ηm
X
w ′ ∈Cv
t
X
s=1
exp −ηm
end
ℓes (v, w)Ixs ∈Xm+1 (σ1:(m+1) )
t
X
s=1
!
ℓes (v, w )Ixs ∈Xm+1 (σ1:(m+1) )
′
!.
(26)
′
Therefore,
by definition (24) of it (w) and it (w ), and since ′projecting two real numbers onto
−M
2 , 1 can only reduce their distance, we get, when w and w are both activated,
M
M
j=1
j=1
it (w) it (w′ )
1 X −j
1 X −j ′
6
2
c
−
2 cj 6 21−m ,
−
+
−
j
2M
2M
2
2
which implies the result since ℓt is 1-Lipschitz.
We are now ready to prove Theorem 6 from the main text.
Proof (of Theorem 6) Let f ∈ F. By Lemma
14, we can fix fM : [0, 1]d → −1/2, 3/2 of the
form (20), with coefficients cm (σ1:m ) ∈ −1, 0, 1 , and such that kfM − f k∞ 6 2−M .
In the proof of Theorem 4, we rewrite the regret with respect to fM as the sum of M regrets
along the path joining the root of TF to the leaf corresponding to fM . Next, we proceed similarly
except that now fM corresponds to a (2d )-ary subtree of TF⋆ indexed by σ1:M ∈ {1, . . . , 2d }M . The
31
C ESA -B IANCHI G AILLARD G ENTILE G ERCHINOVITZ
2dM leaves σ1 ,c1 (σ1 ), σ2 , c2 (σ1:2 ), . . . , σM , cM (σ1:M ) are labeled with the values of fM (before
projecting onto 2−M , 1 ) over the cubes XM (σ1:M ).
We control the regret of HierExp4⋆ with respect to fM by starting from the root v0 of TF⋆ and
by progressively moving
down the tree. We write Pt−1 and Et−1 for conditioning on I1 , . . . , It−1 ,
Pk
∗
∗
and set Pt (k) , i=1 pt (i) = Pt−1 (It 6 k). In the sequel, we repeatedly use Corollary 12 with
parameters Em , 21−m , αm and γ to control the regret on each “Exp4 node". This corollary applies
to loss estimates penalized with pent (k) = −αm /Pt∗ (k) for “Exp4 nodes" at depth m + 1/2, with
m = 0, . . . , M − 1.
Since the root v0 is a “binning node", we have pt (v0 ,k) = pt (σ1 , k) at all rounds t such that
xt ∈ X1 (σ1 ). Therefore, writing ℓt (k) instead of ℓt k/2M ,
"
#
α0 + 4η0 E02
E
pt (v0 , k) ℓt (k) −
Pt∗ (k)
t=1 k=1
K
2
X
X
X
α0 + 4η0 E0
pt (σ1 , k) ℓt (k) −
=
E
Pt∗ (k)
t: xt ∈X1 (σ1 ) k=1
σ1 ∈{1,...,2d }
K
X
X
X
pt (σ1 , c1 ), k ℓt (k) −
6
min E
T X
K
X
σ1 ∈{1,...,2d }
6
X
σ1
E
c1 ∈{−1,0,1}
t: xt ∈X1 (σ1 ) k=1
ln 3 4η0 T (σ1 )α20
+
+
η0
γ2
X
K
X
t: xt ∈X1 (σ1 ) k=1
!
pt [σ1 , c1 (σ1 )], k ℓt (k) −
α0
Pt∗ (k)
(27)
2
d
α0
+ 2 ln 3 + 4η0 T α0 ,
Pt∗ (k)
η0
γ2
(28)
where (27) follows from Corollary 12 with parameters E0 = 2, α0 and γ, and where T (σ1 ) is
the number of rounds t such that xt ∈ X1 (σ1 ). Indeed, the probability distributions pt (σ1 , c1 ), ·
are mixtures of actions it (w) associated with activated leaves w ∈ Lσ1 , whose losses are at most
at distance 2 by Lemma 15 with m = 0. The last inequality (28) above is obtained by choosing
c1 = c1 (σ1 ).
So far we have showed how to move from depth m = 0 to depth 1. Repeating the same
argument at all subsequent depths m = 1, . . . , M − 1, and noting that αm−1 = αm + 4ηm 22−2m =
32
A LGORITHMIC C HAINING AND THE ROLE OF PARTIAL F EEDBACK IN O NLINE N ONPARAMETRIC L EARNING
2 , we get
αm + 4ηm Em
" T K
#
XX
α0 + 4η0 E02
E
pt (v0 , k) ℓt (k) −
Pt∗ (k)
t=1 k=1
K
X
X
X
α
M −1
6
E
pt [σ1 , c1 (σ1 ), . . . , σM , cM (σ1:M )], k ℓt (k) − ∗
P
t (k)
σ1:M
t: xt ∈X1 (σ1:M ) k=1
!
M
−1
X
2(m+1)d ln 3 4ηm T α2m
+
+
ηm
γ2
m=0
!
M
−1
(m+1)d ln 3
2
X
X
X
2
4η
T
α
m
m
ℓt it (v) +
6
+
,
(29)
2
η
γ
m
σ
m=0
t: xt ∈X1 (σ1:M )
1:M
where in the last inequality v denotes the leaf σ1 , c1 (σ1 ), . . . , σM , cM (σ1:M ) , whose probability
distribution pt (v, ·) is the unit mass on its action it (v).
We define feM (x) , min max fM (x), 2−M , 1 to be the projection of fM (x) onto 2−M , 1 ,
and, similarly, fe(x) , min max f (x), 2−M , 1 . Since projection can only reduce the distance
between two points, feM (x) − fe(x) 6 fM (x) − f (x) 6 kfM − f k∞ 6 2−M . But, noting that
fe(x) is 2−M -close to f (x) ∈ [0, 1], this yields feM (x)−f (x) 6 feM (x)− fe(x) + fe(x)−f (x) 6
21−M . Therefore, since ℓt is 1-Lipschitz, and using both (21)
and the definition (24) of it (v), we
get that, when the leaf v = σ1 , c1 (σ1 ), . . . , σM , cM (σ1:M ) is activated,
ℓt it (v) = ℓt feM (xt ) 6 ℓt f (xt ) + feM (xt ) − f (xt ) 6 ℓt f (xt ) + 21−M .
We plug
last bound into (29) and multiply the resulting inequality by 1 − γ. Then, using
the
Et−1 ℓt ybt = (1 − γ)pt (v0 , ·) · ℓt + γℓt (1), together with ℓt (y) ∈ [0, 1], we obtain
#
!
" T K
M
−1
X
XX
2(m+1)d ln 3 4ηm T α2m
α0 + 4η0 E02
RegT (F) 6
+
+ (1 − γ) E
pt (v0 , k)
ηm
γ2
Pt∗ (k)
m=0
t=1
k=1
1−M
+ γT + 2
T.
(30)
Now, we use that by definition p∗t (k) = (1 − γ)pt (v0 , k) + γIk=1 > (1 − γ)pt (v0 , k). Therefore,
similarly to the analysis of Exp3-RTB, we can control the variance term
(1 − γ)
K
K
K Z P ∗ (k)
X
X
X
t
p∗t (k)
Pt∗ (k) − Pt∗ (k − 1)
dx
=
1
+
=
1
+
∗
∗
∗
∗ (k)
∗
Pt (k)
Pt (k)
Pt (k)
P
k=1
k=2
k=2 Pt (k−1) t
Z 1
K Z P ∗ (k)
X
t
dx
1
e
dx
=1+
6 1 − ln Pt∗ (1) 6 1 + ln = ln ,
61+
γ
γ
Pt∗ (1) x
Pt∗ (k−1) x
K
X
pt (v0 , k)
k=1
6
k=2
where we used Pt∗ (1) = p∗t (1) > γ. Hence, substituting into the previous bound (30) and recalling
that E0 = 2, we get
!
M
−1
e
X
2(m+1)d ln 3 4ηm T α2m
+ γT + 21−M T . (31)
+
+
(α
+
16η
)T
ln
RegT (F) 6
0
0
ηm
γ2
γ
m=0
33
C ESA -B IANCHI G AILLARD G ENTILE G ERCHINOVITZ
Optimization of the parameters. In the sequel, we show that the prescribed choices of M , γ,
(αm ), and (ηm ) lead to the stated regret bound. Before that, we explain these particular choices
by approximately optimizing (31), which will lead to (32) and (33). For the sake of readability, we
2 =
will sometimes write f . g instead of f = O(g). First, recall that αm−1 = αm + 4ηm Em
αm + 24−2m ηm . We thus make the approximation αm ≈ 24−2m ηm , and start by optimizing in ηm
the terms inside the sum appearing in (31). To do so, we equalize the terms to get the equality
2(m+1)d ln 3
4ηm T α2m
=
ηm
γ2
which we approximate, using the previous remark, by
3
2md
T 2−4m ηm
≈
ηm
γ2
d
1
1
ηm = c2m( 4 +1) γ 2 T − 4 ,
hence our choice
(32)
where c > 0 will be optimized by the analysis. Now, from these values of ηm , we can compute αm
by choosing αM = 0 and using the recursion αm−1 = αm + 24−2m ηm . This yields
αm =
M
X
M
X
1
1
24−2j ηj = c24 γ 2 T − 4
d
2j( 4 −1) .
(33)
j=m+1
j=m+1
We now upper bound (31) by distinguishing between the two cases 1 6 d 6 4 and d > 4.
• If d 6 4, then the terms inside the sum in (33) are non-increasing, so that
1
1
d
αm 6 c24 M1 γ 2 T − 4 2m( 4 −1) ,
(34)
where M1 = M if 2 6 d 6 4 and M1 = 2 > 1/ 21−d/4 − 1 if d = 1. The sum in the
right-hand side of (31) then becomes, by substituting the definition of ηm (see (32)) and the
above upper bound on αm ,
! M −1
!
M
−1
X 2m( 34 d−1)+d T 14 ln 3 210 c3 M 2 2m( 3d4 −1) T 41
X
2(m+1)d ln 3 4ηm T α2m
1
+
6
+
1
1
2
η
γ
2
m
cγ
γ2
m=0
m=0
=
2d ln 3
c
(∗)
1
−1
1 1M
X
3
+ 210 M12 c3 T 4 γ − 2
2m( 4 d−1)
m=0
1
1
6 27 M12 T 4 γ − 2
6
(
M
−1
X
3
2m( 4 d−1)
m=0
211 T
1
4
γ
3
27 M 2 T
− 12
1
4
γ
− 21 M ( 34 d−1)
2
if d = 1
if 2 6 d 6 4
−1/2
where inequality (∗)Pis by using d 6 4 and choosing c = 2−5/4 M1
, while the last in−m/4 6 23 for d = 1. Plugging this inequality into (31), upper
equality is because ∞
2
m=0
bounding α0 using (34) and η0 using (32), and recalling that M = ⌈ln2 (1/γ)⌉, we get
– case d = 1: choosing γ = T −1/2 / ln(T ),
RegT (F) = O
34
√
T ln T .
(35)
A LGORITHMIC C HAINING AND THE ROLE OF PARTIAL F EEDBACK IN O NLINE N ONPARAMETRIC L EARNING
– case 2 6 d 6 4: choosing γ = T −1/(d+2/3) ,
3
1
1
1
3
3
RegT (F) . ln(1/γ) 2 T 4 γ − 2 γ −( 4 d−1) + ln(1/γ)2 γ 2 T 4 + γT
3d/4
d−1/3
3
d−1/3
3
d−1/3
. (ln T ) 2 T d+2/3 + (ln T )2 T d+2/3 + T d+2/3 . (ln T ) 2 T d+2/3 ,
(36)
where the last inequality is because 3d/4 < d − 1/3 for d > 2.
• If d > 5, then the terms inside the sum in (33) are exponentially increasing, so that
d
2 4 −1
1
1
d
1
d
1
c24
γ 2 T − 4 2M ( 4 −1) 6 c27 γ 2 T − 4 2M ( 4 −1) .
d
−1/4
−1
1−2
24 − 1
(37)
Then, following the lines of the case d 6 4, by substituting ηm (see (32)) and αm in the sum
in the right-hand side of (31)
!
M
−1
X
2(m+1)d ln 3 4ηm T α2m
+
ηm
γ2
m=0
!
3
d
1
1
d
M
−1
X
2m( 4 d−1)+d T 4 ln 3 216 c3 2m( 4 +1)+M ( 2 −2) T 4
6
+
1
1
cγ 2
γ2
m=0
!
1
M −1
M
−1
X
d
d
T 4 2d ln 3 X m( 3 d−1)
2 4
= 1
2m( 4 +1)
+ 216 c3 2M ( 2 −2)
c
2
γ
m=0
m=0
1
4+d
3d
3
T4 2
2M ( 4 d−1) + 216 c3 2M ( 4 −1)
6 1
c
γ2
1
3
T 4 24+d
16 3
= 1
+ 2 c 2M ( 4 d−1)
c
γ2
4
1
2
αm 6 c2 γ T
− 14 M ( d4 −1)
2
1
1
6
3
3
6 T 4 γ − 2 2M ( 4 d−1)+8+ 4 d
where in the last inequality we used c = 2d/4−3 . Plugging into the regret bound (31), upper
bounding α0 using (37) and η0 using (32), and recalling that M = ⌈ln2 (1/γ)⌉, we get
1
1
3
3
d
3
RegT (F) . T 4 γ 2 − 4 d + γ 2 − 4 T 4 ln(1/γ) + γT
d−1/3
d−1
d−1/3
d−1/3
. T d+2/3 + T d+2/3 ln(T ) + T d+2/3 . T d+2/3 , (38)
where the second inequality is by setting γ = T −1(d+2/3) .
Putting together the three cases (35), (36), and (38) concludes the proof of the regret bound.
Running time of HierExp4⋆ . We only address the case d > 2. First note that the total number of
“Exp4 nodes” in TF⋆ is
TExp4 =
=
M
−1
X
m=0
M
−1
X
m=0
n
o
v = (σ1 , c1 , σ2 , c2 , . . . , σm , cm , σm+1 ) : σk ∈ {1, . . . , 2d }, ck ∈ {−1, 0, 1}
3m 2d(m+1)
M +1
− 3 · 2d
1 3 · 2d
d M
=
6
3
·
2
.
3
3 · 2d − 1
35
C ESA -B IANCHI G AILLARD G ENTILE G ERCHINOVITZ
M
M
Similarly, Tbin 6 3 · 2d
and L 6 3 · 2d . Summing the three upper bounds, we get
M
TF⋆ 6 3 3 · 2d . Therefore, using M = ⌈ln2 (1/γ)⌉ and γ = T −1/(d+2/3) , we obtain
d+ln 3
2
ln2 T
⋆
so that
TF = O T d+2/3 .
M 61+
d + 2/3
1
The number of actions in K is 2M = O T d+2/3 . The running time at every round t, which is at
most proportional to the total number
of tuples (v, w, i) for (v, i) ∈ TF⋆ × K and w ∈ Cv , is at most
of the order of TF⋆ × K × max 2d , 3 . Therefore, the running time is at most of the order of
T
d+1+ln2 3
d+2/3
6 T 1.8 for all d > 2.
d+1+ln2 3
A tighter analysis however yields a smaller time complexity than the rate T d+2/3 derived
above. This is because —from Algorithm 5— all elementary computations only need to be computed on activated “Exp4 nodes” nodes, on their (finite number of) children, and on actions in K.
Fix a round t > 1. From a computation similar to the one above, but noting that all activated nodes
share the same coefficients σm for all depths m, we can see that the total number of activated “Exp4
nodes” at round t is at most of
M
−1
X
3M − 1
3m =
2
m=0
1+ln2 3
so that the running time per round is at most of the order of 3M 2M . T d+2/3 , which concludes
the proof for d > 2. Similarly, in the case d = 1, the choice of γ = T −1/2 /(ln T ) entails a running
1+ln2 3
√
time at most of the order of 3M 2M = O T ln T )1+ln2 3 = o T 1+2/3 , so that the stated
running time also holds true for d = 1.
Appendix H. Algorithm HierHedge and proof of Theorem 7
The proof goes along the same lines as the proof of Theorem 4. Let v be any internal node at depth
m = 0, . . . , M − 1. By construction of HierHedge, the distribution pt (v, ·) ∈ ∆(K) is computed
by Hedge and is supported on the set {it (w) : w ∈ Lv }. By Lemma 13, the losses of any pair
of actions in this set differ by at most 2−m+3 . Hence, we may apply (Cesa-Bianchi et al., 2007,
Theorem 5) with N = |Cv | 6 Nm+1 , E = 2−m+3 , and Vet 6 E 2 obtaining
#
" T
T
p
X
X
T ln Nm+1 + ln Nm+1 + 1 .
pt (w, ·) · ℓt 6 2−m+5
pt (v, ·) · ℓt −
max E
w∈Cv
t=1
t=1
As in the proof of Theorem 4, we sum the above along a path v0 → v1 → · · · → vM = w for any
leaf w ∈ L and get
#
" T
M
−1
T
p
X
X
X
2−m
ℓt (it (w)) 6 25
pt (v0 , ·) · ℓt − min
E
T ln Nm+1 + ln Nm+1 + 1 .
t=1
w∈L
m=0
t=1
36
A LGORITHMIC C HAINING AND THE ROLE OF PARTIAL F EEDBACK IN O NLINE N ONPARAMETRIC L EARNING
Algorithm 6: HierHedge (for the full information feedback)
Input
: Tree TF with root v0 and leaves L, learning rate sequences
η2 (v) > η3 (v) > · · · > 0 for v ∈ TF \ L.
Initialization: Set q1 (v, ·) to the uniform distribution in ∆(|Cv |) for every v ∈ TF \ L.
for t = 1, 2, . . . do
1. Get context xt ∈ X ;
2. Set pt (v, i) = Ii=it (v) for all i = 1, . . . , K and for all v ∈ L;
3. Set pt (v, i) = qt (v, ·) · pt (·, i) for all i = 1, . . . , K and for all v ∈ TF \ L;
4. Draw It ∼ pt (v0 , ·);
5. Observe ℓt (i) for all i = 1, . . . , K;
6. For each v ∈ TF \ L and for each w ∈ Cv compute the expert loss ℓet (v, w) = pt (w, ·) · ℓt
and perform the update
P
exp −ηt+1 (v) ts=1 ℓes (v, w)
qt+1 (v, w) = P
(39)
Pt e
′)
exp
−η
(v)
ℓ
(v,
w
′
t+1
s
s=1
w ∈Cv
end
Since L is a (2−M )-covering of F, 2−M ≤ 2ε, and Kε is a ε-covering of Y, we get
#
" T
M
−1
T
p
X
X
X
T ln Nm+1 + ln Nm+1 + 1 .
2−m
ℓt f (xt ) 6 5εT + 25
E
ℓt yIt − inf
t=1
f ∈F
m=0
t=1
Overapproximating the sums with integrals similarly to (18) and (19),
RegT (F) 6 5εT + 27
Z
p
2 T ln N∞ (F, x) + ln N∞ (F, x) dx .
1/2
ε/2
d
p
d
If F is a set of Lipschitz functions f : [0,
1] → [0, 1] , where [0, 1] is endowed with the norm
′
−d
kx − x k∞ , then ln N∞ (F, ε) = O pε
implying
√
O √pT + p ln(1/ε)
if d = 1
1−d
O pT ln(1/ε) + pε if d = 2
RegT (F) 6 5T ε +
√
O pT ε1−d/2 + pε1−d
if d > 3.
Optimizing in ε, we obtain the stated result.
37
| 10math.ST
|
Component Outage Estimation based on Support
Vector Machine
Rozhin Eskandarpour, Amin Khodaei
Department of Electrical and Computer Engineering
University of Denver
Denver, USA
[email protected], [email protected]
Abstract—Predicting power system component outages in
response to an imminent hurricane plays a major role in preevent planning and post-event recovery of the power system. An
exact prediction of components states, however, is a challenging
task and cannot be easily performed. In this paper, a Support
Vector Machine (SVM) based method is proposed to help
estimate the components states in response to anticipated path
and intensity of an imminent hurricane. Components states are
categorized into three classes of damaged, operational, and
uncertain. The damaged components along with the components
in uncertain class are then considered in multiple contingency
scenarios of a proposed Event-driven Security-Constrained Unit
Commitment (E-SCUC), which considers the simultaneous
outage of multiple components under an N-m-u reliability
criterion. Experimental results on the IEEE 118-bus test system
show the merits and the effectiveness of the proposed SVM
classifier and the E-SCUC model in improving power system
resilience in response to extreme events.
Keywords— Support vector machines, extreme events, power
system resilience, resource scheduling, security-constrained unit
commitment.
NOMENCLATURE
Indices and Sets:
b
Index for buses.
B
Set of components connected to bus b.
i
Index for generation units.
s
Index for scenarios.
S
Set of outage scenarios.
t
Index for time.
β
Index for training examples in SVM.
Parameters:
c
Penalty factor of soft margin SVM.
DRi
Ramp down rate of unit i.
DTi
Minimum down time of unit i.
Fi(.)
Generation cost function of unit i.
M
Large positive constant.
Pimin
Minimum generation capacity of unit i.
Pimax
Maximum generation capacity of unit i.
R
System reserve.
URi
Ramp up rate of unit i.
UTi
Minimum up time of unit i.
xl
Line reactance of line l.
This work has been supported in part by the U.S. National Science Foundation
under Grant CMMI-1434771.
Variables:
alb
Element of line l and bus b at line-bus incidence
matrix.
Dbt
Load at bus b at time t.
g
Bias of hyperplane from the origin
h(x)
Hyperplane separating training examples (the
output of the classifier).
Iit
Commitment state of unit i at time t.
LCbts
Load curtailment at bus b at time t in scenario s.
Pits
Real power generation of unit i at time t in
scenario s.
PLlts
Real power flow of line l at time t in scenario s.
on
Number of successive ON hours of unit i at time
Tit
t.
Number of successive OFF hours of unit i at
Titoff
time t.
UXits
Outage state of unit i at time t in scenario s; 0 if
on outage, otherwise 1.
UYlts
Outage state of line l at time t in scenario s; 0 if
on outage, otherwise 1.
vb
Value of lost load at bus b.
w
Normal vector to the hyperplane
xβ
Training example β.
yβ
Class labels of training example β.
Δi
Permissible power adjustment of unit i.
θbts
Phase angle of bus b at time t in scenario s.
I.
INTRODUCTION
Weather events are the number one cause of power system
outages in the United States resulting in millions of dollars
damage every year to critical infrastructure, businesses, and
emergency networks [1]. Among the different types of extreme
weather events, hurricanes are the most frequent in the United
States [2]. The power grid resilience in response to these
imminent hurricanes can be significantly improved by (a) an
efficient forecasting of the likely hurricane impacts on the grid
and to have an idea of which components will be damaged, and
(b) an optimal schedule of available generations in case of
several power component outages. These two factors can
potentially improve pre-event planning and post-event recovery
practices by preparing the system to address likely component
outages and further response to the outages once the hurricane
is passed.
A case study on hurricane planning and rebuilding the
electrical infrastructure along the Gulf Coast for hurricane
Katrina is presented in [3], where the grid resiliency and its
interdependency are estimated with data for Hurricane Katrina
regarding power delivery and telecommunications, collected
post-landfall. Post-disaster restoration is also an important part
of the power system resilience studies. In [5] and [6], a
stochastic pre-hurricane model is proposed which aims at
managing the available resources before the extreme event, as
well as a deterministic post-hurricane model to manage the
system after the extreme event through a proactive recovery. In
[6] an optimal restoration scheduling to minimize the load
curtailment in the case of an extreme event is proposed. The
proposed model is based on AC power flow constraints and
aims at grid restoration after extreme events by using lost load
data. In [7] an optimal repair schedule, system configuration,
and unit commitment model to restore the damaged grid
infrastructure through a post-disaster decision-making model is
proposed. A general informative case study of a variety of
models and algorithms regarding emergency response logistics
in distribution networks is provided in [8]. The study in [9]
predicts the outages in response to extreme events and
investigates several data-driven measurements, such as the total
number of component outages, potential effects on customers,
the type of component outages, and their geographic
distribution and duration. The study is based on large databases
of outages in five hurricanes in Carolina. In [10], to forecast
the power outage durations an ensemble learning method for
regression is provided which are applied for outages resulted
from several hurricanes such as Dennis, Katrina, and Ivan. In
[11], a prediction model to estimate hurricane-related power
outages is proposed and claimed to be viable along the U.S
coastline. This prediction model is applied to publicly available
data and used in estimating several historic storms, such as
hurricane Sandy and typhoon Haiyan.
The relatively considerable number of hurricanes in the
United States have resulted in a substantial amount of data.
Machine learning and data mining methods can therefore be
viable tools to analyze these data and accordingly enhance
power system planning and recovery in response to imminent
hurricanes. In this paper, a multi-class Support Vector Machine
(SVM) is proposed to classify power system components states
after the hurricane into three classes (i.e., operational, outage
and uncertain) based on the distance of each component from
the center of the hurricane and the category of the hurricane. In
the proposed problem three sets of data are available which
need to be classified into three groups using SVM, which is a
binary classifier. As a result, the classical SVM is extended
using one-versus-one multi-classification. During the
evaluation to predict the state of each component, each
component is evaluated by three classifiers where the majority
response of the classifier determines the state of the
component. The obtained component outages from the
proposed model are further integrated into a developed eventdriven security-constrained unit commitment (E-SCUC) model
to find the optimal schedule of available resources that
minimizes the operation cost while enhancing grid resilience.
Unlike the existing work on outage prediction and extended
outage consideration in SCUC, including the previous work of
authors in [12] and [13], this paper uses a multi-class SVM to
determine uncertain component states in addition to outage and
operational. The merit behind considering uncertain states is
that in practical applications there is limited information about
the impact of the hurricane on all components and therefore
considering uncertain states can increase the prediction
accuracy and would help with achieving a better scheduling
solution. To further consider the uncertain states, the
previously defined N-m reliability criterion is extended in this
paper to an N-m-u reliability criterion in which u represents the
number of uncertain components on outage.
The rest of the paper is organized as follows. Sections II
and III respectively present the model outline and the
formulation of the proposed outage estimation model and ESCUC. Section IV presents numerical simulations based on the
IEEE 118-bus test system. Section V concludes the paper.
II.
PROPOSED COMPONENT OUTAGE ESTIMATION MODEL
The proposed model has three stages. At first, an SVM
model is trained to classify the components into three classes
(damaged, operational, and uncertain). The model is trained on
multiple features for each element (wind speed, distance to the
center of the hurricane). Since historical data of the impact of
previous hurricanes on power system components are limited,
synthetic data are generated and used to train the model.
During scheduling for the hurricane, the category and the path
of the potential hurricane that is heading toward the power
system are forecasted, which however can be obtained from
weather forecasting agencies. According to the forecasted
hurricane, the state of each component is determined by the
trained SVM model. Once the state of each system component
is estimated, the E-SCUC problem is solved to provide an
optimal and resilient schedule of the resources in response to
the hurricane.
It is common to use the N-1 criterion for reliability studies
in power systems, where N in the total number of components
in the system. The N-1 criterion simply states that the system
needs to adequately and reliably supply loads in case of a
single component outage at any given time. Although
extremely useful, this criterion cannot be directly applied to the
system to ensure resilience, since more than one component in
response to a hurricane may be out of service. In [12], an
extended criterion, i.e., N-1-m is proposed which
simultaneously considers the power system security in
response to the single component outage, i.e., an N-1 criterion,
and also in response to the outage of m components in
impacted regions. In order to consider uncertainty in
component failure prediction, this paper employs a further
extended criterion, i.e., N-m-u, where m is the number of
components that are predicted to be on outage during the
extreme event, and u is the number of components in uncertain
class.
A. SVM Method for Outage Prediction
The state of a component in response to the hurricane is
considered to be either damaged (i.e., on outage), operational
(i.e., in service), or uncertain (i.e., there is possibility of outage
but it is completely certain). In order to classify the state of
each component, different features can be considered. In this
paper, the wind speed and the distance of the each component
from the center of the hurricane are investigated as two major
features. SVM [14] is applied to determine the decision
boundary between different classes and accordingly
determining the state of each component.
min
åå F (P
Figure 1. One-versus-all and one-versus-one multiclassification
An SVM can generally classify data into two classes by
determining the proper hyperplane that separates training
examples of one class from the other one. The best hyperplane
is considered as the hyperplane with a clear gap that is as wide
as possible. The optimal hyperplane parameters can be found
by the following optimization problem:
m
1 2
min w + c å e b
2
b =1
(
)
yb wT xb + g ³ 1,-e b ,
s.t.
e b ³ 0,
(1)
b = 1,......, m
where w is the normal vector to the hyperplane separating
training examples, |g|/||w|| is the perpendicular distance from
the hyperplane to the origin, and c is a penalty parameter
allowing separating nonlinear examples. This is a quadratic
programming problem which can be solved by Lagrange
duality. By solving the duality, the final hyperplane only
depends on the support vectors (i.e., sample points that are in
the margin) and SVM needs to find only the inner products
between the test samples and the support vectors.
The idea of the maximum-margin hyperplane is based on
the assumption that training data are linearly separable, which
is not the case in many practical applications. In order to apply
SVM to nonlinear data, kernel methods can be used [14]. The
idea of kernel method (or kernel trick) is to map input features
to a higher dimension that can be linearly separable and can fit
the maximum-margin hyperplane in the transformed feature
space. Kernel trick simply states that for all x1 and x2 in the
input space X, a certain function K(x1, x2) can be replaced by
the inner product of x1 and x2 in another space. For example, a
polynomial kernel of degree d can be defined as:
(2)
K x , x = x .x d
(a
b
i
) ( a b)
SVMs are inherently two-class classifiers. The traditional
way to do multiclass classification with SVMs is to use oneversus-rest classifiers (commonly referred to as one-versus-all)
and to choose the class which classifies the test data with the
greatest margin. Another strategy is to build a set of oneversus-one classifiers and to choose the class that is selected by
the most classifiers (majority voting). In one-versus-one
approach, if g is the number of classes, then g(g-1)/2 classifiers
are constructed, and each one trains data from two classes. Fig.
1 shows the difference between one-versus-all and one-versusone approaches in a three class problem.
B. E-SCUC Problem Formulation
The objective of the E-SCUC problem is defined as:
) + ååå vLC bts
t
s
(3)
b
where the first term is the operation cost in normal system
operation (which includes the generation cost and
startup/shutdown costs) and the second term is the cost of
unserved energy during contingency scenarios (which is
defined as the value of lost load times unserved energy). The
value of lost load is defined as the average cost that each type
of customer - residential, commercial, or industrial - is willing
to pay in order to avoid power supply interruptions [15].
Assuming UX and UY as outage states for generation units and
transmission lines, respectively, the proposed objective
function is subject to the following operational constraints:
å P + å PL
"b, "s, "t
(4)
"i, "s, "t
"i, "s, "t
(5)
(6)
Pi (t -1) s - Pits £ DRi
"i, "s, "t
(7)
Titon ³ UTi I it - I i (t -1)
"i, "t
(8)
Titoff
"i, "t
(9)
å Pimax I it ³ Dt + Rt
"i, "t
(10)
0 £ LCbts £ Dbt
"i, "s, "t
(11)
its
iÎ B
l ÎB
min
Pi I itUX its£
Pits - Pi (t -1) s
b = 1,......, m
it 0 , I it
i
t
lts
+LCbts = Dbt
Pits £ Pi max I itUX its
£ URi
(
)
³ DTi (I i (t -1) - I it )
i
Pit 0 - Pits £ D i
-
PLmax
l UY lts£
PLlts -
(12)
"l , "s, "t
(13)
£ M (1 - UYlts )
"l , "s, "t
(14)
PLlts £
åb albq bts
xl
"i, "s, "t
PLmax
l UX lts
Load balance equation (4) ensures that the total injected
power to each bus from generation units and line flows is equal
to the total load at that bus. Load curtailment variable (LCbts) is
further added to the load balance equation to ensure a feasible
solution when there is not sufficient generation to supply loads
(due to an outage of power system components). Load
curtailment will be zero under normal operation conditions.
Generation unit output power is limited by its capacity limit
and will be set to zero depending on its commitment and
outage states (5). Generation units are further subject to
prevailing technical constraints including ramp up and down
rate limits (6)-(7), and minimum up and down time limits (8)(9). The system operating reserve requirement is represented in
(10). Load curtailment at each bus is constrained by the total
load on that bus (11), and the change in a unit generation is
further constrained by the maximum permissible limit between
normal and contingency scenarios (12). Transmission line
capacity limits and power flow constraints are modeled by (13)
and (14), respectively, in which the outage state is included to
model the line outages in contingency scenarios. Further details
can be found in [12], [13].
The E-SCUC problem is formulated using mixed-integer
linear programming (MILP). CPLEX 12.6 [16] is used to solve
the MILP problem. CPLEX uses a branch and cut algorithm
which solves a series of LP sub-problems. In addition, multithreading can be applied to CPLEX solver which makes the
MILP faster and can automatically make adjustments when
memory is limited.
III.
NUMERICAL SIMULATIONS
A. SVM performance
As historical data for the past extreme hurricanes at
component level are limited, a set of synthetic data is generated
to train the SVM model as 300 samples in outage state, 300
samples in the operational state, and 150 samples in the
uncertain area. To define the synthetic data, Saffir-Simpson
Hurricane Scale [17] is used to generate wind speed features of
the synthetic data. In particular, the generated samples follow a
normal distribution function of one-minute sustained windspeed of different categories with a small Gaussian noise.
Initial studies on the generated synthetic data showed that
majority of the components in an anticipated hurricane of
category 4 & 5 will be on outage, the components in a
hurricane of category 1 will be operational, and the majority of
the components in a hurricane of categories 2 & 3 will belong
to the uncertain area, which verifies that the generated data
conform to realistic conditions. The features are normalized to
[0, 1] based on the maximum considered values of wind speed
and distance. Fig. 2 shows the generated synthetic data.
performance, and the average accuracy of classification is
reported in Table I.
Figure 3. Part of the IEEE 118-bus test system and the forecasted hurricane
passing through a hypothetical path
As it is shown, SVM with linear kernel and c=1 offers the
best performance among other settings. The average overall
classification accuracy of the proposed classification model is
85.3%. The relatively small variance (~2%) in the accuracy of
the SVM under various hyper-parameters indicates that the
proposed method is not sensitive to its hyper-parameters and is
not over-fitted to the training data. Fig. 2 shows the decision
boundary of SVM with linear kernel in a one-versus-one multiclassification approach. During the evaluation, to predict the
state of each component, each component is evaluated by the
three classifiers shown in Fig. 2, and the majority response of
the classifier determines the state of the component. Table II
shows the confusion matrix of classifying operational and
outage components into three states. As it is demonstrated, the
proposed method can classify the outage components from the
operational components with a high accuracy.
TABLE I.
A cross-fold validation is a common approach to evaluate a
classifier on a set of data. The cross-fold validation splits the
data into q subsets, trained on (q-1) subsets, and evaluated on
the subset that is left in the training. This process is performed
q times (such that the classifier is evaluated on all samples).
The final classification accuracy is the average of classification
accuracies on all folds. In this paper, five folds are defined on
the generated set of data (q=5). Since our dataset is relatively
small, an off-the-shelf SVM model implemented in LibSVM
[18] is used in this paper. There are other SVM solvers and
techniques such as Liblinear [19] which is optimized for largescale linear classification and regression.
As the number of attributes (features) in the studied case is
smaller than the number of observations (samples), the
probability of overfitting to the training data is low. A series of
hyper-parameter “c” with various well-known kernels are
examined to evaluate its performance and show the sensitivity
of the employed SVM to its parameters. In each setting, a
cross-fold validation with 5 folds is used to evaluate the
ACCURACY OF SVM WITH VARIOUS HYPER-PARAMETERS
“C” AND KERNELS USING 5-FOLD CROSS-VALIDATION
Kernel
Linear
Quadratic
Cubic
Gaussian
c=0.1
84.5
84.7
84.3
81.6
c=1
85.3
84.7
82.9
83.6
c=10
85.1
84.5
78.0
82.1
c=100
85.01
78.0
71.0
81.1
TABLE II.
CONFUSION MATRIX OF CLASSIFYING SYSTEM COMPONENTS
DURING EXTREME EVENT (NUMBER OF SAMPLES AND PERCENTAGE)
Actual
Figure 2. Generated samples for three classes (operational, damaged, and
uncertain), and separating hyper-planes of Linear SVM with c=1 in a oneversus-one multi-classification.
Operational
Predicted
Uncertain
Outage
Operational
274 (91.3%)
25 (8.4%)
1 (0.3%)
Outage
1 (0.3%)
26 (8.7%)
273 (91.0%)
B. E-SCUC with Extended Outage
The proposed E-SCUC problem is applied to the standard
IEEE 118-bus test system [20]. A hurricane is assumed to pass
through a hypothetical path as shown in Fig. 3. The
components in the path and its neighboring areas are classified
into three categories of operational, outage, and uncertain,
according to the wind speed and the distance to the center of
the hurricane. Five components were classified in the outage
class (lines 1, 17, 18, 20, and 30) and six components were
categorized in the uncertain class (lines 7, 8, 13, 16, 22, and
generation unit 1). Table III shows the list of components
examined in this case study, the estimated distance from the
center of hurricane and the wind speed. Three cases are studied
to evaluate the role of considering uncertain area in power
system scheduling:
Case 1: In this case, only the components in the outage class
are considered as possible contingencies and the components in
the uncertain area are considered as operational. The operation
cost is obtained as $1,041,040. No load curtailment has
occurred in this case, so the cost of unserved energy is zero,
and the system is secure against this eight component outage.
Experimental results showed the effectiveness of the proposed
method in separating three classes from each other based on
wind speed and the distance from the center of the hurricane.
The damaged components along with the components in the
uncertain class were then considered in multiple contingency
scenarios of a proposed E-SCUC, which considered the
simultaneous outage of multiple components under an N-m-u
reliability criterion. Experimental results on the IEEE 118-bus
test system indicated that considering uncertain components in
different scenarios could potentially lead to a more resilient
schedule at the expense of marginal increase in operation cost.
Case 2: In this case, the components in the uncertain class are
considered as outages. The operation cost is obtained as
$1,085,516. A total of 1.53 MWh load curtailment has
occurred in this case.
[1]
TABLE III.
CLASSIFCATION OF COMPONENTS IN THE STUDIED CASE
Component
Line 1
Line 17
Line 18
Line 20
Line 30
Line 7
Line 8
Line 13
Line 16
Line 22
Unit 1
Line 60
Line 48
Line 47
Line 44
Line 57
Line 45
Unit 15
Distance
(normalized)
0.632
0.120
0.423
0.351
0.235
0.650
0.361
0.532
0.463
0.342
0.432
0.640
0.497
0.401
0.302
0.252
0.740
0.625
Wind speed
(normalized)
0.838
0.762
0.704
0.656
0.563
0.632
0.543
0.562
0.501
0.423
0.394
0.437
0.395
0.254
0.202
0.184
0.156
0.097
Class
(prediction)
Outage
Outage
Outage
Outage
Outage
Uncertain
Uncertain
Uncertain
Uncertain
Uncertain
Uncertain
Operational
Operational
Operational
Operational
Operational
Operational
Operational
Case 3: In this case, the N-m-u reliability criterion is
considered, where 192 contingency scenarios are defined in the
system. Considering m=5, in each scenario five components in
the outage class and one from the uncertain class are
considered to be out of service. The operation cost is obtained
as $1,087,022. The average load curtailment is calculated as
0.695 MWh. The obtained results advocate that considering
uncertainty can offer a more resilient commitment against
multiple component outages. This can be seen as the amount of
load curtailment is considerably lower than considering all
possible failures as an outage in same contingency scenario.
This significant advantage is obtained at the expense of limited
cost increase, which is less than 0.15%. The proposed approach
to classifying components into three categories can therefore
result in more resilient schedules.
IV.
CONCLUSION
Prediction of a component state in response to an extreme
event is challenging in practice. In this paper, a multiclass
SVM was proposed to categorize system components into three
classes (damaged, operational, and uncertain) in response to an
imminent hurricane. An artificial set of data was generated to
train the SVM, as the publicly available data on the impact of
hurricanes on power system components is limited.
REFERENCES
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
Executive Office of the President, “Economic Benefits of Increasing
Electric Grid Resilience to Weather Outages-August 2013.”
R. E. Brown, “Cost-benefit analysis of the deployment of utility
infrastructure upgrades and storm hardening programs,” Quanta
Technol. Raleigh, 2009.
B. Ball, “Rebuilding electrical infrastructure along the Gulf Coast: A
case study,” BRIDGE-Wash.-Natl. Acad. Eng.-, vol. 36, no. 1, p. 21,
2006.
A. Arab, A. Khodaei, Z. Han, and S. K. Khator, “Proactive Recovery of
Electric Power Assets for Resiliency Enhancement,” Access IEEE, vol.
3, pp. 99–109, 2015.
A. Arab, A. Khodaei, S. K. Khator, K. Ding, and Z. Han, “Posthurricane transmission network outage management,” in Proc. IEEE
Great Lakes Symp. Smart Grid New Energy Econ, 2013, pp. 1–6.
A. Arab, A. Khodaei, S. K. Khator, and Z. Han, “Transmission network
restoration considering AC power flow constraints,” in 2015 IEEE
International Conference on Smart Grid Communications
(SmartGridComm), 2015, pp. 816–821.
A. Arab, A. Khodaei, S. Khator, and Z. Han, “Electric Power Grid
Restoration Considering Disaster Economics,” IEEE Access, vol. 4, pp.
639–649, 2016.
N. Perrier et al., “A survey of models and algorithms for emergency
response logistics in electric distribution systems. Part II: Contingency
planning level,” Comput. Oper. Res., vol. 40, no. 7, pp. 1907–1922,
2013.
R. A. Davidson, H. Liu, I. K. Sarpong, P. Sparks, and D. V. Rosowsky,
“Electric power distribution system performance in Carolina
hurricanes,” Nat. Hazards Rev., vol. 4, no. 1, pp. 36–45, 2003.
R. Nateghi, S. D. Guikema, and S. M. Quiring, “Forecasting hurricaneinduced power outage durations,” in Natural Hazards, 2014, vol. 74, pp.
1795–1811.
S. D. Guikema, R. Nateghi, S. M. Quiring, A. Staid, A. C. Reilly, and
M.-L. Gao, “Predicting hurricane power outages to support storm
response planning,” Access IEEE, vol. 2, pp. 1364–1373, 2014.
R. Eskandarpour, A. Khodaei, and J. Lin, “Event-Driven SecurityConstrained Unit Commitment,” in Innovative Smart Grid Technologies,
2016.
R. Eskandarpour, A. Khodaei, and J. Lin, “Event-Driven SecurityConstrained Unit Commitment with Component Outage Estimation
Based on Machine Learning Method,” in North American Power
Symposium, 2016.
C. Cortes and V. Vapnik, “Support-vector networks,” Mach. Learn., vol.
20, no. 3, pp. 273–297, 1995.
L. Economics, “Estimating the Value of Lost Load,” Brief. Pap. Prep.
ERCOT 17 June 2013, 2013.
“CPLEX
12.”
[Online].
Available:
http://www.gams.com/latest/docs/solvers/cplex/index.html.
T. Schott et al., “The Saffir-Simpson hurricane wind scale,” Natl.
Weather Serv., 2012.
C.-C. Chang and C.-J. Lin, “LIBSVM: a library for support vector
machines,” ACM Trans. Intell. Syst. Technol. TIST, vol. 2, no. 3, p. 27,
2011.
R.-E. Fan, K.-W. Chang, C.-J. Hsieh, X.-R. Wang, and C.-J. Lin,
“LIBLINEAR: A library for large linear classification,” J. Mach. Learn.
Res., vol. 9, no. Aug, pp. 1871–1874, 2008.
“Power Systems Test Case Archive - UWEE.” [Online]. Available:
https://www2.ee.washington.edu/research/pstca/.
| 3cs.SY
|
On the Distribution, Model Selection Properties and
Uniqueness of the Lasso Estimator in Low and High
Dimensions
Ulrike Schneider and Karl Ewald
arXiv:1708.09608v2 [math.ST] 29 Jan 2018
Vienna University of Technology
We derive expressions for the finite-sample distribution of the Lasso estimator in the context of
a linear regression model with normally distributed errors in low as well as in high dimensions by
exploiting the structure of the optimization problem defining the estimator. In low dimensions we
assume full rank of the regressor matrix and present expressions for the cumulative distribution
function as well as the densities of the absolutely continuous parts of the estimator. Additionally,
we establish an explicit formula for the correspondence between the Lasso and the least-squares
estimator. We derive analogous results for the distribution in less explicit form in high dimensions
where we make no assumptions on the regressor matrix at all. In this setting, we also investigate
the model selection properties of the Lasso and show that possibly only a subset of models might
be selected by the estimator, completely independently of the observed response vector. Finally,
we present a condition for uniqueness of the estimator that is necessary as well as sufficient.
1
Introduction
The distribution of the Lasso estimator (Tibshirani, 1996) has been an object of study in the
statistics literature for a number of years. The often cited paper by Knight & Fu (2000) gives the
asymptotic distribution of the Lasso in the framework of conservative model selection in a lowdimensional (fixed-p) framework by listing the limit of the corresponding stochastic optimization.
Pötscher & Leeb (2009) derive explicit expressions of the distribution in finite samples as well
as asymptotically for all large-sample regimes of the tuning parameter (“conservative” as well as
“consistent model selection”) in the framework of orthogonal regressors. More recently, Zhou (2014)
gives high-level information on the finite-sample distribution for arbitrary designs in low and high
dimensions, geared towards setting up a Monte-Carlo approach to infer about the distribution.
In Ewald & Schneider (2015), the large-sample distribution of the Lasso is derived in a lowdimensional framework for the large-sample regime of the tuning parameter not considered in
Knight & Fu (2000). Moreover, Jagannath & Upadhye (2016) consider the characteristic function
of the Lasso to obtain approximate expressions for the marginal distribution of one-dimensional
components of the Lasso when these components are “large”, therefore not having to consider the
atomic part of the estimator.
In this paper, we exactly and completely characterize the distribution of the Lasso estimator
in finite samples in the context of a linear regression model with normal errors. In low dimensions,
we give formulae for the cumulative distribution function (cdf), as well as the density functions
conditional on which components of the estimator are non-zero. We do so assuming full column
rank of the regressor matrix. We also exactly quantify the correspondence between the Lasso and
least-squares (LS) estimator. In a high-dimensional setting, we make absolutely no assumptions
on the regressor matrix. We give formulae for the probability of the Lasso estimator falling into a
given set and exactly quantify the relationship between the Lasso estimator and the data object
X 0 y. Through this relationship, we also learn that the Lasso may never select certain models,
this property depending only on the regressor matrix and the penalization weights and being
independent of the observed response vector. In fact, we can characterize a so-called structural set
1
that contains all covariates that are part of a Lasso model for some response vector. This structural
set can be identified by how the row space of the regressor matrix intersects the cube at the origin
whose side lengths are determined by the penalization weights. The set may not contain all indices
in which case the Lasso estimator will rule out certain covariates for all possible observations of the
dependent variable. Finally, we present a condition for uniqueness of the Lasso estimator that is
both necessary and sufficient, again related to how the row space of the regressor matrix intersects
the above mentioned cube. All our results are based on properties of the optimization problem
defining the estimator and do not hinge on the assumption of Gaussian errors.
The paper is organized as follows. We introduce setting and notation in Section 2. The lowdimension case is treated in Section 3 whereas we consider the high-dimensional case in Section 4.
We conclude in Section 5.
2
Setting and Notation
Consider the linear model
y = Xβ + ε,
(1)
where y is the observed n × 1 data vector, X the n × p regressor matrix which is assumed to be
non-stochastic, β ∈ Rp is the true parameter vector and ε the unobserved error term with independent and identically distributed components that follow a N (0, σ 2 )-distribution. We consider
the weighted Lasso estimator β̂L , defined as the unique solution to the minimization problem
minp L(β) = minp ky − Xβk2 + 2
β∈R
β∈R
p
X
λn,j |βj |,
(2)
j=1
where λn,j , are non-negative user-specified tuning parameters that will typically depend on n. To
ease notation, however, we shall suppress this dependence for the most part and write λn,j = λj
for each j. Note that if λj = 0 for all j, the weighted Lasso estimator is equal to the ordinary
LS estimator β̂LS and that λ1 = · · · = λp > 0 corresponds to the classical Lasso estimator as
proposed by Tibshirani (1996), to which case we also refer to by uniform tuning. For later use, let
λ = (λ1 , . . . , λp )0 and define M0 = {j : λj = 0}, the index set of all unpenalized coefficients. If
M0 6= ∅, we speak of partial tuning. We stress dependence on the unknown parameter β when it
occurs, but do not specify dependence on X, y or λ as these quantities are available to the user.
The following notation will be used throughout the paper. Let ej denote the j-th unit vector
in Rp and let φ(µ,Σ) denote the Lebesgue-density of a normally distributed random variable with
mean µ and covariance matrix Σ and Φ be the cdf of a univariate standard normal distribution.
For a vector m ∈ Rp and an index set I ⊆ {1, . . . , p}, the vector mI ∈ R|I| contains only the
components of m corresponding to the elements of I and we write |I| for the cardinality of I and
I c for {1, . . . , p} \ I, the complement of I. The 1-norm of m is denoted by kmk1 whereas the 2norm is simply denoted by kmk. For x ∈ R, let sgn(x) = 1{x>0} − 1{x<0} where 1 is the indicator
function. For a set A ⊆ Rp , the set m + A = A + m is defined as {m + z :Q
z ∈ A} with a analogous
definitions for A − m and m − A. We denote the Cartesian product by
and the column space
and rank of a matrix C by col(C) and rk(C), respectively. The columns of C are denoted by Cj
whereas CI , for some index set I, is the matrix containing the |I| columns of C corresponding to
the indices in I only. We use R>0 for the positive and R≥0 for the non-negative real numbers.
Let {D− , D+ , D0 } be a partition of {1, . . . , p} into three sets, some of which may be empty. It
will be convenient to also describe this partition by a vector d ∈ {−1, 0, 1}p with dj = 1{j∈D+ } −
1{j∈D− } . For such d, we denote by Od = {z ∈ Rp : sgn(zj ) = dj for j = 1, . . . , p} = {z ∈ Rp : zj <
0 for j ∈ D− , zj > 0 for j ∈ D+ , zj = 0 for j ∈ D0 }. Note that m + β ∈ Od is short-hand notation
+
for mj < −βj for j ∈ D− , mj > −βj for j ∈ D+ and mj = −βj for j ∈ D0 . We write D−
for
D− ∪ D+ .
Finally, the directional derivative of a function g : Rp → R at m in direction r ∈ Rp with
krk = 1 is defined as
g(m + hr) − g(m)
∂g(m)
= lim
.
h&0
∂r
h
2
3
The Low-dimensional Case
Throughout this section, we assume that X has full column rank p, implying that we are considering
the low-dimensional setting where p ≤ n. For our arguments, we use a reparametrized version of
the objective function. Define
Vβ (u) = L(u + β) − L(β) = u0 X 0 Xu − 2u0 W + 2
p
X
λj [|uj + βj | − |βj |] ,
(3)
j=1
where W = X 0 ε ∼ N (0, σ 2 X 0 X) and note that Vβ is minimized at û = β̂L − β. We are interested
in the distribution of the estimation error û. To state the main theorem, we need the following
lemma that characterizes a solution to the minimization problem.
Lemma 1. Let m ∈ Rp . The following two statements are equivalent.
(a)
∂Vβ (m)
∂r
≥ 0 ∀r with krk = 1
(b)
∂Vβ (m)
∂ej
≥ 0 and
∂Vβ (m)
∂(−ej )
≥ 0 for j = 1, . . . , p.
Proof. Only (b) ⇒ (a) needs to be proved. Let d ∈ {−1, 0, 1}p such that m + β ∈ Od and let
{D− , D+ , D0 } be the corresponding partition of {1, . . . , p}. A straight-forward calculation shows
that
p
X
∂Vβ (m)
= 2r0 X 0 Xm − 2r0 W + 2
λj −1{j∈D− } rj + 1{j∈D+ } rj + 1{j∈D0 } |rj |
∂r
j=1
=
p
X
1{rj ≥0} (2X 0 Xm − 2W )j + 2λj (−1{j∈D− } + 1{j∈D+ ∪D0 } ) rj
j=1
+ 1{rj <0} −(2X 0 Xm − 2W )j + 2λj (1{j∈D− ∪D0 } − 1{j∈D+ } ) (−rj )
=
p
X
p
1{rj ≥0}
j=1
X
∂Vβ (m)
∂Vβ (m)
(−rj ) ≥ 0.
rj +
1{rj <0}
∂ej
∂(−ej )
j=1
This complete characterization of minima of Vβ now allows to state the following theorem on
the distribution of the estimation error û = β̂L − β.
Theorem 2. Let z ∈ Rp . Let d = sgn(z + β) ∈ {−1, 0, 1}p and let {D− , D+ , D0 } be the corresponding partition of {1, . . . , p}. Then
P (ûj ≤ zj for j ∈ D− , ûj ≥ zj for j ∈ D+ , ûj = zj for j ∈ D0 )
Z
=
Z
Z
Z Z
Z
···
···
···
sj ∈[−λj ,λj ]
j∈D0
mj ≥zj
j∈D+
mj ≤zj
j∈D−
φ(0,σ2 X 0 X) (X 0 Xmβ + sλ ) dmD− dmD+ dsD0 ,
where mβ and sλ ∈ Rp are given by (mβ )D+ = mD+ , (mβ )D0 = −βD0 and (sλ )D− = −λD− , sD+ =
−
−
λD+ , (sλ )D0 = sD0 , respectively.
∂V (m)
Proof. Since the function Vβ is convex, m ∈ Rp is a minimizer of Vβ if and only if β∂r ≥ 0 for
all r ∈ R with krk = 1. We wish to find all minimizers m satisfying mj ≤ zj for j ∈ D− , mj ≥ zj
for j ∈ D+ and mj = zj for j ∈ D0 . Note that this implies that m + β ∈ Od since z + β ∈ O by
∂Vβ (m)
∂Vβ (m)
assumption. By Lemma 1 together with the fact that the condition ∂e
≥ 0 and ∂(−e
≥0
j
j)
3
∂V (m)
β
= 0 if Vβ is differentiable at m with respect to the j-th component, we get the
reduces to ∂u
j
following necessary and sufficient conditions for such m to be a minimizer of Vβ .
0
for j ∈ D−
Wj = (X Xm)j − λj
0
(4)
Wj = (X Xm)j + λj
for j ∈ D+
0
0
(X Xm)j − λj ≤ Wj ≤ (X Xm)j + λj for j ∈ D0
Therefore, m satisfying m + β ∈ Od is a minimizer of Vβ if and only if W lies in the set
{s ∈ Rp : sj = (X 0 Xm)j − λj for j ∈ D− , sj = (X 0 Xm)j + λj for j ∈ D+ ,
(X 0 Xm)j − λj ≤ sj ≤ (X 0 Xm)j + λj for j ∈ D0 },
which equals
X 0 Xm + {sλ : (sλ )D− = −λD− , (sλ )D+ = λD+ , |sλ,j | ≤ λj for j ∈ D0 }.
Since we are interested in all minimizers m of Vβ that satisfy mj ≤ zj for j ∈ D− , mj ≥ zj for
j ∈ D+ and mj = zj for j ∈ D0 (that is, m − z ∈ Od ), we let
Ad = {X 0 Xm : m − z ∈ Od } + {s ∈ Rp : sD− = −λD− , sD+ = λD+ , |sj | ≤ λj for j ∈ D0 }.
As W follows a N (0, σ 2 X 0 X)-distribution, the sought-after probability is clearly given by
Z
φ(0,σ2 X 0 X) (u) du,
Ad
which is what was claimed.
Results on the distribution of β̂L itself rather than the estimation error are of course a direct
consequence of Theorem 2 and summarized in the following corollaries, the latter one giving the
probability of the extreme event of the Lasso setting all components to zero.
Corollary 3. Let z ∈ Rp and let d = sgn(z) with {D− , D+ , D0 } being the corresponding partition
of {1, . . . , p}.
P (β̂L,j ≤ zj for j ∈ D− , β̂L,j ≥ zj for j ∈ D+ , β̂L,j = 0 for j ∈ D0 )
Z
=
Z
···
Z
Z
···
Z
Z
···
φ(0,σ2 X 0 X) (X 0 Xmβ + sλ ) dmD− dmD+ dsD0 ,
sj ∈[−λj ,λj ] mj ≥zj −βj mj ≤zj −βj
j∈D+
j∈D−
j∈D0
where mβ and sλ ∈ Rp are given by (mβ )D+ = mD+ , (mβ )D0 = −β and (sλ )D− = −λD− , sD+ =
−
−
λD+ , (sλ )D0 = sD0 , respectively.
Corollary 4.
Z
λp
Z
λ1
···
P (β̂L = 0) =
−λp
φ(X 0 Xβ,σ2 X 0 X) (s) ds
−λ1
Remark 1. To illustrate the structure behind the proof of Theorem 2, we rewrite Corollary 3 as
P (β̂ ∈ Bz ) = P (W ∈ Aβ (Bz ))
with Bz = {b ∈ Rp : bj ≤ Szj for j ∈ D− , bj ≥Qzj for j ∈ D+ , bj = 0 for j ∈ D0 }, W ∼
p
N (0, σ 2 X 0 X), and Aβ (Bz ) = b∈Bz X 0 X(b − β) + j=1 Bj (bj ) where
(
{sgn(bj )λj } bj 6= 0
Bj (bj ) =
[−λj , λj ]
bj = 0.
4
The events {β̂ ∈ Bz } and {W ∈ Aβ (Bz )} are shown to be equivalent through Lemma 1. This
equivalence holds due to the structure of the optimization problem defining β̂L and does not depend
on the distribution of W = X 0 ε. In this sense, the distributional results do not hinge on the
normality assumption of the errors and can easily be generalized to other error distributions. The
relationship and shape of the sets Bz and Aβ (Bz ) are illustrated in Figure 1. Note that Aβ depends
on λ whereas Bz does not.
We exploit the structure of the optimization problem by characterizing the minimum through
directional derivatives rather than using the Kuhn-Karush Tucker conditions as in the high-level
approach of Zhou (2014). This has the advantage that our distributional results give the joint
distribution of the estimator only and are not “augmented” by the subdifferential vector.
Theorem 2 now puts us into a position to fully specify the distribution of the Lasso estimator. In
case λj > 0 for all j = 1, . . . , p, one easily sees from the preceding corollary that this distribution is
not absolutely continuous with respect to the p-dimensional Lebesgue-measure and thus no density
exists. One can, however, represent the distribution through Lebesgue-densities after conditioning
on which components of the estimator are negative, positive, and equal to zero – which we shall
do in the sequel.
Proposition 5. The distribution of û = β̂L − β, conditional on the event {β̂L ∈ Od }, can be
represented by a kdk1 -dimensional Lebesgue-density given by
f d (zD+ ) =
1{zβ + β ∈ Od }
P (β̂L ∈ Od )
−
Z
Z
···
φ(0,σ2 X 0 X) (X 0 Xzβ + sλ ) dsD0 ,
sj ∈[−λj ,λj ]
j∈D0
where zβ is defined by (zβ )D+ = zD+ and (zβ )D0 = −βD0 , and sλ is defined by (sλ )D− = −λD− ,
−
−
(sλ )D+ = λD+ , and (sλ )D0 = sD0 . Note that the constants P (β̂L ∈ Od ) can be calculated using
Corollary 3.
Proof. Observe that
d
f (zD+ ) =
−
∂
∂zj
+
j∈D−
+
P ûj ≤ zj for j ∈ D−
|β̂L ∈ Od ,
and note that by Theorem 2 for any z ∈ Rp we have with z + β ∈ Od we have
P ûj ≤ zj for j ∈ D− , ûj ≥ zj for j ∈ D+ |β̂L ∈ Od
1
P ûj ≤ zj for j ∈ D− , ûj ≥ zj for j ∈ D+ , β̂L,j = 0 for j ∈ D0
=
P (β̂L ∈ Od )
Z
Z Z
Z Z
Z
1
···
···
· · · φ(0,σ2 X 0 X) (X 0 Xmβ + sλ ) dmD− dmD+ dsD0 ,
=
P (β̂L ∈ Od )
sj ∈[−λj ,λj ]
j∈D0
mj ≤zj
j∈D+
mj ≥zj
j∈D−
where mβ ∈ Rp is defined by (mβ )D+ = mD− ∪D+ , and (mβ )D0 = −βD0 and sλ ∈ Rp is defined by
−
+
(sλ )D− = −λD− , (sλ )D+ = λD+ , and (sλ )D0 = sD0 . Differentiating with respect to zj : j ∈ D−
and taking the absolute value gives the density, thus completing the proof.
Besides the conditional densities, we can also specify the full cdf of û = β̂L − β which is done
in the following theorem.
Theorem 6. The cdf of û = β̂L − β is given by
X
F (z) = P (û1 ≤ z1 , . . . , ûp ≤ zp ) =
d∈{−1,0,1}p
5
Z
Z
···
mj ≤zj
+
j∈D
−
hd (mD+ ) dνkdk1 ,
−
(a) z1 , z2 > 0
(b) z1 , z2 < 0
(c) z1 = 0 and z2 > 0
Figure 1: The sets Bz are displayed on the left-hand side, the corresponding sets Aβ (Bz ) are
displayed on the right-hand side. Illustrated for p = 2 and various values of z, see Remark 1 for
details.
6
where νk denotes k-dimensional Lebesgue-measure and where
Z
Z
hd (mD+ ) = 1{mβ + β ∈ Od }
···
φ(0,σ2 X 0 X) (X 0 Xmβ + sλ ) dsD0 ,
−
sj ∈[−λj ,λj ]
j∈D0
with mβ ∈ Rp given by (mβ )D+ = mD− ∪D+ , and (mβ )D0 = −βD0 and sλ ∈ Rp given by (sλ )D− =
−
−λD− , (sλ )D+ = λD+ , and (sλ )D0 = sD0 .
Proof. It is easily seen that
P (û1 ≤ z1 , . . . , ûp ≤ zp ) =
X
P (β̂L ∈ Od )
d∈{−1,0,1}p
Z
Z
···
f d (mD+ ) dνkdk1 .
−
mj ≤zj
+
j∈D
−
Plugging in the formula for f d completes the proof.
For illustration of Proposition 5 and Theorem 6, consider Figures 2 and 3 which display an
example of the distribution of û = β̂L − β. One can see that the Lasso estimation error follows a
shifted normal distribution conditional on the event ûj 6= −βj (β̂L,j 6= 0) for each j = 1, . . . , p with
the shift depending on the signs of β̂L as is to be seen in Figure 2. Figure 3 displays the mass which
lies on the set {z ∈ R2 : z1 = −β1 , z2 6= 0}, that is, the density functions h(0,1) and h(0,−1) on
their corresponding domains. The mass on the set {z ∈ R2 : z1 6= 0, z2 = −β2 } looks qualitatively
similar to Figure 3. Note that we also have point-mass at −β, as is pointed out by Corollary 4.
Figure 2: The contour lines of the absolutely continuous part with respect to 2-dimensional
1 0.5 ), λ = (0.75, 0.75)0 and β =
Lebesgue-measure of the distribution of β̂L − β for X 0 X = ( 0.5
1
0
(0, −0.25) . Note that the blue lines as well as the blue point also carry probability mass.
3.1
Shrinkage Areas
Using the conditions for minimality from the proof of Theorem 2, we can establish a direct relationship between the LS and the Lasso estimator in the following sense. For any b ∈ Rp , there
exists a set S(b) ⊆ Rp , such that the Lasso estimator assumes the value b if and only if the LS
estimator lies in S(b). We refer to the set S(b) as shrinkage area since the Lasso estimator can be
viewed as a procedure that shrinks the LS estimates from the set S(b) to the point b. Note that
by shrinkage we mean that kbk1 ≤ kzk1 for each z ∈ S(b), but |bj | > |zj | could hold for certain
components. The explicit form of S(b) is formalized in the following theorem.
7
0
0
1 0.5 ), λ = (0.75, 0.75)0
Figure 3: The functions h(0,−1) (in blue) and h(0,1) (in red) for X 0 X = ( 0.5
1
0
and β = (0, −0.25) , corresponding to the absolutely continuous part.
Theorem 7. For each b ∈ Rp there exists a set S(b) ⊆ Rp , such that
β̂L = b ⇐⇒ β̂LS ∈ S(b).
Moreover, for b ∈ Od , the set S(b) is given by
+
S(b) = {z ∈ Rp : (X 0 Xz)j = (X 0 Xb)j + sgn(bj )λj for j ∈ D−
, |(X 0 X(z − b))j | ≤ λj for j ∈ D0 }
Clearly, the sets S(b) are disjoint for different b’s.
Proof. Note that we have β̂LS − β = (X 0 X)−1 X 0 ε = (X 0 X)−1 W . With the minimality conditions
in (4) from the proof of Theorem 2 together with the fact that W = X 0 X(β̂LS − β) and some
rearranging, we get that m = β̂L − β minimizes Vβ if and only if β̂LS satisfies
0
0
(X X β̂LS )j = (X X β̂L )j − λj for j ∈ D−
(X 0 X β̂LS )j = (X 0 X β̂L )j + λj for j ∈ D+
|(X 0 X(β̂LS − β̂L ))j | ≤ λj
for j ∈ D0 ,
or, β̂LS ∈ S(b) for β̂L = b, as required.
Remark 2. Clearly, if b ∈ Rp satisfies bj 6= 0 for all j = 1, . . . , p, then S(b) is the singleton
S(b) = {b − (X 0 X)−1 λ̃},
where λ̃j = − sgn(bj )λj for j = 1, . . . , p. This implies that in case β̂L,j 6= 0 for all j, the Lasso
estimator is given by
β̂L = β̂LS − (X 0 X)−1 λ̃.
Note that aside from b, S(b) depends on X and λ only.
Given Theorem 7, we can identify areas in which components of the LS estimator are shrunk to
zero by the Lasso. For p = 2, this leads to the image displayed in Figure 4. Clearly, the shrinkage
areas are related to the polyhedral selection areas of Lee et al. (2016) but yield a different level of
information. Our results can identify the regions that lead to a given value b of β̂L rather than to
a general model (with given signs of the non-zero components of the estimator).
8
Figure 4: The shrinkage areas from Theorem 7 for p = 2. The blue parallelogram equals the set
S(0). The dark gray area should consists
of lines parallel to the adjacent edge of the parallelogram
where each line equals a set S b02 for b2 6= 0. Analogously, the light gray area consists of lines
parallel to the adjacent edge of the parallelogram and each of those lines equals a set S b01 for
b1 6= 0. The yellow areas contain all singletons S bb12 with b1 , b2 6= 0 as described in Remark 2.
1 0.5 ) and λ = (0.75, 0.75)0 .
In this example, X 0 X = ( 0.5
1
4
High-Dimensional Case
We now turn to the main case of this this article, the high-dimensional setting where p > n. We
make no assumptions on the regressor matrix X in this section. Using similar arguments as in
the case p ≤ n, we can again start by characterizing the distribution of the Lasso, albeit in a
somewhat less explicit form. Note that we have rk(X) < p and that the true parameter is not
identified without further assumptions. We denote by B0 the set of all β ∈ Rp that yield the model
given in (1), that is, B0 = {β ∈ Rp : Xβ = E(y) = µ}. Furthermore, it is important to note that
the Lasso solution need not be unique anymore. We give necessary and sufficient conditions for
uniqueness later on in Section 4.3.
Note that for any fixed β ∈ B0 , the function Vβ defined in (3) is minimized at β̂L − β, where
β̂L may be any solution of (2). All findings in this section also hold when p ≤ n, but more explicit
results for this case are found in Section 3. We start with a high-level result on the distribution.
Theorem 8. For any set M ⊆ Rp and any β ∈ B0 , we have
P (arg min Vβ (u) ∩ M 6= ∅) = P (W ∈ Āβ (M )),
u∈Rp
S
Qp
where W ∼ N (0, σ 2 X 0 X) and Āβ (M ) = m∈M Āβ (m) with Āβ (m) = X 0 Xm + j=1 Bβ,j (mj )
and
(
{sgn(mj + βj )λj } mj + βj 6= 0
Bβ,j (mj ) =
[−λj , λj ]
mj + βj = 0.
Proof. Using the same necessary and sufficient conditions for m ∈ Rp to be a minimizer of Vβ as
in (4), we see that
m ∈ arg min Vβ (u) ⇐⇒ W ∈ Aβ (m).
u∈Rp
While the distribution of β̂L −β depends on the choice of β ∈ B0 , the distribution of β̂L does not,
as it is determined by y ∼ N (µ, σ 2 In ). This is formalized in the following corollary. As mentioned
9
before, β̂L need not be unique. Also remember that β̂L itself minimizes the function L(β) defined
in (2).
Corollary 9. For any set B ⊆ Rp and any β ∈ B0 , we have
P (arg min L(β) ∩ B 6= φ) = P (W ∈ A0 (B)),
β∈Rp
where W ∼ N (0, σ 2 X 0 X) and A0 (B) =
S
b∈B
(
Bj (bj ) =
A0 (b) with A0 (b) = X 0 X(b − β) +
{sgn(bj )λj }
[−λj , λj ]
Qp
j=1
Bj (bj ) and
bj =
6 0
bj = 0.
In particular, the distribution of the estimator β̂L does not depend on the choice of β ∈ B0 .
Proof. First note that arg minu Vβ (u) = arg minβ L(β) − β. We thus have
b ∈ arg min L(β) ⇐⇒ b − β ∈ arg min Vβ (u).
u
β
Using Theorem 8, we get for any β ∈ B0
b − β ∈ arg min Vβ (u) ⇐⇒ W ∈ X 0 X(b − β) +
u
p
Y
Bj (bj ) = A0 (b).
(5)
j=1
Finally, note that A0 depends on β only through X 0 Xβ which assumes the same value for all
β ∈ B0 .
As the random variable W = X 0 ε has singular covariance matrix, some care needs to be
taken when computing the probability from Corollary 9 through the appropriate integral of the
corresponding density function. This is specified in Corollary 10.
Corollary 10. Let the columns of U form a basis of col(X 0 ). The probability that a Lasso solution
lies in the set B ∈ Rp can be written as
Z
P (arg min L(β) ∩ B 6= φ) = 1{col(X 0 ) ∩ A0 (B) 6= ∅}
φ(0,σ2 U 0 X 0 XU ) (s)ds
β∈Rp
U 0 A0 (B)
Proof. Note that U 0 W ∼ N (0, σ 2 U 0 X 0 XU ) and that U 0 X 0 XU is invertible. Let N be a matrix
whose columns form a basis of col(X 0 )⊥ , then N 0 W has covariance matrix σ 2 N 0 X 0 XN = 0 and
N 0 W = 0 almost surely. We therefore have
W ∈ A0 (B) ⇐⇒ (U, N )0 W ∈ (U, N )0 A0 (B) ⇐⇒ U 0 W ∈ U 0 A0 (B) and 0 ∈ N 0 A0 (B)
⇐⇒ U 0 W ∈ U 0 A0 (B) and col(X 0 ) ∩ A0 (B) 6= ∅,
which proves the claim.
4.1
Shrinkage Areas and Model Selection Properties
For the low-dimensional case, Theorem 7 gives shrinkage areas of the Lasso with respect to β̂LS .
In the high-dimensional case, similar results are derived in Theorem 11 whose proof is based on
Corollary 9. These shrinkage areas are now given with respect to X 0 y as this quantity is always
uniquely defined in the high-dimensional case.
Theorem 11. For each b ∈ Rp there exists a set S̄(b) ⊆ Rp such that
b ∈ arg min L(β) ⇐⇒ X 0 y ∈ S̄(b),
β∈Rp
Moreover, S̄(b) is given by
S̄(b) = X 0 Xb +
p
Y
Bj (bj ).
j=1
with
(
{sgn(bj )λj }
Bj (bj ) =
[−λj , λj ]
10
bj 6= 0
bj = 0.
Proof. As X 0 y = X 0 Xβ + X 0 ε = X 0 Xβ + W , this follows immediately from (5).
Remark 3. Inspecting the sets S̄(b) from Theorem 11 more closely, we see that they are in general
not disjoint for different values of b ∈ Rp . This illustrates the fact that the Lasso solution need not
be unique in high dimensions anymore. (This is of course in contrast to the low-dimensional case
as can be seen in Theorem 7.) Indeed, we can have S̄(b) ∩ S̄(b0 ) 6= ∅ as long as b − b0 ∈ ker(X)
and {sgn(bj ), sgn(b0j )} =
6 {−1, 1} for all j = 1, . . . , p. This also makes apparent that b and b0 may
be Lasso solutions not corresponding to the same model which has been noted by Tibshirani (2013)
for the case of λ1 = · · · = λp > 0. We get deeper into the issue of (non-)uniqueness in Section 4.3.
Moreover, just as for the low-dimensional case, note that aside from b, S̄(b) depends on X and
λ only.
Theorem 11 also sheds some light on which models M ⊆ {1, . . . , p} may in fact be chosen by
the Lasso estimator, where the model chosen by β̂L is given by {j : β̂L,j 6= 0}. We find that some
models will in fact never be selected by the Lasso.
0
Let BM = {b ∈ Rp : bj 6= 0 if and only if j ∈ M}. Since X 0 y lies
S in col(X ), any model
0
M ⊆ {1, . . . , p} that satisfies that col(X ) ∩ S̄(BM ) = ∅ with S̄(BM ) = b∈BM S̄(b) can be never
be selected by the Lasso estimator. Looking at the definition of S̄(b) in Theorem 11 and noting
that X 0 Xb ∈ col(X 0 ), we get
col(X 0 ) ∩ S̄(BM ) = ∅ ⇐⇒ col(X 0 ) ∩ BM = ∅,
where
BM =
p
[ Y
b∈BM
(
sgn(bj )λj
[−λj , λj ]
j=1
(
p
Y
j∈M
{−λj , λj }
=
j∈
/M
[−λj , λj ]
j=1
j∈M
j∈
/ M.
We summarize this in the following corollary.
Corollary 12. Let X ∈ Rn×p and λ ∈ Rp≥0 be given. There exist y ∈ Rn such that the corresponding Lasso solution selects model M ⊆ {1, . . . , p} if and only if
col(X 0 ) ∩ BM 6= ∅,
where
(
BM =
{−λj , λj }
[−λj , λj ]
j∈M
j∈
/ M.
which satisfies BM̃ ⊆ BM for M ⊆ M̃.
What do the sets BM look like? If M0 = ∅, B∅ is the p-dimensional λ-box, B{j} is the union
of two opposite facets of the λ-box, for 1 < |M| < p, BM is the union of p − |M| dimensional
faces of the λ-box, and B{1,...,p} simply contains the corners of the λ-box.
For partial tuning with M0 6= ∅, B∅ is p − |M0 |-dimensional and we have BM ⊆ BM0 for all
M ⊆ {1, . . . , p} as well as
{0} ⊆ col(X 0 ) ∩ BM0 6= ∅,
so that, not surprisingly, there always exist y such that the non-penalized components will be part
of the model chosen by the Lasso solution.
We illustrate the results above by the following simplistic yet instructive example.
Example 1. Suppose X = (1, 2), so that n = 1, p = 2 and let λ1 = λ2 = λ̄ (uniform tuning).
Note that
col(X 0 ) ∩ B{1} = ∅
for all λ̄ > 0, so that by Corollary 12, β̂L,1 = 0 for any value of y, in fact independent of B0 and
σ 2 ! This yields
P (β̂L,1 = 0) = 1.
11
To say something about the remaining distribution of β̂L using Corollary 9, note that W = X 0 ε =
( 12 ) ε. We therefore get P (β̂L = 0) = P (W ∈ A0 (0)) with
λ̄,λ̄]
1 ) + [−λ̄,λ̄] ,
A0 (0) = −X 0 Xβ + [−
=
−(β
+
2β
)
(
1
2
2
[−λ̄,λ̄]
[−λ̄,λ̄]
so that the event {W ∈ A0 (0)} is equivalent to the event that W lies in the set
{(s − µ) ( 12 ) : s ∈ [−λ̄/2, λ̄/2]}
with µ = Xβ = β1 + 2β2 , whose probability is given by
P (β̂L = 0) = Φ (λ̄/2 − µ)/σ − Φ (−λ̄/2 − µ)/σ .
Next, for z < 0, we have
P (β̂L,1 = 0, β̂L,2 ≤ z) = P (W ∈
[
[
λ̄,λ̄]
A0 ( b02 ) ) = P (W ∈
b2 ( 24 ) + [−
)
{−λ̄}
b2 <z
= P (W ∈
[
b2 <z
(2b2 − λ̄/2)( 12 )) = P (ε ≤ 2z − λ̄/2) = Φ(2z − λ̄/2).
b2 <z
Similarly, we get that for z > 0
P (β̂L,1 = 0, β̂L,2 > z) = 1 − Φ(2z + λ1 /2).
The distribution of β̂L is thus given by
β̂L,1 = 0
and β̂L,2 following the distribution given by
dF (z) = (Φ(λ1 /2) − Φ(−λ1 /2)) dδ0 (z) + 1{z<0} 2φ(2z − λ1 /2) dz + 1{z>0} 2φ(2z + λ1 /2) dz.
It is interesting to note that the distribution of β̂L,2 is the same as the one of the Lasso estimator
with the same penalization parameter in the smaller model yi = 2β2 + ε where the first regressor
is taken out. Indeed, using the Lasso in the smaller model is equivalent to using the Lasso in
the larger model in our example, as the procedure only considers models that do not contain the
first regressor. This fact is, of course, only valid for the specific form of X and λ. However, the
models which are considered by the Lasso estimator in the first place do not depend on β and ε
in the sense that certain values of X and λ may immediately rule out certain models completely
independently of y and is thus not to be considered a purely data-driven model selection procedure.
The choice between the location model (when β̂L,1 = 0 and β̂L,2 6= 0) and the pure noise model
(when β̂L,1 = β̂L,2 = 0) in our example does of course very much depend on β and ε. These
considerations are also illustrated in Figure 5 which depicts in which areas the quantity X 0 y has
to fall into in order for the Lasso estimator to choose a certain model. For clarification, note that
the Lasso is always unique in this example.
The considerations sparked by Example 1 suggest that in the high-dimensional setting, model
selection by the Lasso estimator may, indeed, not be a purely data-driven procedure in the sense
that there is a structural model or structural set M ⊆ {1, . . . , p} determined by X and λ only,
satisfying β̂L,j = 0 for any j ∈
/ M for all observations y ∈ Rn . In particular, the true parameter
β ∈ B0 as well as the distribution of ε do not have any influence on this set. In other words,
some models are considered by the model selection procedure, completely independently of the
data vector y. Put yet differently again, this means that for a given regressor matrix X, one can
restrict or choose the class of models considered by choice of λ.
Given all the considerations above, one might ask whether such a structural model M always
satisfies |M | ≤ n under certain conditions. Clearly, uniqueness would be a meaningful requirement
in this context, as then all Lasso solutions will choose models of cardinality of at most n as has been
shown in Tibshirani (2013)1 . If this was indeed the case, the Lasso estimator would be equivalent
to a low-dimensional Lasso procedure restricted to this structural model M this could be used to
employ results from low-dimensional models also for inference in high-dimensional settings.
1 Note that this fact alone does not imply that the structural set has cardinality of at most n since the active sets
may certainly vary over y.
12
Figure 5: The shrinkage areas with respect to X 0 y from Theorem 11 for Example 1. Displayed in
0
0
red is span(X 0 ), the area on which the probability mass of X
y is concentrated. The set S̄ ( 0 ) is
displayed in blue, while the yellow lines correspond to S̄ bb12 with b1 , b2 6= 0. The light gray area
is S̄ b02 with b2 6= 0, whereas the dark gray area equals S̄ b01 with b1 6= 0.
Example 1 (continued). Here, M = {2} and the distribution of the Lasso estimator is equal to
the distribution of the Lasso in the model
y = 2β2 + ε.
Using this property, one can construct a confidence set for β2 based on the Lasso using results from
Pötscher & Schneider (2010).
In Example 1 the Lasso solution is always unique (for any value of y) since the column span
of X 0 does not intersect overlapping shrinkage areas. It is not difficult, however, to construct an
example where the Lasso solution is not unique anymore.
Example 2. Again, take the model from Example 1 with X = (1, 2). This time, choose λ = (1, 2)0 .
It can easily be seen using Theorem 11 that for each y < −1,
0
y+1
β̂L = y+1
,
β̂
=
, and β̂L = ( y+1−2c
) for (y + 1)/2 < c < 0.
L
c
0
2
all are Lasso solutions for the same value of y. Similarly, for y > 1,
0
y−1
β̂L = y−1
,
β̂
=
, and β̂L = ( y+1−2c
) for 0 < c < (y + 1)/2.
L
c
0
2
all are Lasso solutions for the same value of y. (Note that β̂L = 0 for all y with |y| ≤ 1.) The
corresponding shrinkage areas are illustrated in Figure 6.
Example 2 shows an already known property of the Lasso from another perspective: The
solution to the Lasso problem is, in general, not unique. Moreover, if the solution is not unique,
then, by convexity of the problem, there exists an uncountable set of solutions2 . This example
shows, moreover, that the set of y which yield non-unique Lasso solutions is not a null set, in fact,
in this example it occurs with probability 2Φ(−1).
Of course, this problem could be overcome by slightly altering the choice of the tuning parameters, even though, this would imply to make a choice of the class of models under consideration,
as pointed out previously in this section.
2 This
fact has been pointed out by Tibshirani (2013) in Lemma 1 for the case of uniform tuning.
13
Figure 6: The shrinkage areas with respect to X 0 y from Theorem 11 for Example 2. Displayed in
0
0
red is span(X 0 ), the area on which the probability mass of X
y is concentrated. The set S ( 0 ) is
displayed in blue, while the yellow lines correspond to S bb12 with b1 , b2 6= 0. The light gray area
is S b02 with b2 6= 0, whereas the dark gray area equals S b01 with b1 6= 0. The red line passes
through S ( 00 ) where the solution is unique but also through the line where the light gray, the dark
gray and the yellow areas intersect.
4.2
Structural Sets
Clearly, Example 2 shows that the structural set may also be equal to the entire set of explanatory
variables. In fact, It is easy to see that for n = 1 and p = 2, the Lasso estimator will always
have a structural set with cardinality n = 1 whenever we have uniqueness. The question is, of
course, whether the same can be said in general. Before answering this question, we show how the
structural set can fully be determined given X and λ by counting how many faces of the λ-box
B∅ are intersected by col(X 0 ).
Theorem 13. Let X ∈ Rn×p and λ ∈ Rp≥0 be given. Let M be the structural set of X and λ that
contains all j ∈ {1, . . . , p} such that there exist y ∈ Rn so that the corresponding Lasso solution
β̂L satisfies β̂L,j 6= 0, that is, the set of all regressors that are part of a Lasso solution for some
observation y. This set is given by
M = M (X, λ) = j ∈ {1, . . . , p} : B{j} ∩ col(X 0 ) 6= ∅ .
Proof. By Corollary 12, there exist y ∈ Rn such that the corresponding Lasso solution chooses
model M if and only if BM ∩ col(X 0 ) 6= ∅. For any {j} ⊆ M, we have, by Corollary 12 also,
BM ⊆ B{j} , so that B{j} ∩ col(X 0 ) 6= ∅ also.
Remark 4. As indicated in Theorem 13 and as discussed above, the structural set M depends
on X and λ only. Moreover, it can easily be seen that it depends on the tuning parameters λ
only through the penalization weighting in the sense that whenever λ = λ̄ω for some λ̄ > 0 and
ω ∈ Rp≥0 , M (X, λ) = M (X, ω) follows. This implies that, in particular, in the common case of
uniform tuning with λ̄ = λ1 = · · · = λp , the structural set only depends on X!
Coming back to the conjecture whether the structural set always satisfies |M | ≤ min{n, p} in
case the solutions are unique, using Theorem 13, we can list the following simple example with
n = 2 and p = 3 to show that this cannot be the case in general. However, note that Theorem 13
allows to compute the structural set and that whenever |M | ≤ n, the resulting Lasso estimator is,
in fact, just equivalent to a low-dimensional procedure!
14
Figure 7: The intersection of col(X 0 ) (in gray and yellow) with the λ-cube (in orange). The
upper left edge is contained in B{1,2} whereas the upper back edge is contained in B{2,3} . (Each
B{i,j} contains four parallel edges.) To view this figure in terms of shrinkage areas, note that the
areas corresponding to single-regressor models are displayed in gray while the shrinkage areas that
correspond to two-regressor models are displayed in yellow. The intersection of the λ-cube with
col(X 0 ) which which corresponds to the zero estimator is displayed in blue.
Example 3. Let
X=
1
0
1
1
0
1
and λ = (1, 1, 1)0 . Then the structural set is clearly given by
M = {1, 2, 3},
as (1, 1, 0)0 ∈ col(X 0 ) ∩ B{1,2} and (0, 1, 1)0 ∈ col(X 0 ) ∩ B{2,3} and BM ⊆ B{j} for any j ∈ M
by Corollary 12, see Figure 7 for illustration. Yet the Lasso solutions for this X and λ are always
unique which can be checked on the basis of Theorem 14 in the subsequent section.
4.3
A Necessary and Sufficient Condition for Uniqueness
We now turn to some results revolving around uniqueness of the Lasso estimator which can be
obtained with a similar geometric approach, that is, studying the intersection of the λ-cube with
col(X 0 ). Note that by uniqueness we mean that for a given X ∈ Rn×p and a given λ ∈ Rp , the
Lasso solutions are unique for all observations y ∈ Rn .
Tibshirani (2013) showed that for a given regressor matrix X, Lasso solutions are unique
for all observations y if the columns of X are in general position 3 which occurs when no kdimensional affine4 subspace for any k < min(n, p) contains more than k + 1 elements of the
3 Note that general position does not mean that any selection of n columns of X is linearly independent, as has
sometimes been suggested in the literature, these two concepts are in fact unrelated.
4 In Tibshirani (2013), the word “affine” is missing which has caused some confusion in the literature.
15
the set {±X1 , . . . , ±Xp }, excluding antipodal pairs (see p. 1463 in Tibshirani, 2013). In fact, the
solutions are then unique for all choices of the tuning parameter, provided that all components
are tuned uniformly. As the condition is sufficient, one may ask whether it is also necessary. The
answer to this question is, in fact, no, as can easily be seen from the example below.
When can exist non-unique solutions? For a given X ∈ Rn×p and λ ∈ Rp this occurs if and
only if there exist b, b̃ ∈ Rp with b 6= b̃ and
S(b) ∩ S(b̃) ∩ col(X 0 ) 6= ∅.
More concretely, by Theorem 11, and since the Lasso fit Xb is always unique5 , this means that
X 0 Xb + v = X 0 X b̃ + v,
where v ∈ col(X 0 )∩BM for some M ⊆ {1, . . . , p} and b̃Mc = bMc = 0. Moreover, for j ∈ M\M0 ,
we have sgn(bj ) = sgn(vj ) whenever bj 6= 0 as well as sgn(b̃j ) = sgn(vj ) whenever b̃j 6= 0. Note
that we therefore have Xb = XM bM = XM b̃M = X b̃, implying that the columns of XM must be
linearly dependent. So non-uniqueness occurs if and only if col(X 0 ) ∩ BM 6= ∅ for M ⊆ {1, . . . , p}
with linearly dependent columns in XM . The following example now immediately shows that the
columns of X being in general position is not necessary for uniqueness.
Example. Let
X=
1
0
1
0
2
1
0
.
3
Clearly, the columns are not general position, however, all Lasso solutions are unique for any
choice of the tuning parameter when the components are tuned uniformly. This is the case since
col(X 0 ) ∩ BM = ∅ whenever {1, 2} ⊆ M or |M | > 2 which can easily be checked using the fact
that v ∈ col(X 0 ) if and only if v 0 w1 = v 0 w2 = 0 for ker(X) = span{w1 , w2 }.
We find that the following criterion is in fact sufficient as well as necessary for uniqueness of
all Lasso solutions for a given X and λ.
Theorem 14 (Uniqueness). Let X ∈ Rn×p and λ ∈ Rp≥0 . The Lasso solution is unique for all
y ∈ Rn if and only if
col(X 0 ) ∩ BM = ∅ for all M ⊆ {1, . . . , p} with |M| > rk(X).
Proof. ( =⇒ ) Assume the condition is not satisfied. Then there exists v ∈ BM with |M| > rk(X)
and v = X 0 z for some z ∈ Rn . We show that there is a y ∈ Rn such that the corresponding Lasso
problem is not uniquely solvable.
If Xj = 0 for some j ∈ M0 , we are done as the corresponding coefficient may be arbitrary.
Note that Xj = 0 for j ∈ M \ M0 is not possible since that would imply vj = 0 as v ∈ col(X 0 ),
but that contradicts v ∈ BM . We therefore assume that Xj 6= 0 for all j ∈ M.
Since |M| > rk(X), there is a column of XM , say Xj (Xj 6= 0), that can be written as a linear
combination of the other columns, in particular,
X
dXj =
cl Xl
l∈M\{j}
where d = sgn(vj ) if λj 6= 0 and d = 1 if λj = 0.
b ∈ Rp by
d
2c
bl = sgn(vl )
0
Moreover, let c = maxl∈M\{j} |cl | > 0. Define
l=j
l ∈ M \ {j}
l∈
/ M.
Then b is a Lasso solution for y = z + Xb since
X 0 y = X 0 Xb + X 0 z = X 0 Xb + v ∈ S(b).
5 This has been shown in Lemma 1 in Tibshirani (2013) for uniform tuning and can easily be extended to nonuniform and partial tuning.
16
We now construct b̃ ∈ Rp with b̃ 6= b that is also a Lasso solution for the same y by
(
cl
sgn(vl ) + 2c
l ∈ M \ {j}
b̃l =
0
l = j or l ∈
/ M.
Clearly, b 6= b̃, sgn(bl ) = sgn(b̃l ) = sgn(vj ) for l ∈ M \ {j} and
Xb =
X
l∈M
bl Xl =
X
l∈M\{j}
bl Xl +
d
Xj =
2c
X
(bl +
l∈M\{j}
cl
)Xl = X b̃.
2c
We therefore get
X 0 y = X 0 Xb + v = X 0 X b̃ + v ∈ S(b̃)
also, implying that both b and b̃ are Lasso solutions for the the given y.
( ⇐= ) We now prove the other direction. Assume that there exists y ∈ Rn such that non-unique
Lasso solutions b 6= b̃ exist. As discussed above, this implies the existence of v ∈ BM ∩ col(X 0 ) for
some M ⊆ {1, . . . , p} with XM bM = XM b̃M and bMc = b̃Mc = 0, entailing that the columns of
XM are linearly dependent.
If |M| > rk(X), we are done. If |M| ≤ rk(X), we do the following. Since we have rk(XM ) <
|M| ≤ rk(X), we can pick z ∈ Rn such that z ∈ col(XM )⊥ \ col(XMc )⊥ . This is possible since
col(XM )⊥ \ col(XMc )⊥ = ∅ ⇐⇒ col(XM )⊥ ⊆ col(XMc )⊥ ⇐⇒ col(XMc ) ⊆ col(XM )
⇐⇒ col(XM ) = col(XM , XMc ) = col(X) ⇐⇒ rk(XM ) = rk(X),
which is not the case. This z satisfies (X 0 z)M = (XM )0 z = 0 and (X 0 z)Mc = (XMc )0 z 6= 0, so
that we can find c ∈ R such that
ṽ = v + c X 0 z ∈ BM̃ ∩ col(X 0 )
with M ⊆ M̃ and |M| < |M̃|. As long as |M̃| ≤ rk(X), repeat the steps above with v = ṽ and
M = M̃.
Remark. Note that just as for Theorem 13, the result from the above theorem depends on λ only
through the penalization weights, meaning that for any M ⊆ {1, . . . , p}, whenever λ = λ̄ω for
some λ̄ > 0 and ω ∈ Rp≥0 , we have col(X 0 ) ∩ BM (λ) = ∅ if and only if col(X 0 ) ∩ BM (ω) (when
indicating the dependence of BM on the tuning parameters).
As mentioned in the preamble of Section 4, Theorem 14 does not require p > n, so that it
also covers the low-dimensional case. Clearly, the condition for uniqueness is trivially satisfied if
rk(X) = p.
5
Conclusion
We give explicit formulae regarding the distribution of the Lasso estimator in finite-samples assuming Gaussian errors. In the low-dimensional case, we consider the cdf as well as the density
functions conditional on “active sets” of the estimator. Our results exploit the structure of the
underlying optimization problem of the Lasso estimator and do not hinge on the normality assumption of the error term. We also explicitly characterize the correspondence between the Lasso
and the LS estimator: It is shown that the Lasso estimator essentially creates shrinkage areas
around the axes inside which the probability mass of the LS estimator is “compressed” into lowerdimensional densities that can be specified conditional on the active set of the estimator. As a
result, the distribution looks like a pieced-together combination of Gaussian-like densities with
each active set having its own distributional piece whose dimension depends on the number of
nonzero components in that it is given by the number of nonzero components, resulting also in
point mass at the origin and mass being distributed along the axes.
The form of the distribution is even more intricate in the high-dimensional case in which the
estimator may not be unique anymore. We quantify the relationship between a Lasso solution
17
and the quantity X 0 y (rather than the LS estimator as in the low-dimensional case). We gain
valuable insights into the behavior of the estimator by illustrating that some models may never be
selected by the estimator. In fact, a structural set based on the regressor matrix and penalization
weights only can be computed that contains all covariates that are part of a Lasso solution for
some response vector y. In case this structural set has cardinality less than or equal to n, the Lasso
is equivalent to a low-dimensional procedure and results from the p ≤ n-framework can be used
for inference. It remains to be seen whether on can quantify that this is in some sense a generic
case.
Finally, the previous insights allow us to close a gap in the literature by providing a condition
for uniqueness of the Lasso estimator that is both necessary and sufficient.
Acknowledgements
The authors gratefully acknowledge support from the Deutsche Forschungsgemeinschaft (DFG)
through grant FOR 916 and thank Thomas Hack and Nikos Dafnis for very insightful discussions
on the geometric aspects of this paper.
References
Ewald, K. & Schneider, U. (2015). Confidence sets based on the Lasso. Tech. Rep. 1507.05315,
arXiv.
Jagannath, R. & Upadhye, N. S. (2016). The Lasso estimator: Distributional properties. Tech.
Rep. 1605:03280, arXiv.
Knight, K. & Fu, W. (2000). Asymptotics of Lasso-type estimators. Annals of Statistics 28,
1356–1378.
Lee, J. D., Sun, D. L., Sun, Y. & Taylor, J. E. (2016). Exact post-selection inference with
an application to the Lasso. Annals of Statistics 44, 907–927.
Pötscher, B. M. & Leeb, H. (2009). On the distribution of penalized maximum likelihood
estimators: The LASSO, SCAD, and thresholding. Journal of Multivariate Analysis 100, 2065–
2082.
Pötscher, B. M. & Schneider, U. (2010). Confidence sets based on penalized maximum
likelihood estimators in Gaussian regression. Electronic Journal of Statistics 4, 334–360.
Tibshirani, R. (1996). Regression shrinkage and selection via the Lasso. Journal of the Royal
Statistical Society Series B 58, 267–288.
Tibshirani, R. J. (2013). The Lasso problem and uniqueness. Electronic Journal of Statistics 7,
1935–1490.
Zhou, Q. (2014). Monte carlo simulation for lasso-type problems by estimator augmentation.
Journal of the American Statistical Association 109, 1495–1516.
18
| 10math.ST
|
arXiv:1704.07114v3 [cs.IT] 20 Jun 2017
Regular Decomposition: an information and
graph theoretic approach to stochastic block
models∗
Hannu Reittu†, Fülöp Bazsó‡, Ilkka Norros§
Abstract
A method for compression of large graphs and matrices to a block
structure is proposed. Szemerédi’s regularity lemma is used as a generic
motivation of the significance of stochastic block models. Another ingredient of the method is Rissanen’s minimum description length principle
(MDL). We propose practical algorithms and provide theoretical results
on the accuracy and consistency of the method.
Keywords: Szemerédi’s Regularity Lemma, Minimum Description Length
Principle, stochastic block model, big data
Contents
1 Introduction
2
2 Basics and definitions
2.1 Szemerédi’s Regularity Lemma . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2 Stochastic block models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.3 The Minimum Description Length (MDL) principle . . . . . . . . . . . . . . . .
3
3
5
7
3 The
3.1
3.2
3.3
family of stochastic block models in the MDL
Two-part MDL for simple graphs . . . . . . . . . . .
Two-part MDL for matrices . . . . . . . . . . . . . .
Refined MDL and asymptotic model complexity . . .
4 The Regular Decomposition approach to
stochastic block models
4.1 Block model codes . . . . . . . . . . . . . . . . . .
4.2 Algorithms for description length minimization . .
4.3 Accuracy of block structure identification by MDL
4.4 ǫ-regularity vs. stochastic block models . . . . . . .
5 Proof of Theorem 4.13
.
.
.
.
framework
. . . . . . . . . . . . . . .
. . . . . . . . . . . . . . .
. . . . . . . . . . . . . . .
8
8
10
11
.
.
.
.
13
13
15
20
21
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
22
∗ This
work has been submitted to the IEEE for possible publication. Copyright may be
transferred without notice, after which this version may no longer be accessible.
† VTT Technical Research Centre of Finland Ltd, P.O. Box 1000, 02044 VTT. Email
[email protected]
‡ Department of Theory, Institute for Particle and Nuclear Physics, Wigner Research Centre
for Physics, Hungarian Academy of Sciences, P.O. Box 49, H-1525 Budapest, Hungary. Email
[email protected].
§ VTT Technical Research Centre of Finland Ltd, P.O. Box 1000, 02044 VTT. Email
[email protected]
1
A Chernoff bounds and other information-theoretic preliminaries
1
30
Introduction
Szemerédi’s Regularity Lemma (SRL) [24] is a fundamental result in graph
theory. Roughly speaking, SRL states that any large enough graph can be approximated arbitrarily well by nearly regular, pseudo-random bipartite graphs,
induced by a partition of the node set into a bounded number of equal-sized
sets. For many graph problems it suffices to study the problem on a corresponding random structure, resulting in a much easier problem (for a review,
see, e.g., [10]). SRL is fundamental also in theoretical computer science, for instance, in showing the existence of polynomial time approximations for solving
dense graph problems, and in characterizing the class of so-called testable graph
properties [2].
Despite the impressive theoretical applications of SRL, it has had only few
applications to ’real-life’ problems. The main reason might be that SRL has
extremely bad worst cases in the sense that the lower bound of graph sizes, for
which the partition claim holds with reasonable accuracy and without a single
exception, is enormous. Thus, a real-world application of SRL in literal sense
is impossible. However, this drawback does not mean that regular partitions
could not appear in much smaller scales relevant to applications. On the contrary, one could conjecture that regular partitions or structures be commonplace
and worth of revealing. The goal of our work is on such realistic yet preferably
large networks that nowadays appear in almost all imaginable application areas. The regular structure granted by SRL is then replaced by a probabilistic
model, substituting the regular bipartite components with truly random bipartite graphs. This model class is well known as stochastic block models [9], and
it has recently gained much attention in research and popularity in practical
applications like community detection [16, 7, 11]. However, the fundamental
nature of SRL suggests the (heuristic) conjecture that stochastic block models
in fact present a very generic form of the separation of structure and randomness in large real-world systems. These models have good algorithmic properties
in maximum likelihood fitting, and one can use powerful tools like expectation
maximization, simulated annealing and Monte Carlo Markov Chain algorithms.
There are also some examples of graph decomposition applications that have
been explicitly inspired by SRL. The practical contexts are varied: brain cortex
analysis [12], image processing [23], peer-to-peer network[15], analysing Functional Magnetic Resonance (fMRI) data to depict functional connectivity of the
brain [14], a matrix of multiple time series [18] . In the last-mentioned work,
the method was generalized from graphs to arbitrary positive matrices using a
Poissonian construction as an intermediary step.
Interestingly, the authors of the recent work [21] define a ’practical’ variant
of SRL by relaxing algorithmic SRL in a certain way to make it more usable
in machine-learning tasks, see also [17]. Bolla [3] has developed a spectral
approach for finding regular structures of graphs and matrices. Our emphasis
is more information-theoretical by nature and continues the works [12, 15, 18].
It would be very interesting to compare the methods of [21, 3] and those of this
paper in depth.
The third main ingredient in this paper is Rissanen’s Minimum Description
2
Length (MDL) principle (see [6]), according to which the quality of a model
should be measured by the length of the bit-string that it yields for unique
encoding of the data. In our case, the data has the form of a graph or matrix, and
we present the stochastic block models as a modeling space in the sense of the
MDL theory. Within this modeling space, the model corresponding to minimum
code length coding presents the optimally regular decomposition (partition) of
the data. Therefore we call this Szemerédi-motivated and technically MDLbased approach Regular Decomposition.
By information theory, the optimal coding reveals as much redundancy in
the data as is possible from the given modeling point of view. The regular
structure has a high degree of redundancy: a regular pair is an almost structureless subgraph where almost all nodes have similar and uniform connectivity
patterns. By definition, the MDL principle should be able to discover regular
structures. Note also that this principle presents a case of ’Occam’s Razor’, a
general rule of reasoning that has proved fruitful in all areas of science.
In non-hierarchic clustering tasks, it has been a major challenge to select the
’right’ size of a partition (k). Intuitively, the optimal choice of k will strike a
balance between the simplest partition of the data using a single cluster and the
maximal partition assigning each data point to its own cluster, and selecting
something in between. A popular device has been the Akaike information criterion (AIC), which was applied also in [12]. The AIC simply adds the ’number
of model parameters’ to the maximal log-likelihood and chooses the model that
minimizes the sum. Our MDL-based approach solves the corresponding model
selection task in a better founded way.
The contributions of this paper are the following: (i) the linkage of SRL,
stochastic block models and the MDL principle, (ii) the unified handling of
graphs and matrices, (iii) effective practical algorithms for revealing regular
structures in data, and (iv) Theorem 4.13 that characterizes how accurately the
MDL principle identifies a stochastic block model.
The paper is structured as follows. Section 2 presents the definitions of
the central notions: SRL, stochastic block models and MDL. The last topic
is expanded in Section 3. Section 4 presents our main technical results, in
particular the algorithms and Theorem 4.13. The proof of Theorem 4.13 is
given in Section 5. It is structured into several propositions and makes strong
use of information-theoretic tools presented in Appendix A.
2
2.1
Basics and definitions
Szemerédi’s Regularity Lemma
Consider simple graphs G(V, E), where V is the set of nodes (vertices) and E
is the set of links (edges). The link density of a non-empty node set X ⊆ V is
defined as
d(X) =
e(X)
,
|X|
2
where e(X) = {{v, w} ∈ E : v, w ∈ X} ,
3
and |·| denotes the cardinality of a set. Similarly, the link density between two
disjoint non-empty node sets X, Y ⊆ V is defined as
d(X, Y ) :=
e(X, Y )
,
|X| |Y |
where e(X, Y ) = {{v, w} ∈ E : v ∈ X, w ∈ Y } .
Definition 2.1 Let ǫ > 0. A pair of disjoint sets A, B ⊆ V is called ǫ-regular,
if for every X ⊆ A and Y ⊆ B, such that |X| > ǫ |A| and |Y | > ǫ |B| we have
|d(X, Y ) − d(A, B)| < ǫ.
A partition ξ = {V0 , V1 , V2 , · · · , Vk } of V into k + 1 sets, where all except V0
have equal cardinalities, is called ǫ-regular, iff all except at most ǫk 2 pairs are
ǫ-regular and |V0 | < ǫ|V |.
Theorem 2.2 (Szemerédi’s Regularity Lemma, [24]) For every ǫ > 0 and for
any positive integer m, there are positive integers N (ǫ, m) and M (ǫ, m), such
that for any graph G(E, V ) with |V | ≥ N (ǫ, m) there is an ǫ-regular partition of
V into k + 1 classes with m ≤ k ≤ M (ǫ, m).
SRL states, roughly speaking, that nodes of any large enough graph can
be partitioned into a bounded number, k, of equal-sized ’clusters’ and into one
small set in such a way that most pairs of clusters look like random bipartite
graphs, whose link probability equals the link density between the pair.
The claim of SRL is significant for sufficiently dense graphs, i.e., when the
link density is higher than ǫ. However, the result can be modified to apply to
sparse graphs by multiplying ǫ at the right-hand side of the regularity definition
by the link density of the entire graph [22].
The amazing fact in SRL is that the regularity claim holds for all graphs
starting from a lower bound for size that depends only on ǫ. However, it is also
well known that this dependence on ǫ is of extremely bad kind: the known lower
bound for the graph size N (ǫ, m) behaves like a tower of powers of 2:
2
..
2
2.
,
where the height of the tower is upper-bounded by 1/ǫ5 . Such a number is too
big to be considered in any applications. Thus, all real-world networks fall into
a ’gray area’ with respect to SRL.
The magnitude of N (ǫ, m) is a problem for using algorithmic versions of the
SRL, like the one introduced by Alon et al, [1]. Although its time complexity is
only polynomial O(n2.376··· ), corresponding to the time for multiplying two n×n
binary matrices, it requires this enormous size (n ≥ N (ǫ, m)) of graph to be able
to find a regular partition. A considerable improvement was found in the recent
work [5], where the running time is only linear in the graph size. Perhaps even
more importantly, this randomized algorithm works in a more realistic fashion:
for any graph, it either finds an ǫ-regular partition or concludes that such a
partition does not exists. Thus, it does not need to work in the safe region
of graph sizes specified in SRL. Another algorithm with the same feature was
suggested by Tao [25]. In principle, such algorithms could possibly be applicable
for real-world graphs, although some extra-big numbers are hidden in some of
the constants of the algorithm.
4
2.2
Stochastic block models
The notion of an ǫ-regular partition is purely combinatorial. The stochastic
model closest to this notion is the following.
Definition 2.3 Let V be a finite set and ξ = {A1 , . . . , Ak } a partition of V .
A stochastic block model is a random graph G = (V, E) with the following
structure:
• There is a symmetric k × k matrix D = (dij )ki,j=1 of real numbers dij ∈
[0, 1] satisfying the irreducibility condition that no two rows are equal, i.e.,
for all i, j, i < j, there is qij ∈ {1, . . . , k} such that diqij 6= djqij ; (1)
• For every pair {v, w} of distinct nodes of V such that v ∈ Ai , w ∈ Aj , let
evw = ewv be a Bernoulli random variable with parameter dij , assuming
that all evw ’s are independent. The edges of G are
E = {{v, w} : v, w ∈ V, v 6= w, evw = 1} .
Note that the case of the trivial partition ξ = {V } yields the classical random
graph with edge probability d11 .
A graph sequence Gn = (Vn , En ) presenting copies of the same stochastic
block model in different sizes can, for definiteness, be constructed as follows.
Pk
Construction 2.4 Let γ1 , . . . , γk be positive real numbers such that i=1 γi =
1. Divide the interval (0, 1] into k segments
#
k−1
X
γi , 1 ,
I1 = (0, γ1 ], I2 = (γ1 , γ1 + γ2 ], . . . , Ik =
i=1
and denote Γ = {I1 , . . . , Ik }. For n = 1, 2, . . ., let the vertices of Gn be
i
: i ∈ {1, . . . , n} .
Vn =
n
For each n, let ξn be the partition of Vn into the blocks
(n)
Ai
= Ii ∩ Vn ,
i = 1, . . . , k.
For small n, we may obtain several empty copies of the empty set numbered
as
blocks.
However,
from some n0 on, all blocks are non-empty and ξn =
n
o
(n)
(n)
A1 , . . . , Ak
is a genuine partition of Vn . We can then generate stochastic
block models based on (Vn , ξn , D) according to Definition 2.3.
Remark 2.5 A slightly different kind of stochastic block model can be defined
(n)
by drawing first the sizes of blocks Ai as independent Poisson(γi n) random
variables and proceeding then with the matrix D as before. The additional level
of randomness, regarding the block sizes, is however of no interest in the present
paper.
Next, we define the notion of a Poissonian block model in complete analogy
with Definition 2.3.
5
Definition 2.6 Let V be a finite set of vertices, n = |V |, and let ξ = {A1 , . . . , Ak }
be a partition of V . The symmetric Poissonian block model is a symmetric random n × n matrix E with the following structure:
• There is a symmetric k × k matrix Λ = (λij )ki,j=1 of non-negative real
numbers satisfying the irreducibility condition that no two rows are equal,
i.e.,
for all i, j, i < j, there is qij ∈ {1, . . . , k} such that λiqij 6= λjqij ; (2)
• For every unordered pair {v, w} of distinct nodes of V such that v ∈ Ai ,
w ∈ Aj , let evw = ewv be a Poisson random variable with parameter λij ,
assuming that all evw ’s are independent. The matrix elements of E are
evw for v 6= w, and evv = 0 for the diagonal elements.
P
P
Thanks to the independence assumption, the sums u∈A v∈B euv are Poisson distributed for any A, B ∈ ξ.
Remark 2.7 The rest of the technical contents of this paper focus on the simple
binary and Poissonian models of Definitions 2.3 and 2.6. However, the following
extensions are straightforward:
• bipartite graphs: this is just a subset of simple graphs;
• m × n matrices with independent Poissonian elements: a matrix can be
seen as consisting of edge weights of a bipartite graph, where the parts are
the index sets of the rows and columns of the matrix, respectively;
• directed graphs: a directed graph can be presented as a bipartite graph
consisting of two parts of equal size, presenting the input and output ports
of each node.
The following construction, inspired by our conditionally Poissonian graph
model [13], is the key to extend the MDL methodology for data that has the
most common form of a large non-negative matrix.
Construction 2.8 Let C = (cij ) be a non-negative m × n matrix. Let N be a
(rather large) integer and denote J = {1, . . . , N }. Let
(1)
(2)
(1)
V = J1 ∪ · · · ∪ J m
,
W = J1 ∪ · · · ∪ Jn(2) ,
(ι)
Ji n= {(ι,
o i)} × J. Define a bipartite Poissonian block model PN (C) = (V ∪
(·)
(·)
, C) with blocks J· and mean matrix C.
W, J·
Although we have not studied this at the technical level, it is natural to
expect that, with large N , a partition of PN (C) with minimum description
(·)
length would with high probability keep the blocks J· unbroken. Because the
regular decomposition algorithm for Poissonian block models operates only on
means over blocks (see subsection 4.2), Construction 2.8 is a strong heuristic
argument that this algorithm applies as such to regular decomposition of nonnegative matrices.
6
2.3
The Minimum Description Length (MDL) principle
The Minimum Description Length (MDL) Principle was introduced by Jorma
Rissanen, inspired by Kolmogorov’s complexity theory, and an extensive presentation can be found in Grünwald’s monography [6], see also [20]. The basic
idea is the following: a set D of data is optimally explained by a model M,
when the combined description of the (i) model and (ii) the data as interpreted
in this model is as concise as possible. By description we mean here a code that
specifies an object uniquely.
The principle is best illustrated by our actual case, simple graphs. A graph
G = (V, E) with |V | = n can always be encoded as a binary string of length
n
2 = n(n − 1)/2, where each binary variable corresponds to a node pair and a
value 1 (resp. 0) indicates an edge (resp. absense of an edge). Thus, the MDL
of G is always at most n2 . However, G may have a structure whose disclosure
would allow a much shorter description. Our heuristic postulate is that in the
case of graphs and similar objects a good a priori class of models should be
inferred from SRL, which points to stochastic block models.
Definition 2.9 Denote by Mn/k the set of irreducible stochastic block models
(V, ξ, D) with
• |V | = n,
• |ξ| = k, and, denoting ξ = {V1 , . . . , Vk },
• for i, j ∈ {1, . . . , k},
dij =
hij
, hij ∈ N,
|Vi ||Vj |
dii =
hii
,
|Vi |
2
hii ∈ N.
The condition in the last bullet entails that each modeling space Mn/k is
finite.
Remark 2.10 Without the irreducibility condition (1), there would not be a
bijection between stochastic block models and their parameterizations.
The models in Mn/k are parameterized by Θk = (ξ, D). A good model for
a graph G is the one that gives maximal probability for G and is called the
maximum likelihood model. We denote the parameter of this model
Θ̂k (G) := arg max (P (G | Θk )),
(3)
Θk ∈Mn/k
where P (G | Θk ) denotes the probability that the probabilistic model specified
by Θk produces G.
One part of likelihood optimization is trivial: when a partition ξ is selected
for a given graph G, the optimal link probabilities are the empirical link densities:
dij =
e(Vi , Vj )
, i 6= j,
|Vi ||Vj |
dii =
e(Vi )
.
|Vi |
(4)
2
Thus, the nontrivial part is to find the optimal partition for the given graph.
This is the focus of the next sections.
7
3
3.1
The family of stochastic block models in the
MDL framework
Two-part MDL for simple graphs
Let us denote the set of all simple graphs with n nodes as
Ωn = {G : G = (V, E) is a graph, |V | = n} .
A prefix (binary) coding of a finite set Ω is an injective mapping
C : Ω → ∪s≥1 {0, 1}s
(5)
such that no code is a prefix of another code. Recall the following proposition
from information theory (see, e.g., [4]):
Theorem 3.1 (Kraft’s Inequality) For an m-element alphabet there exists a
binary prefix
with code lengths l1 , l2 , · · · , lm iff the code lengths
P coding−lscheme
i
≤
1.
2
satisfy:
i=1,··· ,m
An important application of Theorem 3.1 is the following: if letters are
drawn from an alphabet with probabilities p1 , p2 , · · · , pm , then there exists a
prefix coding with code lengths ⌈− log p1 ⌉, · · · , ⌈− log pm ⌉, and such a coding
scheme is optimal in the sense that it minimizes the expected code length (in this
section, the logarithms are in base 2). In particular, any probability distribution
P on the graph space Ωn indicates that there exists a prefix coding that assigns
codes to elements of G ∈ Ωn with lengths equal to ⌈− log P ({G})⌉.
The code length l(·) is the number of binary digits in the code of the corresponding graph. In case of a large set Ω, most such codes are long and as a result
the ceiling function can be omitted, a case we assume in sequel. A good model
results in good compression, meaning that a graph can be described by much
less bits than there are elements in the adjacency matrix. An incompressible
case corresponds to the uniform distribution on Ωn and results in code length
− log (1/ | Ω |) = n2 , equivalent to writing down all elements of the adjacency
matrix.
For every graph G from Ωn and model P we can associate an encoding with
code length distribution − log P (· | Θ̂k (G)). However, this is not all, since in
order to be able to decode we must know what particular probabilistic model
P is used. This means that also Θ̂k (G) must be prefix encoded, with some
code-length L(Θ̂k (G)). We end up with the following description length:
l(G) = ⌈− log P (G | Θ̂k (G))⌉ + L(Θ̂k (G)).
(6)
Eq. (6) presents the so-called two-part MDL, [6]. In an asymptotic regime with
n → ∞, we get an analytic expression of the refined MDL. A simple way of
estimating L(Θ̂k (G)) is just to map injectively every model in Mn/k to an
integer and then encode integers with l∗ (| Mn/k (G) |) as an upper bound of the
code-length. Here
l∗ (m) = max(0, log (m)) + max(0, log log(m)) + · · · , m ∈ N,
(7)
gives, as shown by Rissanen, the shortest length prefix coding for integers (see
[6, 19]). The size of the graph must also be encoded with l∗ (n) bits (it is assumed
that there is a way of defining an upper bound of the models with given n). In
this point, it is necessary that the modeling space is finite. This results in
8
Proposition 3.2 For any graph G ∈ Ωn , there exists a prefix coding with codelength
l(G) = ⌈− log P (G | Θ̂k (G))⌉ + m
∗
≤ mk := l (n) + l
∗
S2(n, k)
!
(k2)+k
n−k+2
+1
+ 1 + 1,
2
where S2(n, k) is the Stirling number of the second kind.
Proof The expression in (6) corresponds to a concatenation of two binary
codes. The L-part is the length of a code for maximum likelihood parameters
(in the case of a non-unique maximum, we take, say, the one with smallest
number in the enumeration of all such models). The corresponding code is called
the parametric code. The parametric code uniquely encodes the maximumlikelihood model. To create such an encoding, we just enumerate all possible
models, given in Definition 2.9, and use the integer to fix the model. The length
of a prefix code corresponding to an integer is the l∗ -function computed for that
integer, and we add 1 to handle the ceiling function.
To obtain an upper bound for the parametric code length mk , we find an
upper bound for the number of models in the modeling space. The number of
models is upper-bounded by the product of the integers, each presenting the
number of partitions of an n-element set into k non-empty sets (blocks), which
equals S2(n, k). We can view the blocks of a partition as the nodes of a ‘reduced
multi-graph’ (in a multi-graph, there can be several links between a node pair,
as well as self-loops). The range of multi-links is between zero and n−k+2
:
2
if we consider a pair of blocks (or one block internally), there can be at most
n − (k − 2) nodes in such a pair (in one set, slightly less), since there must be
at least k − 2 nodes in the other blocks of the partition. Obviously,
in such a
subgraph of n − (k − 2) nodes there can be at most n−k+2
links. Thus, the
2
number of values each multi-link can take is upper-bounded by n−k+2
+ 1.
2
Since the number of node pairs in the reduced multi-graph is k2 + k, we obtain
the second multiplier in the argument of l∗ in the proposition.
Finally, we should show that the coding of the graph is prefix. We concatenate both parts into one code that has the prescribed length and put first the
prefix code of the integer that defines the parameters of the maximum likelihood model. When we start to decode from the beginning of the entire code, we
first obtain a code of an integer, because we used a prefix coding for integers.
At this stage we are able to define the probabilistic model that was used to
create the other part of the code, corresponding to the probability distribution
P (· | Θ̂k (G)). Using this information we can decode the graph G. It remains
to show that the concatenated code itself is prefix. Assume the opposite: some
prefix of such a code is prefix to some other similar code, say, the first code is
a prefix to the second one. However, the parametric code was prefix, so both
codes must correspond to the same model. Since the first two-part code is a
prefix to the second, they both share the same parametric part, and the code for
the graph of the first is a prefix of the second one. But this is impossible, since
the encoding for graphs within the same model is prefix. This contradiction
shows that the two-part coding is prefix.
9
Finally, we call
Mn :=
[
1≤k≤n
Mn/k
(8)
the full regular decomposition modeling space of Ωn .
3.2
Two-part MDL for matrices
In this section we consider input data in the form of a n × m matrix A = (aij )
with non-negative entries. With such a matrix we associate a random bipartite
multi-graph. The set of rows and the set of columns form a bipartition. Between
row i and column j there is a random number of links that are distributed
according to Poisson distribution with mean ai,j . Such a model was introduced
in [13] and it has been used in various tasks in complex network analysis, see
[8]. The aim of this model is to back up, heuristically, a corresponding practical
algorithm for regular decomposition of matrices. Our approach is closely related
to but slightly different from the Poissonian block model. Assume that A is
used to generate random n × m matrices X with independent integer-valued
elements following Poisson(aij ) distributions. The target is to find a regular
decomposition model that minimizes the expected description length of such
random matrices. We propose the following modeling spaces:
Definition 3.3 For integers k1 , k2 from ranges 1 ≤ k1 ≤ n and 1 ≤ k2 ≤ m,
the parameters of a model Θk1 ,k2 in the modeling space Mk1 ,k2 for an integer
matrix X are partition of rows into k1 non-empty sets V = (V1 , · · · Vk1 ) and
partition of columns into k2 non-empty sets U P
= (U1 , · · · , Uk2 ) and k1 × k2
xi,j
.
block average matrix P , with elements (P )α,β := i∈Vα ,j∈Uβ |Vα ||U
β|
Thanks to the addition rule of Poisson distributions, the likelihood of X in
a model Θk1 ,k2 ∈ Mk1 ,k2 , corresponds to probabilistic models where the elements of X are independent and Poisson distributed with parameters xi,j ∼
P oisson(Pα(i),β(j) ), where i ∈ Vα(i) , j ∈ Uβ(j) in the model Θk1 ,k2 . The corresponding likelihood is denoted as P (X | Θk1 ,k2 ), the actual probability of X
is denoted as P (X | A). The maximum likelihood model is found from the
program that maximizes the expected log-likelihood:
X
Θ∗k1 ,k2 = arg max
P (X | A) log P (X | Θk1 ,k2 )
Θk1 ,k2 ∈Mk1 ,k2 X
=
arg max
Θk1 ,k2 ∈Mk1 ,k2
=
arg max
Θk1 ,k2 ∈Mk1 ,k2
X
X
!
P (X|Θk1 ,k2 )
+ P (X|A) log P (X|A)
P (X|A) log
P (X|A)
(−D(PA || PΘk1 ,k2 ) − H(PA ))
where D is the Kullback-Leibler divergence between distributions, H denotes
entropy and PA and PΘk1 ,k2 are the two families of Poisson distributions for the
matrix elements of X. Since H(PA ) is independent on Θk1 ,k2 , it does not affect
the identification of the maximum likelihood model. Thus, the final program
for finding the optimal model is
Θ∗k1 ,k2 = arg min (D(PA || PΘk1 ,k2 )).
Θk1 ,k2 ∈M
10
(9)
The description length of a model l(Θk1 ,k2 ∈ Mk1 ,k2 ) consist of the description
length l(V ) + l(U ) of the two partitions and the description length of the block
average matrix l(P (X)). For the latter we need to know only the integers
presenting the block sums of X, since the denominator is known for a fixed
partition (U, V ). The code lengths of such integers are, for large matrices,
simply the logarithms of the integers. For l(U ) + l(V ) we use the same entropy
based formula as in (14). As a result we end up with the following expression for
the description length of the random multi-graph model A using the modeling
space Mk1 ,k2 :
lk1 ,k2 (A) = D(PA || PΘ∗k ,k ) + l(V ∗ ) + l(U ∗ )
1 2
X
E(log(eα,β + 1 | PΘ∗ k
+
1 ,k2
),
1≤α≤k1 ;1≤β≤k2
where
X
eα,β =
xi,j .
i∈Vα∗ ,j∈Uβ∗
The star superscript refers to parameters corresponding to the solution of the
program (9). The expectation of logarithm is not explicitly computable. However, we assume large matrices and blocks, and then Jensen’s inequality provides
a tight upper bound that can be used in practical computations. Thus, the final
expression for the description length of A is
X
lk1 ,k2 (A) = D(PA || PΘ∗k ,k ) + l(V ∗ ) + l(U ∗ ) +
log(aα,β + 1), (10)
1
2
1≤α≤k1 ;1≤β≤k2
where
aα,β =
X
ai,j .
i∈Vα∗ ,j∈Uβ∗
The full two-part MDL would now be realized by finding the global minimum
of this expression over various (k1 , k2 ). We return to this case in the algorithm
section 4.2. Although a heuristic one, we believe that our method for matrices
is both reasonable and easy to use and implement, see [18].
3.3
Refined MDL and asymptotic model complexity
Let us next consider Rissanen’s refined MDL variant (see [6]). The idea is
to generate just one distribution on Ωn , called the normalized maximum likelihood distribution Pnml . Then a graph G ∈ Ωn has the description length
− log Pnml (G) which is at most as large as the one given by the two-part code
in (6). The function P (· | Θ̂k (·)) maps
P graphs of size n into [0, 1], and it is not
a probability distribution, because G∈Ω P (G|Θ̂k (G)) > 1. However, a related
true probability distribution can be defined as
Pnml (·) = P
P (· | Θ̂k (·))
G∈Ω P (G|Θ̂k (G))
.
(11)
The problem with this is that a computation of the normalization factor
in (11) is far too involved: finding a maximum likelihood parametrization for a
11
single graph is a ‘macroscopic’ computational task by itself and it is not possible
to solve such a problem explicitly for all graphs. Therefore the two-part variant
is a more attractive choice in a practical context. However, the refined MDL
approach is useful as an idealized target object for justifying various approximations implementations of the basic idea. It appears that in an asymptotic
sense the problem is solvable for large simple graphs. The logarithm of the normalization factor in (11) is called the parametric complexity of the model space
Mn/k :
!
X
COM P (Mn/k ) := log
P (G | Θ̂k (G)) .
(12)
G∈Ωn
In a finite modelling space case like in ours, this can be considered as a definition
of model complexity. We have now the following simple bounds:
Proposition 3.4
log (S2(n, k)) ≤ COM P (Mn/k ) ≤ mk + 1,
where we use the same notation as in Proposition 3.2.
Proof The lower bound follows from the fact that we can have at least this
number of graphs that have likelihood 1 in Mn/k . This corresponds to graphs for
which the nodes can be partitioned into k non-empty sets and inside each set we
have a full graph and no links between the sets. Thus, for every partition there
is at least one graph that has likelihood one and all such graphs are different
from each other since there is a bijection between those graphs and partitions.
For the upper bound, we notice that according to Proposition 3.2, there is
a prefix coding with code lengths that
P correspond to the two-part code. As a
result, Kraft’s inequality yields that G∈Ωn 2−lk (G) ≤ 1, or
X
X
1≥
2−⌈− log P (G|Θ̂k (G))⌉−mk ≥
2log P (G|Θ̂k (G))−1−mk ,
G∈Ωn
G∈Ωn
from which we get
X
G∈Ωn
P (G | Θ̂k (G)) ≤ 2mk +1 .
Taking logarithms, we arrive at the claimed upper bound.
When considering large-scale structures corresponding to moderate k, the
upper and lower bounds in Proposition 3.4 are asymptotically equivalent, and
we have
Corollary 3.5 Assume that k > 1 is fixed. Then
COM P (Mn/k ) ∼ n log k,
n → ∞.
Proof Denoting the lower and upper bound of parametric complexity in Proposition 3.4 respectively by bl and bu , we argue that bu ∼ bl ∼ n log k asymptotically when n → ∞. This follows from the fact that the dominant asymptotic
n
component of both bu and bl is log S2(n, k). Indeed, S2(n, k) ∼ kk! for fixed k,
the asymptotic of log S2(n, k) is linear in n, and all other terms of the asymptotics of both bounds are additive and at most logarithmic in n.
12
Remark 3.6 The speed of convergence of the upper and lower bounds in Proposition 3.4 is of type log n/n.
4
The Regular Decomposition approach to
stochastic block models
4.1
Block model codes
The previous section developed both the two-part and refined variants of the
MDL theory, as presented in [6], for the model space of stochastic block models.
In the following, we formulate a variant of two-part MDL that allows both practical implementations and a proof of consistency, i.e., that the MDL principle
identifies a block model. It was shown above that the most heavy task in the
description of a block model is identifying the partition. The same is true for
the model complexity, which is asymptotically just the logarithm of the number
of partitions. It appears that in order to prove consistency, we need quite a
delicate estimate for the description length of the partition. The asymptotic
model complexity given in Corollary 3.5 seems to be too crude for consistency.
A full resolution of this intriguing question is left for further investigations.
We call our two-part MDL construction a block model code of a graph with
respect to a partition of its nodes that allows the computation of a tight upper bound of the code length. This upper bound is also consistent with a
more generic information theoretic point of view with a semi-constructive coding scheme.
We denote by H(·) both Shannon’s entropy function of a partition and the
entropy of a binomial distribution, i.e.
H(ξ) = −
X |A|
|A|
log
,
|V |
|V |
A∈ξ
H(p) = −p log p − (1 − p) log(1 − p).
Remark 4.1 In the rest of this paper, we define also information-theoretic
functions in terms of natural logarithms, and certain notions like code lengths
should be divided by log 2 to obtain their values in bits.
Definition 4.2 A block model code of a graph G = (V, E) with respect to a
partition ξ of V is a code with the following structure:
The model part:
• first, the sizes of the blocks A ∈ ξ are given as integers;
• second, the edge density d(A) inside each block A ∈ ξ and the edge density
d(A, B) between each pair of distinct blocks A, B ∈ ξ are given as the
denominators of the rational numbers presenting the exact densities.
The aim of these two codes is to describe the parameters of two probability distributions, one for the links and the other for the membership of nodes in the
blocks of the partition.
The data part:
13
• third, the partition ξ is specified by a prefix code corresponding to membership distribution P (i ∈ A) = |A|/n, where all nodes are independent of
each other;
• fourth, the edges inside each block A ∈ ξ are specified by a prefix code
corresponding to a stochastic block model distribution of links inside each
block of ξ;
• fifth, the edges between each pair of blocks A, B ∈ ξ are specified by a prefix
code corresponding to a block model distribution of links between pairs of
blocks in ξ.
The choice of link densities as link probabilities (the second code) is natural,
since conditionally to a partition the stochastic block model is just a collection
of Bernoulli models, where the best choice is to use averages as parameters.
Note that a block model code can be given for any graph with respect to any
partition of its nodes.
From Kraft’s inequality and the above definitions, it follows that there exists
a prefix code for a graph G = (V, E) with respect to a partition ξ = {A1 , . . . , Ak }
of V with length at most (and, for large graphs, typically close to)
L(G|ξ) = L1 (G|ξ) + L2 (G|ξ) + L3 (G|ξ) + L4 (G|ξ) + L5 (G|ξ),
L1 (G|ξ) =
k
X
i=1
L2 (G|ξ) =
k
X
i=1
l∗ (|Ai |),
l∗
X
|Ai |
l∗ (|Ai ||Aj |d(Ai , Aj )) ,
d(Ai ) +
2
i<j
(13)
L3 (G|ξ) = |V |H(ξ),
k
X
|Ai |
H(d(Ai )),
L4 (G|ξ) =
2
i=1
X
|Ai ||Aj |H(d(Ai , Aj )),
L5 (G|ξ) =
i<j
where l∗ (m) was defined by (7). Below we shall approximate l∗ (m) by log m
without further mentioning, because their difference is insignificant in our context. Similarly, we have dropped ceiling functions systematically. Also, recall
Remark 4.1 on the use of natural logarithms.
Next, we shall extend the definition of block model codes to Poissonian block
models. Because the entries of E are integers, the matrix E can be encoded
in the same way as the simple graphs in Definition 4.2. Denote by HP (a) the
entropy of the distribution Poisson(a):
Hp (a) = −
∞
X
ak
k=0
k!
e−k (k log a − log k! − a).
Now, the Poissonian block model code length of E with respect to any partition
14
η = {B1 , . . . , Bm } is defined as
L(E|η) = L1 (E|η) + L2 (E|η) + L3 (E|η) + L4 (E|η) + L5 (E|η),
m
X
log(|Bi |),
L1 (E|η) =
i=1
m
X
X
|Bi |
log (|Bi ||Bj |a(Bi , Bj )) ,
log
a(Bi ) +
L2 (E|η) =
2
i<j
i=1
(14)
L3 (E|η) = |V |H(η),
m
X
|Bi |
HP (a(Bi )),
L4 (E|η) =
2
i=1
X
|Bi ||Bj |HP (a(Bi , Bj )),
L5 (E|η) =
i<j
where
a(Bi ) :=
P
v,w∈Bi
evw
|Bi |(|Bi | − 1)
,
a(Bi , Bj ) :=
P
v,w∈Bi
evw
|Bi ||Bj |
.
Remark 4.3 Definition (14) parallels (13), and it works smoothly in Theorem
4.13 below. However, the expressions of L4 and L5 correspond to the case that
each set of random variables is indeed Poisson distributed with the same parameter, which is not the case when η is not a refinement of ξn . The definition
given in Section 3.2 is based strictly on the notion of description length, and it
is also applied in the practical algorithm in Section 4.2.
4.2
Algorithms for description length minimization
In this section we present algorithms that we have used in actual computations of regular decompositions of graph and matrix data. These are written
for standard two-part MDL, where the code lengths L4 and L5 have a usual
interpretation as a minus log-likelihood of a graph corresponding to a stochastic
block model.
Thus, we use link coding lengths found in the upper bound of Proposition 3.2.
In many cases, this is all that can be computed realistically. We can obviously
describe a partition into k nonempty sets using an n × k binary matrix with all
row sums equal to one and requiring that none of the column sums equals zero.
The space of all such matrices we denote as Rk and the members of this set as
R ∈ Rk .
Definition 4.4 For a given graph G ∈ Ωn with adjacency matrix A and a
partition matrix R ∈ Rk , denote
P1 (R) := RT AR,
where ·T stands for matrix transpose, the column sums of R are denoted as
nα := (RT R)α,α ,
1 ≤ α ≤ k,
and the number of links within each block and between block pairs as
1
eα,β (R) = (1 − δα,β )(P1 (R))α,β .
2
15
Then define
(P (R))α,α := 1{nα >1}
eα,β (R)
,
nα
(P (R))α,β :=
2
eα,β (R)
, α 6= β.
nα nβ
Then the coding length of the graph corresponding to A using the model R
is:
Definition 4.5
lk (G(A) | R ∈ Rk ) :=
X
ni nj H((P (R))i,j ) +
1≤i<j≤k
1≤i≤k
+ lk (R),
where
lk (R) =
X
X ni
H((P (R))i,i )
2
ni H(ni /n) +
1≤i≤k
X
l∗ (ei,j (R))
1≤i≤j≤k
is the code length of the model, according to our theory and notation.
The two-part MDL program of finding the optimal model, denoted as
Rk∗ , can now be written as:
(k ∗ , Rk∗ ) := arg min min lk (G(A) | R ∈ Rk )
1≤k≤n R∈Rk
(15)
To solve this program approximately, we can use the following greedy algorithm.
Algorithm 4.6 Greedy Two-part MDL
Input: G = G(A) ∈ Ωn a simple graph of size n.
Output: (k ∗ , Rk∗ ∈ Rk∗ ), such that the two-part code for G is shortest possible
for all models in Mn by using this pair as a model.
Start: k = 1, l∗ = ∞, R ∈ Rn = {I}, k ∗ = 1, where I is denotes the n × n unit
matrix.
1. Find
R̂k (G) := arg min(lk (G | R)
R∈Rk
using subroutine ARGMAX k (Algorithm 4.8).
2. Compute lk (G) = ⌈lk (G | R̂(G))⌉ + lk (R̂(G))
3. If lk (G) < l∗ then l∗ = lk (G), Rk∗ = R̂k (G) , k ∗ = k
4. k = k + 1
5. If k > n, Print (Rk∗ , k ∗ ) and STOP the program.
6. GoTo 1.
Definition 4.7 A mapping Φ : Rk → Rk is defined as follows:
(LogP (R))α,β := log(P (R))α,β ,
(1 − LogP (R))α,β := 1 − log(P (R))α,β ,
L(R) := −AR(LogP (R))T − (1 − A)RLog(1 − P (R)),
where we set log 0 = 0 (all log 0 values will be later multiplied by 0),
β(i, R) := inf{β : β = arg min(L(R))i,α }, 1 ≤ i ≤ n,
1≤α≤k
and finally
Φ(R)i,α = δα,β(i,R) .
16
The mapping Φ(R) moves each node to a possibly different block such that
the description length would be minimized if all other nodes stay in their current
blocks.
Algorithm 4.8 ARGMAX k
Algorithm for finding optimal regular decomposition for fixed k
Input: A: the adjacency matrix of a graph (an n × n symmetric binary matrice
with zero trace); N : an integer (the number of iterations in the search of a global
optimum); k: a positive integer.
Start: m = 1.
1. i := 0; generate a uniformly random element Ri ∈ Rk .
2. If at least one of the column sums of Ri is zero, GoTo 1. Otherwise, set
Ri+1 := Φ(Ri ).
3. If Ri+1 6= Ri , set i := i + 1 and GoTo
Pn 2.
4. R(m) := Ri ; m = m + 1; l(m) := i=1 min1≤α≤k (L(R(m)))i,α .
5. If m < N , GoTo 1.
6. M := {m : l(m) ≤ l(i), i = 1, 2, ..., N }; m∗ := inf M .
Output optimal solution: R(m∗ ).
For very large graphs, the program may not be solvable in the sense that it is
not possible and reasonable to go through all possible values of k ∈ {1, 2, · · · , n}.
One option is to limit the range of k. In case that no minimum is found, then
use as an optimal choice the model found for the largest k within this range.
Another option is to find the first minimum with smallest k and stop. When
the graph is extremely large, it makes sense to use only a randomly sampled
sub-graph as an input — indeed, when k ∗ << n, a large-scale structure can be
estimated from a sample.
Our algorithm for Poissonian block models and matrices is essentially similar,
with certain differences in formulae as detailed below. Algorithms for other
cases like directed graphs and non-quadratic matrices are written very similarly,
although two partitions are needed, one for rows and one for columns. The logic
of the solution remains the same however.
A semi-heuristic two-part MDL algorithm for finding a regular decomposition for an n × m matrix A with non-negative entries works as follows. The
decomposition takes a form of a bi-clustering: there are two partitions, one for
rows and one for columns. Such partitions are described by binary matrices
with row sums equal to one. The two-part MDL program to find an optimal
regular decomposition is written as follows:
The row partition-matrices are denoted as R ∈ Rk1 with dimensions n × k1 ,
1 ≤ k1 ≤ n, and the column partition matrices as C ∈ Ck2 with dimensions
m × k2 , 1 ≤ k2 ≤ m.
Let us formulate the cost function for the matrix case that is derived from
Eq (9). The number of matrix elements in row group α and column group β
can be written as a matrix element:
(N )α,β = (RT R)α,α (C T C)β,β := nα mβ .
Assuming that all blocks are non-empty, all Nα,β > 0, we can define an average
matrix element of block α, β. First compute the sum of all matrix elements of
17
A over such a block:
eα,β = (RT AC)α,β .
The corresponding block averages form a k1 × k2 P -matrix with elements
(P )α,β =
eα,β
.
(N )α,β
The coding length of the matrix A using a two-part MDL code with partitions
(R, C) can be written as
lk1 ,k2 (G(A) | R ∈ Rk1 , C ∈ Ck2 ))
X
=
eα,β (1 − log((P )α,β ) + l∗ ([eα,β ])
1≤α≤k1 ,1≤β≤k2
+
X
nα H
1≤α≤k1
X
mβ
nα
+
+ k1 × k2 c.
mβ H
n
m
1≤β≤k2
Here we assume a similar handling of log 0’s as in the binary case. nα and mβ
are the sizes of row and column blocks, and [eα,β ] denotes the integer part of
eα,β ; it is assumed that such block sums are large numbers with finite decimal
precision c > 0. The description length of such decimals is the last term and it
is small compared with other terms for large matrices and can be safely ignored.
Similarly to the binary case, the two-part MDL program is defined as:
Definition 4.9
(k1∗ , k2∗ , R, C)
:= arg min
(k1 ,k2 )
min
R∈Rk1 ,C∈Ck2
lk1 ,k2 (G(A) | R ∈ Rk1 , C ∈ Ck2 )
(16)
where
1 ≤ k1 ≤ n,
1 ≤ k2 ≤ m.
The greedy algorithm of solving this program is very similar to the case of a binary matrix. The difference is that two parametric sequences of partitions must
be searched, Ri and Cj , and in the subroutine that finds the optimal partitions
for fixed i and j. One may consider different strategies in the corresponding
search of an optimal pair. For instance, moving first along the diagonal i = j,
and finding the value where the cost function (coding length of A) has a kneepoint, and after that make an off-diagonal search near that value. Another
option could be moving along the steepest descent direction of the cost function, or alternating the directions of increments in i and j, until a saturation is
reached in one direction, and then keeping that parameter fixed and finding the
optimum on the second parameter. This question is a subject to further experimenting with real and artificial data. Therefore we write only the subroutine
that finds the optimal partitioning with fixed k1 and k2 in a greedy fashion.
First we need
Definition 4.10 Define the mappings ΦR : Rk1 × Ck2 → Rk1 and ΦC : Rk1 ×
Ck2 → Ck2 as follows. Let E be an n × m matrix with all elements equal to
18
1. Then, using the definition of P -matrix and LogP -matrix (related to P as in
binary case), define two matrices using a block-matrix notation:
CP T
L(R, C) = E A
−C(LogP )T
and
M (R, C) = E T
AT
Define
RP
−RLogP
.
β1 (i, R, C) = inf{β : β = arg min(L(R, C))i,α },
1 ≤ i ≤ n,
β2 (i, R, C) = inf{β : β = arg min(M (R, C))i,α },
1 ≤ i ≤ m.
1≤α≤k1
and
1≤α≤k2
Then,
ΦR (R, C)i,α = δα,β1 (i,R,C) ,
1 ≤ α ≤ k1 ,
1 ≤ i ≤ n,
ΦC (R, C)i,α = δα,β2 (i,R,C) ,
1 ≤ α ≤ k2 ,
1 ≤ i ≤ m.
and
The main greedy subroutine is:
Algorithm 4.11 ARGMAX (k1 , k2 )
Algorithm for finding optimal regular decomposition for fixed (k1 , k2 ).
Input: A: a real n × m matrix with non-negative entries; N : a positive integer
(the number of iterations in the search of a global optimum); (k1 , k2 ): a pair of
positive integers.
Start: m = 1.
1. i = 0; generate uniformly random elements Ri ∈ Rk1 . and Ci ∈ Ck2 .
2. If at least one of column sums of Ri or Ci is zero, GoTo 1. Otherwise, set
Ri+1 := ΦR (Ri , Ci ),
Ci+1 := ΦC (Ri , Ci ).
3. If Ri+1 6= Ri or Ci+1 6= Ci , set i := i + 1 and GoTo 2.
4. R(m) := Ri ; C(m) := Ci ; m := m + 1;
1 ≤ α ≤ k1 ,
eα,β := T T (m)AC(m) α,β ,
T
T
Nα,β := R (m)R(m) α,α C (m)C β,β ,
1 ≤ α ≤ k1 ,
X
eα,β
.
eα,β 1 − log
l(m) :=
Nα,β
1 ≤ β ≤ k2 ;
1 ≤ β ≤ k2 ;
1≤α≤k1 ,1≤β≤k2
5. If m < N , GoTo 1.
6. M := {m : l(m) ≤ l(i), i = 1, 2, ..., N }; m∗ := inf M .
OUTPUT optimal solution: (R(m∗ ), C(m∗ )).
Remark 4.12 It is also possible to find a regular decompositions in the case
of partly missing matrix elements and also in the case of mixed positive and
negative entries. In the latter case, we can use the idea of directed links already
used in[12]. In the first case, we note that the main characteristic of the regular
decomposition is the P -matrix with elements that are averages of the data matrix
over large blocks that can be estimated, in many cases, despite a portion of data
is missing.
19
4.3
Accuracy of block structure identification by MDL
The general idea of the Regular Decomposition method is to be a generic tool
for separating structure and randomness in large data sets of graph or matrix
form. A partition of a real-world data set that minimizes the (nominal) code
length given in (13) resp. (14) can often not be compared with a ‘true solution’
for the simple reason that there may not be any objective notion of a ‘true
structure’ of the data. However, it is important to analyse and understand
how the method performs when the data really originates from a stochastic
block model. This question is called the consistency of MDL. Our results on
this question are summarized in the following theorem, formulated in terms of
the asymptotic behavior of a model sequence as specified by Construction 2.4.
In such a framework, an event is said to happen with high probability, if its
probability tends to 1 when n → ∞.
Theorem 4.13 Consider a sequence of stochastic block models (Gn , ξn ) based
on a vector (γ1 , . . . , γk ) of relative block sizes and a matrix D = (dij )ki,j=1 of
link probabilities, as described in Definition 2.3 and Construction 2.4. With
high probability, the following hold:
(i) Among all partitions η of Vn such that |η| ≤ k, ξn is the single minimizer
of L(Gn |η).
(ii) For any fixed ǫ ∈ (0, mini γi ), ξn is the single minimizer of L(Gn |η) among
partitions η with minimal block size higher than nǫ.
(iii) No refinement η of ξn with |η| ≤ m improves L(Gn |ξn ) by more than
const(k, m) log n.
The corresponding claims hold for the Poissonian block model mutatis mutandis.
Proof The proof is given in the next section and structured into several propositions. Claim (i) follows from Proposition 5.1, Proposition 5.2 and Corollary
5.6. Claim (ii) is Proposition 5.7. Claim (iii) is Proposition 5.5.
The results of Section 5 offer a richer picture than what was distilled into
Theorem 4.13. For example, Proposition 5.1 shows that if |η| = k and η differs
from ξn only a little, then the remaining misplaced nodes can be immediately
identified by computing their effect to the value of L(Gn |η). On the other hand,
we have not been able to exclude the possibility that a refinement of η could
yield a slight O(log n) improvement of the code length.
Remark 4.14 It is rather obvious that with large n, the identification of the
block structure ξn is robust against independent noise. The simplest case is that
the Poissonian block model is disturbed by additive Poissonian noise to each
matrix element:
ẽij = eij + φij ,
where the φij s are i.i.d. with φij ∼ Poisson(ν), ν > 0. Then (ẽij ) is again
an irreducible Poissonian block model with the same partition. More interesting
cases are binary flips in the graph case and multiplicative noise with mean 1 in
the case of non-negative matrices. We leave these for forthcoming work.
20
4.4
ǫ-regularity vs. stochastic block models
Although the structure that a MDL-based algorithm finds typically looks like
an ǫ-regular structure, there is a principal difference. In particular cases, an
ǫ-regular graph can have a structure that allows much better compression that
that provided by the ǫ-regular partition. In this section we give an explicit
example of such a case.
An important point in SRL is that for any ǫ > 0, there is an upper bound for
the size of regular partition, M (ǫ) so that for any graph with size above some
finite threshold N (ǫ), all such graphs have a regular partition with at most
M (ǫ) sets. Based on this, we show that the ǫ-regular structure of SRL and the
structure induced by the MDL need not coincide. Let us fix an order of graph
2n, large enough so that SRL holds for some ǫ > 0, and that M (ǫ) < n1−α for
some fixed 0 < α < 1/2.
Proposition 4.15 There is a graph of order 2n such that it has a MDL structure with code length o(n2 ) and an ǫ−regular structure that allows only Θ(n2 )
code length.
Proof Take n large enough as prescribed above. Then construct a bipartite
graph (X, Y ) with |X| = |Y | = n such that n is divisible by nα with some
rational α ∈ (0, 1/2). Assume that bothP
parts of the
Pbipartition are further
partitioned into equal size blocks: X =
Xi , Y =
Yi , |Xi | = |Yi | = nα .
Define then a random graph Gp = (X, Y, E) as follows. For each pair (Xi , Yj ),
take e(Xi , Yj ) = ξi,j |Xi |Yj |, where ξi,j ∼ Ber(p) is a Bernoulli random variable
with parameter 0 < p < 1, and the variables for different pairs are independent.
Assume that there are no other edges.
We show that, with high probability, Gp is ǫ-regular with regular partition
{X,
such a structure has a coding length at least
Y }. In the M DL-approach,
n
2
H(p
+
o(1))
=
O(n
).
This
comes
from the log-likelihood part, and the o(1)
2
corresponds to very small deviations of link densities from the expected value
p that can be made arbitrarily small by increasing n. Now we check that the
ǫ-regularity of graph Gp has a positive probability, which implies that such an ǫregular pair exists (actually, it appears that this happens with high probability).
ǫ-regularity means that for any X ′ ⊆ X, Y ′ ⊆ Y , | X ′ |, | Y ′ |> ǫn, the link
density d(X ′ , Y ′ ) deviates from the link density of the pair, d(X, Y ), no more
than by ǫ. By definition,
P
′
′
i,j ξi,j | X ∩ Xi || Y ∩ Yj |
′
′
d(X , Y ) =
,
| X ′ || Y ′ |
and as a result the expectation is
P
Ed(X ′ , Y ′ ) =
Denote
xi,j :=
i,j
Eξi,j | X ′ ∩ Xi || Y ′ ∩ Yj |
| X ′ || Y ′ |
ξi,j | X ′ ∩ Xi || Y ′ ∩ Yj |
| X ′ || Y ′ |
= p.
The range
P of xi,j is interval [0, 1] of unit length. Hoeffding’s inequality yields
for S = xi,j that
2
P (| S − ES |> t) ≤ 2e
21
− Pt 2
b
i,j
,
|X ′ ∩X ||Y ′ ∩Y |
i
j
where bi,j =
is the range of variable xi,j . The denominator of the
|X ′ ||Y ′ |
exponent in the right hand-side of the Hoeffding inequality can be bounded as
n 2
X
1
1
1 X
n2α n2α
= 2 n2α .
| X ′ ∩ Xi |2 | Y ′ ∩ Yj |2 ≤
b2i,j ≤
2
2
(ǫn)
(ǫn)
nα
ǫ
By taking t = ǫ | X ′ || Y ′ |≥ ǫ(ǫn)2 we get for the link density:
P (| d(X ′ , Y ′ ) − p |> ǫ) ≤ 2e−2ǫ
8
n2(1−α)
.
Finally, since there are at most 4n pairs of subsets (X ′ , Y ′ ) from which to choose,
the probability that none of them violates regularity is lower bounded by
1 − 2 × 4n e−2ǫ
8
n2(1−α)
→ 1,
if the exponent has a positive power of n, and this happens when α < 1/2.
Thus, all large subsets have densities that deviate from expectation less than ǫ
with a probability tending to one. Thus, we have shown the ǫ-regularity of the
partition (X, Y ).
On the other hand, using MDL, we could reach the level of small sets Xi
and Yi , and the corresponding log-likelihood is zero. The model complexity is
o(n2 ), as can be easily seen from asymptotic formulas for the upper bound for
mk , with k = n1−α .
5
Proof of Theorem 4.13
Through this section, we consider a sequence (Gn |ξn ) of increasing versions of a
fixed stochastic block model based on a vector (γ1 , . . . , γk ) of relative block sizes
and a matrix D = (dij )ki,j=1 of link probabilities, as specified in Construction
2.4.
Consider partitions η of Vn . Denote
d(η, ξn ) =
1
max min |B \ A|.
n B∈η A∈ξn
Thus, d(η, ξn ) = 0 if and only if η is a refinement of ξn . If v ∈ Vn and B ∈ η,
denote by ηv,B the partition obtained from η by moving node v to block B (if
v ∈ B, then ηv,B = η).
Proposition 5.1 There is a number ǫ0 > 0 such that the following holds with
high probability: if |η| = k and d(η, ξn ) ≤ ǫ0 , then
if A ∈ ξn , B ∈ η,
1
|B \ A| ≤ ǫ0 and v ∈ A \ B, then L(Gn |ηv,B ) < L(Gn |η).
n
Proof Let ǫ, δ > 0 be small numbers and m a positive integer to be specified.
They can be chosen so that the following holds:
• ǫ is small so that η and ξn nearly overlap when d(η, ξn ) ≤ ǫ:
mnǫ ≤ δ min |A|;
A∈ξn
22
(17)
• all the differing link probabilities are widely separated in δ units:
δ≤
1
min {|dij1 − dij2 | : i, j1 , j2 ∈ {1, . . . , k} , dij1 6= dij2 } ;
m
(18)
• the empirical densities are close to their mean values: for any Ai , Aj ∈ ξn
(possibly i = j), we have, with high probability,
|d(Ai , Aj ) − dij | + mǫ ≤ δ.
(19)
Let η be a partition of Vn such that d(η, ξn ) ≤ ǫ. Condition (17) entails that for
each block Ai ∈ ξn there is a unique block Bi ∈ η such that |Bi \ Ai | ≤ ǫ. Let
us now assume that v ∈ Ai ∩ Bj and i 6= j, and compare the partitions η and
ηv,Bi . Denote bi = |Bi |, i = 1, . . . , k, and B̃i = Bi ∪ {v}, B̃j = Bj \ {v}. Then
L4 (Gn |η) + L5 (Gn |η) − (L4 (Gn |ηv,Bi ) + L5 (Gn |ηv,Bi ))
bi + 1
bj − 1
bj
bi
H(d(Bj )) −
H(d(B̃i )) −
H(d(Bi )) +
H(d(B̃j ))
=
2
2
2
2
Xh
+
bi bq H(d(Bi , Bq )) + bj bq H(d(Bj , Bq ))
q6=i,j
i
− (bi + 1)bq H(d(B̃i , Bq )) + (bj − 1)bq H(d(B̃j , Bq ))
+ bi bj H(d(Bi , Bj )) − (bi + 1)(bj − 1)H(d(B̃i , B̃j )).
(20)
Consider first the sum over q. Leaving out the common factor bq , each term of
the sum can be written as
bj − 1
H(d(B̃j , Bq ))
bj H(d(Bj , Bq )) −
bj
bi
H(d(Bi , Bq ))
− (bi + 1) H(d(B̃i , Bq )) −
bj + 1
1
bj − 1
d(B̃j , Bq ) + d({v} , Bq )
= bj H
bj
bj
bj − 1
1
−
H(d(B̃j , Bq )) − H(d({v} , Bq ))
bj
bj
1
bi
− (bi + 1) H
d(Bi , Bq ) +
d({v} , Bq )
bi + 1
bi + 1
1
bi
H(d(Bi , Bq )) −
H(d({v} , Bq ))
−
bi + 1
bi + 1
(note the addition and subtraction of the term H(d({v} , Bq ))). Using Lemma
A.7 and the assumptions on ǫ, δ and m, the last expression can be set to be,
with high probability, arbitrarily close to the number
I(djq : diq ) + I(diq : diq ) = I(djq : diq )
(the function I(·, ·), the Kullback-Leibler divergence of Bernoulli distributions,
is defined in (32)). Thus, the sum over q is, with high probability, close to
X
bq
I(djq , diq ).
q6=i,j
23
Let us then turn to the remaining parts of (20) that refer to two codings of
the internal links of Bi ∪ Bj . Similarly as above, we can add and subtract terms
to transform these parts into
!
"
bj −1
bj
b
−
1
j
2
H
d(B̃j ) + bj d({v} , B̃j )
bj
2
2
2
#
bj −1
bj − 1
2
− bj H(d(B̃j )) − bj H(d({v} , B̃j ))
2
2
"
bi + 1
−
H
2
bi
−
"
bi
2
d(Bi )
bi +1
2
2
H(d(Bi ))
bi +1
2
+ bi bj H
−
+
bi
bi
!
d({v} , Bi )
bi +1
2
#
H(d({v} , Bi ))
bi +1
2
1
bj − 1
d(Bi , B̃j ) + d({v} , Bi )
bj
bj
#
1
bj − 1
H(d(Bi , B̃j )) − H(d({v} , Bi ))
−
bj
bj
"
bi
1
− (bi + 1)(bj − 1) H
d(Bi , B̃j ) +
d({v} , B̃j )
bi + 1
bi + 1
#
bi
1
−
H(d(Bi , B̃j )) −
H(d({v} , B̃j ))
bi + 1
bi + 1
≈ (bj − 1)I(dij : djj ) − (bi + 1)I(dii : dii ) + bi I(dii : dij ) − (bj − 1)I(dij : dij )
= bj I(dij : djj ) + bi I(dii : dij ).
By the above analysis of (20), we have obtained
L4 (Gn |η) + L5 (Gn |η) − (L4 (Gn |ηv,Bi ) + L5 (Gn |ηv,Bi ))
X
≈ bq
I(djq : diq ) + bj I(dij : djj ) + bi I(dii : dij ).
(21)
q6=i,j
By the irreducibility assumption (1), there is a block Aq such that dqi 6= dqj ,
with the possibility that q ∈ {i, j}. It follows that at least one of the I(x : y)’s
in (21) is positive. Denote
κ∗ = min {I(dij1 : dij2 ) : dij1 6= dij2 } .
Thus, with high probability,
L4 (Gn |η) + L5 (Gn |η) − (L4 (Gn |ηv,Bi ) + L5 (Gn |ηv,Bi )) >
1 ∗
(κ min γi )n.
i
2
On the other hand, it is easy to compute that
L3 (Gn |η) − L3 (Gn |ηv,Bi ) = n(H(η) − H(ηv,Bi )) → log
γj
.
γi
The changes of L1 and L2 when moving from η to ηv,Bi are negligible. This
concludes the proof.
24
The proof of Proposition 5.1 showed that when d(η, ξn ) ≤ ǫ0 , moving any
node to its correct block decreases L(Gn |η) at least by (1/2)(κ∗ mini γi )n. In
particular, with high probability, ξn is the unique minimizer of L(Gn |η) among
k-partitions η satisfying d(η, ξn ) ≤ ǫ0 .
Proposition 5.2 For any ǫ ∈ (0, 1) and positive integer m, there is a constant
cǫ such that the following holds with high probability:
if |η| ≤ m and d(η, ξn ) > ǫ, then
1
(L(Gn |η) − L(Gn |ξn ∨ η)) ≥ cǫ .
n2
Proof Fix an ǫ ∈ (0, 1) and let η be a partition of Vn such that d(η, ξn ) > ǫ.
By the concavity of H, we have
L4 (Gn |η) + L5 (Gn |η)
X |B|
1
=
H(d(B)) +
2
2
B∈η
X
|B||B ′ |H(d(B, B ′ ))
B,B ′ ∈η
B6=B ′
X X |A ∩ B|
H(d(A ∩ B))
≥
2
B∈η A∈ξn
+
X1
2
B∈η
X
|A ∩ B||A′ ∩ B|H(d(A ∩ B, A′ ∩ B))
(22)
A,A′ ∈ξn
A6=A′
+
1
2
X
B,B ′ ∈η
B6=B
X
A,A′ ∈ξn
|A ∩ B||A′ ∩ B ′ |H(d(A ∩ B, A′ ∩ B ′ ))
′
= L4 (Gn |η ∨ ξn ) + L5 (Gn |η ∨ ξn ).
By assumption, there is B ∈ η such that |B \ A| > ǫn for every A ∈ ξn . It is
easy to see that there must be (at least) two distinct blocks, say Ai and Aj ,
such that
ǫ
n.
(23)
min {|Ai ∩ B|, |Aj ∩ B|} ≥
k−1
By the irreducibility assumption (1), there is a block Aq such that dqi 6= dqj ,
with the possibility that q ∈ {i, j}. Fix an arbitrary δ > 0 to be specified later.
By ǫ-regularity (claim (ii) of Lemma A.7), with high probability, every choice
of a partition η with |B \ A| > ǫn results in some blocks Ai , Aj , Aq with the
above characteristics plus the regularity properties
|d(Ai ∩ B, Aq ∩ B ′ ) − diq | ≤ δ,
|d(Aj ∩ B, Aq ∩ B ′ ) − djq | ≤ δ,
(24)
where B ′ denotes a block of η that maximizes |Aq ∩ B ′ | (note that because
25
|η| ≤ m, |Aq ∩ B ′ | ≥ |Aq |/m). By the concavity of H,
|Ai ∩ B||Aq ∩ B ′ |H(d(Ai ∩ B, Aq ∩ B ′ ))
+ |Aj ∩ B||Aq ∩ B ′ |H(d(Aj ∩ B, Aq ∩ B ′ ))
|Ai ∩ B|
H(d(Ai ∩ B, Aq ∩ B ′ ))
= |(Ai ∪ Aj ) ∩ B||Aq ∩ B ′ |
|(Ai ∪ Aj ) ∩ B|
|Aj ∩ B|
+
H(d(Aj ∩ B, Aq ∩ B ′ ))
|(Ai ∪ Aj ) ∩ B|
< |(Ai ∪ Aj ) ∩ B||Aq ∩ B ′ |H(d((Ai ∪ Aj ) ∩ B, Aq ∩ B ′ )).
In the case that q ∈ {i, j} and B = B ′ , we obtain a similar equation where
|Aq ∩ B| is partly replaced by |Aq ∩ B| − 1. Because of (24) and (23), the
difference between the sides of the equality has a positive lower bound that
holds with high probability. On the other hand, this difference is part of the
overall concavity inequality (22).
Proposition 5.3 For any refinement η of ξn , we have
(st)
L4 (Gn |ξn ) + L5 (Gn |ξn ) − (L4 (Gn |η) + L5 (Gn |η)) ≤
M(|η|)−M(k)
X
(log 2 + Yi ),
j=1
(25)
where (st) refers to stochastic order, the Yi ’s are i.i.d. Exp(1) random variables,
and
x(x + 1)
.
(26)
M (x) =
2
Proof Here we apply results presented in Appendix A. Denote by η ∩ Ai the
subset of η whose members are subsets of the block Ai of ξn . Writing the edge
code lengths of the coarser and finer partition similarly as in (22), taking the
difference and using (35), we obtain
L4 (Gn |ξn ) + L5 (Gn |ξn ) − (L4 (Gn |η) + L5 (Gn |η))
k
X |B|
X
X
1
I(d(B) : dii ) +
=
2
2
i=1
B∈η∩Ai
|B||B ′ |I(d(B, B ′ ) : dii )
B,B ′ ∈η∩Ai
B6=B ′
!
|Ai |
−
I(d(Ai ) : dii )
2
+
X
i<j
X
X
B∈η∩Ai B ′ ∈η∩Aj
′
′
!
|B||B |I(d(B, B ) : dij ) − |Ai ||Aj |I(d(Ai , Aj ) : dij ) .
(27)
Applying Proposition A.6 to each term of both outer sums now yields the claim,
because
k
X
i=1
(M (|η ∩ Ai |) − 1) +
X
i<j
(|η ∩ Ai ||η ∩ Aj | − 1) = M (|η|) − M (k).
26
Remark 5.4 It is rather surprising that the stochastic bound (25) depends only
on the number of blocks in η — not on their relative sizes, nor on the overall
model size n!
Proposition 5.5 For any positive integer m > k, the following holds with high
probability:
L(Gn |ξn ) −
min
η≥ξn , |η|≤m
L(Gn |η) ≤ (m + M (m − k)) log n,
where the relation η ≥ ξn means that η is a refinement of ξn , and M (·) was
defined in (26).
Proof Let η be a refinement of ξn . Refining the partition w.r.t. ξn yields a
gain, based on the concavity of H, in the code part L4 + L5 , but costs in the
parts L1 , L2 and L3 . We have to relate these to each other.
Consider first the value of L2 (Gn |η). In our analysis, it is important to
distinguish between ‘large’ and ‘tiny’ blocks, where the relative sizes of large
blocks exceed some pre-defined number ǫ and the rest can be arbitrarily small,
even singletons. Now, each block Ai ∈ ξn must contain at least one block
Bi ∈ η ∩ Ai such that |Bi | ≥ |Ai |/m. Define
ǫ := min |Aj |/(2mn).
j
Because no concavity gain can be obtained with an index pair {i, j} such that
dij ∈ {0, 1}, it does not restrict generality to assume that dij ∈ (0, 1) for all i, j.
Then
!
k
X
X
X
1
|B|
′
log(|B||B |dii )
L2 (Gn |η) =
log
dii +
2
2
i=1
B∈η∩Ai
B,B ′ ∈η∩Ai
B6=B ′
+
X
i<j
≥
k
X
i=1
+
X
i<j
X
B∈η∩Ai
log
X
B ′ ∈η∩A
2 2
ǫ n
dii
2
j
!
log(|B||B ′ |dij )
!
+ (|η ∩ Ai | − 1) log(ǫndii )
log ǫ2 n2 dij + (|η ∩ Ai | + |η ∩ Aj | − 2) log(ǫndij )
= (k|η| + k) log n + c1 (D, ǫ).
On the other hand, the bound |Ai | ≤ n yields
L2 (Gn |ξn ) ≤ (k 2 + k) log n + c2 (D).
Thus,
L2 (Gn |η) − L2 (Gn |ξn ) ≥ k(|η| − k) + c1 (D, ǫ) − c2 (D).
(28)
We obviously have also L1 (Gn |η) ≥ L1 (Gn |ξn ), but this difference is insignificant
in the present context.
27
The refinement gain in code parts L4 and L5 was bounded in Proposition 5.3
stochastically by Exp(1) random variables. The rate function (see the beginning
of Appendix A) of the distribution Exp(1) is
IE (x) = x − 1 − log x.
Denoting
Gainn (η) := L4 (Gn |ξn ) + L5 (Gn |ξn ) − (L4 (Gn |η) + L5 (Gn |η)),
Proposition 5.3 yields, for y > log 2,
P (Gainn (η) > y)
M(|η|)−M(k)
X
Yi > y − (M (|η|) − M (k)) log 2
≤P
j=1
y − (M (|η|) − M (k)) log 2
≤ exp −(M (|η|) − M (k))IE
M (|η|) − M (k)
(M(|η|)−M(k))
2e
≤ exp − y + (M (|η|) − M (k)) log y
,
M (|η|) − M (k)
where the second factor is bounded and will be henceforth neglected.
For two refinements of ξn , write η ′ ∼ η if the block sizes of η ′ in each Ai
are identical to those of η. The number of refinements η ′ of ξn with η ′ ∼ η is
upperbounded by
X
exp
|A|H(η|ξn ) = enH(η|ξn ) .
A∈ξn
On the other hand, we have
L3 (Gn |η) − L3 (Gn |ξn ) = nH(η) − nH(ξn ) = nH(η|ξn ).
Write
∆L123 (η) = L1 (Gn |η) + L2 (Gn |η) + L3 (Gn |η)
− (L1 (Gn |ξn ) + L2 (Gn |ξn ) + L3 (Gn |ξn )),
∆L(η) = L(Gn |η|) − L(Gn |ξn |).
Denote z(η) := (|η| + M (|η| − k)) log n. Recalling (28), the union bound yields
P sup Gainn (η ′ ) > ∆L123 (η) + z(η)
η ′ ∼η
≤ exp (nH(η|ξn ) − ∆L123 (η) − z(η) + (M (|η|) − M (k)) log(∆L123 (η) + z(η)))
≤ exp − z(η) + − k(|η| − k) + M (|η|) − M (k) log n + c3 (D, ǫ) .
The number of different block size sequences ℓ1 ≥ · · · ≥ ℓ|η| is upper bounded
28
by n|η|−1 = e(|η|−1) log n . Thus, a second application of the union bound yields
!
P
sup
∆L(η) > z(η)
η≥ξn , |η|≤m
≤ me − k(m−k) + M(m)−M(k) − m+M(m−k) + m−1 log n+const
const
→ 0, as n → ∞.
=
n
Corollary 5.6 ξn is the unique minimizer of L(Gn |η) among partitions η with
|η| = k.
Proof
By Proposition 5.1,
min
|η|=k, d(η,ξn )≤ǫ0
L(Gn |η) > L(Gn |ξn )
with high probability. On the other hand, Proposition 5.2 yields that, with high
probability,
min
|η|=k, d(η,ξn )>ǫ0
L(Gn |η) > L(Gn |ξn ∨ η) + cǫ0 n2 .
By Proposition 5.5,
min L(Gn |ξn ∨ η) > L(Gn |ξn ) − (k + M (k 2 − k)) log n
|η|=k
with high probability. It remains to note that n2 grows faster than log n.
Proposition 5.7 Let ǫ ∈ (0, mini γi ). Consider refinements η of ξn with relative minimal block size ǫ, i.e. the set
Bǫ(n) = {η ≥ ξn : |B| ≥ nǫ ∀B ∈ η} .
(29)
With high probability,
min
(n)
η∈Bǫ
\{ξn }
L(Gn |η) > L(Gn |ξn ).
(n)
(30)
Proof The restriction η ∈ Bǫ implies |η| ≤ ⌊1/ǫ⌋ =: m. The difference to
the conditions of Proposition 5.5 is now just the magnitude of L2 (η). When
η ∈ B (n) ,
L2 (Gn |η) ≥ M (|η|) log n2 + c(D, ǫ, |η|),
so that
L2 (Gn |η) − L2 (Gn |ξn ) ≥ 2(M (|η|) − M (k)) log n + c(D, ǫ, |η|).
29
By a corresponding computation as in the proof of Proposition 5.5, we obtain
(n)
for any fixed η ∈ Bǫ that
1
′
P sup Gainn (η ) > ∆L123 (η) − log n
2
η ′ ∼η
s
≤ exp (−(M (|η|) − M (k)) log n + const ) .
Proceeding with a second union bound like in the proof of Proposition 5.5 yields
!
1
P
min
L(Gn |η) ≤ L(Gn |ξn ) + log n
(n)
2
η∈Bǫ \{ξn }
1
≤ max exp
− (M (q) − M (k)) + q − 1 +
log n + const
q≤m
2
1
= exp − log n + const
2
const
= √ → 0, as n → ∞,
n
where the maximum over q was obtained with the smallest value q = k + 1.
A
Chernoff bounds and other information-theoretic
preliminaries
Consider a random variable X with moment generating function
φX (β) = E eβX ,
and denote DX = {β : φX (β) < ∞}. We restrict to distributions of X for which
DX is an open (finite or infinite) interval. The corresponding rate function is
IX (x) = − inf (log φX (β) − βx).
β∈DX
IX (x) is a strictly convex function with minimum at E X and value +∞ outside
the range of X. As a further restriction on the distribution of X, we assume
that IX (x) approaches smoothly +∞ P
both when x decreases or increases from
n
the point E X. For the mean X̄n = n1 1 Xi of i.i.d. copies of X, we have
IX̄n (x) = nIX (x).
The Chernoff bound (also known as Cramér-Lundberg bound):
Proposition A.1
P (X < x) ≤ e−IX (x)
P (X > x) ≥ e
−IX (x)
30
for x < E X,
for x > E X.
(31)
This has the following simple consequence. Let the support of X be the
closure of (x− , x+ ) and write
−
+
IX (x) = IX
(x)1{(x− ,E X]} (x) + IX
(x)1{[E X,x+ )} (x) + ∞ · 1{R\(x− ,x+ )} (x).
−
+
With the assumptions made above, the functions IX
(x) and IX
(x) are, respec−
+
tively, bijections from (x , E X] and [E X, x ) to [0, ∞).
Lemma A.2 For any z ≥ 0 we have
(st)
IX (X) ≤ log 2 + Y,
(st)
where ≤ denotes stochastic order and Y is a random variable with distribution
Exp(1).
Proof
P (IX (X) > z) = P 1{X<0} IX (X) > z or 1{X>0} IX (X) > z
= P 1{X<0} IX (X) > z + P 1{X>0} IX (X) > z
+(−1)
−(−1)
(z)
(z) + P X > IX
= P X < IX
−(−1)
≤ exp(−IX (IX
=e
−z
=e
−(z−log 2)+
+e
+(−1)
(z))) + exp(−IX (IX
(z)))
−z
,
where the inequality comes from Proposition A.1.
In the case that X has the Bernoulli(p) distribution, we have
IX (x) = I(x : p) := q log
1−q
q
+ (1 − q) log
.
p
1−p
(32)
Lemma A.3 The first and second derivatives of the functions H(x) and I(x :
p) are
H ′ (x) = log
1−x
,
x
1
,
x(1 − x)
1
I ′′ (x : p) =
.
x(1 − x)
H ′′ (x) = −
I ′ (x : p) = H ′ (p) − H ′ (x),
(33)
(34)
In particular,
H(q) − (H(p) + H ′ (p)(q − p)) = I(q : p),
and
1
1
1
1
lim n H (1 − )y + x − (1 − )H(y) + H(x)
n→∞
n
n
n
n
= (x − y)H ′ (y) − (H(x) − H(y))
= I(x : y).
31
(35)
Proposition A.4 Let n ≥ 2 and let X1 and X2 be independent random variables with distributions Bin(m, p) and Bin(n−m, p), respectively. Denote X12 =
X1 + X2 and X̄1 = X1 /m, X̄2 = X1 /(n − m), X̄12 = X12 /n. Then the following
identities hold:
mI(X̄1 : p) + (n − m)I(X̄2 : p) − nI(X̄12 : p)
m − X1 m
X1 m
+ (n − Xn )I
:
:
= Xn I
X12 n
n − X12 n
X12 − X1
= mI(X̄1 : X̄12 ) + (n − m)I
: X̄12 .
n−m
(36)
(37)
(38)
The identities in Proposition (A.4) are obtained by writing the full expression
of (36) and re-arranging the log terms in two other ways. Formulae (37) and (38)
are written without X2 , expressing the fact that any two of the three random
variables X1 , X2 and X12 contain same information as the full triple. Note
that (37) and (38) do not contain p. This reflects the fact that the conditional
distribution of X1 given X12 , known as the hypergeometric distribution, does
not depend on p. The identity of (36) and (38) can be interpreted so that the
two positive terms of (36) measure exactly same amount of information about
p as what is subtracted by the negative term. Moreover, (38) has the additional
interpretation of presenting the rate function of the hypergeometric distribution:
Proposition A.5 Let X have the distribution Hypergeometric(n, m, z), i.e. the
conditional distribution of X1 of Proposition A.4 given that X12 = z. The rate
function of X is
x z
z−x z
+ (n − m)I
.
(39)
:
:
IX (x) = mI
m n
n−m n
Proof
Define the bivariate moment-generating function of (X1 , X2 )
φ(α, β) = E eαX1 +βX2 .
Write
P[ X1 = m | X12 = z] =
P (X1 = m, X2 = z − m)
,
P (X12 = z)
and note that we can assume p = z/n. We can now derive the claim using
φ(α, β) in similar manner as in the well-known proof of the one-dimensional
Chernoff bound.
Proposition A.6 Let k ≥ 2 and let Xi , i ∈ {1, . . . , k}, be independent random
P
variables with distributions Bin(ni , p), respectively. Denote X1...j = ji=1 Xi ,
Pj
X̄i = Xi /ni and X̄1...j = X1...j / i=1 ni . Then
k
X
i=1
X
(st) k−1
ni I(X̄i : p) − nI(X̄12 : p) ≤
(log 2 + Yi ) ,
where Y1 , . . . , Yk−1 are independent Exp(1) random variables.
32
(40)
i=1
(41)
Proof
For k = 2, the left hand side of (40) equals
X12 − X1
: X̄12
n1 I(X̄1 : X̄12 ) + n2 I
n2
(42)
by Proposition A.4. For any N ∈ {0, . . . , n}, consider the conditional distribution of (42), given that X12 = N . By Proposition A.5, this is the distribution
of the Hypergeometric(n, n1 , N ) rate function taken at the random variable X1
with the same distribution. The claim now follows by Lemma A.2, because the
stochastic upper bound does not depend on N , i.e. on the value of X12 .
For k > 2 we proceed by induction. Assume that the claim holds for k − 1
and write
k
X
i=1
=
ni I(X̄i : p) − nI(X̄12 : p)
k−1
X
i=1
ni I(X̄i : p) − (n − nk )I(X̄1...(k−1) : p)
+ nk I(X̄k : p) + (n − nk )I(X̄1...(k−1) : p) − nI(X̄1...k : p).
By the induction hypothesis,
the first row of the second expression is stochastiPk−2
cally bounded by i=1 (log 2 + Yi ), irrespective of the value of X1...(k−1) . Similarly, the second row is stochastically bounded by log 2+Yk , where Yk ∼ Exp(1),
irrespective of the value of X1...k . It remains to note that Yk can be chosen to
be independent of (Y1 , . . . , Yk−2 ), because Xk is independent of (X1 , . . . , Xk−1 ),
and of X̄1...(k−1) in particular.
Lemma A.7 Consider the sequence (Gn , ξn ) of stochastic block models as in
Theorem 4.13. Then the following holds.
(i) For any blocks Ai and Aj such that dij 6∈ {0, 1}, it holds for an arbitrary
ǫ > 0 with high probability that
min
v∈Ai
e({v} , Aj )
1
≥ dij − n− 2 +ǫ ,
|Aj |
max
v∈Ai
e({v} , Aj )
1
≤ dij + n− 2 +ǫ .
|Aj |
(ii) For any ǫ > 0, the partition ξn is ǫ-regular with high probability.
Proof
Claim (i): By Proposition A.1 and (34),
e({v} , Aj )
P max
> dij + h
v∈Ai
|Aj |
X
e({v} , Aj )
> dij + h
≤
P
|Aj |
v∈Ai
≤ |Ai | exp (−|Aj |I(dij + h : dij ))
h2
h3 ′′′
= |Ai | exp −|Aj |
+ I (z : dij )
.
2dij (1 − dij )
6
33
1
The last expression converges to zero with the choice h = n− 2 +ǫ (recall that
|Ai | ∼ nγi and |Aj | ∼ nγj ), which proves the claim on the maximum. The case
of the minimum is symmetric.
Claim (ii): Fix ǫ > 0 and consider any i, j. Let U1 ⊆ Ai and U2 ⊆ Aj such
that |U1 | ≥ ǫ|Ai | and |U2 | ≥ ǫ|Aj |. By Proposition A.1,
P (|d(U1 , U2 ) − dij | > ǫ) ≤ e−|U1 ||U2 |I(dij +ǫ:dij ) + e−|U1 ||U2 |I(dij −ǫ:dij ) .
Let ι(ǫ) = min {I(dij + ǫ : dij ), I(dij − ǫ : dij )}. The union bound yields
P (∃U1 ⊆ Ai , U2 ⊆ Aj : |U1 | ≥ ǫ|Ai |, |U2 | ≥ ǫ|Aj |, |d(U1 , U2 ) − dij | > ǫ)
≤ 2|Ai ||Aj | exp (|Ai | + |Aj |) log 2 − ǫ2 |Ai ||Aj |ι(ǫ)
(γi + γj ) log 2
− γi γj ǫ2 ι(ǫ)
→ 0 as n → ∞,
≤ 2n2 exp n2
n
because ι(ǫ) > 0.
Remark A.8 Because ι(ǫ) ∝ ǫ2 as ǫ → 0, the proof of claim (ii) indicates that
with a fixed ǫ, ǫ-regularity starts to hold when n >> ǫ−4 .
Poissonian block model
Denote by HP (a) the entropy of the distribution Poisson(a)
Hp (a) = −
∞
X
ak
k=0
k!
e−k (k log a − log k! − a)
= a − a log a + e−a
∞
X
ak
k=0
k!
log k!,
(43)
(44)
and by IP (b : a) the Kullback-Leibler divergence between distributions Poisson(a)
and Poisson(b)
IP (b : a) = a − b + b log b − b log a.
(45)
We note the following facts:
Lemma A.9 The entropy function HP (α) is increasing and concave, and its
two first derivatives have the expressions
d
X +1
HP (α) = E log
,
dα
α
1
1
d2
,
HP (α) = − + E log 1 +
dα2
α
X +1
(46)
(47)
where X denotes a random variable with distribution Poisson(α).
Lemma A.10 For any α ∈ [0, 1] and x, y > 0, denote z = αx + (1 − α)y. Then
we have
HP (z) − (αHP (x) + (1 − α)HP (y)) ≤ αIP (x : z) + (1 − α)IP (y : z)
αx
= zIBer (α) ( )
z
= IBin (z,α) (αx),
where the form (50) is applicable when z is a positive integer.
34
(48)
(49)
(50)
Proof
By (44) and (45),
HP (z) − (αHP (x) + (1 − α)HP (y)) − (αIP (x : z) + (1 − α)IP (y : z))
= f (z) − (αf (x) + (1 − α)f (y)) ≤ 0,
because the function
f (u) = e−u
∞
X
uk
k=2
is convex:
f ′′ (u) = e−u
k!
log k!
∞
X
uk
k=1
k+2
log
+ log 2
k!
k+1
!
> 0.
The expression (50) is obtained by writing the right hand side of (48) with the
substitution y = (z − αx)/(1 − α) and re-combining the log terms.
The Poissonian counterpart of Proposition A.6 is the following.
P
Proposition A.11 Let a > 0, k ≥ 2, ni ≥ 1, i = 1, . . . , k, and n = i ni .
Let Xi , i ∈ {1, . . . , k}, be independent random variables with distributions
Pj
Poisson(ni a), respectively. Denote X1...j = i=1 Xi , X̄i = Xi /ni and X̄1...j =
Pj
X1...j / i=1 ni . Then
nHP (X̄12 ) −
k
X
i=1
X
(st) k−1
ni HP (X̄i ) ≤
(log 2 + Yi ) ,
(51)
i=1
(52)
where Y1 , . . . , Yk−1 are independent Exp(1) random variables.
Proof The proof of Proposition A.6 can be imitated, using the following observations:
• Using induction, it suffices to consider the case k = 2.
• By (48), the left hand side of (40) can be upperbounded in terms of
Kullback informations.
• The latter expression can be transformed to the form (50), i.e. the value
of a the rate function of a binomial distribution taken at random variable
X1 .
• Now, the conditional distribution of X1 given X12 is the above binomial
distribution. Thus, we can apply Lemma A.2 in a similar way as in the
proof of Proposition A.6.
Acknowledgment. This work was partly supported by Academy of Finland project 294763 (Stomograph).
35
References
[1] N. Alon, R. Duke, H. Lefmann, V. Rödl, and R. Yuster. The algorithmic
aspects of the regularity lemma. Journal of Algorithms, 16:80–109, 1994.
[2] N. Alon, E. Fischer, I. Newman, and A. Shapira. A combinatorial characterization of the testable graph properties: It’s all about regularity. In
STOC’06, Seattle, U.S., May 2006.
[3] M. Bolla. Spectral Clustering and Biclustering. Wiley, 2013.
[4] T. Cover and J. Thomas. Elements of Information Theory. John Wiley
and Sons, U.S.A., 1991.
[5] E. Fischer, A. Matsliah, and A. Shapira. Approximate hypergraph partitioning and applications. Foundations of Computer Science, pages 579–589,
2007.
[6] P. Grünwald. Minimum Description Length Principle. The MIT Press,
2007.
[7] S. Heimlicher, M. Lelarge, and L. Massoulié. Community detection in the
labelled stochastic block model. In NIPS Workshop on Algorithmic and
Statistical Approaches for Large Social Networks, 2012.
[8] R. v. d. Hofstad. Random graphs and complex networks. Cambridge University Press, 2016.
[9] P. Holland, K. Laskey, and S. Leinhardt. Stochastic blockmodels: First
steps. Social Networks, 5(2):109–137, 1983.
[10] J. Komlós and M. Simonovits. Szemerédi’s regularity lemma and its applications in graph theory. In D. Miklós, V. Sós, and T. Szonyi, editors,
Combinatorics, Paul Erdös is Eighty, pages 295–352. János Bolyai Mathematical Society, Budapest, 1996.
[11] L. Massoulié. Community detection thresholds and the weak Ramanujan
property. In Proceedings of the Forty-sixth Annual ACM Symposium on
Theory of Computing, STOC ’14, pages 694–703, New York, NY, USA,
2014. ACM.
[12] T. Nepusz, L. Négyessy, G. Tusnády, and F. Bazsó. Reconstructing cortial networks: case of directed graphs with high level of reciprocity. In
B. Bollobás, R. Kozma, and D. Miklós, editors, Handbook of Large-Scale
Random Networks, number 18 in Bolyai Society of Mathematical Studies,
pages 325–368. Springer, 2008.
[13] I. Norros and H. Reittu. On a conditionally Poissonian graph process. Adv.
Appl. Prob., 38:59–75, 2006.
[14] I. Pappas and P.Pardalos.
Identifying cognitive states using regularity partitions.
PLoS ONE, 10(8):
e0137012, 2015.
https://doi.org/10.1371/journal.pone.0137012.
36
[15] V. Pehkonen and H. Reittu. Szemerédi-type clustering of peer-to-peer
streaming system. In Proc. Cnet 2011, San Francisco, U.S.A., 2011.
[16] T. P. Peixoto. Entropy of stochastic blockmodel ensembles. Physical Review
E, 85(056122), 2012.
[17] M. Pelillo, I. Elezi, and M. Fiorucci. Revealing structure in large graphs:
Szemerédis regularity lemma and its use in pattern recognition. Pattern
Recognition Letters, 87:4–11, 2017.
[18] H. Reittu, F. Bazsó, and R. Weiss. Regular decomposition of multivariate
time series and other matrices. In P. Fränti, G. Brown, M. Loog, F. Escolano, and M. Pelillo, editors, Proc. S+SSPR 2014, number 8621 in LNCS,
pages 424–433. Springer-Verlag, 2014.
[19] J. Rissanen. A universal prior for integers and estimation by minimum
description length. The Annals of Statistics, 11(2):416–431, 1983.
[20] J. Rissanen. Stochastic Complexity in Statistical Inquiry, volume 15. World
scientific, 1998.
[21] G. Sárkozy, F. Song, E. Szemerédi, and S. Trivedi. A practical regularity partitioning algorithm and its application in clustering. 2012.
arXiv:1209.6540v1 [math.CO ] 28 Sep 2012.
[22] A. Scott. Szemerédi’s regularity lemma for matrices and sparce graphs.
Combinatorics, Probability and Computing, 20(3):455–466, 2011.
[23] A. Sperotto and M. Pelillo. Szemerédi’s regularity lemma and its applications to pairwise clustering and segmentation. In Proc. EMMCVPR 2007,
Italy, 2007.
[24] E. Szemerédi. Regular partitions of graphs. In Problemés Combinatoires
et Théorie des Graphes, number 260 in Colloq. Intern. C.N.R.S., pages
399–401, Orsay, 1976.
[25] T.
Tao.
Szemerédi’s
regularity
lemma
via
random
partitions,
2009.
Blog
entry,
https://terrytao.wordpress.com/2009/04/26/szemeredis-regularity-lemma-via-random-parti
37
| 7cs.IT
|
1
Atypicality for Heart Rate Variability Using a
Pattern-Tree Weighting Method
arXiv:1710.07319v1 [cs.LG] 12 Oct 2017
Elyas Sabeti, Member, IEEE and Anders Høst-Madsen, Fellow, IEEE
Abstract—Heart rate variability (HRV) is a vital measure of
the autonomic nervous system functionality and a key indicator of
cardiovascular condition. This paper proposes a novel method,
called pattern tree which is an extension of Willem’s context
tree to real-valued data, to investigate HRV via an atypicality
framework. In a previous paper atypicality was developed as
method for mining and discovery in “Big Data,” which requires
a universal approach. Using the proposed pattern tree as a
universal source coder in this framework led to discovery of
arrhythmias and unknown patterns in HRV Holter Monitoring.
I. I NTRODUCTION
Information theory is generally a theory of typicality. For
example, compressing data using the Asymptotic Equipartion
Property (AEP) can be done by throwing away all sequences
that are not typical. Our perspective in this paper and prior
work [1], [2], [3], [4], [5], [6], [7], [8], [9] is that the value
of data lies not in these typical sequence, but in the atypical
sequences. Take art: the truly valuable paintings are those that
are rare and atypical. Take online collections of photos, such as
Flickr.com, the photos that of interest are those that are unique.
They are atypical. Of course, as opposed to ’interestingness,’
an atypicality criterion will find photos that are both uniquely
good and uniquely bad, there is no value judgment. A similar
example can be investing: extraordinary gains can be obtained
only by investing in atypical stocks, yet that can also lead to
total ruin. Atypicality [3] is defined by
Definition 1. A sequence is atypical if it can be described
(coded) with fewer bits in itself rather than using the (optimum) code for typical sequences.
In prior work, this framework has been used for data
discovery [1], [2], [3], [4], [5], [6], [7], [8], [9]. Our aim
with atypicality theory is to find ’unknown unknowns’ [10]. To
encode data in itself, we require a universal source coder. In
our discrete case atypicality papers we used Willems’ ContextTree Weighting (CTW) as an universal source coder which
requires data binarization.
With the aforementioned atypicality purpose, in this paper
we are extending CTW method to real-valued data by introducing the pattern-tree weighting (PTW) method with finite
maximum memory depth as binary-structured tree in which
data samples are partitioned by a binary pattern achieved by
comparing consecutive samples and the probability estimation
in each node is done by predictive estimators such as predictive
minimum description length (MDL).
The principal reason for introducing a new method such as
PTW is its strength in pattern mining of real-valued data which
has enormous applications in data science. For instance, the
patterns in Heart Rate Variability (HRV) are informative and
symptomatic of heart diseases; however their detailed recognition in Holter Monitoring is impractical for cardiologists,
due to huge amount of data at hand, i.e., “Big Data.” We
first discovered these HRV patterns in [1], but because of the
data binarization many patterns of real-valued HRV were not
captured. This issue is circumvented by PTW.
II. P REDICTIVE MDL
of parametric models P
=
Consider the family
Pθ (xn ) : θ ∈ Θ ⊂ Rk . Rissanen [11] defined predictive
MDL (we call it ordinary predictive MDL) by
n−1
X
L(xn ) = −
log Pθ̂(xi ) (xi+1 |xi )
i=1
in which using the already observed data xi , the parameters of
the model θ̂(xi ) are estimated and they’re used to estimate the
op
probability of the next sample Pθ̂(x
(xi+1 |xi ). In our comi)
panion paper [9] we have shown that the ordinary predictive
MDL has initialization problems in the redundancy sense that
effects the total codelength, which we solve it by proposing
another predictive approach that is called the sufficient statistic
method, in which using sufficient statistics t, the distribution
of the parameters θ̂(t) are estimated and applied to calculate
ss
the probability of the next sample Pθ̂(t)
(xi+1 |xi ). To compare
our sufficient statistic method and the ordinary predictive
approach, assume the model is N(µ, σ 2 ), then
1
1
2
op
p
exp − 2 (xn+1 − µ
cn )
(1)
P (xn+1 |xn ) =
2Sn
2πSn2
2 n2
r
Γ n+1
nSn
n
ss
2
P (xn+1 |xn ) =
π (n + 1) Γ n2 (n + 1) S 2 n+1
2
n+1
(2)
Pn
Pn
2
1
cn ) .
where µ
cn = n1 i=1 xi and Sn2 = n−1
i=1 (xi − µ
Note that for convenience, we drop the subscript θ̂(xi ) in
probabilities. Even though both ordinary predictive MDL and
sufficient statistic methods have distinct behavior in their initialization performance, they both achieve the same asymptotic
code length. In section III-B we’ll see how this predictor are
going to be used in each node of pattern-tree to calculate
coding distributions.
III. PATTERN TREE : A MODEL FOR HEART RATE
VARIABILITY
As it was mentioned earlier in the introduction, in [1] we
discovered some patterns in HRV that are signs of particular
heart arrhythmias. Due to quantization of the data, our discovery was limited, therefore we are motivated to extend the
2
result to the real-valued time series of HRV and this requires a
model for HRV signal. Even though there are many nonlinear
autoregressive processes and switching state-space models for
HRV, Costa et al [12] showed that all of those methods give a
rise to a multimodal distribution for HRV and they concluded
that the best is using Gaussian mixture model. Existence of
premature beats (heart beats that happen very early, before
heart contraction happens) in the HRV data that we have used
in [1] was indicative of the same fact at the rudimentary level.
Consider a time series of heart rate measurements xn =
{x1 , x2 , · · · , xn } and define zn as the state variable at time
n with K possibilities (K is the number of Gaussian distributions in the mixture model). Let πk = Pr {zn = k} for all
1 ≤ k ≤ K, then based on Costa’s model [12] we have
X
P (xn+1 |xn ) =
P (zn )P (xn+1 |xn , zn )
zn
=
K
X
πk Pθk (xn+1 |xn )
(3)
k=1
PK
K
where π = {πk }k=1 satisfies 0 ≤ πk ≤ 1 and k=1 πk =
1,
and 2Pθk (xn+1 |xn ) is a Gaussian distribution with θ k =
µk , σk .
The model (3) and the biological reason behind mixturemodeling of HRV made us consider a more complex and more
exhaustive model for this process, a tree model. This can be
explained via a simple example. Assume we have a binary tree
with depth one, this tree partitions the samples in the root node
by comparing it with the previous sample and based on having
a rise or a drop, it will be assigned to one of the child nodes.
Now suppose this tree is used to divide the samples in a HRV
signal that contains premature beats. Note that the histogram of
this HRV signal will be a mixture of two Gaussian distributions
with separated means and different variances. Since the time
difference between the two consecutive normal heart beats
is (relatively) much larger than the time interval between a
normal heart beat and a premature one, the depth-one tree will
separate all the premature beats from normal ones (however
some normal beats can be classified as premature, since depthone tree is too simple). By letting this tree structure to be
deeper, more complex arrhythmias such as tachycardia, flutter
and fibrillation will be separated from normal beats; this could
also lead to new discoveries of heart abnormalities, which is
the real goal of our work. Since the same tree structure is
used in CTW and its application led to data discovery ([3],
[1]), here we want to extend the same concept to develop
the pattern tree weighting (PTW) for real-valued data. This
proposed method uses the binary data pattern as context and
apply predictive MDL for real-valued samples to compute the
coding distribution.
A. Pattern tree as an extension of the context tree
The context tree has been shown to be a powerful method
to compute an appropriate coding distribution [13], [14]. A
context of the binary source symbol yn is a suffix of the
semi-infinite sequence · · · , yn−2 , yn−1 that precedes it. The
context tree consists of nodes corresponding to each context
up to certain depth. A pattern tree of depth D has the same
structure of a context tree of the same depth, but since we
are interested to design it for real-valued data, the way it
splits the source sequence is different. Suppose at time n, the
last D samples of the real-valued source sequence (xi ∈ R)
are xn−D , · · · , xn−2 , xn−1 . After putting xn in the root node
η1 = λ, the way we assign xn to any of the root’s children
(η2 and η3 ) at depth one is based on comparing xn and xn−1 ,
for instance if xn−1 > xn we assign xn to node η2 . Next
the way we assign xn to η2 ’s children (η4 and η5 ) at depth
two is based on comparing xn−1 and xn−2 and we keep on
doing this until we reach the maximum depth D. As it can be
seen, at time n, connecting the nodes that are assigned by xn
illustrates a pattern that shows the fluctuation of the (D + 1)most recent source samples from xn−D to xn , and that’s the
reason we call it a pattern tree. Here is an example to show
how the source sequence is portioned by the pattern tree.
Example 2. Consider a tree with depth D = 3 and suppose
that the source generated the sequence [1.7 0.6 -2.6 -0.8 0.7
7.1 5.5 -2.7 6 1.4] while the past sequence was [... -0.9 0.1
-0.4]. Then the sequence is partitioned by the pattern tree, see
Fig. 1 (for now disregard the probabilities in the Fig. 1).
B. Coding for an unknown tree source
In this section we describe how to employ the predictive
MDL of section II to estimate the probability in each node.
Here we want to use the same concept of the weighted coding
distribution in the context tree, adapted to the real-value case.
Therefore to each node η we will assign a sequential predictive
distribution that only depends on the data samples observed by
this particular node. In fact, the weighted coding distribution
of the pattern tree will be exactly like the context tree, but with
this difference that instead of KT-estimator which is designed
for binary data, we use predictive MDL estimators (e.g., for
the case of Gaussian see (1) or (2) ).
From now on, we assume for a tree with depth D, the
initial context of past sequence x1−D , · · · , x0 is available.
At any time n for each node ηi suppose the vectors tηni =
{t < n : xt ∈ ηi } and xηni = {xt : t ∈ tηni } are the set of all
the time indexes of the already observed data samples by that
node and its corresponding data samples, respectively. Clearly
η
for every internal node we have tηni = tηn2i ∪ tn2i+1 , i.e., the set
of all the time indexes of the already observed data samples
by the parent node ηi is the union of the set of all the time
indexes of the already observed data samples by its children
η
(see Fig. 1), and similarly xηni = xηn2i ∪ xn2i+1 . Ergo the
weighted probability in each internal node is
1
Pη (xn+1 |xηni )
2 i
1
+ Pη2i (xn+1 |xηn2i )Pη2i+1 (xn+1 |xηn2i+1 )
2
where Pηi (xn+1 |xηni ) is the predictive distribution (e.g., (1)
and (2)) at node ηi . So in each node, the parameter of the
model is estimated based on the observed samples at that
particular node. One issue here can be in initialization, for
instance for a Gaussian source, at least each node needs
two samples to estimate both mean and variance. This can
be resolved by estimating the initial predictive distribution
Pω,ηi (xn+1 |xηni ) =
3
parameters based on the first D samples of data in the context.
Finally the weighted probability in each internal node will be
(see (4) at the top of the next page).
As an example for predictive distribution, if the source
sequence is generated according to N (µ, σ 2 ) where both µ
and σ 2 are unknown, then using ordinary predictive MDL of
(1) we have
1
1
2
exp − 2 (xn+1 − µ
cn )
(5)
Pηi (xn+1 |xηni ) = p
2Sn
2πSn2
P
1
η
where µ
cn
=
and Sn2
=
η
xi ∈xni xi
|xni |
P
2
1
η
cn ) .
The
following
example
η
xi ∈xni (xi − µ
|xni |−1
calculates the weighted probability of the source sequence of
Example 2 using Gaussian MDL predictor (2):
Example 3. Back to the Example 2, the weighted probabilities
Pω using our sufficient-statistic-based predictor (2) as the
estimator for Gaussian distribution are depicted in the pattern
tree in Fig. 1.
3,8,11
3,8,11
p !=2.7x10 -5
2,3,7,8,10,11
p !=2.7x10 -5
2,7,10
p !=4.1x10
p !=1.7x10 -9
-5
1,2,3,4,5,
6,7,8,9,10,11
p !=3.8x10 -17
1,4,5,6,9
p !=6.3x10 -9
1
2
3
4 5
6
7
8 9 10 11
... -0.9 0.1 -0.4 1.7 0.6 -2.6 -0.8 0.7 7.1 5.5 -2.7 6 1.4 1
Past
Source Sequence
Figure 2. Updated path of the weighted pattern tree of the source sequence
in Example 2 followed by a 1.
terms of coding, Definition 1 can be stated in the following
form
C(x|P) − C(x) > 0
3,8
3,8
p !=9.6x10 -4
2,10
p !=1x10-2
7
p !=3.4x10 -3
4,9
p !=2.6x10 -4
1
p !=3.2x10 -2
5
p !=1.3x10 -1
6
p !=2x10-3
p !=9.6x10 -4
2,3,7,8,10
2,7,10
p !=4.1x10
p !=3.8x10 -8
-5
1,2,3,4,5,
6,7,8,9,10
p !=3.8x10 -16
1,4,9
p !=9.8x10 -6
5,6
1,4,5,6,9
p !=6.3x10 -9
p !=3.7x10 -4
1
2
3
4 5
6
7
8 9 10
... -0.9 0.1 -0.4 1.7 0.6 -2.6 -0.8 0.7 7.1 5.5 -2.7 6 1.4
Past
Source Sequence
Figure 1. The source sequence of Example 2 is split up by PTW. In Example
3 the weighted probabilities are then calculated.
C. Updating the pattern tree and complexity
Suppose a pattern tree of depth D already seen the source
sequence x1 , · · · , xn−1 , xn , now we want to see how complex
the pattern tree evolves when the next sample xn+1 is going
to be processed. This is done by updating all the nodes in only
one path of length D from the root to a leaf node and this
path is determined by the procedure explained in section III-A.
Then the updating steps in all the D + 1 nodes in the evolving
path are: (i) updating the parameter estimation, (ii) updating
predictive probabilities Pηi (xn+1 |xηni ) and (iii) updating the
weighted probabilities Pω,ηi (xn+1 |xηni ). The following example which is the continuation of the series of examples shows
how the source sequence in Example 3 is updated for new
source sample.
Example 4. Suppose that the source has already generated the
sequence of Example 2. This resulted the weighted pattern tree
in Note
Fig. 1.
that by assuming a parametric model (here, a Gaussian
model), there would be no more need to store all the data
samples in each node; instead, only parameters should be
updated and stored in the nodes.
IV. ATYPICALITY USING PTW
Suppose P = Pθ (xn ) : θ ∈ Θ ⊂ Rk is a family of
parametric model class that can be assigned to the data. In
Here C(x|P) is the code length of x encoded with the
optimum coder according to the typical law (known parameters
θ), and C(x) is x encoded ’in itself.’ As argued in [3], we
need to put a ’header’ in atypical sequences to inform the
encoder that an atypical encoder is used. We can therefore
write C(x) = τ + C̃(x), where τ is the number of bits for the
’header,’ and C̃(x) is the number of bits used for encoding the
data itself. For encoding the data in itself an obvious solution
is to use a universal source coder. We have therefore chosen
to use the PTW algorithm.
A. Typical Encoding and Training
In Definition 1 we have assumed that parameters θ of the
typical model of data is exactly known, however, in many
cases the parameters are not known exactly. Let us assume we
are given a single long sequence t for training – rather than
the parameters θ in model P – and based on this we need
to encode a sequence x. To understand what this means, we
have to realize that when x is encoded according to C(x|P)
with a known θ, the coding probabilities are fixed; they are
not affected by x. This is an important part of Definition 1
that reacts to ’outliers,’ data that does not fit the typical model.
But the issue with universal source coders is that they often
easily adapts to new types of data, a desirable property of a
good universal source coder, but problematic in light of the
above discussion. We therefore need to ’freeze’ the source
coder, for example by not updating the dictionary. However,
because the training data is likely incomplete as discussed
above, the freezing should not be too hard. This issue is
precisely described in [3] and due to page limitation we don’t
go over it here, but with a simulation we show why freezing
the encoder is essential in implementing atypicality: The PTW
algorithm trained with a Gaussian process with σ 2 = 4 and
then tested with another Gaussian source with same mean and
different variance σ 2 = 1. It can be seen in Fig. 3 that the
non-frozen algorithm learn the statistic and behavior of the
new source; however, the codelength using frozen algorithm
keeps increasing and that’s a property we are interested in.
4
Pω,ηi (xn+1 |xηni ) =
η2i+1
1
1
ηi
η2i
2 Pηi (xn+1 |xn ) + 2 Pη2i (xn+1 |xn )Pη2i+1 (xn+1 |xn )
ηi : internal node
ηi : leaf node
algorithm outperformed other methods. Here we want to
compare our algorithm with CDM on a simulated data. In
this experiment the training sequence is a Gaussian process
with zero mean and σ 2 = 4, and the test sequence have the
same statistics as the training data but with two episodes of
anomalies embedded in it: a sinusoidal segment (red part in
Fig. 4) and a segment of Gaussian process with zero mean
and σ 2 = 1 (green part in Fig. 4). As can be verified
from the figure, our PTW-based atypicality found both of
the anomalies in the data; however, CDM only detected the
sinusoidal pattern.
The Effect of Freezing: Non-Frozen Typical Coder
Will Learn The Atypical Sequences
10000
Non-Frozen Typical Code Length
Frozen Typical Code Length
Atypical Code Length
9000
(4)
Pηi (xn+1 |xηni )
8000
7000
6000
5000
4000
3000
2000
Test Data
1000
20
A)
0
0
50
100
150
200
250
300
350
400
450
B)
C)
10
500
Figure 3. The importance of freezing source coding when testing for
atypicality.
0
-10
-20
B. Atypical subsequences
50
Let X (n, l) = (xn , xn+1 , . . . , xn+l−1 ) be a subsequence
of {xn , n = 0, . . . , ∞} that we want to test for atypicality.
As mentioned earlier, the start of a sequence needs to be
encoded as well as the length. Additionally the code length
is minimized over the maximum depth D of the context tree.
The atypical code length is then given by
∗
∗
LA (X (n, l)) = min (− log Pω,λ (D) + log D) + log l
D
except for the τ . Here Pω,λ (D) denotes the probability at
the root of the pattern tree of depth D. For typical coding
we use the algorithm in Section IV-A when the parameters
are not known; let L(n) be the codelength for the sequence
x0 , . . . , xn . Then we can put LT (X (n, l)) = L(n + l − 1) −
L(n). We need to test every subsequence of every length, that
is, we need to test subsequences X (n, l) for every value of
n and l. For atypical coding this means that a new PTW
algorithm needs to be started at every sample time. So, if the
maximum sequence length is L, L separate PTW trees need
to maintained at any time. These are completely independent,
so they can be run on parallel processors. The result is that
for every bit of the data we calculate
∆L(n) = min {LA (X (n, l)) − LT (X (n, l))}
l
100
150
200
250
300
350
400
450
500
Atypicality
CDM
Figure 4. A comparison of two different methods. For our proposed method,
the 4L of equation (6) has been plotted. A) Gaussian process with zero mean
and σ 2 = 4, exactly the same as training (typical) data, B) A sinusoidal
segment C) A segment of Gaussian process with zero mean and σ 2 = 1 .
B. Detection of Paroxysmal Atrial Fibrillation (PAF)
Paroxysmal Atrial Fibrillation (PAF) is an irregular, often
rapid heart rate that commonly causes poor blood flow, which
may lead to severe consequences. PhysioNet [16] has provided
a database for this arrhythmia [17]. Part of this database
includes 30-minute ECG records of different subjects (who
have PAF) during a period that is distant from any episode of
PAF. Each 30-minute record is then followed by two 5-minute
record of the same subject, one of which contains an episode
of PAF, and the other one has no such an episode. Fig. 5 shows
HRV signal for a subject in the database.
30 minutes HRV segment: Distant from any episode of PAF
(6)
1.5
1
and we can state the atypicality criterion as ∆L(n) < −τ .
V. A NOMALY D ETECTION
In [15] authors came up with the Compression-based Dissimilarity Measure (CDM) that uses the dissimilarity measure
for anomaly detection. They showed by that their proposed
0
200
400
600
800
1000
1200
1400
1600
1800
2000
5 minutes HRV segment: Distant from any episode of PAF
Time (sec)
In [3] we have used CTW for anomaly detection in binary
data and we got unique results. In order to verify the performance of our algorithm for real-valued data, here we want
to use PTW as encoder in atypicality framework for same
purpose.
A. Comparison with an alternative method
0.5
1.5
1
0.5
0
50
100
150
200
250
300
350
5 minutes HRV segment: Contains episode of PAF
1.5
1
0.5
0
50
100
150
200
250
Samples
Figure 5. HRV signal of a PAF subject in the database
300
350
5
As we mentioned earlier, the ultimate goal of atypicality is
to find “unknown unknowns” in Big Data. One example can be
the data that achieved by Holter Monitoring, i.e., a continuous
tape recording of a patient’s ECG for 24 hours. For such a
data we used the MIT-BIH Normal Sinus Rhythm Database
(nsrdb) which is provided by PhysioNet [16]. Even though the
subjects included in this database were found to have had no
significant arrhythmias, there exist many arrhythmic beats and
patterns to look for. We want to apply our algorithm to find
interesting parts of the data in the dataset.
Since the data is assumed to be “Normal Sinus Rhythm,”
we trained and froze our typical PTW with Gaussian process
of the same mean and variance of the data. Using the process
of the section IV-B, we managed to find interesting results
on the dataset. As an example, we provide Fig. 6. An can
be seen in the figure, for that particular data we came up
with three major atypical subsequences: the first one is denotes
by “S,” second one is marked with “V” and the last one as
“pattern.” After looking up in the annotation file which is
provided for the HRV, the “S” group corresponds to onset
of some supraventricular beats and the “V” groups was the
result of ventricular contraction; however, there were no label
for the segment that we call “pattern” in the HRV annotations.
Looking closer in the data shows existence of some repetitive
patterns that was not even seen by the cardiologist who
annotated the HRV data. Taking a deeper look into ECG (not
HRV) annotation shows that those patterns are happening at
the same time that there are either some isolated QRS-like
artifact or signal quality change [16]. This shows our algorithm
was able to find something that was missed by the eye of
the expert who was only looking into the HRV data without
considering the ECG signals; nevertheless, it could be the sign
of some heart malfunction after the recording was over.
R EFERENCES
[1] A. Høst-Madsen, E. Sabeti, and C. Walton, “Information theory for
atypical sequences,” in IEEE Information Theory Workshop (ITW’13),
Seville, Spain, 2013.
[2] A. Host-Madsen and E. Sabeti, “Atypical information theory for realvalued data,” in 2015 IEEE International Symposium on Information
Theory (ISIT). IEEE, 2015, pp. 666–670.
[3] A. Høst-Madsen, E. Sabeti, and C. Walton, “Data discovery and anomaly
detection using atypicality: Theory,” IEEE Transactions on Information
Theory, submitted 2016, available at http://arxiv.org/abs/1709.03189.
[4] A. Host-Madsen, E. Sabeti, C. Walton, and S. J. Lim-Higbie, “Universal
data discovery using atypicality,” in 3rd International Workshop on
Pattern Mining and Application of Big Data (BigPMA - Big Data 2016),
2016 IEEE International Conference on. IEEE, 2016.
1.5
HRV: An episode of 24 hours
Atypicality: "L(n) (bits)
Time (sec)
In our experiment, for each subject we train the PTW with
HRV of 30-minute records and freeze it, then it was used as
encoder for its two 5-minute segments. Our goal was to detect
the 5-minute record with PAF episode, since we believed using
the PTW trained on 30-minute records, smaller codelength is
needed to encode the 5-minute record without PAF episode.
After applying the same procedure on all the 25 subjects, we
were able to detect all the 5-minute record containing PAF
episode correctly, with 100 accuracy (consequently due to data
structures, zero false alarm), in fact we have improved the
best results of other researches on the same database about 10
percent [18], [19]. Again, as we discussed in section IV-A it
verifies the importance of freezing in our notion of atypicality.
C. Detection of anomaly in Holter Monitoring
1
0.5
S
Patterns
V
0
-0.5
-1
1
2
3
4
5
Samples
6
7
8
9
10
#10 4
Figure 6. Atypicality using PTW for Holter Monitoring HRV: “S” stands for
supraventricular arrhythmia and “V” stands for ventricular contraction
[5] E. Sabeti and A. Host-Madsen, “Atypicality for vector gaussian models,”
in 2015 IEEE Global Conference on Signal and Information Processing
(GlobalSIP). IEEE, 2015, pp. 328–332.
[6] ——, “Atypicality for the class of exponential family,” in 2016 54rd Annual Allerton Conference on Communication, Control, and Computing
(Allerton). IEEE, 2016.
[7] ——, “How interesting images are: An atypicality approach for social
networks,” in Big Data (Big Data), 2016 IEEE International Conference
on. IEEE, 2016.
[8] E. Sabeti and A. Høst-Madsen, “Data discovery and anomaly
detection using atypicality: Signal processing methods,” IEEE
Transactions on Signal Processing, submitted 2017, available at
http://arxiv.org/abs/1709.03191.
[9] E. Sabeti and A. Host-Madsen, “Enhanced mdl with application to
atypicality,” in 2017 IEEE International Symposium on Information
Theory (ISIT). IEEE, 2017.
[10] D. Rumsfeld, Known and Unknown: A Memoir. Penguin, 2011.
[11] J. Rissanen, “Stochastic complexity and modeling,” The Annals of
Statistics, no. 3, pp. 1080–1100, Sep. 1986.
[12] F. M. Costa T, Boccignone G, “Gaussian mixture model of heart rate
variability,” PLoS ONE, no. 5, 2012.
[13] F. M. J. Willems, Y. Shtarkov, and T. Tjalkens, “The context-tree weighting method: basic properties,” Information Theory, IEEE Transactions
on, vol. 41, no. 3, pp. 653–664, 1995.
[14] F. Willems, Y. Shtarkov, and T. Tjalkens, “Reflections on "the context
tree weighting method: Basic properties",” Newsletter of the IEEE
Information Theory Society, vol. 47, no. 1, 1997.
[15] E. Keogh, S. Lonardi, and C. A. Ratanamahatana, “Towards parameterfree data mining,” in Proceedings of the tenth ACM SIGKDD international conference on Knowledge discovery and data mining. ACM,
2004, pp. 206–215.
[16] A. L. Goldberger, L. A. N. Amaral, L. Glass, J. M. Hausdorff,
P. C. Ivanov, R. G. Mark, J. E. Mietus, G. B. Moody, C.-K.
Peng, and H. E. Stanley, “Physiobank, physiotoolkit, and physionet,”
Circulation, vol. 101, no. 23, pp. e215–e220, 2000. [Online]. Available:
http://circ.ahajournals.org/content/101/23/e215
[17] G. Moody, A. Goldberger, S. McClennen, and S. Swiryn, “Predicting
the onset of paroxysmal atrial fibrillation: The computers in cardiology
challenge 2001,” in Computers in Cardiology 2001. IEEE, 2001, pp.
113–116.
[18] E. Sabeti, M. B. Shamsollahi, and F. Afdideh, “Prediction of paroxysmal
atrial fibrillation using empirical mode decomposition and rr intervals,”
in Biomedical Engineering and Sciences (IECBES), 2012 IEEE EMBS
Conference on. IEEE, 2012, pp. 750–754.
[19] T. Thong, J. McNames, M. Aboy, and B. Goldstein, “Prediction of
paroxysmal atrial fibrillation by analysis of atrial premature complexes,”
IEEE transactions on biomedical engineering, vol. 51, no. 4, pp. 561–
569, 2004.
| 7cs.IT
|
The power of deeper networks for expressing natural functions
David Rolnick∗ , Max Tegmark†
Dept. of Physics, Massachusetts Institute of Technology, Cambridge, MA 02139 and
Dept. of Mathematics, Massachusetts Institute of Technology, Cambridge, MA 02139
arXiv:1705.05502v1 [cs.LG] 16 May 2017
(Dated: May 17, 2017)
It is well-known that neural networks are universal approximators, but that deeper networks tend
to be much more efficient than shallow ones. We shed light on this by proving that the total number
of neurons m required to approximate natural classes of multivariate polynomials of n variables
grows only linearly with n for deep neural networks, but grows exponentially when merely a single
hidden layer is allowed. We also provide evidence that when the number of hidden layers is increased
from 1 to k, the neuron requirement grows exponentially not with n but with n1/k , suggesting that
the minimum number of layers required for computational tractability grows only logarithmically
with n.
I.
INTRODUCTION
Deep learning has lately been shown to be a very powerful
tool for a wide range of problems, from image segmentation to machine translation. Despite its success, many
of the techniques developed by practitioners of artificial
neural networks (ANNs) are heuristics without theoretical guarantees. Perhaps most notably, the power of feedforward networks with many layers (deep networks) has
not been fully explained. The goal of this paper is to shed
more light on this question and to suggest heuristics for
how deep is deep enough.
It is well-known [1–3] that nonlinear neural networks with
a single hidden layer can approximate any function under
reasonable assumptions, but it is possible that the networks required will be extremely large. Recent authors
have shown that some functions can be approximated by
deeper networks much more efficiently (i.e. with fewer
neurons) than by shallower ones. However, many of the
functions in question are complicated or arise from “existence proofs” without explicit constructions, and the
results often apply only to types of network rarely used
in practice.
Deeper networks have been shown to have greater representational power with respect to various notions of complexity, including piecewise linear decision boundaries [4]
and topological invariants [5]. Recently, Poole et al. [6]
and Raghu et al. [7] showed that the trajectories of input
variables attain exponentially greater length and curvature with greater network depth. Work including [8] and
[9] shows that there exist functions that require exponential width to be approximated by a shallow network.
Mhaskar, Liao, and Poggio [10], in considering compositional functions with this property, inquire whether explicit examples must be pathologically complicated.
Various authors have also considered the power of deeper
∗ [email protected]
† [email protected]
network of types other than the standard feedforward
model. The problem has also been posed for sum-product
networks [11] and restricted Boltzmann machines [12].
Cohen, Sharir, and Shashua [13] showed, using tools from
tensor decomposition, that shallow arithmetic circuits
can express only a measure-zero set of the functions expressible by deep circuits. A weak generalization of this
result to convolutional neural networks was shown in [14].
In summary, recent years have seen a wide variety of theoretical demonstrations of the power of deep neural networks. It is important and timely to extend this work
to make it more concrete and actionable, by deriving
resource requirements for approximating natural classes
of functions using today’s most common neural network
architectures. Lin, Tegmark, and Rolnick [15] recently
proved that it is exponentially more efficient to use a
deep network than a shallow network when approximating the product of input variables. In the present paper,
we will greatly extend these results to include broad natural classes of multivariate polynomials, and to tackle the
question of how resource use depends on the precise number of layers. Our results apply to standard feedforward
ANNs with general nonlinearities and are borne out by
empirical tests.
The rest of this paper is organized as follows. In §II B,
we consider the complexity of approximating a multivariate polynomial p(x) using a feedforward neural network
with input x = (x1 , . . . , xn ). We show (Theorem II.4)
that for general sparse p, exponentially many neurons are
required if the network is shallow, but at most linearly
many for a deep network. For monomials p, we calculate
(Theorem II.1) exactly the minimum number of neurons
required for a shallow network. These theorems apply for
all nonlinear activation functions σ with nonzero Taylor
coefficients; a slightly weaker result (Theorem II.2) holds
for an even broader class of σ.
In §II C, we present similar results (Propositions II.5
and II.6) for approximating univariate polynomials. In
this case, shallow networks require linearly many neurons
while for deep networks it suffices to use only logarithmically many neurons. In §II D, we tie the difficulty of
2
approximating polynomials by shallow networks to the
complexity of tensor decomposition (Proposition II.7).
In §III A, we consider networks with a constant number
k of hidden layers. For input of dimension n, we show
(Theorem III.1) that products can be approximated with
a number of neurons exponential in n1/k , and justify our
theoretical predictions with empirical results. While still
exponential, this shows that problems unsolvable by shallow networks can be tractable even for k > 1 of modest
size.
Finally, in §III B, we compare our results on feedforward neural networks to prior work on the complexity of
Boolean circuits. We conclude that these problems are
independent, and therefore that established hard problems in Boolean complexity do not provide any obstacle
to analogous results for standard deep neural networks.
B.
Multivariate polynomials
The following theorem generalizes a result of Lin,
Tegmark, and Rolnick [15] to arbitrary monomials. By
setting r1 = r2 = . . . = rn = 1 below, we recover their
result that the product of n numbers requires 2n neurons
in a shallow network but can be done with linearly many
neurons in a deep network.
Theorem II.1. Let p(x)
Pn denote the monomial
Suppose that the
xr11 xr22 · · · xrnn , with N =
i=1 ri .
nonlinearity σ(x) has nonzero Taylor coefficients up to
xN . Then:
m1 (p, σ) =
m(p, σ) ≤
n
Y
(ri + 1),
i=1
n
X
(7dlog2 (ri )e + 4),
(1)
(2)
i=1
where dxe denotes the smallest integer that is at least x.
II.
THE INEFFICIENCY OF SHALLOW
NETWORKS
In this section, we compare the efficiency of shallow networks (those with a single hidden layer) and deep networks at approximating multivariate polynomials.
A.
Definitions
Let σ(x) be a nonlinear function, k a positive integer, and
p(x) a multivariate polynomial of degree d. We define
mk (p, σ) to be the minimum number of neurons (excluding input and output) required to approximate p with
a neural net having k hidden layers and nonlinearity σ,
where the error of approximation is of degree at least
d + 1 in the input variables. Thus, in particular, m1 (p, σ)
is the minimal integer m such that:
m
X
j=1
wj σ
n
X
!
aij xi
= p(x) + O(xd+1
+ . . . + xd+1
n ).
1
i=1
Note that approximation up to degree d allows us to approximate any polynomial to high precision as long as
the input variables are small enough. In particular, for
homogeneous polynomials of degree d, we can adjust the
weights so as to scale each variable by a constant λ 1
before input, and then scale the output by 1/λd , in order
to achieve arbitrary precision.
We set m(p, σ) = mink≥0 mk (p, σ). We will show that
there is an exponential gap between m1 (p, σ) and m(p, σ)
for various classes of polynomials p.
Proof. Without loss of generality, suppose that ri > 0 for
i = 1, . . . , n. Let X be the multiset1 in which xi occurs
with multiplicity ri .
Qn
We first show that i=1 (ri + 1) neurons are sufficient to
approximate p(x). Appendix A in [15] demonstrates that
for variables y1 , . . . , yN , the product y1 · · · · · yN can be
approximated as a linear combination of the 2N functions
σ(±y1 ± · · · ± yN ).
Consider setting y1 , . . . , yN equal to the elements of multiset X. Then, we conclude that we can approximate p(x)
as a linear combination of the functions σ(±y1 ±· · ·±yN ).
However, these functions are not all distinct: there are
ri + 1 distinct ways to assign ± signs to ri copies of xi
(ignoring
Qn permutations of the signs). Therefore, there
are i=1 (ri + 1) distinctQfunctions σ(±y1 ± · · · ± yN ),
n
proving that m1 (p, σ) ≤ i=1 (ri + 1).
We now adapt methods introduced in [15] to show that
this number of neurons is also necessary for approximating p(x). Let m ≡ m1P
(p, σ) and suppose that σ(x) has
∞
k
the Taylor expansion
k=0 σk x . Then, by grouping
terms of each order, we conclude that there exist constants aij and wj such that
!N
m
n
X
X
σN
wj
aij xi
= p(x)
(3)
j=1
σk
m
X
j=1
1
i=1
wj
n
X
!k
aij xi
=0
for 0 ≤ k ≤ N − 1. (4)
i=1
That is, a collection of elements of X, in which each element is
allowed to occur multiple times.
3
For each S ⊆ X, let us take the derivative of equations
(3) and (4) by every variable that occurs in S, where we
take multiple derivatives of variables that occur multiple
times. This gives
!N −|S|
m
n
Y
X
Y
σN · N ! X
wj
ahj
aij xi
=
xi , (5)
|S|! j=1
i=1
h∈S
i6∈S
!k−|S|
m
n
Y
X
σk · k! X
wj
ahj
aij xi
=0
(6)
|S|! j=1
i=1
7dlog2 (ri )e neurons arranged in a deep network. Thereri
fore,
P we can approximate all of the xi using a total of
i 7dlog2 (ri )e neurons. From [15], we know that these
n terms can be multiplied
using 4n additional neurons,
P
giving us a total of i (7dlog2 (ri )e + 4). This completes
the proof.
In order to approximate a degree-N polynomial effectively with a single hidden layer, we must naturally assume that σ has nonzero N th Taylor coefficient. However, if we do not wish to make assumptions about the
other Taylor coefficients of σ, we can still prove the following weaker result:
h∈S
Qn
for |S| ≤ k ≤ N − 1. Observe that there are i=1 (ri + 1)
choices for S, since each variable xi can b included
anywhere
from 0 to ri times. Define A to
Qn
Q be the
( i=1 (ri + 1))×m matrix with entries AS,j = h∈S ahj .
We claim that A has full row rank. This would show that
the number of columns m is at least the number of rows
Q
n
i=1 (ri + 1), proving the desired lower bound on m.
Theorem II.2. Once again, let p(x) P
denote the monon
· · xrnn . Then, for N =
mial xr11 xr22 · Q
i=1 ri , we have
n
m1 (p, σ) ≤
(r
+
1)
if
σ
has
nonzero
N th Tayi
i=1
lor coefficient. For any σ, we have that m1 (p, σ) is at
least the maximum
coefficient in the univariate polynoQ
mial g(y) = i (1 + y + . . . + y ri ).
Suppose towards contradiction that the rows AS` ,• admit
a linear dependence:
r
X
Proof outline. The first statement follows as in the proof
of Theorem II.1. For the second, consider once again
the equation (5). The left-hand side can be written, for
any S, as the linear combination of basis functions of the
form:
!N −|S|
n
X
aij xi
.
(7)
c` AS` ,• = 0,
`=1
where the coefficients c` are nonzero and the S` denote distinct subsets of X. Set s = max` |S` |. Then,
take the dot product of each side of the above equation by the vector with entries (indexed by j) equal to
Pn
N −s
wj ( i=1 aij xi )
:
0=
r
X
c`
`=1
=
+
m
X
wj
j=1
X
c`
Y
ahj
wj
aij xi
Y
ahj
j=1
h∈S`
i=1
X
m
X
Y
n
X
c`
j=1
wj
ahj
!N −|S` |
n
X
`|(|S` |=s)
`|(|S` |<s)
Letting S vary over multisets of fixed size s, we see that
the right-hand side of (5) attains every degree-s monomial that divides p(x). The number of such monomials is
the coefficient Cs of the term y s in the polynomial g(y).
Since such monomials are linearly independent, we conclude that the number of basis functions of the form (7)
above must be at least at least Cs . Picking s to maximize
Cs gives us the desired result.
!N −s
i=1
h∈S`
m
X
n
X
i=1
aij xi
!(N +|S` |−s)−|S` |
aij xi
.
Corollary II.3. If p(x) is the product of the inputs
(i.e. r1 = · · ·p= rn = 1), we conclude that m1 (p, σ) ≥
n
n
bn/2c ∼ 2 / πn/2, under the assumption that the nth
Taylor coefficient of σ is nonzero.
|S` |!
·0
σk · k!
It is natural now to consider the cost of approximating general polynomials. However, without further constraint, this is relatively uninstructive because polynomials of degree d in n variables live within a space of dimension n+d
d , and therefore most require exponentially
many neurons for any depth of network. We therefore
consider polynomials of sparsity c: that is, those that
can be represented as the sum of c monomials. This includes many natural functions.
i=1
h∈S`
We can use (5) to simplify the first term and (6) (with
k = N + |S` | − s) to simplify the second term, giving us:
0=
X
c` ·
h6∈S`
`|(|S` |=s)
=
X
`|(|S` |=s)
Y
|S` |!
·
xh +
σN · N !
c` ·
|S` |!
·
σN · N !
Y
X
`|(|S` |<s)
c` ·
xh .
h6∈S`
Q
Since the monomials xi ∈S` xi are linearly independent,
this contradicts our assumption that the c` are nonzero.
We conclude thatQA has full row rank, and therefore that
n
m1 (p, σ) = m ≥ i=1 (ri + 1). This completes the proof
of equation (1).
For equation (2), it follows from Proposition II.6 part
(2) below that for each i, we can approximate xri i using
The following theorem, when combined with Theorem
II.1, shows that general polynomials p with subexponential sparsity have exponential large m1 (p, σ), but subexponential m(p, σ).
4
Theorem II.4. Suppose that p(x) is a multivariate polynomial of degree N , with c monomials
q1 (x), q2 (x), . . . , qc (x). Suppose that the nonlinearity
σ(x) has nonzero Taylor coefficients up to xN . Then,
1. m1 (p, σ) ≥ 1c maxj m1 (qj , σ).
P
2. m(p, σ) ≤ j m(qj , σ).
Proof outline. Our proof in Theorem II.1 relied upon the
fact that all nonzero partial derivatives of a monomial
are linearly independent. This fact is not true for general polynomials p; however, an exactly similar argument
shows that m1 (p, σ) is at least the number of linearly independent partial derivatives of p, taken with respect to
multisets of the input variables.
Consider the monomial q of p such that m1 (q, σ) is maximized, and suppose that q(x) = xr11 xr22 · · ·Q
xrnn . By Theon
rem II.1, m1 (q, σ) is equal to the number i=1 (ri + 1) of
distinct monomials that can be obtained by taking partial derivatives of q. Let Q be the set of such monomials,
and let D be the set of (iterated) partial derivatives corresponding to them, so that for d ∈ D, we have d(q) ∈ Q.
Consider the set of polynomials P = {d(p) | d ∈ D}. We
claim that there exists a linearly independent subset of
P with size at least |D|/c. Suppose to the contrary that
P 0 is a maximal linearly independent subset of P with
|P 0 | < |D|/c.
Since p has c monomials, every element of P has at most
c monomials. Therefore, the total number of distinct
monomials in elements of P 0 is less than |D|. However,
there are at least |D| distinct monomials contained in elements of P , since for d ∈ D, the polynomial d(p) contains
the monomial d(q), and by definition all d(q) are distinct
as d varies. We conclude that there is some polynomial
p0 ∈ P \P 0 containing a monomial that does not appear
in any element of P 0 . But then p0 is linearly independent of P 0 , a contradiction since we assumed that P 0 was
maximal.
We conclude that some linearly independent subset of
P has size at least |D|/c, and therefore that the space
of partial derivatives of p has rank at least |D|/c =
m1 (q, σ)/c. This proves part (1) of the theorem. Part (2)
follows immediately from the definition of m(p, σ).
only logarithmically many neurons in a deep network.
Thus, depth allows us to reduce networks from linear
to logarithmic size, while for multivariate polynomials
the gap was between exponential and linear. The difference here arises because the dimensionality of the space
of univariate degree-d polynomials is linear in d, which
the dimensionality of the space of multivariate degree-d
polynomials is exponential in d.
Proposition II.5. Let σ be a nonlinear function with
nonzero Taylor coefficients. Then, we have m1 (p, σ) ≤
d + 1 for every univariate polynomial p of degree d.
Proof. Pick a0 , a1 , . . . , ad to be arbitrary, distinct real
numbers. Consider the Vandermonde matrix A with
entries Aij = aji . It is well-known that det(A) =
Q
0
i<i0 (ai − ai ) 6= 0. Hence, A is invertible, which means
that multiplying its columns by nonzero values gives another invertible matrix. Suppose that we multiply
P the
jth column of A by σj to get A0 , where σ(x) = j σj xj
is the Taylor expansion of σ(x).
Now, observe that the ith row of A0 is exactly the coefficients of σ(ai x), up to the degree-d term. Since A0
is invertible, the rows must be linearly independent, so
the polynomials σ(ai x), restricted to terms of degree at
most d, must themselves be linearly independent. Since
the space of degree-d univariate polynomials is (d + 1)dimensional, these d+1 linearly independent polynomials
must span the space. Hence, m1 (p, σ) ≤ d + 1 for any
univariate degree-d polynomial p. In fact, we can fix the
weights from the input neuron to the hidden layer (to be
a0 , a1 , . . . , ad , respectively) and still represent any polynomial p with d + 1 hidden neurons.
Proposition II.6. Let p(x) = xd , and suppose that σ(x)
is a nonlinear function with nonzero Taylor coefficients.
Then:
1. m1 (p, σ) = d + 1.
2. m(xd , σ) ≤ 7dlog2 (d)e.
Proof. Part (1) follows from part (1) of Theorem II.1
above, by setting n = 1 and r1 = d.
For part (2), observe that we can approximate the square
x2 of an input x with three neurons in a single layer:
1
C.
Univariate polynomials
As with multivariate polynomials, depth can offer an exponential savings when approximating univariate polynomials. We show below (Proposition II.5) that a shallow
network can approximate any degree-d univariate polynomial with a number of neurons at most linear in d.
The monomial xd requires d + 1 neurons in a shallow network (Proposition II.6), but can be approximated with
2σ 00 (0)
(σ(x) + σ(−x) − 2σ(0)) = x2 + O(x4 ).
We refer to this construction as a square gate, and the
construction of [15] as a product gate. We also use identity
gate to refer to a neuron that simply preserves the input
of a neuron from the preceding layer (this is equivalent
to the skip connections in residual nets).
Consider a network in which each layer contains a square
gate (3 neurons) and either a product gate or an identity gate (4 or 1 neurons, respectively), according to the
5
following construction: The square gate squares the output of the preceding square gate, yielding inductively a
k
result of the form x2 , where k is the depth of the layer.
Writing d in binary, we use a product gate if there is a 1
in the 2k−1 -place; if so, the product gate multiplies the
output of the preceding product gate by the output of
the preceding square gate. If there is a 0 in the 2k−1 place, we use an identity gate instead of a product gate.
k
k−1
Thus, each layer computes x2 and multiplies x2
to
the computation if the 2k−1 -place in d is 1. The process
stops when the product gate outputs ud .
This network clearly uses at most 7dlog2 (d)e neurons,
with a worst case scenario where d + 1 is a power of
2.
Proof. Suppose that we can approximate p(u) using a
neural net with m neurons in a single hidden layer. Then,
there exist vectors a1 , a2 , . . . , am of weights from the input to the hidden layer such that
p(u) ≈ w1 σ(a1 · u) + · · · + wm σ(am · u).
If we consider only terms of degree d, we obtain
p(u) = σd
m
X
wi (ai · u)d
(8)
i=1
The symmetric tensor
T=
m
X
(σd wi )ai ⊗ · · · ⊗ ai
i=1
D.
Tensor decomposition
We conclude this section by noting interesting connections between the value m1 (p, σ) and the established hard
problem of tensor decomposition. Specifically, we show
(Proposition II.7) that the minimum number of neurons
required to approximate a polynomial p equals the symmetric tensor rank of a tensor constructed from the coefficients of p.
Let T be an order-d symmetric tensor of dimensions n ×
· · · × n. We say that T is symmetric if the entry Ti1 i2 ···id
is identical for any permutation of i1 , i2 , . . . , id . For T
symmetric, the symmetric tensor rank RS (T) is defined
to be the minimum r such that T can be written
T=
r
X
λi ai ⊗ · · · ⊗ ai ,
i=1
with λ ∈ R and ai ∈ Rn for each i [16]. Thus, for d = 1,
T is simply a real number and RS (T) = 1. For d = 2, T
is a symmetric matrix, and the symmetric tensor rank is
simply the rank of T, where one can take the values λi
above to be eigenvalues of the matrix T.
For a multiset S, we let π(S) denote the number of distinct permutations of S. Thus, if q1 , . . . , qs are frequencies for elements in S, we have
q1 + · · · + qs
π(S) =
.
q1 , . . . , qs
For p(u) = p(u1 , . . . , un ) a homogeneous multivariate
polynomial of degree d, we define the monomial tensor
T(p) as follows:
T(p)j1 ···jd = [uj1 · · · ujd ]p /π({j1 , . . . , jd }),
where [•]p denotes the coefficient of a monomial in p.
Proposition II.7. For p(u) = p(u1 , . . . , un ) a homogeneous multivariate polynomial of degree d, we have
m1 (p) ≥ RS (T (p)).
has symmetric tensor rank at most m. Furthermore, by
equation (8), we have
X
p(u) =
Tj1 ···jd (uj1 · · · ujd ).
1≤j1 ,...,jd ≤m
By the definition of T, the entries Tj1 ···jd are equal up to
permutation of j1 , . . . , jd . Therefore, T must equal the
monomial tensor T(p), showing that m ≥ RS (T(p)) as
desired.
Corollary II.8. Let p(u) = p(u1 , . . . , un ) be a multivariate polynomial of degree d. For c = 0, 1, . . . , d, let
pc (u) be the homogeneous polynomial obtained by taking
all terms of p(u) with degree c. If RS is the maximum
symmetric rank of T(pc (u)), then m1 (p) ≥ RS .
The proof of this statement closely follows that of Proposition II.7.
Various results are known for the symmetric rank of tensors over the complex numbers C [16]. Notably, Alexander and Hirschowitz [17] showed in a highly non-trivial
proof that the symmetric rank of a generic symmetric tensor
of dimension n and order k over C equals
d n+k−1
/ne, with the exception of a few small values
k
of k and n. However, this result does not hold over the
real numbers R, and in fact there can be several possible
generic ranks for symmetric tensors over R [16].
III.
HOW EFFICIENCY IMPROVES WITH
DEPTH
We now consider how mk (p, σ) scales with k, interpolating between exponential in n (for k = 1) and linear
in n (for k = log n). In practice, networks with modest
k > 1 are effective at representing natural functions. We
explain this theoretically by showing that the cost of approximating the product polynomial drops off rapidly as
k increases.
6
A.
ditions
Networks of constant depth
By repeated application of the shallow network construction in Lin, Tegmark, and Rolnick [15], we obtain the
following upper bound on mk (p, σ), which we conjecture
to be essentially tight. Our approach is reminiscent of
tree-like network architectures discussed e.g. in [10], in
which groups of input variables are recursively processed
in successive layers.
0 = −λ
Y
bj +
j6=i
i−1
X
Qk
j=h+1 bj
bi
h=1
+ (log 2)
k
Y
!
2bh
bj 2bi , for 1 ≤ i ≤ k
(11)
j=i+1
0=n−
k
Y
bj .
(12)
j=1
Theorem III.1. For p(x) equal to the product
x1 x2 · · · xn , and for any σ with all nonzero Taylor coefficients, we have:
1/k
mk (p, σ) = O n(k−1)/k · 2n
.
(9)
Dividing (11) by
recursion
Qk
j=i+1 bj
and rearranging gives us the
bi = bi−1 + log2 (bi−1 − 1/ log 2).
(13)
Thus, the optimal bi are not exactly equal but very slowly
increasing with i (see Figure 1).
Proof. We construct a network in which groups of the
n inputs are recursively multiplied. The n inputs are
first divided into groups of size b1 , and each group is
multiplied in the first hidden layer using 2b1 neurons (as
described in [15]). Thus, the first hidden layer includes
a total of 2b1 n/b1 neurons. This gives us n/b1 values to
multiply, which are in turn divided into groups of size
b2 . Each group is multiplied in the second hidden layer
using 2b2 neurons. Thus, the second hidden layer includes
a total of 2b2 n/(b1 b2 ) neurons.
We continue in this fashion for b1 , b2 , . . . , bk such that
b1 b2 · · · bk = n, giving us one neuron which is the product
of all of our inputs. By considering the total number of
neurons used, we conclude
mk (p, σ) ≤
k
X
i=1
n
Qi
j=1 bj
2bi =
k
X
k
Y
i=1
bj 2bi . (10)
j=i+1
FIG. 1: The optimal settings for {bi }ki=1 as n varies are shown
for k = 1, 2, 3. Observe that the bi converge to n1/k for large
n, as witnessed by a linear fit in the log-log plot. The exact
values are given by equations (12) and (13).
Setting bi = n1/k , for each i, gives us the desired bound
(9).
The following conjecture states that the bound given in
Theorem III.1 is (approximately) optimal.
In fact, we can solve for the choice of bi such that the
upper bound in (10) is minimized, under the condition
b1 b2 · · · bk = n. Using the technique of Lagrange multipliers, we know that the optimum occurs at a minimum
of the function
L(bi , λ) :=
n−
k
Y
i=1
!
bi
λ+
k
X
k
Y
i=1
Conjecture III.2. For p(x) equal to the product
x1 x2 · · · xn , and for any σ with all nonzero Taylor coefficients, we have
mk (p, σ) = 2Θ(n
1/k
)
,
(14)
bj 2bi .
j=i+1
Differentiating L with respect to bi , we obtain the con-
i.e., the exponent grows as n1/k as n → ∞.
We empirically tested Conjecture III.2 by training ANNs
to predict the product of input values x1 , . . . , xn with
7
n = 20 (see Figure 2). The rapid interpolation from
exponential to linear width aligns with our predictions.
In our experiments, we used feedforward networks with
dense connections between successive layers, with nonlinearities instantiated as the hyperbolic tangent function.
Similar results were also obtained for rectified linear units
(ReLUs) as the nonlinearity, despite the fact that this
function does not satisfy our hypothesis of being everywhere differentiable. The number of layers was varied,
as was the number of neurons within a single layer. The
networks were trained using the AdaDelta optimizer [18]
to minimize the absolute value of the difference between
the predicted and actual values. Input variables xi were
drawn uniformly at random from the interval [0, 2], so
that the expected value of the output would be of manageable size.
B.
Circuit complexity
It is interesting to consider how our results on the inapproximability of simple polynomials by polynomial-size
neural networks compare to results for Boolean circuits.
Recall that T C 0 is defined as the set of problems that
can be solved by a Boolean circuit of constant depth and
polynomial size, where the circuit is allowed to use AND,
OR, NOT, and MAJORITY gates of arbitrary fan-in.2 It is
an open problem whether T C 0 equals the class T C 1 of
problems solvable by circuits for which the depth is logarithmic in the size of the input.
In this section, we consider the feasibility of strong general no-flattening results. It would be very interesting if
one could show that general polynomials p in n variables
require a superpolynomial number of neurons to approximate for any constant number of hidden layers. That
is, for each integer k ≥ 1, we would like to prove a lower
bound on mk (p, σ) that grows fast than polynomially in
n.
Such a result might seem to address questions such as
whether T C 0 and T C 1 are equal. However, Boolean circuits compute using 0/1 values, while the neurons of our
artificial neural networks take on arbitrary real values.
To preserve 0/1 values at all neurons, we can restrict inputs to such values and take the nonlinear activation to
be the Heaviside step function:
(
σ(x) =
0
1
if x ≤ 0
if x > 0.
This gives us essentially a multi-layered perceptron, as
inspired by McCulloch and Pitts [19].
FIG. 2: Performance of trained networks in approximating
the product of 20 input variables, ranging from red (high
mean error) to blue (low mean error). The curve w =
1/k
n(k−1)/k · 2n
for n = 20 is shown in black. In the region
above and to the right of the curve, it is possible to effectively
approximate the product function (Theorem III.1).
Eq. (14) provides a helpful rule of thumb for how deep is
deep enough. Suppose, for instance, that we wish to keep
typical layers no wider than about a thousand (∼ 210 )
neurons. Eq. (14) then implies n1/k <
∼ 10, i.e., that the
number of layers should be at least
k>
∼ log10 n.
We assume also that each neuron has access to a
fixed bias constant: that is, a neuron receiving inputs
x1 , x2 , . . . , xn is of the form σ(a0 + a1 x1 + a2 x2 + · · · +
an xn ) where a0 , a1 , a2 , . . . , an are real constants. Such
a neuron corresponds to a weighted threshold gate in
Boolean circuits.
It follows from the work of [20] and [21] that such artificial
neural nets (ANNs), with constant depth and polynomial
size, have exactly the same power as T C 0 circuits. That
is, weighted threshold gates can simulate and be simulated by constant-depth, polynomial-size circuits of AND,
OR, NOT, and MAJORITY gates. The following are simple
constructions for these four types of gates using weighted
2
An AND gate evaluates True if all its inputs do so, an OR gate
evaluates True if any of its inputs do so, a NOT gate evaluates
True if its input is False and vice versa, and a MAJORITY gate
evaluates True if at least half of its inputs do so.
8
thresholds.
n
X
!
1
xi − n −
2
i=1
!
n
X
1
OR(x1 , x2 , . . . , xn ) = σ
xi −
2
i=1
1
NOT(x) = σ
−x
2
!
n
X
n−1
MAJORITY(x1 , x2 , . . . , xn ) = σ
xi −
2
i=1
AND(x1 , x2 , . . . , xn ) = σ
Thus, we should not hope easily to prove general noflattening results for Boolean functions, but the case
of polynomials in real-valued variables may be more
tractable. Simply approximating a real value by a
Boolean circuit requires arbitrarily many bits. Therefore,
performing direct computations on real values is clearly
intractable for T C 0 circuits. Moreover, related work such
as [4, 8, 13] has already proven gaps in expressivity for
real-valued neural networks of different depths, for which
the analogous results remain unknown in Boolean circuits.
IV.
CONCLUSION
We have shown how the power of deeper ANNs can be
quantified even for simple polynomials. We have proven
that there is an exponential gap between the width of
shallow and deep networks required for approximating
a given sparse polynomial. For n variables, a shallow
network requires size exponential in n, while a deep network requires at most linearly many neurons. Networks
with a constant number k > 1 of hidden layers appear
to interpolate between these extremes, following a curve
exponential in n1/k . This suggests a rough heuristic for
the number of layers required for approximating simple
functions with neural networks. For example, if we want
no layers to have more than 103 neurons, say, then the
minimum number of layers required grows only as log10 n.
of convolutional neural nets. That is, each neuron in a
layer is assumed to be connected only to a small subset of
neurons from the previous layer, rather than the entirety
of them (or some large fraction). In fact, we showed
(e.g. Prop. II.6) that there exist natural functions that
can be computed in a linear number of neurons, where
each neuron is connected to at most two neurons in the
preceding layer, which nonetheless cannot be computed
with fewer than exponentially many neurons in a single
layer, no matter how may connections are used. Our
construction can also easily be framed with reference to
the other properties mentioned in [13]: those of sharing
(in which weights are shared between neural connections)
and pooling (in which layers are gradually collapsed, as
our construction essentially does with recursive combination of inputs).
This paper has focused exclusively on the resources (notably neurons and synapses) required to compute a given
function. An important complementary challenge is to
quantify the resources (e.g. training steps) required to
learn the computation, i.e., to converge to appropriate
weights using training data — possibly a fixed amount
thereof, as suggested in [22]. There are simple functions
that can be computed with polynomial resources but require exponential resources to learn [23]. It is quite possible that architectures we have not considered increase
the feasibility of learning. For example, residual networks
(ResNets) [24] and unitary nets (see e.g. [25, 26]) are no
more powerful in representational ability than conventional networks of the same size, but by being less susceptible to the “vanishing/exploding gradient” problem,
it is far easier to optimize them in practice. We look
forward to future work that will help us understand the
power of neural networks to learn.
V.
ACKNOWLEDGMENTS
It is worth noting that our constructions enjoy the property of locality mentioned in [13], which is also a feature
This work was supported by the Foundational Questions
Institute http://fqxi.org/, the Rothberg Family Fund for
Cognitive Science and NSF grant 1122374. We thank
Scott Aaronson, Surya Ganguli, David Budden, and
Henry Lin for helpful discussions and suggestions.
[1] G. Cybenko, Mathematics of Control, Signals, and Systems (MCSS) 2, 303 (1989).
[2] K. Hornik, M. Stinchcombe, and H. White, Neural networks 2, 359 (1989).
[3] K.-I. Funahashi, Neural networks 2, 183 (1989).
[4] G. F. Montufar, R. Pascanu, K. Cho, and Y. Bengio,
in Advances in Neural Information Processing Systems
(NIPS) (2014), pp. 2924–2932.
[5] M. Bianchini and F. Scarselli, IEEE transactions on neu-
ral networks and learning systems 25, 1553 (2014).
[6] B. Poole, S. Lahiri, M. Raghu, J. Sohl-Dickstein, and
S. Ganguli, in Advances In Neural Information Processing Systems (NIPS) (2016), pp. 3360–3368.
[7] M. Raghu, B. Poole, J. Kleinberg, S. Ganguli,
and J. Sohl-Dickstein, arXiv preprint arXiv:1611.08083
(2016).
[8] M. Telgarsky, Journal of Machine Learning Research
(JMLR) 49 (2016).
9
[9] R. Eldan and O. Shamir, in 29th Annual Conference on
Learning Theory (2016), pp. 907–940.
[10] H. Mhaskar, Q. Liao, and T. Poggio, arXiv preprint
arXiv:1603.00988v4 (2016).
[11] O. Delalleau and Y. Bengio, in Advances in Neural Information Processing Systems (NIPS) (2011), pp. 666–674.
[12] J. Martens, A. Chattopadhya, T. Pitassi, and R. Zemel,
in Advances in Neural Information Processing Systems
(NIPS) (2013), pp. 2877–2885.
[13] N. Cohen, O. Sharir, and A. Shashua, Journal of Machine
Learning Research (JMLR) 49 (2016).
[14] N. Cohen and A. Shashua, in International Conference
on Machine Learning (ICML) (2016).
[15] H. W. Lin, M. Tegmark, and D. Rolnick, arXiv preprint
arXiv:1608.08225v3 (2016).
[16] P. Comon, G. Golub, L.-H. Lim, and B. Mourrain, SIAM
Journal on Matrix Analysis and Applications 30, 1254
(2008).
[17] J. Alexander and A. Hirschowitz, Journal of Algebraic
Geometry 4, 201 (1995).
[18] M. D. Zeiler, arXiv preprint arXiv:1212.5701 (2012).
[19] W. S. McCulloch and W. Pitts, The bulletin of mathematical biophysics 5, 115 (1943).
[20] A. K. Chandra, L. Stockmeyer, and U. Vishkin, SIAM
Journal on Computing 13, 423 (1984).
[21] N. Pippenger, IBM journal of research and development
31, 235 (1987).
[22] C. Zhang, S. Bengio, M. Hardt, B. Recht, and O. Vinyals,
in Proceedings of the International Conference on Learning Representations (ICLR) (2017).
[23] S. Shalev-Shwartz, O. Shamir, and S. Shammah, arXiv
preprint arXiv:1703.07950 (2017).
[24] K. He, X. Zhang, S. Ren, and J. Sun, in Proceedings of
the IEEE Conference on Computer Vision and Pattern
Recognition (2016), pp. 770–778.
[25] M. Arjovsky, A. Shah, and Y. Bengio, in International
Conference on Machine Learning (2016), pp. 1120–1128.
[26] L. Jing, Y. Shen, T. Dubček, J. Peurifoy, S. Skirlo,
M. Tegmark, and M. Soljačić, arXiv preprint
arXiv:1612.05231 (2016).
| 9cs.NE
|
arXiv:1509.06332v2 [cs.DS] 23 Sep 2015
A note on linear fractional set packing problem
Pooja Pandey ∗
Department of Mathematics, Simon Fraser University
250 - 13450 102nd Avenue, Surrey, BC, V3T 0A3, Canada
Abstract
In this note we point out various errors in the paper by Rashmi Gupta and R.
R. Saxena, Set packing problem with linear fractional objective function, International
Journal of Mathematics and Computer Applications Research (IJMCAR), 4 (2014) 9
- 18. We also provide some additional results.
1
Introduction
The set packing problem, set covering problem, and set partitioning problem are among
the most well-studied problems in combinatorial optimization and they have wide range of
real life applications [1, 2, 3, 4, 5, 8]. Arora, Puri and Swarup [1, 2] developed couple of
solution algorithms for the set covering problem with the linear fractional objective function,
where they exploited the structural properties of the set covering problems. Later, Gupta
and Saxena [7] extended these results for the set packing problems with the linear fractional
objective function.
In this note, we show that the properties established in [7] are incorrect. Gupta and
Saxena [6] extended results of [1, 2] to the linear fractional set packing problems, but these
extensions suffer many drawbacks since they overlooked the structural properties of the set
packing problem and ignore the required conditions for their results to be correct.
2
The linear fractional set packing problem
Let E = {1, 2, . . . , m} be a finite set and F = {S1 , S2 , . . . , Sn } be a family of subsets of E.
The index set for elements of F is denoted by G = {1, 2, . . . , n}. For each element j ∈ G, a
cost cj and a weight dj are prescribed. We refer to cj as the linear cost of the set Sj and
c = (c1 , . . . , cn ) as the linear cost vector. Similarly dj is referred to as the linear weight of
the set Sj and d = (d1 , . . . , dn ) as the linear weight vector. α and β are constants where β > 0.
∗
Corresponding author. Email: [email protected]
1
A subset H of G is said to be a pack of E if
T
Sj Sk = ∅.
S
j∈H
Sj = E, and j, k ∈ H, j 6= k, implies
Then
Ph the linear set packing problem (LSPP) is to select a pack H = {π(1), . . . , π(h)} such
that i=1 cπ(i) is maximized. Likewise the linear fractional set packing problem (LFSPP) is
Ph
cσ(i) + α
to select a pack H = {σ(1), . . . , σ(h)} such that Phi=1
is maximized.
d
+
β
σ(i)
i=1
For each i ∈ E, consider the vector ai = (ai1 , ai2 , . . . , ain ) where
(
1 if i ∈ Sj
aij =
0 otherwise.
and A = (aij )m×n be an m × n matrix. Also, consider the decision variables x1 , x2 , . . . , xn
where
(
1 if j is in the pack
xj =
0 otherwise.
The vector of decision variables is represented as x = (x1 , . . . , xn )T and is column
vector of size n where all entries equal to 1. Then the LSPP and LFSPP can be formulated
respectively as 0-1 integer programs
LSPP:
Maximize cx
Subject to Ax ≤
x ∈ {0, 1}n
(1)
(2)
and
LFSPP:
cx + α
dx + β
Subject to Ax ≤
x ∈ {0, 1}n
Maximize
(3)
(4)
It is assumed that c ≥ , α = 0 and β is a scalar such that dx + β > 0. Throughout the
paper we will assume that dx + β > 0 for any feasible solution of LFSPP.
The continuous relaxations of LSPP and LFSPP, denoted respectively by LSPP(C) and
LFSPP(C), are obtained by replacing the constraint set x ∈ {0, 1}n by x ≥ , respectively
in LSPP and LFSPP.
The family of feasible solutions of both LSPP and LFSPP is given by S = {x|Ax ≤
, x ∈ {0, 1}n } and the family of feasible solutions for their continuous relaxations is given
by S̄ = {x|Ax ≤ , x ≥ }.
2
Following are some definitions given in [7]. A solution x ∈ S which satisfies (3) and (4)
is said to be a pack solution. For any pack H, a column of A corresponding to j ∈ G is said
to be redundant if H + {j} is also a pack. If a pack corresponds to one or more redundant
columns, it is called a redundant pack. A pack H ∗ is said to be a prime pack, if none of the
columns corresponding to j ∗ ∈ G is redundant. A solution corresponding to the prime pack
is called a prime packing solution.
The linear fractional set covering problem (LFSCP) is obtained by replacing (3) with
Ax ≥
(5)
and changing the problem from maximization to minimization problem in the 0-1 integer
formulation of LFSPP. Any x ∈ {0, 1}n satisfying (5) is called a cover solution.
Any x ∈ S is called a cover solution and an optimal solution to the underlying problem
LFSCP is called an optimal cover solution. Note that each cover solution corresponds to a
cover and viceversa. A cover P is said to be redundant if P − {j} with j ∈ P is also a cover.
A cover which is not redundant is called a prime cover. The incidence vector x corresponds
to prime cover is called a prime cover solution.
For the linear fractional set covering problem, Arora, Puri, and Swarup [1] proved that
every optimal cover is a prime cover, if cj ′ s and dj ′ s satisfy certain conditions.
Gupta and Saxena [7] claimed an extension of the above result to LFSPP, assuming
c ≥ , d ≥ and dx + β > 0. More precisely, they claimed:
Theorem 2.1. (Theorem 2 of [7]) If the objective function in LFSPP has finite value then,
there exists a prime pack solution where this value is attained.
This result is not true as established by the following example. Let
1 1 0
β = 2, c = (1, 2, 5), d = (4, 4, 6), and A =
1 0 1
For the LFSPP with A, c, d and β defined as above, it can be verified that x∗ = (0, 0, 1)T
5
is an optimal solution with the objective function value
= 0.625. The optimal pack
8
corresponding to x∗ is H ∗ = {3} which is a redundant pack since H ∗ + {2} = {2, 3} is also a
pack. All other pack solutions and their respective objective function values are listed below:
x1 = (1, 0, 0)T prime pack solution
x2 = (0, 1, 1)T prime pack solution
x3 = (0, 0, 0)T redundant pack solution
x4 = (0, 1, 0)T redundant pack solution
3
1
= 0.16666
6
7
= 0.5833
f (x2 ) =
12
f (x3 ) = 0
1
f (x4 ) = = 0.3333
3
f (x1 ) =
None of these corresponds to an optimal solution for LFSPP. In particular, no prime
pack solution is optimal for the instances of LFSPP constructed above, which contradicts
Theorem 2.1. However, a variation of the Theorem 2.1 can be proved as noted below.
Theorem 2.2. There always exists a prime pack optimal solution for LFSPP if
(a) c > and d < or
(b) any ratio of the partial sums of ci ’s or di ’s is greater than the value of the objective
function at the pack solution and also the partial sum of di ’s is positive.
Proof. Let us assume that there exists an optimal pack H1 for LFSPP which is a redundant
optimal pack of LFSPP. Since H1 is a redundant pack, a prime pack H2 can be derived from
H1 by adding redundant columns of H1 .
The objective function value of LFSPP for H1 and H2 are ZH1 and ZH2 respectively:
P
P
P
cj
j∈H1 cj +
j∈H1 cj
P j∈{H2 −H1 }
Z H1 = P
and ZH2 = P
j∈H1 dj + β
j∈H1 dj +
j∈{H2 −H1 } dj + β
(a) Since c > and d < , therefore,
cj +
j∈H1
X
j∈H1
dj + β >
X
dj +
X
and
X
j∈H1
cj <
j∈H1
X
cj
(6)
X
dj + β
(7)
j∈{H2 −H1 }
j∈{H2 −H1 }
Since denominator of the objective function value is always positive, dividing inequality
(6) by (7), we get:
which gives
P
P
P
cj
j∈H1 cj +
j∈H1 cj
P
P j∈{H2 −H1 }
<P
j∈H1 dj + β
j∈H1 dj +
j∈{H2 −H1 } dj + β
Z H1 < Z H2 .
This shows that H2 is an optimal pack for LFSPP instead of H1 , contradicts the optimality of H1 , therefore H1 is a prime pack of LFSPP . This completes the proof of
part (a).
Note: In this case if we relax c > , then above theorem is no longer true.
4
(b) It is given that any ratio of the partial sums of ci ’s or di ’s is greater than the value of the
objective function at the pack solution and also the partial sum of di ’s is positive and
denominator of the objective function is positive for any feasible solution of LFSPP,
therefore:
P
P j∈{H2 −H1 }
j∈{H2 −H1 }
P
cj
> ZH1 , which implies
dj
P
j∈H1 cj
>P
, cross multiplication gives
j∈H1 dj + β
X
X
cj )
>(
dj )(
j∈{H2 −H1 } cj
P
j∈{H2 −H1 }
(
X
cj )(
X
dj
dj + β)
j∈H1
j∈{H2 −H1 }
adding both sides (
P
j∈H1
X
(
cj )(
cj )(
P
X
j∈H1
dj + β) will give
X
cj )(
X
cj )(
dj + β) + (
X
X
dj + β) >
j∈H1
j∈{H2 −H1 }
j∈H1
j∈H1
X
dj + β) + (
j∈H1
j∈H1
(
j∈H1
j∈{H2 −H1 }
dj )(
j∈{H2 −H1 }
X
cj )
j∈H1
after simplifying, we get
(
X
j∈H1
dj + β)(
X
cj +
j∈H1
X
cj ) > (
X
cj )(
j∈H1
j∈{H2 −H1 }
X
j∈H1
dj + β +
X
dj )
j∈{H2 −H1 }
which is equivalent to
P
P
P
( j∈H1 cj + j∈{H2 −H1 } cj
( j∈H1 cj )
P
P
.
> P
( j∈H1 dj + β)
j∈H1 dj + β +
j∈{H2 −H1 } dj )
which gives
Z H2 > Z H1 .
Therefore, the objective function value of any prime pack is always greater than any
corresponding redundant pack, therefore, the optimal pack is a prime pack. This
completes the proof of part (b).
Lemma 2.3. If c and d are integer numbers, k, l = 1, . . . , n, ld + β > 0 and k + l ≤ n, then
kc + lc
kc
<
,
kd + β
kd + ld + β
5
if c > 0
and
kc + lc
kc
>
,
kd + β
kd + ld + β
if c < 0.
Proof. For given integer numbers c and d, if c > 0 then
lcβ > 0
(8)
add (kc)(kd + β) + (kc)(ld) both sides of the inequality (8), we get
(kc)(kd + β) + (kc)(ld) + lcβ > (kc)(kd + β) + (kc)(ld)
(9)
after rearranging inequality (9) we get
(kc)(kd + β) + (lc)(kd + β)
(kd + β)(kc + lc)
> (kc)(kd + β) + (kc)(ld) which is same as
> (kc)(kd + ld + β)
(10)
since (kd + β) > 0 and (kd + ld + β) > 0, divide both sides of the inequality (10) by
(kd + β)(kd + ld + β) , we get
kc
kc + lc
<
kd + β
kd + ld + β
(11)
which proves the first part.
Now if c < 0 then
kc + lc
kc
>
,
kd + β
kd + ld + β
this can be proved in a similar manner as we did the first part.
This completes the proof.
Theorem 2.4. If ci = c, and di = d, ∀i = 1, . . . , n, and H ∗ is an optimal pack of LFSPP:
(a) if c > 0, then H ∗ is a largest cardinality prime pack solution of LFSPP, (b) if c < 0,
then H ∗ is a smallest cardinality pack solution of LFSPP.
Proof. (a) If c > 0 :
if H ∗ is not a prime pack then we can always add redundant columns for H ∗ and find a
prime pack H ∗∗ of LFSPP and corresponding objective function values for H ∗ and H ∗∗ are
following:
P
P
P
∗∗
∗ c
j∈H ∗ c +
j∈H ∗ c
P j∈{H −H }
and ZH ∗∗ = P
ZH ∗ = P
j∈H ∗ d + β
j∈H ∗ d +
j∈{H ∗∗ −H ∗ } d + β
using first part of Lemma 2.3 we can claim that
ZH ∗ < ZH ∗∗
6
which contradicts the optimality of H ∗ . Therefore, H ∗ is a prime pack of the given LFSPP.
If among all prime packs of LFSPP H ∗ is not of the largest cardinality then there exist
a prime cover H o of LFSPP such that |H ∗| < |H o | and
P
P
j∈H o c
j∈H ∗ c
and ZH o = P
ZH ∗ = P
j∈H ∗ d + β
j∈H o d + β
P
P
since |H ∗| < |H o| , then j∈H ∗ c < j∈H o c, along with the first part of Lemma 2.3 we
can claim that
ZH ∗ < ZH o
which contradicts the optimality of H ∗ . Therefore, H ∗ if the prime pack of the largest
cardinality among all prime packs of LFSPP.
(b) If c < 0 :
If among all packs of LFSPP H ∗ is not of the smallest cardinality then there exist a pack
H of LFSPP such that |H ∗| > |H o | and
P
P
j∈H ∗ c
j∈H o c
ZH ∗ = P
and ZH o = P
j∈H ∗ d + β
j∈H o d + β
P
P
but since |H ∗ | > |H o| , then j∈H ∗ c < j∈H o c, along with the second part of Lemma
2.3 we can claim that
ZH ∗ < ZH o
o
which contradicts the optimality of H ∗ . Therefore, H ∗ is the pack of LFSPP of the
smallest cardinality. This completes the proof.
3
Conclusion
I would like to explore non-linear set packing problems in future research. I would like to
thank Prof. Abraham P. Punnen for his valuable suggestions during the preparation of this
note.
References
[1] Arora, S.R., Swaroop, Kanti, and Puri, M.C., The set covering problem with linear
fractional functional, Indian Journal of Pure and Applied Mathematics, 8, (1977) 578588.
[2] Arora, S.R., and Puri, M.C., Enumeration Technique for the Set Covering Problem with
Linear Fractional Functional as its Objective Functions, ZAMM - Journal of Applied
Mathematics and Mechanics, 57, (1977) 181-186.
7
[3] Bazaraa, Mokhtar S.; Goode, Jamie J, A cutting-plane algorithm for the quadratic setcovering problem, Operations Research, 23, (1975) 150 - 158.
[4] Bector, C. R. and Bhatt. S. K., A linearization technique for solving integral linear
fractional program, Proc. fifth Manitoba Conference on Numerical Mathematics, (1975)
221 - 229.
[5] Garfinkel, M. and Nemhauser. G. L, Integer Programming, A Wiley-Interscience Publication, John Wiley and Sons (1973).
[6] Gupta, Rashmi, and Saxena, R. R., Linearization technique for solving quadratic set
packing and partitioning problems, International Journal of Mathematics and Computer
Applications Research, 4, (2014) 9 - 20.
[7] Gupta, Rashmi, and Saxena, R. R., Set packing problem with linear fractional objective
function, International Journal of Mathematics and Computer Applications Research, 4,
(2014) 9 - 18.
[8] Lemke, C. E., Salkin, H. M. and Spielberg K., Set covering by single branch enumeration
with linear programming sub-problem, Operations Research, 19, (1971) 998 - 1022.
8
| 8cs.DS
|
THE AGGREGATED UNFITTED FINITE ELEMENT METHOD FOR ELLIPTIC
PROBLEMS
arXiv:1709.09122v1 [cs.CE] 26 Sep 2017
SANTIAGO BADIA, FRANCESC VERDUGO, AND ALBERTO F. MARTÍN
Abstract. Unfitted finite element techniques are valuable tools in different applications where the generation of body-fitted meshes is difficult. However, these techniques are prone to severe ill conditioning
problems that obstruct the efficient use of iterative Krylov methods and, in consequence, hinders the practical usage of unfitted methods for realistic large scale applications. In this work, we present a technique
that addresses such conditioning problems by constructing enhanced finite element spaces based on a cell
aggregation technique. The presented method, called aggregated unfitted finite element method, is easy to
implement, and can be used, in contrast to previous works, in Galerkin approximations of coercive problems with conforming Lagrangian finite element spaces. The mathematical analysis of the new method
states that the condition number of the resulting linear system matrix scales as in standard finite elements
for body-fitted meshes, without being affected by small cut cells, and that the method leads to the optimal finite element convergence order. These theoretical results are confirmed with 2D and 3D numerical
experiments.
Keywords: unfitted finite elements; embedded boundary methods; ill-conditioning.
Contents
1. Introduction
2. Embedded boundary setup and cell aggregation
3. Aggregated unfitted Lagrangian finite element spaces
4. Approximation of elliptic problems
5. Numerical analysis
5.1. Stability of the coordinate vector extension matrix
5.2. Mass matrix condition number
5.3. Inverse inequality
5.4. Coercivity and Nitsche’s coefficient
5.5. Well-posedness of the unfitted finite element (FE) problem
5.6. Error estimates
6. Numerical experiments
6.1. Setup
6.2. Moving domain experiment
6.3. Convergence test
7. Conclusions
References
1
3
5
7
9
9
10
10
11
11
13
14
14
15
16
17
19
1. Introduction
Unfitted FE techniques are specially appealing when the generation of body-fitted meshes is difficult.
They are helpful in a number of contexts including multi-phase and multi-physics applications with moving
interfaces (e.g., fracture mechanics, fluid-structure interaction [1], or free surface flows), or in situations
in which one wants to avoid the generation of body-fitted meshes to simplify as far as possible the preprocessing steps (e.g., shape or topology optimization frameworks, medical simulations based on CT-scan
Date: September 27, 2017.
SB gratefully acknowledges the support received from the Catalan Government through the ICREA Acadèmia Research
Program. E-mails: [email protected] (SB), [email protected] (FV), [email protected] (AM).
1
AGGREGATED UNFITTED FINITE ELEMENT METHOD
2
data, or parallel large-scale simulations). In addition, the huge success of isogeometrical analysis (splinebased discretization) and the severe limitations of this approach in complex 3D geometries will probably
increase the interest of unfitted methods in the near future [2]. Unfitted FE methods have been named in
different ways. When designed for capturing interfaces, they are usually denoted as eXtended FE methods
(XFEM) [3], whereas they are usually denoted as embedded (or immersed) boundary methods, when
the motivation is to simulate a problem using a (usually simple Cartesian) background mesh (see, e.g.,
[4, 5, 6]).
Yet useful, unfitted FE methods have known drawbacks. They pose problems to numerical integration,
imposition of Dirichlet boundary conditions, and lead to ill conditioning problems. Whereas different
techniques have been proposed in the literature to address the issues related with numerical integration
(see, e.g., [7]) and the imposition of Dirichlet boundary conditions (see, e.g., [8]), the conditioning problems
are one of the main showstoppers still today for the successful use of this type of methods in realistic large
scale applications. For most of the unfitted FE techniques, the condition number of the discrete linear
system does not only depend on the characteristic element size of the background mesh, but also on
the characteristic size of the cut cells, which can be arbitrary small and have arbitrarily high aspect
ratios. This is an important problem. At large scales, linear systems are solved with iterative Krylov subspace methods [9] in combination with scalable preconditioners. Unfortunately, the well known scalable
preconditioners based on (algebraic) multigrid [10] or multi-level domain decomposition [11] are mainly
designed for body-fitted meshes and cannot readily deal with cut cells. Different preconditioners for
unfitted FE methods have been recently proposed, but they are mainly serial non-scalable algorithms
(see, e.g., [12, 13, 14, 15]). Recently, a robust domain decomposition preconditioner able to deal with
cut cells has been proposed in [16]. Even though this method has proven to be scalable in some complex
3D examples, it is based on heuristic considerations without a complete mathematical analysis and its
application to second (and higher) order FEs is involved. This lack of preconditioners for unfitted FEs
can be addressed with enhanced formulations that provide well-posed discrete systems independently of
the size of the cut cells. Once the conditioning problems related to cut cells are addressed, the application
of standard preconditioners for body-fitted meshes to the unfitted case is strongly simplified, opening the
door to large-scale computations.
The main goal of this work is to develop such an enhanced unfitted FE formulation that fixes the problems associated with cut cells. The goal is to achieve condition numbers that scale only with the element
size of the background mesh in the same way as in standard FE methods for body-fitted meshes. Our
purpose is to implement it in FEMPAR, our in-house large scale FE code [17]. Since FEMPAR is a parallel multi-physics multi-scale code that includes different continuous and discontinuous FE formulations
and several element types, it is crucial for us that the novel formulation fulfills the following additional
properties: 1) It should be general enough to be applied to several problem types, 2) it should deal with
both continuous and discontinuous FE formulations, 3) it should deal with high order interpolations, and
4) it should be easily implemented in an existing parallel FE package.
To our best knowledge, none of the existing unfitted FE formulations fulfill these requirements simultaneously. For instance, one can consider the ghost penalty formulation used in the CutFEM method [4, 18]
However, it leads to a weakly non-consistent algorithm, and it requires to compute high order derivatives
on faces for high order FEs, which are not at our disposal in general FE codes and are expensive to
compute, certainly complicating the implementation of the methods and harming code performance. Alternatively, for finite volume and discontinuous Galerkin (DG) formulations, one can consider the so-called
cell aggregation (or agglomeration) techniques [19, 20]. E.g., for DG formulations, the idea is simple: cells
with the small cut cell problem, i.e., the ratio between the volume of the cell inside the physical domain
and the total cell volume is close to zero, are merged with neighbor full cells forming aggregates. A new
polynomial space is defined in each aggregate that replaces the local FE spaces of all cells merged in it.
This process fixes the conditioning problems, since the support of the newly defined shape functions is at
least the volume of a full cell. Even though this idea is simple and general enough to deal with different
problem types and high order interpolations, the resulting discrete spaces are such that the enforcement
of continuity through appropriate local-to-global degrees of freedom (DOFs) numbering, as in standard
FE codes (see, e.g., [17]), is not possible, limiting their usage to discontinuous Galerkin or finite volume
formulations. Up to our best knowledge, there is no variant of cell agglomeration currently proposed
AGGREGATED UNFITTED FINITE ELEMENT METHOD
3
in the literature producing conforming FE spaces, which could be used for classical continuous Galerkin
formulations. It is the purpose of this work.
In this article, we present an alternative cell aggregation technique that can be used for both continuous
and discontinuous formulations, the aggregated unfitted FE method. We start with the usual (conforming)
Lagrangian FE space that includes cut cells, which is known to lead to conditioning problems. The main
idea is to eliminate from this space all the potentially problematic DOFs by introducing a set of judiciously
defined constraints. These constraints are introduced using information provided by the cell aggregates,
without altering the conformity of the original FE space. Alternatively, the method can be understood
as an extension operator from the interior (well-posed ) FE space that only involves interior cells to a
larger FE space that includes cut cells and covers the whole physical domain. Discontinuous spaces can
also be generated as a particular case of this procedure, which makes the method compatible also with
DG formulations. In contrast to previous works, we also include a detailed mathematical analysis of the
method, in terms of well-posedness, condition number estimates, and a priori error estimates. For elliptic
problems, we mathematically prove that 1) the method leads to condition numbers that are independent
from small cut cells, 2) the condition numbers scale with the size of the background mesh as in the standard
FE method, 3) the penalty parameter of Nitsche’s method required for stability purposes is bounded above,
and 4) the optimal FE convergence order is recovered. These theoretical results are confirmed with 2D
and 3D numerical experiments using the Poisson equation as a model problem.
The outline of the article is as follows. In Section 2, we introduce our embedded boundary setup and
the strategy to build the cell aggregates. In Section 3, we describe the construction of the novel FE spaces
based on the cell aggregates. In Section 4, we introduce our elliptic model problem. The numerical analysis
of the method is carried out in Section 5. Finally, we present a complete set of numerical experiments in
Section 6 and draw some conclusions in Section 7.
2. Embedded boundary setup and cell aggregation
Let Ω ⊂ Rd be an open bounded polygonal domain, with d ∈ {2, 3} the number of spatial dimensions.
For the sake of simplicity and without loss of generality, we consider in the numerical experiments below
.
that the domain boundary is defined as the zero level-set of a given scalar function ψ ls , namely ∂Ω = {x ∈
Rd : ψ ls (x) = 0}.1 We note that the problem geometry could be described using 3D CAD data instead of
level-set functions, by providing techniques to compute the intersection between cell edges and surfaces
(see, e.g., [21]). In any case, the way the geometry is handled does not affect the following exposition. Like
in any other embedded boundary method, we build the computational mesh by introducing an artificial
domain Ωart such that it has a simple geometry that is easy to mesh using Cartesian grids and it includes
the physical domain Ω ⊂ Ωart (see Fig. 1a).
internal cells
cut cells
external cells
(a)
(b)
Figure 1. Embedded boundary setup.
Let us construct a partition of Ωart into cells, represented by Thart , with characteristic cell size h. We
are interested in Thart being a Cartesian mesh into hexahedra for d = 3 or quadrilaterals for d = 2, even
1Analogous assumption have to be made for body-fitted methods.
AGGREGATED UNFITTED FINITE ELEMENT METHOD
4
though unstructured n-simplex background meshes can also be considered. Cells in Thart can be classified as
follows: a cell K ∈ Thart such that K ⊂ Ω is an internal cell ; if K ∩ Ω = ∅, K is an external cell ; otherwise,
K is a cut cell (see Fig. 1b). The set of interior (resp., external and cut) cells is represented with Thin and
its union Ωin ⊂ Ω (resp., (Thext , Ωext ) and (Thcut , Ωcut )). Furthermore, we define the set of active cells as
.
Thact = Thin ∪ Thcut and its union Ωact . In the numerical analysis, we assume that the background mesh is
quasi-uniform (see, e.g., [22, p.107]) to reduce technicalities, and define a characteristic mesh size h. The
maximum element size is denoted with hmax .
We can also consider non-overlapping cell aggregates AK composed of cut cells and one interior cell
K such that the aggregate is connected, using, e.g., the strategy described in Algorithm 2.1. It leads to
another partition Thagg defined by the aggregations of cells in Thact ; interior cells that do not belong to any
aggregate remain the same. By construction of Algorithm 2.1, there is only one interior cell per aggregate,
denoted as the root cell of the aggregate, and every cut cell belongs to one and only one aggregate. For a
cut cell, we define its root cell as the root of the only aggregate that contains the cut cell. The root of an
interior cell is the cell itself. Thus, there is a one-to-one mapping between aggregates (including interior
cells) AK ∈ Thagg and the root cut cell K ∈ Thin . As a result, we can use the same index for the aggregate
and the root cell. We build the aggregates in Thagg with Algorithm 2.1. In any case, other aggregation
algorithms could be considered, e.g., touching in the first step of the algorithm not only the interior cells,
but also cut cells without the small cut cell problem. It can be implemented by defining the quantity
.
ηK = |K∩Ω|
|K| and touch in the first step not only the interior cells but also any cut cell with ηK > η0 > 0
for a fixed value η0 .
Algorithm 2.1 (Cell aggregation scheme).
(1) Mark all interior cells as touched and all cut cells as untouched.
(2) For each untouched cell, if there is at least one touched cell connected to it through a facet F such
that F ∩ Ω 6= ∅, we aggregate the cell to the touched cell belonging to the aggregate containing
the closest interior cell. If more than one touched cell fulfills this requirement, we choose one
arbitrarily, e.g., the one with smaller global id.
(3) Mark as touched all the cells aggregated in 2.
(4) Repeat 2. and 3. until all cells are aggregated.
touched
(a) Step 1
untouched
(b) Step 2
Aggregates’ boundary
(c) Step 3
∂Ω
(d) Step 4
Figure 2. Illustration of the cell aggregation scheme defined in Algorithm 2.1.
Fig. 2 shows an illustration of each step in Algorithm 2. The black thin lines represent the boundaries
of the aggregates. Note that from step 1 to step 2, some of the lines between adjacent cells are removed,
meaning that the two adjacent cells have been merged in the same aggregate. The aggregation schemes
can be easily applied to arbitrary spatial dimensions. As an illustrative example, Fig. 3 shows some of
the aggregates obtained for a complex 3D domain.
In the forthcoming sections, we need an upper bound of the size of the aggregates generated with
Algorithm 2.1. To this end, let us consider the next lemma.
AGGREGATED UNFITTED FINITE ELEMENT METHOD
5
Figure 3. 3D aggregates.
Lemma 2.2. Assume that from any cut cell K0 ∈ Thact there is a cell path {K0 , K1 , . . . , Kn } that satisfies:
1) two consecutive cells share a facet F such that F ∩ Ω 6= ∅; 2) Kn is an interior cell; 3) n ≤ γmax , where
γmax is a fixed integer. Then, the maximum aggregate size is at most (2γmax + 1)hmax .
Proof. By construction, an aggregate can grow at most at a rate of one layer of elements per each iteration.
Thus, after n iterations the aggregate size will be at most (2n + 1)hmax considering that the aggregate can
potentially grow in all spatial directions. It is obvious to see that the aggregation scheme finishes at most
after γmax iterations. Thus, the aggregate size will be less or equal than (2γmax + 1)hmax .
From Lemma 2.2, it follows that the aggregate size will be bounded if so is the value of γmax . In what
follows, we assume that γmax is fixed, e.g., eliminating any cut cell that would violate property 3) in
Lemma 2.2. One shall assume that each cut cell shares at least one corner with an interior cell (this is
usually true if the grid is fine enough to capture the geometry). In this situation, we can easily see that
γmax = 2 for 2D and γmax = 3 for 3D. Then, by Lemma 2.2, the aggregate size is at most 5hmax in 2D and
7hmax in 3D. Even though it is not used in the proof of Lemma 2.2, the fact that we aggregate cut cells
to the touched cells belonging to the aggregate containing the closest interior cell (see step 2 in Algorithm
2.1) contributes to further reduce the aggregate size. Indeed, the actual size of the aggregates generated
in the numerical examples (cf. Section 6) is much lower than the predicted by these theoretical bounds.
In 2D, the aggregate size tends to 2hmax as the mesh is refined, whereas it tends to 3hmax in the 3D case.
This shows that the aggregation scheme produces relative small aggregates in the numerical experiments.
3. Aggregated unfitted Lagrangian finite element spaces
Our goal is to define a FE space using the cell aggregates introduced above. To this end, we need to
.
introduce some notation. In the case of n-simplex meshes, we define the local FE space V (K) = Pq (K),
i.e., the space of polynomials of order less or equal to q in the variables x1 , . . . , xd . For n-cube meshes,
.
V (K) = Qq (K), i.e., the space of polynomials that are of degree less or equal to k with respect to each
variable x1 , . . . , xd . In this work, we consider that the polynomial order q is the same for all the cells in
the mesh. We restrict ourselves to Lagrangian FE methods. Thus, the basis for V (K) is the Lagrangian
basis (of order q) on K. We denote by N (K) the set of Lagrangian nodes of order q of cell K. There is a
one-to-one mapping between nodes a ∈ N (K) and shape functions φa (x); it holds φa (xb ) = δab , where xb
are the space coordinates of node b. We assume that there is a local-to-global DOF map such that the
resulting global system is C 0 continuous. This process can be elaborated for hp-adaptivity as well, but it
is not the purpose of this work.
With this notation, we can introduce the active FE space associated with the active portion of the
background mesh
.
Vhact = {v ∈ C 0 (Ωact ) : v|K ∈ V (K), for any K ∈ Thact }.
We could analogously define the interior FE space
.
Vhin = {v ∈ C 0 (Ωin ) : v|K ∈ V (K), for any K ∈ Thin }.
AGGREGATED UNFITTED FINITE ELEMENT METHOD
6
The active FE space Vhact (see Fig. 4c) is the functional space typically used in unfitted FE methods (see,
e.g., [16, 15, 6]). It is well known that Vhact leads to arbitrary ill conditioned systems when integrating the
FE weak form on the physical domain Ω only (if no extra technique is used to remedy it). It is obvious
that the interior FE space Vhin (see Fig. 4a) is not affected by this problem, but it is not usable since it
is not defined on the complete physical domain Ω. Instead, we propose an alternative space Vhagg that is
defined on Ω but does not present the problems related to Vhact . We can define the set of nodes of Vhin and
.
Vhact as Nhin and Nhact , respectively (see Fig. 4). We define the set of outer nodes as Nhout = Nhact \ Nhin
(marked with red crosses in Fig. 4b). The outer nodes are the ones that can lead to conditioning problems
due to the small cut cell problem (see (9)). The space Vhagg is defined taking as starting point Vhact , and
adding judiciously defined constraints for the nodes in Nhout .
◦ nodes in Nhin
• nodes in Nhact
× nodes in Nhout
(a) Vhin
(b) Vhagg
(c) Vhact
Figure 4. Finite Element spaces.
In order to define Vhagg we observe that, in nodal Lagrangian FE spaces, there is a one-to-one map
between DOFs and nodes (points) of the FE mesh (for vector spaces, the same is true for every component
of the vector field). On the other hand, we can define the owner vertex, edge, or face (VEF) of a node
as the lowest-dimensional VEF that contains the node. Furthermore, we can construct a map that for
every VEF F such that F 6⊂ Ω, gives a cell owner among all the cells that contain it. This map can be
arbitrarily built. E.g., we can consider as cell owner the one in the smallest aggregate. As a result, we
have a map between DOFs and (active) cells. Every active cell belongs to an aggregate, which has its own
root (interior) cell. So, we also have a map between DOFs and interior cells. This map between b ∈ Nhout
and the corresponding interior cell is represented with K(b) (see Fig. 5).
×
aggregate
node in Nhout
node to cell map
Figure 5. Map from outer nodes to interior cells.
The space of global shape functions of Vhin and Vhact can be represented as {φb : b ∈ Nhin } and
: b ∈ Nhact }, respectively. Functions in these FE spaces are uniquely represented by their nodal
in
values. We represent the nodal values of uh ∈ Vhin as uin ∈ R|Nh | , whereas the nodal values of uh ∈ Vhact
act
as uact ∈ R|Nh | . Considering, without loss of generality, that the interior nodal values are labeled the
out
same way for both FE spaces, we have that uact = [uin , uout ]T , where uout ∈ R|Nh | .
{φb
AGGREGATED UNFITTED FINITE ELEMENT METHOD
7
Now, we consider the following extension operator. Given uh ∈ Vhin and the corresponding nodal values
we compute the outer nodal values as follows:
X
φa (xb )uin a , for b ∈ Nhout .
uout b =
(1)
uin ,
a∈N (K(b))
That is, the value at an outer node b ∈ Nhout is computed by extrapolating the nodal values of the interior
cell K(b) associated with it. In compact form, we can write it as uout = Cuin , where C is the global
in
act
matrix of constraints. We define the global extension matrix E : R|Vh | → R|Vh | as Euin = [uin , Cuin ]T .
Let us also define the extension operator E : Vhin → Vhact , such that, given uh ∈ Vhin represented by its
nodal values uin , provides the FE function E(uh ) ∈ Vhact with nodal values Euin . We define the range of
.
this operator as Vhagg = range(E(Vhin )) ⊂ Vhact . This FE space is called the aggregated FE space since the
map K(·) between outer nodes and interior cells is defined using the aggregates in Thagg . The motivation
behind the construction of such space is to have a FE space covering Ωact (and thus Ω) with optimal
approximability properties and without the ill-conditioning problems of Vhact .
As one can observe, the new space is defined only by interior nodal values, whereas the conflictive outer
nodes are eliminated via the constraints in (1). These constraints are cell-wise local. Thus, they can be
readily applied at the assembly level in the cell loop, making its implementation very simple, even for nonadaptive codes that cannot deal with non-conforming meshes. We consider as basis for Vhagg the extension
of the shape functions of Vhin , i.e., {E(φa )}a∈N in . The fact that it is a basis for Vhagg is straightforward,
h
due to the fact that the extension operator is linear. The extension of a shape function is easily computed
as follows:
X
E(φa ) = φa +
Cba φb , for a ∈ Nhin ,
b∈C(a)
where C(a) represents the set of outer nodes in Nhout that are constrained by a.
Remark 3.1. We note that one could consider an alternative aggregated space,
Vhagg,∗ = {v ∈ C 0 (Ω) : v|A ∈ V (A), for any A ∈ Thagg },
where V (A) denotes the space of q order Lagrangian polynomials on n-simplices or n-cubes. It is obvious
to check that in fact Vhagg,∗ ⊂ Vhagg , but it is not possible to implement the inter-element continuity for this
space using standard FE techniques. On the other hand, the FE space Vhagg has the same size as the interior
problem and the implementation in existing FE codes requires minimal modifications. Furthermore, it is
also easy to check that the two approaches coincide for DG formulations, where all DOFs belong to the
cells. In fact, a DG method with Vhagg,∗ has been proposed in [20].
4. Approximation of elliptic problems
For the sake of simplicity, we consider the Poisson equation with constant physical diffusion as a model
problem, even though the proposed ideas apply to any elliptic problem with H 1 -stability, e.g., the linear elasticity problem and heterogeneous problems. The Poisson equation with Dirichlet and Neumann
boundary conditions reads as (after scaling with the diffusion term): find u ∈ H 1 (Ω) such that
− ∆u = f
in Ω,
u = gD
on ΓD ,
∇u · n = g N
on ΓN ,
(2)
where (ΓD , ΓN ) is a partition of the domain boundary (the Dirichlet and Neumann boundaries, respectively), f ∈ H −1 (Ω), g D ∈ H 1/2 (ΓD ), and g N ∈ H −1/2 (ΓN ).
For the space discretization, we consider H 1 -conforming FE spaces on the conforming mesh Thact that
are not necessary aligned with the the physical boundary ∂Ω. For simplicity, we assume that, for any cut
cell K ∈ Thact , either K ∩ Γ ⊂ ΓD or K ∩ Γ ⊂ ΓN . We consider both the usual FE space Vhact as well as
the new aggregated space Vhagg in order to compare their properties. We will simply use Vh when it is not
necessary to distinguish between Vhact and Vhagg .
For unfitted grids, it is not clear to include Dirichlet conditions in the approximation space in a strong
manner. Thus, we consider Nitsche’s method [23, 24] to impose Dirichlet boundary conditions weakly on
AGGREGATED UNFITTED FINITE ELEMENT METHOD
8
ΓD . It provides a consistent numerical scheme with optimal converge rates (also for high-order elements)
that is commonly used in the embedded boundary community [6]. We define the FE-wise operators:
Z
Z
.
(τK uv − v (n · ∇u) − u (n · ∇v)) dS,
∇u · ∇v dV +
AK (u, v) =
ΓD ∩K
K∩Ω
Z
.
`K (v) =
τK vg D − (n · ∇v) g D dS,
ΓD ∩K
defined for a generic cell K ∈ Thact . Vector n denotes the outwards normal to ∂Ω. The bilinear form
AK (·, ·) includes the usual form resulting from the integration by parts of (2) and the additional term
associated with the weak imposition of Dirichlet boundary conditions with Nitsche’s method. The righthand side operator `K (·) includes additional terms related to Nitsche’s method. The coefficient τK > 0 is
a mesh-dependent parameter that has to be large enough to ensure the coercivity of AK (·, ·).
The global FE operator A : Vh → Vh 0 and right-hand side term ` ∈ Vh 0 are stated as the sum of the
element contributions, i.e.,
. X
. X
A(u, v) =
AK (u, v), `(v) =
`K (v),
for u, v ∈ Vh .
(3)
K∈Thact
K∈Thact
We will make abuse of notation, using the same symbol for a bilinear form, e.g., A : Vh → Vh 0 ,
.
and its corresponding linear operator, i.e., hAu, vi = A(u, v). Furthermore, we define b : Vh 0 → Vh as
.
b(v) = f (v) + g N (v) + `(v), for v ∈ Vh . With this, the global problem can be stated as: find uh ∈ Vhagg
such that
(4)
A(uh , vh ) = b(vh ),
for any vh ∈ Vhagg .
By definition, this problem can analogously be stated as: find uh ∈ Vhin such that A(E(uh ), E(vh )) =
b(E(vh )) for any vh ∈ Vhagg . After the definition of the FE basis (of shape functions) that spans Vhagg , or
alternatively the extension operator E(·), the previous problem leads to a linear system to be solved.
A sufficient (even though not necessary) condition for A to be coercive is to enforce the element-wise
constant coefficient τK to satisfy
BK (v, v)
.
τK ≥ CK = sup
,
(5)
v∈V (K) DK (v, v)
for all the mesh elements K ∈ Thact intersecting the boundary ΓD . In the previous formula, DK (·, ·) and
BK (·, ·) are the forms defined as
Z
Z
.
.
DK (u, v) =
∇u · ∇v dV, and BK (u, v) =
(n · ∇u) (n · ∇v) dS.
K∩Ω
ΓD ∩K
Since Vh is finite dimensional, and DK (·, ·) and BK (·, ·) are symmetric and bilinear forms, the value CK
.
(i.e., the minimum admissible coefficient τK ) can be computed numerically as CK = λ̃max , being λ̃max the
largest eigenvalue of the generalized eigenvalue problem (see [15] for details): find uK ∈ Vh |K and λ̃ ∈ R
such that
BK (uK , vK ) = λ̃DK (uK , vK ) for all vK ∈ Vh |K .
(6)
For standard FEs for body-fitted meshes, it is enough to compute coefficient τK as τK = β/h to
satisfy condition (5), where β is a sufficiently large (mesh independent) positive constant (see, e.g., [25]).
However, for standard unfitted FE methods using the usual space Vhact without any additional stabilization,
coefficient τK cannot be computed a priori; in fact, the minimum cell-wise value that assures coercivity is
not bounded above. In this case, a value for τK ensuring coercivity has to be computed for each particular
setup using the cell-wise eigenvalue problem (6). The introduction of the new space Vhagg solves this
problem and τK is bounded again in terms of the element size as expected in the body-fitted case (see
Section 5.4 for more details). In this case, we have taken τK = 100/hK in the numerical experiments
below.
The linear system matrix that arises from (4) can be defined as
.
Aab = A(E(φa ), E(φb )),
for a, b ∈ Nhin .
(7)
AGGREGATED UNFITTED FINITE ELEMENT METHOD
The mass matrix related to the aggregated FE space Vhagg is analogously defined as
Z
.
Mab =
E(φa )E(φb ),
for a, b ∈ Nhin .
9
(8)
Ω
It is well known that the usual FE space Vhact is associated with conditioning problems due to cut cells.
The condition number of the discrete system without the aggregation, i.e., considering Vhact instead of
Vhagg in (4), scales as
−(2q+1−2/d)
κ(A) ∼ min ηK
,
(9)
K∈Thact
where κ(A) is the 2-norm condition number of A (see [15] for details). Thus, arbitrarily high condition
numbers are expected in practice since the position of the interface cannot be controlled and the value ηK
can be arbitrarily close to zero. This problem is solved if the new aggregated space Vhagg is used instead
of Vhact (cf. Corollary 5.9).
5. Numerical analysis
In this section, we analyze the well-posedness of the agregated unfitted FE method (4), the condition
number of the arising linear system, and a priori error estimates. As commented above, we assume
that the background mesh is quasi-uniform. Therefore, the number of neighboring cells of a given cell is
bounded above by a constant ncell independently of h. In a mesh refinement analysis, we also assume that
the coarser mesh level-set function already represents the domain boundary.
In the following analysis, all constants being used are independent of h and the location of the cuts in
cells, i.e., ηK . They may also depend on the threshold η0 in the aggregation algorithm if considered; we
have considered η0 = 1 for simplicity. The constants can depend on the shape/size of Ω and ΓD , the order
of the FE space, and the maximum aggregation distance γmax , which are assumed to be fixed in this work.
In turn, due to Lemma 2.2, the maximum size of an aggregate is bounded by a constant times h. As a
result, the following results are robust with respect to the so-called small cut cell problem. When we have
that A ≤ cB for a positive constant c, we may use the notation A . B; analogously for &.
For the analysis below, we need to introduce some extra notation. Given a function uh ∈ Vhin (or Vhact ),
the nodal vector u will be used without any superscript, as soon as it is clear from the context. For
a given cell K, the cell-wise coordinate vector is represented with uK . On the other hand, given a FE
function uh ∈ Vhin , for every interior cell K ∈ Thin , let us define define the cell-wise extension operator
EAK u = [uK , CÂK uK ], where CÂK is the cell-wise constraint matrix, whose entries can be computed in
P
the reference space (see (1)), such that Cu · Cu = K∈T in CÂK uK · CÂK uK . We denote with k · k2 the
h
Euclidean norm of a vector and the induced matrix norm. Standard notation is used to define Sobolev
spaces (see, e.g., [26]). Given a Sobolev space X, its corresponding norm is represented with k · kX .
5.1. Stability of the coordinate vector extension matrix. We start the analysis of the scheme by
proving bounds for the norm of the global and cell-wise coordinate vector extension matrix. Therefore,
their norms can be bounded independently of the cut location and the size of the aggregate.
Lemma 5.1. The cell-wise and global coordinate vector extension matrices hold the following bounds:
1 ≤ kEAK k22 ≤ 1 + kCÂK k22 ,
for every AK ∈ Thagg ,
and
1 ≤ kEk22 ≤ 1 + kCk22 ≤ Ce ,
for a positive constant Ce .
Proof. Using the definition of the extension operator in Section 3, we have that kEuk22 = kuk22 + kCuk22 .
We proceed analogously for the cell-wise result, to get kEAK uk22 = kuk22 + kCÂK uk22 . It proves the first
result. On the other hand, we have,
X
X
kCuk22 =
kCÂK uK k22 ≤
kCÂK k22 kuK k22 ≤ ncell sup kCÂK k22 kuk22 ,
AK ∈Thagg
AK ∈Thagg
AK ∈Thagg
where we have used the fact that the constraint matrix is aggregate-wise and that the maximum number of
cell neighbors of a vertex/edge/face is bounded above by a constant ncell . The value supAK ∈T agg kCÂK k22
h
(or an upper bound) can explicitly be computed prior to the numerical integration and its entries are
independent of the aggregate cut and the geometrical mapping, i.e., h. In fact, given a polynomial order
AGGREGATED UNFITTED FINITE ELEMENT METHOD
10
and γmax , one can precompute the maximum value of kCÂK k22 among all possible aggregate configurations
and explicitly obtain an upper bound Ce of the global extension matrix norm. It proves the lemma.
5.2. Mass matrix condition number. In order to provide a bound for the condition number of the
mass matrix, we rely on the maximum and minimum eigenvalues of the local mass matrix in the reference
cell K̂:
λ− kuK k22 ≤ kuh k2L2 (K̂) ≤ λ+ kuK k22 ,
for uh ∈ V (K).
The values of λ− and λ+ only depend on the order of the FE space and can be computed for different
orders on n-cubes or n-simplices (see [27]). Using typical scaling arguments, one has the following bound
for the local mass matrix of the physical cell:
λ− hdK kuK k22 ≤ kuh k2L2 (K) ≤ λ+ hdK kuK k22 .
(10)
In the next lemma, we prove the equivalence between the L2 (Ωact ) norm and the interior DOF Euclidean
norm, for functions in Vhagg .
Lemma 5.2. The following bounds hold:
hd kuk22 . kE(uh )k2L2 (Ωact ) . hd kuk22 ,
for any uh ∈ Vhin .
(11)
Proof. By definition, every function in Vhagg can be expressed as E(uh ) for some uh ∈ Vhin . Using (10), the
fact that Ωin ⊂ Ω, and the quasi-uniformity of the background mesh, we obtain the lower bound in (11)
as follows:
X
X
kE(uh )k2L2 (Ωact ) ≥ kuh k2L2 (Ωin ) =
kuh k2L2 (K) ≥
hdK λ− kuK k22 & hd kuk22 .
K∈Thin
K∈Thin
On the other hand, using Ω ⊂ Ωact , Lemma 5.1, (10), and the fact that the number of surrounding cells
of a node is bounded above by a positive constant, we get:
X
X
kE(uh )k2L2 (Ωact ) =
kE(uh )k2L2 (K) ≤
hdK λ+ kuK k22 . hd kEuk22 . hd kuk22 .
K∈Thact
K∈Thact
It proves the lemma.
The upper and lower bounds in (11) lead to the continuity of the extension operator and a bound for
the condition number of the mass matrix of the aggregated FE space.
Corollary 5.3 (Continuity of the extension operator). The extension operator satisfies the following
bound:
kE(uh )k2L2 (Ωact ) . kuh k2L2 (Ωin ) ,
for any uh ∈ Vhin .
Corollary 5.4 (Mass matrix condition number). The mass matrix M in (8), related to the aggregated FE
space Vhagg , is bounded by κ(M) ≤ C, for a positive constant C.
5.3. Inverse inequality. In order to prove the condition number bound for the system matrix arising
from (4), we need to prove first an extended inverse inequality. We rely on the fact that an inverse
inequality holds for the FE space Vhact , i.e.,
k∇uh kL2 (Ωact ) . h−1 kuh kL2 (Ωact ) ,
for any uh ∈ Vhact .
(12)
This standard result for conforming meshes can be found, e.g., in [22, p. 111].
Lemma 5.5 (Inverse inequality). The following inverse inequality holds:
k∇E(uh )kL2 (Ωact ) . h−1 kuh kL2 (Ωin ) ,
for any uh ∈ Vhin .
Proof. Using the fact that Ωin ⊆ Ω ⊆ Ωact , E(uh ) ∈ Vhact , the standard inverse inequality (12), and the
stability of the extension operator in Lemma 5.3, we get:
k∇E(uh )kL2 (Ωact ) . h−1 kE(uh )kL2 (Ωact ) . h−1 kuh kL2 (Ωin ) .
It proves the lemma.
AGGREGATED UNFITTED FINITE ELEMENT METHOD
11
5.4. Coercivity and Nitsche’s coefficient. In this section, we consider a trace inequality that is needed
to prove the coercivity of the bilinear form in (3). Given a cell K ∈ Thact , let us consider the set of
constraining interior cells K1 , . . . , KmK , mK ≥ 1, i.e., the interior S
cells that constraint at least one DOF
.
.
K
of the cut cell. Let us also define Kcut = K ∩ Ω and ΩK = Kcut ∪ m
i=1 Ki ⊂ Ω.
Lemma 5.6. For any uh ∈ Vhagg and K ∈ Thact , the following bound holds
−1
kn · ∇uh kL2 (ΓD ∩K) ≤ C∂ hK 2 k∇uh kL2 (ΩK ) ,
for a positive constant C∂ .
Proof. For interior cells, the left-hand side is zero and the bound trivially holds. Let us consider a cut cell
.
K. Let us also consider a FE function uh ∈ Vhagg and its gradient ξ h = ∇uh . Assuming that all the cells
have the same order, we have that ξ h belongs to the discontinuous Lagrangian FE space of order q − 1,
and we represent the corresponding coordinate vector with ξ K .
First, we use the equivalence of norms in finite dimension and a scaling argument to get:
kξ h k2L2 (ΓD ∩K) . |ΓD ∩ K|kξ h k2L∞ (K) ,
where the constant can only depend on the FE space order. Analogously, we have kξ h k2L∞ (K̂) . kξ K k22 .
Following the same ideas as above, ξ can be expressed as an extension of the corresponding nodal values of
the q − 1 order FE spaces on top of the interior cells Ki , represented with ξ K ; we represent this extension
i
with the matrix DK , i.e., ξ K = DK [ξ 1 , . . . , ξ m ]T . Using an analogous reasoning as above for matrix C,
P K
2
the norm of this matrix cannot depend on the cut or h. Thus, we have that kξ K k22 . m
i=1 kξ Ki k2 . On the
other hand, using again the equivalence of norms in finite dimension, we get kξ K k2 . kξ h kL2 (K̂i ) . As a
i
result, using typical scaling arguments, and using the fact that |K| . |Ki | . |K| for constants independent
of mesh size or order, we get:
mK
X
2
−1
kξ h kL∞ (K) . |K|
kξ h k2L2 (Ki ) .
i=1
Combining these results, we get:
kn ·
ξ h k2L2 (Γ∩AK )
≤ |Γ ∩
K||K|kξ h k2L∞ (K)
.
h−1
K
mK
X
kξ h k2L2 (Ki ) ,
i=1
where we have used the fact that |Γ ∩
lemma.
K||K|−1
.
h−1
K
holds for a quasi-uniform mesh. It proves the
5.5. Well-posedness of the unfitted FE problem. In this section, we prove coercivity and continuity
of the bilinear form (3). First, we prove coercivity with respect to the following mesh dependent norm in
Vhagg :
X
.
2
|||uh |||2h = k∇uh k2L2 (Ω) +
βK h−1
for uh ∈ Vhagg ,
K kuh kL2 (ΓD ∩K) ,
K∈Thact
which is next proved to bound the L2 (Ω) norm.
Theorem 5.7. The aggregated unfitted FE problem in (4) satisfies the following bounds:
i) Coercivity:
A(uh , uh ) & |||uh |||2h , for any uh ∈ Vhagg ,
(13)
ii) Continuity:
A(uh , vh ) . |||uh |||h |||vh |||h , for uh , vh ∈ Vhagg ,
(14)
if βK > C, for some positive constant C. In this case, there exists one and only one solution of (4).
Proof. For cut cells, we use
Z
−1
2
−1
2
uh (n · ∇uh ) ≤ αK C∂ h−1
hK kn · ∇uh k2L2 (ΓD ∩K)
K kuh kL2 (ΓD ∩K) + αK C∂
ΓD ∩K
−1
2
2
≤ αK C∂ h−1
K kuh kL2 (ΓD ∩K) + αK k∇uh kL2 (ΩK ) .
(15)
AGGREGATED UNFITTED FINITE ELEMENT METHOD
12
Using the fact that the mesh is quasi-uniform and that the number of neighboring cells and γmax is
bounded, one can take a value for αK large enough (but uniform with respect to h and the cut location)
such that:
Z
X
1
2
2
2
uh (n · ∇uh ) ≤
αK C∂ h−1
K kuh kL2 (ΓD ∩K) + k∇uh kL2 (Ω) .
2
ΓD
act
K∈Th
As a result, we get:
X
1
2
(βK − αK C∂ ) h−1
A(uh , uh ) ≥ k∇uh k2L2 (Ω) +
K kuh kL2 (ΓD ∩K) .
2
act
K∈Th
For, e.g., βK > 2αK C∂ , A(uh , uh ) is a norm. By construction, this lower bound for βK is independent of
the mesh size h and the intersection of Γ and Thact . It proves the coercivity property in (13). Thus, the
bilinear form is non-singular. The continuity in (14) can readily be proved by repeated use of the CauchySchwarz inequality and inequality (15). Since the problem is finite-dimensional and the corresponding
linear system matrix is non-singular, there exists one and only one solution of this problem.
Lemma 5.8. If Ω has smoothing properties, the following bound holds:
for any uh ∈ Vhagg .
kuh kL2 (Ω) . |||uh |||h ,
Proof. Let us consider uh ∈ Vhagg and let ψ ∈ H01 (Ω) solve the problem −∆ψ = uh with the boundary
conditions ψ = 0 on ΓD and n·∇ψ = 0 on ΓN . Using the fact that the domain Ω has smoothing properties,
it holds kψkH 2 (Ω) . kuh kL2 (Ω) . We have, after integration by parts:
Z
Z
Z
kuh kL2 (Ω) = − uh ∆ψ =
∇uh · ∇ψ −
uh n · ∇ψ.
(16)
Ω
Ω
ΓD
The first term in the right-hand side of (16) is easily bounded using the Cauchy-Schwarz inequality:
Z
∇uh · ∇ψ ≤ k∇uh kL2 (Ω) k∇ψkL2 (Ω) . k∇uh kL2 (Ω) kuh kL2 (Ω) .
Ω
On the other hand, the following trace inequality holds kn · ∇ψk2L2 (ΓD ) . |ψ|2H 2 (Ω) for a constant that
depends on the size of ΓD (see [26]). Using the Cauchy-Schwarz inequality and the previous trace inequality,
we readily get:
1
1
2
2
Z
X
X
−1
2
2
uh n · ∇ψ ≤
h kuh k 2
−
hK kn · ∇ψk 2
K
ΓD
L (ΓD ∩K)
L (ΓD ∩K)
K∈Thact
K∈Thact
1
≤
2
X
2
h−1
K kuh kL2 (ΓD ∩K)
kn · ∇ψkL2 (ΓD )
K∈Thact
1
.
2
X
2
h−1
K kuh kL2 (ΓD ∩K)
kuh kL2 (Ω) .
K∈Thact
Combining these bounds, we prove the lemma.
Corollary 5.9 (Stiffness matrix condition number). The condition number of the linear system matrix A
in (7) is bounded by κ(A) . h−2 .
Proof. To prove the corollary, we have to bound u · Au = A(E(uh ), E(uh )) above and below by kuk22
times some constant. The lower bound follows from the coercivity property in Th. 5.7, Lemma 5.8, the
lower bounds in Lemmas 5.2 and 5.1, which lead to A(E(uh ), E(uh )) & kE(uh )k2L2 (Ω) & hd kuk22 . The upper
bound is readily obtained from the continuity property in Lemma 5.7 and the upper bound in Lemma 5.8,
i.e., u · Au = A(E(uh ), E(uh )) . |||E(uh )|||2h . Using scaling arguments and the equivalence of norms for
2
finite-dimensional spaces, we get kE(uh )k2L2 (ΓD ∩K) . hd−1
K kEuK k2 . Adding up for all cells, invoking the
AGGREGATED UNFITTED FINITE ELEMENT METHOD
13
fact that the number of neighbour cells is bounded, and using the upper bound of the coordinate vector
extension operator in 5.8, we obtain:
X
d−2
2
βK h−1
kuk22 .
(17)
K kE(uh )kL2 (ΓD ∩K) . h
K∈Thact
Using the inverse inequality in Lemma 5.5 and the upper bound in Lemma 5.2, we obtain:
k∇E(uh )k2L2 (Ω) . h−2 kE(uh )k2L2 (Ω) . hd−2 kuk22 .
Combining (17)-(18), we get u · Au ≤ chd−2 kuk22 . It proves the corollary.
(18)
5.6. Error estimates. In this section, we get a priori error estimates for the aggregated FE scheme (4).
In order to do that, we prove first approximability properties of the corresponding spaces.
Lemma 5.10. Let us consider an aggregated FE space of order q, m ≤ q, 1 ≤ s ≤ m ≤ q + 1, 1 ≤ p ≤ ∞,
and m > dp . Given a function u ∈ H m (Ω), it holds:
inf
uh ∈Vhagg
ku − uh kWps (Ω) . hm−s |u|Wpm (Ω) .
Proof. Under the assumptions of the lemma, we have that the following embedding Wpm (Ω) ⊂ C 0 (Ω̄) is
continuous (see, e.g., [28, p. 486]). Thus, given a function u ∈ Wpm (Ω), let represent with σ(u) the
vector ofP
nodal values in Ωin , i.e., σ(u)a = uh (xa ) for a ∈ N (Ωin ). We define the interpolation operator
.
Ih (u) = a∈N (Ω) φa [Eσ(u)]a .
Given a cut cell K ∈ Thact , the fact that its DOFs values only depend on interior DOFs in Ω̄K , and since
m (K) ⊆ W m (K), it follows from the upper bound of the norm of the
each shape function φa belongs to W∞
p
nodal extension operator in Lemma 5.1 that kIh (u)kWpm (K) ≤ CkukC 0 (Ω̄K ) (see also [22, Lemma 4.4.1]).
On the other hand, we consider an arbitrary function π(u) ∈ Wpm (Ω) such that π(u)|K ∈ Pq (ΩK ). We
note that, by construction, π(u)|K = Ih (π(u))|K . Thus, we have:
ku − Ih (u)kWpm (K) ≤ ku − π(u)kWpm (K) + kIh (π(u) − u)kWpm (K)
. ku − π(u)kWpm (K) + kπ(u) − ukC 0 (Ω̄K ) . ku − π(u)kWpm (ΩK ) ,
where we have used in the last inequality the previous continuous embedding. Since ΩK is an open bounded
domain with Lipschitz boundary by definition, one can use the Deny-Lions lemma (see, e.g., [28]). As a
result, the π(u) that minimizes the right-hand side holds:
ku − Ih (u)kWpm (ΩK ) . |u|Wpm (ΩK ) .
Using standard scaling arguments, we prove the lemma.
Theorem 5.11. If Ω has smoothing properties and the solution u of the continuous problem (2) belongs
to Wpm (Ω) for dp < m ≤ q, the solution uh ∈ Vhagg of (4) satisfies the following a priori error estimate:
ku − uh kH 1 (Ω) ≤ hm−1 |u|H m (Ω) .
Proof. Combining the consistency of the numerical method, i.e., A(u, vh ) = `(vh ), and the continuity and
coercivity of the bilinear form in Th. 5.7, we readily get, using standard FE analysis arguments:
|||wh − uh |||2h . A(wh − uh , wh − uh ) = A(wh − u, wh − uh ) . |||wh − u|||h |||wh − uh |||h ,
for any wh ∈ Vhagg . On the other hand, we use the trace inequality (see [29])
kψk2L2 (∂T ) . |∂T |−1 kψk2L2 (T ) + |∂T |kψk2H 1 (T ) ,
for any ψ ∈ H 1 (T ).
Using this trace inequality, we get:
−1
−2
2
2
2
2
h−1
K ku − wh kL2 (ΓD ∩K) ≤ hK ku − wh kL2 (∂ΩK ) . hK ku − wh kL2 (ΩK ) + ku − wh kH 1 (ΩK ) .
Combining the previous bound with the approximability property in Lemma 5.10, we readily get
|||wh − u|||h . hm−1 |u − wh |H m (Ω) .
It proves the theorem.
AGGREGATED UNFITTED FINITE ELEMENT METHOD
14
6. Numerical experiments
6.1. Setup. The numerical examples below consider as a model problem the Poisson equation with nonhomogeneous Dirichlet boundary conditions. The value of the source term and the Dirichlet function are
defined such that the PDE has the following manufactured exact solution:
1/2
2
2
2
,
u(x, y, z) = sin 4π (x − 2.3) + y + z
(x, y) ∈ Ω ⊂ R2 , z = 0 in 2D,
(x, y, z) ∈ Ω ⊂ R3 in 3D.
We consider two different geometries, a 2D circle and a 3D complex domain with the shape of a popcorn
flake (see Fig. 6). These geometries are often used in the literature to study the performance of unfitted
FE methods (see, e.g., [4], where the definition of the popcorn flake is found). In all cases, we use the
cuboid [0, 1]d , d = 2, 3 as the bounding box on top of which the background Cartesian grid is created. For
the sake of illustration, Fig. 6 displays both the considered geometries, numerical solution and bounding
box.
(a) 2D case
(b) 3D case
Figure 6. View of the problem geometries, numerical solution and bounding box.
The main goal of the following tests is to evaluate the (positive) effect of using the aggregation-based FE
space Vhagg instead of the usual one Vhact . In the next plots, the results for the usual (un-aggregated) FE
space are labeled as standard, whereas the cases with the aggregation are labeled as aggregated (or aggr.
in its short form). In all the examples, we use Lagrangian reference FEs with bi-linear and bi-quadratic
shape functions in 2D, and tri-linear and tri-quadratic ones in 3D.
Both the standard and the aggregated formulations have been implemented in the object-oriented HPC
code FEMPAR [17]. The system of linear equations resulting from the problem discretization are solved
within FEMPAR with a sparse direct solver from the MKL PARDISO package [30]. Condition number
estimates are computed outside FEMPAR using the MATLAB function condest.2 For the standard
unfitted FE space Vhact , we expect very high condition numbers that can hinder the solution of the discrete
system using standard double precision arithmetic. To address this effect and avoid the breakdown of sparse
direct solvers, we bound from below the minimum distance between the mesh nodes and the intersection
of edges with the boundary Γ to a small numerical threshold Dmin proportional to the cell size, namely
Dmin = εh, where ε is a (mesh independent) user defined tolerance. If the edge cut-node distance is
below this threshold , the edge cut is collapsed with the node, perturbing the geometry. In the numerical
experiments, we take ε = 10−6 and ε = 10−3 in 2D and 3D respectively. Using the fact that ηK ∼ εd , we can
rewrite the condition number estimate (9) in terms of the user-defined tolerance ε as κ(A) ∼ ε−d(2q+1−2/d) .
2MATLAB is a trademark of THE MATHWORKS INC.
AGGREGATED UNFITTED FINITE ELEMENT METHOD
15
For instance, we have κ(A) ∼ ε−7 and κ(A) ∼ ε−13 for first and second order interpolations, respectively,
in 3D. This illustrates that the condition numbers expected for second order interpolation are extremely
high as it is confirmed below unless very large values of ε are considered. However, the value of ε cannot be
increased without affecting the numerical error, since it perturbs the geometry, and destroys at some point
the order of convergence of the numerical method. Similar perturbation-based techniques with analogous
problems have been used in the frame of the finite cell method in [8]. We note that the tolerance ε is not
needed at all when using the aggregated FE space.
6.2. Moving domain experiment. In the first numerical experiment, we study the robustness of the
unfitted FE formulations with respect to the relative position between the unfitted boundary and the
background mesh. To this end, we consider two moving domains that can travel along one of the diagonals
of the bounding box (see Fig. 7). The considered geometries are obtained by scaling down the circle
and the popcorn flake depicted in Fig. 6 by a factor of 0.25. In both cases, the position of the bodies is
controlled by the value of the parameter ` (i.e., the distance between the center of the body and a selected
vertex of the box). As the value of ` varies, the objects move and their relative position with respect to
the background mesh changes. In this process, arbitrary small cut cells can show up, leading to potential
conditioning problems. In this experiment, we consider a background mesh with element size h = 2−5 .
(a) 2D case
(b) 3D case
Figure 7. Setup of the moving domain experiment.
Fig. 8 shows the condition number estimate of the underlying linear systems varying the position of
the physical domain Ω. The plot is generated using a sample of 200 different values of `. It is observed
that the condition numbers are very sensitive to the position of the domain for the standard unfitted
FE formulation, whereas the condition numbers are nearly independent of the position when using the
aggregation-based FE spaces. Note that the standard formulation leads to very high condition numbers
specially for second order interpolations and the 3D case. Moving from 1st order to 2nd order leads to a
rise in the condition number between 10 and 15 orders of magnitude. The same disastrous effect is observed
when moving from 2D to 3D. In contrast, the condition number is nearly insensitive to the number of
space dimensions, and mildly depends on the interpolation order (as for body-fitted methods) when using
aggregation-based FE spaces.
From the results shown in Fig. 8, it is clear that the aggregation-based FE spaces are able to dramatically
improve the condition numbers associated with the standard unfitted FE formulation. The next question
is how cell aggregation impacts on the accuracy of the numerical solution. In order to quantify this effect,
Fig. 9 shows the computed energy norm of the discretization error. It is observed that the error is slightly
increased when using the aggregation-based FE spaces. This is because the considered meshes in this
moving domain experiment are rather coarse. The error increments become negligible for finer meshes
AGGREGATED UNFITTED FINITE ELEMENT METHOD
p=1, standard
p=2, standard
p=1, aggr.
p=2, aggr.
log10(condest(A))
25
20
15
10
30
p=1, standard
p=2, standard
p=1, aggr.
p=2, aggr.
25
log10(condest(A))
30
16
5
20
15
10
5
0
0
0.3
0.4
0.5
0.6
0.7
0.3
0.4
0.5
ℓ
0.6
0.7
ℓ
(a) 2D case
(b) 3D case
Figure 8. Condition number vs. domain position.
(see Section 6.3 below). In this example, we cannot compute a solution for all the values of ` for 3D and
2nd order interpolation without using cell aggregation (see the discontinuous fine red curve in Fig. 9b).
The condition numbers are so high (order 1030 ) that the system is intractable, even with a sparse direct
solver, using standard double precision floating point arithmetic.
1
1
log10(Error energy norm)
0.5
0
-0.5
-1
-1.5
-2
-2.5
-3
0.3
0.4
0.5
0.6
0.7
p=1, standard
p=2, standard
p=1, aggr.
p=2, aggr.
0.5
log10(Error energy norm)
p=1, standard
p=2, standard
p=1, aggr.
p=2, aggr.
0
-0.5
-1
-1.5
-2
-2.5
-3
0.3
0.4
0.5
ℓ
(a) 2D case
0.6
0.7
ℓ
(b) 3D case
Figure 9. Error energy norm vs. domain position.
6.3. Convergence test. The second experiment is devoted to study the asymptotic behavior of the
methods as the mesh is refined. To this end, we consider the geometries and bounding boxes displayed in
Fig. 6, which are discretized with uniform Cartesian meshes with element sizes h = 2−m , m = 3, 4, . . . , 9
in 2D, and m = 3, 4, 5, 6 in 3D.
First, we study how the size of the aggregates scales when the mesh is refined. Fig. 10 shows that the
aggregate size is 2h in 2D, whereas it tends to 3h in the 3D case. These results agrees with the theoretical
bounds for the aggregate size discussed in Section 2.
Then, we study the scaling of the condition numbers with respect to the mesh size (see Fig. 11). For the
aggregation-based FE spaces, the condition numbers of the stiffness matrix scales as h−2 , like in standard
FE methods for body fitted meshes. This confirms the theoretical result of Corollary 5.9. Conversely,
the condition number has an erratic behavior if cell aggregation is not considered. The reason is that, as
shown in the previous experiment (cf. Section 6.2), the standard unfitted FE formulation leads to condition
numbers very sensitive to the position of the unfitted boundary. Several configurations of cut cells can show
up when the mesh is refined, leading to very different condition numbers. As in the previous experiment,
the condition number is very sensitive to the interpolation order and number of space dimensions for the
standard unfitted FE formulations. This effect is reverted when using cell aggregates.
AGGREGATED UNFITTED FINITE ELEMENT METHOD
17
8
H/h
6
4
2D
3D
Bound in 2D
Bound in 3D
2
0
-3.5
-3
-2.5
-2
-1.5
-1
log10(h)
Figure 10. Scaling of the ratio maximum aggregate size (H) / element size (h).
40
40
p=1, standard
p=2, standard
p=1, aggr.
p=2, aggr.
slope -2
35
30
30
25
25
log10(condest(A))
log10(condest(A))
35
20
15
20
15
10
10
5
5
0
p=1, standard
p=2, standard
p=1, aggr.
p=2, aggr.
slope -2
0
-2.5
-2
-1.5
-1
log10(h)
-2.5
-2
-1.5
-1
log10(h)
(a) 2D case
(b) 3D case
Figure 11. Scaling of the condition number upon mesh refinement.
Finally, we study the convergence of the discretization error. To this end, Figs. 12 and 13 report the
discretization errors measured both in the energy norm and in the L2 norm. Like in the previous experiment
(cf. Section 6.2), the discrete system could not be solved when using the finest meshes in 3D for 2nd order
interpolation without using cell-aggregation due to extremely large condition numbers (see the incomplete
curve in Fig. 12b). The results show that the error increment associated with the aggregation-based FE
space becomes negligible when the mesh is refined. Moreover, the theoretical results of Section 5.6 are
confirmed: optimal order of convergence is (asymptotically) achieved in all cases when using aggregationbased FE spaces both for the energy and the L2 norm, for 1st and 2nd order interpolations, and for 2 and
3 spatial dimensions.
7. Conclusions
We have proposed a novel technique to construct FE spaces designed to improve the conditioning
problems associated with unfitted FE methods. The spaces are defined using cell aggregates obtained
by merging the cut cells to interior cells. In contrast to related methods in the literature, the proposed
technique is easy to implement in existing FE codes (it only involves cell-wise constraints) and it is general
enough to deal with both continuous and DG formulations. Another novelty with respect to previous works
is that we include the mathematical analysis of the method. For elliptic problems, we have proved that 1)
the novel FE space leads to condition numbers that are independent from small cut cells, 2) the condition
1
1
0
0
-1
-1
-2
-2
log10(Error energy norm)
log10(Error energy norm)
AGGREGATED UNFITTED FINITE ELEMENT METHOD
-3
-4
-5
-6
-3
-4
-5
-6
p=1, standard
p=2, standard
p=1, aggr.
p=2, aggr.
slope 1
slope 2
-7
18
p=1, standard
p=2, standard
p=1, aggr.
p=2, aggr.
slope 1
slope 2
-7
-8
-8
-2.5
-2
-1.5
-1
-2.5
log10(h)
-2
-1.5
-1
log10(h)
(a) 2D case
(b) 3D case
1
1
0
0
-1
-1
-2
-2
log10(Error L2 norm)
log10(Error L2 norm)
Figure 12. Convergence of the discretization error in energy norm.
-3
-4
-5
-6
p=1, standard
p=2, standard
p=1, aggr.
p=2, aggr.
slope 2
slope 3
-7
-8
-3
-4
-5
-6
p=1, standard
p=2, standard
p=1, aggr.
p=2, aggr.
slope 2
slope 3
-7
-8
-2.5
-2
-1.5
log10(h)
(a) 2D case
-1
-2.5
-2
-1.5
-1
log10(h)
(b) 3D case
Figure 13. Convergence of the discretization error in L2 norm.
number of the resulting system matrix scales with the inverse of the square of the size of the background
mesh as in standard FE methods, 3) the penalty parameter of Nitsche’s method is bounded from above,
and 4) the optimal FE convergence order is recovered. These theoretical results are confirmed with 2D
and 3D numerical experiments using both first and second order interpolations.
AGGREGATED UNFITTED FINITE ELEMENT METHOD
19
References
[1] S. Badia, F. Nobile, and C. Vergara.
Fluid–structure partitioned procedures based on
Robin transmission conditions. Journal of Computational Physics, 227(14):7027–7051, 2008.
doi:10.1016/j.jcp.2008.04.006.
[2] D. Kamensky, M.-C. Hsu, D. Schillinger, J. A. Evans, A. Aggarwal, Y. Bazilevs, M. S. Sacks, and
T. J. R. Hughes. An immersogeometric variational framework for fluid–structure interaction: Application to bioprosthetic heart valves. Computer Methods in Applied Mechanics and Engineering, 284:
1005–1053, 2015. doi:10.1016/j.cma.2014.10.040.
[3] T. Belytschko, N. Moës, S. Usui, and C. Parimi. Arbitrary discontinuities in finite elements. International Journal for Numerical Methods in Engineering, 50(4):993–1013, 2001. doi:10.1002/10970207(20010210)50:4<993::AID-NME164>3.0.CO;2-M.
[4] E. Burman, S. Claus, P. Hansbo, M. G. Larson, and A. Massing. CutFEM: Discretizing geometry
and partial differential equations. International Journal for Numerical Methods in Engineering, 104
(7):472–501, 2015. doi:10.1002/nme.4823.
[5] R. Mittal and G. Iaccarino. Immersed Boundary Methods. Annual Review of Fluid Mechanics, 37
(1):239–261, 2005. doi:10.1146/annurev.fluid.37.061903.175743.
[6] D. Schillinger and M. Ruess. The Finite Cell Method: A Review in the Context of Higher-Order
Structural Analysis of CAD and Image-Based Geometric Models. Archives of Computational Methods
in Engineering, 22(3):391–455, 2014. doi:10.1007/s11831-014-9115-y.
[7] Y. Sudhakar, J. P. Moitinho de Almeida, and W. A. Wall. An accurate, robust, and easy-to-implement
method for integration over arbitrary polyhedra: Application to embedded interface methods. Journal
of Computational Physics, 273:393–415, 2014. doi:10.1016/j.jcp.2014.05.019.
[8] J. Parvizian, A. Düster, and E. Rank. Finite cell method: H- and p-extension for embedded domain
problems in solid mechanics. Computational Mechanics, 41(1):121–133, 2007. doi:10.1007/s00466007-0173-y.
[9] Y. Saad. Iterative Methods for Sparse Linear Systems. Other Titles in Applied Mathematics. Society
for Industrial and Applied Mathematics, 2003.
[10] W. Briggs, V. Henson, and S. McCormick. A Multigrid Tutorial, Second Edition. Other
Titles in Applied Mathematics. Society for Industrial and Applied Mathematics, 2000.
doi:10.1137/1.9780898719505.
[11] S. Badia, A. F. Martín, and J. Principe. Multilevel Balancing Domain Decomposition at Extreme
Scales. SIAM Journal on Scientific Computing, pages C22–C52, 2016. doi:10.1137/15M1013511.
[12] A. Menk and S. P. A. Bordas. A robust preconditioning technique for the extended finite element method. International Journal for Numerical Methods in Engineering, 85(13):1609–1632, 2011.
doi:10.1002/nme.3032.
[13] L. Berger-Vergiat, H. Waisman, B. Hiriyur, R. Tuminaro, and D. Keyes. Inexact Schwarz-algebraic
multigrid preconditioners for crack problems modeled by extended finite element methods. International Journal for Numerical Methods in Engineering, 90(3):311–328, 2012. doi:10.1002/nme.3318.
[14] B. Hiriyur, R. Tuminaro, H. Waisman, E. Boman, and D. Keyes. A Quasi-algebraic Multigrid Approach to Fracture Problems Based on Extended Finite Elements. SIAM Journal on Scientific Computing, 34(2):A603–A626, 2012. doi:10.1137/110819913.
[15] F. de Prenter, C. V. Verhoosel, G. J. van Zwieten, and E. H. van Brummelen. Condition number
analysis and preconditioning of the finite cell method. Computer Methods in Applied Mechanics and
Engineering, 316:297–327, 2017. doi:10.1016/j.cma.2016.07.006.
[16] S. Badia and F. Verdugo. Robust and scalable domain decomposition solvers for unfitted finite element
methods. Computational and Applied Mathematics, In press, 2017.
[17] S. Badia, A. F. Martín, and J. Principe. FEMPAR: An object-oriented parallel finite element framework. Archives of Computational Methods in Engineering, In press, 2017.
[18] E. Burman and P. Hansbo.
Fictitious domain finite element methods using cut elements: II. A stabilized Nitsche method. Applied Numerical Mathematics, 62(4):328–341, 2012.
doi:10.1016/j.apnum.2011.01.008.
[19] C. Helzel, M. Berger, and R. Leveque. A High-Resolution Rotated Grid Method for Conservation
Laws with Embedded Geometries. SIAM Journal on Scientific Computing, 26(3):785–809, 2005.
AGGREGATED UNFITTED FINITE ELEMENT METHOD
20
doi:10.1137/S106482750343028X.
[20] F. Kummer. Extended Discontinuous Galerkin methods for multiphase flows: The spatial discretization. Center for Turbulence Research Annual Research Briefs, pages 319–333, 2013.
[21] O. Marco, R. Sevilla, Y. Zhang, J. J. Ródenas, and M. Tur. Exact 3D boundary representation in
finite element analysis based on Cartesian grids independent of the geometry. International Journal
for Numerical Methods in Engineering, 103(6):445–468, 2015. doi:10.1002/nme.4914.
[22] S. C. Brenner and R. Scott. The Mathematical Theory of Finite Element Methods. Springer, softcover
reprint of hardcover 3rd ed. 2008 edition, 2010.
[23] R. Becker. Mesh adaptation for Dirichlet flow control via Nitsche’s method. Communications in
Numerical Methods in Engineering, 18(9):669–680, 2002. doi:10.1002/cnm.529.
[24] J. Nitsche. Über ein Variationsprinzip zur Lösung von Dirichlet-Problemen bei Verwendung von
Teilräumen, die keinen Randbedingungen unterworfen sind. Abhandlungen aus dem Mathematischen
Seminar der Universität Hamburg, 36(1):9–15, 1971. doi:10.1007/BF02995904.
[25] D. N. Arnold, F. Brezzi, B. Cockburn, and L. D. Marini. Unified analysis of discontinuous Galerkin
methods for elliptic problems. SIAM Journal on Numerical Analysis, 39(5):1749–1779, 2002.
[26] S. Agmon and N. S. F. (U.S.). Lectures on elliptic boundary value problems. Van Nostrand mathematical studies. Van Nostrand, 1965.
[27] H. C. Elman, D. J. Silvester, and A. J. Wathen. Finite elements and fast iterative solvers: with
applications in incompressible fluid dynamics. Oxford University Press, 2005.
[28] A. Ern and J.-L. Guermond. Theory and Practice of Finite Elements. Springer, 2004.
[29] D. Arnold. An Interior Penalty Finite Element Method with Discontinuous Elements. SIAM Journal
on Numerical Analysis, 19(4):742–760, 1982. doi:10.1137/0719052.
[30] Intel MKL PARDISO - Parallel Direct Sparse Solver Interface. https://software.intel.com/
en-us/articles/intel-mkl-pardiso.
Department of Civil and Environmental Engineering.
Girona 1-3, Edifici C1, 08034 Barcelona, Spain.
Universitat Politècnica de Catalunya, Jordi
CIMNE – Centre Internacional de Mètodes Numèrics en Enginyeria, Parc Mediterrani de la Tecnologia,
UPC, Esteve Terradas 5, 08860 Castelldefels, Spain.
| 5cs.CE
|
Why It’s Nice to be Quoted: Quasiquoting for Prolog
Jan Wielemaker1 and Michael Hendricks2
arXiv:1308.3941v1 [cs.PL] 19 Aug 2013
1
Web and Media group, VU University Amsterdam,
De Boelelaan 1081a, 1081 HV Amsterdam, The Netherlands,
[email protected]
2
Hendricks Solutions, LLC, Hanna, Wyoming, USA,
[email protected]
Abstract. Prolog’s support for dynamic programming, meta programming and
text processing using context free grammars make the language highly suitable
for defining domain specific languages (DSL) as well as analysing, refactoring
or generating expression states in other (programming) languages. Well known
DSLs are the DCG (Definite Clause Grammar) notation and constraint languages
such as CHR. These extensions use Prolog operator declarations and the {. . . }
notation to realise a good syntax. When external languages, such as HTML, SQL
or JavaScript enter the picture, operators no longer satisfy for embedding snippets
of these languages into a Prolog source file. In addition, Prolog has poor support
for quoting long text fragments.
Haskell introduced quasi quotations to resolve this problem. In this paper we
‘ported’ the Haskell mechanism for quasi quoting to Prolog. We show that this
can be done cleanly and that quasi quoting can solve the above mentioned problems.
1 Introduction
Prolog is commonly used for tasks where it needs to manage snippets of code written
in the syntax of an external language such as HTML, SQL or JavaScript. Such code
snippets often do not comply with the standard Prolog syntax for various reasons [5]:
– Using operator syntax, with only prefix, infix and postfix operators are insufficient.
For example, consider the JavaScript case statement.
– Basic lexical primitives are incompatible. For example, consider textual content in
HTML documents.
– The Prolog term that results from parsing an expression is insufficient for recovering the intent in the target language. For example, consider identifier names that
start with an uppercase letter, which will read as a Prolog variable.
In [5], we claim that some simple languages can be covered well by tweaking the
Prolog syntax using operators, while it is possible to realise acceptable representations
for some other, e.g., XML based languages by using a Prolog syntax to represent the
(simple) data model of these language. In other cases, embedding of external languages
can be supported partially using semantic transformations. For example, calls to Prolog
predicates whose functor matches a table in an RDBMS and whose arity matches the
Proceedings of WLPE 2013, arXiv:1308.2055, August 2013.
2
Jan Wielemaker and Michael Hendricks
number of columns can be translated into SQL queries [2]. In [5], we propose a number
of syntactic extensions to broaden the range of languages that can be supported using
one of the above means.
In this paper, we propose a solution for complex external languages for which none
of the above achieves a satisfactory embedding. The solution is called quasi quotations,
where we borrow the integration into the language from Haskell [4]. A Haskell quasi
quotation, as a syntactic element, has a syntax identifier (a function symbol) and a
snippet of quoted text. The function associated with the syntax identifier is called during
parsing and can manipulate the text and massage it to fit further processing by Haskell.
It is good practice for the function to create an abstract syntax tree by parsing the text
according to the rules of the external language. As we will see in this paper, more
lightweight approaches can also be sufficient.
This paper is organised as follows. After examining related work, we revisit the
problem statement and describe why current support in Prolog is insufficient to solve
this problem in a satisfactory way. In section 6 we introduce the notion of quasi quoting
and how it fits into managing strings that present data structures from other languages.
In section 7, we introduce the concrete solution proposed for SWI-Prolog. This is followed by examples, future work and the conclusions.
2 Related work
Quasi quotations are used for embedding snippets in many ‘scripting’ languages because they form an easy to understand mechanism to create long strings or ASTs (Abstract Syntax Tree) with embedded variables that are given values determined by the
environment. Pure textual replacement carries the risk of injection attacks, which triggered research towards more safe template engines.3 Notably modern scripting languages such as Python and Ruby provide a rich pallete of such techniques. This paper follows the reasoning and technique described in “Why It’s Nice to be Quoted:
Quasiquoting for Haskell” [4]. We refer to this article for further background information. Another interesting system is Camlp4 [1], which provides parsing and ‘anti
parsing’ infrastructure for OCaml. Both approaches allows for extending the host language. The main difference with our approach is that we merely propose a hook into
the Prolog parser that enable smart processing of quasi quotations without dictating how
these quotations are processed.
3 The need to embed snippets of external languages in Prolog
We will explain this need by an example from SWI-Prolog’s web page generating facilities [6]. Web pages are by no means the only place where these problems arise [4],
but the domain is quite familiar to many programmers and is of particular interest to us
3
TAL (http://www.owlfish.com/software/simpleTAL/tal-guide.html)
Cheetah (http://www.cheetahtemplate.org/examples.html)
Django (https://docs.djangoproject.com/en/dev/ref/templates/)
Jinja2 (http://jinja.pocoo.org/docs/templates/).
Why It’s Nice to be Quoted: Quasiquoting for Prolog
3
because generating web pages is an important application area for SWI-Prolog. A modern web page typically consists of HTML that creates the initial DOM structure, CSS
that provides styling and JavaScript for creating interactive components. Our solution
for generating HTML is similar to PiLLoW [3] and based on representing the HTML
DOM structure as a nested Prolog term. The SWI-Prolog HTML page generation library provides good solutions for
–
–
–
–
Generating the page structure.
Definition and reuse of DCGs that generate page fragments.
Create references (URLs) to other pages on the server.
Modular inclusion of resources, such as style and JavaScript files including dependency tracking and automatic ordering by computing a partial order based on the
dependencies.
Pages generated with this infrastructure are guaranteed to be syntactically correct,
regardless of the data from which the page is created. This built-in protection against
injection attacks is considered a key asset of the server infrastructure and a property that
we wish to maintain when integrating JavaScript into the page.
In earlier publications [5, 6] we already identified two missing pieces: (1) longer
(HTML) text fragments and (2) JavaScript fragments. The first problem can be solved
using PWP4 . PWP was developed by Richard O’Keefe and can be considered the opposite of the above described Prolog based page generator. The page is described in XML
and reserved XML elements and tags trigger the generation of dynamic page components based on the evaluation of Prolog goals. Like the Prolog page generator, PWP
protects against injection attacks because it considers solutions from Prolog as data that
is used to extend the XML DOM, which is subsequently serialized for generating the
final page. PWP simplifies the specification of pages with large amounts of text, but the
rule format is less natural and it lacks the modularity of the Prolog based generator.
We have been struggling with JavaScript, trying various approaches. Currently,
there are a large number of JavaScript frameworks that provide ‘widgets’ for use on
web pages. Popular examples are YUI5 and jQuery6 . Including these widgets typically
demands creating a DOM structure that carries an id attribute and represents the data
(e.g., a menu may be represented as a ul list or an advanced text editor can be represented as a textarea). This DOM element is transformed into the target widget
by calling a JavaScript initialisation function that modifies the DOM of the target element and installs event handlers that provide the dynamic behavior of the element.
The initialisation function often takes a configuration object (a JavaScript object literal, {name:value, . . . }), where some of the values are callback functions, represented
as JavaScript lambda functions.
The HTML framework is well equipped to create the initial DOM and pull in the
JavaScript and CSS resources using its dependency framework. The initialisation call
4
5
6
Prolog Well-formed Pages, http://www.swi-prolog.org/pldoc/package/
sgml.html
http://yuilibrary.com/
http://jquery.com/
4
Jan Wielemaker and Michael Hendricks
is easy enough to abstract, except when the configuration object contains lambda functions. These are often required to massage JSON data from the server into the desired
visualisation and support AJAX based operations, such as computing auto-completion
candidates. It is quite common for such objects to require dozens of lines of JavaScript.
Representing JavaScript lambda functions requires support for the complete complexity of the language. One way to do this is to represent the JavaScript fragment as a list
of literals, mixed with DCG rules that insert content generated from Prolog, such as
URLs or JavaScript literal objects. This leads to code as below, which we consider hard
to type on a keyboard, hard to read and, because it is so hard to read, often subject to
syntax errors. Finally, the code below is subject to injection attacks, unless we hand the
variables URL and Id to a grammar that generates valid JavaScript string content from
any Prolog atom.
[ ’$.ajax({ url: "’,URL,’",\n’,
’
data: { id: "’,Id,’",\n’,
’
}\n’,
’
});\n’
]
We tried several designs to improve on the above, none of which we considered satisfactory. Below are the main directions that we tried.
– Abstract away. This implies using the HTML page generation facilities to generate
e.g., make a JavaScript call, initialize a variable from Prolog data, etc. To remain
managable, a fairly high level of abstraction is needed that is geared towards the
JavaScript framework used. This is problematic because it makes it hard for the
programmer to relate the JavaScript examples from the framework documentation
to the Prolog code. Although it is possible to abstract some of the lambda functions,
there is too much variation to deal with all of them.
– Put the JavaScript in a separate file. This creates many short files that typically
only support a specific generated HTML page because the details of the required
JavaScript, such as locations on the server that must be addressed vary from page to
page. Because there is no formal relation between the two pages, it is hard to relate
them and keep them in sync.
– Create application specific JavaScript resources that can be configured without the
need for lambda functions in the configuration object. The problem here is that the
application developers create their own refinement of the external widget library
that requires understanding and documentation, while the externally provided libraries are already quite high level. In other words, the newly created layer adds
mostly new complications for managing and understanding the code.
We have come to the conclusion that (1) we need a mechanism that allows for
including JavaScript into the Prolog page generation source code, (2) the representation
of JavaScript in the Prolog source must be easy to type on a keyboard and understand,
(3) the generated JavaScript should be safeguarded against injection attacks without
explicit calls to encode data and (4) syntactic validation of the generated fragments are
likely to improve productivity.
Why It’s Nice to be Quoted: Quasiquoting for Prolog
5
4 Using Prolog syntax
First, we investigated to what extent the syntax extensions proposed in [5] would help
to represent JavaScript naturally using Prolog syntax. We realised support for the empty
argument lists (e.g., name()), array notation and function bodies using {. . . }. These
were resolved after a proposal by Jose Morales, which extends the notion of operators to
lists and curly-bracket blocks. In addition to our hope of improving JavaScript support,
list subscription and curly-bracket attribute lists are in use with B-Prolog and ECLiPSe.
We have realised a prototype that can express a fair deal of the JavaScript syntax.
However, the following problems remain unresolved:
– It requires ! and . to be defined as operators. These operators are known to cause
ambiguity issues.
– Quoting of identifiers may be needed (e.g., ’String’(...))
– A special symbol is needed to distinguish identifiers from strings. (”string” cannot
be used because it is a list of integers).
– Restrictions are needed with regard to spaces. No space is allowed after function
symbols and keywords such as while. In some places additional spaces are required to prevent Prolog from reading two JavaScript tokens as a single Prolog
token.
– Some symbols clash with Prolog. Consider | or %, which play a totally different
role in the Prolog syntax.
– Some JavaScript snippets can be turned into valid Prolog syntax, but the resulting
AST is ambiguous. For example, ++a is different from a++, but both result in the
Prolog term ++(a).
We believe that the result would have been more usable than using Prolog quoted
atoms for short (1-10 lines) JavaScript snippets, that include simple lambda functions
such as making trivial AJAX callbacks. Our main point of doubt is that it is hard to
convey the restrictions to casual Prolog users, which makes it likely to get surprised
by output that differs from the expectations and can only be understood through deep
knowledge of Prolog’s syntax and underlying term representation.
5 The three problems
If we cannot include JavaScript using Prolog syntax with user-defined operators, the
other option is to use (quoted) text. For this, we need to solve three orthogonal problems:
1. Allow for long quoted text. Current (ISO) Prolog only provides single and double
quoted text, which follow the same rules. In particular
– Quoted text cannot span more than one line, unless the newline is escaped with
a backslash.
– The backslash and quote must be escaped with a backslash.
– If good layout of the output is desirable, there is no way to indent continuation
lines in the Prolog source such that the indentation does not show up in the
output.
6
Jan Wielemaker and Michael Hendricks
With these restrictions, simple copy and paste of example code must be followed
by a tedious process of making the text fit the Prolog syntax, while the reverse is
needed to test the code, for example in a JavaScript console.
2. Relate Prolog variables (data) to references in the quoted material.
3. Establish a safe way to embed data into the template string.
These problems are independent. We illustrate this with a small HTML example
that presents a page with the current time. We use an HTML skeleton to make the
comparison fair. In pure Prolog, this could be achieved using the following code snippet:
...,
get_time(Now),
format_time(atom(Date), ’%+’, Now),
print_html([date=Date],
’<h1>My digital clock</h1>\n\
<p>It is now {{date}}</p>’).
The task of print html/2 is to relate the template variable date to the Prolog variable
Date (third item above), determine the lexical context of {{date}} to know the proper
encoding that needs to be done, realise this encoding on the Prolog atom and create the
proper output string. This is the second item on the list above. Ideally, we would like an
error if the HTML fragment is malformed to begin with.
In pure ISO Prolog, all work must be done at runtime. Many Prolog systems provide
goal expansion/2 or similar, which allows for doing the analysis of the template at
compile time. What remains on our wishlist is to (1) get rid of [date=Date] and
(2) get rid of complicated escape sequences for languages that frequently use quotes or
backslashes. We propose to solve both issues using quasi quotations for Prolog.
6 Quasi Quotations
Quasi quotations find their origin in linguistics7 and introduces variables into textual
expressions. They are commonly used in scripting languages. For example:
a=world; echo "Hello $a"
// Unix shell
$a = "world"; print "Hello $a"; // Perl
This approach is natural to the (novice) programmer. Unfortunately it does not work for
Prolog because nothing in the normal Prolog compilation process provides access to
the names of variables. In addition, plain text insertion is a direct invitation to injection
attacks. As described in section 5, the template replacement code must be aware of the
syntax to perform a safe, i.e., properly escaped, insertion of the variable.
Haskell quasi quotes resolve the injection problem (which is described as a typing
problem in [4]) by associating a syntax identifier directly with the quoted data. The
syntax identifier is associated with a function which typically parses the quoted material
into an abstract syntax tree that can represent the target language.
7
http://en.wikipedia.org/wiki/Quasi-quotation
Why It’s Nice to be Quoted: Quasiquoting for Prolog
7
7 Quasi Quotations in Prolog
Realising quasi quotations requires for a syntactic construct that (1) provides long
quoted strings, (2) associates the quotation to a predicate that can act on it according to the requirements of processing the external language and (3) provides access to
the clause’s variable dictionary.
There are few options for adding a syntax extension to Prolog because Prolog ‘symbol’ characters glue together to form an atom to which operator properties can be assigned. According to Ulrich Neumerkel,8 taking sequences of Prolog solo character that
do not form names is a good starting point. Among others, this allows for {|...|}.
Combined with a term that identifies the syntax, we propose the following syntax for
Prolog quasi quotations, where Syntax is an arbitrary (callable) Prolog term:
{|syntax-identifier ||quoted-material |}
In the rather unlikely event that |} needs to be embedded in the quoted material,
this can be realised in two ways: (1) define the predicate that processes the quotation to
respect some escape sequence or (2) use the existing flexibility of the target language
to avoid |}. An example of (1) could be to introduce the mapping \\ → \, \{ → {,
\} → }, after which the user can write |\}. An example of (2) can be to insert a space
between the two if this does not change the semantics or write \u007c} inside a string
if the target language supports \uXXXX escapes inside strings.
Orthogonal to the syntax is the mapping of the quoted material to a Prolog term,
which can be a full AST of the snippet or a simplified representation as used in section 8.2. This mapping is defined by the syntax-identifier, a callable term to which we
refer as Syntax from now on. At the same time, quasi quotation merges (Prolog) variables from the environment into the quoted material. Quasi quotations are processed as
follows:
– The predicate read/1 is modified to recognize the Syntax term and the quoted text.
– After all normal processing is finished, read/1 performs the following call:
call(+SyntaxName, +Content, +SyntaxArgs, +VarDict, -Result)
Here, SyntaxName is the functor of the Syntax term and SyntaxArgs is the list of arguments, i.e., Syntax =.. [SyntaxName|SyntaxArgs]. In section 8, we
will see why it is useful to split the functor from the arguments. Content is an
opaque handle to the quoted material as we will see later. VarDict is Name = Var
list conforming to read term/3 and write term/3. Result is determined by the call
to the deterministic predicate SyntaxName/4. The predicate read/1 inserts Result at
the location of the quasi quotation in the output term.
– The Result term (and thus the quasi quotation) must be a goal (see section 8.3) or
appear as the argument of a goal that processes (often serializes) the result (often
an AST). For example, reply html page/3 in section 8.1 is designed to serialize
the HTML DOM (AST) produced by the HTML parser called by the quasi quoter.
8
https://lists.iai.uni-bonn.de/pipermail/swi-prolog/2013/
010422.html
8
Jan Wielemaker and Michael Hendricks
We provide two support predicates for SyntaxName/4 to process the result. Note
that there are no restrictions on how the called predicate combines the quoted text with
the syntax arguments and variable dictionary to construct the final term.
phrase from quasi quotation(:Grammar, +Content)
Calls the grammar Grammar on the list formed by Content. This predicate uses the
pure input library described in [7] to parse the content. Syntax errors may be raised
using the non-terminal syntax error//1, which produces a precise syntax
location that consists of the file, line number, line position and character count.
with quasi quotation input(+Content, -Stream, :Goal)
Calls Goal on the Prolog stream Stream. The stream position information reflects
the location in the source file, except for the byte count.
8 Examples of using quasi quotes in Prolog
In this section we provide three examples to illustrate quasi quotations. The first example (HTML) is based on the existing SWI-Prolog libraries for parsing and serializing
HTML. The second example demonstrates how JavaScript can be handled safely (but
with limited syntax checking support) by only tokenising the quotation. The third example concerns SPARQL,9 demonstrates the value of the quasi quotation approach in a
scenario where the parsed quasi quotation is used as a Prolog query.
8.1 Quasi quoting HTML
The first, complete, example illustrates safe embedding of long HTML texts into a Prolog web page. The code of the quasi quoter is given in figure 2. First, we give a fully
working webserver based on this quasi quoter in figure 1. We notice that the quoted
material can contain multiple lines, does not require any line endings and may contain
quotes. The content of the quoted material is valid HTML and because an editor can
easily detect the {|html and |} indicators, it is not hard for development tools to provide support for the embedded HTML, such as highlighting, indentation or completion.
The HTML quoter is defined to limit Prolog variables that are replaced to those that
appear as arguments to the html syntax indicator (Date in figure 1). The convention to
pass variables that are subject to replacement explicitly has the following advantages:
(1) it avoids a singleton variable warning on Date10 , (2) Prolog clauses have a relatively large number of variables due to the lack of functional notation which can avoid
variables and destructive assignment which allows for reusing a variable, (3) it makes
the substitution more explicit and (4) it allows the quoter to check that all intended
relacements were made.
The HTML quasi quoter as defined in figure 2 replaces Prolog variables indicated
by its arguments (html(Date)) if they appear as value for an attribute or content of an
element. It performs the following steps:
9
10
http://www.w3.org/TR/sparql11-overview/
This problem can also be resolved by the quasi quoter by removing singleton variables that
appear inside the result term of the quasi quoter.
Why It’s Nice to be Quoted: Quasiquoting for Prolog
9
:- use_module(library(http/thread_httpd)).
:- use_module(library(http/http_dispatch)).
:- use_module(library(http/html_write)).
server(Port) :- http_server(http_dispatch, [port(Port)]).
:- http_handler(/, clock, []).
clock(_Request) :get_time(Now),
format_time(atom(Date), ’%+’, Now),
reply_html_page(
title(’My digital clock’),
{|html(Date)||
<h1>My digital clock</h1>
<p>It is now <span class="time">Date</span>
|}).
Fig. 1. Example web server with embedded HTML
1. Parse the quoted HTML text using load html/3. The option max errors(0)
causes the parser to throw a syntax error exception and abort on the first error.
As this processing happens while reading the source, the HTML syntax error is
reported during compilation and includes line, line position and character count
information.
2. The XML DOM structure is recursively traversed and attributes that have a Prolog
variable as value or content that matches a Prolog variable from Syntax is replaced
by this variable.
When the page is generated (figure 1), binding of the variable Date completes the
XML DOM structure. This structure is serialised by reply html page/3. The serialisation ensures type safety and the generation of correct HTML syntax.11
8.2 Embedding JavaScript
The HTML quasi quoter of the previous section was easily implemented because the
SWI-Prolog infrastructure already contains a parser and serializer for HTML. We do
not have these for JavaScript. We are likely to develop this in the future, but here we
want to illustrate that it is possible to achieve safe template replacement by only using
a tokeniser.
11
The HTML infrastructure has a global option to select between HTML and XHTML serialisation, which implies that HTML in the Prolog source may be serialised as XHTML to the
client.
10
:::::-
Jan Wielemaker and Michael Hendricks
module(html_quasi_quotations, [ html/4 ]).
use_module(library(sgml)).
use_module(library(apply)).
use_module(library(lists)).
use_module(library(quasi_quotations)).
:- quasi_quotation_syntax(html).
html(Content, Vars, Dict, DOM) :include(qq_var(Vars), Dict, QQDict),
with_quasi_quotation_input(
Content, In,
load_html(In, DOM0,
[ max_errors(0)
])),
xml_content(QQDict, DOM0, DOM).
qq_var(Vars, _=Var) :- member(V, Vars), V == Var, !.
xml_content(Dict, [Name], [Var]) :atom(Name),
memberchk(Name=Var, Dict), !.
xml_content(Dict, Content0, Content) :maplist(xml_content_element(Dict), Content0, Content).
xml_content_element(Dict,
element(Tag, Attrs0, Content0),
element(Tag, Attrs, Content)) :- !,
maplist(xml_attribute(Dict), Attrs0, Attrs),
xml_content(Dict, Content0, Content).
xml_content_element(_, Element, Element).
xml_attribute(Dict, Attr=Name, Attr=Var) :memberchk(Name=Var, Dict), !.
xml_attribute(_, Attr, Attr).
Fig. 2. Source for the HTML quasi quoter
Because JavaScript is generated as part of the HTML page generation, the
JavaScript quasi quoter produces output for the HTML backend. It translates the
JavaScript into a list of two types of elements: (1) plain atoms (that will be emitted
in the context of a script element) and calls to a grammar js expression//1,
which defines a translation of native Prolog data into JavaScript literals according to
table 1.
The quasi quoter tokenises the quoted material using an ECMAScript compliant
tokeniser, implemented using a Prolog grammar. The quoter (figure 4) replaces identi-
Why It’s Nice to be Quoted: Quasiquoting for Prolog
Prolog
number
atom
@true
@false
@null
List
object(NameValueList)
{ Name:Value, ...}
11
JavaScript
number
string (escaped using JavaScript syntax)
boolean true
boolean false
null constant
array
object literal
object literal
Table 1. Prolog to JavaScript conversion
fier tokens that match with a Prolog variable with a call to js expression//1, and
translates the remainder into plain atoms. Figure 3 shows a shortened predicate from the
SWI-Prolog website that applies the JavaScript quasi quoter to initialise a jQuery widget called tagit. The actual predicate contains a larger configuration object, just ‘more of
the same’. In this example, the first block is used to compute server URLs and properties
for Obj that we need in the remainder. Next, the html//1 call creates the DOM needed
for the tagit widget. Here, we could also have used the HTML quasi quoter. The choice
is rather arbitrary in this case because the fragment is short and defines only structure
and data that is passed in from a variable. Next, we see html requires//1, which
ensures that the page head is extended to load the required JavaScript and CSS resources
and finally, there is the script fragment with embedded Prolog variables. Because Prolog
variables are valid JavaScript identifiers, the fragment contains valid JavaScript syntax.
8.3 Embedding SPARQL
SPARQL12 is the query language for the semantic web RDF language. In most of this
section, one may replace SPARQL with SQL. ClioPatria is SWI-Prolog’s semantic web
framework, which contains a SPARQL ‘endpoint’. The SPARQL engine compiles a
SPARQL query into a Prolog query, optimizes and executes this query and serializes
the results according to the SPARQL result specification.
When writing middleware (reasoning) in the ClioPatria, one typically uses direct
queries to the embedded RDF store for the reasoning. In some cases, one would like
to use SPARQL for specifying the query. Consider cases where the query is already
available in SPARQL, the query must also be used with external servers or the author
is much more fluent in SPARQL than in Prolog. The quasi quotation syntax can be
used to write down the code below. The SPARQL query is parsed at compile time and
the SPARQL projection variables can naturally integrate with the Prolog variable of
the enclosing clause. Note that in this example, we think it is better not to pass the
substituted variables as arguments to the sparql syntax term because this list is made
explicit in the variable projection clause that starts the query.
12
http://www.w3.org/TR/sparql11-overview/
12
Jan Wielemaker and Michael Hendricks
prolog:doc_object_page_footer(Obj, _Options) -->
{ http_link_to_id(complete_tag, [], Complete),
http_link_to_id(show_tag, [], OnClick),
http_link_to_id(remove_tag, [], Remove),
object_id(Obj, ObjectID),
object_tags(Obj, Tags),
atomic_list_concat(Tags, ’,’, Data)
},
html(div(class(’user-annotations’),
input([id(tags), value(Data)]))),
html_requires(tagit),
js_script({|javascript(Complete, OnClick, ObjectID, Remove)||
$(document).ready(function() {
$("#tags").tagit({
autocomplete: { delay: 0.3,
minLength: 1,
source: Complete
},
onTagClicked: function(event, ui) {
window.location.href = OnClick+"tag="+
encodeURIComponent(ui.tagLabel);
},
beforeTagRemoved: function(event, ui) {
$.ajax({ dataType: "json",
url: Remove,
data: { tag: ui.tagLabel,
obj: ObjectID
}
});
}
});
});
|}).
Fig. 3. Shortened code fragment from the SWI-Prolog website that illustrates the embedding of
JavaScript for initializing a widget.
...,
{|sparql||
SELECT ?Name, ?Place WHERE {
...
}
|}>
Why It’s Nice to be Quoted: Quasiquoting for Prolog
13
javascript(Content, Vars, Dict, \Parts) :include(qq_var(Vars), Dict, QQDict),
phrase_from_quasi_quotation(
js(QQDict, Parts), Content).
qq_var(Vars, _=Var) :- member(V, Vars), V == Var, !.
js(Dict, [Pre, Subst|More]) -->
here(Here0), js_tokens(_), here(Here1),
js_token(identifier(Name)),
{ memberchk(Name=Var, Dict), !,
Subst = \js_expression(Var),
diff_to_atom(Here0, Here1, Pre)
},
js(Dict, More).
js(_, [Last]) -->
string(Codes), \+ [_], !,
{ atom_codes(Last, Codes) }.
js_tokens([]) --> [].
js_tokens([H|T]) --> js_token(H), js_tokens(T).
%!
%
%
%
%
diff_to_atom(+Start, +End, -Atom)
True when Atom is an atom that represents the
characters between Start and End, where End must
be in the tail of the list Start.
diff_to_atom(Start, End, Atom) :diff_list(Start, End, List),
atom_codes(Atom, List).
diff_list(Start, End, List) :- Start == End, !, List = [].
diff_list([H|Start], End, [H|List]) :diff_list(Start, End, List).
here(Here, Here, Here).
Fig. 4. Partial source for the javascript syntax quoter.
A similar approach is feasible with SQL, offering a more convenient way to interact
with the database than using prepared statements and both a safer and more convenient
way than direct text-based SQL queries.
14
Jan Wielemaker and Michael Hendricks
9 Implementation
The implementation of quasi quotation handling in an existing Prolog system is straightforward and should not require more than a couple of days. The tokeniser is extended
to recognise {|. On encountering this token, the parser builds a list of terms, each of
which represents a quasi quotation. If the list of quasi quotations is not empty when
read/1 reads the fullstop token, it materializes the variable dictionary and calls a routine that calls the quasi quotation parsers. Note that quasi quotations are parsed after
read/1 completes reading the term. This allows for quasi quotations to refer to variables
that appear after the quasi quotation (see section 8.3). This does imply that read/1 must
buffer the quoted text. The functionality is made available to the user through the library
quasi quotations.pl, which appeared in SWI-Prolog 6.3.17.
The Prolog tokeniser is extended with two tokens. {| is a new token that starts a
quasi quotation. The sequence ||...|} is processed as a single token.
10 Future work
Quasi quotations are young in SWI-Prolog. We believe that the core functionality described in this article is largely future-proof.
One of the future tasks is to establish libraries that facilitate syntax handling and
safe replacement. Another is to establish design patterns for using this technology. What
can be learn from the Haskell, Python and Ruby communities here? Quasi quotations
typically ‘evaluate’ to a Prolog term that shares variables with the clause into which it is
embedded. We do not have a functional syntax that we can exploit to force ‘evaluation’
of this term with instantiated variables. On the other hand, we can also use the nonground abstract syntax tree and use it for e.g., matching tasks.
11 Conclusions
In this article we have motivated why it is necessary to be able to embed longer snippets
of code written in another language in Prolog source code. This requires for a syntax that
permits embedding, practically without the need for Prolog escaping in the embedded
text. Next, this text must be related to Prolog data. The integration of Prolog data into the
snippet must be done according to the syntax and data model of the external language.
This implies we need a programmable component that is related to the quoted material.
Combining these requirements into one syntactic extension that is executed by
read/1 simplifies support by tooling such as editors, provides natural access to Prolog
variables and allows for seemless integration of error messages.
We have demonstrated quasi quotations using two implemented quoters. These
quoters allow for easy copying and pasting material in their native syntax to and from
Prolog source. Using these quoters is likely to reduce the learning curve for embedding
snippets into Prolog, while the quasi quoter can guarantee that the integrated material
is (syntactically) correct and that Prolog material is safely integrated.
Why It’s Nice to be Quoted: Quasiquoting for Prolog
15
Acknowledgements
Jan Pobrislo provided the insights that were used to write section 5. Michiel Hildebrand
and Jacco van Ossenbruggen (VU University Amsterdam) have tried many of the pure
Prolog based alternatives which were needed to formulate section 3 and section 4. They
also help shaping the current JavaScript quasi quoter and commented on drafts of this
text. Ulrich Neumerkel has propose the combination of | with brackets and pointed at
the tokenisation issues described in section 9.
This publication was supported by the Dutch national program COMMIT/
References
1. Jake Donham and Nicolas Pouillard. Camlp4 and template haskell. In ACM SIGPLAN Commercial Users of Functional Programming, CUFP ’10, pages 6:1–6:1, New York, NY, USA,
2010. ACM.
2. Christoph Draxler. Accessing relational and higher databases through database set predicates
in logic programming languages. Phd thesis, Zurich University, 1991.
3. Daniel Cabeza Gras and Manuel V. Hermenegildo. Distributed WWW programming using
(ciao-)prolog and the piLLoW library. TPLP, 1(3):251–282, 2001.
4. Geoffrey Mainland. Why it’s nice to be quoted: quasiquoting for haskell. In Gabriele Keller,
editor, Haskell, pages 73–82. ACM, 2007.
5. Jan Wielemaker and Nicos Angelopoulos. Syntactic integration of external languages in Prolog. In Proceedings of WLPE 2012, 2012.
6. Jan Wielemaker, Zhisheng Huang, and Lourens van der Meij. Swi-prolog and the web. TPLP,
8(3):363–392, 2008.
7. Jan Wielemaker and Ulrich Neumerkel. Precise garbage collection in Prolog. In Proceedings
of CICLOPS 2008, 2008.
| 6cs.PL
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.